name
stringlengths
5
6
title
stringlengths
8
144
abstract
stringlengths
0
2.68k
fulltext
stringlengths
1.78k
95k
keywords
stringlengths
22
532
592966
Continuous Data Block Placement in and Elevation from Tertiary Storage in Hierarchical Storage Servers.
Given the cost of memories and the very large storage and bandwidth requirements of large-scale multimedia databases, hierarchical storage servers (which consist of disk-based secondary storage and tape-library-based tertiary storage) are becoming increasingly popular. Such server applications rely upon tape libraries to store all media, exploiting their excellent storage capacity and cost per MB characteristics. They also rely upon disk arrays, exploiting their high bandwidth, to satisfy a very large number of requests. Given typical access patterns and server configurations, the tape drives are fully utilized uploading data for requests that fall through to the tertiary level. Such upload operations consume significant secondary storage device and bus bandwidth. In addition, with present technology (and trends) the disk array can serve fewer requests to continuous objects than it can store, mainly due to IO and/or backplane bus bandwidth limitations. In this work we address comprehensively the performance of these hierarchical, continuous-media, storage servers by looking at all three main system resources: the tape drive bandwidth, the secondary-storage bandwidth, and the host's RAM. We provide techniques which, while fully utilizing the tape drive bandwidth (an expensive resource) they introduce bandwidth savings, which allow the secondary storage devices to serve more requests and do so without increasing demands for the host's RAM space. Specifically, we consider the issue of elevating continuous data from its permanent place in tertiary for display purposes. We develop algorithms for sharing the responsibility for the playback between the secondary and tertiary devices and for placing the blocks of continuous objects on tapes, and show how they achieve the above goals. We study these issues for different commercial tape library products with different bandwidth and tape capacity and in environments with and without the multiplexing of tape libraries.
Introduction Future multimedia information systems are likely to contain large collections of delay sensitive data objects (e.g. audio and video) with various lengths and requirements. Some of the main data characteristics which present serious challenges when building multimedia servers are: their large size, the delay-sensitivity of their data, and their high display bandwidth requirements. Multimedia objects can be very large in size: As one example, 100 minutes of MPEG-2 video may require up to 6 GB of storage. Thus, the storage requirements of future continuous media servers for many different applications will easily exceed several terabytes. The cost of memory strongly depends on the type of memory being em- ployed. Currently, for RAM memory the cost is about $1.0 per MB; for individual magnetic disk storage, the cost is about $0.1 per MB (see for ex- ample, the cost of high performance disk drives, such as the Quantum Atlas (http://www.quantum.com)and the Seagate Cheetah (http://www.seagate.com)). When a large number of disk drives are grouped using disk array technology to store large amounts of data, the cost for disk storage becomes greater than $0.3 per MB (see for example Maximum Strategy's Gen-5 disk array product (http://www.maxstrat.com)). For magnetic tape storage the cost is less $0.004 per MB. For higher-end tape libraries, the tape storage cost per MB ranges from per MB (e.g., for the high-performance Ampex DST812 library with 50GB tapes (http//www.ampex.com/html)) to significantly lower figures for 330GB tapes. These "capital" costs for different types of memories should also include an organization's costs for managing their storage systems, which can be quite significant: for disk storage it is reported to be $7 per MB; for tape library storage, due to its automated functions, this cost is minimal [Sim97]. As an indication, in absolute dollar costs, a company's capital investment for the storage required for a full-fledged video server (which for a few thousand videos is currently around 12TB) would be around $400,000 if a high-performance tape library is chosen (such as the 12.8TB Ampex DST812 library with four drives and 256 50GB tapes) or more than $4,000,000 for eight Gen-5 disk arrays (each storing about 1.7TB). Given that small and mid-sized companies are reported to be willing to spend annually less than $50,000 and $120,000 for mass storage, respectively the above cost difference ensures a dominant place in the market for tape libraries. In fact, the annual tape library market has been reported as growing at a pace of 34%, with some submarkets growing with a pace around 50% [Sim97]. Thus, tape library storage is very attractive when storage cost is of primary concern. On the other hand, the access speeds of memory units point to a different direction. Tertiary storage is characterized by very slow access times, in the order of tens of seconds (or up to minutes in lower-end products) for magnetic tape libraries. Magnetic disks have access times in the order of tens of milliseconds, whereas RAM memories' access times are in the order of tens of nanoseconds. The above are strong arguments for employing a hierarchy of memory tech- nologies, forming a Hierarchical Storage Management System (HSMS), including RAM-based Primary Storage (PS), magnetic disk-array-based Secondary Storage (SS), and tape-based Tertiary Storage (TS). A key idea in a HSMS is to store all objects in the inexpensive TS, from where they will be elevated to the SS and PS levels before being accessed. This solves the problems owing to the large storage space requirements of continuous media servers and the costs of memory units. A complementary idea is to use the higher levels as a cache for the levels below; all objects will reside permanently on TS, while popular objects, for ex- ample, will reside for long periods of time in SS and portions of them will reside for long periods of time in PS. This idea addresses the problems owing to the memories' access times, and to the delay-sensitivity and high display bandwidth requirements of many continuous objects. Given the reported market growth, tape libraries are becoming increasingly popular for the low level of such a HSMS. In fact, with respect to continuous media applications, several companies have developed video storage solutions based on tape libraries for applications such as broadcasting [Hen97]. In most envisaged solutions, tape libraries are occupying the lowest level of the hierarchy, are used mainly for storing the vast amounts of data, exploiting their excellent cost per MB characteristics, while disk arrays are used for their high bandwidth in order to serve as many requests as possible. At the same time, colder (hoter) data objects migrate from disks (tapes) to tapes (disks). With this in mind, this paper addresses how to cleverly exploit the bandwidth offered by Table Storage and display parameters explanation typical values bD display bandwidth (consumption rate) of blocks in an object varies; user-defined s block size varies; user-defined d "time display time for 1 block varies; depends on s bT tape drive bandwidth 0.5 - 20 MB/sec - # of tape drives 1 - 64 e robot exchange time 6 - ?40 sec c switch search - # of tape cartridges r bT =bD 0.5 - 50 of jobs multiplexed varies t # of blocks in a "time slice" (in multiplexing) varies modern tape drives employed in such HSMSs for continuous media applications, so to help increase the overall system's performance. 1.1. Hierarchical Storage Management Systems We will now discuss HSMSs in some detail, paying attention to their specifications and functionality in order to gain the relevant insights and put the contributions of this work in context. A summary of the HSMS's relevant parameters and some typical values is given in Table 1. An appropriate candidate for SS appears to be arrays of magnetic disks [PGK88]. Given the high bandwidth requirements of many multimedia objects and the typical, relatively low, effective bandwidth of magnetic disks (nominal bandwidths are in the range of 4-15 MB/sec), striping techniques [SGM86,BMGJ94,TF98] are likely to prove beneficial, since, when striping objects across several disks, the effective SS bandwidth for these objects is several tape- / CDROM- based elevate Primary Storage (PS) Secondary Storage (SS) disk-based Tertiary Storage (TS) Figure 1. Conceptual model of a Hierarchical Storage Management System (HSMS) times that of a single disk. Currently, average seek and rotational delays are approximately 8-10 and 3-5 msec, respectively. With respect to TS, technological developments have led to the emergence of robot-based magnetic-tape libraries, making them the best candidate for the TS devices, given their desirable cost per megabyte figures. Tape libraries consist of a series of shelves (storing a number of tape cartridges), a number of tape drives onto which the tapes must be loaded before their objects can be accessed, and a number of robot arms (usually 1) which are responsible for loading and unloading the tapes to and from the drives. Despite TS's high data transfer rates (see Table these devices remain comparatively slower than magnetic disks due to the high exchange costs (unload a tape, put it on the shelf, get another tape, load it) and to the cost of searching within a tape (proceeding at a maximum pace of less than 1.6 GB/sec). The conceptual model of a HSMS is illustrated in Figure 1. The conceptual model shows the three levels of the HSMS. Data is elevated one level at a time. In the rest of this paper, the term "elevate" has been reserved for "elevate from TS to SS". Figure 2 illustrates the physical architectural view of the HSMS. The key feature that must be noted here is that PS serves as an intermediate staging area between the TS and the SS device and that an object is not required to be SS-resident in order to be displayed; it can also be displayed from TS. Display consume upload Tape Controller disk array Secondary Storage (SS) robotic tape library Tertiary Storage (TS) Disk Controller Primary Storage (PS) retrieve flush Figure 2. Physical model of a Hierarchical Storage Management System (HSMS) 1.2. Related Work, Motivations, and the Problem 1.2.1. The Big Picture There is great consensus that the I/O subsystem of a computer system has become the performance bottleneck [PGK88,Pat93]. This observation is the motivation for a large body of research in the area of storage servers, aiming to increase the available SS bandwidth in the system. One thread of this research led to the development of disk arrays [PGK88] and to the development of new methods for placing data objects on disk arrays, which attempt to exploit their inherent potential for high performance and reliability [SGM86,PGK88,CP90, Gea94], mainly through data placement techniques, such as striping. Multimedia data, such as video and audio, require very large storage capacities and very large I/O bandwidths, making disk arrays the natural choice for the SS medium. Researchers in multimedia and video storage servers then started paying attention to the placement of multimedia data on SS devices in such a way so to increase the system's performance by effectively utilizing the available SS bandwidth of the system [RV93,LS93,GR,Tea93,BMGJ94,KK95, ORS96,TF98]. Another thread concentrated on developing techniques to reduce the number of required secondary storage I/O's by exploiting main-memory buffer caches and the characteristics of several emerging applications (such as video on demand servers). Through buffer and data sharing techniques, such as batching, bridging, and adaptive piggybacking, and prefetching into smart-disk caches this research attained a reduction in the required number of SS I/O streams required to support video object requests [KRT94,RZ95,GLM96,DS94,Dan95,ASS96,TH99]. Thus, so far related research has concentrated on techniques for cleverly exploiting the PS and SS resources of the system in order to increase its perfor- mance. With respect to tertiary storage, there have been efforts at modelling the performance characteristics of individual tape drives and robotic tape libraries [HS96b,HS96a,JM98a,JM98b] and [Joh96,JM98a,JM98b] respectively, to predict future references and prefetch documents in the SS [KW98] for digital library applications, to derive intelligent placement strategies of data in the library [CTZ97], as well as efforts to derive intelligent scheduling algorithms for multiplexed video streams over tape drives [KMP90,LLW95]. As mentioned in the introduction, disk-based storage is viewed as presently too costly to store the thousands of video objects in a full-fledged video server, and thus TS is employed for storage augmentation purposes. A continuous object can be played either from TS or from SS [Che94, KDST95,GR98,Che98]. One may Play from TS an object by issuing upload requests to TS. The uploaded blocks of the object are placed into PS buffers, from where they are subsequently consumed: 1 either they are transmitted over a network to a remote client, or they are displayed to a local client. However, note that the bandwidth of TS tape drives is typically significantly greater than the display bandwidth of objects. Thus, playing objects directly from TS can in general create serious PS buffer space problems. As a result, a much more preferable choice is to first elevate the object from TS to SS (that is, physically upload it from TS to PS, and then flush it from PS to SS), and subsequently Play from SS the object, by issuing retrieval requests to SS for the blocks to be displayed [GS94,KDST95]. Furthermore, given that very few users (if any) would tolerate high response times, the playback process must be started immediately after enough data has been elevated to SS; hence, typically, the retrieval 1 The terms "consume", "display", "playback" and "play" will be used interchangeably in this paper. operations executing on behalf of the playback process are executed in parallel with the elevate operations which move the future blocks of the object to SS [GDS95]. Let us refer to the above procedure (i.e., elevate from TS to SS, and simultaneously play from SS), which represents the research state of the art, (for displaying continuous objects residing in TS) as the Conventional Play method. 1.2.2. The Problem This paper assumes the above framework of a continuous media server based on a hierarchical storage manager. The research reported here first addresses the important problem of continuous data elevation, from its permanent place in TS to the higher levels [CT95]. In particular, we concentrate on on-demand elevation (i.e., elevation occurs when a display request for the object arrives). On-demand elevation is a non-trivial problem since it must deal with the bandwidth mismatch problem: firstly, the display bandwidth requirements of objects are different from the bandwidths of TS and SS devices; and secondly, the sustained (available, effective) TS bandwidth vary with time, depending on the requirements of the current workload (e.g., whenever multiplexing is employed). An Example System Configuration and Problem Setup Let us consider an example HSMS system configuration which can be employed for a media server for movie-viewing, tele-teaching, digital libraries, and other video-based multimedia applications. At the lower level there is a high performance tape library (based on the Ampex DST812 tape library) with 12.8 of storage, with four drives and 256 50GB tapes, costing about $400,000. The SS level, acts as a cache, and is a high performance disk array (based on Maximum Strategy's Gen-5 array) consisting of 95 magnetic disks, each storing totaling about 10% of the tertiary storage, and costing about $500,000 (or 125% of the tertiary cost). Considering 90 minute MPEG-2 videos that require about 3GB of storage, about 6 videos can fit in a disk. Thus, the SS level can store 570 videos in total. Assuming for simplicity 500KB video disk blocks, each with a display time of one second, with present technology each disk drive of the array, in isolation, can support the uninterrupted display of about 16 videos (e.g., with 12MB/s transfer rate, 8ms average seek cost, 4 ms average rotational delay, and 1 ms track/cylinder switch cost). However, due to the bandwidth limitations in the front-end and back-end I/O buses, the combined data transfer rate of the array is about 250MB/sec, or about 20 times that of a single member drive. Thus, although the whole SS can store 570 videos, it can support the playback of about (20 \Theta 16 =) 320 video requests. Assuming a real-world system, with several hundred video requests in the system at any time, and a Zipf access-request distribution to videos, a few (e.g., the 10-15) most popular videos will be receiving more than 70% of all requests. Thus it is important to create copies of the most popular videos and employ a clever video placement of the copies to the SS disks, so their load can be balanced and serve as many video requests as possible. Given that considerable SS bandwidth is (in addition to sustaining the playback of the SS-resident videos) required for elevating colder video blocks from TS to SS, it is easy to see that all of the system's SS I/O bandwidth can be consumed. In fact, please note that products for servers supporting large numbers of disks are currently constrained by backplane bus limitations, which is many times smaller the combined raw disk bandwidth [www98,www97]. This fact has been recognized by many storage system researchers [RGF98,KP98] and is a key problem on which they focus. In this framework, and using the Conventional Play method, the TS is used for storage augmentation purposes, simply to store the large number of colder videos and elevate them to SS when a request for them arrives. Since typically the TS bandwidth is sufficient to support only few requests that fall through to the TS level [Che98] the TS drives will be fully utilized elevating colder video data to SS. The SS I/O bandwidth (including the backplane bus bandwidth) is also fully utilized supporting retrieval requests for video blocks and elevation requests for TS blocks. As mentioned, even without the bandwidth consumed by the data elevation from TS, the SS stores many more videos than it can support. When SS bandwidth is consumed by the need to elevate colder videos from the TS, the available SS bandwidth decreases significantly. The key goal of this work is to develop techniques which cleverly exploit the bandwidth of the tape drives during elevation so to save IO bandwidth. This saved bandwidth can allow the SS to serve additional requests for some of the other videos it stores and thus improve the overall system performance. This is not an easy task because, despite the high transfer rates of modern tape drives, the average access cost in tape libraries remains very high (currently, about three orders of magnitude higher) due to the high costs for robotic movements and head positioning delays. Thus, we focus on requests for delay-sensitive data elevation from TS to higher levels, with the goals of achieving IO bandwidth savings and, in addition, hiccup-free displays of video streams, and low start-up latencies. The saved IO bandwidth can be used to accept and serve additional requests for SS-resident data. At the same time, we handle the bandwidth mismatch problem during data elevation, in a manner that allows the above savings in IO bandwidth while not requiring any extra PS buffer space. The remainder of the paper is structured as follows. In Section 2 we will present "Alternate Play", a novel algorithm for on-demand data elevation through the levels of the hierarchy and we will discuss its benefits with respect to IO bandwidth savings and its compromises with respect to PS buffer space. In Section 3, we contribute a set of novel techniques which place the blocks of video objects on the tertiary storage media in a manner which alleviates the need for extra PS buffer space and companion play algorithms which achieve the same SS bandwidth savings as the Alternate Play algorithm. Subsequently, we contribute the notion of strips of streams, the use of which makes our techniques applicable even when the display bandwidth requirements are larger than the available TS bandwidth (e.g., for low-end tape drives). In Section 4 we revisit these issues, contributing algorithms and analyses for the case when the tape library is multiplexed across streams (to avoid experiencing unacceptably-long start-up latencies and to be able to offer the SS savings of the earlier sections to many more video requests). In Section 5 we revisit the same issues, only now the assumption of a fixed and known, a priori, multiprogramming degree is removed. In Section 6 we discuss some real-world pitfalls and how our algorithms can be used to avoid them. Finally, in Section 7 we will present the conclusions of this work. 2. Data Elevation with Alternate Play In this section we will present a novel elevation method, achieving IO band-width savings (when compared with the Conventional Play). This new method will be an integral part of many of our algorithms in the rest of this paper. 2.1. Alternate Play The obvious advantage of the Conventional Play method is that it requires no additional PS buffer space. 2 On the other hand, its obvious deficiency, is that SS is additionally taxed as a result of the execution of flush and retrieval operations, reducing thus the available IO bandwidth significantly. Since in our target environment we expect IO bandwidth to be one of the (two) scarcest resources, the above observation must be noted seriously. In the following, we will present some approaches aiming to alleviate this problem. An initial attempt to overcome the shortcomings of the Conventional Play method centers on the following idea: uploaded blocks from TS can be maintained in PS buffers and made available to the playback process from them. This can save significant IO bandwidth. Of course, blocks belonging to popular objects may still be flushed to SS, in addition to maintaining them in PS buffers from where they will be consumed. This alteration of the Play from TS method (i.e., Play from TS, and then flush to SS), when compared with the Conventional Play, saves bandwidth for the current request (since no retrievals from SS are needed). Also, future requests for the same object will not additionally tax the TS or SS (since no uploads or flushes for the SS-resident blocks are needed). Despite these unquestionable benefits, there is an obvious concern regarding the amount of PS buffer space which is needed to realize the aforementioned bandwidth savings. For example, if 2, the PS space requirements are 50% of the size of the entire object to be displayed (since when the last block is uploaded, only half of the blocks will have been consumed). In general, the PS space requirements are (r \Gamma 1)=r of the object's size. Given that PS buffer space is another scarce system resource, care must be taken to use it wisely. In an effort to reconcile this trade-off between the PS space requirements and the IO bandwidth savings, we can use hybrid techniques, so that for some blocks the Conventional Play method - essentially: Play from SS - is employed, while for other blocks a Play from TS method is followed. We refer to the newly- 2 Throughout the paper, "PS buffer space requirements" will refer to the maximum number of PS buffers (each holding 1 block) required at any instance during the display of an object. The buffers holding blocks, which are currently being displayed, to be displayed during the next time unit, or currently being elevated to SS are not counted, since in each of these cases, using a buffer is unavoidable. Table Example of Alternate Play (r =2; B=13) time (elevated) (uploaded) block unit to SS to PS displayed proposed method as the Alternate Play algorithm. For the special case, an example of the Alternate Play algorithm on a 13-block object, playing odd-numbered blocks from TS and even-numbered blocks from SS, is shown in Table 2. Newly arriving blocks have been underlined. The Alternate Play algorithm terminates when all blocks have been read off the TS. The display of the requested object continues by consuming all remaining blocks in strict alternation from PS and SS. At the end, half of the blocks will have been played from TS and half from SS. Two features of the above described (for the Alternate Play algorithm must be noted. First, during its execution (i.e., during the first l time units of the object's display) the SS bandwidth requirements of even time units are twice the SS bandwidth requirements of odd time units; for the remaining time units, the SS bandwidth requirements are equal to those of the odd time units. Second, during the algorithm's execution, every two time units the number of PS buffers (needed to hold the Play from TS blocks) increases by 1; at the end of the algorithm's execution, half of the 'B=2 Play from TS blocks will still be in PS. It thus follows that the PS space requirements are 25% of the object's size. 2.2. Generalizing Alternate Play 2.2.1. Integer r The r=2 Alternate Play algorithm of the last section, can be generalized to handle arbitrary r ratios as follows: In every time unit, r blocks are read from TS. From these r blocks, k blocks are uploaded to PS (and played from there) and the remaining r\Gammak blocks are elevated to SS, for some (arbitrarily/appropriately chosen) This gives us a family of Alternate Play algorithms, one for each different value of k. Pseudo-code for the above described Alternate Play is given as Algorithm 1. The given algorithm uses routines to upload an indicated block to a PS buffer, Algorithm 1 Alternate Play INPUT: Object's blocks in TS next upld blk next displ blk upload(next upld blk++) for time unit i:= 1 to l r do parbegin 1. if 1 then Play-from-SS(next displ blk++) else consume(next displ blk++) 2. for j:= 1 to r \Gamma k do elevate(next upld blk++) for j:= 1 to k do upload(next upld blk++) parend or to elevate (i.e., upload and then flush) an indicated block to SS. The parbegin/parend construct is intended to indicate that the display (consume and Play-from-SS) of a block occurs in parallel with the readings (elevate and upload) of the next two blocks from TS. The reader may verify that Table 2 is a trace of Algorithm 1 for the r=2, k=1 case. The obvious question now is "what is the best value of k?". To answer this question, we will examine the impact of the choice of k on the SS bandwidth savings and on the required PS buffer space. For any value of k, the SS bandwidth savings are Bk=r blocks (since for k out of every r blocks neither a flush to, nor a retrieve from, SS is required). Also for any value of k, the PS buffer space requirements are Bk(r \Gamma 1)=r 2 . This holds because the algorithm runs for B=r time units, during which Bk=r Table Example of brc Alternate Play (r =3:5; k=1; B=22) time (elevated) (uploaded) block unit to SS to PS displayed blocks are uploaded to PS (k blocks per time unit) and Bk=r 2 blocks (= k=r blocks for each of the B=r time units) have been played from TS. Therefore, blocks are still in PS when the algorithm terminates. We thus see that there is a trade-off between SS bandwidth savings and required PS buffer space: the higher the # of blocks played from SS # of blocks played from TS ratio is, the lower the PS space requirements become, but the lower SS bandwidth savings are achieved. The k parameter in the general algorithm acts as a "knob", fine-tuning this trade-off. 2.2.2. Real r The alert reader may have realized that, so far, we have made the implicit assumption that r is an (arbitrary) integer. If r is a real number, as it will most likely be the case in real-world situations, one may apply the above ideas, uploading k out of every brc blocks to PS (and playing them from there). An example is shown in Table 3. Again, newly arriving blocks have been underlined. A block in brackets indicates that only part of the block is uploaded/elevated. Clearly, the SS bandwidth savings of the above outlined algorithm (when compared with the Conventional Play algorithm) are Bk= brc blocks. The PS space requirements of the above algorithm are Bk(r \Gamma 1)=(r brc) Bk=r. This holds because the algorithm runs for B=r time units, displaying B=r blocks. During this time, Bk= brc blocks are uploaded to PS, and Bk=(r brc) blocks (= k= brc blocks for eac h of the B=r time units) are played from TS. Therefore, when the algorithm terminates, the number of blocks still in PS is One might of course chose to upload k out of every dre blocks. Alternatively, if PS space is very scarce, one might chose to upload sometimes k out of every brc blocks and some other times k out of every dre blocks. This method will be further discussed in Section 3.2.2. 3. Data Elevation without Multiplexing The Alternate Play methods of the preceding section (some blocks are played from TS, and some from SS) will be the starting point for the derivation of better playing algorithms. 3.1. APWAT As discussed earlier, the high PS buffer requirements of the Alternate Play method are due to the long time that a PS buffer is dedicated to holding a particular Play from TS block. Observe also that this long time is dependent on r: if r is large, then the Play from TS blocks will occupy a PS buffer for a long time (since they will come into a PS buffer too far ahead of their consumption times). In trying to attain further reductions in the PS space requirements, our key idea is that, by altering the order with which the blocks of an object are recorded on the TS media, large values of r (i.e., bandwidth mismatches between the TS and the display) can be accounted for in a way that reduces the occupancy time of PS buffers by Play from TS blocks. Algorithm 2 is a placement algorithm, which, when given as inputs B and r, Algorithm 2 Twisted Placement (integer r) INPUT: B, integer r OUTPUT: r-twisted sequence of object's blocks (to be placed on TS) for i:=0 to l r randomly place blocks B d B r to blank tape positions determines an r-twisted placement/sequence of an object's blocks locations on a tape. Examples of twisted sequences are given in Figure 3 (the "randomly place blocks . " of Algorithm 2 is the "sequentially (b) (a) Figure 3. (a) 2-twisted and (b) 4-twisted placements of a 13-block object place blocks . " in the given examples). One may now immediately see that a produced r-twisted sequence is such that, for given r, when the object's blocks are read off the tape sequentially, every r-th block (boldfaced in Figure 3) will be uploaded into PS precisely before the time it has to be consumed. These ' B=r blocks can thus be played from TS without any PS buffer space requirements; the remaining blocks will be elevated to SS, and will be played from SS when needed. Algorithm 3 materializes the above idea. The notational conventions of Algorithm 1 have been used here, Algorithm 3 APWAT on r-twisted sequence, when system's bandwidth ratio is r INPUT: r-twisted placement of object's blocks in TS next upld blk next displ blk upload(next upld blk++) for time unit i:= 1 to l r do parbegin 1. consume(next displ blk++) 2. for j:=1 to r \Gamma 1 do elevate(next upld blk++) upload(next upld blk++) parend too. Note that the given Alternate Play With A Twist (APWAT) algorithm is essentially an Alternate Play algorithm (that is, some blocks are played from TS and some others from SS), but applied on a specific arrangement of the object's blocks, so that boldfaced blocks are played from TS. The APWAT algorithm terminates when all blocks have been read off the TS. The display of the requested object continues by consuming one more block from PS, and all remaining blocks exclusively from SS. The SS bandwidth requirements remain constant (r \Gamma 1 blocks per time unit) throughout the algorithm's execution; the display of the remaining blocks requires a constant bandwidth of 1 block per time unit. An example of APWAT's action on the 2-twisted sequence of Figure shown in Table 4. Again, newly arriving blocks have been underlined. One can immediately see that the SS bandwidth savings of the APWAT method are B=r blocks (since 1=r of the blocks are played from TS) for each object. Clearly also, the PS buffer requirements are zero. 3.2. Generalizing APWAT 3.2.1. Disassociating integer r and twisting Given the system parameter r, APWAT can be applied not only on the r- twisted sequence, but also on any r -twisted sequence, where k is a divisor of r. In this case, APWAT will play all boldfaced blocks (of the r -twisted sequence) from TS. An example of APWAT's action on the 2-twisted sequence of a 13-block object (see Figure 3(a)) when r=4 is given in Table 5. Once more, newly arrived blocks have been underlined. Clearly, when the system's bandwidth ratio is r, the SS bandwidth savings of our newest APWAT algorithm applied on the r -twisted sequence, are Bk=r blocks, since k=r (boldfaced) of the B blocks are never flushed to or retrieved from the SS. The PS buffer space requirements of the new algorithm are B(k\Gamma1)=r. This is true because the algorithm runs for B=r time units, during each of which the number of blocks in PS is increased by k (uploaded from TS) and decreased by Note that when space requirements are zero. Note also that APWAT is always better than Alternate Play, in the sense that for any attainable SS bandwidth savings, APWAT's PS requirements are less that Alternate Play's PS requirements. This follows, since Bk(r\Gamma1) r is equivalent to r k, which is always true. 3.2.2. Real r As it were the case with our Alternate Play algorithm (in Section 2.2.2), our APWAT algorithm and its companion Twisted Placement algorithm do not require r to be an integer. One may generate the brc-twisted sequence, and apply APWAT on it (playing all boldfaced blocks from TS). Let's call this algorithm the brc APWAT algorithm. The obvious drawback of the brc APWAT algorithm is that the PS buffer space requirements are not zero any more. Indeed, one may routinely verify that, in the worst case, the PS buffer requirement will increase to B=(r(r \Gamma 1)). This is so, because the APWAT algorithm runs for B=r time units, displaying B=r blocks. During this time, B= brc blocks are uploaded to PS. Since all blocks are played from TS, when the algorithm terminates, the number of blocks still left in PS is B brc . If PS space is scarce, an alternative to the brc APWAT is possible. The main idea is to generate a twisted sequence such that, when it is APWAT-played, Figure 4. 2.3-twisted sequence of a 20-block object no boldfaced block will be ever uploaded earlier than the time it will be consumed (actually, since r may not be an integer, a block is allowed to be uploaded earlier than the time it will be consumed but only by a fraction of a time unit). An example of such a twisted sequence is given in Figure 4. Pseudo-code is given Algorithm 4. Since the distances between the boldfaced blocks of our new Algorithm 4 Twisted Placement (for real r) INPUT: B, real r OUTPUT: r-twisted sequence of object's blocks (to be placed on TS) for i:=1 to l r randomly place blocks B d B r to blank tape positions method are sometimes 1 brc and sometimes 1 dre , we will refer to this new twist and play method as the brc&dre APWAT algorithm. Clearly, the SS bandwidth savings of our newest brc&dre APWAT algorithm are B=r blocks, and the PS requirements is 1 block (needed to hold parts of the partially uploaded blocks). 3.2.3. Disassociating real r and twisting If the system parameter is a real number r, APWAT can be applied not only on the r-twisted sequence, but also on a r 0 -twisted sequence, for any r 0 !r. Again, APWAT will play all boldfaced blocks (of the r 0 -twisted sequence) from TS. Reasoning as in Section 3.2.1, one may easily verify that the SS bandwidth savings in this case are B=r 0 blocks, and that the PS requirements are B( 1 Table 6 summarizes the performance of all examined play algorithms. Table Play algorithms' performance comparison PS SS bandwidth requirements savings Remarks Conventional Play 0 - Alternate Play r (integer r) r (integer r r divisor of r (integer r (real \Gammar (real 3.3. SS-resident Objects & Strips of Streams Although not mentioned explicitly so far, all previously described algorithms work only if r 1. One may notice that a r !1 situation may indeed arise (e.g., low-end products - see Table 1). One may also immediately see that the r 1 restriction is obviously not a deficiency of our algorithms: if r!1, and an object resides exclusively on TS, no algorithm can display it without hiccups, or without a long response time. If r ! 1, and the object resides in SS as well, one may use the ideas presented in the previous sections to display the object in a manner requiring less SS bandwidth than the obvious Play from SS algorithm, while still requiring zero PS buffer space. To this goal, we contribute the concept of a strip of a stream. 3.3.1. Integer r A strip of a stream is a subset of the blocks of a (stream) object. We define a 1 r -strip (of a stream) to be the strip consisting of every 1 r -th block of an object, except for the 1st block of the object. For example, the 2-strip is [B3, B5, B7, B9, . ], the 3-strip is [B4, B7, B10, B13, . ], etc. We store such a 1 r -strip (as a partial replica of the object) permanently in TS. One may now see that a continuous uploading of TS blocks results in each such block being brought into PS precisely before it has to be consumed. Pseudo-code displaying an object in such a manner is given as Algorithm 5. The nota- Algorithm 5 Strips of Stream Playing (r !1; 1 r -strip (br(B\Gamma1)c blocks) in TS ffl (rest of) object's blocks in SS next upld blk := first block on strip next displ blk upload(next upld blk++) for time unit i:= 1 to 1 r do Play-from-SS(next displ blk++) while time unit i B do parbegin 1. consume(next displ blk++) for Play-from-SS(next displ blk++) 2. upload(next upld blk++) parend tional conventions of Algorithms 1 and 3 have been used here, too. A trace of the algorithm is given in Table 7. Unlike all previous algorithms, the given Strips of Stream Play algorithm terminates when all blocks have been displayed (not simply when they have been read off the TS). Eventually, rB of the object's B blocks will have been played from TS, and the remaining (1\Gammar)B from SS. When compared with the Conventional Play algorith, the Strips of Stream Play algorithm achieves rB blocks SS bandwidth savings. The PS buffer space requirement of the algorithm is zero. Table Example of Strips of Stream Play (B=12; r=0:5; time uploaded block unit from TS strip displayed 5-6 B7 B5 B6 7-8 B9 B7 B8 Table Example of Strips of Stream Play 2:3 time uploaded block unit from TS strip displayed 1.0-2.3 B4 B1 B2 [B3 2.3-4.6 6.9-9.2 14.8-16.1 17.1-18.4 3.3.2. Real r Once more, the given definition of a 1 r -strip is meaningful if 1 r is an integer. r is a arbitrary real, a 1 r -strip can be defined so that each block in the strip is uploaded precisely before it is to be consumed. As an example, the reader is invited to verify that if 2:3 , then the2:3 -strip of a 20-block object should be [B 4 The analogy between this 1 2:3 -strip and the boldfaced blocks of the 2.3-twisted sequence of Figure 4 should be obvious. The Strips of Stream Play, when applied on the above 1 2:3 -strip, will play the object as shown in Table 8. Here, notation that 0.3 of block B 3 is displayed. Pseudo-code generating a r -strip, for arbitrary r, is given in Algorithm 6. Algorithm 6 1 r -strip generation (real r!1) INPUT: B, real r!1 r -strip S[1], . , S[B] s :=r for i:= 1 to l r 4. Data Elevation With Known Multiplexing Degree In many environments, consisting of high-end tape library products with very large bandwidth, the tertiary storage level of a multimedia server will be satisfying more than one requests simultaneously. In order to avoid unacceptably high start-up latencies, it is desirable for the TS to be multiplexed. When the bandwidth is multiplexed across several concurrent streams, the effective TS bandwidth b T;eff and the effective ratio r eff are reduced. We will assume a Round Robin scheduling discipline for the multiplexed requests. A moment's thought reveals that using, say, the APWAT algorithm on the twisted sequence for r, does not work. Consider, for example, the twisted sequence for r =4, given in Figure 3(b). If the number of multiplexed jobs (multi- plexing degree) is for each of the two multiplexed jobs, 3 the APWAT algorithm, with Round Robin scheduling of a time slice equal to d, will suffer from hiccups. For instance, while the B1's of the two objects are displayed, APWAT will elevate/upload the next 4 blocks of the first object (which will take time unit) and it will then switch to elevate/upload the next 4 blocks of the second object (which will also take 1 time unit). Thus, it is apparent that the displays will starve. Another approach, is to use the APWAT algorithm on the twisted sequence of Figure 3(a), instead of Figure 3(b). For our example, the APWAT algorithm, with Round Robin scheduling of a time slice equal to d=2, 3 For now, for readability purposes, we assume no exchange overhead. will display both objects without hiccups. It will elevate/upload a pair of blocks of the first object (which will take 1/2 time unit), and it will then switch to elevate/upload a pair of blocks of the second object (which will take another 1/2 time unit). This method (playing boldfaced blocks from TS, and the rest from SS) requires no extra PS buffers, and it achieves 50% SS bandwidth savings (since half of the blocks are played from TS), when compared with the Conventional Play algorithm. A crucial factor to the last approach's success (besides the use of the twisted sequence for r was the fact that its time slice was the time needed to blocks. To see why this is indeed crucial, consider, for example, the extreme case of a time slice equal to the time needed to upload 12 blocks: the last approach could not achieve 50% SS bandwidth savings with no extra PS buffers (for the same sequence of r This suggests that small time slices are better than big ones, and one should thus use time slices as small as possible. Unfortunately, as we discussed earlier, the above approach is over-simplified, since there is a switch cost c (which includes, possibly, a tape exchange cost, plus a search within a tape) associated with each Round Robin's switch from one object to another. Therefore, the just derived minimum time slice will not be sufficient to display the objects without hiccups. Nevertheless, we are now in a position to describe a general approach and to derive the minimum time slice. 4.1. APWAT with Round Robin The basic idea is to split each object's blocks into groups of t blocks each (called t-tuples, or, simply tuples), for some appropriate value of t, as follows: the first t-tuple will contain the blocks the second t-tuple will contain the blocks B t+1 , . , B 2t , etc. In order to process (upload, elevate, display) each t-tuple in the most efficient way, we will use an r-twisted placement of the blocks within each t-tuple. The desired twisting here is slightly different than the one produced by Algorithm 2. The Twisted Placement, produced by Algorithm 2, places the first blocks of an object in locations 1; r+1; 2r+1; 3r+1, etc. (for integer r). The Twisted Placement that is useful in our case places the first blocks of a tuple in locations r, 2r, 3r,. within the tuple. Such an arrangement of an object's blocks will be called a (t; r)-organization, or simply a t-organization. An example is shown in Figure 5. The pseudo-code for deriving the (t; r)-organization is given in Algorithm 7. Figure 5. (8,4)-organization of a 17-block object The number of "full" t-tuples is , with possibly one more t-tuple with fewer than t blocks. Note that there are \Xi t r bold blocks in each t-tuple. In general, the contents of the n-th t-tuple is the set of blocks which will occupy the tape positions The first bold block in the n-th t-tuple will be B (n\Gamma1)t+2 . The i-th bold block of a t-tuple is placed at position ir within the positions occupied by this t-tuple, r Algorithm 7 Twisted (t; r)-organization INPUT: B, integer r, t OUTPUT:(t; r)-organization for n:=1 to do /* place the bold blocks in their proper position */ for i:=1 to \Xi t r do randomly place un-assigned blocks of the set fB on the blank tape positions in the set f(n \Gamma 1)t As stated earlier, the j objects to be displayed will be multiplexed in a Round Robin fashion: each time slice will process one t-tuple by employing the APWAT algorithm. As an example, suppose that r=4, and consider the (8,4)-organization of Figure 5. While displaying B 1 of the first object, the Round Robin algorithm will elevate [B 4 . While displaying B 2 , it will elevate It will then start displaying B 3 , and it will switch to a second object (processing its first 9 blocks), possibly then to a third object (also processing its first 9 blocks), etc. During the uploads/elevations of the blocks of the second, third, etc., objects, the display process of the first object will finish off consuming B 3 , and it will subsequently start playing 9 from SS. To guarantee a play of all objects with no hiccups, Round Robin should return to the first object before its display of B 9 has started. The last remark can be used to derive a relationship between j and t. As explained, the time required to read j t-tuples (one tuple from each object) should be less than or equal to the time needed to display the t blocks of the tuple read time slices earlier. Since the time needed to upload a t-tuple is ts=b T , or td=r, and the switch cost is c, it follows that the time required to read j t-tuples is (td=r+c)j, or (ts=b T also, the time needed to display t blocks (of one object) is td. Thus td td r should hold. This allows us to derive the minimum allowable value of t, given the system's characteristics c, r and d, - or equivalently (since d=r=s=b T r, s and b T - and the number j of multiplexed jobs, as cjr The "j !r" condition is intuitively correct (if the number of jobs to be multiplexed is greater than the number of blocks that can be uploaded in a time unit, then no multiplexing is possible), and is mathematically a consequence of the "t ? 0" condition that must hold together with (1). Note also that if (2) gives t min (j)B, then no multiplexing is possible. Example 1: Consider an object consisting of 12,000 blocks, each 0.5 MB, residing on a TS having bandwidth 20 MB/sec and exchange cost 10 sec. Further- more, assume that the time needed to display one block is 1 sec. These figures imply that r=40. Then, eqn. (2) says that the display of objects can be multiplexed using APWAT with Round Robin if each object is split into at most 30 tuples of at least 400 blocks each. Inequality (1) may be seen under a different angle. If we assume that the size t of each tuple has been (somehow) fixed, it gives us the maximum number of jobs that can be multiplexed as rtd cr tsr ts Example 2: Given the configuration described in Example 1, eqn. (3) says that if each object is split into 30 tuples of 400 blocks each, then up to 20 such objects can be multiplexed. Given a system bandwidth ratio r, the APWAT algorithm using a (t; r)- configuration, for any t, has zero PS buffer requirements, since each uploaded (boldfaced in Figure 5) block of each of the j objects is uploaded precisely before it is to be consumed. Clearly also, the SS bandwidth savings are B=r for each of the j jobs yielding a total savings of jB=r blocks for all j jobs, when compared with the Conventional Play algorithm (elevating all blocks to SS, and playing them from there). Note that as discussed in Section 3.2.1, given the system's ratio r, one need not use a (t; r)-configuration; one may use instead a (t; r k )-configuration, for any dividing r. In such a case (see Table 6), the PS buffer requirements will be and the SS bandwidth savings will be jBk=r blocks for all j jobs. Similarly, we can also handle the case where r is real. We omit the relevant details, for space reasons, since they are very similar to the description presented in Section 3.2.2. 4.2. Start-Up Latency Considerations In the algorithms of Sections 2 and 3, the issue of the start-up latency was not addressed in detail since it was not important (we were assuming only one display request at a time: start-up latency was equal to the time needed to upload the 1st block of the object). However, in a multiplexed environment, this issue deserves more attention. Assuming that all requests for the j objects arrive simultaneously, the display of the 1st object will start with a delay of s=b T (i.e., after B 1 of the 1st object has been uploaded), the display of the 2nd object will start with a delay of s=b T +ts=b T +c (i.e., after the 1st t-tuple of the first object has been uploaded, has switched to the beginning of the 2nd object, and B 1 of the 2nd object has been uploaded), etc. Therefore, the display of the j-th object will start with 4 This is the worst case scenario, assuming that techniques such as bridging, batching, and adaptive piggy backing are not employed. a delay of r The precise effect of the choice of t now becomes more clear. Given a certain multiplexing degree j, if one chooses t ? t min , the start-up latency will be higher (as implied by (4)), but the TS will be idling (since when the time slice of the 1st object arrives again, not all of its blocks, uploaded during its previous time slice, will have been consumed), allowing it to be used for other purposes. Nonetheless, the display of the object will be done correctly (i.e., without hiccups), the PS space requirement will still be zero, and the total SS bandwidth savings will still be jB=r for all j jobs. It is worth noting that, for the configuration in Example 1 20, the maximum start-up latency witnessed by a multiplexed stream is smaller than 7 minutes. If multiplexing was not performed, to serve 20 streams the maximum start-up latency witnessed would be over one hour and thirty eight minutes, which is of course unacceptable for video-on-demand applications. 5. Data Elevation With Unknown Multiplexing Degree The algorithms of the previous section (generating the (t; r)-organizations and subsequently APWAT-playing them) work if the number j of multiplexed jobs in known in advance. In this section, we will discuss methods of overcoming this limitation. Assuming that r?1, if j ?r, then no multiplexing of all j display requests is possible. We may thus assume that there is a known upper bound j M on the maximum value of the multiplexing degree j for the system. Obviously, In order to display j objects, for arbitrary 2 j j M , it suffices to use the t min (j M )-organization. This is so, because, according to (2), t min (j M ) (j). The discussion now of the last paragraph of Section 4.2 implies that using the t min (j M )-organization will work correctly, without changing the attained SS bandwidth savings of jB=r blocks, or the required zero PS buffer space. One drawback of the last proposal is the fact that the TS may be idling as discussed in Section 4.2. Additionally, this proposal suffers from increased start-up latency, as shown by (4). If startup latency is of great importance, it can be traded-off with additional TS media storage space, as follows: Create and store a t min (j 1 )-organization, a t min (j 2 )-organization, etc., for various values of Each used t min (j i )-organization is a full replica of an object in TS. Frequently, the workload can be guessed, using past experience, and this can be used as a guide to decide which replicas to create. At run time, given the current multiplexing degree j curr , use the smallest such t min (j i ) value which is larger than t min (j curr ). This will result in the smallest possible start-up latency. On the other hand, as shown by (3), selecting the smallest such t min (j i ) value (which is larger than t min (j curr )) results in a smaller value for the maximum allowable number of multiplexed jobs. Since the total SS bandwidth savings offered by our methods are potentially greater when the allowable multiprogramming degree is higher, it is most beneficial from this point of view to use the t min value for the highest possible number of multiplexed jobs. To resolve this trade-off between start-up latencies and achievable SS bandwidth savings, we suggest the following algorithm: use the smallest such t min (j i ) value which is larger than t min (j curr ) ffl as more requests arrive to the TS for multiplexing, (or depart from TS service) dynamically switch to the smallest such t min (j i ) value which is larger than t min (j curr ) for the new value of j curr . With this algorithm we exploit the replicas (with the different t-tuple organi- zations) in order to dynamically adapt to the new j curr value and allow the maximum number of requests to be multiplexed by the TS, while at the same time enforcing the smallest possible penalty for their start-up latencies. 6. Pragmatic Concerns In this section we focus on pragmatic concerns arising from tape technology and MPEG-based video-server application peculiarities and which present difficulties for the implementation of the placement strategies developed in the earlier sections. 6.1. Streaming Mode Recording The first concern is with regard to the fact that tape drives are usually recorded in a streaming mode. Because of their relatively high error rates, this means that, when errors are detected during recording, the re-write of the affected block occurs without stopping and rewinding the tape, creating essentially a number of "bad spots" or "holes" amidst the useful data on the tape. This of course complicates the requirements for our twisted placements. The proposed solution is technical. Recall from Section 3 that given the system parameter r, our play algorithms can operate on any r 0 -twisted sequence, for any r 0 ! r. Note that essentially the aforementioned "holes" on a tape have the effect of reducing the transfer rate of the real/useful data. In other words, given r and an estimation of the write-error rate of a tape drive, we can (perhaps pessimistically, to avoid hiccups) estimate the parameter r 0 and create the twisting sequence for r 0 . As an example, if our estimations of the tape write-error rate lead us to r instead of using the 4-twisting sequence shown by Figure 3(b), we can use the 2-twisting sequence of Figure 3(a). When recorded on tape, there will be "holes" in between the blocks of this sequence so that on average, in each time unit the tape head will have scanned as much tape is required to hold data blocks will have been read off the tape, one of which will be played from TS directly. 6.2. Twisting Variable Sized Blocks Our champion application is video servers. The predominant encoding technique for videos is MPEG. Given the well-established practice of normalizing video block lengths with respect to the playback duration time, the recorded MPEG video blocks necessarily become of variable size. The variability in the block sizes, of course, complicates our twisted placements. In fact, it complicates all video-data placement problems. For this reason most related work (placement algorithms (and their analyses) for video data on SS) [ assume constant block sizes, as we have done up to now. In this section we solve this problem. We adopt the following approach. Let us assume that b denotes the size of block with s now denoting the average block size. Recalling that each time unit lasts d seconds, in each time unit T \Delta d MB are read off the tape. Assuming that none of the B blocks has a size that is greater than T \Delta d, the pseudocode for the twist algorithm is given as Algorithm 8. Observations 1. During each time unit, by design, one partition is read. 2. Blocks will be the Play-from-TS blocks. Algorithm 8 Twisted Placement with Variable Block Sizes 1. Store B 1 at the beginning of the tape 2. Split the remaining movie size into partitions Each partition has size T \Delta d MB There will be L such partitions, where l \Deltad 3. Store blocks BL+1 at the right end of partition 1, 2, ., L, respectively 4. Distribute the remaining blocks into the empty spaces of the partitions - these will be the Play-from-SS blocks. 3. As long as each Play-from-TS block is "flushed right" in each partition, there will be no additional PS space requirements - these blocks will arrive just in time before their display starts. 4. Step 4 with the above requirement is NP-Complete. It can be reduced to the well-known bin packing problem, where each partition corresponds to a bin and we want to employ as few (bins) partitions as possible during Step 4. That is, we wish to fit the remaining Play-from-SS blocks as tightly as possible in the partitions, so to have the Play-from-TS blocks closer to the right end of their partitions, so to minimize the additional requirements for PS space. Thus, finding the solution with the minimum PS space requirements is an NP-Complete problem. We want an approximation algorithm for Step 4 that is computationally inexpensive and performs close to the optimal solution. A simple algorithm that meets these demands is the Decreasing First Fit (DFF) algorithm [GJ79]. DFF, when "translated" into our problem setting, works as follows: First it sorts the blocks in decreasing order of their sizes. Then the blocks are placed in this order into the first partition they fit. It can be shown [GJ79] that the performance of DFF is never worse than 22% of the optimal solution, which in our problem setting means that: 1. On occasion DFF will result in requiring 22% more tape storage for a video, because of "holes" in each partition between the Play-From-SS and the Play- 3from-TS blocks. This also implies that, in the worst case, 22% more time will be required to read the blocks of a video from the tape. 2. However, each Play-from-TS block will require zero additional PS space. The obvious solution to the above problem is to, first, shift left each Play- block in its partition to fill the empty space and, second, shift left the partitions to cover the now empty spaces between the partitions. This will eliminate all "holes" but will also result, in the worst case, in 22% additional PS space. Thus, we need an algorithm that will strike an effective balance between read time and PS space overhead. Pseudocode for this algorithm is given as Algorithm 9. The essence of the algorithm's contribution is in Step 6, which guarantees that there will never be a requirement for additional PS space that is greater than MB. This is achieved at the expense of having fewer Play-from-TS blocks (i.e., lower SS bandwidth savings, since some partitions "lose" their Play-from-TS blocks). A following example, using typical parameter values, will illustrate the usefulness of the above algorithm. 6.3. Average I/O Bandwidth Savings of Algorithm 9 We now provide an analysis showing the average I/O bandwidth savings of the above algorithm, in the worst case performance of the DFF algorithm. The total empty space, in the worst case, that would result from the running of Algorithm 8 is 22% of the total space, that is: total s: (5) The number of partitions is The average empty space per partition is: Thus, on average, one in every k partitions "loses" its Play-from-TS block, where Algorithm 9 Polynomial-Time Twisted Placement under Variable Block Sizes INPUT: All Play-from-SS and from-TS blocks, after B 1 is placed at the beginning. next remaining blocks = set of all Play-from-SS blocks next available tape into partitions of size T \Delta d MB. 1. Put at the right end of partition next part the block next TS block 2. Sort remaining blocks in decreasing order 3. Run DFF algorithm until partition next part can be filled no more i.e., until its empty space is smaller than minimum block size in remaining blocks 4. inherit empty += amount of empty space in partition next part 5. shift left the Play-from-TS block in partition next part to cover the partition's empty space 6. If (inherit empty T \Delta d) then f the last block in partition next part will now be marked as a Play-from-SS block (i.e., partition next part lost its Play-from-TS block) 7. remove from remaining blocks the ones just placed 8. next part += 1 9. next TS block += 1 10. Repeat until all blocks are placed. Therefore, the number of partitions without a Play-from-TS block is L k and the I/O bandwidth savings are k . Example 7: Consider an object consisting of 3,000 blocks, each 0.5 MB, while MB/sec. For this example we have: total 5. The number of partitions without a Play-from-TS block is about 60. So, in the worst case, we require at most 5MB PS space while about 80% of all partitions retain their Play-from-TS block. 7. Conclusions In this paper we have addressed the problem of continuous data elevation in multimedia servers which are based on HSMSs; a problem which, in our view, has not received adequate attention despite the facts that: (i) the current cost per megabyte figures make tape libraries the most cost-efficient medium for such applications, (ii) several real-world products are already employing tape libraries for storing continuous media, and (iii) the significant continuous annual increase in the tape library market promises near-future technological improvements that will improve their performance. We first contributed the notion of alternating the playback of delay-sensitive data between the TS and the SS and discussed how this idea can save significant SS bandwidth (but also pointed out that it requires non-zero buffer space). Sub- sequently, we contributed the Twisted Placement algorithm with a companion play algorithm, called Alternate Play With A Twist; the Placement algorithm determines the proper placement/recording order for the blocks of objects on the tapes so that the Play algorithm achieves the same SS bandwidth savings as before but with zero PS buffer space requirements this time. Subsequently, we contributed the notion of strips of streams, which are special partial replicas of stream objects, residing permanently in TS, and which consist of the blocks which are to be played from TS. Strips of streams allow the previous contributions to be enjoyed even when (1) the bandwidth of TS is smaller than the display bandwidth of objects; and (2) the objects also reside in SS (which will be the case for the most popular objects). Later we considered the subject of multiplexing TS-resident video streams over tape drives. Multiplexing is desirable for two reasons. First, it can drastically reduce the average start-up latency of videos residing in the TS. Second, employing our placement and Play algorithms in such environments can offer the associated SS bandwidth savings for all streams being multiplexed, increasing the total benefits of our approach. We derived algorithms which employ the previously-developed techniques to continue offering SS bandwidth savings, at no additional PS buffer space. We presented an algorithm showing how to store the stream on tapes so that a high multiplexing degree j is maintained. For each of the j jobs the algorithm continues to achieve the aforementioned savings; thus, the total savings are significantly greater. For many applications we expect that the data objects will exhibit skewed access distributions (e.g., popular movies in a movie-on-demand application). For a large majority of cases the most popular objects will have been uploaded and will be residing in SS. In such scenarios, the SS will be responsible for satisfying a great percentage of requests, referring to the most popular streams. With present technology, however, even the high-end SS server products do not have enough bandwidth to satisfy all requests for all streams they store. (This is expected to hold in the future as well, given the pace of improvement of storage capacities.) When taking into account the IO bandwidth required for uploading colder streams from TS to SS, the available IO bandwidth becomes even smaller. In this paper, we have shown techniques which exploit the ever-increasing bandwidth of modern tape drives, for increasing the available IO bandwidth for satisfying requests. We made an effort to view the problem comprehensively in the sense that we focus on all main resources (host RAM memory, SS and IO bus bandwidth, and TS bandwidth). With our techniques the overall IO bandwidth is increased without requiring extra host RAM, while still avoiding video hiccups. In a sense, the aforementioned contributions suggest the collaboration of TS and SS in order to improve the system's throughput. The essence of our proposed techniques is that the TS of a HSMS, is not simply used to augment the storage capacity of SS (as its traditional role would indicate) but to augment the SS bandwidth as well. This is the high level contribution of this research. --R Dynamic batching policies for an on-demand video server Staggered striping in multimedia information systems. Tertiary Storage: An Evaluation of New Applications. Challenges for tertiary storage in multimedia servers. Maximizing performance in a striped disk array. Research and development issues for large-scale multimedia information systems Principles of optimallly placing data in tertiary storage libraries. Buffering and caching in large-scale video servers Buffer management policy for an on-demand video server A pipelining mechanism to minimize the latency time in hierarchical multimedia storage managers. Computers and Intractability: A Guide to the Theory of NP-completeness Adaptive piggybacking: A novel technique for data sharing in video-on-demand storage servers Continuous retrieval of multimedia data using parallelism. A Study on the Use of Tertiary Storage in Multimedia Systems. On multimedia repositories The role of data storage in broadcasting future. On the modeling and performance characteristics of a serpentine tape drive. Random I/O scheduling in online tertiary storage systems. Benchmarking tape system performance. Performance measurements and models of tertiary storage devices. An analytical performance model of robotic storage libraries. Using tertiary storage in video-on-demand servers Evaluating video layout strategies for a high performance storage server. The optimum execution order of queries in linear storage. A case for intelligent disks (idisks). Buffer management for continuous media sharing in multimedia database systems. Integrated document prefetching and caching in hierarchical storage based on markov-chain predictions The design of a storage server for continuous media. Keynote speech. A case for redundant arrays of inexpensive disks (RAID). Active storage for large-scale data mining and multimedia applications Efficient storage techniques for digital continuous multime- dia Buffer management for video database systems. striping. Untangle your tape storage costs. Overlay striping and optimal parallel I/O in modern applications. Prefetching into smart-disk caches for high-performance continuous media servers Tpc executive summaries. --TR --CTR simulated annealing approach for multimedia data placement, Journal of Systems and Software, v.73 n.3, p.467-480, November-December 2004 Athena Vakali , Evimaria Terzi , Elisa Bertino , Ahmed Elmagarmid, Hierarchical data placement for navigational multimedia applications, Data & Knowledge Engineering, v.44 n.1, p.49-80, January
tape libraries;video playback;hierarchical storage servers;placement algorithms
592967
Cactus Tools for Grid Applications.
Cactus is an open source problem solving environment designed for scientists and engineers. Its modular structure facilitates parallel computation across different architectures and collaborative code development between different groups. The Cactus Code originated in the academic research community, where it has been developed and used over many years by a large international collaboration of physicists and computational scientists. We discuss here how the intensive computing requirements of physics applications now using the Cactus Code encourage the use of distributed and metacomputing, and detail how its design makes it an ideal application test-bed for Grid computing. We describe the development of tools, and the experiments which have already been performed in a Grid environment with Cactus, including distributed simulations, remote monitoring and steering, and data handling and visualization. Finally, we discuss how Grid portals, such as those already developed for Cactus, will open the door to global computing resources for scientific users.
Introduction Cactus [1], [2] is an open source problem solving environment designed to provide a unied modular and parallel computational framework for physicists and engineers. The Cactus Code was originally developed to provide a framework for the numerical solution of Einstein's Equations [3], one of the most complex sets of partial dierential equations in physics. These equations govern such cataclysmic events as the collisions of black holes or the supernova explosions of stars. The solution of these equations with computers continues to provide challenges in the elds of mathematics, physics and computer science. The modular design of Cactus enables people and institutes from all these disciplines to coordinate their research, using Cactus as the collaborating and unifying tool. The name Cactus comes from the design of a central core (or which connects to application modules (or thorns) through an extensible interface. Thorns can implement custom developed scientic or engineering applications, such as the Einstein solvers, or other applications such as computational uid dynamics. Other thorns from a standard computational toolkit provide a range of capabilities, such as parallel I/O, data distribution, or checkpointing. Cactus runs on many architectures. Applications, developed on standard workstations or laptops, can be seamlessly run on clusters or supercomputers. Parallelism and portability are achieved by hiding the driver layer and features such as the I/O system and calling interface under a simple abstraction API. The Cactus API supports C/C++ and F77/F90 programming languages for the thorns. Thus thorn programmers can work in the language they nd most convenient, and are not required to master the latest and greatest computing paradigms. This makes it easier for scientists to turn existing codes into thorns which can then make use of the complete Cactus infrastructure, and in turn be used by other thorns within Cactus. Cactus provides easy access to many cutting edge software technologies being developed in the academic research community, such as the Globus Meta-computing Toolkit, HDF5 parallel le I/O, the PETSc scientic computing library, adaptive mesh renement, web interfaces, and advanced visualization tools. 2. The Need for the Grid Of many applications using the Cactus framework, an important one which continues to drive its development is the solution of the Einstein Equations. The large and varied computational requirements of solving these equations for scenarios such as black hole or neutron star collisions, make them a good example for demonstrating the need for Grid computing, and an ideal testbed for developing new techniques. In developing the Cactus infrastructure to make full use of the Grid for such problems these advances are then immediately available for all applications. Implementing the full Einstein Equations in a nite dierence code amounts to a memory requirement of around one hundred 3D arrays, and a CPU requirement of thousands of oating point operations per grid point and timestep. Considering that a su-ciently accurate solution of a full 3D black hole problem will require at least 1000 grid points in each spatial dimension, this implies computers. Further, to analyze the large data sets created during a simulation requires advanced techniques in le management and visualization To date, the resources of the individual supercomputers available have limited simulations to around 200-300 grid points in each spatial dimension. Even simulations generate huge amounts of data, and negotiating the curiosities of dierent supercomputers, such as batch queues and le systems, is not something that physicists relish. The Grid provides a way to access the resources needed for these simulations. It provides a uniform access layer to supercomputers and computing resources, making these resources far more useful to scientists who want to use them for simulations. Given the appropriate permissions, networks, allocations and Grid enabling software, a scientist could in principle run a simulation on a set of supercomputers, all connected by the Grid, and thus be able to run far larger problems than would be possible on a routine basis without the Grid. With proper software tools, the Grid provides the necessary infrastructure to connect the machines, and to deal with the resulting large data sets, and, ultimately, the resources to analyze such volumes of data. The dream for physicists is that Grid computing will provide a scientic programming environment similar to that shown in Figure 1, allowing working scenarios such as: A physicist, sitting in a cafe in Berlin has an idea for a colliding black hole run, maybe to try a new initial data set, or to test a new evolution method. She uses a web portal on her PDA to select the needed Cactus thorns, and to estimate the required computer resources. Her Grid software takes over, selecting the most appropriate machine or set of machines to use from those available to her. This Grid software automatically creates or transfers executables and parameter les and starts the run on the remote resources. After several coees, she connects to the running Cactus simulation, using one of the remote access thorns, and sees that things are going better than expected. She rings up colleagues in the who watch the isosurfaces being streamed out from Cactus. They want to save some 3D data sets to analyze later, so they connect to the Cactus run using their web browser, and turn on output for the grid functions they are interested in. As futuristic as such a scenario sounds, all the pieces already exist in a prototype form, and are being further developed and integrated, as described below and in [1], [4], [5] and [6]. Figure 1. The dream of Grid computing: Grid infrastructure provides a transparent and exible working environment providing access to global computing resources. 3. Grid Computing with Cactus Cactus was designed with the Grid and Grid applications in mind. It provides a layer on top of the Grid, giving a programming interface which allows the user to be completely ignorant of the nature of the machine or machines that the simulation runs on. The code provides access to Grid resources such as distributed I/O and parallelization across any number of supercomputers with precisely the same interface as it does to the resources of a single machine [7]. Cactus thus provides a convenient laboratory for computer scientists to develop metacomputing techniques, which can then be tested with real physics applications and also by real users without requiring changes to the physics application code. When a new technique is perfected, it can immediately be made available to the whole community of Cactus Users. Grid Computing developments and experiments have been performed using Cactus for several years, some of which are described in the sections below. Capabilities are being further developed in connection with Cactus through several projects. A DFN-Verein project [4] at the AEI in Germany is funded to exploit high speed networks for colliding black hole simulations, and is concentrating on remote visualization [8], remote steering and distributed I/O [9]. A project funded by the so-called KDI program of the American National Science Foundation (NSF) joins ve institutes to develop an Astrophysics Simulation Collaboratory [5] which will provide an environment for physicists to utilize Grid computing for problems such as the collisions of neutron stars. The GrADs project [10], also funded by the NSF in the USA, is using Cactus as one of its applications for developing a Grid based computing environment. The European Grid Forum [11] has chosen Cactus as one of its applications running on the European Grid-TestBed. These technologies are being brought into the scientic and engineering communities as they are developed. 3.1. Distributed Simulations on the Grid We are actively working to develop techniques which allow researchers to harness computational resources wherever they may be on the Grid. This could include a distributed set of machines connected by high speed networks, allowing larger or faster simulations than would be possible on a single machine. At Super-computing a neutron star collision was run with Cactus, using the Globus [6] metacomputing toolkit to split the domain across two T3Es on dierent sides of the Atlantic, one in Munich, Germany and one in San Diego, California. In this simulation the neutron stars collided somewhere in cyberspace, over the Atlantic Ocean. The simulations were launched, visualized, and steered from the show oor in Orlando. The scaling across the two machines used for this simulation was roughly 50%, which we believe is excellent considering the large amount of communication required between the machines to solve these equations and the latencies in the transatlantic link. The latency and bandwidth are characteristic features, which determine the speed of a network. Cactus is aware of these features and can be ne-tuned in order to optimize communication. For example, if a network has a high latency but also a high bandwidth, many small messages can be coalesced into fewer bigger ones. When running in a metacomputing environment, one has to deal with dier- ent types of networks (shared-memory, distributed-memory, high-speed-network, LAN, WAN/internet) with dierent latency/bandwidth characteristics. Here, we also have the possibility to distinguish between these dierent types of network- connections in one single distributed run and tune Cactus communication patterns adequately. Partly this is already achieved by using MPICH-G2, the next-generation MPI-implementation, which can distinguish between processors located on one host (with native MPI installed) and processors separated by a LAN or WAN. According to the location, MPICH-G2 can choose dierent protocols (TCP or vendor's MPI) for communication in one single distributed parallel run. Cactus can be used with this new technology without problem, which was demonstrated in many metacomputing experiments last year, including Supercomputing 2000 in Dallas. A further aspect is load-balancing. Since dierent architectures provide dierent types of processors at dierent speeds, Cactus provides the ability to decompose the whole computational problem into sub-problems of dierent size, which t the local processor power. For running in a metacomputing environment at a real production-level, a User Portal has been built (described in Section 5), making it possible to congure and start Cactus runs from one machine via a special web-based GUI. This greatly simplies the situation for the scientist, since she does not have to deal with every detail of the local supercomputer, such as batch-systems, username/password. The portal provides automatic staging and compilation of the code on the local supercomputer, the distributed machines appearing as a single virtual machine. 3.2. Checkpointing Distributed Simulations on the Grid Grid computing events in the past have been often understood as one-time, well prepared attempts to harness several machines at the same time. In a more realistic setting, compute resources of considerable size can not be expected to be available at a given time, instead their availability is a dynamic process. A true Grid application has to be capable of dealing with such dynamic allocations of resources. The Cactus framework addresses this challenge by providing a sophisticated cross-platform checkpointing mechanism. In general, checkpointing technology allows the user to freeze the state of an application by writing a checkpoint le to disk, from which the application can be restored and continued at a later time. single MPP supercomputer cluster or compute farm tightly coupled heterogeneous supercomputers Figure 2. Migration scenario for a distributed simulation: the simulation starts on three tightly coupled supercomputers from where it is checkpointed and migrated to a single machine. The computation migrates again to nish on a cluster. In Cactus the checkpoint is not just the memory image of the application written to disk, as found in several other checkpointing systems, but the total set of user dened objects (variables, scalars, etc. While memory images tend to be quite huge and are only compatible within the same class of operating systems and architectures, this approach allows for smaller, architecture-independent check- points. The cross-platform checkpoints of Cactus can be transferred between arbitrary architectures, operating systems and numbers of processors for restarting and continuing simulations. The checkpointing mechanism is completely transparent to the user, who can request a checkpoints to be written at regular timestep intervals, at the end of the requested compute time allocation, or using a steering interface immediately at the current timestep. All of the internal technicalities of parallel I/O are hidden from the user. The user can control checkpoint behavior (as frequency or parallel I/O) by means of steerable parameters. The checkpoint mechanism allows for the output of a single, global check-point le as well as multiple checkpoint les for each of the distributed machines. The mechanism makes use of parallel I/O where possible. For restarting, the multiple checkpoint les can be recombined into a single le which can be used to restart on an arbitrary set of machines. The parallel restart operation from multiple les is currently restricted to the same topology of machines. Future developments will add intelligent components to immediately restart from multiple checkpoint les across arbitrary machine topologies. With respect to distributed simulations, a Cactus user has the ability to perform a distributed run and checkpoint this simulation even though is it being run on a heterogeneous machine set. A checkpoint le can then be transferred to a new conguration of machines to continue the simulation. The new pool of machines can dier from the previous one in type and number of machines involved as well as the number of processors. This exible chain of distributed simulations is illustrated in Figure 2: an initial simulation run across three tightly coupled supercomputers is checkpointed. The checkpoint le is transferred to a single MPP machine and restarted. After a second checkpointing event the third stage of the simulation is continued on a cluster system. 4. Grid-enabled Communication and I/O Techniques The parallel driver layer in Cactus, which manages the allocation and domain decomposition of grid variables as well as their synchronization between processor boundaries, is provided by a thorn. This means that dierent thorns can be used to implement dierent parallel paradigms, such as PVM, Pthreads, OpenMP, CORBA, etc. Cactus can be compiled with as many driver thorns as required (subject to availability), with the one actually used chosen by the user at run time through the parameter le. The current standard driver thorn is called PUGH, which uses MPI to provide parallelism. In order to perform distributed Cactus simulations on the Grid, this PUGH thorn is simply linked against the Grid-enabled MPICH-G [12] implementation of MPI which is available with the Globus toolkit. Thus, preparing a Grid-enabled version of Cactus is a compilation choice, and it is completely transparent for application thorn programmers to add their own code to a Grid-enabled 9Cactus. Using the Globus job submission tools, Cactus users can start their Cactus runs in a Grid environment just as easily as they do on a single machine. The Cactus I/O subsystem is implemented in a similar, generic manner: the esh provides a runtime interface for arbitrary I/O thorns to register their own, specic I/O methods. These methods can then in turn be invoked by the esh or any application thorn to read external data into Cactus variables or dump their contents to a storage medium for postprocessing analysis and visualization purposes. The I/O thorns currently available in the computational toolkit provide methods to write simulation data in dierent formats (1D traceline plots, 2D slices and JPEG images, full N-dimensional arrays, arbitrary hyperslabs of N-dimensional arrays, reduction scalars (e.g. minimum/maximum values), isosurface geometry data, particle trajectories, runtime standard output) also using dierent I/O libraries (FlexIO [13], HDF5 [14], JPEG, ASCII). Further methods or libraries can easily be added by thorn programmers. In the following sections we will describe in more detail the Grid software techniques we have developed to date which allow Cactus users to easily perform postprocessing analysis on data produced by a remote Cactus simulation, and also to monitor and steer running Cactus jobs remotely. A general overview of the nal proposed architecture of our Grid-enabled I/O system is shown in Figure 3. The Hierarchical Data Format version 5 (HDF5) plays a key role in this overall picture. HDF5 has become a widely accepted standard in the scientic computing community for storing data. It denes a very exible le format and provides an e-cient software library for managing arbitrary multidimensional datasets of various types. Raw data access is accomplished via a generic Virtual File Driver (VFD) layer in HDF5. Beneath this abstraction layer exists a set of low-level I/O drivers which provide dierent ways of accessing the raw data of an HDF5 le, either located on a local disk or on other storage media. We have added our own drivers to this layer which enable existing applications to have the additional capability of accessing remote data residing anywhere on the Grid. 4.1. Direct Remote File Access HDF5 already has a GASS driver (Global Access to Secondary Storage) which Local File System Virtual File Driver Layer Amira Data DPSS GridFTP Virtual File Driver Layer GASS GridFTP Unix DPSS Stream Handling Cactus Simulation Thorns Thorn Hyperslab I/O Methods Methods Visualization (Amira) Amira Modules Steering GUI Steering DPSS Server GridFTP Server HDF5 I/O Library HDF5 I/O Library Figure 3. General overview of the Grid-enabled I/O architecture automatically stages complete remote les rst to the local machine and then operates on their local copies via standard UNIX le I/O. This method is feasible for small or medium sized data les. However, large-scale computer simulations often generate large-scale data sets | single simulations may generate les containing several hundreds of GBytes, up to the order of a TByte as machine resources increase. Conventional postprocessing analysis then becomes prohibitively resource-intensive when remote simulation data must be staged for local processing. Further, in many cases, for example for rst-sight visualization purposes, only a small fraction of the overall data is really needed. For example, in a simulation of the evolution of two colliding black holes, the output may contain a dozen variables representing the state of the gravitational eld at perhaps 1000 time steps during the evolution. For visualization one might want to analyze only the rst time step of one or two of the variables. Or, in order to perform a quick pre-analysis of high-resolution data, it might be su-cient to downsample the array variables and fetch data at only every other grid point. By enhancing the HDF5 VFD layer with a driver that builds on top of the Data Grid software components [15] from the Globus toolkit we enable existing I/O layers to operate on remote HDF5 les directly. These are uniquely addressed by their URL, and after opening them with the appropriate driver, all read and write operations are performed as network transactions on the Grid | completely transparent to the application. Using the data selection capabilities of (dening so-called hyperslabs as arbitrary rectangular subregions in the multidimensional data sets, optionally with downsampling and type conversion applied) individual time steps and zones of interesting data can be read and visualized in a very e-cient and convenient way. The Data Grid client software only supports remote partial le access to Distributed Parallel Storage Systems (DPSS) [16]. During Supercomputing 1999 in Portland and at CeBIT 2000 in Hannover we successfully demonstrated the feasibility of such a DPSS Data Grid Infrastructure. In these demonstrations, Cactus simulation data residing on remote DPSS data servers was visualized by an HDF5-enabled version of the visualization package Amira [17]. This is illustrated in Figure 4. Remote Visualization with Amira CeBIT 2000 Hannover AEI Potsdam HDF5 datasets on DPSS prototype Rechenzentrum Garching Cactus Simulation on T3E Infrastructure Gigabit Network German Figure 4. Remote le access and visualization demo presented at CeBIT 2000 Remote access to les which are located anywhere on the Grid will soon be provided by a GridFtp driver [18] which supports the standard FTP protocol, enhanced with partial le access, parallel streaming capabilities, and Grid security mechanisms. Another challenge occurs when simulations are carried out on a distributed computer and generate physically distributed les. This would occur, for exam- ple, in order to exploit parallel I/O methods. It is desirable to access and transfer such distributed data sets as consistent single les, using a global address space having pointers to pieces at other locations. We plan to also tackle these problems with the Data Grid components, by organizing related les as collections of logical le instances. The DataGrid project of the Globus group is investigating such techniques [15]. 4.2. Remote Online Data Streaming and Visualization Cactus also provides the capability to stream online-data from a running simulation via TCP/IP socket communications. This can be used for many pur- poses. To date, the most common use for live data streaming is for remote visualization, which is our focus here. However, in our vision of future Grid sim- ulations, we expect running simulations to communicate with each other, migrate from machine to machine, spawn additional processes on the Grid, etc. Hence, we expect that data streaming will be a fundamental enabling technology for future Grid simulations. We are now exploiting the data streaming capabilities we describe here to enable such advanced Grid simulations, as demonstrated in our \Cactus Worm" scenario where a running Cactus simulation was able to migrate itself, using the data streaming techniques described below, from site to site across the European Egrid [19]. This is a simple example of more sophisticated types of Grid simulations, based on data streaming, that we will be developing in the future. But in the remainder of this section we focus on data streaming for use in remote visualization. Multiple visualization clients can then connect to a running Cactus executable via a socket from any remote machine on the Grid, request arbitrary data from the running simulation, and display simulation results in real-time, visualizing for example photons falling into a black hole, or isosurfaces of gravitational waves which are emitted during a black hole collision. Data streaming is integrated into Cactus in several dierent ways. One method is to access Cactus output les while they are being written by the running simulation. Those les are registered with the HTTP control interface, described in the following section, and can be downloaded to any web browser. For example, simple 1D data graphs can be viewed by simply clicking on a down-load le and ring o, for example, an xgraph program. Two-dimensional JPEG images can be viewed directly in a web browser, and continuous time sequences of JPEGs can be displayed using the auto-refresh option of the capable browsers. Another technique implements a proprietary communication protocol for sending specic geometry data such as isosurfaces or particle trajectories down a raw socket connection to a visualization program [20]. This is illustrated in Figure 6. Precomputing such data at the simulation side not only allows for parallel rendering of images but also reduces the amount of data to be transferred to remote visualization clients. Figure 5. Online Visualization: The Amira visualization toolkit [17] allows a user to visualize slices through a complete 3D data set streamed from a running Cactus simulation, and at the same time to display an isosurface obtained online for the same 3D eld. The most generic approach for streaming arbitrary data of any type is again based on the HDF5 I/O library and its VFD layer. We have developed a Stream driver which holds the HDF5 data to be streamed out of the Cactus simulation as an in-memory HDF5 le. On a ush/close operation the entire le is sent through Figure 6. Trajectories of freely falling particles in the vicinity of a rotating black hole. The particle positions are streamed to the visualization tool in real-time during computation. This was demonstrated with Cactus and Amira at IGrid 2000 in Yokohama [20]. a socket to the connected client. In the client application, the same driver is used to reconstruct the in-memory le which then can be accessed as usual to read the datasets. Since the VFD layer hides all low-level I/O operations from the upper layers of the HDF5 library and from the application that builds on top of it, applications can use their existing HDF5 le-based I/O methods immediately for online remote data access without changing their I/O interfaces. This has been demonstrated using dierent visualization toolkits, including Amira [17], the IBM Data Explorer [21], and LCA Vision [22]. The Stream driver is capable of sending data simultaneously to multiple clients. This is one key component for building a collaborative visualization environment where scientists at dierent sites can analyze the results of a remote simulation either by looking simultaneously at the same data or by requesting dierent views of it. We are working on a design for a more sophisticated I/O request protocol and the implementation of an external data server which will handle multiple clients and can also serve requests individually. By integrating intelligent data management and caching strategies, such a server would relieve the simulation from communication overhead and help to reduce data tra-c in general. 4.3. Remote Monitoring and Steering The Cactus Computational Toolkit contains a thorn HTTPD which can be added to any Cactus simulation to provide an inbuilt HTTP server. Pointing their web browsers to a URL identifying a running Cactus job on a remote machine, any number of collaborators can connect to monitor and steer the simulation online. The provided Cactus web interface allows users to query certain information about the run, such as the current iteration step, a list of available thorns and variables, and a full description of all parameters and their current settings. After successful authorization a user can also interactively change parameters which are marked as steerable. At each simulation cycle these parameters are checked, and the appropriate thorns may react on changes individually. Most of the I/O parameters are steerable. This enables users to selectively switch on or specic output at runtime, dynamically choosing which variables are output using which I/O method. I/O options such as hyperslabbing or downsampling parameters may also be modied in order to adjust online data streaming to remote visualization clients. The web interface can also be used to pause the simulation, optionally when a chosen condition is satised, and to advance the simulation by single timesteps. The Web interface provided by thorn HTTPD is dynamically extensible in that any thorn can register and update its own HTML pages at runtime. Besides a download page for Cactus output les there is a also viewport available which embeds dynamically generated JPEG images. Another steering interface again builds on top of HDF5 and the Stream driver described above. For this interface the data streaming is simply used in a bidirectional way: Cactus writes parameters into an HDF5 le which is then streamed to any connected steering client. After some user interaction, this client sends back a modied version of the parameter le which is read and evaluated by Cactus. Because of its self-describing data format and the exibility to add addi- tional, user-dened information, HDF5 also provides the possibility to build more advanced steering clients with graphical user interfaces. As an example, mini- mum/maximum values could be assigned to numerical parameters to create sliders gfor more convenient user interactions. Parameters belonging to one thorn could be sorted into a group hierarchy for building menus. These features of HDF5 make it relatively easy to implement dynamic graphical user interfaces for arbitrary Cactus parameter sets which adapt themselves to the current Cactus conguration. We are actively working on including such user interfaces into existing visualization clients. Their steering capabilities would then not only be limited to exchanging HDF5 parameter les but could also be extended to feed back any kind of data elds into Cactus, for instance to add photons to a black hole simulation to locate an event horizon. 5. Portals onto the Grid The Grid is only useful as a concept if its services can be used to create the illusion that all of the resources are centralized to the user's workstation. So the most successful distributed applications on the Grid will paradoxically be those which make the user least aware that they are in fact operating in distributed fashion. The motivation for producing a Grid Portal interface to a like Cactus is derived from the desire to hide distributed applications and immensely complex distributed/parallel software architectures behind a single point of presence, and to make them accessible through comparatively simple client-side interfaces. A portal is a single point of presence (typically hosted on the web) which can be customized for a particular user and remembers particular aspects of the customizations regardless of where the user accesses it from. Yahoo and HotMail are typical consumer-oriented examples of this capability and are in fact the originators of this new meaning for the term portal. It doesn't matter where you are, when you login to the URL of these portals, you get access to the same view of your personalized environment and data (ie. your email). Placing the PSE like Cactus within a portal creates a universally accessible interface to your scientic computing platform. The GUI is user-customizable, as if it were a desktop application on the user's own workstation, except that the same customized environment is accessible from virtually any location by simply connecting to the same URL address. A science portal has the additional implied function of automating the entire work ow for a particular scientic application, from initial data generation, to selecting resources to run the application, to archival storage management and analysis of the results of those simulations. This replaces a series of perhaps loosely (or usually poorly) integrated tools with a comprehensive environment which is customized around a particular application. Finally a collaboratory provides additional facilities for sharing information either online or asynchronously among users of the portal. Cactus has several key advantages which make it very suitable as the basis for a portal design. Its modular design supports dynamic assembly of applications online through a simplied GUI. Its sophisticated multiplatform compilation makes it very simple to run the code on any available Grid resource without the complexities of Imake or the performance penalty of a Virtual Machine. Its centralized revision control mechanism permits e-cient sharing of code, software updates, and bug xes. Finally the integrated visualization for remote monitoring and steering of the code through a web interface allows seamless integration of these capabilities with the portal's web-GUI. The Astrophysics Simulation Collaboratory Portal [5] is a concrete use of Cactus within a web portal GUI which leverages o of technology which was developed originally for e-commerce applications. The architecture utilizes a commercial-grade StrongHold (Apache) webserver which oers SSL encryption using a site certicate from a commercial Certicate Authority. Running side- by-side with the webserver is a TomCat JSP engine which oers a cleaner means to manage automation in an elegant and easily maintainable fashion. JSP allows us to directly execute methods of server-side Java beans rather than the typical CGI-script methodology of parsing the state of form elements individually after an HTTP 'POST' event. The Java beans directly call Java CoG [6] which is a pure Java implementation of the Globus toolkit to extend the automation to Grid resources. The user-state within the system is maintained by a back-end database system (OpenLDAP or mySQL) which allows simple replication of portal state allowing the web services to be scaled through server replication. The science portals and collaboratories will play an increasingly important role in HPC as the Grid evolves. The natural point of organization for user communities in an HPC environment is a particular eld of science or a particular application code, just as experimental laboratories bring together top researchers who are interested in the same or similar lines of research. The internet has provided us with access to enormous remotely located resources, but this has shifted the center of focus to the particular HPC site and its operating environment, batch queues and security policies rather than the science that is computed there. The single point of presence oered by an Grid portal recreates gthe traditional laboratory environment where scientists who share similar interests and applications are brought together under the umbrella of a shared a collaboratory. The portal itself is a distributed Grid application for a specic community of scientists rather than a general-purpose resource. So unlike traditional web-portals where you have an implicit oer that if you go to www.<my portal location>.org we'll do everything for you here using our compute resources, the Grid portal's business plan can be simply stated as go to www.<my portal application>.org and we will do everything you need for <my application> regardless of your location and that of your compute resources. This returns the focus of a scientic community to the scientic application rather than the location of the HPC resources. If the Grid is really working, in another 5 years we will no longer think of the, for example, NSF supercomputing centers as distinct sites like SDSC, NCSA, or PSC. We will instead think only of the particular application collaboratories which have been set up to study dierent scientic applications. That resource will merely be a name rather than a place. Summary The Cactus Code and Computational Toolkit and the large scale applications which it serves provide a ideal laboratory for developing and testing new Grid techniques and working practices. Cactus can be very easily congured and run in a Grid environment, and the tools developed so far already provide many capabilities for exploiting global computing resources. The infrastructure and tools developed are immediately available to the user community for testing, and many are already being successfully and benecially used by collaborations researching computationally intensive problems such as black hole and neutron star collisions. Acknowledgements The development of the Cactus Code is a highly collaborative eort, and we are indebted to a great many experts at dierent institutions for their advice, visions and support. The original design of Cactus was by Joan Masso and Paul Walker, since when it has been extensively developed at the AEI, NCSA and Washington University. It is a pleasure for us to thank Ian Foster, Steve Tuecke, Warren Smith, Brian Toonen, and Joe Bester from the Globus team at Argonne National Labs (ANL) for their Globus and Data Grid work; Mike Folk and his HDF5 development group at NCSA who helped us in implementing the requirements of remote le access into their HDF5 code; Brian Tierney from Lawrence Berkeley Labs for his DPSS support; Jason Novotny at NLANR for his help with Globus and graphical user interfaces; and Michael Russell at the University of Chicago for his Portal work. Computing resources and technical support have been provided by AEI, ANL, NCSA, Rechenzentrum Garching/Germany, and ZIB. We greatly acknowledge nancial support for Andre Merzky and Thomas Radke as well as provision of a gigabit network infrastructure in the course of the TIKSL research project by DFN (German Research Network). --R http://www. Numerical Relativity As A Tool For Computational Astrophysics J. DFN Gigabit Project: Toolkit: http://www. Grid Adaptive Development Software (GrADS): http://www. The European Grid-Forum: http://www Implementation: http://www. FlexIO: http://zeus. Hierachical Data Format Version 5: http://hdf. Distributed Parallel Storage System http://www-didc The Globus Project: GridFTP: Universal Data Transfer for the Grid. Geodesics in Kerr Space-Time IBM Data Explorer: http://www. http://zeus. --TR --CTR Michael Russell , Gabrielle Allen , Greg Daues , Ian Foster , Edward Seidel , Jason Novotny , John Shalf , Gregor von Laszewski, The Astrophysics Simulation Collaboratory: A Science Portal Enabling Community Software Development, Cluster Computing, v.5 n.3, p.297-304, July 2002 Henri Casanova , Francine Berman , Thomas Bartol , Erhan Gokcay , Terry Sejnowski , Adam Birnbaum , Jack Dongarra , Michelle Miller , Mark Ellisman , Marcio Faerman , Graziano Obertelli , Rich Wolski , Stuart Pomerantz , Joel Stiles, The Virtual Instrument: Support for Grid-Enabled Mcell Simulations, International Journal of High Performance Computing Applications, v.18 n.1, p.3-17, February 2004 Gabrielle Allen , David Angulo , Ian Foster , Gerd Lanfermann , Chuang Liu , Thomas Radke , Ed Seidel , John Shalf, The Cactus Worm: Experiments with Dynamic Resource Discovery and Allocation in a Grid Environment, International Journal of High Performance Computing Applications, v.15 n.4, p.345-358, November 2001 James A. Kohl , Torsten Wilde , David E. Bernholdt, Cumulvs: Interacting with High-Performance Scientific Simulations, for Visualization, Steering and Fault Tolerance, International Journal of High Performance Computing Applications, v.20 n.2, p.255-285, May 2006 Katarzyna Rycerz , Alfredo Tirado-Ramos , Alessia Gualandris , Simon F. Portegies Zwart , Marian Bubak , Peter M. A. Sloot, Regular Paper: Interactive N-Body Simulations On the Grid: HLA Versus MPI, International Journal of High Performance Computing Applications, v.21 n.2, p.210-221, May 2007 Lingyun Yang , Jennifer M. Schopf , Ian Foster, Conservative Scheduling: Using Predicted Variance to Improve Scheduling Decisions in Dynamic Environments, Proceedings of the ACM/IEEE conference on Supercomputing, p.31, November 15-21, Wes Bethel , John Shalf, Grid-Distributed Visualizations Using Connectionless Protocols, IEEE Computer Graphics and Applications, v.23 n.2, p.51-59, March Ming Wu , Xian-He Sun, Grid harvest service: a performance system of grid computing, Journal of Parallel and Distributed Computing, v.66 n.10, p.1322-1337, October 2006
grid portals;cactus;grid computing
592973
Event Services in High Performance Systems.
The Internet and the Grid are changing the face of high performance computing. Rather than tightly-coupled SPMD-style components running in a single cluster, on a parallel machine, or even on the Internet programmed in MPI, applications are evolving into sets of cooperating components scattered across diverse computational elements. These components may run on different operating systems and hardware platforms and may be written by different organizations in different languages. Complete applications are constructed by assembling these components in a plug-and-play fashion. This new vision for high performance computing demands features and characteristics not easily provided by traditional high-performance communications middleware. In response to these needs, we have developed ECho, a high-performance event-delivery middleware that meets the new demands of the Grid environment. ECho provides efficient binary transmission of event data with unique features that support data-type discovery and enterprise-scale application evolution. We present measurements detailing ECho's performance to show that ECho significantly outperforms other systems intended to provide this functionality and provides throughput and latency comparable to the most efficient middleware infrastructures available.
Introduction area distributed computing has been a strong focus of research in high performance computing, resulting in the development of software infrastructures like PVM, MPI, and Globus, and in the creation of the National Machine Room and the Grid by DOE and NCSA/Alliance researchers. Increasingly, research focus in this domain has turned towards component architectures which facilitate the development of complex applications by allowing the creation of generic reusable components and by easing independent component development. Some of the earliest requirements for component architectures in high-performance computing were derived from systems that attach scientific visualizations to running computations, but continuing research has generalized such models to include the ability to flexibly link general purpose computational elements as well [22,24,2]. Component-based software development has been proposed by the software engineering community over the last decade [25,29] and its advantages have been widely recognized in industry, resulting in the development of systems such as Enterprise Java Beans, Microsoft's Component Object Model and its distributed extension (DCOM), and the developing specification of the CORBA Component Model (CCM) in OMG's CORBA version 3.0. A common technique for integrating the different components of a system is event-based invocation, also known as implicit or reactive invocation, which has historical roots in systems based on actors [15], daemons, and packet-switched networks. Event-based integration is attractive as it strongly supports software reuse and facilitates system evolution [12,11]. In bringing the benefits of component-based software development to the domain of high-performance computing, our work does not seek to create a complete component framework. Instead, we have concentrated on providing the integration mechanism that will allow the community to obtain the advantages of such architectures while maintaining high performance. This paper discusses the results of our work, an efficient event-based middleware, ECho, through which systems of distributed collaborating components can be constructed. Several attributes of ECho distinguish it from related work: High performance sharing of distributed data. ECho transports distributed data with performance similar to that achieved by systems like MPI. This level of performance is required if the integration mechanism is to support the normally large data flows that are part of high performance applications. For a distributed vi- sualization, for example, this level of performance enables end users to interact via meaningful data sets generated at runtime by the computational models being employed. This paper demonstrates ECho's high performance across heterogeneous hardware platforms, using networked machines resident at Georgia Tech. In previous work, we have used ECho in Internet-wide collabora- tions[17], and we have demonstrated its ability to represent both the control and the data events occurring in distributed computational workbenches. s Dynamic data provision and consumption. ECho supports the publish/subscribe model of communica- tion. Thus new components can be introduced into an ECho-based system simply by registering them to the right set of events in the system, without need for re-compilation or re-linking. In addition, components can be dynamically replaced without affecting other components in the system, facilitating system evolution. Event-based publish/subscribe models like the one offered by ECho have become increasingly popular and their utility within a variety of other environments, including Internet- and E-commerce applications[30], extensible systems[23], collaborative systems[14], distributed virtual reality[19] and mobile systems[32], has been well-established. ECho differs from such ongoing or past research in its efficient support for event transmission across heterogeneous machines, derived from its ability to recognize and provide runtime translation for user-defined event formats. While systems like In- foBus[20] and Schooner[16] have demonstrated the utility of making type information available to middleware, neither have attempted to attain the high performance achieved by ECho. Dynamic type extension and reflection. One of the major features differentiating component-based applications from their tightly-coupled kin is the relative lack of a priori knowledge about data flows. In order to be able to "drop" a component into place in a system, the component must be able to discover the contents of the data flows it is to operate upon. Even the parts of an application that were designed to work together face difficulty maintaining a priori knowledge in a wide area Grid environment. As different pieces of an application are changed or upgraded over time it may be necessary to modify their data flows, invalidating other pieces that rely on previous knowledge and/or requiring their simultaneous upgrade. Because of these diffi- culties, component-based systems typically provide an integration mechanism that offer some degree of type extension and reflection. Those terms, borrowed from object-oriented systems, express the ability to transparently extend existing data types while preserving the validity of code using the old type (type extension) and the ability for third parties to discover the contents of and operate upon a data type without a priori knowledge (reflection). One of the most important contributions of ECho is that it provides these features without compromising performance, as measurements in this paper will demonstrate. Interoperability. ECho-based applications can also interoperate with CORBA- or Java-based components, like those used in the Diesel Combustion Collaboratory or the Hydrology workbench. Thus, end users can continue to employ tools like the Java-based VisAD data visualization system or the CORBA-based collaboration services in Deepview, but gain high performance for data movement (in contrast to event rates attained for CORBA- or Java-based event systems[3,31]). Inter-operability with Java- and CORBA-based systems will be demonstrated elsewhere. ECho has been available since October 1997, and our group has used it for various large-scale, ongoing development and research efforts. Among such efforts, of principal interest to the high performance community are the atmospheric and hydrology applications mentioned earlier as well as two additional ones now being developed by our group: (1) a distributed materials design workbench, where multiple end users interact with each other and with computational tools in order to jointly design high performance materials, and (2) a distributed implementation of an NT-Unix-spanning system for molecular dynamics and/or for crystal plasticity studies done by collaborators in the departments of Mechanical Engineering and Physics in Georgia Tech. Finally, ECho events are one of the key building blocks of the DARPA-funded Infosphere Information Technology Expedition[26]. The remainder of this paper is organized as follows. Section 2 describes ECho's basic functionality. Section 3 compares ECho's event delivery performance to that of other communication systems which offer some form of type extension and reflection. In particular, we examine the performance of a set of middleware systems which might be considered as alternative candidates for the integration mechanism of a component infrastruc- ture, including CORBA event channels, event distribution via Java's RMI, and an XML-based communication scheme; comparing the basic latency of each to that of ECho and using an MPI message exchange as a baseline for measurement. We also study the impact of machine heterogeneity on ECho's performance and explore the effects of its type extension features. Finally, Section 4 discusses some key areas of future work and summarizes our conclusions. 2. ECho Functionality ECho shares semantics common to a class of event delivery systems that use channel-based subscriptions. That is, an event channel is the mechanism through which event sinks and sources are matched. Source clients submit events to a specific channel and only the sink clients subscribed to that channel are notified of the event. Channels are essentially entities through which the extent of event propagation is controlled. The CORBA Event Service[13] is also channel-based, with channels being distributed objects. Channel Process C Process B Process A Channel Event Channel Event Event (a) Abstract view of Event Channels. Channel Event Channel Channel Event Event Channel Event Event Channel Event Event Channel Process C Process B Process A Channel (b) ECho Realization of Event Channels. Figure 1. Using Event Channels for Communication. 2.1. Efficient Event Propagation Unlike many CORBA event implementations and other event services such as Elvin[28], ECho event channels are not centralized in any way. Instead, channels are light-weight virtual entities. Figure 1a depicts a set of processes communicating using event channels. The event channels are shown as existing in the space between processes, but in practice they are distributed entities, with bookkeeping data residing in each process where they are referenced as depicted in Figure 1b. Channels are created once by some process, and opened anywhere else they are used. The process which creates the event channel is distinguished, in that it is the contact point for other processes wishing to use the channel. The channel ID, which must be used to open the channel, contains the contact information for the creating process (as well as information identifying the specific channel). However, event distribution is not centralized and there are no distinguished processes during event propagation. Event messages are always sent directly from an event source to all sinks and net-work traffic for individual channels is multiplexed over shared communications links. ECho is implemented on top of DataExchange[10] and PBIO[8], packages developed at Georgia Tech to simplify connection management and heterogeneous binary data transfer. As such, it inherits from these packages portability to different network transport layers and threads packages. DataExchange and PBIO operate across the various versions of Unix and Windows NT, have been used over the TCP/IP, UDP, and ATM communication protocols and across both standard and specialized network links like ScramNet[6]. In addition to offering interprocess event delivery, ECho also provides mechanisms for associating threads with event handlers allowing a form of intra-process communication. Local and remote sinks may both appear on a channel, allowing inter- and intra-process communication to be freely mixed in a manner that is transparent to the event sender. When sources and sinks are within the same address space, an event is delivered by directly placing the event into the appropriate shared-memory dispatch queue. While this intra-process delivery can be valuable, this paper concentrates on the aspects of ECho relating to remote delivery of events. 2.2. Event Types and Typed Channels One of the differentiating characteristics of ECho is its support for efficient transmission and handling of fully typed events. Some event delivery systems leave event data marshalling to the application. ECho allows types to be associated with event channels, sinks and sources and will automatically handle heterogeneous data transfer issues. Building this functionality into ECho using PBIO allows for efficient layering that nearly eliminates data copies during marshalling and unmarshalling. As others have noted[18], careful layering to minimize data copies is critical to delivering full network bandwidth to higher levels of software abstrac- tion. The layering with PBIO is a key feature of ECho that makes it suitable for applications which demand high performance for large amounts of data. Base Type Handling and Optimization. Functionally, ECho event types are most similar to user defined types in MPI. The main differences are in expressive power and implementation. Like MPI's user defined types, ECho event types describe C-style structures made up of atomic data types. Both systems support nested structures and statically-sized arrays. ECho's type sys- smilliseconds (logscale)Data Structure Size Round-trip Latency1Kb 100b 100Kb10.0.1 Java CORBA MPICH ECho Figure 2. A comparison of latency in basic data exchange in event infrastructures tems extends this to support null-terminated strings and dynamically sized arrays. 1 While fully declaring message types to the underlying communication system gives the system the opportunity to optimize their transport, MPI implementations typically do not exploit this opportunity and often transport user defined types even more slowly than messages directly marshalled by the application. In contrast, ECho and PBIO achieve a performance advantage by avoiding XDR, IIOP or other 'wire' representations different than the native representation of the data type. Instead, ECho and PBIO use a wire format that is equivalent to the native data representation (NDR) of the sender. Conversion to the native representation of the receiver is done upon receipt with dynamically generated conversion routines. As the measurements in [4] show, PBIO 'encode' times do not vary with data size and 'decode' times are much faster than MPI. Because as much as two-thirds of the latency in a heterogeneous message exchange is software conversion overhead[4], PBIO's NDR approach yields round-trip message latencies as low as 40% of that of MPI. Type Extension. ECho supports the robust evolution of sets of programs communicating with events, by allowing variation in the data types associated with a single channel. In particular, an event source may submit an event whose type is a superset of the event type associated with its channel. Conversely, an event sink may have a type that is a subset of the event type associated with its channel. Essentially this allows a new field to be added to an event at the source without invalidating existing event receivers. This functionality can be extremely valuable when a system evolves because it means that event contents can be changed without the need to simultaneously upgrade every component 1 In the case of dynamically sized arrays, the array size is given by an integer-typed field in the record. Full information about the types supported by ECho and PBIO can be found in [8].1Kb Data Structure Size Delivered Bandwidth 100b Mbits/secCORBA Java MPICH ECho Figure 3. A comparison of delivered bandwidth in event infrastructure to accommodate the new type. ECho even allows type variation in intra-process communication, imposing no conversions when source and sink use identical types but performing the necessary transformations when source and sink types differ in content or layout. The type variation allowed in ECho differs from that supported by message passing systems and intra- address space event systems. For example, the Spin event system [23] supports only statically typed events. Similarly, MPI's user defined type interfaces do not offer any mechanisms through which a program can interpret a message without a priori knowledge of its con- tents. Additionally, MPI performs strict type matching on message sends and receives, specifically prohibiting the type variation that ECho allows. In terms of the flexibility offered to applications, ECho's features most closely resemble the features of systems that support the marshalling of objects as mes- sages. In these systems, subclassing and type extension provide support for robust system evolution that is substantively similar to that provided by ECho's type vari- ation. However, object-based marshalling often suffers from prohibitively poor performance. ECho's strength is that it maintains the application integration advantages of object-based systems while significantly out-performing them. As the measurements in the next section will show, ECho also outperforms more traditional message-passing systems in many circumstances. 3. ECho Performance 3.1. Breakdown of Costs Figures 2 and 3 represent the basic performance characteristics of a variety of communication infrastructures that might be used for event-based communication in high performance applications. The values are of basic event latency and bandwidth in an environment con- 5ECho CORBA MPICH XML Total Round-Trip 30.6 53.0 80.1 1249 Network Transfer 13.9 13.9 13.9 182 x86 Decode 1.6 1.6 11.6 276 x86 Encode 0.015 0.64 8.9 124 Network Transfer 13.9 13.9 13.9 182 Table Cost breakdown for heterogeneous 100Kb event exchange (times are in milliseconds). sisting of a x86-based PC and a Sun Sparc connected by 100 Mbps Ethernet. 2 Note the use of a logarithmic vertical scale in Figure 2. This is useful in presenting latencies for a range of message sizes on the same graph, but it tends to minimize the substantial performance advantage that ECho demonstrates as compared to the other infrastructures. The infrastructures compared don't all share the same characteristics and features, a fact that accounts for some of their performance differences. ECho's strength is that it provides the important features of these systems while maintaining the performance achieved by traditional high-performance systems like MPICH. In particular, ECho provides for event type discovery and dynamic type extension in a manner similar to that of XML, or that which can be achieved by serializing objects as events (as in Java RMI). CORBA is also gaining acceptance as distributed systems middleware and its Event Services provide similar features. This section will examine ECho's performance characteristics in more detail and contrast them with these other infrastructures. Table 1 shows a breakdown of costs involved in the roundtrip event latency measures of Figure 2. We present round-trip times because they naturally show all the combinations of send/recv on two different architectures in a heterogeneous system. The time components labeled "Encode" represent the span of time between an application submitting data for transmission and the point at which the infrastructure invokes the underlying network 'send()' operation. The "Net- work Transfer" times are the one-way times to transmit the encoded data from sending to receiving ma- chines. The "Decode" times are the time between the end of the "recv()" operation and the point at which the data is presented to the application in a usable form. This breakdown is useful for understanding the different costs of the communication and in particular, how they might change with different networks or processors. 2 The Sun machine is an Ultra 30 with a 247 MHz cpu running Solaris 7. The x86 machine is a 450 MHz Pentium II, also running Solaris 7. 100Kb.00110.0 millisecond (logscale) .1 CORBA MPICH Sender Encode Times Data Structure Size Figure 4. Send-side data encoding times. We have excluded Java RMI from the breakdown in Table 1 because it performs its network 'send()' operations incrementally during the marshalling process. This allows Java to pipeline the encode and network send operations making a simple cost breakdown impos- sible. However, as a result of this design decision, Java RMI requires tens of thousands of kernel calls to send a 100Kb message, seriously impacting performance. Additionally, while the round-trip times listed in Table are near the sum of the encode/xmit/decode times, this is not true for the CORBA numbers. This is because implementations of the CORBA typed event channel service typically rely on CORBA's dynamic invocation interface to operate. In the ORBs we have ex- amined, DII does not function for intra-address-space invocations. The result of this is that the CORBA typed event channel must reside in a different address space than either the event source or event sink, adding an extra hop to every event delivery. This could be considered an implementation artifact that might be handled differently in future CORBA event implementations 3.1.1. Sending side costs ECho's most significant performance feature is its use of the native data format on the sending architecture as its 'wire format'. The effects of this approach are most noticeable when comparing the "Encode" times for the different communication infrastructures. Figure 4 expands upon the summary data of Table 1 and shows sender-side data encoding costs for a variety of communications systems. For example, MPICH uses a very slow interpreted marshalling procedure for heterogeneous communication of MPI user-defined data types. That this has a significant impact on MPICH performance sis apparent in Table 1 which shows MPICH devoting as much as 60% of its round-trip message time to encoding and decoding. CORBA's IIOP wire format differs from the architec- tures' native data layout in its alignment requirements. As a result, CORBA must copy all application data before sending. In ORBacus, this copy is performed by compile-time-generated stub code, resulting in better performance than with the MPICH approach. How- ever, ECho is significantly faster because it performs very little processing prior to the network send operation Using XML as a wire format is obviously a decision which has a significant performance impact on an event system. Table 1 makes clear two of the most significant issues: the large encode/decode times, and the expanded network transmission times. The former is a result of the distance between the ASCII representation used by XML and the native binary data repre- sentation. XML encoding costs represent the processing necessary to convert the data from binary to string form and to copy the element begin/end blocks into the output string. Just one end of the encoding time for XML is several times as expensive as the entire round-trip message exchange for the other infrastructures. Net-work time is also significantly higher for XML because the ASCII-encoded data (plus the be- gin/end labels) can be much larger than the equivalent binary representation. How much larger depends upon the data, the size of the field labels and other details in the encoding. Thus, XML-based schemes transmit more data than schemes which rely on binary encoding. 3.2. Receiving side costs ECho's NDR approach to binary data exchange eliminates sender-side processing by transmitting in the sender's native format and isolating the complexity of managing heterogeneity in the receiver. As a result, the receiver must perform conversion of the various incoming 'wire' formats to its `native' format. Such a conversion may require byte-order changes (byte-swapping), movement of data from one offset to another, or even a change in the basic size of the data type (for example, from a 4-byte integer to an 8-byte integer). This conversion is another form of the "marshal- ing problem" that occurs widely in RPC implementa- tions[1] and in network communication. Marshaling can be a significant overhead[7,27], and tools like USC[21] attempt to optimize marshaling with compile-time so- lutions. Unfortunately, the dynamic form of the marshaling problem in ECho, where the layout and even the complete field contents of the incoming record are unknown until run-time, rules out such static solutions. The conversion overhead is nil for some homogeneous data exchanges, but as Table 1 shows, can be significantly high for some heterogeneous exchanges (up to 66%). Generically, receiver-side overhead in communication middleware has several components: ffl byte-order conversion, ffl data movement costs, and ffl control costs. Byte order conversion costs are to some extent unavoid- able. If the communicating machines use different byte orders, the translation must be performed somewhere regardless of the capabilities of the communications package. Data movement costs are harder to quantify. If byte- swapping is necessary, data movement can be performed as part of the process without incurring significant additional costs. Otherwise, clever design of the communications middleware can often avoid copying data. How- ever, packages that define a 'wire format' for transmitted data have a harder time being clever in this area. One of the basic difficulties is that the native format for mixed-datatype structures on most architectures has gaps, unused areas between fields, inserted by the compiler to satisfy data alignment requirements. To avoid making assumptions about the alignment requirements of the machines they run on, most packages use wire formats which are fully packed and have no gaps. This mis-match forces a data copy operation in situations where a clever communications system might otherwise have avoided it. costs represent the overhead of iterating through the fields in the record and deciding what to do next. Packages that require the application to marshal and unmarshal their own data have the advantage that this process occurs in special-purpose compiler- optimized code, minimizing control costs. However, to keep that code simple and portable, such systems uniformly rely on communicating in a pre-defined wire for- mat, therefore incurring the data movement costs described in the previous paragraph. Packages that marshal data themselves typically use an alternative approach to control, where the marshaling process is controlled by what amounts to a table-driven interpreter. This interpreter marshals or unmarshals application-defined data, making data movement and conversion decisions based upon a description of the structure provided by the application and its knowledge of the format of the incoming record. This approach to data conversion gives the package significant flexibility in reacting to changes in the incoming data and was our initial choice when implementing the PBIO technology ECho is based on. necessarily takes a different approach to receiver-side decoding. Because the 'wire' format is a continuous string, XML is parsed at the receiving end. The Expat milliseconds (logscale) 100Kb 10Kb 1Kb 100b10.0Data Structure Size MPICH CORBA ECho Receiver Decode Times Figure 5. Receiver side costs for XML, MPI and NDR. (Logarith- mic scale used in vertical dimension) XML parser 3 [5] calls handler routines for every data element in the XML stream. That handler can interpret the element name, convert the data value from a string to the appropriate binary type and store it in the appropriate place. This flexibility makes XML extremely robust to changes in the incoming record. The parser we have employed is quite fast, but XML still pays a relatively heavy penalty for requiring string-to-binary conversion on the receiving side. (We assume that for most high performance computing functions, data is being sent somewhere for processing and that processing requires the event data to be in other than string form. Thus, XML decoding is not just parsing, but also the equivalent of a C strtod() or similar operation to convert the data into native representation.) Comparison of receiver-side costs for XML, ECho, and non-optimized ECho wire formats. Figure 5 shows a comparison of receiver-side processing costs on the Sparc for interpreted converters used by XML, MPICH (via the MPI-Unpack()) call, and NDR. receiver conversions are clearly expensive, typically between one and two orders of decimal magnitude more costly than our NDR-based converter for this heterogeneous exchange. (For an exchange between homogeneous architectures, ECho and MPI would have substantially lower costs, while XML's costs would remain unchanged.) Our NDR-based converter is relatively heavily optimized and performs considerably better than MPI, in part because MPICH uses a separate buffer for the unpacked message rather than reusing the receive buffer (as we do). However, ECho's receiver-side conversion costs still contribute roughly 20% of the cost 3 A variety of implementations of XML, including both XML generators and parsers, are available. We have used the fastest known to us at this time, Expat [5]. milliseconds (logscale) 100b 100Kb1.0 .1 Data Structure Size ECho DCG ECho MPICH Receiver Decode Times Figure 6. Receiver side costs for interpreted conversions in MPI and ECho and DCG conversions in ECho. of an end-to-end message exchange. While a portion of this conversion overhead must be the consequence of the raw number of operations involved in performing the data conversion, a significant fraction of this overhead is due to the fact that the conversion is essentially being performed by an interpreter. Optimizing receiver-side costs for ECho. Our decision to transmit data in the sender's native format results in the wire format being unknown to the receiver until run-time, requiring a somewhat costly interpreted conversion. Our solution to this problem to the problem is to employ dynamic code generation to create a customized conversion subroutine for every in-coming record type 4 . These routines are generated by the receiver on the fly, as soon as the wire format is known. The execution times for these dynamically generated conversion routines are shown in Figure 6. (We have chosen to leave the XML conversion times off of this figure to keep the scale to a manageable size. Again, please note the use of logarithmic scale.) The dynamically generated conversion routine operates significantly faster than the interpreted version. This improvement removes conversion as a major cost in communication, bringing it down to near the level of a copy operation. In particular, without dynamic code generation for conversion routines ECho's practice of eliminating the sender-side encoding cost by transmitting in the sender's native format might not be viable. It is the combination of these two techniques, ffl transmitting in the sender's native format and ffl using dynamic code generation for conversion routines on the receiving side, that is the key to ECho's efficiency. The PBIO binary code generation facilities exploited by ECho exist for most popular machine architectures, including Sparc, 4 More details on the nature of the ECho infrastructure dynamic code generation can be found in [9]. s ORBacus ECho Send Receive Send Receive Data size side side side side overhead overhead overhead overhead Table Cost breakdown for homogeneous event exchange. (Times are in milliseconds.) MIPS, x86, and i960 machines. 3.3. Costs for Homogeneous Exchanges Because ECho has virtually no sender-side encoding costs and because its dynamic code generation achieves performance similar to that achieved through compile-time stub generation, ECho tends to outperform other communication infrastructures. This is particularly apparent in heterogeneous message exchanges because the encode/decode time can play a significant role in overall message costs. However, ECho's approach also yields performance gains for transfers between homogeneous systems, as shown in Table 2. For simplicity, this table concentrates on the ECho and ORBacus infrastructures. The higher ORBacus costs for large data sizes represent the cost of the required data copy in converting the IIOP wire format to the native data representation. ECho requires no such copy. 5 As in the heterogeneous case, ECho does not pre-process data prior to sending, and because the 'wire format' corresponds to the native data representation, ECho can deliver received data directly to the application without copying it from the message buffer. This is not possible with IIOP because of potential data alignment conflicts between IIOP and the native data representation. At common 100Mbps network speeds, these additional data copy operations are account for a relatively small fraction of the total exchange costs. How- ever, minimizing data copies is critical to delivering full network bandwidth to higher levels of software ab- straction[18]. As gigabit networks and specialized low-latency communications mechanisms come into more common use, the additional copy operations imposed on even homogeneous communications by fixed wire formats will become a more important limitation on communication speeds, increasing ECho's performance advantage. 5 For the smaller data sizes, the extra copy overhead is small compared to the fixed delivery costs in these systems. .1 .01Mismatching fields000000000000111111111111000000001111111100001111Matching fields00001111 100Kb Hetergeneous Receive Times Data Structure Size milliseconds (logscale) Figure 7. Receiver-side decoding costs with and without an unexpected field - heterogeneous case. 3.4. Costs for Type Extension In addition to efficient operation in basic event fer, ECho supports the creation and evolution of sets of collaborating programs through event type discovery and dynamic type extension. ECho events carry format meta-information, somewhat like an XML-style description of the message content. This meta-information can be an incredibly useful tool in building and deploying enterprise-level distributed systems because it (1) allows generic components to operate upon data about which they have no a priori knowledge, and (2) allows the evolution and extension of the basic message formats used by an application without requiring simultaneous upgrades to all application components. In other terms, ECho allows reflection and type extension. Both of these are valuable features commonly associated with object systems. data type information is represented during transmission with format tokens which can be used to retrieve full type information. These tokens are small and are included in every ECho event transmission as part of the header information. As such they do not affect performance significantly. ECho supports type extension by virtue of doing field matching between incoming and expected records by name. Because of this, new fields can be added to events without disruption because application components which don't expect the new fields will simply ignore them. Most systems which support reflection and type extension in messaging, such as systems which use XML as a wire format or which marshal objects as messages, suffer prohibitively poor performance compared to systems such as MPICH and CORBA which have no such sup- port. Therefore, it is interesting to examine the effect of exploiting these features upon ECho performance. In particular, we measure the performance effect of type extension by introducing an unexpected field into the incoming message and measuring the change in receiver-side processing. Figures 7 and 8 present receive-side processing costs .1 .010000111100001111Matching fields0000111100001111Mismatching fields 100Kb 10Kb 1Kb 100b milliseconds (logscale) Data Structure Size Homogeneous Receive Times Figure 8. Receiver-side decoding costs with and without an unexpected field - homogeneous case. for an exchange of data with an unexpected field. These figures show values measured on the Sparc side of heterogeneous and homogeneous exchanges, respectively, using ECho's dynamic code generation facilities to create conversion routines. It's clear from Figure 7 that the extra field has no effect upon the receive-side per- formance. Transmitting would have added slightly to the network transmission time, but otherwise the support of type extension adds no cost to this exchange. Figure 8 shows the effect of the presence of an unexpected field in the homogeneous case. Here, the overhead is potentially significant because the homogeneous case normally imposes no conversion overhead in ECho. The presence of the unexpected field creates a layout mismatch between the wire and native record formats and as a result the conversion routine must relocate the fields. As the figure shows, the resulting overhead is non-negligible, but not as high as exists in the heterogeneous case. For smaller record sizes, most of the cost of receiving data is actually caused by the overhead of the kernel select() call. The difference between the overheads for matching and extra field cases is roughly comparable to the cost of memcpy() operation for the same amount of data. The results shown in Figure 8 are actually based upon a worst-case assumption, where an unexpected field appears before all expected fields in the record, causing field offset mismatches in all expected fields. In general, the overhead imposed by a mismatch varies proportionally with the extent of the mismatch. An evolving application might exploit this feature of ECho by adding any additional at the end of existing record formats. This would minimize the overhead caused to application components which have not been updated. 4. Conclusions and Future Work This paper examines ECho, an event-based middle-ware designed to meet the demands of a new generation of Grid applications. In particular, we consider the communication/integration demands of component-based systems in a high-performance computing environment and how they might be different from those of more tightly-coupled applications. ECho meets those requirements by providing a publish-subscribe communication model that supports type extension and type discovery. While object-based and XML-based systems provide similar functionality, the measurements in Section 3 show that ECho does it with significantly better performance, both in terms of delivered bandwidth and end-to-end latency. The measurements also show that ECho matches and, in most cases, outperforms MPICH in both metrics supporting our assertion that ECho is suitable for use in the main data flows of Grid applications Future work will examine aspects of ECho which are beyond the scope of this paper. Those features include derived event channels, which support for source-side event filtering and remote data transformation, and proto-channels, a mechanism through which receivers can themselves control and customize source-side event generation. We will also expand upon ECho's ties to other systems, including CORBA and Java. --R The impact of language and system on remote procedure call design. Katarzyna Kea- hey Asynchronous notifications among distributed objects. Efficient wire formats for high performance computing. Scramnet networks. Architectural considerations for a new generation of protocols. Portable self-describing binary data streams Fast heterogeneous binary data interchange. performance communication in distributed laboratories. Formalizing design spaces: Implicit invocation mechanisms. An introduction to software architecture. Object Management Group. A language for proving theorems in robots. Configuring scientific applications in a heterogeneous distributed system. ACDS: Adapting computational data streams for high performance. Efficient layering for high speed communication: Fast messages 2. The information bus - an architecture for extensible distributed systems Universal stub compiler. The diesel combustion collaboratory: Combustion researchers collaborating over the internet. Dynamic binding for an extensible system. A channel for distributed microscopy and informatics. Foundations for the study of software architecture. Performance or Firefly RPC. Elvin has left the building: A publish/subscribe notification service with quenching. Software Architecture: Perspectives on an Emerging Discipline. An information flow based approach to message brokering. The jini An infrastructure for collaborating mobile agents. --TR --CTR James A. Kohl , Torsten Wilde , David E. Bernholdt, Cumulvs: Interacting with High-Performance Scientific Simulations, for Visualization, Steering and Fault Tolerance, International Journal of High Performance Computing Applications, v.20 n.2, p.255-285, May 2006
publish/subscribe communication;event services;component architectures;middleware;dynamic extensibility
592978
Scalable Session Locking for a Distributed File System.
File systems provide an interface for applications to obtain exclusive access to files, in which a process holds privileges to a file that cannot be preempted and restrict the capabilities of other processes. Local file systems do this by maintaining information about the privileges of current file sessions, and checking subsequent sessions for compatibility. Implementing exclusive access in this manner for distributed file systems degrades performance by requiring every new file session to be registered with a lock server that maintains global session state. We present two techniques for improving the performance of session management in the distributed environment. We introduce a distributed lock for managing file access, called a i>semi-preemptible lock, that allows clients to cache privileges. Under a semi-preemptible lock, a file system creates new sessions without messages to the lock manager. This improves performance by exploiting locality the affinity of files to clients. We also present data structures and algorithms for the i>dynamic evaluation of locks that allow a distributed file system to efficiently manage arbitrarily complex locking. In this case, complex means that an object can be locked in a large number of unique modes. The combination of these techniques results in a distributed locking scheme that supports fine-grained concurrency control with low memory and message overhead and with the assurance that their locking system is correct and avoids unnecessary deadlocks.
Introduction Distributed file systems have become the principal method for sharing data in distributed applications. Programmers understand file system semantics well, and use them to easily gain access to shared data. For exactly the same reason that distributed file systems are easy to use, they are difficult to implement. The distributed file system takes responsibility for providing synchronized access and consistent views of shared data, shielding the application and programmer from these tasks, but moving the complexity into the file system. A distributed client-server file system presents a local file system interface to remote and shared data. The file system client takes responsibility for implementing the semantics of the local file system and translating the local interface onto a client/server network protocol. For heterogeneous distributed file systems (many client operating systems), the system may be translating the semantics of several different local file systems onto a single network protocol. In this work, we present a locking construct, called file session locks, that implement sessions on files defined by open and close calls from an application. File sessions enforce concurrency constraints; e.g., if one client opens a file for exclusive writing, permitting no concurrent readers, opens for read on other clients must be forbidden. These locks are not designed to provide data consistency or cache coherency - a suitable cache coherency protocol is required in addition to file session locking. Instead, the locks allow clients to choose from among the many exclusive access and sharing options available in its native file system interface and have the semantics of the local open enforced throughout a distributed system. To help encode and enforce sessions in a distributed environment, we contribute the semi-preemptible lock, which allows file system client to cache session privileges. A client holding a semi-preemptible lock on a file has the right to access a file in any of the modes specified by its held lock. Clients maintain their own file open (session) state locally, and do not need to transact with the file system server when opening or closing a file. Clients may continue to hold such a lock even when they have no open instances. In this way, a client can cache access privileges, the right to open a file, and service subsequent open requests without a message to the server. This mechanism reduces server traffic by eliminating open and close messages, and consequently reduces latency by avoiding message round trip time. Clients cache access privileges to a file on the belief that the file will be used again locally before being used by another client. Semi-preemptible locks also reduce distributed lock state. Clients often hold multiple open instances of a single file concurrently. Clients locally create concrete locks for each session which are held under (or in the context of) a distributed lock. All of these open instances can be granted under a single semi-preemptible lock, rather than holding a separate lock for every open. A single semi-preemptible lock summarizes all of the client's open state to the distributed system. Semi-preemptible locks combine two concepts in lock management. First, clients manage locks hierarchically to separate session state from lock state. Second, clients use semi-preemptibility, also called lazy-revocation or sticky locks [24], to retain privileges on files in the absence of active sessions. We use the term hierarchical differently than most of the locking literature. Generally, the term hierarchical describes locks held concurrently on different levels of abstraction on the same data object to achieve granular locking [12, 27]. Instead, we use the term to describe two levels of locking - the higher level is abstract, summarizing client lock state to the server, and the lower level is concrete, representing sessions. The interface to open a file (create a session) has many options that allow an application to specify its intended actions and restrict concurrent actions by other clients. To express these options, locks on these data generally have multiple locking modes each with unique semantics. The number of different possible locking modes increases exponentially with the number of access methods and locking for a complex data object using existing methods quickly becomes unmanageable. We take the Windows-NT interface as an example to establish this point. The file system takes six binary arguments when opening a file which requires 64 unique locking modes to describe all sessions. Considering locking modes pairwise for compatibility, there are 4096 combinations. Locks are compatible if they can be held concurrently. Furthermore, the large number of modes in this interface do not fully specify all possible access modes. For example, they do not address concurrent or exclusive access to file system metadata. Existing methods for lock management fall short in that they either fail to scale well as the number of locking modes become large, or they are ad hoc systems that are poorly specified. Database systems employ a static data structure, called a lock compatibility table [12], that describes the relationships between all locking modes pairwise. This data structure fully specifies the interactions between locking modes, but grows quadratically in size with respect to the number of locking modes. Alternatively, modern file systems use ad hoc rule-based methods to evaluate locks that are efficient and compact [16]. However, without formally specified semantics, the interactions between locking modes can be difficult to reason about and implement correctly. We present a formal specification of locking modes and derive from this a data structure and algorithms for the management of distributed locks called dynamic evaluation. Our methods have the advantage of scaling well with the number of locking modes. In addition, they specify fully the interactions among all locking modes, which eases the implementation of a correct locking protocol. Our system obviates the need for static data structures, such as the lock compatibility tables used in database systems, which require memory in respect to the number of legal locking modes. In our system, lock compatibility is evaluated algorithmically based on small static lock structures that require memory in O(log n). While, a distributed file system provides a good example of a complex locking system, we feel the dynamic lock evaluation has wide applicability to distributed systems. 2 A Storage Area Network File System A brief digression into the file system architecture in which we implement file session locking helps to motivate the performance advantages. In the Storage Tank project at IBM research, we are building a distributed file system on a storage area network (SAN) (Figure 1). A SAN is a high speed network that Storage Area Network Control Network Tape Server Server Server Client Client Client Server Cluster Figure 1: Schematic of the Storage Tank distributed file system on a storage area network (SAN). gives computers shared access to storage devices. Currently, SANs are being constructed on Fibre Channel networks [4]. In the future, we expect network attached storage devices to be available for general purpose data networks, so that SANs can be constructed for networks such as Gigabit Ethernet [9]. A distributed file system built on a SAN removes the server bottleneck for I/O requests by giving clients a direct data path to disks. File system clients on a SAN access data directly over the storage area network. In contrast, most traditional client-server file systems [26, 14, 16, 7] store data on the server's private disks. Clients function ship all data requests to a server that performs I/O on their behalf. Unlike traditional file systems, Storage Tank clients perform I/O directly to shared storage devices. This direct data access model is similar to the file system for network attached secure disks (NASD) [11], using shared disks on an IP network, and the Global file system [21], for SAN attached storage devices. Clients communicate with Storage Tank servers over a general purpose network to obtain file metadata. In addition to serving file system metadata, the servers manage cache coherency protocols, authentication, and the allocation of file data (managing data placement and free space). Unlike most file systems, metadata and data are stored separately. Metadata, including the location of the blocks of each file on shared storage, are kept at the server. The SAN storage devices contain only the blocks of data for the files. In this way, the shared devices on the SAN can be optimized for data traffic, block transfer of data, and the server private storage can be optimized for a metadata workload, frequent small reads and writes. The SAN environment simplifies the distributed file system server by removing its data tasks, and radically changes the server's performance characteristics. Previously, server performance was measured by data rate. Performance was occasionally limited by network bandwidth, but more often limited by the server's ability to read data from storage and write it to the network. In the SAN environment, a server's performance is more properly measured in transactions per second, analogous to a database server. Without data to read and write, the Storage Tank server performs many more transactions than a traditional file server with equal processing power. Without the relatively slow process of shipping data from the client to the server to hide protocol over- head, minimizing the message traffic for file system operations becomes important. Protocol overhead is the network and server resources and added latency used for client-server messages. In traditional client-server file systems, clients go to the server to obtain data. Because shipping data to the client takes significantly longer and uses many more resources than a single server message, the overhead associated with the messages for opening and closing a file are hidden by the cost of shipping data. In Storage Tank, protocol overhead limits performance. When the semi-preemptible lock allows a client to open a file without contacting the server, a message is avoided and time is saved on the critical path. 3 Semi-Preemptible File Locks For distributed files systems, little data sharing occurs in practice [2, 17], where data sharing indicates two clients concurrently accessing the same file. Additionally, clients often access data that they have recently used. These claims are supported by the effectiveness of data caching in this environment [19, 15]. More mature distributed file systems [14, 7, 24, 16, 1, 13, 19, 23] take advantage of this observation and cache file data at clients even when no process actively uses the data. The design decision to cache file data after a file has been closed improves performance when a subsequent open from the same system is more likely than a request from another client to access the same data. For subsequent accesses, caching improves performance by avoiding a server message and data read from disk. However, if another client attempts to access the same data, it sees additional latency while the file system server invalidates the cache of the client that holds data before granting access to the new client. For the same reasons that caching improves performance on data access, the semi-preemptible lock improves performance on file open. When a local process requests an open for a file, if no lock is held, the client obtains a semi-preemptible lock before granting the open. When the local process closes the file, the client records that there are no open file instances currently using the lock, but holds onto the lock awaiting future opens of the same file. Analogous to data caching, holding access locks past close decreases latency by avoiding a server message on subsequent opens. But, caching locks adds latency to opens from other clients, because the server must revoke the lock before granting access to the other client. Server Client A Client B Open Instance Locks Held Open Files Client A Semi-Preemptible Demand Messages Open State at Clients Client B Figure 2: Demands for the semi-preemptible lock are accepted or denied depending upon client file system open state. By recording the open instances associated with each access lock, a client differentiates locks that are held to protect open files, and therefore cannot be released, from locks that are held to improve performance on subsequent opens. Consider that a second client wishes to obtain exclusive access to a file that is already locked by a first client. The server processes the second client's request by demanding 1 the lock from the first client, sending a message that requests the release of the held lock. If the first client holds an open instance of that file, it requires the held lock to protect that instance and denies demand requests from the server (Figure 2). However, if no process holds an open instance, the client no longer utilizes the held lock and releases it safely. These access locks are called semi-preemptible because the server must demand them, as if they were preemptible locks. However, a client can refuse a demand request. The semi-preemptible locking system is particularly appropriate for our SAN-based distributed file sys- tem, because data are not obtained through the server. When a lock is held, a client can directly access the data from shared storage, and need not interact with the server at all. Without direct access storage, clients must interact with servers for data, and these file systems cannot save a message from semi-preemptible locking. We use the term "demand" for consistency with existing terminology despite the fact that clients can refuse a demand. A term like "request" would more appropriately describe the server's action, but request is often used to describe the client's process for acquiring locks from the server. Managing Complexity Lock management includes both the locking protocol and the lock evaluation. Protocol describes how locks are obtained and released and, in our system, we use the semi-preemptible protocol. Lock evaluation is the enforcement of semantics of the locks themselves and the semantics are encoded in the lock mode. For file session locks, the mode defines the actions that are allowed under a lock and actions that are restricted by concurrent sessions. In a file system, there are a large number of possible locking modes, describing all possible combinations of actions and restrictions. Evaluation of locks becomes difficult as the number of locking modes grows large. Techniques that reduce the number of locking modes have been explored as an alternative to implementing the full complement of possible modes. Although these techniques reduce complexity, they negatively affect concurrency and reduce performance. In Storage Tank, we implement all possible locking modes for the greatest degree of concurrency and application correctness. However, we describe other techniques for comparison. 4.1 Implementing Subsets By implementing a chosen subset of all possible locking modes, some systems trade reduced complexity for minor semantic violations. File systems often select a set of distributed locks for file access that are simpler than the semantics of the underlying file open system call [7, 15]. The mismatch between open semantics and locking results in either concurrent opens being allowed that violate open semantics, or concurrent opens being disallowed that open semantics would permit. File systems opt to disallow legal opens because this policy impinges on concurrency rather than correctness. An extreme example of a subset is to protect an object with a single exclusive lock. This lock is very easy to reason about, implement, and manage. One client holds it at a time, and it gives that client total control over the locked object. However, this lock allows no concurrent action by other clients. For sessions, this would mean that there can only be one open instance of a file at a time. Single exclusive locks operate reasonably for a data consistency protocol, in which locks can be preempted. However, for non-preemptible locks, the concurrency restrictions are intolerable. 4.2 Multiple Simple Locks Breaking a single complex lock into multiple locks with simpler semantics is another technique for reducing complexity. While this strategy is desirable, because each individual lock has few modes and therefore is easy to manage, it introduces either deadlock or livelock and ultimately limits performance. Problems arise as the locks are not truly independent. Applications require multiple locks to be held concurrently for their operation, and application semantics tie seemingly independent locks together. Any complex set of locking modes for a single lock can be implemented as multiple locks each with simple semantics that are obtained individually. For example, a lock that protects a data file for both writing data and updating metadata can be broken down into two separate locks, one for writing and one for metadata update. However, both locks cannot be obtained atomically. For example, clients that need to obtain multiple preemptible locks to conduct an operation can experience livelock and have no guarantee that they will ever make any progress. Consider two clients trying to obtain the same set of locks. Both clients hold some locks currently and request the remaining locks. As each client obtains some locks, other locks are revoked. Neither client ever has the guarantee that it will get the full complement of locks needed to continue. Increases in the number of clients, resource contention, and number of locking modes all exacerbate livelock. Techniques that eliminate livelock introduce deadlock. To eliminate livelock, clients that need multiple locks can deny or ignore revocation on its currently held locks while awaiting the other locks it requires. The same two clients in our livelock example, each requesting the same set of locks, would be at an impasse if they ignored revocation. Neither client releases the locks it holds, and neither client obtains the locks it needs. Deadlock is generally handled through either avoidance or detection [12]. Regardless of the chosen technique, dealing with deadlock always incurs a performance penalty. Protecting all accesses using a single lock with more complex semantics, rather than multiple individual locks, avoids deadlock, livelock, and the associated inefficiencies. Rather than needing to obtain multiple locks, applications atomically obtain a single lock. The request can be evaluated immediately at the server and granted or denied. Application and lock state always progress. 5 Compatibility Tables For comparison against our system of dynamic evaluation, we present the compatibility table which has been developed to manage distributed concurrency in databases [12]. Our key criticism of this structure is that it does not scale well when the number of locking modes becomes large. To be fair, in databases the number of locking modes tends to be very small (in contrast to file systems) and these static data structures are more than adequate. A compatibility table determines whether incoming lock requests can be granted in conjunction with currently outstanding locks. A system defines a set of locks and fills out this data structure, which defines Symbol Name Description r Read Reader lock w Write Reader and writer lock s Shared Reader lock, no writers Update Writer lock, no writers Exclusive No readers or writers Held Requested None r s w u x Table 1: Locking modes and the compatibility table for file system open mode locks. how these locks are managed. When opening a file, clients select a locking mode that best matches the semantics of its open and requests the corresponding lock from a locking server. The server uses the compatibility table to evaluate whether the incoming request can be serviced. We present an example based on the locking modes and compatibility table in Table 1, which are a subset of all possible locks. The table consists of rows that index the mode of the incoming lock request and columns that index locks currently held by other clients. The table cells that hold a plus indicate that the request is compatible with the outstanding lock, or a minus indicating that the requested and held lock are in conflict. For example, if a server receives a request for a w (write) lock on a file with outstanding locks r (read) and s (shared), it evaluates the locks as follows. It first looks up the cell for requested mode w and held mode r and sees that the lock modes are compatible. It then continues to evaluate w against s and, seeing that they are incompatible, the w lock cannot be immediately granted given current lock state. The server demands the s lock and grants the w lock if the demand succeeds. For n locking modes, the compatibility table contains n 2 entries. For databases and other systems with few locks, programming a compatibility table presents no problems. However, when we looked at implementing this structure for all possible locks in a distributed file system, the amount of state was daunting, and we were concerned with the correctness and maintenance of our implementation. 6 Dynamic Lock Evaluation To reduce the memory used by static locking data structures, we define simple algorithms for evaluating the compatibility of lock requests. These algorithms require the client and server to store only the number of allowable access methods, k, and need not record every lock mode and the associated compatibility table. These algorithms support all possible locking modes for a given set of access modes, unlike systems that implement a subset of all possible locks to limit complexity. Our algorithms use the following quantities: L Set of locking modes A Set of valid access modes permitted access modes disallowed sharing modes Each lock X can be appropriately thought of as an ordered pair hP X ; DX i of the access modes that it permits and the access modes it disallows. Permitted modes are the access methods that the lock holder can perform. Disallowed modes are the access methods that the lock holder forbids other clients from performing concurrently. Both the permitted and disallowed access modes vary over all possible subsets of the set A. Thus for A containing k different access modes, there are 2 2k different potential locks, because the number of distinct subsets of A equals 2 k . To develop algorithms for evaluating lock state, we begin by casting the definition of lock compatibility in set theoretic terms. From this definition, we create simple algorithms for evaluating lock requests at the server and processing local open requests and server lock demands at the client. Lock compatibility is evaluated in a dynamic fashion; i.e., compatibility need not be pre-computed and stored in tables. Definition 6.1 (Compatibility) Lock X and Y are compatible iff PX \ D Two locks are compatible if they do not forbid the access modes that the other lock protects. Compatibility must be symmetric so that the lock state in a distributed system has no dependence on the order in which locks are acquired. This helps avoid non-deterministic behavior due to race conditions. For illustration, consider that compatibility were not symmetric. We could have two locks, X and Y , with X compatible with Y but not vice versa. If two clients were obtaining these locks and lock X were obtained first, then another client could get lock Y . On the other hand, if Y were obtained first, X would be unavailable. The final lock state would vary depending upon the order in which the lock requests arrive. In addition to lock compatibility, used to evaluate lock requests, clients require the concepts of lock strength and weakness to manage locks hierarchically and determine lock transitions. Strength and weakness are used by clients when either upgrading a held lock to service a local file open request or downgrading a lock in response to the server's demand. Definition 6.2 (Strength) Lock X is stronger than lock Y iff P Y PX and D Y Definition 6.3 (Weakness) Lock X is weaker than Y iff Y is stronger than X . A stronger lock permits more access methods and restricts sharing when compared with a weaker lock. The strength and weakness definitions include identical locks; i.e., two identical locks are both mutually stronger than each other and weaker than each other. Although this abstraction seems incongruous, it simplifies locking algorithms by addressing boundary conditions. The following theorems help show that these definitions match our intuition for file access locking and have semantics identical to the compatibility, strength, and weakness of static locking data structures. First, a lock hierarchy can only be valid if the strength and weakness relations are transitive, required conditions for an ordering. Theorem 6.1 X stronger than Y and Y stronger than Z implies X stronger than Z . Proof. X stronger than Y implies that P Y PX and D Y DX . Also, Y stronger than Z implies that From these, we observe that P Z PX and DZ DX , which is exactly the condition for X stronger than Z . Corollary 1 X weaker than Y and Y weaker than Z implies X weaker than Z . Lock strength is closely related to compatibility and will be used to help determine transitions among locking modes in our system. This next theorems express key concepts that clients use to summarize locks and evaluate lock transitions. Theorem 6.2 If X is weaker than Y then all locks compatible with Y are compatible with X . Proof. When X weaker than Y , PX P Y and DX D Y . Also, for any lock Z compatible with Y , This implies that PX \ We conclude that X is compatible with Z . As we have not constrained Z , X is compatible with any lock compatible with Y . Corollary 2 If X is not compatible with Y then all locks stronger than X are also incompatible. Proof. For the sake of contradiction, conjecture an Z , stronger that X , that is compatible with Y . By Theorem 6.2 Z compatible with Y implies that X is compatible with Y . This establishes the contradiction and no such Z can exist. The presented description of locks defines the semantics that must be implemented by lock evaluation algorithms. The algorithms for dynamic lock evaluation compare lock requests and transitions to these definitions, rather than performing a look up in a preset data structure. This eliminates the memory required to store lock tables. O L1 O L2 O L3 O L4 O L3 O L3 O L2 O L3 O L4 O L4 Summary Bitlocks bit 1: bit 2: bit 3: bit 4: bit 5: bit bit 7: bit 8: Outstanding O O O O O Figure 3: Lock management data structure. 7 A Summarizing Data Structure for Lock Evaluation In addition to the formal definition of lock semantics, we require data structures at the lock requester (client) and lock granter (server) to fully specify algorithms. For lock evaluation, the server has the task of receiving incoming lock requests and evaluating the compatibility of these requests against currently outstanding locks. Obvious implementations might elect a data structure that enumerates all outstanding locks and, when receiving a new lock request, the server iterates over existing locks comparing compatibility with the requested lock. Implementations based on lock tables work in exactly this fashion, performing a table look up that compares the requested lock against each currently outstanding lock. We define a data structure that allows our algorithms to evaluate lock compatibility directly, without iteration. The fundamental concept of this data structure is to summarize the sharing restrictions and permitted access modes of all outstanding locks in a single location and evaluate incoming lock requests against the summary. The data structure also allows the server to find all locks that need to be demanded directly, without searching. This concept of summarizing lock state is similar, but more general than, using the strongest outstanding lock as a representative for all outstanding locks [12]. However, our system makes provisions for locks that are compatible yet not strength-related, which occurs when all possible combinations of access modes are allowed. The lock summary data structure contains a list of all outstanding locks in addition to the summary itself, which indexes the list of locks (Figure 3). This allows the data structure to evaluate new lock requests against the summary, but also permits the contribution of any individual lock to the summary to be calculated and removed when the lock is released. In the lock summary data structure, we use a bit vector to represent any individual lock. For a locking system with k unique access modes, the bit vector contains 2k bits. The first k bits correspond to the set of access modes that a lock permits. The next k bits indicate the set of access modes that the lock disallows. For example, in a locking system with metadata read, metadata write, read, and write access modes (m r the bit vector h1011 : 0001i describes a lock that permits metadata read and data read and write while disallowing other clients from holding a write privilege. We punctuate the bit vector with a colon to indicate the semantic differences between the first and second k bits. The data structure contains a list, called the Outstanding list, of all individual locks currently held by clients. The list holds a description of each lock, its 2k bits describing permitted access and concurrently disallowed locking modes (h i), and is indexed by lock identifier (L i ). stands for the bit vector representation of set P . The server picks lock identifiers for incoming lock requests, and all subsequent lock operations are conducted by lock identifier. Since all client/server lock operations are conducted in the context of lock identifier, the server can always look up entries in this list quickly. It is possible to maintain this list as an ordered index, so that locks can be looked up by name in O(1), but engineering restrictions 2 often require unique, non-decreasing lock identifiers. These lock identifiers are maintained using extendible hashing [8] for scalability and fast look up, in time O(log n) for n outstanding locks. The data structure also contains a summary of all outstanding locks. The lock summary consists of the union of all permitted access modes and the union of all disallowed concurrently held modes. For a system with locks L outstanding, the summary is: i: (1) In our data structure, we represent the summary using bit vectors for sets, as if it were a single lock. The set operations for union and intersection are computed using logical operators on bit vectors: "bit-wise or" _ for union and "bit-wise and" ^ for intersection. These bit operators can be legally applied to two vectors of the same length and operate pairwise on the bits in each vector, producing a result bit vector of the same length. Our implementation encodes the two unions of the summary in terms of bit operations, whose bit components can be expanded to arrive at an equivalent logical expression. _ _ By maintaining the union of the permitted and disallowed lock modes at all times, the server can implicitly evaluate the compatibility of an incoming lock request against all locks by evaluating the request against the summary. A requested lock it does not disallow any access modes permitted recovery [3] benefit from unique and non-decreasing lock identifiers. by the summary, and the summary does not disallow any permitted modes it accesses, i.e. The intuition behind the summary is that all the lock state in the system can be represented by an equivalent single lock, constructed from the union of all permitted and concurrently prohibited access modes. In addition to the summary and Outstanding lists, we require additional lists to aid in processing lock requests and efficiently maintaining the summary. For each bit in the lock summary, we keep a BitLocks structure that contains a list of the locks in Outstanding that set that bit high. Each BitLocks list describes the set of locks that contribute to setting an individual bit of the summary high. Our algorithms for lock management use this list to efficiently determine what locks need to be revoked when a requested lock is not compatible with current lock state. Each lock in Outstanding also has references (pointers) to the locations in all BitLocks lists in which the lock appears. When a lock is released, the lock management algorithms use these references to quickly remove the released lock from all BitLocks lists. The references actually point from the entry in the Outstanding list to the actual entry in the BitLocks list and not to the head of a list as the diagram might indicate. This allows the algorithm to look up an entry in the BitLocks list in unit time. 7.1 Correctness To establish that evaluating lock requests based on this summary (Equation 3) is correct, we show that the summary expression can be derived as a logical simplification of the compatibility criteria: a requested lock is compatible with outstanding lock state and can be granted if it is compatible (Definition 6.1) with all outstanding locks. For a system with locks L outstanding, the requested lock must be evaluated for compatibility against locks L _ _ _ This transformation shows that evaluation against the summary is equivalent to evaluation against all outstanding locks and is the original motivation for the summarizing data structure. We first developed the formalism for locking presented in Section 6. Upon seeing the expression for evaluating the compatibility of a single (requested) lock against a set of compatible (outstanding) locks, we determined that it can be simplified. The summarizing data structure for lock evaluation merely captures the logical simplification. 8 Algorithms for Lock Management Based on both dynamic lock evaluation and the semi-preemptible protocol, we develop a set of algorithms for managing a distributed set of locks. These algorithms are based on the summarizing data structure and logical rules for lock evaluation. They also capture our design goal of eliminating lock state using hierarchical locking. In Storage Tank's semi-preemptible protocol, we restrict each client's lock holdings to a single distributed session lock for each file. Clients may have many local open instances of that given file protected by a single lock. Clients must be able to modify their currently held lock - change the protected access and sharing without releasing the lock. This need arises in two instances: (1) when the client holds a semi-preemptible lock protecting local open instances and another client requests a lock for that file that does not conflict with the open instances, but is incompatible with the held semi-preemptible lock; and (2) when a client holding a semi-preemptible lock that protects open instances has a local process request another open instance, compatible with current open instances, that has access and sharing requirements the semi-preemptible lock cannot provide. In the first case, the held lock is too strong and the client must convert it to a weaker and compatible lock that still protects the open instances. This process is a lock downgrade. The client cannot release its lock outright and obtain a weaker lock because it has current open instances to protect. For downgrades, the lock demands received from a server must contain the type of file access lock requested so that clients can resolve compatibility and the appropriate downgrade. In the second case, the held lock cannot provide the access and sharing requirements of the new request. The held lock is too weak and the client attempts to obtain a stronger lock that can protect all of the current open instances and the new request. Obtaining the stronger lock while continuing to hold the old lock is called a lock upgrade. The concepts of strength and weakness help the system determine what action to take, i.e. how to change lock state to service local open requests or server demand messages. We note that upgrade does not lead to deadlock for file session locks. In databases, the term upgrade is sometimes used interchangeably with the term promotion [12], which describes a process reading a data object under one lock and promoting the lock before writing. This type of promotion leads to deadlock. For example, if two processes reading the same object both hold the weaker lock and require the stronger lock for writing, then neither will receive the promoted lock, because neither will release the weaker lock which protects the contents of the object that it has read. In contrast, for file session locks an upgrade changes the abstract distributed lock and does not modify the mode of an actual session; i.e., upgrades are not promotions. Conflicting upgrades of session locks do not result in deadlock, because the semi-preemptible protocol allows the server to deny conflicting lock requests. 8.1 Server Algorithm The server acts as a central management authority for granting and revoking locks. It takes incoming lock requests from clients, takes action by revoking outstanding locks to make the requested lock compatible, and grants the request when possible. To perform these actions, the server uses the summarizing locking data structure. The server receives lock requests from multiple clients concurrently and processes these requests seri- ally. Distributed locking requires that all lock state changes occur atomically. For this reason, the server considers one lock operation at a time. Multiple requests for the same lock are queued at the server and executed serially. Processing one lock request may result in multiple changes in lock state. That is, before granting a lock a server may demand multiple locks and reflect many state changes on the summary. How- ever, these are server-driven state changes rather than client requests and, therefore, are conducted in the context of a single client-initiated lock request. Upon receiving an incoming lock request, the server must evaluate the requested lock for compatibility and demand incompatible locks before granting the request. The server follows the steps in Figure 4 to EvaluateLockRequest Inputs: The requested lock DR i and the lock summary data structure. Evaluate the compatibility of the requested lock. 1. If ( 0 , the lock is compatible. Proceed to step 5 to grant the lock. 2. Else Demand all incompatible locks. (a) For all non-zero bits in demand the locks from the summary that disallow the access modes desired by the requested locks, i.e. for then demand all locks in BitLocks(i (b) Similarly, for all non-zero bits in DR demand the locks from the summary that permit access for the bit the requested lock disallows, i.e. for DR (i) then demand all locks in BitLocks(i). 3. Receive and process all responses to demand requests using function ReleaseLock or by downgrading the lock. 4. If any lock holder refuses the demand, deny the lock request and Return. 5. Grant the requested lock. Reflect the change in lock state on the summary data structure using GrantLock and Return. Figure 4: Server routine for evaluating incoming lock requests. determine the compatibility of the incoming lock request with respect to current system state. The server first evaluates the lock against the summary (step 1) to determine if the lock is compatible with all outstanding locks and can be granted immediately. If the request cannot be immediately fulfilled, one or more currently outstanding locks conflict with the request and must be demanded. The server determines and demands the set of conflicting locks in step 2. Outstanding locks can be incompatible because they disallow one of the access modes required by the requested lock (step 2(a)). Alternatively, outstanding locks can be incompatible because they permit an access mode the requested lock disallows (step 2(b)). The routine evaluates the conflicting access modes using the requested lock and the summary. For all high bits in DR , the routine demands all locks that set those bits. The remainder of the routine implements the semi-preemptible lock protocol through client/server interactions. If any client denies a demand, the requested lock is refused. Otherwise, if all demands succeed, the server grants the lock request (step 5). The routine to evaluate lock requests results in outstanding locks being released and a new lock being granted, which changes the global lock state. The server updates the locking data structure to reflect these changes using subroutines when granting a lock (Figure 5) and when clients release locks (Figure 6). When granting a lock, the server updates the locking data structures to reflect the new outstanding lock Figure 5). The new lock is added to the list of outstanding locks (step 1). The summary is modified to reflect the change in lock state (step 2). Finally, for every access mode permitted or disallowed, the new lock must be added to the BitLocks list and references added from the lock entry in Outstanding to the entries in BitLocks. GrantLock Inputs: The lock to be granted DG i and the lock summary data structure. 1. Register the lock to be granted in the list of outstanding locks. Create entry O in the list Outstanding. 2. Calculate the contribution of the lock to be granted to the summary. (a) PG . (b) DG . 3. Add a reference to G in the BitLocks data structures for each bit it permits and each it disal- lows. For each entry add a link from OG to the BitLocks list. (a) For (b) For to BitLocks(i). Figure 5: Server routine for updating the summarizing data structure when granting a lock. When clients release locks, the server updates its data structure using the routine in Figure 6. Releasing a lock consists of removing the lock from the list of Outstanding locks (step 3), updating the summary to reflect the change of state (step 2), and removing references to the lock (step 1). We do not present algorithms for processing upgrade requests and downgrade requests because these operations can be treated as combinations of release and request. The important difference between upgrade ReleaseLock Inputs: The lock to be released DR i and the lock summary data structure. 1. Look at the references in OR to find and remove all occurences of O R in the BitLocks lists. 2. If any of the BitLocks lists become empty, set the corresponding bits low in the summary, i.e. 3. Remove OR from the Outstanding list. Figure Server routine for processing released locks. Summary Client 1 Client 2 Client 3 2,3 Summary Held Summary Held Summary Server Client 1 Client 2 Client 3 Held Figure 7: Hierarchical lock management among client and servers. (or downgrade) and a combination of simpler operations is that the upgrade (or downgrade) must be performed atomically. Atomicity for multiple operations can be achieved at the server by allowing only a single process to access the locking data structure at a time. However, unlike the server algorithms, the client/server protocol must contain upgrade and downgrade primitives, as the protocol has no atomicity guarantee across multiple operations. 8.2 Client Algorithms The semi-preemptible locking protocol allows clients to hold locks even when they do not use them. By holding locks when not in use, caching locks, the client optimistically retains privileges on the premise that the last client to lock a resource is the most likely to lock it again. This design aims to minimize changes in lock state and the associated messaging and transactional overhead. The client uses two techniques to manage distributed locks and local processes that operate under these locks. First, the client allows local processes to use any subset of the access modes allowed by the locked i.e., the client is not required to use the full strength of the held lock. Also, the client allows multiple local processes to operate under the same distributed lock. To implement these techniques, the client manages its held lock and local processes hierarchically. The client holds a distributed lock (H) with a server (Figure 7). H encapsulates all actions that local processes can take against the locked resource. The local processes only access data using the modes specified by the lock and only prohibit the concurrent access modes encoded in H . Under this held lock, the client maps the access modes and sharing limitations of each local process to a local lock. The local lock is a local management abstraction and is not visible to other clients or the server. A client holds many local locks and uses the same summarizing data structure as the server (Figure 3) to evaluate and manage local locks. For the client, the summary describes all in-use access modes and disallowed modes for local processes accessing the data. The summary is restricted to be always weaker than the distributed held lock. This allows the held lock to protect concurrency guarantees expressed in the summary and, therefore, needed by the local locks. The management of the client's summarizing data structure is different than the server's. In many ways, management is simpler because the local locks held by local processes are non-preemptible. This eliminates demands or revocations of held locks. In other ways, management is more complex because when the client changes local lock state, it must consider the distributed system state, encapsulated by the held distributed lock H . The main principle in the management of multiple local locks under a single distributed lock is Theorem 6.2, which states that if a lock is compatible with all other outstanding locks in the distributed system then all locks weaker than that lock are also compatible. This ensures that if the remainder of the distributed system acts in accord with the held lock, the permitted access and sharing of the local locks are respected. Locking clients respond to several stimuli, requests to obtain and release local locks from local pro- cesses, and demand requests from the lock server. Clients obtain and upgrade distributed locks on behalf of local processes that require stronger local locks and downgrade and release locks in response to server demands. In effect, the client acts as a server of locks to local processes, under the constraint that the client holds a suitable distributed lock. For this reason, lock acquisition, release, upgrade, and grant operations at the client use routines much like the server's. Owing to similarities, we omit detailed descriptions of these operations. The remaining client action is to process demand requests from the server (Figure 8). Demands are initiated by a server attempting to alter the distributed lock state in order to grant a lock to another client. The lock mode requested by the other client is included in the server's demand message so that the requested lock can be evaluated against current open state. In step 1, the client evaluates whether the requested lock can be granted given local lock state. Only the clients know local state and they must evaluate the request and determine the conflicting outstanding locks. The clients themselves make the final determination as to the compatibility of the request against current system state by comparing the requested lock against their current local lock state. Clients with compatible local locks downgrade their distributed lock. When choosing the locking mode for the downgrade, there are many possible choices. We present the two extreme choices as possible heuris- tics. First, a client can elect to retain as strong a lock as possible, the minimum downgrade heuristic. The ProcessDemand Inputs: The remotely requested lock included in the demand message DR i and the client lock summary data structure. 1. Evaluate whether the requested lock is compatible with the currently held local locks. If 0 proceed to step 2, Else deny the demand and Return. 2. Determine the downgraded lock mode required to protect the currently held locks and the requested lock using one of the two following heuristics. (a) Maximum downgrade heuristic, (b) Minimum downgrade heuristic, 3. Inform the server of the downgrade and Return. Figure 8: Client routine to determine the needed lock downgrade. client reduces its holding as little as possible, in the lock strength sense, to make its lock compatible with the requested lock. This choice would perform well if future access to the object were more likely to occur locally than from other clients. However, if the interest in the object has moved to other clients, the policy may result in future demands. The other alternative is to downgrade the lock holding as much as possible: the maximum downgrade heuristic. The downgraded locking mode is exactly the lock summary-the fewest permitted access modes and disallowed modes needed by local locks. 9 Comparing Dynamic Evaluation with Traditional Locking We present the execution time and memory usage bounds of lock evaluation algorithms to illustrate how dynamic evaluation asymptotically outperforms evaluation based on static data structures if locking systems are complex. The amount of space used by a locking system can be divided into static and dynamically used space. The static space contains the information the system needs to know about locking modes to properly run the locking protocol. The dynamic space contains information about current outstanding locks. Conventional wisdom holds that the dynamic costs dominate the static costs. However, as locking systems become more complex, this assertion is not always true because the static space costs can grow exponentially in the number of access modes. For static data structures, k different access modes results in 2 2k unique locks. Locking with traditional data structures requires space in 4k ). There are 2 4k entries in a lock compatibility table for 2 2k locks. For dynamic lock evaluation, the algorithms only need to know the different locking modes, which uses space in O(k). The dynamic space usage of locking using traditional data structures is (nk) for n outstanding locks. The system stores a list of the k modes of the n outstanding locks. Dynamic evaluation uses the summarizing data structure to hold dynamic lock state, which is an (nk) data structure. The Outstanding list has n entries each of size k, and there are k different BitLocks lists with a total of n entries in all lists and n incoming pointers referring to these entries. For execution time, we look at how long operations take with respect to the amount of outstanding lock state n and the amount of static state k. Lock evaluation with static data structures takes time in n), because the request must be checked against all outstanding locks. Dynamic lock evaluation uses execution time in O(k) to evaluate the lock, update the lock summary, and fill out the BitLocks and Outstanding lists. Similarly, releasing a lock takes time in O(k). When concurrency is high, there are many outstanding locks, dominates k, and dynamic evaluation operates more efficiently. The total space required for dynamic evaluation of locks is always asymptotically less than or equal to that required for evaluation based on static data structures. However, the dynamic space requirements of dynamic evaluation of locks are larger by a constant factor. The list of locks used for static evaluation is equivalent in size and content to the Outstanding list. All other portions of the summarizing data structure are overhead. The true space savings are realized on the static data structures and become important when k grows large. Perhaps more important than the space and time savings is the improvement in manageability and ease of implementation. An implementer using a lock compatibility table must reason about each pair of locks and complete the entry in the table. This task can be difficult even for as few as 64 locks in Windows NT. Additionally, all locking modes must be pre-programmed. With dynamic evaluation, the implementer need only identify how the object will be accessed. The dynamic evaluation algorithms manage concurrency and compatibility. Related Research Many distributed file systems clients transact with a server on every open and close of a file for synchronization [25, 5, 26]. This is the simplest technique to implement local file system open semantics in the distributed environment. However, all file open and file close requests require a network operation. The Andrew file system [15] interacts with a server at every open and close and uses open and close as points to synchronize cached data. Andrew does implement a data cache that can hold data past close and a general callback (demand) mechanism. However, callbacks apply only to preemptible data locks. Like Storage Tank, the Calypso file system [7] uses open mode synchronization locks, called tokens, to implement local file system semantics. In Calypso, tokens are fully preemptible at the client and cannot be held past close. Every file system open and close generates a token request or release. File system open state and token request conflicts are managed completely at the server. Calypso uses a simple lock hierarchy for its data locks, but the hierarchy does not apply to open mode synchronization tokens. The DFS file system [16] describes a token mechanism similar to semi-preemptible locks for the management of data, metadata, and open state. Like semi-preemptible locks, a client can refuse or permit revocation of a token, depending upon local state. Token management in DFS differs in that all elements of system locking, including file access locks, data locks, and byte-range locks are managed with the single token mechanism. The DFS treatment of token management is less concrete than our discussion of locking, and does not address mapping local file system semantics to a distributed locking system. This work omits a discussion of synchronization in the presence of failure. A distributed file system that presents a local file system interface to remote and shared storage must continue to do so when components fail. As do many modern file systems [24, 21, 18], Storage Tank uses a lease-based [6] protocol to ensure operational safety and high availability in the presence of client and server failures, and network partitions. Conclusions Distributed file systems need to manage open state for referential integrity and synchronized access to files. Existing distributed systems address this problem by either relaxing local file system semantics, or by sending every file open request to a server. We have introduced a locking construct, the semi-preemptible lock, that permits file system clients to grant most file open requests locally, without a server transaction. By avoiding server messages on open, our client improves performance by exploiting locality of access to files. At the file system client, semi-preemptible locks are used to summarize open state, so that many open files may be granted under the protection of a single semi-preemptible lock. This reduces global lock state and further reduces client-server messages. We have also presented a set of algorithms and data structures for the dynamic evaluation of locks. Dynamic evaluation provides a formal and compact framework for evaluating lock compatibility at runtime. Dynamic evaluation obviates lock compatibility tables, exponentially reducing the space requirements to describe a locking system. It also obviates ad hoc systems for lock evaluation, providing a formalism that manages locks correctly and without deadlock. --R Serverless network file systems. Measurements of a distributed file system. Fast Crash Recovery in Distributed File Systems. Fibre Channel: Gigabit Communication and I/O for Computer Networks. A universal file server. An analytical study of opportunistic lease renewal. Recovery in the Calypso file system. Extendible hashing: A fast access method for dynamic files. Gigabit ethernet: From 100 to 1 File server scaling with network-attached secure disks Filesystems for network-attach secure disks Transaction Processing: Concepts and Techniques. The design of a multicast-based distributed file system Scale and performance in a distributed file system. Synchronization and caching issues in the Andrew file system. DEcorum file system architectural overview. Disconnected operation in the Coda file system. A coherent distributed file cache with directory write- behind Caching in the sprite network file system. An authentication service for computer networks. Authenticating network attached storage. A highly available file system for a distributed workstation environment. A scalable distributed file system. The LOCUS distributed operating system. Overview of the Sun network file system. --TR
session locking;distributed file systems;concurrency control;lock evaluation
592979
Web Proxy Acceleration.
Numerous studies show that miss ratios at forward proxies are typically at least 4050%. This paper proposes and evaluates a new approach for improving the throughput of Web proxy systems by reducing the overhead of handling cache misses. Namely, we propose to front-end a Web proxy with a high performance node that filters the requests, processing the misses and forwarding the hits and the new cacheable content to the proxy. Requests are filtered based on hints of the proxy cache content. This system, called Proxy Accelerator, achieves significantly better communications performance than a traditional proxy system. For instance, an accelerator can be built as an embedded system optimized for communication and HTTP processing, or as a kernel-mode HTTP server. Scalability with the Web proxy cluster size is achieved by using several accelerators. We use analytical models, trace-based simulations, and a real implementation to study the benefits and the implementation tradeoffs of this new approach. Our results show that a single proxy accelerator node in front of a 4-node Web proxy can improve the cost-performance ratio by about 40%. Hint-based request filter implementation choices that do not affect the overall hit ratio are available. An implementation of the hint management module integrated in Web proxy software is presented. Experimental evaluation of the implementation demonstrates that the associated overheads are very small.
Introduction Proxy caches for large Internet Service Providers can receive high request volumes. Numerous studies show that miss ratios are often at least 40%-50%, even for large proxy caches [7,9,12]. Consequently, a reduction of the cache miss overheads can result in significant throughput improvements for proxy caches experiencing high request volumes. Based on this insight and on our experience with Web server acceleration [15], this paper proposes and evaluates a new approach to improving the through-put of Web proxy systems. Namely, we propose to front-end a Web proxy system with a high performance node that filters the requests, processing the misses and forwarding the hits and new cacheable content to the proxy. The Web proxy's miss-related overheads are reduced to receiving the cacheable objects pushed by the accelerator over permanent connections. Consequently, the proxy can use the released resources to service an increased number of hits. Figure 1 illustrates the interactions of the proxy accelerator with Web proxy nodes, clients, and Web servers. Hints Cache WebServer Client ~30% WebProxy GBytes ~35% ~35% Non-cacheable Content Hits Proxy Accelerator Figure 1. Accelerated Web Proxy Architecture. The accelerator is an extended HTTP server running under an embedded operating system optimized for communication, like the Web server accelerator described [15], or under a general-purpose operating system in kernel-mode at service-interrupt level, like the Netfinity Web Server Accelerator[16]. The accelerator has a main memory cache, mainly used to store content to be forwarded to the Web proxy. The accelerator filters the requests based on a summary of the proxy cache content, called hints. The hints are maintained based on information provided by proxy nodes and information derived from the accelerator's own decisions. The hint representation and the update mechanism are important factors for the performance gains enabled by the proposed proxy acceleration method. This is due to the related requirements for system and network resources and to the intrinsic tradeoff between resource usage and information accuracy. Previous research has considered the use of location hints in the context of cooperative Web proxies [8,19,22,10]. Our system applies this approach to the management of a cluster-based Web proxy. In addition, we experiment with new hint maintenance mechanisms, specifically designed for the interaction between a front end and a cluster of homogeneous nodes. Our work extends the set of solutions to proxy cache redirection [1,13,14,17]. Most relevant is the comparison with the proposal in [14], which has the front end identifying and processing the requests for non-cacheable content. Extending this functionality, a proxy accelerator processes all of the requests for content not cached at the Web proxy, including the cacheable content not currently in the proxy cache. Building on top of previous research on cluster-based network services [4,2], our proxy acceleration method overcomes the intrinsic limitations of content-based routing components by integrating several accelerator nodes in front of a Web proxy cluster. In this paper, we use analytical models in conjunction with simulations to study the potential for performance improvement of our proxy acceleration scheme. We evaluate the impact of several hint representation and maintenance methods on throughput, hit ratio, and resource usage. In addition, we present and evaluate the performance impact of an actual implementation of the hint- management component in Web proxy application. Our analytical study reveals that the throughput improvement can be signif- icant. For instance, when the accelerator can achieve about an order of magnitude better throughput than a Web proxy node for communications operations, such as the embedded system considered in [15], a single accelerator node can boost the throughput of a four-node Web proxy by about 100%. Our simulations validate that, with appropriate hint representations and update protocols, the proposed acceleration method can offload more than 50% of the WP load with no impact on the observed hit ratio. In addition, the evaluation of our implementation shows that the overhead incurred with hint management has no significant impact on the client-perceived response times and throughput. Paper Overview . The rest of the paper is structured as follows. Section 2 describes the architecture of an accelerated Web proxy. Section 3 analyzes the expected throughput improvements with an analytical model. Section 4 evaluates the impact of the hint management on several Web proxy performance metrics with a trace-driven simulation. Section 5 presents the implementation of the hint-management module in a Web proxy application and evaluates its performance impact. Section 6 discusses related work, and Section 7 summarizes our results. 2. Accelerated Web Proxy Typical high-performance Web proxy systems consist of a cluster of nodes running a Web proxy application and a front-end node that distributes the requests to the application nodes. Front end-based solutions to improving the performance of Web proxy systems are restricted mainly to balancing the load based on criteria like number of active connections and CPU utilization. Because of the inherently unpredictable content mixtures in Web proxy caches, typical content-based routing policies that distinguish origin servers and/or object types are less beneficial. However, a content-based scheme distinguishing between cacheable and non-cacheable content has been proved to bring significant performance improvements to Web proxy systems when combined with off-loading the processing of requests for non-cacheable content from proxy nodes to the front end [14]. In this paper, we propose to further offload proxy node overheads by extending the functionality of the front-end. Namely, the extended front-end, called proxy accelerator (PA), is a content-based router that can distinguish between cache hits and misses based on information about proxy cache content. In the case of a cache hit, the PA forwards the request to a Web proxy (WP) node that contains the object either by TCP splicing or handoff [17,21]. In the case of a cache miss, the PA sends the request to the origin Web site, bypassing the WP. When it receives the object from the Web site, the PA replies to the client, and if appropriate, pushes the object to a WP node which is selected based on some load balancing criteria. WP WP WP WP PA PA Web Proxy Cluster Requests Hints Router Servers Clients Figure 2. Interactions in a Web Proxy (WP) cluster with several Proxy Accelerators (PA). The enabling element for the new front-end functionality is the PA's information about the content of the WP cache, henceforth called hints. Possible solutions for hint representation and update mechanism will be discussed later in this section and evaluated by trace-based simulation in Section 4. The PA includes a main memory cache for storing content waiting to be pushed to the WP. In addition, the PA cache can store duplicates of objects in the WP cache. Requests satisfied from a PA cache never reach the WP, further reducing the load of the WP. Figure 1 illustrates the interactions of a PA in a four-node WP cluster. PA and WP nodes are connected by permanent TCP connections for request handoff [17], hint updates, and content push. Notice that WP nodes still interact with Web sites. This occurs when the PA filter is not accurate, or cache objects have to be revalidated. For scalability, the system may include several PAs (Figure 2) and a router that balances the load among them. Each PA has hints about each of the WP nodes. Therefore, a PA can appropriately forward any of the incoming hits. The policy for new content distribution may be as simple as round-robin. A PA can be built from stock hardware and runs under an embedded operating system optimized for communication. Alternatively, the PA can be built as an extended kernel-mode HTTP server. Either way, a PA can achieve an order of magnitude better throughput than a Web proxy node for communications operations. A WP node would typically run software extended for hint maintenance and distribution, for handoff of TCP/IP connections and URL requests, and for pushing objects to PA caches. Hint Representation. Previous research has considered several hint representations and consistency protocols. Hint representations vary from schemes that provide accurate information [22,10] to schemes that trade accuracy for reduced costs (e.g. memory and communication requirements, look-up and update over- In this study, we consider two representations - one that provides accurate information and one that provides approximate (inaccurate) information. The accurate-information scheme, called the directory scheme, uses a list of object identifiers including an entry for each object known to exist in the WP cache. In our implementation, the list is represented as a hash table [10,22]. Hint Space Hint Entry Figure 3. Hint representation based on Bloom Filters. The approximate-information scheme is based on Bloom Filters [5]. The filter is represented by a bitmap, henceforth called hint space (see Figure 3), and several hash functions. To register an object in the hint space, the hash functions are applied to the object identifier, and the corresponding bits, which are generically called hint entries, are set in the hint space. A hint look-up is positive if all of the entries associated with the object of interest are set. A hint entry is cleared when no object is left in the cache whose representation includes the entry. Therefore, the cache owner (i.e., the WP node) maintains a "collision" counter for each entry, which is updated whenever a related object is added or removed from the cache. When a collision counter reaches its maximum value, it is no longer updated. The entire hint space is recomputed when the number of overflowed counters reaches a threshold. The PA maintains a Bloom Filter for each of the WP nodes. Hint Consistency. The hint consistency protocol ensures that the information at the PA site reflects the content of the WP cache. Protocol choices vary in several dimensions: the entity that identifies the updates, the protocol initiator, the rate of update messages, and the exchanged information. In the previously proposed solutions, updates are identified either by the WP (henceforth called WP-only updates) [8,19] or by the content-based router (PA) [17]. Both methods have a significant likelihood of wrong request redirection decisions. The first method may result in false misses because of the inherent delay of updates. The second method may result in false hits because information about objects removed from the cache is not propagated to the router. False misses are also possible because entries in the router directory are discarded periodically due to space limitations. In this paper, we propose and evaluate a method, called eager hint regis- tration, which attempts to address the drawbacks of previous methods. Namely, both PA and WP identify and record hints. When the PA pushes an object to a WP node, it "eagerly" registers that the object exists in the corresponding cache. As a side effect, the method enables a reduction of the update traffic sent by a WP node. Namely, a WP node does not send update notifications for objects received from the PA (a small exception, due to the asynchronous communication between PA and WP, is discussed in Section 5). With respect to the protocol initiator, previously proposed solutions address only WP-only updates. Updates are pulled by the hint owner (i.e., PA) [19] or pushed by the cache owner (i.e., WP) [8,22,10]. With respect to the update rate, previous proposals consider update messages sent at fixed time intervals [19] or after a fixed number of cache updates [8]. The exchanged information may be incremental, addressing only the updates occurred since the previous message [10,8,19], or a copy of the entire hint space [8]. Proxy Accelerator Cache. The PA cache may include objects received from Web sites or pushed by the associated WP nodes. In selecting the PA cache replacement policy, we have to consider that all objects in the PA cache also exist in the WP cache. Consequently, the PA cache replacement policy should not be focused on minimizing the network-related costs, as shown appropriate for WP caches. Policies that maximize the hit ratio based on hotness statistics and object sizes are expected to enable better PA performance than policies like Greedy-Dual-Size [6]. In the following sections, we evaluate the throughput potential of the proposed proxy acceleration method and the impact of several hint implementation choices. 3. Expected Throughput Improvements The major benefit of the proposed hint-based acceleration of Web Proxies is throughput improvement. We analyze the throughput properties of this scheme with an analytic model by varying the number of WP and PA nodes, the PA functionality and cache size, and the WP's and PA's CPU speeds. The system is modeled as a network of M/G/1 queues. The servers represent PAs, WP CPUs and disks, and a Web site. The serviced events represent (1) stages in the processing of a client request (e.g. PA Miss, WP Hit, Disk Access) and (2) hint and cache management operations (e.g. hint update, object push). Event arrival rates derive from the rate of client requests and the likelihood of a request to switch from one processing stage to another. The likelihood of various events, such as cache hits and disk operation, are derived from recent studies on cache performance and Web content models [23,14,6,18]. The overhead associated with each event includes a basic compute overhead, I/O overheads (e.g. connect, receive message, handoff), and operating system overheads (e.g. context switch). The basic overheads are identical for the PA and WP, but the I/O and operating system overheads are different. Namely, the PA overheads correspond to the embedded system used for Web server acceleration in [15] - a 200 MHz Power PC PA which incurs a 514-sec miss overhead (i.e., 1945 misses/sec) and 209-sec hit overhead (4783 hits/sec). The WP overheads are derived from [17] - a 300 MHz Pentium II which incurs a 2518-sec miss overhead (i.e., 397 misses/sec) and 1392-sec hit overhead (718 hits/sec). These overheads assume 8 KByte objects. The following parameters are fixed for this study: (1) 40% non-cacheable content ratio and 25% miss ratio, (2) TCP handoff is used by the PA for client connections upon positive hint look-up; (3) directory-based hints with 1 hour incremental updates; (4) disk I/O is required for 75% of the WP hits; (5) one disk I/O per object; (6) four network I/Os per object transmission; (7) 25% of WP hits are pushed to the PA, when the PA caches hot objects; (8) 30-sec WP context switch overhead. Unless otherwise specified, the PA nodes have 450 MHz CPUs, and the WP nodes have 300 MHz CPUs. The costs of the hint-related operations are derived from our implementation and the simulation-based study (see Section 4). Hint-based acceleration improves throughput and cost-performance ratio Figure 4 presents the expected performance for a traditional WP-cluster and Throughput WP Nodes traditional Figure 4. Throughput with number of WP and PA nodes. (no service from PA cache)1030507090Cost/Performance Improvement WP Nodes WP300MHz PA300MHz WP300MHz PA450MHz WP300MHz PA700MHz Figure 5. Cost-performance improvement with number and speed of WP and PA nodes. (no service from PA cache) for clusters enhanced with up to four PAs when the PAs do not service objects from their cache. The plot illustrates that the hint-based acceleration enables significant WP throughput improvements. For instance, a single PA node can increase the performance of a 2-node WP about three times and that of a 4 node WP about two times. With two PAs, a 4-node WP can achieve higher throughput than a 16-node WP. However, consistent with previous studies on content-based routing [4,2,21], the plot illustrates that a PA becomes a bottleneck for large WP clusters. The cost-performance improvement enabled by the proposed scheme is rele- vant. For instance, Figure 5 illustrates that if we assume that the cost of a PA is Throughput WP Nodes traditional 1 PA only non-cacheable 2 PAs only non-cacheable Figure 6. Throughput of hint-based acceleration and non-cacheable redirection. (no service from PA cache)200040001 Throughput WP Nodes traditional PA hit 0% PA hit 10% Figure 7. Throughput when PA services objects from its cache. almost the same as for a WP node, the cost-performance improvement achieved with a single PA is -60% in a 2-node WP and -40% in a 4-node WP. The performance benefits of the scheme increase with the difference between WP and PA's CPU speeds, as illustrated by Figure 5 for the case of 700 MHz PA CPU's. Figure 5 also illustrates that a 2-node accelerator benefits relatively large WP configurations. For instance, the cost-performance improvement with a 2 PAs is lower than with 1 PA for a 2-node WP. Hint-based acceleration can benefit more than the redirection of non-cacheable objects. Figure 6 illustrates that, for appropriately sized WP clus- ters, the proposed hint-based acceleration can enable better performance than the previously proposed policy of redirecting only the non-cacheable objects [14]. For a workload with only 25% miss ratio, a 2-node WP cluster can attain more than double throughput levels if the front-end can identify and process the request for cache misses along with the requests for non-cacheable content. PA-level caches benefit low-performance WPs. Figure 7 illustrates that the effect of a PA cache is marginal in particular when the ratio of non-cacheable content is significant, like in our selection of experiment parameters. To summarize, the hint-based Web Proxy acceleration may lead to significant performance and cost-performance improvements when the PA nodes can achieve a much higher throughput network I/O than the WP nodes. The new scheme improves upon the previously proposed redirection of non-cacheable content for appropriately sized WP clusters. 4. Impact of Hint Management In this section, we evaluate how the hint representation and the consistency protocol can affect the performance of an accelerated Web proxy. We focus on performance metrics, such as hit ratio and throughput, and on cost metrics, such as memory, computation, and communication requirements. These performance and cost metrics depend on various characteristics of the hint management mechanism (see Table 1). For instance, the look-up overhead affects the long-term PA throughput. Also, the false miss decisions of the PA filter cause undue network loads and response delays. Table 2 summarizes some of the differences between the representation methods and the update protocols introduced in Section 2. For instance, the look-up overhead is constant for the Bloom Filter-based scheme while for the directory scheme it depends on the con- text, namely on how well the hash table is balanced. Similarly, eager registration results in no false miss, while for WP-only registration the amount of false misses increases with the update period. In this study, we identify and compare relevant trends of these representation methods and update protocols. The study is conducted with a trace-driven simulator and a segment of the Home IP Web traces collected at UC Berkeley in 1996 [11]. The traces include about 5.57 million client requests for cacheable objects over a period of 12.67 days. The number of objects is about 1.68 million with a total size of about 17 GBytes. Factors. The factors considered in this study are the following: (1) PA and Table Impact of Hint Management Choices. Characteristic Affects memory at WP WP memory hit ratio memory at PA PA cache hit ratio update period long-term throughput update period WP/PA short-term throughput look-up time PA long-term throughput false hits WP load false misses network usage Table Selected Method Characteristics. Characteristic Representation Directory Bloom Filters memory at WP none O(hint space) memory at PA O(objs WP) O(hint space) look-up overhead context-dependent constant false hits none O(objs WP) Protocol WP-only Eager Reg. false misses period-dependent none PA computation on-off, large bursts smaller bursts WP memory, (2) hint representation, (3) hint update protocol, and (4) update period. Namely, we consider configurations with 0.25-1 GByte PA memory and memory. For the Bloom Filter-based scheme, we vary the size of the hint space (0.5-10 MBytes) and the number of hash functions; each hash function returns the remainder of the hint space size divided by an integer obtained as a combination of four bytes in the object identifier. Due to limitations associated with trace encoding, object identifiers are composed of the 16-byte MD5 representation of the URL and a 4-byte encoding of the server name (this restriction makes impossible an accurate evaluation of the resource requirements associated with the Bloom Filter-based scheme). Finally, we experiment with update periods in the range of 1 sec to 2 hours. Fixed Parameters. We experiment only with "push" updates; in comparison to "pull" updates, "push" updates allow better control of hint accuracy [8] and lower I/O overheads. In addition, we experiment only with incremental updates, as we consider it more appropriate than "entire copy" updates in the context of the hint representations and the system parameters considered in this study. The WP cache replacement policy is LRU. The PA cache replacement is "LRU with bounded replacement burst". According to this policy, an incoming object is not cached if this would cause more than a given number of objects to be removed from the cache. This policy is expected to perform well based on studies that show that small objects have a large contribution to the hit ratio [18,3]. For both caches, garbage collection is invoked when a new object has to be accommodated. 4.1. Cost Metrics Memory Requirements. The main memory requirements of the hint mechanism derive from the hint meta-data and the message buffers used for the hint consistency protocol. Hint Meta-data. For the directory scheme, no meta-data is maintained at the WP, and at the PA, the requirements increase linearly with the population of the WP cache. For instance, in our implementation, with 68-byte entries, the meta-data is -70 MBytes for a 10 GByte WP cache. By contrast, for the Bloom Filter-based scheme, meta-data is maintained at both PA and WP, and is independent of the cache size. The PA meta-data is as large as the hint space multiplied by the number of WP nodes, and the WP meta-data is as large as the hint space multiplied by the size of the collision counter (e.g. 8 bits). Hint Consistency Protocol. Message buffer storage space required at the WP site for the hint consistency protocol increases with the rate of WP cache updates and the period of the consistency protocol. For the directory scheme, an update entry has unpredictable length, as it includes the full name of the object. For the Bloom Filter-based scheme, an update corresponds to a hint entry update, represented by a 5-byte data structure; a cache update can result in as many hint entry updates as hash functions in the filter. Computation Requirements. The computation requirements associated with our acceleration scheme result mainly from request look-up and the processing of update messages. In addition, minor computation overheads may be incurred at WP and PA upon each cache update. Look-up Overhead. For the directory scheme, the look-up overhead depends on the distribution of objects to hash table buckets. For the Bloom Filter-based Maximum Entries Per Update Message WP Memory (MBytes) BloomFilter, WP-only BloomFilter, eager WP-only Figure 8. Variation of maximum number of entries per update message. 1 hour period, 256 MByte PA, 4 MByte hint space, 4 entries per object.0.10.30.50.7 1000 2000 3000 4000 5000 6000 7000 8000 9000 1000011000 Total Entries Per Update Payload (MEntries) WP Memory (MBytes) BloomFilter, WP-only BloomFilter, eager WP-only Figure 9. Variation of total number of update entries. 1 hour update period, 256 MByte PA, 4 MByte hint space, 4 entries per object. scheme, the look-up overhead is more predictable, depending mainly on the filter parameter. More specifically, the overhead includes the transformation of the object name into one or more integer values, the computation of the hash functions, the selection, and the test of the corresponding bitmap entries. On a 332 MHz PowerPC, the transformation of the object name into an MD5 representation takes -9-sec per 55 byte string segment. For the four simple hash functions used in our Bloom Filter implementation, the overhead of hash function computation and bitmap look-up is -1-sec. Overall, the look-up overhead is lower than 10% of the expected cache hit overhead on a 300MHz PA (see Section 3). Hit Memory (MBytes) Directory BloomFilter Figure 10. Variation of overall hit ratio with WP configuration. 512 MByte PA, Bloom Filter Hint Consistency Protocol. The amount of computation triggered by the receipt of a hint consistency message depends on the number of update entries included in the message. Because of hint entry collisions, the Bloom Filter-based scheme results in about half the load of the directory scheme (with 160-byte object iden- tifiers). The variation of the maximum and average message sizes presented in Figure 8 and Figure 9 illustrates this trend. These figures also illustrate the benefit of the eager registration approach. For both hint schemes, on average, the eager hint registration reduces by half the computation loads. Similar trends are observed for the communication requirements of the consistency protocol. Note that the very large maximums observed in Figure 8 are due to a few very large WP cache replacement bursts (-2300 objects). 4.2. Performance Metrics Hit Ratio. The characteristic of the hint mechanism that most affects the (observed) hit ratio is the likelihood of false misses. Typically, false misses are due to hint representation and update delays. In our experiments, because of the selected hint representations, false misses are only the result of update delays. As expected, the likelihood of false misses increases with the update period. Figure 11 illustrates this trend indirectly: the ratio of requests directed to the WP with a traditional update mechanism decreases with the increase of the update period. Eager hint registration reduces the number of observed false misses, with less impact on the Bloom-Filter scheme because of collisions. 262728290 1000 2000 3000 4000 5000 6000 7000 8000 of Requests to WP Hint Update Period (secs) Eager, BloomFilter WP-only, BloomFilter Eager, Directory WP-only, Directory Figure 11. Variation of WP requests with update protocol and period. 256 MByte PA, 4 GByte of False Hits Hint Space (MBytes) 6 GByte WP 8 GByte WP Figure 12. Variation of false hit ratio with hint space and WP cache. 512 MByte PA. The amount of hint meta-data at the WP can have a relevant impact on the overall hit ratio (i.e., hits in both PA and WP caches) when the ratio of the meta-data in the WP memory is large, as for small WP cache sizes (see Figure 10). PA meta-data has no significant impact because objects removed from the PA cache are pushed to the WP rather than being discarded. Throughput. The extent of throughput improvement depends on how well the PA identifies the WP cache misses. This characteristic, reflected by the likelihood of false hits, is affected by the representation method and the consistency protocol. Hint Representation. While the directory scheme has no false hits, for the Bloom of False Hits WP Memory (MBytes) Figure 13. Variation of false hit ratio with hint entries per-object. 512 MByte PA. Filter-based scheme, the likelihood of false hits increases with the hint-space collision. As illustrated in Figure 12, this occurs when the hint space decreases or the number of objects in the WP cache increases. We notice that the false-hit ratio is constant when the hint space increases beyond a threshold. This threshold depends on the WP cache size, and the effect is due to the characteristics of the selected hash functions. The likelihood of false hits also depends on the number of hint entries per object. Figure 13 illustrates the variation observed in our experiments with filters with 3-5 hash functions. While, typically, the ratio of false hits decreases with the number of hint entries per object, when hint space occupancy is high, inversions may occur due to the collisions. Hint Consistency Protocol. The impact of update period on false hit ratio is more significant in configurations with frequent WP cache updates (e.g., systems with small WP caches). Figure 14 illustrates this trend for both hint schemes. For instance, for Bloom Filters, the difference between the false-hit ratios observed for 1-hour and for 1-sec updates is 0.07% for a 4 GByte cache and 0.04% for a 6 GByte cache. Comparing the two schemes, the relative impact of the update period is almost identical. Request Response Time. Besides the overhead of hint look-up and update, the accelerated WP architecture can affect the average request response time by the reduction of the hit ratio in the WP and PA main-memory caches. Besides the amount of meta-data in the available main memory, the hit ratios depend on the cache replacement policies. Considering the PA cache, Figure 15 illustrates that 12340 1000 2000 3000 4000 5000 6000 7000 8000 of False Hits Hint Update Period (secs) BloomFilter, 4 GB WP BloomFilter, 6 GB WP Figure 14. Variation of false hits with update period. 256 MByte PA, 4 MByte hint space.40506070 Hit Replacement Bound Total Hits, 256 PA PA Hits, 256 PA Total Hits, 512 PA PA Hits, 512 PA Figure 15. Variation of hit rate with replacement bound. '-1' for unbounded replacements. Bloom Filter, 6 GByte WP, 4 MByte hint space. the bounded replacement policy enables better PA hit ratios without reducing the overall hit ratio. For instance, limiting the replacement burst to 10 objects enables more than 5% larger PA hit ratios than with unlimited burst (see bound -1). This is because the average replacement burst is significantly smaller than the maximum (e.g. 3 vs. 2300 in our experiments). Summary . By trading off accuracy, the Bloom Filter-based scheme exhibits lower computation and communication overheads. However, the meta-data at the WP node may reduce the memory-hit ratio. While, in general, the hint look-up overheads are small with respect to the typical request processing overheads, the overheads are more predictable for the Bloom Filter-based scheme. Eager hint registration reduces update-related overheads and prevents hit ratio reductions caused by update delays. False hit ratio (and, consequently, throughput) is more significantly affected by hint representation than by update period. For Bloom Filter-based schemes, this ratio increases exponentially with hint space contention. 5. Implementation of Hint Management in Web Proxy Nodes The implementation of the hint-based accelerated Web proxy infrastructure is built by extending a thread-based Web proxy application and a kernel-mode HTTP server with functionality for hint management and content push. The implementation of the hint mechanism is based on Bloom Filters with eager registration and periodic updates. The implementation makes possible a flexible system configuration. Namely, the Web proxy node can interact with multiple accelerators and the accelerator node can interact with multiple proxy nodes. Hint Representation. The size of the hint space is customizable at initialization time. The collision counters take up one byte. The Bloom Filter has six hash four applied to the object name and two to the origin server IP address. The object name, s 0 s 1 :::, is transformed into an integer value by evaluating the polynomial P fixed prime number. The overhead of name transformation is -0.95-sec for the first 10 bytes, and -0.7-sec for each additional 10 bytes, measured on a 332 MHz PowerPC. Note that, although less effective than MD5 in avoiding collisions, this solution is characterized by lower computation overheads, which clearly benefit the proxy accelerator performance. The hash functions are equal to mod(HintSpace) applied to the resulting 4-byte integer and to three permutations of its bytes. Similarly, the mod(HintSpace) function is applied to the 4-byte integer representing the IP address and to one permutation of its bytes. The overhead of computing these hash functions is -5.5-sec. When the WP has several nodes, the corresponding Bloom Filters are represented at the PA by interleaved bitmaps. In this representation, the corresponding bits all of the filters are placed in consecutive memory locations (see Figure 16). In comparison to the original representation in which each Bloom Filter is represented in a separate memory segment, the interleaved representa- /Hint Entry Hash1(obj, Hint Space Figure 16. Hint representation as interleaved Bloom filters for 4 WP nodes. tion is characterized by predictable look-up overheads, independent of the cluster size. With the interleaved representation, the look-up overhead for a 4-bit entry is constant at -1.28-sec, on a 332 MHz PowerPC. For the original representa- tion, the overhead varies from -1.19-sec for a single node look-up to -4.39-sec for four node look-up. Hint Update Protocol. The hint management implementation uses the eager hint registration protocol introduced in Section 2. Namely, an accelerator registers hints as it pushes objects towards a proxy node. A proxy node collects the hint updates and periodically sends them to each accelerator such that an accelerator will not receive updates resulting from cache updates for content it has pushed to the WP node. An exception is made when a hint entry is set by a push operation when the previous clear has not reached the accelerator before it pushed the object (and set the entry in its representation). If not appropriately handled, this situation may lead to false misses. To identify this type of situation, proxy updates are labeled, and the pushed content sent by an accelerator carry the label of the most recent update enacted by the proxy. Web Proxy Software Extension. The extension of the Web proxy software has three components. The first component represents the update interface invoked by the proxy threads handling client requests when objects are added or permanently removed from the cache (cache operations that remove objects only for accommodating their newer versions should not result in hint updates). The calling thread indicates the object name, origin server, and the corresponding cache operation. The resulting hint updates, if any, are recorded in update buffers. When an update buffer gets full, it is released for transmission. Update buffers are organized in an active set and a waiting set. Updates are recorded in buffers of the active set. When such a buffer gets full, it is replaced with one in the waiting set. The active set includes a buffer for each active accelerator and the local proxy node. Updates are registered in the active buffer associated with the node that originated the operation. More specifically, updates resulting from a cache- add operation are registered in the buffer of the proxy node when it has received the content from the origin Web server; otherwise, the update is registered in the buffer of the accelerator that pushed the object to the proxy cache. Updates for cache-remove operations are registered in the buffer of the proxy node. This scheme prevents replication of update representations when the proxy interacts with several accelerators and permits the use of multicast if available. The second component represents the proxy accelerator interface. This component consists of three threads: one thread listens for incoming connections from accelerators, one thread handles accelerator requests for transmission of the entire hint space, and the third thread handles the periodic updates. Accelerators may request a complete copy of the hint space at their initialization. The periodic update thread disseminates to the registered accelerators the update buffers released for transmission. If no buffer is released, the thread transmits the currently active batch update buffer. The third component of the Web proxy extension is the push handler. This component receives the new content and stores it in the cache. The content is sent in HTTP format as a POST request. While this requires the proxy node to duplicate the header processing that was already done in the accelerator, it decouples the implementations of the proxy cache and the accelerator. In order to hide the latency of the related I/O operations, multiple push connections can be concurrently active between accelerator and proxy. Experimental Evaluation. In the remainder of this section, we present an experimental evaluation of the overhead of hint management. The experimental evaluation is done with the Web Polygraph proxy performance benchmark [20]. Each component of the benchmark, polyclt, polysrv, Web Proxy, and proxy accelerators run on separate nodes of a switched Ethernet-based LAN. The selected traffic model is meant to stress the hint-related component of the Web proxy application. Namely, the request model is characterized by a 75% miss ratio. The content size is fixed to 2 bytes, which results, with the headers, in about 370 bytes per request. Requests are issued by the polyclt node in best-effort mode, meaning that a new request is sent as soon as the reply for the previous one is received. Each experiment consists of 100,000 requests. During experiments, the Web proxy configuration is varied from original (i.e., no hint collection and updates), to hint collection but no updates, and to both hint collection and updates to one and two accelerators. During these experiments, the role of accelerators is reduced to receiving hint updates; requests are received directly by the Web proxy and, therefore, they do not pass through the accelerator. The update period is 5 min. The Web Polygraph statistics about the response times observed by the client show that the distributions of hit and miss response times are almost identical up to the 98-percentile for all the tested configurations. For the 99-percentile, the miss response times experienced with the configurations with active hint management differ by less than 1% from the performance of the original configuration. For hit response times, the 99-percentile difference is less than 10% of the performance in the original configuration. The throughput observed by the client is unchanged for the experiment with no active accelerators and -1.8% lower for the experiments with active accelerators. With Web Proxy internal monitoring instrumentation, we collected statistics about the overhead of hint collection when no proxy accelerator was active. This overhead includes hint computation and segment of the collection into batch update buffers. The statistics show a minimum of 31-sec, a median of 4-sec, and a 95-percentile of 123-sec. The above results demonstrate that hint collection and update protocol have negligible impact on the performance of a Web proxy node. 6. Related Work Previous research has considered the use of location hints in the context of cooperative Web Proxies [8,19,10,22]. In these papers, the hints help reduce network traffic by more efficient Web Proxy cache cooperation. Extending these approaches, we use hints to improve the throughput of an individual Web proxy. In addition, we propose and evaluate new hint maintenance techniques that reduce overheads by exploiting the interactions between a cache redirector and the associated cache. Web traffic interception and cache redirection relates our approach to architectures such as Web server accelerators [15], ACEdirector (Alteon) [1], Dy- naCache (InfoLibria) [13], Content Smart Switch [14], and LARD [17]. The ability to integrate a Proxy Accelerator-level cache renders our approach similar to architectures in which the central element is an embedded system-based cache, such as a hierarchical caching architecture [13], or the front end of a Web server [15]. Our proxy accelerator extends the approach to bypassing the Web proxy proposed in [14]; besides non-cacheable objects, our accelerator can identify the misses and process them locally. Trading the throughput of the routing compo- nent, which is not the bottleneck resource, our approach can boost even more the throughput of the bottleneck proxy nodes. Furthermore, our approach is similar to the redirection methods used in architectures focused on content-based redirection of requests to Web Proxy nodes [1,17]. However, our approach enables dynamic rather than fixed mappings of objects to WP nodes[1]. In contrast to the method in [17], redirection does not cause caching of multiple object replicas and is independent of client request patterns. 7. Conclusions Based on the observation that miss ratios at proxy caches are often relatively high [7,12,9], we have developed a method to improve the performance of a cluster-based Web proxy by shifting some of its cache miss-related functionality to be executed on a Proxy Accelerator - an extended content-based router implemented on an embedded system optimized for communication. Consequently, the Web proxy can service a larger number of hits, and the Proxy Accelerator-based system can achieve better throughput than traditional Web proxies [1,8,19,17] or systems in which the Web proxy is bypassed only for non-cacheable objects [14]. In addition, under moderate load, response time can be improved with a Proxy Accelerator main memory cache [15,13]. Our study shows that a single Proxy Accelerator node with about an order of magnitude better throughput for communication operations than a proxy node [15,16] can improve the cost-performance ratio of a 4-node Web proxy by -35%. Our study shows that eager registration is a "must" and that the Bloom Filter-based scheme is more appropriate than the directory scheme for large WP clusters or when PA and WP nodes have comparable power. The implementation of the accelerated Web proxy node demonstrates that the overhead added by hint management is not significant in comparison to typical proxy node overheads. --R Alteon Web Systems Scalable Content-aware Request Distribution in Cluster-based Network Servers in: <Proceedings>Proc Changes in Web Client Access Patterns. Distributed Packet Rewriting and its Application to Scalable Server Architectures in: Space/time trade-offs in hash coding with allowable errors The Measured Access Characteristics of World- Wide-Web Client Proxy Caches in: <Proceedings>Proc A Scalable Wide-Area Web Cache Sharing Protocol in: <Proceedings>Proc Performance of Web Proxy Caching in Heterogeneous Bandwidth Environments in: A Taste of Crispy Squid in: UC Berkeley Home IP HTTP Traces in: http://www. System Design Issues for Internet Middleware Services: Deductions from a Large Client Trace in: DynaCache: Weaving Caching into the Internet in: 3-rd International WWW Caching Workshop (<Year>1998</Year>) Design and Performance of a Web Server Accelerator in: IBM Netfinity Web Server Accelerator V2. A Performance Study of the Squid Proxy on HTTP/1.0 in: WWW Journal 1-2 (<Year>1999</Year>) Cache Digests in: http://squid. High Performance Benchmarking with Web Polygraph in: http://ircache. Design Alternatives for Scalable Web Server Accel- erators Beyond Hierarchies: Design Considerations for Distributed Caching on the Internet in: On the scale and performance of cooperative Web proxy caching in: --TR --CTR Chi-Hung Chi , HongGuang Wang, A generalized model for characterizing content modification dynamics of web objects, Web content caching and distribution: proceedings of the 8th international workshop, Kluwer Academic Publishers, Norwell, MA, 2004 Yu, The state of the art in locally distributed Web-server systems, ACM Computing Surveys (CSUR), v.34 n.2, p.263-311, June 2002
web performance;proxy cache;hints
592983
A Comparison of Protocols for Updating Location Information.
Detailed location information of mobile objects, for example that of a user with a mobile computer or phone, is an important input for many location-aware applications. However, constantly updating the location information for thousands of mobile objects is not feasible. Therefore, special update protocols for location information are required that transmit the information as efficiently as possible, that is requiring only few update messages, while still being effective in returning the location information with the desired accuracy. Different classes of such update protocols are described in this paper and a new combined protocol is proposed. To be able to compare their effectiveness and efficiency, we present an analysis for the minimum and average resulting accuracy of the location information in comparison with the number of messages transmitted. We also present the results of simulations that we have performed to back up our analysis.
Introduction As detailed information about the location of a mobile device or user has become widely available, mainly through the satellite positioning system GPS, more and more applications are using this information. Examples are car navigation or fleet management systems, as well as location-aware information services for cellular phones (which today use the less accurate information of the communication cell a mobile phone is currently located in). Most of these applications only consider the location of a single user. Further function- ality, like determining all mobile objects inside a given area or generating an event whenever a mobile object enters a room, requires a special location service, which collects and manages the location information for all of these objects. A universal global location service has been proposed for example by [Leo98] and [Maa97]. Such a location service has to deal with all sorts of mobile objects with various movement characteristics, for example objects that move continuously for a longer period of time, like cars and trucks, or objects that only move seldomly, like pieces of office equipment, for example a printer. The location information for these objects may have been acquired by different types of sensor systems and therefore has different degrees of accuracy. If the location information is acquired on the mobile object itself (e.g., via GPS), it has to be transmitted from the mobile object to a location server using wireless communi- cation. In case of a distributed service, where the location information is cached or replicated on a number of servers, the information also has to be transmitted between location servers. To control the transmission of location informa- tion, different update protocols with varying properties can be used. In case of a querying protocol the information is pulled by the receiver, while with a reporting protocol it is pushed by the sender. A combined protocol is also possible, where an optimal ratio between the number of updates and the number of queries has to be found. In this paper we will discuss the characteristics of these classes of update protocols and their subclasses as well as their strengths and weaknesses according to different areas of application. To be able to guarantee the receiver a certain accuracy of the returned location information, it usually has to be transmitted rather frequently. If the location information is transmitted from a mobile device to a location server, a wireless channel has to be used, where bandwidth is low and expensive. It is therefore important to use an update protocol that works efficiently, that is requires as few messages as possible, and effectively, that is achieves a desired accuracy of the location information on the server. In this paper we discuss the efficiency of the basic types of update protocols by means of an analysis of the number of transmitted messages and their effectiveness by analyzing the resulting minimum and average accuracy of the location information at the receiver. Although this work has been performed in the context of updating the location information on a location service, its results are applicable wherever location information is transmitted from a sender to a receiver The remainder of this paper is structured as follows: In Section 2 we look at related work and in Section 3 we describe the background for this paper as well as its technical environment. The different classes of update protocols and their properties are discussed in Section 4, while Section 5, besides defining a data-model for the location information and its accuracy, compares their efficiency and effectiveness by means of an analysis. In Section 6 we present some simulation results, to support the outcome of our analysis. Fi- nally, Section 7 contains our conclusion, as well as the plans for future work. Related Work Architectures for a location service have been discussed in the area of location-aware applications as an important component of such systems. However, different update protocols for the transmission of the location information have not been compared in detail. Early location services were usually developed for a particular positioning system, like the location service for the Active Badge system, which is described in [HH94]. For a similar positioning system, the ParcTab system developed at Xerox Parc, a location service has been created, where the location information of a mobile object or user is protected by a dedicated Home Agent, which controls access to it [ST94]. In this context the problem of disseminating the location information to a large number of clients has been discussed in [ST94]. Clients with similar queries receive the location information over dynamically created multicast groups in which the responses to these queries are grouped. In his PhD-thesis, Leonhardt ([Leo98]) proposes a universal location service, which is independent from the types of applications accessing it and the types of positioning systems it gets the location information from. A detailed location model is presented that integrates different types of location information. Another main concern of this work is privacy and security. For this purpose it defines requirements and policies for access control of a location service. The thesis also contains some thoughts about the design of a global, general-purpose location service but does not propose or evaluate a specific architecture. The management of the location information of their mobile phones is a very important issue in the area of mobile communication networks. When a call has to be forwarded to a mobile phone, a mobile communication network sends a paging message to all the cells the mobile phone may be in, whereupon the corresponding mobile phone answers to accept the call. The mobile phone on the other hand periodically reports its current cell, which is updated in the Visitor Location Register and Home Location Register. Various update strategies have been discussed with the goal of minimizing the overall number of paging and update mes- sages. In [BNKS95] a distance-based, a movement-based and a time-based protocol are compared. In this special application area the location information is a discrete iden- tifier, describing a certain cell of the communication net- work. Today, these cells have a size of about 1 to 10 miles but are expected to become smaller with the introduction of future technologies. Update policies for location management have been optimized for these special requirements. In [BD99] the LeZi-update approach is described, which uses the cell names as an alphabet for minimizing the updates by an mechanism similar to the Lempel-Ziv compression. In our work we have assumed that the location information is much more accurate, acquired for example by a GPS sensor, and that the location information can have different levels of accuracy. The location service, these update protocols are intended to be used for, also allows different types of queries (e.g., range queries), which are not supported by the location management components of mobile communication networks. In the DOMINO project a database for the tracking of mobile objects is being developed (see [WSCY99]). One intended area of application is fleet management for the trucks of a transportation company. To reduce the number of up- dates, it is assumed that the route, on which the mobile object is traveling, is known to the database in advance and that the object informs the database when it changes its route. Different dead-reckoning strategies are proposed and compared, where the database estimates the current location of a mobile object on this route based on its speed and the object sends an update of its location when it differs from this estimation by more than a certain threshold. The different strategies have different ways for determining this threshold or for adapting it dynamically. These dead-reckoning protocols have been designed on the assumption that the route of the mobile objects is known beforehand. In this paper we examine update protocols from a more general point of view, where we assume no knowledge of the mobile objects' future movements. 3 Background This section gives an overview of the technical environment our work is based on, namely the sensor systems the location information is acquired from and the network environment it is transmitted over. Finally, we describe the architecture for a universal distributed location service, which is being developed at our Institute and where we plan to apply the results of this work. 3.1 Positioning Sensors The location information for a mobile object can be determined through various types of positioning systems. A basic distinction can be made between tracking systems, where a system of stationary sensors determines the location of a mobile object, and positioning systems, where the location information is determined by a sensor on the mobile object itself. A typical tracking-system is the Active Badge system (see [WHFG92]), where an Active Badge carried by a user or attached to an office device periodically emits an infrared beacon, which is detected by an infrared sensor placed in each room of a building. The widely used global positioning system (GPS) is a positioning system, where a sensor determines its own position by taking a bearing on four of a whole of twenty-four special satellites (see [HWLC97]). Different types of positioning systems return the location information with varying formats and degrees of accuracy. It can be symbolic (i.e., the identifier of a region, like a room or communication cell) or geometric, where it is described by a global geodetic coordinate system like WGS84 (see [IA97]). The classifications in this paragraph have been taken from [Leo98]. Because no positioning system offers a global coverage (e.g., GPS works only outdoors) a universal location service has to be able to integrate the location information acquired by all types of positioning systems. 3.2 Network Environment In our work, we have assumed the following network environ- ment: Location servers and stationary clients, which query the location information, communicate over a fixed network. Mobile devices need to be connected to this network by a wireless link. They have to be able to communicate and to be contacted independently from their geographical location (e.g., using a mechanism like Mobile IP [Pe96]), because the device has to transmit location information to a server via the wireless link or is queried by the location server through it, if the information is determined locally. Mobile devices may also act as a client of the location service by posting queries over the wireless link. The wireless network can either be a wireless WAN, like GSM [MP92], a MAN, like the Metricom Ricochet network, or a LAN, for example according to the IEEE 802.11 standard [Com97]. A common characteristic of these wireless networks is that a connection can be temporarily lost while the device is at an unfavorable location (e.g., inside of a tunnel), a state which is called a disconnection. In most cases a wireless network can also not offer as good a bandwidth and latency as a fixed networks (see [Sat96]). 3.3 A Universal Location Service This comparison of update protocols has been done as part of our efforts to design and develop a universal distributed location service (see [LK99]). As mentioned before, this service will have to collect and integrate the location information from various sensor systems. Furthermore, it will provide a general interface through which different types of applications can access the location information in a general way. For reasons of scalability and to achieve short response times as well as high update rates, the information will be distributed between a number of location servers. The location service shall provide the following function- ality: It will support range queries, that is finding all mobile objects inside a given area, as well as position queries, which request the current location of a certain mobile object. A more specialized query is, for example, the search for the nearest object to a certain location. It is intended, that the location service will additionally support an event mecha- nism, where the client can define a certain predicate and is notified by the service whenever this predicate becomes true. A client program may for example be notified, whenever a mobile object enters or leaves a given area or when a mobile object meets another one, that is comes within a certain distance from it. location server client programs location register object register sensor systems object sightings location service API queries, events results, notifications lookups lookups updates updates Figure 1: Main components of a universal distributed location service We have proposed an architecture for such a location service, which is shown in Figure 1. Its components and their functionality are described next: ffl Clients and sensor systems: Clients of the location service are location-aware applications, which run either on a mobile or on a stationary computer and query the location servers using the interface mentioned earlier in this subsection. A client on a mobile device is usually also equipped with a positioning sensor or is being tracked by one. Sensor systems, which, depending on their type, determine the location information for one or more mobile objects, are associated with a certain location server and update the location information for these objects. ffl Location server: Each location server is responsible for a certain geographical area and stores the location information for all mobile objects that are currently inside of this area. If a mobile object leaves the area, a hand-over has to be performed with the server, into whose area the object has moved. A location server answers the clients' queries concerning the mobile objects it is responsible for and forwards other queries to the appropriate server(s). This is done by accessing one of the two following registers: ffl Object register: The data, which belongs to a single mobile object, is stored in the object register. It can contain general information about the object or information that is determined by the location service, like the maximum speed of the object. The object register also stores the addresses of the location servers the location information for this object is currently stored on. It can easily be distributed, so that each register server is responsible for a certain set of objects. ffl Location register: The location register stores the association between the location servers and the area they are responsible for. It is used to find the appropriate location servers for a range query. If necessary, the location register can be distributed across different servers, which are organized in a tree structure similar to the Domain Name System (DNS) of the Internet. Because of privacy concerns, security and access control for such a location service are very important issues. A user should be able to control in detail who and to what degree has access to his location information. Some solutions for this problem have been presented in [Leo98]. In this paper we do not discuss privacy any further, but will address the issue in future work. 4 Update Protocols In this section the different classes of update protocols are introduced and their main properties are discussed. The protocols are used to update a remote secondary copy of the location information for a certain mobile object based on the contents of a primary copy. The goal is to guarantee a given accuracy of the location information in the secondary copy. The information of the primary copy is either determined directly by a positioning system or is the information stored on another location server. We call the component which manages the primary copy the source of the location information. The secondary copy is usually stored on a special or general location server, which is queried by local or remote applications. Figure 2 shows the components which are involved in the transmission of the location information. source/ primary copy update protocol fixed networkor mobile communication queries location server mobile deviceor location server sensor system application server/ secondary copy Figure 2: Components involved in the updating of location information. 4.1 Classification The update protocols can be divided into three main classes, namely querying, reporting and combined protocols, where each class has a number of typical variants. Each of these protocols has its characteristic properties and is suitable for a given environment or for certain requirements. Some of the reporting protocols are similar to the paging/update schemes that are used for Location Management in the re-search area of Personal Communication Service (PCS) networks and are summarized, for example, in [BNKS95]. The classes of update protocols and their relationships are shown in more detail in Figure 3. querying updateprotocols reporting simple simple caching periodic distance- based time- based dead- reckoning combined Figure 3: Classification of update protocols for the transmission of location information. 4.1.1 Querying Protocols A protocol is called a querying protocol, if the server decides when to request the location information from the source. In this case, the source can be very simple, as it does not need to keep extensive information about its state or realize a complicated logic. This may be important for small mobile devices. The simple and the cached querying protocols transmit the location information on-demand, that is only when it is queried by an application. If the location information is queried only seldomly, these protocols are more efficient than the reporting protocols described in 4.1.2, because the information is not transmitted unnecessarily (see Section 5). Simple: In the most simple form the server requests the location information from the source each time it is queried by an application and therefore does not need a secondary copy at all. This leads to the highest possible accuracy of the location information, but also to a large number of mes- sages, if the information is queried very often. The response time of the server is also comparatively large, as the server has to contact the source for each query. On the other hand, this protocol has to be used, if because of privacy concerns the user does not allow his/her location information to be stored on any location server other than on his/her personal device. In this case, the personal device has to be accessed for each query to check for access rights. An example for this is the location service presented in [ST93], where a special program, a so-called User Agent, stores and protects the location information of a certain user. Here, we use this protocol mainly for comparison. Cached: The cached querying protocol is an optimization of the simple protocol, where the server stores a cached copy of the last transmitted location information. When the location information for this mobile object is queried by an application, the server estimates the accuracy of the cached copy and returns it, if it is considered to be accurate enough. Otherwise, the server has to send a request to the source like in the simple protocol. A pessimistic cached querying protocol uses the distance, a mobile object may have traveled at its maximum speed, for the estimation of the accuracy. If the maximum speed of the mobile object is much higher than its average speed, the cached copy is often unnecessarily considered not accurate enough. An optimistic cached querying approach could use the average speed of the mobile object instead of the maximum one. However, with any optimistic estimation for the accuracy of the location infor- mation, the actual location of the mobile object can in the worst case differ from the reported location by more than the requested accuracy. With a cached querying protocol the response time of the server is varying and depends on whether the server can use the cached copy or has to contact the source. Periodic: If the server queries the location information periodically from the source with a certain time interval D, the protocol is called a periodic querying protocol. Although here the initiative is reversed, this protocol has the same properties as the time-based variant of the reporting protocols described next. 4.1.2 Reporting Protocols In case of a reporting protocol the initiative is on the side of the source. It remembers which location information it has sent last to the server and therefore knows which location is stored there. An update of the location information is sent, whenever a comparison to the current position of the mobile object exceeds a certain distance or time threshold. The server can therefore conclude the maximum uncertainty of its copy from the value of this threshold. With an unmodified reporting protocol the server always answers the queries of the applications by directly returning the information currently stored in its copy. It can therefore only return the location information with an accuracy specified by the value of the threshold, even if a more accurate information is queried by an application. The response time of the server on the other hand is shorter, as it does not have to contact the source. A reporting protocol is usually more efficient for a given maximum accuracy than a querying protocol, if the location information is queried often (see Section 5) or if the server has to check periodically for the occurrence of events. Simple: The most simple reporting protocol sends the location information each time the value of the primary copy Message rate is adjusted to mobility of object. Message rate is adjusted to query rate and requested accuracy. Upper bound for spatial uncertainty of returned information. Allows applications to specify a desired accuracy. Can detect disconnections querying: simple \Theta (\Theta) \Theta \Theta cached \Theta (\Theta) \Theta \Theta periodic \Theta reporting: simple (\Theta) (\Theta) time-based \Theta distance-based \Theta \Theta dead reckoning \Theta \Theta combined \Theta \Theta \Theta \Theta (\Theta) Table 1: Summary of properties for different update protocols. changes, for example because a sensor system has determined a new location sighting. In this case the number of messages depends on the update rate of the sensor system and can be rather high. Depending on the type of sensor system, the simple reporting protocol also falls into one of the two following categories. We therefore do not consider it any further. Time-based: With a time-based protocol the location information is transmitted periodically, after a certain interval of time T has elapsed. The update rate is fixed and does not depend on the behavior of the mobile object, which guarantees a certain temporal but not a spatial uncertainty of the information. If the object moves slowly or not at all, there is little or no difference in the location information transmitted by the messages to the server. If the object moves fast, not enough messages are sent to achieve a high accuracy. Distance-based: The distance-based protocol sends an up-date of the location information whenever the geographic distance between the current location of the mobile object and the last reported location becomes greater than a given threshold D. As this protocol sends more messages if the mobile object is moving fast and less messages if it is slower or stationary, it is often more efficient for objects that perform sporadic movements between periods of immobility. This is, for example, typical for users in an office environment or for all sorts of office equipment. It is also possible to integrate the time-based and the distance-based protocol to combine their properties. Dead-reckoning: Dead-reckoning is an optimization of the distance-based protocol. Here, the server estimates the current location of the mobile object based on its old location, its speed and the direction of its movement or on information about the route of the object. The source also calculates this estimated location and sends an update when it differs from the actual location by more than a certain distance threshold. A dead-reckoning approach performs very well, if the object is moving with constant speed in a given direction for some time or if the route of the mobile object is known in advance. In the second case the update costs can, according to [WSCY99], be reduced by 85% compared to other reporting protocols. 4.1.3 Combined Protocol While a plain querying protocol can not be adjusted to different mobility characteristics of the mobile objects, a reporting protocol does not consider the query rate and the accuracy requested by the applications. With a combined protocol, which integrates the distance-based reporting protocol and the cached querying protocol, both of these features can be achieved. Similar to the distance-based reporting protocol, the source may update a secondary copy on the server to achieve a given spatial accuracy D. If the location information stored in the secondary copy is not accurate enough for a certain query, the server requests the information from the source as in a querying protocol. To minimize the total number of messages consisting of updates and queries, it has to be decided whether to update a secondary copy at all and what distance threshold D has to be used, depending on the mobility properties of the mobile object and on the queries by the applications (see Section 5). If the source monitors the mobility properties and the server those of the queries, they can adapt the properties of the protocol dynamically by increasing/decreasing the distance threshold D or by starting/stopping the updating of the secondary copy altogether. With the combined protocol, the response time is again varying and depends on whether the secondary copy on the server is accurate enough or if the server has to contact the source. 4.2 Behavior in Case of Disconnection A problem that frequently occurs with mobile devices and wireless data transmission is a temporal disconnection of the communication link. A protocol which is intended to transmit location information from a mobile device to a server via a wireless communication link has therefore to be able to deal with such disconnections. In the following paragraphs we discuss the properties of the basic protocols with regard to disconnections, namely how long it takes to detect a disconnection and the maximum uncertainty of the location information returned during that time. If the server has detected a disconnection it can, according to the preferences of its users, either return an error message when queried for the location information or return the last known position together with an indication that it is obsolete. Where necessary we sketch appropriate modifications to these protocols that enable them to deal with disconnections. Querying protocols: In case of a querying protocol, the disconnection is detected as soon as the server receives a query and requests the location information from the source (and, in case of a cached querying protocol, also can not answer the query from its cached copy). Therefore, the server does not at any time return less accurate information than in the normal case. Time-based reporting protocol: When using a time-based protocol a disconnection can be detected if no update messages have arrived after the time threshold T has elapsed since the last update. Again the server only returns location information with the specified accuracy. Distance-based reporting protocol: With the basic distance-based protocol the server is not able to detect a disconnec- tion. Instead, it assumes that the mobile object has not moved by more than the distance threshold D and returns the old location information as being up-to-date. To be able to detect disconnections, the distance-based protocol can be combined with a time-based one, by having it sent a location update at least every time interval Tmax . A suitable value for Tmax has to be found by considering the message overhead versus the maximum uncertainty, the user is willing to tolerate in case of an error. Dead-reckoning reporting protocol: A basic dead-reckoning protocol has the same problems with disconnections as the distance-based protocol and the same mechanism can be used to deal with them. In [WSCY99] another technique for detecting disconnections has been introduced, were the distance threshold D is being continuously decreased until the source is forced to send an update message or it has become increasingly likely that a disconnection has occurred. Combined protocol: How the combined protocol reacts to disconnections depends on how accurate a secondary copy is kept on the server. A disconnection is detected as soon as the server has to request the location information from the source. If this is not the case, the combined protocol behaves like the distance-based protocol and can return inaccurate location information. Again, a maximum time interval Tmax between update messages should be specified. A summary of the properties for the different update protocols discussed in this section is contained in Table 1. It shows, whether the message rate of a protocol depends on the mobility characteristics of the mobile objects and whether it depends on the queries of the applications. It also indicates, whether the protocol can guarantee an upper bound for the returned location information and if it allows the applications to request a certain accuracy of the infor- mation. The final column shows, whether the protocol is able to deal with disconnections. Analytical Comparison of the Protocols In this section, the effectiveness and efficiency of the update protocols are compared in more detail, by means of an analy- sis. First, the variables that appear in the following analysis are described and a general data model for the location information and its accuracy is defined. The analysis considers the number of messages required for updating the location information on the server as opposed to the minimum and average accuracy of the location information returned to an application. Corresponding formulas are shown first for the querying and reporting protocols, then for a combined pro- tocol. Finally, we discuss the characteristic properties of these protocols, based on the results of the analysis. Only the updates and queries for the location information of a single mobile object are considered, as the results can be easily applied to several objects. For reasons of simplicity we also do not consider the delays for the transmission of the messages, which would lead to a further (temporal) uncertainty of the location information. Because it is added uniformly to all messages, the delay does not substantially affect the comparison of the protocols. The efficiency of an update protocol is here considered to be the average number of messages m that is transmitted per second between the source and the server. Messages can be location updates generated by the source as well as location requests from the server. The accuracy of the location information on the server, which describes the effectiveness of a protocol, is defined by the maximum and average deviation between the information returned as a result of a query and the actual position of the mobile object. Up to now, we have assumed that applications are only interested in the maximum uncertainty of the location information, for example if they want to be able to determine which room of a building a certain user is in. However, other types of applications may be more concerned with the average accuracy of the information. An example for this could be a map, where the current locations of a number of mobile objects is shown. In our analysis we have therefore considered both, the maximum and the average accuracy, dmax and davg . The calculations for m, dmax and davg are based on the behavior of the mobile object, defined by its average and maximum speed, vavg and vmax , and on the uncertainty up with which the location information is available at the source. Moreover, the calculations depend on how frequently and to what accuracy the location information is queried by the applications, which is described by the average number of queries per second q and the average of the requested uncertainty uq . The calculations are also affected by characteristic parameters of the different protocols, namely the maximum uncertainty required of the secondary copy us , the speed that is assumed for the mobile object by the cached querying protocol vasd , and the time threshold T for a time-based or the distance threshold D for a distance-based reporting protocol. Table 2 shows a summary of these variables. Some important restrictions between these variables are as follows: Because the information transmitted to the server can not be more accurate than the information of the source, the uncertainty of the secondary copy is always equal to or greater than the uncertainty of the primary copy, us - up . Also, it is not practical for an application to request more accurate location information than the one available at the source. Hence, we also assume that uq - up . For the calculation of the average deviation we assume that the average uncertainty of the location information on the source is half of the maximum uncertainty. This may not be true for all sensor-systems (e.g., the average uncertainty is less than half of the maximum if the source is a GPS sensor), but does not have much influence on the outcome of the analysis. The restrictions described here apply to all formulas in this section and are not explicitly stated again. 5.1 Location and Uncertainty Model For the following analysis we use a uniform model of the location information returned by the different positioning systems and of its accuracy. The location information is supposed to be given by global geodetic coordinates, for example of the WGS84 stan- Maximum speed of mobile object. vavg Average speed of mobile object. q Average number of times per sec- ond, the location information of the mobile object is queried. uq Average of requested accuracy for this location information. q Accuracy requested in one certain query. up Uncertainty of primary copy. Tp Update interval of primary copy. l Time at which a certain location sighting l has been acquired. us Uncertainty of secondary copy. vasd Speed that is assumed for estimating the accuracy of the location information. Time threshold of the time-based reporting protocol. Distance threshold of the distance-based reporting protocol. Average number of messages transmitted per second between source and server. Maximum deviation between the location information the server returns as result of a query and the actual position of the mobile object. davg Average deviation. Table 2: Variables used in the analysis. dard, where the distance between two locations can easily be calculated. Symbolic location information (e.g., the name of a room) returned by some of the positioning systems can easily be converted to geometric coordinates. The geometric coordinate for a region described by a certain symbolic name is given by the center of the region, while the spatial accuracy is the distance from the center to the farthest point in the region. A location sighting can have a temporal as well as a spatial accuracy. The temporal accuracy is given by the time that has elapsed since the location sighting has been acquired, while the spatial accuracy is defined by the maximum distance between the position reported by the sighting and the actual position of the mobile object. For many location-aware applications the spatial accuracy of a sighting is more important, because the applications are concerned with the spatial relationship between (mobile) objects. The uncertainty u l (t) of a certain location sighting l describes the spatial accuracy at a given time t - t l . At the time of the sighting the uncertainty is determined by the accuracy up of the sensor system. The uncertainty at a later time t can be estimated by the distance the mobile object may have traveled during the time . If a maximum bound for the velocity of the mobile object (vmax) ex- ists, the maximum uncertainty of the location sighting can be calculated by adding the distance the object can have traveled to the uncertainty of the sensor system (see Figure 4). This is described by the following equation: For example, the maximum velocity of a car can be set to 300 km/h (to be on the safe side) and the maximum velocity locationsighting uncertainty Figure 4: Uncertainty of location information depending on the accuracy of the sensor system and the elapsed time. of a person to 36 km/h. By evaluating the history of the location sightings of a mobile object, it is also possible to determine a maximum velocity separately for each object. In many cases, however, the location sighting will be more accurate than the uncertainty given by this equation, as the mobile object will not be moving at its maximum speed or not in a straight line. A person in an office will usually remain relatively stationary at a certain location (e.g., his office or a conference room) for a longer interval of time, while moving at a comparatively high speed between these locations. If the uncertainty does not need to be limited by an upper bound or the maximum velocity is not known, an assumed velocity vasd (e.g., the average velocity vavg of the mobile object) can be used to predict the uncertainty instead of the maximum velocity vmax . In some cases this will lead to an actual deviation, which is greater than the predicted one. This error will be worse for the sporadic movement of a person in an office, compared to the more steady movement of a traveling car. 5.2 Basic Protocols The equations for an approximate calculation of the number of transmitted messages as well as for the maximum and average uncertainty of the location information are shown next for the basic querying and reporting protocols. We have not yet performed the analysis for the dead-reckoning pro- tocol, because it depends to a great extent on the movement characteristics of the mobile object and we did not want to assume prior knowledge about the route of the object. In general the dead-reckoning protocol can be expected to behave similar to the distance-based reporting protocol. Simple querying protocol: In the simple querying protocol, which is presented here only for reasons of comparison, the number of exchanged messages is equal to the number of queries as every query is forwarded to the source. If we do not consider the communication delay, the location information presented to the applications has the same accuracy as the location information of the source. Cached querying protocol: The cached querying protocol uses Formula 1 to estimate the uncertainty of the location information stored on the server and if accurate enough returns it without contacting the source. For the speed of the mobile object used in this formula it can take a pessimistic approach with vmax (as shown in Formula 1), or an optimistic approach with vavg instead of vmax . In the following considerations we represent the assumed speed of the mobile object by vasd , which is then replaced by the speed used in a concrete version of the protocol. Compared to the simple protocol, the caching approach eliminates all messages where the query falls into the period of time in which the cached copy of the location information is considered accurate enough, (derived from Formula 1). The number of messages transmitted between source and server is therefore the total number of queries, multiplied by the fraction of queries that can not be answered from the cache, which is approximately The maximum uncertainty of the returned location information is given by the time \Deltat the cached copy is considered accurate, multiplied with the maximum speed of the mobile object. To this the initial uncertainty of the location information of the primary copy has to be added. The average uncertainty can be approximated by the uncertainty of the sensor system, if the location information has to be requested from the source. Otherwise, the average uncertainty is half the distance that the mobile object can have traveled at average speed during the time in which a cached copy is considered valid. The corresponding Formula 7 is shown in a simplified form. If the assumed speed vasd of the mobile object is set to its maximum speed vmax , the uncertainty of the location information presented to the querying application is limited by uq (substituting the variables in Formula 6). The protocol can therefore in this case guarantee the accuracy requested by the applications. Time-based reporting protocol: As the time-based protocol sends a location update periodically, the number of messages depends only on the time threshold T . To achieve a given uncertainty us for the secondary copy, T has to be set to the time, the mobile object needs to cover the distance given by us minus the uncertainty of the primary copy at maximum The maximum deviation can be calculated by adding the distance that the object may have traveled during time T to the uncertainty of the location information at the source. For the average deviation the average speed is used, instead of the maximum one. Distance-based reporting protocol: The number of messages transmitted with the distance-based protocol is also independent of the number of queries. To guarantee a given uncertainty us of the location information on the server, the distance threshold D is set to us \Gamma up . The average number of messages per second can then be calculated by the time interval that the mobile object requires to cover this distance at average speed. us \Gamma up If no messages are lost due to problems with the network connection (compare Section 4), the distance-based protocol achieves the requested maximum uncertainty us for the secondary copy on the server. On average the uncertainty is half of it. us (12) us 5.3 Combined Protocol The combined protocol comprises elements of the distance-based reporting protocol and the pessimistic cached querying protocol, described in the previously. Therefore, it also combines elements of their behavior. The transmitted messages are the updates sent to keep the secondary copy up- to-date as well as the requests for the location information on the source sent by the server, if the secondary copy is not accurate enough. The number of updates can be calculated similar to the number of messages in the distance-based reporting protocol. The number of requests is the number of messages in the cached querying protocol, where the query rate is reduced by the probability that the uncertainty demanded in a query is lower than the accuracy of the location information already stored in the secondary copy, P (u us ). The probability distribution depends on how the location information is queried and used by the applications. us \Gamma up Because the combined protocol queries the source whenever the location information stored in the server is less accurate than the accuracy demanded in a query, it can always return the desired accuracy similar to a cached querying protocol. The maximum uncertainty returned is therefore the uncertainty demanded in the queries. The average uncertainty is again a combination of the average accuracy of the reporting protocol, in case the accuracy requested in a query can be met by the secondary copy, and of the average accuracy of the querying protocol, otherwise. us The behavior of the combined protocol can be controlled through the uncertainty of the secondary copy. If it is set to a low value, the protocol behaves like the distance-based reporting protocol, if it is set to infinity, the combined protocol can be made to behave like a cached querying protocol. 5.4 Discussion To compare the properties of the protocols, we first look at their efficiency, that is the number of messages that are transmitted between source and server. These are shown in Figure 5 for the querying and reporting protocols depending on the number of queries per second. For the other parameters we have assumed values taken from GPS traces that we have obtained from a Differential GPS sensor during a car ride and which are described in more detail in the next section. The average and maximum speed for the mobile object is m/s. The uncertainty of the primary copy up is that of the sensor system and is set to 5 m. For the queries we have assumed a fixed demanded uncertainty of messages per second queries per second simple querying pessimistic cached querying optimistic cached querying time-based reporting distance-based reporting Figure 5: Analytical results: A comparison of the number of messages transmitted with the querying and reporting protocols for a given query rate. Generally, a distance-based reporting protocol performs better than the time-based one and the optimistic cached querying protocol better than the pessimistic one. While the properties of the reporting protocols are independent from the query rate, the number of messages increases with a querying protocol with the number of queries. If the query rate is low, a querying protocol is therefore better than a reporting protocol. For very low query rates a simple protocol is sufficient. If the query rate is higher, the distance-based reporting protocol requires fewer messages than the pessimistic cached querying protocol, which is in many cases to be preferred to the optimistic protocol as it can guarantee a maximum uncertainty of the returned location information. The query rate at which the performance of the reporting protocol becomes better than that of the querying protocol is given by the following inequation: us \Gammau p If vavg is low compared to vmax , which is the case for mobile objects with sporadic movements, the limit for the query rate in Formula 17 approaches 0. For such objects, the distance-based reporting protocol is always better than the cached querying one. If vavg approaches vmax , the limit for the query rate approaches infinity. In this case, the cached querying protocol always performs better. However, the number of messages for high query rates is only a little higher with the reporting than with the querying protocol. The other aspect of the update protocols, which is discussed in this paper, is their effectiveness, that is to what degree the server can fulfill the accuracy demanded in a query. The maximum and average uncertainty (dmax and davg) of the location information returned by the server is shown separately for the two cached querying and the two reporting protocols in Figure 6. The uncertainty is shown depending on the speed ratio between the average and the maximum speed of the mobile object (vavg=vmax ), which gives a good indication of its mobility characteristics. A mobile object with a low ratio moves sporadic, an object with a ratio approaching 1 moves steadily. For the graphs we have assumed a query rate q of 0:1 queries per second. The pessimistic cached querying approach can meet the required accuracy independently from the mobility of the object. For a low speed ratio it also has a very low average uncertainty. In comparison, the optimistic cached querying protocol, which is much more efficient, results in an almost constant average uncertainty, a little below the uncertainty requested by the client. However, for a low speed ratio the maximum uncertainty can be much higher than the one demanded by the client. This protocol is therefore not suitable for objects with sporadic movement, if the querying application can not tolerate variations in the uncertainty. Both reporting protocols can meet a fixed uncertainty demanded by the client. The time-based protocol has a low average uncertainty for a small speed ratio, compared to the distance-based protocol, which offers a constant average uncertainty for all types of mobile objects. As the former requires more messages, the later is to be preferred. The reporting protocols, however, can only return the location information with the accuracy currently stored on the server. They can not meet the demand of applications which query the information with a lower requested accuracy, u us . From the viewpoint of their effectiveness, the distance-based reporting protocol is equal to the cached querying protocol if only a fixed requested accuracy is considered (al- though, the later has a better average uncertainty, especially for a low speed rate). The reporting protocols are not suit- able, if the location server allows the client to flexibly specify the requested accuracy and it is important to meet these requirements As mentioned before, the performance of the combined protocol depends on the uncertainty with which the secondary copy is updated as well as on the query rate. In Figure 7, the number of transmitted messages is shown for the combined protocol depending on the uncertainty of the secondary copy and the query rate. For the uncertainty demanded in the queries we have assumed a Gaussian distribution with an expected value of and a standard deviation of 20 m. All other parameters have the same values as before. Figure 7 indicates an optimal value for the uncertainty of the secondary copy for higher query rates, whereas for lower query rates the number of messages is lowest, if the uncertainty is set to infinity. It is therefore possible to use Formula 14 to decide, whether it is useful to update a secondary copy at all and to find an optimal uncertainty value for it. The uncertainty value can either be calculated with the necessary parameters set to the average values determined for a certain environment, the protocol is going to be used in. Or it can be adjusted dynamically according to changes in the environment. For comparison with the other protocols, Figure 8 shows the number of messages transmitted with the combined protocol depending on the query rate. Optimal values for the uncertainty of the secondary copy have been taken from Fig- uncertainty of returned location information in meters ratio between average and maximum speed of mobile object pessimistic cached querying maximum uncertainty average uncertainty101000 uncertainty of returned location information in meters ratio between average and maximum speed of mobile object optimistic cached querying maximum uncertainty average uncertainty101000 uncertainty of returned location information in meters ratio between average and maximum speed of mobile object time-based reporting maximum uncertainty average uncertainty101000 uncertainty of returned location information in meters ratio between average and maximum speed of mobile object distance-based reporting maximum uncertainty average uncertainty Figure Analytical results: Maximum and average uncertainty of the location information on the server for the optimistic and pessimistic cached querying protocols and the time-based and distance-based reporting protocols. ure 7. The combined protocol integrates the positive features of the more simple querying and reporting protocols. Al- though, for a higher query rate it is less efficient than the distance-based reporting protocol, it requires much less messages for a low query rate, similar to the querying protocols. Unlike the reporting protocols, it is also able to meet an arbitrary requested accuracy. Its disadvantages are that it is more complicated to implement, because an optimal value for the uncertainty of the secondary has to be found, and that communication errors are more difficult to handle. In summary, the combined protocol is suitable for most cases, especially if a protocol is needed that performs equally well in different environments or even can adapt to them. Of the querying and reporting protocols the pessimistic cached querying protocol and the distance-based protocol are to be preferred, because they combine an upper bound for the spatial accuracy of the returned location information with a good efficiency. Of these two, the former is decidedly better for objects with sporadic movements and for higher query rates, but does not allow the accuracy of the location information to be requested on application-level. 6 Simulation Results For the update protocols discussed in Section 5, we have performed various simulation runs based on ten actual GPS traces. To this end we have developed a simple simulator, which implements a location source with a primary copy and a location service with a secondary copy for the location information of one mobile object. The information on the source is updated from the GPS trace, while the server receives randomly created location queries with an exponentially distributed time interval and a fixed or, in case of the combined protocol, a Gaussian distributed accuracy. The mean value for the exponential distribution of the time interval between queries is the inverse of the query rate q. The fixed accuracy requested in the queries is 100 m for the basic protocols. The Gaussian distributed accuracy for the combined protocol has a mean value of 100 m and a standard deviation of 20 m. Into this simulation environment different update protocols can be inserted that specify, how the source and server react to incoming updates or queries and how the location information is transmitted. As in the analysis, the delays for transmitting the messages are not considered. The ten GPS traces used in our simulations have been acquired on a typical car ride in the commuting traffic around the city of Stuttgart. These traces have an average length of about 28 minutes and an average speed of 44.68 km/h. A Differential GPS sensor with an accuracy between 2 and 5 meters has been used to determine the location information, which is written to a file every second. For each trace we have performed various simulation runs, where the number of transmitted messages has been calculated depending on a given query rate. Figure 9 shows the results of these simulation runs as the mean values of 100 separate simulations. The parameter values for the protocols (e.g., the average and maximum speed) have been taken speed ratio pessimistic optimistic distance-based time-based combined cached querying cached querying reporting reporting max. avg. max. avg. max. avg. max. avg. max. avg. Table 3: Simulation results: Maximum and average accuracy of the location information for the different update protocols. combined protocol uncertainty of secondary copy in meters queries per second0.40.6 messages per second Figure 7: Analytical results: Number of messages transmitted with the combined protocol depending on the number of queries and the uncertainty of the secondary copy on the server. from the respective trace. In general, these simulation results confirm the assumptions we have made for the analysis. The number of messages for the cached querying protocols is a little higher than in the analysis. One reason for this is that the location information queried at the source is not determined exactly when queried, as we have assumed there. Instead, it has already an average age of Tp=2 where Tp is the update interval of the sensor system, in our case 1 second. Therefore, the location information on the server has to be updated a little sooner than calculated. Also, the maximum velocity of the mobile object is sometimes difficult to obtain from the GPS traces, because the uncertainty of the positioning systems can have a great influence on the distance of two neighboring entries. We therefore take the average of 5 seconds to determine the maximum speed. It happens a number of times that the the flow of location information from the DGPS sensor is interrupted for a short time, which may also lead to variances in the behavior of the protocols. For the combined protocol, we have again determined the message count depending on the uncertainty value for the secondary copy as well as on the query rate. The results of this simulation are shown in Figure 10. Although the results are again a little higher than the values determined in our analysis, the analysis reflects their characteristics nicely. For five GPS traces with different speed ratios ranging from 0.26 to 0.354 the maximum and average resulting accuracies are depicted in Table 3. The query rate has been set to 0.1, as the results tend to vary more for higher rates, while the other parameters have the same values as before. For the accuracies, as compared to the number of transmitted messages, the results depend more on the characteristics0.011100 messages per second queries per second combined Figure 8: Analytical results: Number of messages transmitted with the combined protocol depending on the number of queries. For the uncertainty of the secondary query the optimal values taken from Figure 7 are used. of a certain trace. Because we have averaged the maximum speed over 5 seconds, the apparent speed of two neighboring entries in the trace may be higher than the one we have assumed for the parameters of our protocols. This can lead to a maximum accuracy that is greater than the requested one (100 m), when using a cached querying or the combined protocol. 7 Conclusion and Future Work In this paper, we have described different protocols for transmitting location information and have compared them according to their effectiveness and efficiency. These protocols may be used, wherever location information is queried or updated continuously. With the help of the analysis a suitable protocol can be found for a given environment. Fur- thermore, the analysis gives an estimation for the expected network load and the average uncertainty of the transmitted location information, as well as an upper bound for its maximum uncertainty. Based on the querying and reporting protocols, we have proposed a combined protocol that integrates most of the advantageous properties of the basic protocols. Here, the analysis can give optimal parameter settings for a certain environment. For our future work, we plan to run further simulations with different types of mobile objects, for example a person while window-shopping in a city or riding a bicycle, to look at various movement characteristics. The results will then be used, to further improve the analysis, where nec- essary. We also plan to extend the combined protocol by adding a mechanism that allows it to adapt to changes in messages per second queries per second simple querying pessimistic cached querying optimistic cached querying time-based reporting distance-based reporting combined Figure 9: Simulation results: Number of messages transmitted with the querying, reporting and combined protocols determined by simulation. combined protocol uncertainty of copy in meters queries per second0.40.60.8 messages per second Figure 10: Simulation results: Number of messages transmitted with the combined protocol determined by simulation depending on uncertainty of secondary copy as well as on the query rate. the environment, dynamically. This mechanism will have to monitor critical parameters (e.g., the average speed of the mobile object) and to adjust the uncertainty of the secondary copy or to stop updating it at all, accordingly. Such a mechanism can increase the efficiency of the protocol in a dynamic environment. It can also be the basic mechanism for an adaptive location service, which is able to adjust its whole architecture. Because we wanted to look at the protocols from a general viewpoint, we have not examined the dead reckoning protocols in more detail, as they work best, when they have some knowledge about the future movement of their users. Nevertheless, they can increase the efficiency of a location service to a great deal. The prediction of a mobile object's future movements can be based on four possible sources: much further effort, a simple protocol could use the current speed and direction of the mobile object to predict its future movements. Also, information about the route of the mobile object can come from the user or an application program, for example a car navigation system (com- pare [WSCY99]). In a more flexible solution, the route can be predicted based on the former movements of a mobile object, as for example a person will mostly use the same route when commuting to his/her office (compare [BD99]). Finally, a map-based approach can find the street a user is moving on and predict his future movements according to the course of the street, possibly also considering the usual movement pattern of mobile objects on this street. We plan to integrate a dead-reckoning protocol into our comparison of update protocols and to implement a map-based protocol for the planned location service. Our main goal is the development of a universal distributed location service, as it has been described shortly in Section 3. We have already developed a prototype for a centralized service, and are currently looking at and comparing different architectures for distributing this service efficiently. Future work includes aspects of efficiently managing the location data and of creating a security mechanism, which allows a user to control in detail who and to what degree has access to the information about his location. --R An information-theoretic approach to track mobile users in PCS networks Mobile users: To update or not to update? IEEE Computer Society LAN MAN Standards Committee. A distributed location system for the active office. Global Positioning System: Theory and Practice. DoD World Geodetic System Supporting Location-Awareness in Open Distributed Systems An architecture for a universal The GSM System for Mobile Communications. IP mobility support. Fundamental challenges in mobile computing. Providing location information in a ubiquitous computing environment. Disseminating active map information to mobile hosts. The active badge location system. Updating and querying databases that track mobile units. --TR --CTR Reynold Cheng , Kam-Yiu Lam , Sunil Prabhakar , Biyu Liang, An efficient location update mechanism for continuous queries over moving objects, Information Systems, v.32 n.4, p.593-620, June, 2007
location-awareness;location update protocols;location service
593003
Organized Terminode Routing.
We consider the problem of routing in a wide area mobile ad hoc network called Network. Routing in this network is designed with the following objectives. First, it should scale well in terms of the number of nodes and geographical coverage; second, routing should have scalable mechanisms that cope with the dynamicity in the network due to mobility; and third, nodes need to be highly collaborative and redundant, but, most of all, cannot use complex algorithms or protocols. Our routing scheme is a combination of two protocols called Routing (TLR) and Terminode Remote Routing (TRR). TLR is used to route packets to close destinations. TRR is used to route to remote destinations. The combination of TLR and TRR has the following features: (1) it is highly scalable because every node relies only on itself and a small number of other nodes for packet forwarding; (2) it acts and reacts well to the dynamicity of the network because as a rule multipath routing is considered; and (3) it can be implemented and run in very simple devices because the algorithms and protocols are very simple and based on high collaboration. We performed simulations of the TLR and TRR protocols using the GloMoSim simulator. The simulation results for a large, highly mobile ad hoc environment demonstrate benefits of the combination of TLR and TRR over an existing protocol that uses geographical information for packet forwarding.
Introduction We focus on the problem of routing in a large mobile ad-hoc network that we call terminode net- work. We call nodes in a terminode network, terminodes, because they act as network nodes and terminals at the same time. Our routing solution is designed with three requirements in mind: rstly, it should scale well in a very large mobile ad-hoc network; secondly, it should cope with dynamically changing network connectivity owing to mobility; and thirdly terminodes need to be highly collaborative and redundant, but, most of all, cannot use complex algorithms or protocols. For the rst requirement, our solution is designed such that a terminode relies only on itself and a small number of other terminodes for packet forwarding. The second requirement, uncertainty in the network due to mobility, is addressed in our work by considering multipath routing as a rule, and not as an exception. We note that the target of our work is dierent from MANET[10] proposals that focus on networks consisting of up to several hundreds of nodes. Each terminode has a permanent End-system Unique location-dependent address (LDA). The LDA is simply a triplet of geographic coordinates (longitude, lati- tude, altitude) obtained, for example, by means of the Global Positioning System (GPS) or the GPS-free positioning method[5]). In this paper, we concentrate on the problem of unicast packet forwarding, assuming that the source terminode knows or can obtain the LDA of the destination. A packet sent by a source terminode contains, among other elds, the destination LDA and EUI, and possibly some source routing information, as mentioned later. Mobility management in a terminode network may be performed by a combination of the following functions. Firstly, a location tracking algorithm is assumed to exist between communicating terminodes; this allows a terminode to predict the location (LDA) of corresponding terminodes. Secondly, LDA manage- ment, which is based on the distributed location database, allows a terminode A to obtain a probable location of terminode B (LDAB ) that A is not tracking by the previous method. Mobility management is out of scope of this paper, (see for example [7],[9]). Here we assume that nodes move with the speed that corresponds to pedestrian or car speed, so that we can obtain (with the mobility manage- ment) the destination LDA with the precision of approximately one transmission range and validity of about ten seconds. We also assume that terminode network is mainly connected, although temporary partitions can occur. Even if our protocol uses geographical locations, it is independent from the physical infrastructure (i.e, it does not assume directional antennae) and from the physical underlying layer. We performed simulations with IEEE 802.11 MAC protocol, because it is a commonly used MAC protocol. We anticipate that our routing protocol would have the similar results with other MAC protocols. We further assume that multipath routing is acceptable for the transport protocol. However, with the current TCP this is not acceptable since there are problems with managing a large number of timers due to many paths. We envision either to bring enhancements to the current TCP, or to use of multiple description coding techniques. In this latter case, the source data is encoded and sent over multiple paths in order to provide better load balancing and path failure protection. We use a combination of two routing protocols: Routing (TLR) and Terminode Remote Routing (TRR). TLR is a mechanism that allows to reach destinations in the vicinity of a terminode and does not use location information for making packet forwarding decisions. In contrast, TRR is used to send data to remote destinations and uses geographic information; it is the key element for achieving scalability and reduced dependence on intermediate systems. TRR consists of the following elements: Anchored Geodesic Packet Forwarding (AGPF) is a method that allows for data to be sent to remote terminodes. AGPF is solely based on locations. AGPF sends data along the anchored path. An anchored path denes a rough shape of the path from the source to the destination and is given with a list of anchors. Anchors are points described by geographical co-ordinates and do not, in general, correspond to any terminode location. A good anchored path should avoid obstacles and terminode \deserts" from the source to the destination. Between anchors geodesic packet forwarding is performed; this is a greedy method that follows successively closer geographic hops to an anchor or the nal destination. Friend Assisted Path Discovery (FAPD) is the path discovery method used to obtain anchored paths. A terminode keeps a list of other termin- odes, that it calls friends, to which it maintains one or several good path(s). In FAPD, a terminode may contact its friends in order to nd an anchored path to the destination of interest. FAPD is based on the concept of small world Path Maintenance is a method that allows a terminode to improve acquired paths, and delete obsolete or mal-functioning paths. Multipath Routing. A terminode normally attempts to maintain several anchored paths to any single destination of interest. In a highly mobile environment, anchored paths can be broken or become congested. A path that worked well suddenly can deteriorate. As a response to such uncertainty in the network, TRR uses multipath routing. TRR is used to send data to a remote destina- tion. However, when a packet gets close to the destination, if only locations are used for making packet forwarding decisions, positional errors and inconsistent location information can result in routing errors and loops. This happens if the destination has considerably moved from the location that is known at the source. In order to cope with this problem, in our approach when close to destination, the packet forwarding method becomes TLR. TLR does not use location informa- tion. Once a packet has been forwarded with TLR, the \use TLR" bit is set within the packet header, and downstream terminodes should not use TRR again. This avoids loops due to the combination of the two routing methods. The rest of this paper is organized as follows. The following section gives a short overview of some existing mobile ad-hoc routing protocols that are relevant for our work. In Section 3 we give a complete description of TLR and TRR . This is followed by the description of the simulation results in Section 4. Finally, we give some conclusions. 2. Related Work Many routing protocols have been proposed for mobile ad-hoc networks. A recent overview can be found for instance in [15]. The existing routing protocols can be classied either as proactive or reactive. Proactive protocols attempt to maintain routes continuously, so that the route is already available when it is needed for a packet to be forwarded. In those protocols, routing tables are exchanged among neighbouring nodes each time a change occurs in the network topology. As a consequence, proactive protocols are not suitable when the mobility rate in the network is high. An attempt to overcome these limitations is to look for a route only on demand. This is the basic idea of reactive protocols such as DSR[4], TORA[11] and AODV[12]. In reactive protocols a control message is sent to discover a route to a given destination. In DSR, when a source S needs a route to a destination node D, S rst checks if some of its neighbours possesses the route in question. If this is not the case, S oods the network with a route request for the destination node. When the request reaches the destination, the destination returns a route reply to the request's originator. The reply message contains the list of all intermediate nodes from S to D. Then S uses source routing with the acquired source route to send packets to D. Several methods are proposed for limiting the propagation of requests. One of these is that nodes cache the route that they learn or overhear, so that intermediate nodes can reply on behalf of the destination if the route to the destination is known. Reactive protocols have smaller control tra-c overhead than proactive protocols. However, since a route has to be discovered before the actual transmission of the data, these protocols can have a longer delay. Further more, due to mobility, the discovered route may be unusable since some links of the route may be broken. ZRP[13] is a protocol that combines both a proactive and reactive approach. Every node proactively maintains routes to other nodes whose distance is less than a certain number of hops (its zone). Within a zone, an arbitrary proactive routing scheme can be applied. For inter-zone routing, on demand routing is used. ZRP avoids ooding the network in order to nd routes and routing protocol overhead is limited. However, the on-demand solution for inter-zone routing poses the latency problem typical to on-demand routing schemes. There are a number of proposed geographical routing protocols that use location information to reduce propagation of control messages, to reduce intermediate system functions or for making packet forwarding decisions. Geographical routing allows nodes in the net-work to be nearly stateless; the information that nodes in the network have to maintain is about their one-hop neighbours. Location Aided Routing (LAR)[19] in an optimization of DSR where the knowledge of the destination location is used to limit the propagation of route request control packets. Those packets are propagated to the geographical region around the last known destination location. LAR does not use location for data packet forwarding. With LAR, end-to-end routes are still DSR's source routes. Location Distance Routing Eect Algorithm for Mobility (DREAM)[2] is a routing protocol in which the information about location and speed of the destination is used to obtain the direction of the destination. A node that has a packet to send determines the direction of the destination. Then it forwards data to all one hop neighbours in the calculated direction of the destination. DREAM proposes how to disseminate location information in the network in the scalable way. In Geographical Routing Algorithm (GRA)[22], every node has only a partial knowledge of a net- work. It knows about its immediate neighbors and a small number of remote nodes to which it has discovered a path. When an intermediate node receives a packet to forward, it checks which of the nodes that it knows is closest to the destination. Then the packet is forwarded to the neighbour that is next hop towards the node that is closest to the destination. Each node thus forwards the packet in the similar way till the packet reaches the des- tination. If it happens that some node S does not know about any node that is closer to the destination D than itself, a route discovery method is invoked. This method nds an acyclic path from S to D and all intermediate nodes on update their routing tables with the next hop information in order to reach D. GEDIR[16], GPSR[8] and GFG[21] routing protocols are very important to this paper, and we present them in more detail. These protocols propose to use a greedy method for making packet forwarding decisions. Packet forwarding decisions are made using only information about a node's immediate neighbours and the location of destina- tion; packet is forwarded to the neighbour that is closest to destination. The GEDIR paper proves that packet forwarding is loop-free provided that location information is accurate. However, a packet may reach a node that does not have any neighbours closer than itself to the ultimate destination. This situation indicates that there is a hole in the geographic distribution of nodes. As a solution to this situation, GPSR and GFG use a planar subgraph of the wireless net- work's graph to route around the perimeter of a hole. This method is rst proposed by Bose et al. in [20]. Packet forwarding for such a packet is switched from greedy to perimeter mode. The knowledge of identities and locations of its one-hop neighbours is su-cient for a node to determine the edges of the planar subgraph. Packets that are in perimeter mode are forwared using a simple planar graph traversal. A perimeter-mode packet is forwarded on progressively closer to destination faces of the planar graph. As soon as a packet reaches the node that is closer to destination than the node that initiated perimeter mode forwarding, a packet is then forwarded in a greedy way. In a dense network, packets are normally forwarded in greedy way, and the perimeter mode is used occasionally when a packet is stuck when a node does not have a one-hop neighbour that is closer to the destination. Then a packet is for- wared in perimeter mode for only a very few (2-3) hops, before a node closer than the point of entry into perimeter mode is reached, and then greedy forwarding resumes. On sparser networks, perimeter mode tends to be used for longer sequences of hops. routing does not maintain strict source routes. Dierently from DSR, in terminode routing, there is no need for ooding of the whole network to discover the route or react when some link is broken. FAPD is a way to discover loose source paths without ooding of the network. When a path with anchors is known, AGPF is used. AGPF is a greedy method that uses locations for packet forwarding, and recovers from a link failure relaying only on terminodes' local information. If there is a hole in nodes's distribution, GPSR uses routing around the perimeter of a hole. In our ap- proach, if anchors are correctly set, AGPF avoids holes in terminodes distribution and uses perimeter method only occasionally. As it is discussed in the introduction, if only geographical locations are used for making packet forwarding decisions, this may result in looping problems due to positional errors and inconsistent location information. In our design we use TLR in order to alleviate looping problems. 3. Terminode Routing routing uses geographical information in making routing decisions. However, when the packet gets close to the destination, packet forwarding uses local routing tables that every terminode proactively maintains for its close terminodes Our routing scheme is a combination of two pro- tocols, Terminode Local Routing (TLR) and Terminode Remote Routing (TRR). TLR determines a route to destination D when D is in the local routing table (of a source or of an intermediate terminode). Otherwise, TRR uses mainly the locations of S and D to discover a path from S to D. Below we describe elements of terminode routing in more details. 3.1. Terminode Local Routing (TLR) Routing (TLR) is used by terminodes to build their local routing tables, and for forwarding packets to destinations in their vicinity. The idea of TLR was inspired by the intrazone routing protocol (IARP) in ZRP[13]. We say that terminode D is TLR-reachable for has a means to reach D with the TLR protocol. The TLR-reachable area of S includes the terminodes whose minimum distances in hops from S are at most equal to local radius. The local radius is a measure, in number of hops, of the TLR-reachable area. In the current implementation of TLR, the local radius is set to two hops. Building of local routing tables In order to build its local routing table, every terminode proactively maintains the identity (EUI) and location 1 (LDA) of the terminodes that are no more than two hops away. This is done by means of HELLO messages that every terminode periodically broadcasts at the MAC layer. A terminode announces in a HELLO message its own EUI and LDA, as well as EUIs of its immediate neighbours. Upon reception of a HELLO mes- sage, a node updates its local routing table with EUI and LDA of its one-hop and two-hop distant terminodes. For two-hop distant terminodes, a terminode also keeps in its table the next hop terminode via which a two-hop distant terminode can be reached. If a node does not hear from its neighbour for some amount of time, it removes from the routing table the entry that corresponds to the lost neighbour, as well as all entries that correspond to two-hop distant terminodes that were reachable via the lost neighbour. TLR packet forwarding TLR uses a simple distance vector routing protocol to send data to TLR- reachable destinations. The only addressing information used by TLR is the EUI of the destination. When the source, or an intermediate node, has to forward a packet to the destination that is TLR- reachable, the \use TLR" bit in the packet header is set, if not already set. If the destination is two-hop away, the packet is sent to the next hop ter- minode, as from the routing table. Otherwise, the packet is sent directly to the destination. It is also possible to use a local radius larger than two. However, this would increase the TLR overhead because of the update tra-c required for every node to maintain its TLR-reachable area. In 1 The LDA information is not used for TLR, but it is added because it is used by TRR, as explained below. addition, for a larger local radius, the slow convergence problem, typical to distance vector routing protocol, would aect TLR. 3.2. Terminode Remote Routing (TRR) Remote Routing (TRR) allows data to be sent to non TLR-reachable destinations. TRR primarily forwards packets on anchored paths. In contrast with traditional routing algorithms, an anchored path does not consist of a list of nodes to be visited to reach the destination. An anchored path is a list of xed geographical points, called anchors. In traditional paths made of lists of nodes, if nodes move far from the location they had at the time when the path was computed, the path cannot be used to reach the destination. Given that geographical points do not move, the advantage is that an anchored path is always \valid". In order to forward packets along an anchored path, TRR uses the method called Anchored Geodesic Packet Forwarding (AGPF). With AGPF the packet is sent in the direction of an anchor, thus trying to reach some terminode in proximity of this anchor. Thereon, the packet is forwarded in the direction of the next anchor on the anchored path. Anchored paths are obtained by the source with the method called Friend Assisted Path Discovery (FAPD). If the source does not have an available anchored path to the destination, it uses a default method called Geodesic Packet Forwarding (GPF). 3.2.1. Geodesic Packet Forwarding (GPF) GPF is a simple method for sending data in the direction of a geographical point. This point can be an anchor (see AGPF below) or the destination location. Unlike TLR, GPF is based solely on lo- cations. A similar method is used in GFG[21] and in GPSR [8]. Source S can use GPF to send data to remote destination D in the following way. At rst, S acquires some approximate value of the D's LDA. Then S sends packets by GPF in the greedy man- ner: the packet is sent to some neighbour X within a transmission range of S where the distance to D is the most reduced. In turn, X checks whether D is TLR-reachable. If this is not the case, X sends the packet to its neighbour that is closest to the destination. Otherwise, X uses TLR to forward the packet. In this simplest form, GPF will often not work. If there is no connectivity along the shortest line due to obstacles or a terminode desert, then the method fails. The packet may be \stuck" at some terminode that does not have a neighbour that is closer to the destination. One possible solution to this problem is to use the method of a planar graph traversal, where a packet is routed around the perimeter of the region where there are no terminodes closer to the destination (this solution is proposed in GFG and GPSR). In this way, a packet is routed until it arrives at the terminode that reduces the distance to the destination, and thereon the packet is forwarded in a greedy manner, as described above. GPF works well when there is a good connectivity along the shortest line from the source to the destination. However, this may not always work. In order to circumvent holes in terminodes distribution in a large area mobile ad-hoc network, we introduce the method called AGPF. 3.2.2. Anchored Geodesic Packet Forwarding (AGPF) The key element of the AGPF is anchored path. Anchors are computed by source nodes, using the path discovery method called FAPD, as presented below. A source terminode adds to the packet the anchored path that is used as loose source routing information. With AGPF the packet is forwarded such that it loosely follows an anchored path. The sequence of intermediate terminodes on the way to the destination depends on the actual terminode distribution. AGPF works as follows. At the source, the packet is sent in the direction of the rst anchor (AP1) on the anchored path by applying geodesic packet forwarding: the source sends data to its immediate 7neighbour that has the minor distance to AP1. When an intermediate terminode receives a packet with the anchored path, it checks whether geographically falls within its transmission range. If so, it deletes AP1 from the anchored path and sends the packet in the direction of the next anchor (AP2). This is repeated until all anchor points are deleted from the anchored path. Then the packet is sent in direction of the nal destination by using GPF as described in the previous section. If the anchors are correctly set, then there is a high probability that the packet will arrive at the destination. This is veried in our simulations (see Section 4.2). A good anchored path directs packets along regions with good terminode connectiv- ity. Occasionally, when there is a hole in terminode distribution between two anchors, routing around the perimeter of a hole is used. Figure 1 presents an example of AGPF. 3.2.3. How to expedite termination of TRR With TRR the packet gets closer to the desti- nation. TRR is used until some intermediate node nds that the destination can be reached by means of TLR. In this case the \use TLR" bit in the packet header is set. Thereon, TLR can be only used for packet forwarding. Termination of TRR is important in order to avoid loops that may happen if packet forwarding from TLR reverts to TRR. However, if accuracy of location management is low or if the packet has been delayed due to congestion or bad paths, it may happen that the condition to set \use TLR" bit is never met and a packet may start looping. Our approach is to discover such loops and to drop looping packets. We recognize a possible loop when a terminode nds that the destination location written in the packet header is within its transmission range, but the destination is not TLR-reachable. In order to address this case, a terminode X performs the following action: if than the transmission range of X, and D is not TLR-reachable for X, X sets the TTL eld within a packet header to the Figure 1. The gure presents how AGPF works when a terminode with EUIS has some data to send to a terminode with EUID , and there is no connectivity along the shortest line from S to D. S has a path to D given by a list of geographical locations called anchors: fAP1, AP2g. First, geodesic packet forwarding in the direction of AP1 is used. After some hops the packet arrives at a terminode A that nds that AP1 falls within its transmission range. At A, the packet is forwarded by using geodesic packet forwarding in the direction of AP2. Second, when the packet comes to B, that is close to AP2, it starts sending the packet towards D. Last, when the packet comes to C it nds that D is TLR-reachable and forwards the packet to D by means of TLR. value equal to min(term loop; TTL). term loop is a xed value, which indicates that a loop due to destination location inaccuracy is always limited to term loop 2 hops. 3.2.4. Path Discovery Friend Assisted Path Discovery (FAPD) is a method to obtain anchored paths. It is based on the concept of small world graphs[18]. Small world graphs are very large graphs that tend to be sparse, clustered, and have a small diameter. The small-world phenomenon was inaugurated as an area of experimental study in social science through the work of Stanley Milgram in the 60's. These experiments have shown that the acquintanceship graph connecting the entire human population has a diameter of six or less; small world phenomenon al- In our current implementation of TRR term loop is equal to 3. lows people to speak of the \six-degrees of separa- tion". We view a terminode network as a large graph, with edges representing the \friend relationship". B is a friend of A if (1) A thinks that it has a good path to B and (2) A decides to keep B in its list of friends. A may have a good path to B because A can reach B by applying TLR, or by geodesic packet forwarding, or because A managed to maintain one or several anchored paths to B that work well. Every terminode has a knowledge of a number of close terminodes in its TLR-reachable re- gion; this makes a graph highly clustered. In ad- dition, every terminode has a number of remote friends to which it maintains a good path(s). We conjecture that this graph has the properties of a small world graph. In a small world graph, roughly speaking, any two vertices are likely to be connected through a short sequence of intermediate vertices. This means that any two terminodes are likely to be connected with a small number of intermediate friends. With FADP, each terminode keeps the list of its friends with the following information: location of friend, path(s) to friend and potentially some information about the quality of path(s). FAPD is composed by two elements: Friends Assisted Path Discovery Protocol (FAPDP) and Friends Management (FM). Friends Assisted Path Discovery Protocol FAPDP is a distributed method to nd an anchored path between two terminodes in a terminode network. When a source S wants to discover a path to destination D, it requests assistance from some friend, let's say F . If F is in condition to col- laborate, it tries to provide S with some path to D (it can have it already or try to nd it, perhaps with the help of its own friends). Figures 2 and 3 present FAPDP in pseudocode at the source and at an intermediate friend. When source S, which has some data to send to if (S has a friend F1 where dist(F1,D)<dist(S,D) ) {S sets "F" bit in the packet header; send a packet to F1;} else if (S has a friend F3 such that dist(S, F3) < max_dist ) {S sets "F" bit in the packet header; send the packet to F3;} else apply geodesic packet forwarding to D; Figure 2. Friend Assisted Path Discovery Protocol at the source F1 is intended receiver of a path discovery packet needs a path to D if path reply with fapd_anchored_path to S;} else if (F1 has a path to D) append this path in fapd_anchored_path and send the packet to D; else if (tabu_index > 0 ) //packet in tabu mode { if ( F1 has a friend F2 where dist(F2, D) < min_dist) {tabu_index=0; send the packet to F2} else if (tabu_index < 2 and F1 has a friend F3 such that dist(F1, F3) < max_dist ) { tabu_index++; send a packet to F3} else // tabu_index reached the maximum value {send a packet to D by geodesic packet forwarding} else //packet not in tabu mode { if (F1 has a friend F2 where dist(F2,D)<dist(F1,D) ) send a packet to F2; else if (F1 has a friend F3 such that dist(F1, F3) < max_dist ) {tabu_index=1; min_dist=dist(F1,D); send a packet to F3} else apply geodesic packet forwarding to D; Figure 3. Friend Assisted Path Discovery Protocol at the intermediate friend and at the destination D, has some friends that are closer to D than S itself, it selects friend F1 that is closest to D, and starts FAPDP with F 1. S sends the data packet to F1 according to the existing path that S maintains to F1 because F1 is a friend of S. S sets, within the data packet header, \F" bit 3 . This denotes that the corresponding packet is a path discovery packet. When F1 receives this packet it recognizes the packet as a request of a path to D. The fapd anchored path eld inside the path discovery packet progressively contains anchor points from S to D. If S has an anchored path to F 1, S simply put anchors of this path in fapd anchored path eld (S sends data to F1 with AGPF). Otherwise, S leaves this eld empty (in this case S sends to F1 with geodesic packet forwarding). Upon reception of the path discovery packet, F1 puts its geographical location inside fapd anchored path eld as one anchor. If F1 has an anchored path to D, F1 appends this path into fapd anchored path eld and 3 \F" bit is not reset before reaching D sends the packet to D by AGPF. If F1 does not have a path to D, it recursively uses FAPDP. In this case, F1 checks if it has a friend F2 closer to D, and then it performs the same steps as S. This is repeated until the packet is received by some intermediate node that nds D to be TLR-reachable and it forwards the packet to D by TLR. However, there are situations where the source or an intermediate friend does not have a friend closer to the destination. In some topologies with obstacles, at some point, going in the direction opposite from the destination may be the only way to reach the destination. FAPDP permits that some terminode T (the source or an intermediate friend) sends a path discovery packet to a friend even though the packet is not getting closer to the destination. However such a friend must not be distant from T more than distance dist 4 . Here is where \tabu mode" mechanism of FAPDP starts. With the tabu mode mechanism, the packet can be sent in direction opposite to D for a limited number of times. Tabu mode is denoted at T by setting the tabu index eld inside the packet to 1 (default value of tabu index is 0). Tabu mode mechanism uses a eld called min dist, where the terminode that started tabu mode puts its distance to the destination. When an intermediate friend F1 receives the path discovery packet, which is in tabu mode, it rst checks if it has a friend whose distance to D is smaller than min dist. If this is the case, the packet is sent to such a friend, and tabu index is reset to 0. Other- wise, F1 may forward the packet to its friend F2 whose distance to D is more than min dist and F2 increments tabu index. In FAPDP, the number of times that the packet is forwarded to a friend that is further from D than min dist is limited to two (i.e, the value of tabu index must not be larger than two). Tabu mode mechanism stops either because a friend that is distant from D less than min dist is found, or because tabu index is equal 4 we use max dist equal to ve times the transmission range of a terminode to 2. In the second case the packet is forwarded directly to D by geodesic packet forwarding. Finally, when D receives the packet with \F" bit equal to one, D must send back to S a \path reply" control packet with the acquired anchored path from S to D. This packet is sent to S by reverting the anchored path and applying AGPF. Once S receives from D a packet with the anchored path, S stores this path in its route cache. If S does not receive a anchored path within some time, or if S wants more paths to D, S starts FAPDP with some other friend. FAPDP is illustrated with two examples. The rst example, presented in Figure 4 shows the case where the path from S to D is found by using two intermediate friends. The second example, in Figure 5 illustrates the tabu mode of FAPDP. Friends Management Friends Management (FM) is a set of procedures for selecting, monitoring and evaluating friends. For each node A, FM maintains a (xed-size) set of nodes: the set of friends. The set of friends contains the nodes that are contacted with FAPDP for discovering paths. These nodes are periodically evaluated in order to assure their availability and their validity as friends. For that reason, the Friends Monitoring component of FM keeps under control, for a node A, a set of parameters for each friend of A. These parameters ranges from technical characteristics (as, for example, the active time, the average distance, etc.) to more \social" information (as being at disposal for supporting FAPDP or forwarding packets). Based on these parameters, the Friends Evaluation component periodically evaluates whether it is benecial to keep a node in the friend set or it is better to discard it. It is the responsibility of Friend Selection to identify a set of nodes that are suitable to act as friends. A node B can be selected to become a friend of a node A because there are frequent communications between A and B, because B is often in a strategic position for A's communications, be- Figure 4. The gure presents how FAPDP works when source S, has a friend F1 that is closer to D than S. sends data packet to F1 and sets \F" bit in the packet header in order to denote that this is a \path discovery packet". Upon reception of the path discovery packet, F1 puts its geographical location inside the fapd anchored path eld of the path discovery packet as one anchor. In this example F1 does not have path to D, but has a friend F2 whose distance to D is smaller than the distance from F1 to D. F1 sends path discovery packet to F2. Once F2 receives the packet, it nds out that D is TLR-reachable and F2 forwards the packet to D by TLR. When D receives the packet with set \F" bit, it should send back to S a \path reply" control packet with the acquired anchored path from S to D. Assuming that the path from S to F1 and from F1 to F2 does not contain any anchors, the anchored path from S to D is thus a list of anchors (LDAF1 ; LDAF2 ). cause B is supposed to be helpful for discovering good paths or for some other reasons. Once a potential friend has been identied, it will be evaluated by the Friend Evaluation com- ponent, and included in the set of friends if there is still space in set or if the node's evaluation is better than the one of some other node of the set. FM is critical in the initial phase (bootstrapping). When a node bootstraps, it does not have any information of (possible) friends. In this phase, the Friend Selection component will use the HELLO messages for identifying possible friends (that is, in the initial phase, will be limited to 1-2 hops neigh- bours). At run-time, these initial friends will dis- Figure 5. The gure presents how FAPDP works when source S does not have a friend that is closer to D than itself. S contacts its friend F1 that is farther from D in geometrical distance than S is, but such that dist. As in the previous example, S sends data packet to F1 with \F" bit set. In addition S sets the tabu index eld to 1 and thus starts the tabu mode of FAPDP. S puts dist(S; D) within min dist eld. Upon reception of the path discovery packet, F1 nds out that it does not have a friend whose distance to D is smaller than min dist. F1 forwards the path discovery packet to its friend F2 where dist(F and sets tabu index to 2. Upon reception of the packet, F2 checks that tabu index is equal to its maximum value equal to 2, and F2 cannot forward the packet to its friend that does not reduce the distance min dist. In our exam- ple, F2 has a friend F3 whose distance to D is smaller than min dist and forwards the packet to it. At F3, tabu index is reset to 0. This means that FAPDP is not longer in tabu mode. From F3 packet is forwarded to its friend F4 and from there to D by using the TLR protocol. The anchored path from S to D is thus a list of anchors appear very likely, for being substituted by more valid friends as described above. 3.2.5. Path Maintenance Every terminode normally attempts to maintain multiple anchored paths to the destinations that it communicates with. Multipath routing is a way to cope with uncertainty in a terminode network; the paths that a source has acquired by FAPDP can deteriorate due to mobility and packets can be lost. We advocate that the source data is encoded and sent over multiple independent paths in order to provide better load balancing and path failure protection. Diversity of paths is essential for taking advantage of multipath routing [14]. Path maintenance consists of three main func- tions: independent path selection, path simplica- tion, path monitoring and deletion and congestion control. independent path selection A terminode analyzes all acquired paths to a destination. Then its selects a set of independent paths. They are paths that are as diverse as possible in geographical points (anchors) that they consist of. path simplication One method consists in approximating an existing path with a path with fewer anchors. Such an approximation yields a candidate path, which may be better or worse than the old one. We use a heuristic based on curve tting. path monitoring and deletion A terminode constantly monitors existing paths in order to collect necessary information to give the value to the path. The value of the path is given in terms of congestion feedback information such as packet loss and delay. Other factors like ro- bustness, stability and security are also relevant to the value of a path. This allows a terminode to improve paths, and delete mal-functioning paths or obsolete paths (e.g, the path that corresponds to two terminodes that do not communicate any more). congestion control The value of the path given in terms of congestion feedback information is used for a terminode to decide how to split the tra-c among several paths that exist to the des- tination. A terminode gives more load to paths that give least congestion feedback information. 4. Performance Evaluation We simulated the terminode routing protocol in GloMoSim[17]. GloMoSim is a scalable simulation environment for wireless network systems. It is based on the parallel, discrete-event simulation language PARSEC[1]. The IEEE 802.11 Medium Access Control(MAC) protocol is used; it implements the Distributed Co-ordination Function (DCF)[6]. In all simulations, radio range is the same for every terminode, and is equal to 250 meters. The channel capacity is 2Mbits/sec. In order to build its local routing table, in our implementation, every terminode sends a HELLO message every 1 second. Each entry in the neighbour table expires after 2 seconds if it is not up- dated. To avoid synchronization of HELLO mes- sages, a terminode jitters each HELLO message transmission by 50% of the period of a HELLO message. 802.11 MAC layer noties when a unicast packet exceeds the maximum number of retransmissions and the acknowledgement has not arrived. This means that the intended neighbour has left the sender's transmission range and that the entry that corresponds to that neighbour is invalid and can be removed from a sender's neighbour table. In our implementation such a packet is sent back to the routing layer where a new neighbour to send a packet is chosen. In this section we evaluate both components of the terminode routing, namely TLR and TRR. 4.1. Evaluation of TLR To assess the relevance of TLR, we analyze a performance of terminode routing when TLR is used and when TLR is not used. We performed simulations in a scenario where geodesic packet forwarding towards the destination works well, and there is no need for the AGPF technique. We wanted to evaluate solely the performance of geodesic packet forwarding with TLR, against the case when TLR is not used. Geodesic packet forwarding (GPF) uses destination location for making packet forwarding de- cisions. Therefore, it is important that the source knows this information accurately enough. We recall that in our proposal, once the communication has begun, terminodes are assumed to use location tracking to exchange their current locations. This is enough to assure valid location information in several situations. However, there are situations where location tracking is not able to give regular periodic location information (e.g, when GPS is temporary unavailable or when location tracking packets are lost). In these cases, we expect TLR to perform better than when TLR is not used. The rst simulation case study is when packets are forwarded towards the destination based only on the destination loca- tion. Every intermediate node forwards the packet to its neighbour, which reduces the distance to the nal destination. This is performed until the destination is reached. In the second simulation study case, TLR is used. Every terminode keeps a list of its TLR- reachable destinations. Similar to the rst case, the packet is forwarded to the neighbour closer to the destination. But, when some intermediate node nds that the destination is TLR-reachable, it uses the TLR to send the packet to the destina- tion. Note that TLR is used in two-hop neighborhood and does not need additional routing overhead compared to the case when TLR is not used. The only requirement for TLR is that all terminodes keep in their routing tables information not only about immediate neighbours, but also about their two-hop neighbours. We assume that the source can not know an exact destination location all the times. In our sim- ulations, the source learns a destination location and uses this information for the time that we call location information lifetime. After this time, the source again acquires an exact destination location and uses it for another location information life-time interval. In our simulations we used network of 600 ter- minodes. The simulation area is a rectangle of the size 5400m X 1000m. The simulated network is quite dense; in this case we veried that geodesic packet forwarding is working well. It is very rare situation where packets are stuck at some node be- Location information lifetime Packet delivery fraction without TLR with TLR Figure 6. Demonstration of TLR utility cause it has no closer neighbour to the destination. We simulated ows. Each CBR ow sends at 2Kbps and uses 256-byte packets. Terminodes are moving according to to the \random way- point"[4] model. In the \random waypoint" mobility model, a node chooses one random destination in the simulation area. Then it moves to that destination at a speed distributed uniformly between 0 and some maximum speed. Upon reaching its destination, the node pauses for the pause time, selects another random destination inside the simulation area, and proceeds as previously described. In our simulations a maximum speed is 20m/s and the pause time is 10s. In the performed simula- tions, we veried that the average end-to-end of data packets in less than 1s. The two study cases (geodesic packet forwarding with TLR and without TLR) are evaluated according to packet delivery fraction. This is the ratio of the data packets delivered to the destinations to data packets generated by the CBR sources. We evaluate this metric under various location information lifetimes. We simulated ve dierent randomly generated motion patterns. Figure 6 presents an average of packet delivery fraction for ve simulation runs. This g- ure shows that for smaller location information life-times (less than 20s), the packet delivery fraction is similar with TLR and without TLR. However, for higher location information lifetimes (lower precision of location information) routing with TLR gives better delivery fraction than without TLR. Therefore, we conclude that when using TLR in addition to geodesic packet forwarding, routing is more robust in the case of positional errors and inconsistent location information. Moreover, TLR can be the only possible packet forwarding method in small ad-hoc networks, where it could be di-cult to obtain location infor- mation. In this case all nodes are TLR-reachable, and geographical positioning is not needed. 4.2. Evaluation of TRR We analyze AGPF by comparing a performance of TRR when AGPF is used and when AGPF is not used. Before presenting the simulation and simulations results, we introduce the new mobility model called \restricted random waypoint" that is used in evaluation of TRR. Mobility Model In the most recent papers about mobile ad-hoc networks simulations, nodes in the simulation move according to the \random waypoint" model as described in Section 4.1. We nd this model unrealistic for a wide area mobile ad-hoc network such as a terminode net- work. In this network, terminodes are small personal devices that are distributed geographically within a very large area. It is less probable that for each movement a terminode selects a random destination within a very large geographical area. On the contrary, the random destination is selected within a small area for a number of movements, and then a movement is made over a long distance. This better represents the fact that most people move for a certain period within one area, and then they move away to another distant area. We have implemented a new mobility model that we call \restricted random waypoint". This model is town town 2 town 3 Figure 7. Model of the simulation area with four towns closer to a real-life situation for a wide-area mobile ad-hoc network than the random waypoint model. For the restricted random waypoint mobility model, we introduce the topology based on towns and highways. Towns are areas that are connected with highways. Inside town areas, terminodes move with the random waypoint mobility model. After a certain number of movements in the same town, a terminode moves to another town. Terminodes that are moving between the town areas, simulate highways between towns. The model of the simulated area that consists of four towns is presented in Figure 7. In our simulations, we dene inside the conguration le, the pairs of towns that are connected by highways. For example in Figure 7 those pairs are (town 0, town 1), (town 0, town 2), and (town 1, town 3). This information is used by terminodes when they move from one town to another. We distinguish two types of restricted mobility that represent the \ordinary terminode" and the \commuter terminode". At the beginning of the simulation, terminodes are placed at randomly chosen locations inside one of four towns. An \ordinary terminode" begins the simulation by selecting at random one destination inside the town where it is placed. Then it moves to that destination at the speed distributed uniformly between 0 and some maximum speed. Upon reaching that destination, the ordinary terminode pauses for the pause time, selects another destination within the same town, and proceeds as previously described. Thus, the ordinary terminode's movement inside a town is the random waypoint mobility model. It repeats such movements for a number of times set by the stay in town parameter. Then a terminode selects at random a destination within a new town and move there (the new town is randomly chosen from a list of towns that are connected with the current town by a highway). Once it reaches the new location, a terminode applies inside the new town the random waypoint mobility model for another stay in town time. There are also a number of terminodes that frequently commute from one town to another. Those termnodes are called \commuters" and they insure the connectivity between towns. The commuter's movement model is the restricted random waypoint where stay in town parameter is equal to one. A commuter selects a random destination within one town area and moves to that destination with a speed distributed uniformly between some minimum speed and some maximum speed. Once this destination is reached, a commuter pauses for a pause time that is smaller than ordinary termin- odes' pause time. Then it selects at random another town (such that is connected with the current town) and the random destination inside the chosen town, and moves to this destination. It pauses in the new town for a small interval of time and then again moves to another town. An example of a network that comes out when terminodes are moving according to the restricted random waypoint mobility model is presented in Figure 7. In such a network, not all town areas are connected with a highway (e.g, town 2 and town 3). Scenario Characteristics A source terminode normally tries to acquire several anchored paths to the destination of interest by means of FAPD. In our simulations, we do not implement FAPD. In order to obtain anchored paths without using FAPD, in our simulation model based on \towns and highways", we assume that a high level geographical view of the network is available at every terminode. This means that each terminode has a knowledge of a map of towns. A map denes town areas and existence of highways between towns. Thus, for example, the map of towns presented in Figure 7 denes those towns that are directly connected by highways as well as towns' areas. In our simulations, a town area is a square around the town center with the given width around the town center. When source S has some data to send to destination D, S rst determines the \destination town" (DT ). This is the town in which area D's location falls. If D is not inside any town area, then the destination town is the town whose center is closest to D's location. Similarly, S determines the \source town" (ST ), the town where S is situated, or the closest town to S if S is on the highway. Once S determines DT and ST , S contacts the map of towns to check if DT and ST are the same, or they are directly connected with a highway. If so, then geodesic packet forwarding (GPF) towards D has a good chance of working. Then, S does not add to the packet an achored path and S sends a packet using GPF. Otherwise, if there is no highway from ST to DT , S nds out from the map, those town areas that a packet has to pass in order to reach D. Then S adds to the packet the anchor path. This anchored path is given by a list of centers of towns that the packet has to pass. Then S starts AGPF in order to deliver the packet. For example, it S is in the area of town 2 and D is the area of town 3, then anchors on the anchored path are centers of town 0 and town 1. In this case, AGPF works as follows: the packet is rst forwarded in the direction of the rst anchor (the center of town 0). Once the packet arrives at some terminode that nds that the rst anchor falls within its transmission range, the packet is then forwarded in the direction of the second anchor (the center of town 1). As before, when a packet comes to a terminode that is close to the second anchor, the packet is then forwarded in the direction of D's location. Assuming that there are terminodes to ensure network connectivity in town areas and on high- ways, the packet is forwarded with AGPF mostly in the greedy way: packets are forwarded to terminodes that are always progressively closer to an anchor point or the destination. If however, occasionally there are regions of the network where such a greedy path does not exist (i.e, it is required that the packet moves temporarily farther away from an anchor or destination), we use the approach proposed in GPSR[8]. With this approach a packet is forwarded in perimeter mode: a packet traverses successively closer faces on a planar subgraph of the full network connectivity graph, until reaching a node closer to an anchor or the destination and then greedy forwarding resumes. Given the simplicity of the network topology based on four towns, in our simulations we do not use multipath routing. We only have one path from source to destination. Thus, in our simulations, TRR uses GPF towards D if ST and DT are the same or they are connected by a highway. Other- wise, AGPF is used. In order to assess the relevance of AGPF, we use simulations to evaluate packet forwarding in two cases: the rst case corresponds to when both GPF and AGPF are used, while the second case is when only GPF towards the destination is used. The only dierence between GPSR and GPF is as fol- lows: GPSR uses the destination location for making packet forwarding decisions for the whole way until the packet arrives at the destination; with GPF, an intermediate node switches to TLR if the destination is TLR-reachable. We illustrate packet forwarding when both GPF and AGPF are used, and when only GPF is used in the example presented in Figure 8. Here, S is in town 0 and D is town 3. In the case of AGPF, S sets the anchored path to consist of one an- Figure 8. Figure presents the path of the packet from source S to destination D in case of two routing protocols: GPF that does not use anchors and AGPF with anchors. AGPF gives shorter path than GPF. chor: center of town 1. AGPF forwards the packet along the path that goes to town 1. Once the packet is close to the center of town 1, the packet is forwarded towards D by using GPF. Packet forwarding is almost always in greedy mode; however, there are cases where perimeter mode is used for a very few (2-3) hops, before greedy forwarding resumes Geodesic packet forwarding (GPF) without anchors uses a much longer path across town 2. Figure 8 illustrates that the packet is rst forwarded in the greedy mode toward D until it reaches terminode starts because P1 does not have neighbour closer than itself to D. The packet is thus forwarded in perimeter mode until greedy mode resumes at node G1 (G1 that is closer to D than P 1). At P 2, packet forwarding starts again perimeter mode. In this mode, a packet is forwarded from town 2 back to town 0, and from there through town 1 and town 3. Fi- nally, when the packet arrives at G2, which is closer to D then P2 (where perimeter mode is started), greedy mode resumes until the packet is received by D. Figure 8 clearly illustrates the case where usage of anchors gives shorter paths than when anchors are not used. Simulation Results In order to evaluate TRR, we conducted simulations of 500 terminodes forming an ad-hoc network. The size of the simulated area is 3000m x 2500m. Terminodes move between 4 towns inside the simulation area (see Figure 7). Centers of four towns have coordinates: (550 m, 550 m), (2500 m, 500 m), (1000 m, 2000 m) and (2500 m, 2000 m) respectively. The town area is a square around the town center with the width of 500 m from the town center. There are 200 ordinary terminodes and 300 commuters. The mobility model is the restricted waypoint mobility model. An ordinary terminode begins its journey from a random location inside the random town. As described above, it moves stay in town times inside the same town and then selects another random town to move. For each movement, a terminode takes a random speed that is uniformly distributed between 0-20m/s; before each move- ment, a terminode pauses for some pause time. We ran simulations with dierent pause times and dierent stay in town parameter of ordinary ter- minode. These parameters dene dierent degrees of ordinary terminode mobility. A longer pause time means that ordinary terminodes are less mo- bile. For a xed pause time, a larger stay in town means that a terminode is staying longer within a geographical region that corresponds to a single town. We consider dierent mobility rates of ordinary terminodes because this is set of nodes where all tra-c sources and destinations come from. In our simulations, commuters are moving faster than ordinary terminodes. For their movements they take a random speed that is uniformly distributed between 10-20m/s. The pause time for commuter terminodes is equal to 1 second; once they reach a town, another random destination inside a dier- ent town is chosen for the subsequent movement. We checked by simulations that if terminodes move following restricted the random waypoint mobility Packet delivery fraction sources, stay in town=10 sources, stay in town =10 95% confidence interval Figure 9. Packet Delivery Fraction with 40 sources; stay in town parameter is 10 model, there is good network connectivity (Figure 7). Tra-c sources are continuous bit rate (CBR). The source-destination pairs are spread randomly over the network. All CBR sources send at 2kbps, and uses 256-byte packets. All communication patterns are peer-to-peer, and CBR connections are started at times uniformly distributed between 400 and 500 seconds, and they last until the end of sim- ulation. All simulations last for 1200 seconds. All source destination pairs are chosen from the group of ordinary terminodes. We carried out a performance study of two routing protocols for the network in Figure 7. The rst protocol is TRR with both elements: geodesic packet forwarding (GPF) and anchored geodesic packet forwarding (AGPF). The second protocol uses only GPF. Below we present the simulation results for packet delivery fraction. Each data point presents an average of at least ve simulations with identical tra-c models, but dierent randomly generated movements patterns. We evaluated the two protocols by varying mobility and tra-c load lev- els. In order to examine the performance of the routing protocol under dierent degrees of conges- tion, we varied the number of CBR sources in the network. The rst set of experiments (Figure shows packet delivery fraction when there are 40 CBR Pause time Packet Delivery Fraction AGPF&GPF, 50 sources, stay in town =10times GPF, 50 sources, stay in town =10times 95% confidence interval Figure 10. Packet Delivery Fraction with 50 sources; stay in town parameter is 10 Packet Delivery Fraction AGPF&GPF, 50 sources, stay in town =2 times GPF, 50 sources, stay in town =2 times 95% confidence interval Figure 11. Packet Delivery Fraction with 50 sources; stay in town parameter is 2 sources. The stay in town parameter is set to 10 for ordinary terminodes (CBR sources and sinks). Dierent degrees of mobility are obtained for dier- ent pause times of ordinary terminodes. For higher pause times, since stay in town is high, ordinary terminodes for most of the simulation time move inside the same town area. For smaller pause times they perform moving to dierent town areas more frequently. Figure 9 shows that the combination of GPF and AGPF delivers about 20 percent more packets compared to the case where only GPF is used. This result is explained as follows. GPF can give complex and long paths for those source-destination pairs that are situated in towns not connected with a highway (Figure 8). For those packets there is a higher probability to be dropped. Keeping in mind that in this experiment CBR sources and destinations do not frequently change town areas, there are several ows where GPF loses most of the packets, while there AGPF have a higher success. We have also observed end-to-end delays for the two routing protocols. It includes all possible de- lays: queuing at the interface queue, retransmission delays at the MAC, and the propagation and transfer times. However, we found unfair to compare average delay of packets delivered by GPF to the ones delivered by AGPF. The reason is that GPF has lower packet delivery fraction than AGPF and then average delay counts only for delivered packets. We observed that with GPF a large number of the packets that take long paths are dropped, and that most of the packets that are received at the destination experienced short paths, with consequent short delays. A packet delivery fraction of the second set of experiments is presented in Figure 10. It diers from the previous example in that the number of CBR sources in the network is increased to 50. We observe that the combination of AGPF and GPF again delivers more packets than when only GPF is used. However, AGPF decreases its delivery fraction compared to the previous case when there are sources. Indeed in this experiment, we observed an increased level of congestion in the region of the network between town 0 and town 1. Since AGPF directs most of its anchored paths across this region, that result in congestion and increased packet drops. GPF performs as in the previous experiment with smaller number of sources. We explain this on the example in Figure 8. Unlike AGPF, GPF does not directly forward the packet from S in direction of town 1. Firstly, GPF forwards it to town 2 and than back to town 0 where the packet is forwarded in direction of town 1. Provided that the packet is not lost during the journey between towns 0 and 2, it contributes in the congestion along the highway between town 0 and town 1. However, in our simulations we have observed that there are many packets that are lost before taking this highway. Thus this explains why GPF is less susceptible to the increased number of CBR sources than AGPF. Here is where multipath routing for AGPF would be benecial. In our simple network topology based on four towns, AGPF uses only one anchored path to the destination. If, however, there were several paths over which packets can be sent, this would result in load balancing. It is left for future work to investigate mechanisms for choosing routes in the network so that data tra-c is more evenly distributed in the network. The last set of experiments is presented in Figure 11. Here there are 50 CBR sources, but ordinary terminodes are moving more frequently from one town area to another. Here stay in town parameter is set to 2. We observe that GPF delivers more packets than in the previous two experiments. This can be explained: with increased mobility, those source-destination pairs for which GPF gives a small fraction delivery in the previous two cases can move to towns where GPF gives a better path. In this way bad situations, where GPF gives long complex paths, do not last for the whole duration of the simulation. This is especially true for lower pause times. For higher pause times, again we observe GPF decreases in the packet delivery fraction We conclude that AGPF results in higher packet delivery fraction than GPF in all our experiments. However, we observe that the improvement of AGPF over GPF is more important when nodes stay in single town areas or close to these areas for most of the simulation time. Keeping in mind that AGPF is intended for large area mobile ad-hoc net- works, we believe that this assumption would there be satised. We performed our simulations for a relatively small simulation area and small number of nodes. We believe that within a larger area, the benets of good anchored paths over complex, long GPF paths will be more evident. Anchors dene a rough shape of a path from the source to the des- tination. The source should monitor all anchored paths it is using, and react if the value of a path is deteriorated. Finally, we discuss about routing overhead of the terminodes routing. Routing overhead is the total number of routing packets transmitted during the simulation. Terminodes routing generates two types of protocol packets. TLR uses HELLO mes- sages, whereas TRR uses control messages that are needed for FAPD. Every terminode proactively generates HELLO messages every second and those messages are received but not forwarded by its neighbours. Overhead due to HELLO messages is independent of the mobility rate of terminodes and the number of tra-c ows. As the size of the network increases, the network-wide count of HELLO messages in- creases. However, at a constant terminode density, the size of the network does not have an eect on TLR overhead per node, since HELLO messages are not propagated beyond a single hop. One possible optimization to reduce the HELLO message overhead is that nodes that have some data to forward defer the sending of HELLO mes- sages. Then a sender piggybacks in every data packet the information it would send via a HELLO message. This is possible when the network interface is used in a promiscuous mode, and a node receives all packets from all terminodes within its transmission range. This optimization has not been implemented so far. for our simulations we have not implemented FAPD, we have not evaluated FAPD over- head. This is task left for further work. 5. Conclusions We focused on the problem of routing in a wide area mobile ad-hoc network called work. Routing in this network is designed with the following objectives. First, it should scale well in terms of the number of nodes and geographical cov- erage. Second, routing should have scalable mechanisms that cope with load balancing and the dynamicity in the network due to mobility. Our routing scheme is a combination of two protocols called Routing (TLR) and Terminode Remote Routing (TRR). TRR is activated when the destination D is remote and uses the location of the destination obtained either via location management or by location tracking. TLR acts when the packet gets close to the destination and uses routing tables built with hello messages. The use of TRR results in a scalable solution that reduces dependence on the intermediate systems, while TLR allows us to reduce problems of loops due to location inaccuracy. Anchor Geodesic Packet Forwarding (AGPF) is a component of TRR that provides paths when there are holes in terminodes distribution and the source can not reach the destination over the direct geodesic path. We performed simulations to assess relevance and performance improvements when TLR and AGPF are used. In our simulations, we introduced the topology based on four towns where nodes move between towns according to the mobility model that we called \restricted random waypoint". Our simulation results demonstrate improvements obtained with TLR and AGPF over GPSR that uses geographical information for packet forwarding. --R PARSEC: A Parallel Simulation Environment for Complex Systems. A performance comparison of multi-hop wireless ad hoc network routing protocols IEEE Computer Society LAN MAN Standards Com- mittee Towards Mobile Ad-Hoc WANs: Terminodes GPSR: Greedy Perimeter Stateless Routing for Wireless Networks. A Highly Adaptive Distributed Routing Algorithm for Mobile Wireless Net- works Determining the Optimal Con On the impact of alternate path routing for load balancing in mobile ad hoc networks. A Review of Current Routing Protocols for Ad-Hoc Mobile Wireless Networks A Scalable Network Simulation Environment. In Small Worlds Routing with guaranteed delivery in ad hoc wireless networks Workshop on Discrete Algorithms and Methods for Mobile Computing and Communications (DialM'99) Internal node and shortcut based routing with guaranteed delivery in wireless networks Geographical Routing Using Partial Information for Wireless Ad Hoc Networks --TR --CTR Christian Lochert , Martin Mauve , Holger Fler , Hannes Hartenstein, Geographic routing in city scenarios, ACM SIGMOBILE Mobile Computing and Communications Review, v.9 n.1, January 2005 Karim Seada , Ahmed Helmy , Ramesh Govindan, Modeling and analyzing the correctness of geographic face routing under realistic conditions, Ad Hoc Networks, v.5 n.6, p.855-871, August, 2007 Wook Choi , Sajal K. Das, Design and performance analysis of a proxy-based indirect routing scheme in ad hoc wireless networks, Mobile Networks and Applications, v.8 n.5, p.499-515, October Silvia Giordano , Ivan Stojmenovic, Position-based ad hoc routes in ad hoc networks, The handbook of ad hoc wireless networks, CRC Press, Inc., Boca Raton, FL, Niklas Carlsson , Derek L. Eager, Non-Euclidian geographic routing in wireless networks, Ad Hoc Networks, v.5 n.7, p.1173-1193, September, 2007 Xiaojing Xiang , Xin Wang , Zehua Zhou, An Efficient Geographic Multicast Protocol for Mobile Ad Hoc Networks, Proceedings of the 2006 International Symposium on on World of Wireless, Mobile and Multimedia Networks, p.73-82, June 26-29, 2006 Ljubica Blazevic , Jean-Yves Le Boudec , Silvia Giordano, A Location-Based Routing Method for Mobile Ad Hoc Networks, IEEE Transactions on Mobile Computing, v.4 n.2, p.97-110, March 2005 Laura Galluccio , Giacomo Morabito , Sergio Palazzo, Spontaneous group management in mobile Ad Hoc networks, Wireless Networks, v.10 n.4, p.423-438, July 2004 Badri Nath , Drago Niculescu, Routing on a curve, ACM SIGCOMM Computer Communication Review, v.33 n.1, p.155-160, January Seungjoon Lee , Bobby Bhattacharjee , Suman Banerjee, Efficient geographic routing in multihop wireless networks, Proceedings of the 6th ACM international symposium on Mobile ad hoc networking and computing, May 25-27, 2005, Urbana-Champaign, IL, USA Anchor-based Routing Protocol for Mobile Ad Hoc Networks, Wireless Personal Communications: An International Journal, v.42 n.2, p.277-300, July 2007 Dragos Niculescu , Badri Nath, Trajectory based forwarding and its applications, Proceedings of the 9th annual international conference on Mobile computing and networking, September 14-19, 2003, San Diego, CA, USA
self organized routing;mobile;wide-area network;ad hoc
593175
A Controlled Skip Parser.
Real-world natural language sentences are often long and complex, and contain unexpected grammatical constructions. They even include noise and ungrammaticality. This paper describes the Controlled Skip Parser, a program that parses such real-world sentences by skipping some of the words in the sentence. The new feature of this parser is that it controls its behavior by finding out which words to skip, without using domain-specific knowledge. The parser is a priority-based chart parser. By assigning appropriate priority levels to the constituents in the chart, the parsers behavior is controlled. Statistical information is used for assigning priority levels. The statistical information (n-grams) can be thought of as a generalized approximation of the grammar learned from past successful experiences. The control mechanism gives a great speed-up and reduction in memory usage. Experiments on real newspaper articles are shown, and our experience with this parser in a machine translation system is described.
Introduction Parsing real-world natural language text is a difficult task. The difficulty comes from several sources. One is that a real-world sentences are long and complex, especially in newspaper articles. The average length of a sentence is 20 to 30 words, and it often exceeds 100 words. The sentence structure is complex and syntactically ambiguous due to the frequent usage of relative clauses, appositives, disjunctions, and other constructions. Also, there are always unexpected syntactic phenomena in real-world texts, which are beyond the formal grammar's coverage. Another difficulty comes from noise and ill-formedness of the input, typically in spoken language. Interjected words (ah, well ), false starts (I'd like to do you have any flights from .), or errors in speech recognition devices make parsing with a rigid grammar impossible. This kind of noise and ill-formedness is also seen in written text, for example, misspelled words, unbalanced parentheses, incorrect punctuation, characters, and special formatting characters which fail to be filtered by text preprocessing modules. One extreme way to deal with these difficulties is to abandon syntactic analysis completely, instead relying purely on statistics for machine translation [Brown, et al. 1990], or using template matching for information extraction [Jackson, et al. 1991]. With these methods, no grammar rules are used, and no analysis of syntactic and semantic structure of sentences is performed. Target translation or extracted information is directly obtained from the surface input, using context sensitive word-by-word translation statistics or target-dependent database templates. These approaches, however, cannot fully utilize the semantic and pragmatic knowledge even when the syntactic structure is easy to parse. When domain-specific semantic and pragmatic knowledge is available, a certain level of loose syntactic analysis has been shown to be useful, such as in [McDonald 1992], [Hobbs, et al. 1992], [Seneff 1992], and [Stallard and Bobrow 1993]. In these systems, a regular parser first attempts to parse a sentence, and when the parser fails, the parse fragments, which represent syntactic structure of parts of the sentence, are combined together using domain-specific semantics and pragmatics. The upper tree is the result of parsing with two words skipped. The lower trees are from a non-skipping producing 10 fragments. The meaning of those fragments (from left) are; 1. According to a source near the president, 2. on 14th, to the parliament, 3. in advance (skipped), 4. being sent, 5. the presidential message is, 6. domestic and foreign policy, 7. comma (skipped), 8. inserted-marker, 9. consists of six chapters by category such as economics, 10, period. Figure 1: Skip Parsing versus Non-skip Parsing Further, if domain-specific lexical and syntactic knowledge is available, grammar rule can be designed to create parse fragments which can be easily combined and used, as in [Ayuso, et al. 1994]. However, these methods, which rely heavily on domain-specific knowledge, are useful only when such knowledge is available. When such knowledge is not available, or in systems which need to handle texts in any domain, domain independent syntactic methods are more useful. For example, [Mellish 1989] shows how to enhance chart-based techniques for parsing ill-formed sentences. In his method, when bottom-up chart parsing fails, top-down predictions, derived from grammar rules, make hypotheses to add/delete/substitute words to parse the sentence. [Lavie and Tomita 1993] show an extension to the GLR (Tomita) parser, which skips words by allowing shift operations at inactive states. Skipping words in the input sentence to obtain a complete parse is effective in processing spoken language with noise. As mentioned above, noise is also seen in written text. Moreover, some unexpected grammatical constructions can be parsed by skipping words. Figure 1 is one example of skip-parsing a Japanese sentence in a newspaper article. In this example, a complete parse (shown as an upper tree in the figure) is obtained by skipping an adverb and a comma. skipping these two words, the parser is left with six major parse fragments (shown as lower trees in the figure). This example is taken from the parser described in this paper. This parser is used in a broad-coverage Japanese-to-English machine translation system [Knight, et al. 1995]. Because this system operates across domains, it is impossible to rely on domain-specific semantic and pragmatic knowledge to complete a parse. Syntax-only domain-independent mechanism for unparsable sentence is the most effective method for a broad-coverage system. There is another reason for using a skip parser. In the early stage of the development, we wanted to have a simple general fallback mechanism which absorbs the inadequacy of text preprocessing module and low coverage of grammar rules. Text formatting of the input can be arbitrary. It is difficult to have a preprocessing module which perfectly filters out symbol characters and special formatting characters, especially in the early stage of development. 1 It takes several months or even years to build sufficiently many grammar rules. 2 Skip parsing is a convenient way to absorb the immaturity of text preprocessing and grammar rules. The basic idea of a skip parser is very simple; just allow the parser to create a syntactic constituent with a gap. But the naive implementation would make it practically unusable. Consider the English noun phrases noun noun . These word sequences often appear in English texts. Notice that most of the noun + "of"+ noun sequences would still be grammatically correct even if the word "of"were skipped. The parser wastes time and space in processing all the possible skips. In other examples, skipping a certain word is locally grammatical but will not lead to a complete successful parse. Due to undesirable or unsuccessful skipping, a practical implementation of a skip parser has to deal with time and space limitations. In other words, the exhaustive search must be avoided by using efficient technique for reducing the search space. [Lavie and Tomita 1993] reduce the search space by using beam-search which limits the number of active nodes. Although this is a heuristic and is not guaranteed to find a optimal skipping, it is reported to be useful. Following this idea, we investigated more powerful heuristics for controlling the search space: using statistics of past successful parses. The statistics are a form of knowledge generalized from past experiences. The heuristics and mechanism of the parser are described in the next section. Our experimental results are shown in Section 3. Discussion and conclusions follow. 2.1 Skip Control using N-gram To avoid grammatically correct but undesirable skipping, we use N-gram information obtained from past successful parsing. As we saw in the last section, it is generally fruitless to try skipping the preposition of between two nouns. To get this kind of knowledge, we feed our parser thousands of unannotated sentences, select sentences which are successfully parsed, and record which sequences of words can be parsed. The frequency statistics of the sequences of words are called N-gram. Note that we select parsable sentences using our own grammar and parser (without skipping), because we need knowledge about which word sequences can be parsed under the current grammar, to find out which words should be skipped for unparsable sentences. The N-gram information obtained from past successful parsing is a handy measure to guess which words are good candidates to be skipped. Normally N-gram frequency information records how often two or more words appear adjoining. We use a variant of N-gram information. As is seen in the previous example, it is enough to know the sequence noun + "of"+ noun is likely to be parsed successfully. Therefore, for content words such as nouns, adjectives, and verbs, the Also, word skipping can work well with markers inserted by a preprocessing module. We use a separate pre- parsing module which inserts phrase boundaries and noun compound markers for efficient parsing (examples are shown in Figure 1). Since this module is not always accurate, the markers must be treated only as hints. The mechanism of skip parsing works as a hint handling facility. It skips the marker if the parse fails and does the parsing without the marks. Word skipping can give good feedback to the grammar writer. By knowing which words are skipped, a grammar writer can easily identify an error or a missing rule in the grammar. grammatical category name is used as an N-gram index. Then, the preposition "of"in noun "of"+ noun is guessed not to be skipped, estimated by the N-gram value of noun The N-gram can be also thought as an approximation of the grammar rules. This approximation can be seen as generalized knowledge learned from past successful experiences. Although long-distance syntactic relations and global phrase structures cannot be captured by the N-grams, intra- phrase local constraints and specific patterns at phrase boundaries can be approximated. 2.2 Extension to Chart Parser We applied this idea of controlled skip parsing on our bottom-up chart parser. A bottom-up chart parser ([Kay 1980]) builds up larger constituents by concatenating adjoining constituents as prescribed by grammar rules. By recording created constituents in a chart, a parse fragment (constituent) is shared by multiple different parse trees, avoiding repetition of the analysis of the same part of the input sentence. Two extensions are needed to the standard chart parser to be able to skip words using the ranking information provided by an N-gram model. One is controlling the order of parsing by the cost (ranking) information, and another is extending lexical constituents to skip words. Each constituent in the chart is given a cost value. If the cost is lower, the parser gives precedence to it, and delays the processing if the cost is higher. The N-gram information is one of the costs used for controlling skip parser. Other factors such as the total number of words skipped, and the isolation factor (described later), are also considered and single cost as an integer value is assigned to each constituent in the chart. The cost is cumulative. When a new constituent is created from several children, the sum of each child's cost (with an adjustment function) is assigned to the new constituent. 3 If it is a normal (non-skipping) one, the cost is assigned zero. Thus, if the parser can parse the input without skipping a word, the final parse tree has a cost of zero. If the normal parsing fails, skipped lexical constituents are created, by extending normal lexical constituents. A normal lexical constituent corresponds to one word in the input sentence. By extending it, it covers more than a two-word span, then words under the extended span are effectively skipped. 4 Skipped lexical constituents are assigned non-zero costs. A value from the N-gram database and other factors decide the cost. The cost for extended lexical constituents are decided only when the such constituents are created, and the cost of non-lexical constituents is just the sum of the cost of its children. 2.3 Cost assignment The following two subsections describe the details of how the cost is assigned. There are two kinds of cost assignments; one is the initial cost assignment, and another is the compositional 3 It is possible to use a complicated function rather than summing the childrens' cost. In our implementation, we only adjust the value when skipping a pair of brackets in the input sentence to treat it as if only one word is skipped. Generally, just a simple summing of childrens' cost is enough, because recalculating the cost of each constituent is costly. 4 A normal lexical constituent can be extended leftward or rightward, but to skip a word, only either left or right word should be extended. It may be possible to extend non-lexical constituents for efficiency, though we have not done tests. See Discussion. Initial cost := if normal lexical constituent then 0 else /* skipped lexical constituent */ number-of-words-skipped Compositional cost := if the skip broke newly balanced brackets then 9999 else sum-of-its-children - number-of-balanced-pair-of-new-brackets * 100 Figure 2: Cost Assignment cost assignment. The initial cost assignment is for a newly created lexical (normal and skipped) constituent, and the compositional cost assignment is for a newly created non-lexical constituent. These costs assignment changes the total behavior of the skip parser. Our particular implementation is shown in Figure 2. This implementation enforces parsing without skipping first, then if it fails, trying to skip one word. 5 If it still fails, it tries skipping two words, and it tries successively until maximum cost (determined by the length of the sentence) is reached. 6 A total cost of zero means no skipping has been attempted. A total cost of 1 to 100 means one word was skipped, a total cost of 101 to 200 means two words were skipped, and so on. 7 We use an N-gram value which is normalized, quantized, and applied an transformation function. The detail is described in the next subsection. The chart isolation factor is given by the variable part-of-bigger-constit?. It takes a value of 0 or 1. If there is a bigger constituent whose span covers the word, it becomes 1. This roughly means that the word successfully becomes a part of a bigger constituent, so that it seems bad idea to skip the word. The variable becomes 0 when a word cannot attach to other constituents due to a grammar problem, and becomes an isolated constituent in the chart. The variable non-skipping-words? prevents the skipping of some important words. Currently, dummy words START-OF-TEXT and END-OF-TEXT are set not to be skipped. 2.4 N-gram construction We next describe assigning costs derived from N-grams, which is a part of initial cost assignment. The N-gram information is important for spotting which words to skip. Effective and efficient construction of N-gram information is one of the key points for the controlled skip parser. First, we mentioned that the grammatical category name is enough for indexing open-class words in the N-grams, as seen in the example noun + "of"+ noun . However, using just the category name is sometimes too coarse. For example, our grammar is an unification-based feature grammar. Following the trend of contemporary linguistics, fewer grammatical categories are used and complicated variations are expressed using a feature representation. In this setting, N-gram information from the sequence of grammatical categories would not be useful enough to control 5 A pair of balanced brackets is considered as one word, and skipping only one of a pair is prohibited by a high cost. These brackets are dummy words inserted by the pre-parsing bracketing module, and are not parentheses or quotation marks in the original input sentence. As mentioned before, these brackets are inserted as hints for the parser, and are sometimes wrong. The skip parser effectively removes those wrong hint marks. 6 It also has a limit of total CPU time and a limit of the number of constituents in the chart. 7 This is not strictly true. For example, if there are four part-of-bigger-constit words skipped in the constituent, the cost will be greater than 130 lowest five-word skipped one. But we usually do not allow five words to be skipped, and practically we do not need strict cost calculation. skipping. Therefore, in our implementation, some of the syntactic features are used to index open-class words in addition to the category name. 8 For closed-class words such as particles and affixes, we used the surface form and the grammatical category. 9 Second, the collected N-gram counts (how many such sequences of words, grammatical cate- gories, or features appear in parsable sentences), should be normalized and quantized. The purpose of normalization is to make the N-gram count independent of the sampling size. To do this, the raw N-gram count is divided by the maximum count, so that the range of normalized count is zero to one. Quantization is done for efficiency at run time. Floating point data is converted to integer data for speed. A transformation function is applied at the same time. Since the N-gram count does not linearly correspond to the parsing cost, some transformation is necessary. The function is arbitrary. In our implementation, we use a step function with five thresholds 0, 1/1000, 1/100, 1/20, and 1/10. The range values are 0, 1, 2, 5, 8, and 10. Finally, but most importantly, we describe how to mimic N-grams by just bigrams (N=2). Since obtaining arbitrary length of N-gram is costly in terms of computation and preparing sample data, we only use bigrams. A bigram value is obtained for an adjacent pair of words. This value is assigned to the word boundary of the two words. Note that if the words have multiple part-of- speech tags, there are multiple possible bigram values at the boundary. If two words have N tags and M tags, the possible number of bigram value is N * M. We select the maximum of these possible values and assign it as the bigram value at the boundary. This value shows how frequently those two words appeared in the past successful parsing. Therefore, the existence of an unseen tagging sequence, which has low value, is irrelevant if there is a more frequently seen tagging sequence. This justifies the use of the maximum of possible bigram values. When assigning cost to a skipped lexical constituent, the parser must decide the cost for the skipped word, using the word boundary bigram value. The minimum of right adjacent boundary value and left adjacent boundary value is used as a cost for skipping the word. A low bigram value at a boundary means that the words either the left or right of the boundary should be skipped. To select one of the two, each word's opposite boundary value can be used. If the word fits well in the sentence, the opposite side's value are likely to be high, which justifies the use of minimum value of the right and left adjacent boundary value. 3 Experiments 3.1 Coverage We have gathered performance data during the development of a large-scale Japanese grammar. The grammar and the parser was built as a part of JAPANGLOSS Japanese-to-English machine 8 For Japanese nouns, we have eight syntactic subcategories, which are represented in a subcategory feature. For Japanese verbs, we put the inflectional form as a feature. These feature values together with its grammatical category are used for an index to the N-gram database. 9 Multiple part-of-speech tags (grammatical category) and multiple parse trees must be handled properly. We use JUMAN ([Matsumoto, et al. 1993]) part-of-speech tagger before parsing. Unlike other English part-of-speech taggers, it gives multiple possible tags. Recall that we collect N-gram information from sentences which can be parsed by the current grammar. When a sentence is parsed with multiple different parse trees due to multiple part-of-speech tagging of a word, the N-gram value must be discounted by the number of trees. For example, a specific word in a sentence has three possible part-of-speech tags and each tag assignment leads to two parse trees, (thus obtaining six parse trees), each word sequence involving such tag must be counted 1/6 times. The same treatment has to be done for a disjunctive feature if such a feature is used for indexing N-grams. sentences parsed without skip parsed with skip Total parsed sentences Table 1: Coverage of Skip Parsing translation system [Knight, et al. 1995]. Table 1 is the summary of the performance as the raw grammar coverage expanded from around 40% to 60%, as seen in the third column of the table, which shows the number of sentences parsable without skipping. The fourth column shows sentences which were parsed only by skipping, and the rightmost column shows the total number of parsed sentences. In other words, the second column indicates the raw grammar coverage, and the rightmost column is the coverage of the skip parser. Five hundred test sentences were randomly 10 taken from actual Japanese newspapers. 11 Bi-grams were collected from four thousand parsable sentences from the same corpus (not overlapping with the test sentences). The average sentence length was 22.4 words (86.4 bytes). The longest sentence was 199 words (696 bytes) long. 12 On Test4n, 13 the average number of skipped words among the sentences which were parsed only by skipping was 1.67 words. The maximum number of words skipped was limited to five words. As the raw grammar coverage grew from 41.3% to 59.0%, the number of sentences parsed by skipping declined from 49.1% to 29.4%, while the total parsed sentences still grew from 90.0% to 96.2%. This shows that some of the sentences parsed by skipping became covered as the raw coverage expanded. Also some sentences which were not parsed even by skipping became parsable as those sentences became easy to parse by skipping. Since the skip parser skips practically only two or three words at most, if the original sentence had three or more problems, it would not be parsed even by skipping until the problems uncovered by the raw grammar had become one or two. 3.2 Correctness Although the coverage of the skip parser is satisfactorily high, the practical question is how good the output of skipped parse is. Skipping a word such as adverb or symbol character is almost harmless for the whole translation, but skipping an important noun or verb would be disastrous. We investigated what kind of words were skipped in the skipped parse in Test4n. The results is shown in Table 2. There were 147 sentences which were parsed by skipping one or more words. The total number of skipped words were 246 words. 14 The first column of the table shows the category of skipped It even included sentences from an article which only consisted of listings of upcoming events. We used the same 500 test sentences for all the tests. Test2 and Test3 failed to test some of the sentences due to some system trouble. The percentages for those two tests are calculated by dividing by the number of sentences which didn't have system trouble (458 for Test2, and 482 for Test3). This number ignores markers inserted by the preprocessing chunker module. If the markers were included, the average was 35.3 words and the longest was 248 words. 13 Test4 and Test4n used the same grammar, but Test4n had a stricter limit on the number of constituents. The detailed analysis in the following sections was done on Test4n. 14 A pair of brackets and chunker marks are counted as one word. Category correct almost wrong Total verb particle 15 11 15 41 other categories 3 1 7 11 parenthesis chunker marker segment error 0 0 11 11 Total (in percentage) (49.4%) (11.1%) (39.4%) (100.0%) Table 2: Correctness of word skipping words. A human reader rated the skip correctness subjectively (correct, almost, or wrong).These are shown in the second to the fourth column of the table. As expected, if the skipped word was a noun or a verb, most of the sentences were broken. However, for particles, skipping those words worked well in some cases. It included particle "no" used as subject marker, which was relatively a rare case. The next row "other categories" includes adverbs, pronouns, and affixes. Symbol character and parenthesis were usually safe to be skipped. Those symbol characters included a dot or a small circle at the beginning of a sentence, or a pair of parentheses enclosing an entire sentence. Chunker markers inserted by a pre-parsing module were always safe to be skipped, because these were just optional markers for the parser. On the other hand, if the front-end word segmentation module, which broke a sentence into a sequence of words 15 , made an error, it was almost always no help skipping some of the words from a erroneously segmented sequence. The overall correctness is shown in the bottom row, indicating how percentage of skipped words were marked as correct, almost, or wrong, in total. It shows 49.4% of the skipping was correct, and 39.4% was wrong. This means that about half of the skipping did not harm seriously in carrying out the syntactic analysis of the sentence. Even though more than half of the correct skipped words were symbol characters, parentheses, and chunker markers, this does not mean that skip parsing mainly works as a compensation of bad preprocessing effects. Symbol characters and parentheses are vital part of our grammar even in the early stage of the development. Newspaper articles make extensive use of commas for lexical, phrasal, and sentential conjunctions, and small dots and other symbol characters often appear in compound nouns or phrasal itemizations. It is very difficult to exhaustively enumerate these non-alphabetic usage as grammar rules, since these are not well studied in linguistic theory, and the variations are diverse. Fortunately, some of the non-alphabetic parts of a sentence are optional, and the skip parser works with them very well. However, these are not totally optional. Human readers benefit from comma and other symbol characters when they read long texts. Similar thing applies to chunker markers. These markers are completely optional in theory, but without them, the CPU and memory usage would be enormous and many newspaper sentences, which tend to be long, are practically unparsable. Therefore, skip parsing not only works as a compensation of wrong but indispensable preprocessing (as in chunker marking), but also works really well for the gaps that the grammar rules could not fill (as in commas and symbol characters, as well as adverbs and particles). 15 There are no overtly marked boundaries between words in Japanese. The results observed in our experiment are satisfactory. As expected, we experienced the good effect of skip parser when the grammar was being developed, and it boosted the grammar coverage from about 60% to 90%. And about half of the extra coverage, the skipping was reasonable. The heuristics presented here are more powerful than beam-search as in [Lavie and Tomita 1993] or than top-down prediction as in [Mellish 1989], yet our method is still simple, efficient and domain- independent. Using the N-gram as the primary heuristic is a convenient way to approximate rules, generalized from past successful experiences. Note that using N-grams is not a method dependent on the parsing mechanism. This method can be applied to virtually all parsers. It is further possible to enhance it by creating constituents with skipped holes from non-lexical larger constituents. In our implementation, we only extend lexical constituents and re-build larger constituents with skipped words. This will save the cost of creating larger constituents from scratch, although bookkeeping of constituents might become complicated. Another possible extension is to use more complicated N-grams, such as collecting data for N=3 or more, or using finer-grained grammatical categories and features for indexing open-class words. Using variable length N-grams and more generalized N-grams as in [Pereira, et al. 1995] would be an interesting extension. Skipping words in a sentence means some kind of loss of information unless the skipped words are noise. However, since human language has some redundancy and all the information in a sentence is not equally important, parsing by skipping words gives reasonably good results if the skipped words are not particularly important. An important direction of further research is to find a way to identify an unreasonable skipping after parsing. One possible strategy is to reject a parse which skipped words with rich content information in the dictionary. Ultimately, it would be best handled by semantic and pragmatic level. [Lavie 1994] uses semantic coherence for ranking N-best results of his skip parser. 5 Conclusion We presented a controlled skip parser, which selectively skips words to parse an unparsable sentence. The control information comes from heuristics obtained by statistical information. The statistical information (N-grams) can be seen as generalized knowledge learned from past successful parses. We presented a parsing algorithm based on extending a standard bottom-up chart parser. Experiments show that it significantly improves the grammar coverage, and the result of skipping is satisfactory. This method is very useful for a system which cannot utilize domain-specific knowledge, such as a broad-coverage machine translation system. 6 Acknowledgments We would like to thank Eduard Hovy and Kevin Knight for their support of this work. We would also thank all the JAPANGLOSS machine translation group. In fact, we first experimented the first version of the controlled skip parser by writing a program surrounding our original non-skipping parser. --R "Pattern Matching in a Linguistically-Motivated Text Understanding System" "A Statistical Approach to Machine Translation" "Robust Processing of Real-World Natural Language Texts" "A Template Matcher for Robust NL Interpretation" "Algorithm Schemata and Data Structures in Syntactic Processing" "Filling Knowledge Gaps in a Broad-Coverage Machine Translation System" "An Integrated Heuristic Scheme for Partial Parse Evaluation" "GLR* - An Efficient Noise-skipping Parsing Algorithm For Context Free Grammars" "Japanese Morphological Analysis System JUMAN Manual" "An Efficient Chart-based Algorithm for Partial-Parsing of Unrestricted Texts" "Some Chart-based Techniques for Parsing Ill-formed Input" "Beyond Word N-grams" "A Relaxation Method for Understanding Spontaneous Speech Utterances" "The Semantic Linker - A New Fragment Combining Method" --TR
skip parser;n-grams;bigrams;syntactic parser;statistical natural language processing;robust processing
593299
How to Determine the Expressive Power of Constraints.
Some constraint languages are more powerful than others because they allow us to express a larger collection of problems. In this paper, we give a precise meaning to this concept of expressive power for constraints over finite sets of values. The central result of the paper is that the expressive power of a given set of constraint types is determined by certain algebraic properties of the underlying relations. These algebraic properties can be calculated by solving a particular constraint satisfaction problem, which we call an indicator problem. We discuss the connection between expressive power and computational complexity, and show that indicator problems provide a simple method to test for tractability.
Introduction Some constraint languages are more powerful than others because they allow us to express a larger collection of problems. For example, if we are dealing with applications involving real-valued variables, then it is possible to express more numerical relationships using arbitrary polynomial equations than if we were restricted to using just linear equations. Similarly, if we are dealing with applications involving Boolean variables, then it is possible to express more logical relationships using ternary clauses than if we were restricted to using just binary clauses. Of course, the penalty for increased expressive power is generally an increase in computational complexity. For example, the satisfiability problem with ternary clauses is NP-complete [6], whereas the satisfiabiliy problem involving only binary clauses can be solved in polynomial time [6]. More generally, the finite constraint satisfaction problem with arbitrary constraints is known to be NP-complete [13], whereas many families of restricted constraints have been identified which give rise to tractable subproblems [3, 9, 10, 11, 14, 17, 18]. For any given application, it therefore pays to select a constraint system which has sufficient expressive power to express the desired constraints, but is sufficiently restrictive to allow an efficient solution technique. In this paper, we focus on problems with finite sets of possible values and examine how the expressive power of a given set of constraint types may be defined. The central result of the paper is that the expressive power of a given set of constraint types is determined by certain algebraic properties of the corresponding relations. We also show how these algebraic properties can be calculated by solving a particular constraint satisfaction problem, which we call an 'indicator problem'. Finally, we discuss the connection between expressive power and computational complexity, and show that indicator problems provide a simple method to test for tractability. The paper is organised as follows. In Section 2 we give the basic definitions, and describe a general form of algebraic closure property for a set of relations. In Section 3 we describe how to calculate these closure properties and in Section 4 we establish the close link between closure properties and expressive power. In Section 5 we discuss the relationship between expressive power and tractability. Finally, we summarise the results presented and indicate some directions for future research. 2. Definitions 2.1. The constraint satisfaction problem A central notion in the study of constraints and constraint satisfaction problems is the idea of a relation, so we will make extensive use of the following notation. Notation 1 For any set D and any natural number n, we denote the set of all n-tuples of elements of D by D n . For any tuple t 2 D n and any i in the range 1 to n, we denote the value in the ith coordinate position of t by t[i]. The tuple t will be written in the form ht[1]; Definition 1. A subset of D n is called an n-ary relation over D. Example 1 We now describe four relations which will be used as examples through-out the paper. Each of these relations is a set of tuples of elements from the set as defined below: In order to manipulate and combine relations, we will make use of the following standard operations from relational algebra [2]: Definition 2. We define the following operations on relations. ffl Let R 1 be an n-ary relation over a set D and let R 2 be an m-ary relation over D. The Cartesian product R 1 \Theta R 2 is defined to be the (n m)-ary relation ffl Let R be an n-ary relation over a set D. Let 1 - n. The equality selection oe i=j (R) is defined to be the n-ary relation oe ffl Let R be an n-ary relation over a set D. Let i be a sequence of indices chosen from the values n. The projection - (R) is defined to be the k-ary relation We now define the (finite) constraint satisfaction problem which has been widely studied in the Artificial Intelligence community [12, 13, 14]. Definition 3. An instance of a constraint satisfaction problem consists of ffl a finite set of variables, ffl a finite set of values, D; ffl a set of constraints fC g. Each constraint C i is a pair (s i list of variables of length called the constraint scope, and R i is an m i -ary relation over D, called the constraint relation. The tuples of the constraint relation R i indicate the allowed combinations of simultaneous values for the variables in the corresponding scope s i . The length of these tuples will be called the arity of the constraint. In particular, unary constraints specify the allowed values for a single variable, and binary constraints specify the allowed combinations of values for a pair of variables. Notation 2 For any set of relations, \Gamma, the class of all constraint satisfaction problem instances in which the constraint relations are elements of \Gamma will be denoted CSP(\Gamma). Example be the set of four relations, fR 1 as defined in Example 1. One element of CSP(\Gamma) is the constraint satisfaction problem instance, P, which has 6 variables and 3 constraints, defined as follows: ffl The set of variables is ffl The set of values is ffl The set of constraints is fC The structure of this problem instance is illustrated in Figure 1. 2 oe ae x x x x x x Figure 1. The CSP instance defined in Example 2 A solution to a constraint satisfaction problem is a function from the variables to the values such that the combination of values assigned to the variables in each constraint scope is allowed by the corresponding constraint. For any given problem instance we may wish to: ffl decide whether a solution exists; or ffl find one solution, if any exist; or ffl count the number of solutions; or ffl find all solutions. Each of these tasks may have a different complexity, so it is important to be precise about which version of the problem we are trying to solve. It is often convenient to view the solutions to a constraint satisfaction problem as tuples in a relation, as illustrated in the next example. This relation is the largest relation, S, such that the projection of S onto each constraint scope is a subset of the corresponding constraint relation. Example 3 The solutions to the constraint satisfaction problem instance P defined in Example 2 are functions from the set of variables V to the set of values D which satisfy all of the constraints. There are four such functions, which may be expressed in tabular form, as follows: Variables Solution Solution Solution Solution This set of solutions may also be viewed as a relation, S, of arity 6, which has four tuples, corresponding to the four rows in the table. 2 2.2. Expressive power of constraints In any constraint satisfaction problem instance some of the constraints are explicit, whilst others are generally present only as implicit constraints. For any instance in CSP(\Gamma), the explicit constraint relations must be elements of \Gamma, but there may be implicit constraints on some subsets of the variables for which the corresponding relations are not elements of \Gamma. These implicit constraints result from the combined effects of the explicit constraints, and so are said to be 'derived' from these explicit constraints. We now define exactly what it means to say that a constraint relation can be derived from other constraint relations, and what is meant by 'expressive power'. Definition 4. A relation R can be derived from a set of relations \Gamma if it is equal to some projection of the set of solutions to some constraint satisfaction problem instance in CSP(\Gamma). Example 4 Reconsider the set of solutions S for the constraint satisfaction problem instance P, defined in Example 2, as described in Example 3. Note that - 1;2 binary relation can be derived from g. 2 One way to view Definition 4 is as follows. For any relation R of arity r, and any collection of r variables, R can be derived from \Gamma if and only if it is possible to construct a constraint satisfaction problem instance in CSP(\Gamma) which imposes the relation R on the values of those variables. In other words, R can be derived from precisely when there is some combination of the relations in \Gamma that can be used to 'express' a constraint with relation R. Definition 5. The set of all relations that can be derived from \Gamma will be called the expressive power of \Gamma. We now show that the notion of a derived relation corresponds precisely to the notion of a relation which may be obtained using the operations described in Definition 2. be a set of relations over a finite set D. A relation R can be derived from \Gamma if and only if R can be obtained from the elements together with the unary relation D 1 , using some finite sequence of Cartesian product, equality selection, and projection operations. Proof: We first need to show that if R is a projection of the set of solutions to some instance 1 in CSP(\Gamma), then R can be obtained from \Gamma [ using some finite sequence of Cartesian product, equality selection, and projection operations. It is shown in [7] that the set of solutions to any constraint satisfaction problem instance can be obtained by performing a relational join operation on the constraint relations (see Definition 2.8 of [7]). For any instance in CSP(\Gamma) the relations to be joined will either be elements of \Gamma, or the unary relation D 1 (in the case of variables with no explicit constraint specified). Furthermore, it is a standard result of relational database theory [2] that the join operation can be calculated by performing a sequence of Cartesian product, equality selection and projection operations. Hence R can be obtained using the specified relational operations. Conversely, if R can be obtained from \Gamma [fD 1 g using some sequence of Cartesian product, equality selection, and projection operations, then we need to show that R is equal to some projection of the set of solutions to some instance in CSP(\Gamma). To establish this, we first note that, for each of the specified operations, it is straightforward to construct an associated constraint satisfaction problem instance P in CSP(\Gamma) such that the result of that operation is equal to some projection of the set of solutions to P. (The effect of the equality selection operation is achieved by identifying the corresponding variables in P.) By combining these constructions we can derive R from \Gamma. In view of this result, it will be convenient to introduce the following notation. Notation 3 Let \Gamma be a set of relations over a set D. The set of all relations which may be obtained from \Gamma, together with the unary relation D 1 , using some finite sequence of Cartesian product, equality selection, and projection operations, will be denoted states that the set of relations \Gamma precisely equal to the expressive power of \Gamma, as defined above. 2.3. Closure properties of relations We will show in Section 4 that the set of relations \Gamma + is determined by certain algebraic properties of \Gamma. In order to describe these properties, we need to consider arbitrary operations on a set D, in other words, arbitrary functions from D k to D, for arbitrary values of k. Any such operation on D may be extended to an operation on tuples over D by applying the operation in each coordinate position separately (i.e., pointwise). Hence, any operation defined on a set may be used to define an operation on the tuples in a relation over that set, as follows: Definition 6. Let R be an n-ary relation over a set D, and any k-ary operation on D. For any collection of k tuples, t necessarily all distinct) the defined as follows: Using this definition, we now define the following closure property of relations. Definition 7. Let R be a relation over a set D, and D be a k-ary operation on D. R is said to be closed under\Omega if, for all (not necessarily all distinct),\Omega Example 5 Let 4 denote the ternary operation which returns the first repeated value of its three arguments, or the first value if they are all distinct. The relation R 2 , defined in Example 1, is closed under 4, since applying the 4 operation to any 3 elements of R 2 yields an element of R 2 . For example, The relation R 1 , defined in Example 1, is not closed under 4, since applying the 4 operation to the last 3 elements of R 1 yields a tuple which is not an element of positive integer m, and any positive integer i - m, let denote the operation of arity m which returns the value of its ith argument. In other words, for all d Every relation R over D is closed under @ m;i for all values of m and i, since applying the @ m;i operation to any m elements of R simply returns the ith one. 2 For any set of relations \Gamma, and any operation\Omega , if every R 2 \Gamma is closed under \Omega , then we shall say that \Gamma is closed under\Omega . The next lemma indicates that the property of being closed under some operation is preserved by each of the operations on relations described in Definition 2. R and R 0 be relations which are closed under\Omega , for some operation \Omega . The following relations are also closed 1. the Cartesian product, R \Theta R 0 ; 2. any projection of R or R 0 ; 3. any equality selection from R or R 0 . Proof: Follows immediately from the definitions. Relations which are closed under the same operations have many common proper- ties, so we introduce the following notation. Notation 4 Let \Gamma be a set of relations over a set D. The set of all relations over D which are closed under all of the operations under which \Gamma is closed will be denoted \Gamma. 3. Calculating Closure Operations In view of the results presented later in this paper, it is important to be able to calculate the closure operations for a given set of relations. In this section, we show that it is possible to obtain these operations, for each possible arity, by obtaining all solutions to a single constraint satisfaction problem instance. First note that, for any set of relations \Gamma over a set D, the operations under which \Gamma is closed are simply mappings from D k to D, for some k, which satisfy certain conditions, as described in Definition 7. This suggests that it should be possible to define a constraint satisfaction problem in which the variables are the elements of and the constraints impose the necessary conditions to ensure that the solutions are closure operations. This is the motivation for the idea of an 'indicator problem' which is defined as follows. Definition 8. Let \Gamma be a set of relations over a finite set D. For any natural number m ? 0, the indicator problem for \Gamma of order m is defined to be the constraint satisfaction problem instance IP (\Gamma; m) with ffl set of variables D m ; ffl set of values D; ffl set of constraints fC g, such that for each R 2 \Gamma, and for each sequence of tuples from R, there is a constraint C is the arity of R and Note that for any set of relations \Gamma over a set D, IP (\Gamma; m) has jDj m variables and constraints. Example 7 Consider the relation R 1 over defined in Example 1. The indicator problem for fR 1 g of order 1, IP(fR 1 g; 1), has 3 variables and 4 constraints. The set of variables is and the set of constraints is The indicator problem for fR 1 g of order 2, IP(fR 1 g; 2), has 9 variables and 16 constraints. The set of variables is and the set of constraints is Example 8 Consider the relation R 2 over defined in Example 1. The indicator problem for fR 2 g of order 1, IP(fR 2 g; 1), has 3 variables and 3 constraints. The set of variables is and the set of constraints is The indicator problem for fR 2 g of order 2, IP(fR 2 g; 2), has 9 variables and 9 constraints. The set of variables is and the set of constraints is Consider the relations R 1 and R 2 over defined in Example 1. The indicator problem for fR 1 of order 1, IP(fR 1 and 7 constraints. The set of variables is and the set of constraints is equal to the union of the set of constraints of IP(fR 1 as defined in Example 7, and the set of constraints of IP(fR 2 g; 1), as defined in Example 8. The indicator problem for fR 1 of order 2, IP(fR 1 and 25 constraints. The set of variables is and the set of constraints is equal to the union of the set of constraints of IP(fR 1 as defined in Example 7, and the set of constraints of IP(fR 2 g; 2), as defined in Example 8. 2 Solutions to the indicator problem for \Gamma of order k are functions from D k to D, or in other words, k-ary operations on D. We now show that they are precisely the k-ary operations under which \Gamma is closed. Theorem 1 For any set of relations \Gamma over a set D, the set of solutions to IP (\Gamma; is equal to the set of k-ary operations under which \Gamma is closed. Proof: By Definition 7, we know that \Gamma is closed under the k-ary operation\Omega if and only if\Omega satisfies the condition for each possible choice of R (not necessarily all distinct). By Definition 8, this is equivalent to saying satisfies all the constraints in IP (\Gamma; k), so the result follows. Example 10 Consider the relation R 1 over defined in Example 1. The indicator problem for fR 1 g of order 1, defined in Example 7, has 2 solutions, which may be expressed in tabular form as follows: Variables Solution Solution Hence, by Theorem 1, fR 1 g is closed under precisely two unary operations: the constant operation with value 0, and the identity operation. The indicator problem for fR 1 g of order 2, defined in Example 7, has 4 solutions, which may be expressed in tabular form as follows: Variables Solution Solution Solution Solution Hence, by Theorem 1, fR 1 g is closed under precisely four binary operations. 2 Example 11 Consider the relation R 2 over defined in Example 1. The indicator problem for fR 2 g of order 1, defined in Example 8, has 3 solutions, which may be expressed in tabular form as follows: Variables Solution Solution Solution Hence, by Theorem 1, fR 2 g is closed under precisely three unary operations. The indicator problem for fR 2 g of order 3, has a very large number of solutions, including the operation 4, defined in Example 5. Hence fR 2 g is closed under this ternary operation, as well as a large number of others. 2 4. Closure Operations and Expressive Power We have so far defined two sets of relations which are associated with a given set \Gamma: the set of relations which can be derived from \Gamma, denoted and the set of relations which share the same closure operators as \Gamma, denoted \Gamma. In this section, we will show that these two sets of relations, which are defined in very different ways, are actually identical. Theorem 2 For any set of relations, \Gamma, over a finite set D, Proof: By Lemma 2, we know that the property of being closed under any operation is preserved by Cartesian product, projection, and selection operations. It follows immediately that To obtain the reverse inclusion, let any relation in \Gamma. Construct the indicator problem for \Gamma of order m, IP (\Gamma; m), and let S be the set of solutions to IP (\Gamma; m), viewed as a relation. Since every constraint relation in IP (\Gamma; m) is an element of \Gamma, we know that S can be derived from \Gamma, and hence We claim that R can be obtained as a projection of S, and hence R also belongs which gives the result. To establish this claim, let r be the arity of R, and let c Each c j is an element of D m , and hence corresponds to some variable of IP (\Gamma; m), and hence to some coordinate position, p j , of S. Let (S). By the construction of T , each tuple of T is equal to the result of applying one solution of IP (\Gamma; m) to the sequence of tuples belongs to \Gamma, we know that R is closed under all solutions of IP (\Gamma; m), by Theorem 1, so each tuple in T must belong to R. Hence, T ' R. Conversely, it was shown in Example 6 that every relation is closed under the operations @ m;i , for m. Hence, by Theorem 1, each @ m;i is a solution of IP (\Gamma; m), and so gives rise to a tuple in T . By the construction of T , and the definition of @ m:i , this tuple is equal to t i . Hence, R ' T . Hence which establishes the claim. Combining this result with Lemma 1, we have shown that the expressive power of a set of relations is determined by the corresponding set of closure operations. Corollary 1 A relation R can be derived from a set of relations \Gamma over a finite set D if and only if R is closed under all of the operations under which \Gamma is closed. Example 12 Reconsider the relations R 1 and R 2 over defined in Example 1. It was shown in Examples 10 and 11 that R 1 is not closed under all the operations under which R 2 is closed. Hence, by Corollary 1, R 1 cannot be derived from R 2 . Similarly, the results given in Examples 10 and 11 also establish that R 2 cannot be derived from R 1 . 2 5. Expressive Power and Complexity We have seen that the expressive power of a set of relations is determined by their closure operations. To illustrate one application of these results, we now examine how they can be used to obtain information about the complexity of the class of problem instances involving a given set of constraint types. Deciding whether or not a given constraint satisfaction problem instance has a solution is NP-complete in general [13] even when the constraints are restricted to binary constraints. In this section, we shall consider how restricting the allowed constraint relations to some fixed subset of all the possible relations affects the complexity of this decision. We shall therefore regard CSP(\Gamma) as a decision problem in which the question to be decided in each instance is the existence of a solution. If there exists an algorithm which decides every instance in CSP(\Gamma) in polynomial time, then we shall say that \Gamma is a tractable set of relations. Example 13 The binary inequality relation over a set D, denoted 6= D , is defined as Note that CSP(f6=D g) corresponds precisely to the Graph jDj-Colorability problem [6]. This problem is tractable when jDj - 2 and NP-complete when The relation R 4 , defined in Example 1, is equal to the relation 6= D for 2g. Hence, CSP(\Gamma) is NP-complete for any set of relations \Gamma which contains R 4 . 2 It turns out that certain special operations play a major role in distinguishing tractable and intractable sets of relations. These operations are the 'essentially unary operations', which are defined as follows. Definition 9. An called essentially unary if there exists some non-constant unary operation f : D ! D and some i in the range 1 to k such If f is a permutation, then\Omega is called essentially permuting. Example 14 The operations @ m;i , defined in Example 6, are all essentially unary operations. 2 Note that constant functions are excluded from Definition 9, and so are not essentially unary. Theorem 3 Let \Gamma be a set of relations over a finite set D. If every operation under which \Gamma is closed is essentially unary, then CSP(\Gamma) is NP-complete. Proof: Assume that \Gamma is only closed under essentially unary operations. First, we note that if \Gamma is closed under some essentially unary operation\Omega , then it is also closed under the corresponding (non-constant) unary operation f . Let f 0 be a unary operation under which \Gamma is closed such that jf 0 (D)j is minimal. \Gammag. By the choice of f 0 , and the assumption about \Gamma, it follows that f 0 (\Gamma) is only closed under essentially permuting operations. \Gamma is not closed under any constant operation (by assumption), we know that jf 0 (D)j ? 1. There are 2 cases to consider. ffl In the case where jf 0 2, we consider the ternary 'not-all-equal' relation, N , over D, (i.e., the relation containing all 3-tuples over D in which the 3 values are not all equal). The relation N is closed under all essentially permuting operations on D, which means that N 2 f 0 (\Gamma). By Theorem 2 and Lemma 1, this implies that N may be derived from f 0 (\Gamma). Hence, CSP(fNg) may be reduced to CSP(f 0 (\Gamma)) in polynomial time. Now, CSP(fNg) is equivalent to the Not-All-Equal Satisfiability problem, which is known to be NP-complete [6, 15], so it follows that CSP(f 0 (\Gamma)) is also NP-complete. ffl In the case where jf 0 (D)j ? 2, we consider the binary disequality relation, , defined in Example 13. The relation 6= f0(D) is closed under all essentially permuting operations on D, which means that 6= f0(D) 2 f 0 (\Gamma). By Theorem 2 and Lemma 1, this implies that 6= f0(D) may be derived from f 0 (\Gamma). Hence, CSP(f6= f0(D) g) may be reduced to CSP(f 0 (\Gamma)) in polynomial time. is equivalent to the jf 0 (D)j-Colorability problem, which is known to be NP-complete when jf Finally, any instance in CSP(f 0 (\Gamma)) has a solution if and only if the corresponding instance in CSP(\Gamma) has a solution, because \Gamma is closed under f 0 . Hence CSP(f 0 (\Gamma)) is reducible to CSP(\Gamma) in polynomial time, which establishes that CSP(\Gamma) is also NP-complete. If we assume that NP-complete problems do not have polynomial-time algorithms, then this result provides a necessary condition for tractability: a tractable set of relations must be closed under some operation which is not essentially unary. In fact, it was shown in [9] that all the currently known examples of tractable constraint types can be characterised as sets of relations which are closed under particular non-unary operations, even in those cases where they were originally defined using very different criteria. For example, the '0/1/all' relations 2 , first described in [3, 11], can be characterised as the set of all relations closed under the operation 4, defined in Example 5. As a second example, the various sets of 'max-closed' relations, first described in [10], can each be characterised as the set of relations closed under a particular binary associative, commutative, idempotent operation [9]. We may sharpen the result of Theorem 3 a little further by bounding the possible arity of the closure operations on \Gamma which we need to consider, as the next result shows. Theorem 4 For any set of relations \Gamma over a finite set D, if \Gamma is closed under some operation\Omega which is not essentially unary, then it is also closed under some operation -\Omega of arity at most maxf3; jDjg, which is not essentially unary. Proof: (This proof is adapted from the proof of Lemma 1.14 in [16]). Let \Gamma be a set of relations which is closed under some operation\Omega which is not essentially unary. Let -\Omega be an operation of the smallest possible arity, such that \Gamma is closed under -\Omega and -\Omega is not essentially unary, and let k be the arity of -\Omega . If k - 3, then the result holds, so we only need to consider the case where k - 4. Now consider the operations which are obtained from -\Omega by identifying two arguments (i.e., requiring them to have the same value). Since \Gamma is closed under these operations, which have a lower arity than -\Omega , they must all be essentially unary. Hence, if we identify the first two arguments we have z - for some non-constant unary operation f 12 and some i 2 kg. Similarly, if we identify the third and fourth arguments we have z - for some non-constant unary operation f 34 and some j 2 kg. This means that z - z - for all possible choices of x which implies that either i 62 f1; 2g or j 62 f3; 4g. It follows from this that we can permute the order of the arguments of -\Omega to obtain a function \Phi which satisfies the identity z- for some non-constant unary operation f . In particular, we have distinct pairs of indices we know that the is an essentially unary function. Since we have just shown that for the case x the value of this function is equal to f(x 1 ), we know that it must equal cases. Similarly, for all indices is an essentially unary function. When the remaining arguments all have the same value, then this function is equal to f(x), by the arguments above. Hence it must be equal to f(x) in all cases. there are more arguments than values, so at least one argument value must be repeated somewhere, and in all such cases we have shown that just -\Omega with the order of the arguments permuted, we have contradicted the fact that -\Omega is not essentially unary. This means that we must have k - jDj, and the result follows. In order to use Theorem 3 to draw conclusions about the tractability of a set of constraints we need to be able to distinguish closure operations which are essentially unary from those which are not. By Definition 9, this means examining each solution to the appropriate indicator problem to check whether the value of the solution at each variable is determined by a single coordinate position in the tuple describing that variable. For an operation of arity m, this checking may be carried out in mjDj m steps. Alternatively, the next result shows that the number of solutions to certain indicator problems provides a sufficient condition for establishing NP-completeness, without needing to examine the individual solutions in detail. be a set of relations over a set D, let S 1 be the set of non-constant solutions to IP (\Gamma; 1), and let Sm be the set of solutions to IP (\Gamma; m), where then CSP(\Gamma) is NP-complete. Proof: By Theorem 3 and Theorem 4, we know that either CSP(\Gamma) is NP- complete, or \Gamma is closed under some operation which is not essentially unary, and which has arity at most m. By Theorem 1, Sm is the set of all m-ary operations under which \Gamma is closed. By Definition 9, the number of operations in Sm which are essentially unary is equal to the number of non-constant unary operations under which \Gamma is closed, multiplied by m. By Theorem 1, this number is mjS 1 j. Hence, jS m j - mjS 1 j in all cases. In the limiting case, when jS we know that every element of Sm is essentially unary, so \Gamma is not closed under any m-ary operation which is not essentially unary. It follows that \Gamma is not closed under any operation of arity lower than m which is not essentially unary, so CSP(\Gamma) must be NP-complete. Corollary 2 establishes NP-completeness for many sets of constraints without the need for individually constructed, ad-hoc reduction arguments, as the following examples illustrate. Example 15 Consider the relations R 1 and R 2 over defined in Example 1. The indicator problem for fR 1 of order 1, defined in Example 9, has 1 solu- tion, corresponding to the identity operation. The indicator problem for fR 1 of order 3, has 3 solutions (which are all essentially unary). Hence is NP-complete, by Corollary 2. On the other hand, it was shown in Example 10 that R 1 is closed under a constant operation, so CSP(fR 1 g) is tractable, by Proposition 9 of [9]. In fact, any problem in CSP(fR 1 g) has the solution which assigns the value 0 to each variable, so this class of problems is trivial. Furthermore, it was shown in Example 11 that R 2 is closed under the operation 4, defined in Example 5. Hence, CSP(fR 2 g) is tractable, by Theorem 13 of [9]. 2 Example Consider the relations R 1 defined in Example 1. By counting the solutions to the indicator problems of order 1 and order 3 for each relation and each pair of distinct relations in this set, we are able to complete the analysis of the complexity of CSP(\Gamma) for each possible subset \Gamma of these relations. Relations # Solutions # Solutions Complexity (Corollary (Corollary (Corollary For all larger sets of relations we have that \Gamma contains at least one of the pairs of relations shown in the table, and so CSP(\Gamma) is NP-complete.How practical is the test proposed here in general? For small values of jDj, the relevant indicator problems have very few variables, and the solutions may be found easily. This remains true even when the arity of the relations in \Gamma is large. In the simplest case, when 2, the constraint satisfaction problem corresponds to the classical Satisfiability problem [6], for which all possible tractable sets of relations are known, and are fully described in [15]. The only non-trivial examples in this case are the sets of Boolean relations which can be expressed by Horn clauses, or the duals of Horn clauses, or clauses containing at most two literals, or linear equations over the two-element field [15]. It has been shown that any set of relations that is closed under an operation which is not essentially unary belongs to one of these tractable classes [8]. Hence the condition described in Corollary 2 is both necessary and sufficient for NP-completeness in this case (assuming that P is not equal to NP). It follows from this that solving the indicator problems of order 1 and order 3 provides a simple and complete test for tractability of any set of relations over any set with 2 elements. This provides a very practical answer to a question posed by Schaefer in 1978 [15] concerning the existence of an efficient test for tractability in the Generalized Satisfiability problem. Note that carrying out the test requires finding the number of solutions to a constraint satisfaction problem with just 8 Boolean variables. As the size of D increases, the size of the relevant indicator problems increases rapidly, and it may become impractical to compute all solutions, or even to calculate the number of solutions, as required by Corollary 2. On the other hand, for cases of interest, it may be possible to establish from known properties of the constraints that the relevant indicator problems will have particular types of solu- tion, without carrying out a complete solution algorithm. This question is currently being investigated. 6. Conclusion In this paper we have shown how the algebraic properties of relations can be used to determine the expressive power of a set of possible constraint types. Furthermore, we have proposed a method for determining these algebraic properties by solving a particular form of constraint satisfaction problem, which we have called an indicator problem. Finally, we have shown that there are simple algebraic properties which distinguish between sets of relations which give rise to tractable constraint satisfaction problems and those which give rise to NP-complete problems. We have used these results to obtain a sufficient condition for NP-completeness, which we have shown is widely applicable and easy to test. For problems where the set of possible values contains just two elements these results provide a necessary and sufficient condition for NP-completeness (assuming that P is not equal to NP), and an efficient test to distinguish the tractable sets of relations. We are now investigating the application of these results to particular problem types, such as temporal problems involving subsets of the interval algebra. We are also attempting to determine how the presence of particular algebraic closure properties in the constraints may be used to derive appropriate efficient algorithms for tractable problem classes. Acknowledgments We are grateful to Victor Dalmau for helpful discussions which greatly simplified the proof of Theorem 2 Notes 1. Since D is finite, and hence there are only finitely many possibilities for R at each arity, we may assume, without loss of generality, that this instance contains only a finite number of constraints. 2. Also called 'implicational' relations [11]. --R "Constraint Satisfaction From a Deductive Viewpoint" "A Relational Model of Data for Large Shared Databanks" "Characterizing tractable constraints" "Structure identification in relational data" "A sufficient condition for backtrack-bounded search" Computers and intractability: a guide to NP-completeness "Decomposing constraint satisfaction problems using database techniques" "An algebraic characterization of tractable constraints" "A unifying framework for tractable constraints" "Tractable constraints on ordered domains" "Fast parallel constraint satisfaction" "On binary constraint problems" "Consistency in networks of relations" "Networks of constraints: fundamental properties and applications to picture processing" "The complexity of satisfiability problems" Clones in Universal Algebra "On the Minimality and Decomposability of Row-Convex Constraint Net- works" "A generic arc-consistency algorithm and its specializations" --TR --CTR Peter Jeavons , David Cohen , Justin Pearson, Constraints and universal algebra, Annals of Mathematics and Artificial Intelligence, v.24 n.1-4, p.51-67, 1998 Lane A. Hemaspaandra, SIGACT news complexity theory column 43, ACM SIGACT News, v.35 n.1, March 2004 Richard Gault , Peter Jeavons, Implementing a Test for Tractability, Constraints, v.9 n.2, p.139-160, April 2004 Peter Jonsson , Andrei Krokhin, Recognizing frozen variables in constraint satisfaction problems, Theoretical Computer Science, v.329 n.1-3, p.93-113, 13 December 2004 David A. Cohen , Martin C. Cooper , Peter G. Jeavons , Andrei A. Krokhin, The complexity of soft constraint satisfaction, Artificial Intelligence, v.170 n.11, p.983-1016, August 2006
constraint satisfaction problem;complexity;NP-completeness;relational algebra;indicator problem
593308
A Methodological View of Constraint Solving.
Constraints are an effective tool to define sets of data by means of logical formulae. Our goal here is to survey the notion of constraint system and to give examples of constraint systems operating on various domains, such as natural, rational or real numbers, finite domains, and term domains. We classify the different methods used for solving constraints, syntactic methods based on transformations, semantic methods based on adequate representations of constraints, hybrid methods combining transformations and enumerations. The concepts and methods are illustrated via examples. We also discuss applications of constraints to various fields, such as programming, operations research, and theorem proving.
Introduction Constraints allow to define sets of data by means of logical formulae. For example, the set of even numbers y given in extension as the set f0; 2; can be described in comprehension by the formula fy stating that there exists a natural number x such that y is twice x. In this formula, the subformula x is the constraint, and N is its domain of interpretation. Constraints have become very successful in computer science and operation research. There are several reasons for this, some of which are listed below. A convenient notation. Constraints provide a compact, simple, and familiar representation of possibly infinite sets, which is a first reason for their success. Expressivity. A second reason proceeds from a more pragmatic argument: constraints allow to express and manipulate large sets of data that could not be reasonably handled if given in extension. Describing the infinite set of even numbers in extension means to give a procedure than can enumerate this set by outputing the value 2n upon input n. Defining the set of multiples of 3 can be done in the same manner. The intersection of the two sets can now be computed by enumerating them both al- ternatively, and outputing the common values: 6; 12; There is a deadly risk, of course, when the intersection is empty. In contrast to this procedural style, constraints are declarative: they allow to state properties independently of an actual computation mechanism, called solving. In the above example, the set of triples is specified as 9z z with N for domain of interpretation. The intersection of the two sets can be simply expressed by accumulating both constraints, resulting in 9x z. Constraints are indeed targetted at specifying set-theoretic properties. Efficiency. Declarative languages are often considered as inefficient. A third important reason why constraints are so useful, is that they reconcile declarativeness with efficiency. A constraint is solved in a given domain of interpretation, hence very efficient algorithms can be chosen for this purpose and can in practice be fetched from a library. Other declarative languages would use instead a general purpose solving algorithm called search. For example, a standard Prolog implementation would solve a system of linear equations over the rationals by using SLD resolution, while a language allowing for rational constraints can call Gaussian elimination, and therefore be much more efficient. Easy combination. There is yet a fourth reason for the success of constraints: they can be very easily combined with other computation models. For example, combining logic programming and constraints is achieved by replacing the unification component of a logic programming language by constraint solving algo- rithms. In this combination, the logic programming computation produces new constraints which are consumed by the constraint solving algorithms, which in turn produces new inputs for the logic programming computation. The ability to efficiently compute with sets of constraints which evolve along a computation is a last major property of constraints called incrementality. Constraints play a key role since the work of Alan Colmerauer [6], their abstract formulation by Joxan Jaffar and Jean-Louis Lassez [26], and the introduction of finite domain constraints in CHIP [16]. So far, we have not explained a major aspect of constraints, directly related to the title of this paper. There are (in general infinitely) many ways to represent a set of data by means of constraints. For example, the set of even numbers could have been defined as the set of non-odd natural numbers z, that is by the constraint This is of course a more complicated equivalent but what does more complicated mean? There is no definite answer to this question, unless the formula we are interested in defines the empty set. Indeed, it is A METHODOLOGICAL VIEW OF CONSTRAINT SOLVING 3 a major problem to avoid useless computations by efficiently checking constraints for emptiness. To summarize, a constraint system comes in three parts: 1. The syntax in which the logical formulae are expressed. In most cases, it is a first-order language, or a fragment thereof closed under the conjunction operator &. Very common are the existential fragment, which is free of universal quantifier, and the positive fragment, for which negation and disjunction are not allowed. Constraints without logical connectives are called atomic. A typical example consists in the conjunctions of linear equations over rational numbers, such as 2. The structure in which the formulae are interpreted, composed of a domain of values on one hand, and, on the other hand, of operations and relations interpreting all symbols in the syntax. In our example, the domain of values is the set of rational numbers, with the usual operations. A solution of a constraint assigns its free variables to values of the domain in such a way that the formula evaluates to True. In the above example, there is only one solution: x := 27; y := 36. 3. A constraint solving algorithm takes as input a constraint C and returns as output a particular representation of the set of all its solutions. As an important particular case, this algorithm allows to decide whether C has at least one solution, this is called constraint satisfaction. A more general algorithm is often needed, called constraint entailment, which decides whether the set of solutions of a constraint C is included into the set of solutions of a constraint C 0 . Gaussian elimination is one algorithm used to solve linear equations over rational numbers. We describe later in detail constraints on finite domains, which pop up in many application areas. Such constraints are NP-hard, which makes it impossible to deal with large application problems. As a consequence, practice sometimes favors incomplete algorithms, which may sometimes fail to detect that a given constraint has no solution. On the other hand, a constraint solving algorithm must always be sound: it should not pretend that there is no solution when there is some. So far, all our examples deal with numbers. Indeed, constraints over numbers or Booleans have many industrial applications as operations research and hardware verification. But, there are also applications to computer science, for which the domain is itself made of expressions, called terms in this context. A typical example is the problem of solving equations over terms, called unification, which is at the heart of logic programming. We call symbolic, constraints over terms. Symbolic constraints have been used for years in computer science in order to: ffl represent sets of formulae. A constrained formula is a pair (OE j C) which stands for the set of instances of OE by the solutions of C. Consider arithmetic expressions built over the addition + and the multiplication as they are found in 4 H. COMON, M. DINCBAS, J.-P. JOUANNAUD, C. KIRCHNER programming languages. The subset of these expressions which are sums of two different expressions can be expressed as the formula 9xy Here, x 6= y means that x and y stand for two syntactically different expres- sions. This example shows the expressive power of constraints, since the above set cannot be defined by finite automata operating on trees, as the ones used later in this article. ffl avoid useless processing of identical subterms by sharing them. This is used in most interpreters or compilers under the name of an environment binding the variables. For example, may be a very large expression. the structure used by most implementations of logic programming languages. For example, the infinite list of zeros can be represented by the constraint interpreted in the domain of lists of natural numbers. ffl express search strategies, as in automated deduction, allowing to prune further the search space. The need for more complex combined constraints involving several domains of interpretation arises in many situations. For example, integers are often used to describe parameterized families of terms, by indicating the repetition of a given subexpression. Constraints operating on such terms are then made of two compo- nents, an integer constraint, and a symbolic constraint. Another example, that we will work out in more detail in the sequel, is provided by bibliographic data bases. Querying the data base for all titles containing a given word can be expressed as an entailment problem over a language combining feature constraints for expressing record definitions, with word constraints for expressing the search of a given word. Our goal in this paper is to describe the different methods on which constraint solving algorithms are based. We distinguish among: (complete) syntactic methods, which do not commit to any particular representation of the problem domain or the set of solutions; (complete) semantic methods, which rely on a particular representation of the problem domain as well as of the set of solutions; hybrid methods which combine (incomplete) syntactic or semantic steps together with non-deterministic steps based on a partial enumeration of the set of solutions. These three methods are described respectively in sections 2, 3, 4. Applications areas are then considered in section 5. 2. Syntactic Transformations 2.1. The general setting Syntactic methods are based on a very simple principle: to repeatedly transform a constraint into an equivalent one until a so-called solved form is obtained. Most computations can be seen as such a rewriting process. However, we insist here on the fact that our transformation rules are applied on the formulas themselves, as A METHODOLOGICAL VIEW OF CONSTRAINT SOLVING 5 opposed to some coding of them: the rules simply match the (sub)formulas against some given patterns and replace them with simpler formulas. We will see some techniques which do not fall in this category in the next two sections. Assume that I have twice the age you had when I had the age you have. Since I'm not 48, and you are not 38, how old are we? This can be expressed by the following constraint over rational numbers: The transformation proceeds as you expect, by choosing a variable, say x and replacing it within the other equations, yielding the new equivalent system, that is having the same set of solutions: The system is then simplified according to elementary rational number theory, resulting in: We now proceed by eliminating t together with its existential quantifier: which is in solved form. Hence, there are solutions, such as, e.g. Figure 1. What is their age? Figure these concepts by developping an example of solving linear constraints over rationals. This technique is attributed to Gauss, the celebrated German mathematician of the 19th century, but was actually in use before that. A different problem is obtained by using inequalities instead of equalities. This problem can be reduced to the first by adding the so-called slack variables which express the difference between the two sides of the equation. For example, the constraint x - y is equivalent to the constraint 9z x This is of course another formula transformation. With linear programming comes a more difficult problem, in which a linear function must be optimized with respect to the set of solutions of a constraint of the above form. An example is described in figure 2, with the general principles of Dantzig's simplex algorithm [14]. Although exponential, it is often prefered to more recent polynomial algorithms such as the "interior point methods" [32] for two main reasons: it performs quasi linearly in the average, and has an efficient incremental version. Our examples show the ingredients of the transformation method. The constraint is a logical formula. Each rule transforms a constraint into an equivalent one, the 6 H. COMON, M. DINCBAS, J.-P. JOUANNAUD, C. KIRCHNER Vincent Mongeard-Mugneret, wine maker in Burgundy, just bought 20 acres of good soil at Vosne Roman'ee, planted with red pinot. He has two options, which cannot be changed once they are chosen. By bringing manure from his friend's farm, his investment per acre-year will be equal to $1000, and will rise up to $3000 if he uses chemical fertilizer. One acre of the first kind will produce 1000 bottles a year sold $25 dollars each, while the second will produce 50% more sold $20 each. His investment for the next 5 years is limited to $200000. How many acres should be fertilized with chemicals, and how many with manure in order to maximize the profit? Let us take x 1 for the number of acres cultivated with manure, x 2 for the number of acres cultivated with chemicals, z 1 (number of non-cultivated acres) and z 2 (unused are the slack variables. All variables are positive.! The process repeats the following steps: (i) choose one distinct variable per equa- tion, (ii) solve the system with respect to the choosen variables, (iii) replace the left-hand side variables in the profit function. The choice of the variables at step (i) must ensure that the right-hand side constants in the equations obtained at step (ii) are all positive, and the constant in the profit function obtained at step (iii) has not decreased.! Pivoting on x 1 and x 2 yields! The profit function is clearly maximized by taking z since their coefficients are negative. This yields a profit of $255000 1 . Figure 2. How good is your Vosne Roman'ee? A METHODOLOGICAL VIEW OF CONSTRAINT SOLVING 7 equivalence proof relying on some algebraic property of the constraint domain. Then each rule comes with some conditions on its application, which is supposed to ensure the termination of the overall transformation process. The rules are not necessarily deterministic: it may happen that several rules can be applied to the same constraint, or it may happen that two distinct instances of the same rule can be applied at the same time. For example, solving systems of linear equations, at each step we may choose to eliminate any (unsolved) variable, while keeping the termination property. This non-determinacy is quite different from e.g. non-determinism in PRLOG programs: this is a "don't care'' non-determinism. We can commit to any choice, without taking care of other possible transformations. Then, an actual implementation has to consider in addition a strategy which restores the determinism. For instance, in the case of linear equation solving, the strategy will specify an ordering on the pivoting variables. This distinction between the strategy and the rules themselves is important as it allows to reuse the same set of rules (and the same termination proof) together with distinct strategies, which yield efficient implementations depending on the particular properties of the constraints. For example, the ordering on variables to be eliminated may depend on the form of the system. Finally, a particular set of constraints called solved forms has to be specified. Such solved forms are supposed to have some additional properties. For example, in the case of a system of linear equations, a solved form is a system of equations such that every left member is a variable not occurring anywhere else. Strategies should be designed in such a way that they are complete, i.e. every irreducible constraint should be in solved form. Formalizing this approach, the following steps are necessary: 1. Choose a set of solved forms, that is a set of formulae that have no solution if and only if they are syntactically equal to the formula ?, standing for the empty set. 2. Design a set of transformation rules and show that it has the following properties (A) Correctness: applying a rule from the set to an arbitrary input constraint results in a new constraint which has the same set of solutions. Proving correctness is usually manageable since it breaks down into elementary correctness proof for each transformation rule. given an arbitrary constraint, every sequence of transformation originating from it results in an irreducible constraint after a finite number of steps. This property may be hard to prove. 3. Design a strategy which yields a deterministic set of rules and having the following property: (C) Completeness: there are enough rules, that is, for each constraint which is not a solved form, at least one rule applies. Therefore, the 8 H. COMON, M. DINCBAS, J.-P. JOUANNAUD, C. KIRCHNER algorithm never fails to detect that a given constraint has no solution. This property is usually simple to prove and easy to enforce. As part of its conceptual simplicity, an important advantage of this approach is to provide a systematic guide for constraints solving. When it applies, this method yields an algorithm which is inherently incremental, since new constraints can always be added (by using conjunction) to a solved form, yielding a new constraint equivalent to the conjunction of the starting constraint and the new one. 2.2. Unification and its extensions The approach described in the previous section was indeed successfully used for a variety of problems, starting with J. Herbrand [24] who was interested in solving equations over terms, for automated deduction purposes. This now classical abstract formulation of the well known unification algorithm was rediscovered in the late 70's by Martelli and Montanari [37], who formulated it from a more operational point of view. The presentation of the unification transformation rules given in figure 3, is inspired from [6] and [28]. This general approach has been applied in a systematic way since the beginning of the eighties in two different directions which both show the power of the to reformulate existing constraint solving algorithm; to generalize the unification algorithm in a number of directions. Let us mention now some of these generalizations. Equational unification consists in solving equations over terms when the function symbols considered satisfy certain equational axioms. The problem here is that using the Decompose rule is no longer complete. For example, if the head symbol f in the rule is commutative (i.e. the order of its two arguments is irrelevant), then the equation should be splited in the two possibilities: either s Although it is still correct in this particular case, the Check rule is no more correct in general. Many investigations in equational unification were motivated by the practical need of building equational knowledge in the resolution inference rule of theorem provers. This is the case, for example, of commutativity and associativity which occurs in many useful algebraic structures, or distributivity (unification modulo distributivity has been shown recently to be decidable [45]). But automated deduction is not the only application area where these techniques are eventually used. Let us mention among many: - a theory arising in the context of library search, related to the axiomatization of cartesian closed categories; - a theory originating from type inference for records in object oriented functionnal languages; - the theory of Boolean rings, which study was motivated by harware verification problems. Roughly speaking, unification constraints are formulae where quantifier and negation do not appear. Quantifiers occur of course naturally in the expression of various problems, hence more general constraints arise in many applications. Arbitrary first-order constraints built upon the equality predicate interpreted over terms are called equational constraints [7]. For example, a method for building A METHODOLOGICAL VIEW OF CONSTRAINT SOLVING 9 The unification transformation rules are parameterized by the vocabulary used for building expressions, called terms. We use f and g for arbitrary function sym- bols, x and y for variables, and s and t for arbitrary terms. Conjunctions of equations between terms are called unification constraints. Solved forms are chosen to be assignements of terms to variables, that is constraints of the form in which any of the variables x i occurs exactly once (see [28] for other choices). To obtain a solved form from an arbitrary unification constraint, several transformation rules are needed, among which two are most im- portant: decomposition simplifies an equation whose left and right hand side terms are rooted by the same function symbol; elimination propagates when necessary the value of a variable to the rest of the unification constraint. The whole set of transformation rules for unification is the following: Delete Decompose Conflict Check if x occurs in the non-variable term s Eliminate which x is replaced by s if x does not occur in s and occurs in P if x is a variable, s is not a variable In the above rules, the conjunction & is supposed to be associative and commuta- tive, allowing us to single out any equation from the constraint. As an example of use, the unification constraint f(x; g(a; can be transformed as follows: which is in solved form. The associated assignment called a most general unifier of the starting unification constraint. Any other assigment solution of the constraint can indeed be obtained from a most general unifier by an appropriate specialization. Figure 3. Unification constraints counter examples in theorem proving uses equational constraints with quantifiers and negations [4]. Jouannaud and Kounalis method for inductive theorem proving [30] uses equational constraints with universal quantifiers and negation (see [7]). Equational constraints also appear in learning from examples and counter-examples as described by Lassez and Marriott [35]. Equational constraints can be reduced to a solved form, as shown independenlty by Comon and Maher [11, 36]. These solved forms are purely existential, hence quantifier elimination rules become necessary. Although there is still a limited amount of negations in these solved forms, there are also negation elimination rules. And indeed, it has been an important problem in this area, whether negations could always be eliminated when an equivalent negation solved form existed. A positive answer has been given by Tajine [48] to this important problem which has applications to compilation of pattern-matching def- initions, and also to inductive inference from both examples and counter-examples. Of course, some operators may again satisfy certain axioms, and again associativity and commutativity are important for applications. Unfortunately, it is undecidable to know whether an arbitrary equational constraint has solutions in this particular case [52], and this is true as well for most algebraic structures. 2.3. Some other constraint systems There are other important predicates in practice. Term orderings, for example, surface naturally in mechanizing deduction since they allow reducing the search space by eliminating those inferences which do not satisfy certain ordering conditions (see [9, 31, 41, 12] among others for term ordering constraint solving). Membership predicates are important in typed languages where they are interpreted as set membership (see e.g. [10] for constraint solving). Set predicates have been used for defining and computing partial interpretations of logic programs (see [23]. Again, the problem becomes more difficult, in general undecidable, when some operators additionally satisfy certain equational axioms. First-order terms (that we have used up to now) lack the possibility to bind (possibly functional) variables as in higher-order logic and functional program- ming. Unfortunately, even unification constraints become undecidable in this set- ting. However, they are semi-decidable, and it turns out that Huet's semi decision procedure [25] is used in most proof development systems. An important decidable subcase, called pattern unification [39], is used in the language -PROLOG, a powerful extension of ordinary PROLOG. Function symbols have a fixed number of arguments. It is sometimes convenient to represent a term as a record, by using numeric keys for retrieving subterms, as in f(1 expressions have been used for quite some time in natural language processing [44], and more generally in knowledge representation, under the name of features. They have also been used together with subtyping mechanims in the area of programming languages and data bases to model inheritance [?]. The figure 4 describes a possible syntax for feature constraints, and suggests how the unification rules may be adapted to this case. The area of feature constraints has been very active in the last 10 years, and resulted in a complete A METHODOLOGICAL VIEW OF CONSTRAINT SOLVING 11 constraint solver for first-order features on one hand [3, 53], and in many prototype implementations of feature-based languages on the other hand [2, 46]. We refer to [53] for more information on feature constraints. Finally let us note that most of the efforts in symbolic constraint solving are surveyed in [28, 7] See also [29, 42] for more recent works. 3. Semantic Methods Semantic methods, as opposed to syntactic ones, do not operate directly on the constraint syntax. Instead, they use another representation of the constraint and of its solutions, so that constraint solving relies then on the use of a specific data structure. 3.1. Automata techniques A typical example is the use of automata. The idea, which goes back to B-uchi in the early sixties, consists in associating with each formula defining a constraint an automaton recognizing the solutions of the constraint (see e.g. [49, 8]). An automaton is a very simple finite transition system which reads a string of input symbols sequentially, and upon each reading, moves from one state to another depending on its current state. A word is recognized if the automaton ends up in a distinguished final state. Once an automaton has been associated with each atomic constraint, the construction proceeds by replacing logical operations on constraints by set operations on automata: conjunction yields intersection, disjunction yields union, negation yields complement, existential quantification yields projection. This is so because logical operations on constraints correspond to set operations on their solution sets. For instance, the solutions of a conjunction of two constraints C 1 and C 2 are in the intersection set of the solutions of C 1 and C 2 respectively. Since automata are closed under set operations, the automaton associated with A 1 and A 2 accept the solutions of C 1 and C 2 respectively. Projection has to be explained a little bit further: the automata have to recognize tuples with as many elements as the number of free variables in the constraint. Projection consists in forgetting one of the components of the tuple, while keeping the same control in the automaton. Once an automaton A is associated with the constraint C, the satisfaction problem for C is equivalent to the emptiness decision problem for A that is: C has at least a solution iff A accepts at least one word (or tree or graph). The automaton A is then a representation of the set of solutions. It can be cleaned, which corresponds to reduce it to a solved form, and reused for further computations: this method is incremental. We did not precise so far on which objects the automata are working, which depends on the representation of the constraints domain values by words, trees or graphs. The idea is the same in each case: the class of automata has to possess several closure properties and emptiness should be decidable. Several automata 12 H. COMON, M. DINCBAS, J.-P. JOUANNAUD, C. KIRCHNER Features are constructed from three given sets, a S set of sorts (also called basic types), a set K of keys (also called features), and a set X of sort variables. The set F of feature expressions (or terms) follows the syntax given in the form of the following For example the feature term person(age is meant to represent the set of all persons whose age is the natural number X . For our purpose, feature constraints will be conjunctions of equalities between feature terms. For example, one can specify as a constraint the set of all persons who have blue eyes, and are married with a person of the same age: Unlike ordinary trees, feature terms may have an arbitrary number of subterms for each sort name. Still, feature terms unification looks pretty much like ordinary unification. Let us give the decomposition and conflicting rules only: Decompose x(Key KeyConflict if Key 1 and Key 2 are different SortConflict if s and s 0 are different sort names In these rules, x; stand for constants or variables of the appropriate categories. For example, x may be the sort person and y the variable X above. As an example of use, the constraint: X(age Decompose confines X to describe a set X of persons having the same age. Figure 4. Feature constraints A METHODOLOGICAL VIEW OF CONSTRAINT SOLVING Once we read a 0, then we know that the number (which is written in base 2) is even. Hence we enter a final state marked with a double circle. Then, whatever we read, the number will be even and we stay in that final state. Figure 5. An automaton accepting even numbers have been designed in the literature, each solving some particular kind of constraint. Let us mention the classical finite state automata on words, the finite tree automata, the automata with equality and disequality constraints, the tree automata with free variables, the tree matching automata,. All possess the appropriate closure properties by set operations (see e.g. [8] for more details). The efficiency of automata techniques will follow directly from the efficiency of operations on automata and emptiness decision. To get an idea of the computational complexity, let us recall that, for classical word automata, union is computed in constant time, intersection in quadratic time, projection in linear time and complement in linear time for deterministic automata and exponential time for non-deterministic ones. Emptiness decision is linear. This behaviour can be found again in most automata classes: determinization and complement are usually the most expensive steps. Let us give a few examples of constraint solving using automata. Presburger arithmetic consists in first-order formulae over the atomic formulae where s; t are built using constants (0; addition and multiplication by a constant. The interpretation domain is the set of natural numbers. For instance, defines the set of even numbers. Assume now that the natural numbers are written in base two, from right to left. For example, the number thirteen will be written 1011. With each formula can be associated a word automaton. For example, 9x associated with the automaton of figure 5. In general, the construction might be a bit more complicated because some formulae may have In such a case, the n-uples of numbers are encoded as words over f0; 1g n . For example, the pair (thirteen, four) will be represented by the word1011 : reading from right to left the lower word we find thirteen in base 2 and reading from right to left the upper word we find four. The height of the stack of numbers is the number of free variables. With such a convention, it is possible to build the automaton accepting the pairs of numbers which satisfy and the set of triples x; y; z of numbers 14 H. COMON, M. DINCBAS, J.-P. JOUANNAUD, C. KIRCHNER The only possible transitions are those which are labeled with pairs of identical symbols. Figure 6. An automaton for the formula 1The two states correspond to "no carry" (the final state) and "carry" the non-final state. Initially we enter without carry. Since reading the triples1 1 , we stay in the final state. the transition by1 0 to the state carry. All other transitions are built in the same way. Figure 7. An automaton for the formula such that z (see figure 7). Now, combining these two automata, by intersection and projection (this last operation simply forgets a component of the tuples), we get back the automaton of figure 5. Automata techniques have been used for a large variety of constraints, but few of them are really used in actual software, either because of their high computational complexity or because of their relatively recent discovery. A simple example is the use of tree automata in membership constraints, which express typing properties; the formulae consist in combinations of atomic formulae t 2 i where i is recognized A METHODOLOGICAL VIEW OF CONSTRAINT SOLVING 15 by a finite tree automaton. For example, we might want to restrict to integers the values that a given variable x can take. Assuming that we have the function int, . We can express the typing condition using a constraint x 2 int where int is interpreted as the set of trees recognized in the state int by the above specified automaton. This will include is a real value, but not t E(t). Such constraints are used, e.g. in the constraint concurrent logic programming language Oz [47], developed at DFKI in Saarbr-ucken. Typing constraints can also be inferred at compile time in a logic programming language, taking advantage of the structure of the program [19]. Then automata techniques can be devoted to this particular piece of program, which yields more efficient execution. The first-order theory of typing (also called membership) constraints has been proved decidable by H. Comon and C. Delor [10]. Related to typing constraints, set constraints have been introduced by Mishra in 84 [40] in the area of program analysis. Set constraints are combinations of formulae are set expressions using e.g. intersection, union, complement, application of a function symbol, etc. They are interpreted as (possibly infinite) subsets of the set of all terms. The contribution of the French school in this area was to demonstrate that some kind of tree automata (tree automata with free are adequate for the representation of solutions and for constraint solving [15], [22, 21, 51]. Set constraints can also be used as part of a constraint logic programming language and there are other techniques than tree automata to solve them [34, 5] 3.2. Other algebraic representations Besides automata techniques there are other (complete) constraints methods which extensively use the representation of the constraints. A first example is equational unification as already introduced in section 2, and we will consider here the case where there are associative and commutative function symbols. An alternative to the syntactic methods already described is to associate a semi-ring 2 with the equational axioms, and then to reduce unification constraints to the solving of equations in the semi-ring. Let us also mention an old semantic algorithm by L-owenheim for Boolean ring unification which has been rediscovered by Martin and Nipkow [38]. Another example of importance concerns the linear diophantine equations that consist of linear polynoms with integer coefficients. They are used in particular in another constraint solving problem, which again uses semantic methods: solving equations between terms in presence of associative and commutative function symbols. A typical example of diophantine equations is the system: ae where we are searching only the positive integer solutions. We have seen already that this can be solved by means of automata. However, in this particular situation, specific methods can be more efficient. The idea of Contejean and Devie [13] is to represent the solutions as vectors in the real space of dimension n (the number of @ @ @ @R \Theta \Theta \Theta \Theta \Theta \Theta \Gamma\Psi Figure 8. vectors equations, 2 in our example). The constraint solving algorithm then consists of starting from the origin and repeat the addition of one of the default vectors until we reach back the origin. Default vectors are computed from the equation system as the values of the two left members of the equations on the vector basis. In our example, the default vectors are represented on figure 8. After the first step, the addition of default vectors d to the current vector v is only considered when the scalar product d \Delta v is negative. Then the process is shown to be always terminating and complete, in the sense that every solution of the system will be a combination of the solutions obtained in this way. For example, figure 9 shows how the solution (4,2,1,0) is reached by the algorithm. 4 is the number of times we have added the default vector a(e 1 ), 2 is the number of times we have added the default vector the number of times we have added the default vector a(e 3 ), 0 is the number of times we have added the default vector a(e 4 ). Other semantic methods have been proposed for linear Diophantine equations solving; see e.g. [1, 18, 17, 43, 50]. 4. Hybrid Methods As we have seen in the previous sections, solving a constraint consist in finding some "canonical representation" of it. It is canonical in the sense that all unsatisfiable constraints have a unique representation. This corresponds to what we call complete constraint solvers. The completeness is not really required in many sit- uations. Consider for instance applications of constraints in automated deduction or logic programming. Constraints allow to represent (in an efficient way) sets of A METHODOLOGICAL VIEW OF CONSTRAINT SOLVING 17 @@ @@ @@ @@ @ @ @ @ @ @- O @ @ @ @ @ @ @ @ @ @ @R \Theta \Theta \Theta \Theta \Theta \Theta \Theta \Theta \Theta \Theta \Theta \Theta \Theta \Theta \Theta \Thetaffi \Gamma\Gamma\Psi \Gamma\Psi \Theta \Theta \Theta \Theta \Theta \Theta \Theta \Theta \Theta \Theta \Theta \Theta \Theta \Theta \Theta \Thetaffi @ @ @ @ \Gamma\Psi @ @ @ @ @ @ @ @ Figure 9. Computation of the solution (4,2,1,0) of the system formulas. Then inferences take constrained formulas as premisses and yield constrained formulas as conclusion. As long as the proof is not completed, or as soon as a contradiction is not found (the empty clause in logic programming), it does not matter for the deduction process whether or not the constraint of a formula is satisfiable or not. Of course, not checking for satisfiability may yield a huge number of junk formulas. But, on the other hand, checking satisfiability can be quite expensive in time. That is why in most constraint logic programming languages or constraint automated deduction systems do not solve the constraints eagerly; there is a trade-off between time efficiency and space efficiency. Usually some constraint simplifications are performed eagerly and the complete constraint solving is performed only when the empty clause is generated (or, more generally when a proof is obtained). As a typical example, the incomplete solver checks, at each step of the computa- tion, a relaxed version of the set of constraints, an idea quite common in Operations Research, where the name of relaxation comes from. In practice, rather than checking a conjunction of a large number n of elementary constraints, a possibility is to check all possible combinations of conjunctions of two elementary constraints, a technique known as local consistency. Since local consistency does not imply the existence of solutions, this technique is usually used in combination with an enumeration procedure operating on locally consistent constraints in order to restore the completeness of the whole constraint solving process. Enumeration should be taken here in a broad sense: relations as well as values may be enumerated. A hybrid method is therefore a constraint solving technique which comes in two parts: one syntactic constraint solving technique, which is incomplete in general, and a (syntactic or semantic) technique which completes the first part of the constraint solving. While the first simplification is performed eagerly, the second part of the constraint solving is performed only when necessary. Hybrid methods are actually used in all applications of constraints in automated deduction and logic programming, as soon as the constraint solving is expensive. In logic programming, a typical example is the finite domain constraints. In automated deduction, this is the case of ordering constraints [20]. Combination mechanisms can also be seen as an instance of this scheme. Let us consider now these constraints in more detail. 4.1. Finite domains Finite domain constraints are described in figure 10, together with a constraint solving method based on local consistency checking. These techniques are at the root of the success of Constraint Programming environments like CHIP [?] or Ilog Solver [?] for solving combinatorial problems. In practice, the efficiency of these hybrid methods for finite domains heavily depends on the enumeration procedure for which various heuristics can be used. Although some of them work pretty well, such as the first-fail principle for variables and domain-splitting for values in the context of finite domains constraints, there is no best strategy applicable for all types of problems. Finite domain constraints as they are defined in figure 10 do not allow to easily specify and solve complex problems in areas like planning, scheduling, packing and placement. To this end, new predicates (like 6=) and logical connectives (like negation and disjunction) can be introduced to the price of efficiency problems. More- over, these particular predicates are too low-level to allow modelling problems in an easy and natural way. An alternative, first explored in CHIP with the cumulative constraint, is to introduce the so called global constraints to reach more expressivity by means of high-level abstractions. The cumulative constraint generalises the disjunctive constraints in order to model finite capacity scheduling problems for which there may be several copies of each given kind of resource or these resources can be shared. The second goal of these global constraints is to take advantage of the interactions among different constraints at runtime in order to reach a better pruning of the search tree. Instead of just using values or bound propagation as in the previous method, global constraints take into account structural properties of A METHODOLOGICAL VIEW OF CONSTRAINT SOLVING 19 We consider here a structure FD whose domain is the set N of natural numbers, hence variables will range over N . In this setting, Finite domain constraints are existential positive formulae built up with the five predicates =; ?; - interpreted in N , and infinitely many membership predicates 2 [a b], one for each finite interval [a b] of N . The precise syntax is given by the following grammar: where x denotes a variable ranging over N , a and b natural numbers. The atomic constraint x 2 [a b] is interpreted by the empty set when b ! a. The atomic constraints a and x 2 [a a] are identified. Solving finite domain constraints is both NP-complete and very important for practice, which has therefore favored the use of a practically efficient technique, constraint propagation, described here in the form of a set of transformation rules: each atomic ordering or equality constraint C between the variables in whose domains are defined by membership constraints x new restrictions on these domains, resulting in new membership constraints. We call reduced domain RD(x j ; C) of the variable x j for C, the smallest interval [a 0 in N such that the constraint obtained by substituting x j by m in C & (& i=n satisfiable in FD. Forward Checking x 2 [a Look Ahead x 2 [a Eliminate ag Falsity Enumerate x 2 [a ag or x 2 [a if no other rules applies Using the Forward Checking rule is easy: since x is its only variable, C can be solved which modifies the original domain of x and allows to eliminate C. There are several uses of the Look Ahead rule whose principle is to lift the constraint from the variables to their domain. For example, the constraint cx 1 ! dx 2 +e entails the relation ca 1 ! db 2 +e. We can therefore choose for b 0 2 the largest natural number m such that cm ! db Consider now the constraint: Continued on Figure 11 Figure 10. Finite domains: 1/2 20 H. COMON, M. DINCBAS, J.-P. JOUANNAUD, C. KIRCHNER Continued from Figure 10 Now, we enumerate, and we obtain first (the or must be understood as a non-deterministic which again needs applying Enumerate until we get the solution: 5: We can see the crucial role of enumerations, and it is the case here that enumerating from the largest values would speed up the process. A popular schema, Domain splitting, is similar to binary search. Figure 11. Finite domains: 2/2 the constraints in order to deduce and propagate more information. This is usually done via a semantic representation of the constraints based on graphs. While simple arithmetic and logic knowledge is used in the former methods, deeper finite mathematics, graph theory and operations research knowledge is required and used in the latter. Other global constraints, like "diffn", a generalization of disequality constraint to n-dimensional objects, have been recently introduced in CHIP [?] which permitted to solve quite difficult placement problems. 4.2. Ordering constraints Symbolic constraints also may use enumerations in the form of a domain splitting rule since the Herbrand domain is usually infinite. This is the case with ordering constraints, conjunctions of atomic constraints of the form s ? t or t. They have been studied by Comon [9], Jouannaud and Okada [31], and Nieuwenhuis [41]. In this context, simplification rules are obtained by expressing the recursive definition of the ordering on terms as a set of transformation rules operating on atomic constraints, while enumeration rules linearize the constraint considering all possible orderings between the componebts. SATURATE, a system developed by Nieuwenhuis and Ganzinger [20], implements constraints solving mechanism of this form, with a lazy enumeration rule. 4.3. Combination techniques Another kind of hybrid method arises when an application uses complex constraints operating on several computational domains, each of them being supposed to have its own syntax and constraint solving method. Such problems are called combina- A METHODOLOGICAL VIEW OF CONSTRAINT SOLVING 21 tion problems. In this case, the constraint syntax is of course built with different pieces coming from the different constraints involved in the combination. And, of course, it is desirable to modularly derive a constraint solving algorithm for the combination from the algorithms already known for the elementary constraints. It turns out that this problem was thoroughly investigated for the unification case, since it was actually the main difficulty of associative commutative unification: when there are associative commutative function symbols, the problem can be seen as a combination of several unification algorithms, one for each of these symbols. Querying a bibliographic data base for articles satisfying certain constraints is an interesting practical example of application for this kind of techniques. Assume each item in the data base records information about a particular article, the au- thors, the title, the affiliation, etc., as in the bibtex format. Each item can be represented by a so called feature term, and the whole (finite) data base becomes a (big) conjunction \Phi of feature terms. Note that it will be easy to augment the data base, by adding a new element in the conjunction. Querying the data base for a particular entry, for example all papers containing the word "unification" in their title, will be expressed as an entailment problem of the form \Phi entails OE, if OE represents the query. The problem here is that the feature terms are not homogeneous, since they contain subexpressions which are strings over a certain vocabulary, which involves solving associative pattern matching constraints, see figure 12. Solving a constraint over an heterogeneous domain starts with a variable abstraction consisting in purifying the constraints: the goal is to get a conjunction of constraint systems each of which is homogeneous. Then each constraint solving algorithm can be applied to the pure parts. Still, interactions between the homogenous parts have to be considered. This can be done through an enumeration procedure which, roughly, guesses the shared variables dependencies. 5. Application Contexts Because constraints are a natural way to specify mathematical problems, there are a number of potential applications of the ideas presented in this paper. Numerical constraints are used in Operations Research, in Robotics, and more generally in the area of applied mathematics, physics and mechanical engineering. Symbolic constraints arise naturally in computer science applications like type verification and inference, abstract interpretations, logic programming, deduction, and artificial intelligence. We concentrate briefly now on two main applications of the tools presented in this paper: constraints for deduction and constraints for computations. The use of constraints in deduction has recently developed into a new promising field of research, of which constraint logic programming is a specific instance. Constraints are of particular importance in theorem proving for two main reasons. Constraints allow to (possibly exponentially) reduce the number of formulae resulting from a single deduction step, by packing them all in a constrained formula. For example, automated deduction techniques require solving equations between terms possibly involving associative and commutative operators, as subgoals of a given logical deduction. But, an equation as simple as x 22 H. COMON, M. DINCBAS, J.-P. JOUANNAUD, C. KIRCHNER We will use the syntax of features as defined in figure 4. Our sorts (or types) will be the elements set farticle; person; identity; university; addressg, augmented by "built-in types", list-of[], string and Nat. Using a set of keys that are clear from the context, here is the syntax of our basic types: authors adr postalcode name first last When implicit, we do not recall the constant sorts in a feature term. We now consider the query: title authors where the letters A, X, Y, Z, U denote variables of the appropriate types. The solutions of this query are all the values for A, X, Y, Z, U for which the constraint above is entailed by the data base expression. Of course, this entailment problem will cause an enumeration of all elementary formulas in the data base corresponding to the various bibliographical data, allowing then to resolve the entailment problem for each bibliographical datum in turn. Assume now that the bibliographical data base is the one at the end of this paper and thus contains the following entry: title ) An Efficient Unification Algorithm; authors last last In order to make the variables match the appropriate information, the constraint will first be split into homogeneous pieces, that is pieces of a given domain. Here, there are three domains, feature terms, strings, and natural numbers. We therefore obtain the new constraint: title Unification Algorithm authors authors [person(name last person(name last which in turn is decomposed in: by using an appropriate string matching algorithm. Figure 12. Bibliographic search A METHODOLOGICAL VIEW OF CONSTRAINT SOLVING 23 has over 30 billions of incomparable solutions. It is of course not realistic to engage current computers in such a number of further deductions. The remedy, keeping the equation as a constraint for the further deductions, was initiated in [33]. Constraints allow also to record information on the computation as a component of the deduced formulae, hence they permit to formulate inference rules which reason about this information for efficiently pruning the search space. Constraint solving techniques have found their first key application within the Constraint Logic Programming framework, especially for solving Combinatorial Search Problems [26]. As examples of these problems occurring in different economical areas, we can mention project management, production scheduling, crew assignment, tour planning, etc. Not only are these problems NP-hard in general, they are also hard to model and therefore hard to program. Constraint Logic Programming brings an ideal solution to this problem by supporting different types of constraint systems (symbolic and numeric), and by allowing the use of powerful constraint solving algorithms combined with domain-oriented heuristic search. Several successful industrial applications have been developed in these areas, which are described in this volume. An overview of Constraint Logic Programming can be found in [27], where are also described applications to circuit design (symbolic ver- ification, test pattern generation, logic synthesis, diagnosis) and decision problems in management (option trading analysis, portfolio management). Acknowledgments This work was partly supported by Vincent Mongeard-Mugneret, 21700 Vosne Ro- man'ee, France, Fax: 33-80-62-35-75 and by the ESPRIT BRA CCL. We thank Carlos Castro for his comments on earlier version of this paper. Notes 1. Vincent Mongeard-Mugneret actually prefers supreme quality to big profit, hence uses manure only. We recommand his Clos Vougeot Grand Cru. 88, 89 and 90 are extremely good years. Feel free to use the 3rd author's recommandation. 2. A semi-ring is a structure which has the same properties as a ring, except that there might be no inverse for addition. (N; +; \Theta; 0; 1) is a typical example of a semi-ring. --R General solutions of systems of linear diophantine equations and inequations. A complete and recursive feature theory. A method for simultaneous search for refutations and models by equational constraint solving. Some notes on rational spaces. Disunification: a survey. Tree automata techniques and applications. Solving symbolic ordering constraints. Equational formulae with membership constraints. Equational Problems and Disunification. Ordering constraints on trees. An efficient algorithm for solving systems of diophantine equations. Linear Programming and Extensions. Rewriting and tree automata. The constraint logic programming language CHIP. Outils pour la d'eduction automatique dans les th'eories associatives- commutatives A fast method for finding the basis of nonnegative solutions to a linear Diophantine equation. Logic programs as types for logic programs. The Saturate System Solving systems of set constraints with negated subset relationships. Solving systems of set constraints using tree automata. based program analysis. Recherches sur la th'eorie de la d'emonstration. R'esolution d Constraint logic programming. Constraint logic programming: A survey. Solving equations in abstract algebras: a rule-based survey of unification First International Conference on Constraints in Computational Logics Automatic proofs by induction in theories without constructors. Satisfiability of systems of ordinal notations with the subterm property is decidable. A new polynomial-time algorithm for linear programming Deduction with symbolic con- straints Set constraints in logic programming. Explicit representation of terms defined by counter examples. Complete axiomatizations of the algebras of finite An efficient unification algorithm. Boolean unification. A logic programming language with lambda-abstraction Towards a theory of types in prolog. Simple LPO constraint solving methods. Constraint Programming: Basics and Trends Minimal solutions of linear diophantine systems A complete logical calculus for record structures representing linguistic information. The oz programming model. A foundation for higher-order concurrent constraint programming The negation elimination from syntactic equational formulas is decidable. Automata on infinite objects. Solving linear diophantine equations using the geometric structure of the solution space. Automates et contraintes ensemblistes. A new method for undecidability proofs of first order theories. Feature trees over arbitrary structures. --TR
constraint solving;rule based solving;constraint satisfiability;unification
593311
Localizer.
Local search is a traditional technique to solve combinatorial search problems which has raised much interest in recent years. The design and implementation of local search algorithms is not an easy task in general and may require considerable experimentation and programming effort. However, contrary to global search, little support is available to assist the design and implementation of local search algorithms. This paper describes the design and implementation of Localizer, a modeling language for implementing local search algorithms. Localizer makes it possible to express local search algorithms in a notation close to their informal descriptions in scientific papers. Experimental results on Boolean satisfiability, graph coloring, graph partitioning, and job-shop scheduling show the feasibility of the approach.
Introduction Most combinatorial search problems are solved through global or local search. In global search, a problem is divided into subproblems until the subproblems are simple enough to be solved directly. In local search, an initial configuration is generated and the algorithm moves from the current configuration to a neighborhood configuration until a solution (deci- sion problems) or a good solution (optimization problems) has been found or the resources available are exhausted. The two approaches have complementary strengths, weaknesses, and application areas. The design of global search algorithms is now supported by a variety of tools, ranging from modeling languages such as AMPL [2] and Numerica [10] to constraint programming languages such CHIP, Ilog Solver, CLP(!), Prolog-IV, and Oz to name only a few. In contrast, little attention has been devoted to the support of local search, despite the increasing interest in these algorithms in recent years. (Note however there are various efforts to integrate local search in CLP languages, e.g., [9]). The design of local search algorithms is not an easy task however. The same problem can be modeled in many different ways (see for instance [4]), making the design process an inherently experimental enterprise. In addition, efficient implementations of local search algorithms often require maintaining complex data structures incrementally, which is a tedious and error-prone activity. Localizer [5] is a domain-specific language for the implementation of local search al- gorithms, combining aspects of declarative and imperative programming, since both are important in local search algorithms. Localizer makes it possible to write local search algorithms in a notation close to the informal presentation found in scientific publications, while inducing a reasonable overhead over special-purpose implementations. Localizer offers support for defining traditional concepts like neighborhoods, acceptance criteria, and restarting states. In addition, Localizer also introduces the declarative concept of invariants in order to automate the most tedious and error-prone aspect of local search procedures: incremental data structures. Invariants provide a declarative way to specify what needs to be maintained to define the neighborhood and the objective function. This paper is a progress report describing the status of Localizer as of February 1998. Its main focus is on the language and its implementation. 1 It is not intended as a final word on the language, since new, higher-level, extensions are currently under evaluation. The paper however describes the core of Localizer which will probably not evolve in significant ways. The paper is organized in four main parts. Section 2 gives readers a quick tour of Localizer. Section 3 describes the language in more detail. Section 4 describes the implementation of invariants which are the cornerstone of Localizer. Section 5 summarizes some experimental results from several applications. Finally, Section 6 concludes the paper. A Tour of Localizer This section gives an overview of the main features of Localizer. It starts by reviewing the computational model of Localizer and the general form of Localizer statements. It then considers the two main contributions of Localizer: invariants and neighborhoods. 2.1 The Computation Model To understand statements in Localizer, it is best to consider the underlying computational model first. Figure 1 depicts the computational model of Localizer for decision problems. The model captures the essence of most local search algorithms. The algorithm performs a number of local searches (up to MaxSearches and while a global condition is satisfied). Each local search consists of a number of iterations (up to MaxTrials and while a local condition is satisfied). For each iteration, the algorithm first tests if the state is satisfiable, in which case a solution has been found. Otherwise, it selects a candidate move in the neighborhood and moves to this new state if this is acceptable. If no solution is found after MaxTrials or when the local condition is false, the algorithm restarts a new local iteration in the state restartState(s). The computation model for optimization problems is similar, except that line 5 needs to update the best solution so far if necessary, e.g. in the case of a minimization, 5.1 bestBound := value(s); 5.2 best := s; companion paper for the Operations Research community is oriented around applications and modeling aspects. The optimization algorithm of course should initialize f properly and return the best solution found at the end of the computation. procedure Localizer begin 2 for search := 1 to MaxSearches while Global Condition do 3 for trial := 1 to MaxTrials while Local Condition do 5 return s; 6 select n in neighborhood(s); 7 if acceptable(n) then 9 s := restartState(s); Figure 1: The Computation Model of Localizer 2.2 The Structure of Localizer Statements The purpose of a Localizer statement is to specify, for the problem at hand, the instance data, the state, and the generic parts of the computation model (e.g., the neighborhood and the acceptance criterion). A Localizer statement consists of a number of sections as depicted in Figure 2. The instance data is defined by the Type, Constant, and Init sections, using traditional data structures from programming languages. The state is defined as the values of the variables. The neighborhood is defined in the Neighborhood section, using objects from previous sections. The acceptance criterion is part of the definition of the neighborhood. The initial state is defined in section Start. The restarting states are defined in section Restart, the parameters (e.g. MaxTrials) are given in the Parameter section, and the global and local conditions are given in sections Global Condition and Local Condition. Note that all the identifiers in boldface in the description of computation model (e.g., search and trial), are in fact keywords of Localizer. As mentioned previously, the most original aspects of Localizer are in the specifications of the neighborhood and the acceptance criterion. Of course, some of the notations are reminiscent of languages such as AMPL and Claire at the syntactical level but the underlying concepts are fundamentally different. In the rest of this section, we describe the most original aspects of Localizer without trying to be comprehensive. 2.3 The Running Example This overview mostly uses Boolean satisfiability to illustrate the concepts of Localizer. A Boolean satisfiability problem amounts to finding a truth assignment for a first-order boolean formula expressed in disjunctive normal form. The input is given as a set of clauses, Model i ::= [solve j optimize] hObjective Functioni] Figure 2: The Structure of Localizer Statements each clause consisting of a number of positive and negative literals. As is traditional, a literal is simply an atom (positive atom) or the negation of an atom (negative atom). A clause is satisfied as soon as at least one of its positive atoms is true or at least one of its negative atoms is false. The local search model considered for Boolean satisfiability is based on the GSAT algorithm by Selman et al. in [8], where the local search moves consist of flipping the truth value of an atom. A local improvement model for Boolean satisfiability is described in Figure 3. In the model, atoms are represented by integers 1 to n and a clause is represented by two sets: the set of its positive atoms p and the set of its negative atoms n. This data representation is specified in the Type section. A problem instance is specified by an array of m clauses over n variables. The instance data is declared in the Constant section and initialized in the Init section which is not shown. The state is specified by the truth values of the atoms and is captured in the array a of variables in the Variable section. Variable a[i] represents the truth value of atom i. The Invariant section is the key component of all Localizer statements: it describes, in a declarative way, the data structures which must be maintained incrementally. Invariants are reviewed in detail in Section 2.4. In the model depicted in Figure 3, they maintain the number of true literals nbtl[c] in each clause c and the number of satisfied clauses nbClauseSat. The Satisfiable section describes when the state is a solution (all clauses are satisfied), while the Objective Function section describes the objective function (maximize the number of satisfied clauses) used to drive the search. The Neighborhood section describes the actual neighborhood and the acceptance criterion. The neighborhood consists of all the states which can be obtained by flipping the truth value of an atom and a move is accepted if it improves the value of the objective function. The Neighborhood section is another important part of Localizer and is reviewed in Solve Type: Constant: cl: array[1.m] of clause = .; Variable: a: array[1.n] of boolean; Invariant: in 1.m] of int = sum(i in cl[i].p) a[j] + sum(j in cl[i]:n) !a[j]; Satisfiable: Objective Function: maximize nbClauseSat; Neighborhood: move a[i] := !a[i] where i from f1.ng accept when improvement; Start: forall(i in 1.n) a[i] := random(ftrue,falseg); Restart: forall(i in 1.n) a[i] := random(ftrue,falseg); Figure 3: A Local Improvement Model for Boolean Satisfiability more detail in Section 2.5. The Start and Restart sections describe how to generate an initial state and a new state when restarting a search. They both use a simple random generation in the model. It is interesting at this point to stress the simplicity of the model, since it is difficult to imagine a more concise formal statement of the algorithm. 2.4 Invariants Invariants are probably the most important tool offered by Localizer to support the design of local search algorithms. They make it possible to specify what needs to be maintained incrementally without considering how to do so. Informally speaking, an invariant is an expression of the guarantees that, at any time during the computation, the value of variable v of type t is the value of the expression exp (also of type t of course). For instance, the invariant in 1.m] of int = sum(i in cl[i].p) a[j] + sum(j in cl[i]:n) !a[j]; in the Boolean satisfiablity model specifies that nbtl[c] is equal to the sum of all true positive atoms and all false negative atoms in clause c, for all clauses in 1::n. Localizer uses efficient incremental algorithms to maintain these invariants during the computation, automating one of the tedious and time-consuming tasks of local search algorithms. For instance, whenever a value a[k] is changed, nbtl[c] is updated in constant time. Localizer allows a wide variety of invariants over complex data structures. The invariant (also from the Boolean satisfiability model) illustrates the use of relations inside an invariant. A relation, when used inside an expression, is considered a 0-1 integer, i.e., the relation evaluated to 1 when true and 0 otherwise. The excerpt C: array[1.n] of int = distribute(x,1::n,1.n); select i from 1.n where select i from 1.n where size(C[i]) ? 0g; unused Candidates B: array[k in 1.n] of select i from C[k] & select j from C[k] where A[i; j] is taken from a graph-coloring model implementing an algorithm in [4]. The graph-coloring problem amounts to finding the smallest number of colors to label a graph such that that adjacent vertices have different colors. For a graph with n vertices, the algorithm considers n colors which are the integers between 1 and n. Color class C i is the set of all vertices colored with i and the bad edges of C i , denoted by B i , are the edges whose vertices are both colored with i. The main idea of the algorithm is to minimize the objective function whose local minima are valid colorings. To minimize the function, the algorithm chooses a vertex and chooses a color whose color class is non-empty or one of the unused colors. It is important to consider only one of the unused colors to avoid a bias towards unused colors. The invariant B: array[k in 1.n] of select i from C[k] & select j from C[k] where Adj[i; j] that B[k] is the set of edges obtained by selecting two adjacent vertices in color class k. It illustrates that Localizer can maintain queries over sets varying in time (since evolves during the local search). The invariant C: array[1.n] of int = distribute(x,1::n,1.n); is equivalent to, but more efficient than, C: array[i in 1.n] of int select i from 1.n where This primitive function is provided, since it is useful in a large variety of applications. The invariant select i from 1.n where size(C[i]) ? 0g; defines the non-empty classes. Note that here the set 1.n does not vary but the condition Once again, it is important to emphasize the significant support provided by Localizer with invariants. These invariants maintain complex data structures incrementally, but users only have to specify them in a declarative way. 2.5 The Neighborhood Many strategies such as local improvement, simulated annealing, and tabu search have been proposed in the last decades for local search algorithms. This section reviews how they are modeled in the neighborhood section, which is the other fundamental tool provided by Localizer. 2.5.1 Local Improvement and Local Nondegradation The model depicted in Figure 3 uses a stochastic local improvement approach. The neighborhood section Neighborhood: move a[i] := !a[i] where i from f1.ng accept when improvement; specifies the following strategy: select a value i in 1::n (i.e., select an atom), flip a[i], and, if the resulting state improves the value of the objective function, take the move. If the state does not improve the value of the objective function, the move is not taken and Localizer proceeds to the next iteration of the innermost loop in the computational model. This strategy illustrates the structure of the neighborhood definition in Localizer. The move part specifies a state transformation: it uses traditional imperative constructs to show how to transform a state into another state. The where part specifies the set of objects used to specify the state transformation. The accept part describes when to accept the move. The local improvement strategy can be made greedy by adding the keyword best in front of the move instruction, as in Neighborhood: best move a[i] := !a[i] where i from f1.ng accept when improvement; Solve Type: Constant: cl: array[1.m] of clause = .; Variable: a: array[1.n] of boolean; Invariant: in 1.m] of int = sum(i in cl[i].p) a[j] + sum(j in cl[i]:n) !a[j]; Satisfiable: Objective Function: maximize nbClauseSat; Neighborhood: best move a[i] := !a[i] where i from f1.ng accept when noDecrease; Start: forall(i in 1.n) a[i] := random(ftrue,falseg); Restart: forall(i in 1.n) a[i] := random(ftrue,falseg); Figure 4: A GSAT-based Model for Boolean Satisfiability This excerpt specifies the following strategy: consider each value i in 1::n which, when flipped, produces an improvement and select the one with the best improvement. Note that this strategy explores the neighborhood in a systematic way, while the previous strategy was selecting a random move and testing it for improvement. The neighborhood section Neighborhood: first move a[i] := !a[i] where i from f1.ng accept when improvement; is another approach that explores the neighborhood systematically until a move improving the value of the objective function is found. It should be contrasted with the random walk strategy presented previously. Sometimes it is important to allow more flexibility in the local search and to allow moves which may not improve the objective function. The neighborhood Neighborhood: best move a[i] := !a[i] where i from f1.ng accept when noDecrease; accepts the best move which does not decrease the value of the objective function. The resulting model, depicted in Figure 4, captures the essence of the GSAT algorithm. 2.5.2 Simulated Annealing Simulated annealing is a well-known stochastic strategy to enhance a local improvement search. It is easily expressed by the Localizer neighborhood Neighborhood: move a[i] := !a[i] where i from f1.ng accept when improvement ! ch++; cor noDecrease; The key novelty here is that the accept statement may have a number of acceptance conditions which are tried in sequence until one succeeds or all fail. In addition, each acceptance condition can be associated with an action. The simulated annealing neighborhood specifies that a move is accepted when it improves the objective function, when it does not decrease the objective function, or with the standard probability of simulated annealing, which depends on a temperature parameter and the variation delta of the objective function. Note that the variable ch is incremented when there is an improvement or a decrease in the objective function. The complete model is given in Figure 5. The model illustrates also several new features of Localizer. The Operator section describes two procedures which are used subsequently in the Start and Restart sections. Operators in Localizer uses traditional constructs from imperative programming languages (e.g., loops and conditions) as well as some new primitives for randomization. These features are once again described in more detail in Section 3. Note also the variables t (the temperature) and ch (the change counter) which are used in various places in the model. 2.5.3 Tabu Search Tabu search is another strategy to escape local optima which, in contrast to simulated annealing, does not resort to stochastic moves. The neighborhood Neighborhood: best move a[i] := !a[i] where i from f1.ng such that !tabu(i) accept when always ! t[v] := trial; Solve Type: Constant: cl: array[1.m] of clause = .; Variable: a: array[1.n] of boolean; t: real; Invariant: in 1.m] of int = sum(i in cl[i].p) a[j] + sum(j in cl[i]:n) !a[j]; Operator: void forall(i in 1.n) x[i] := random(ftrue,falseg); ch := 0; void lowTemp() f ch := 0; Satisfiable: Objective Function: maximize nbClauseSat; Neighborhood: move a[i] := !a[i] where i from f1.ng accept when improvement ! ch++; cor noDecrease Start: Restart: lowTemp(); Figure 5: A Simulated Annealing Model for SAT indicates how a simple tabu search can be expressed in Localizer. The key idea here is to select an atom which is not tabu. The where clause is generalized to include this condition. All the moves so-defined are accepted and the model also keeps track of when an atom was last flipped by using the keyword trial. An atom is then tabu if it has been flipped recently, which can be expressed as f return where tl is a parameter specifying the time an atom stays on the tabu list. The complete model is described in Figure 6. Of course, more complicated tabu search algorithms (e.g., using aspiration criteria to overwrite the tabu status or a tabu list whose size varies over time) can be implemented easily. 2.5.4 Composing Neighborhoods Localizer makes it also possible to compose neighborhoods. For instance, the following neighborhood try move where i from OccurInUnsatClause accept when default: best move where i from f1.ng accept when noDecrease; implements the random walk/noise strategy of GSAT. Here, Localizer flips an arbitrary variable in an unsatisfied clause with a probability of 0.1 and applies the standard strategy with a probability of 0.9. Note that Localizer simply goes to the next iteration if the selected neighborhood is empty, since other neighborhoods may be non-empty. 2.5.5 Incrementality Issues In the models presented so far, Localizer needs to simulate the move to find out how the objective function evolves. This simulation can become very expensive when few moves are accepted. In practice, local search implementations often try to evaluate the impact Solve Type: Constant: cl: array[1.m] of clause = .; Variable: a: array[1.n] of boolean; t: array[1.n] of int; tl: int; Invariant: in 1.m] of int = sum(i in cl[i].p) a[j] + sum(j in cl[i]:n) !a[j]; Operator: void tl := 10; forall(i in 1.n) x[i] := random(ftrue,falseg); forall(i in 1.n) t[i] := -tl; return Satisfiable: Objective Function: maximize nbClauseSat; Neighborhood: best move a[i] := !a[i] where i from f1.ng such that !tabu(i) accept when always ! t[v] := trial; Start: Restart: Figure Model for Boolean Satisfiability of the move in the current state. Localizer supports this practice by allowing to specify acceptance criteria which are evaluated in the current state. For instance, the neighborhood definition Neighborhood: first move a[i] := !a[i] where i from f1.ng accept when in current state evaluates the condition gain[i] ?= 0 in the current state to determine whether to take the move. Of course, this requires to generalize the invariants to maintain gain[i] incrementally. The invariants now become Invariant: nbtl: array[ i in 1.m ] of int = sum(i in cl[i]:p) a[j] + sum(j in cl[i]:n) !a[j]; The informal meaning of the new invariants are the following. g01[i] represents the change in the number of satisfied clauses when changing the value of atom i from false to true, assuming that atom i is currently false. Obviously, the flip produces a gain for all unsatisfied clauses where atom i appears positively. It also produces a loss for all clauses where i appears negatively and is the only atom responsible for the satisfaction of the clause. g10[i] represents the change in satisfied clauses when changing the value of atom i from true to false, assuming that atom i is currently true. It is computed in a way similar to g01. gain[i] represents the change in satisfied clauses when changing the value of atom i. It is implemented using a conditional expression in terms of g01[i], g10[i], and the current value of atom i. No simulation is necessary in the resulting model. The GSAT model can be made even more incremental. Since GSAT only selects the move with the best objective value, it is possible to maintain these candidate moves incrementally. The only change is to add the two invariants Candidates select i from 1.n where Here maxGain is simply the maximum of all gains and Candidates describes the set of candidates for flipping as the set of atoms whose gain is positive and maximal. Once the invariants have been described, the neighborhood is defined by flipping one of the candidates. There is no need to use the keyword best or a noDecrease acceptance criteria, since they are already enforced by the invariants. The complete model is depicted in Figure 7. Of course, the same transformation can be performed for the tabu search model. Solve Data Type: Constant: cl: array[1.m] of clause = .; po: array[ i in 1.n] of fintg := select c from 1.m where i in cl[c]:p no: array[ i in 1.n] of fintg := select c from 1.m where i in cl[c]:n Variable: a: array[1.n] of boolean; Invariant: nbtl: array[ i in 1.m ] of int = sum(i in cl[i]:p) a[j] + sum(j in cl[i]:n) !a[j]; gain: array[ i in 1.n ] of int = if a[i] then g10[i] else g01[i]; Candidates select i from 1.n where Satisfiable: Neighborhood: move a[i] := !a[i] where i from Candidates; Start: forall(i in 1.n) Restart: forall(i in 1.n) Figure 7: A More Incremental Model of GSAT 3 The Language As mentioned previously, Localizer is an hybrid language embedding aspects of declarative programming within an imperative language. The main declarative tool is, of course, the concept of invariants which specify expressions whose values must be maintained in- crementally. Imperative constructs are mostly used to specify state transformations and the starting and restarting states. This section reviews Localizer in more detail and is essentially organized along the textual ordering of Localizer statements. 3.1 The Type Section The Type section of Localizer is used to define record types. Records within Localizer are identical to records found in conventional imperative languages like Pascal or C. They aggregate a number of named fields, possibly of different types. For instance, the excerpt declares a record type Edge consisting of two integers (the origin and destination nodes), while the excerpt declares a record type clause with the two fields pl and nl of type "sets of integers". 3.2 Constants The Constant section declares the input data and, possibly, some derived data which are useful in stating the model. All constants are typed, read only (i.e., they cannot be modified by assignments), and are initialized. As shown later on, there are various ways to initialize data in Localizer. 3.2.1 Data Types Basic Data Types The basic data types supported by Localizer are integers, booleans and floats. The excerpt declares an integer n whose value is 10, a float whose value is 3:14, and a Boolean. Integers can range from \Gamma2 floats are double-precision floating-point numbers. Arrays Localizer supports multi-dimensional arrays of arbitrary types. The declara- tion defines a one-dimensional array a of integers which is initialized by the vector [6,1,4,5,7]. The declaration declares a matrix of integers. Records As mentioned previously, records can be used in Localizer to cluster together related data. For instance, the declaration of Edge := [! defines p as an array of 3 edges. Each edge is initialized with a tuple. A tuple is compatible with a record type if it has the same number of fields and the type of each field is compatible with the type of the corresponding field. Sets Finally, Localizer supports sets of arbitrary types. For instance, the declaration declares and initializes a set of edges. 3.2.2 Inline and Offline Initializations Constants can be initialized inline as in all previous examples. They can also be initialized offline in the Init section to separate the model from the instance data, which is usually a good practice. The excerpt declares a float f whose initialization is given in the Init section. The Init section consists of a set of pairs (identifier,value) and, of course, the type of the initializationmust match the type of the declaration. Offline initializations can be used for arbitrary types. For instance, the Boolean satisfiability models may contain an initialization section of the form Init: cl Here, cl is initialized with a vector of 11 tuples. Each tuple is a pair of sets. The first set of each pair is associated with the first field of the record of type clause and denotes the set of positive literals in that clause. The second set is matched with the second field of the record and denotes the negative atoms of the clause. 3.2.3 Generic Data Localizer also supports the concepts of generic data which was introduced in Numerica [10]. The basic idea here is to initialize the data using an expression which may depend on parameters of the declaration. Genericity is especially attractive to define derived data which are then used to simplify the model. In the case of the fully incremental version of GSAT (see Figure 7, it is important to know the clauses where an atom i appears positively (and negatively). This information is derived from the data cl using the generic declarations Constant: po: array[i in 1.n] of fintg select c from 1.m where i in cl[c]:pg; no: array[i in 1.n] of fintg select c from 1.m where i in cl[c]:ng; There are a couple of important points to stress here. First, the declarations use parameters which range over the index sets of the array. For instance, parameter i ranges over 1::n, the set of atoms. Second, these parameters are used in the expression defined on the right-hand side of := to specify the value of the array at the given position. In the GSAT example, po[i] is defined as the set of clauses where atom i appears positively. The expressions allowed for the right-hand side are very general and their syntax is given in Figure 8. Figure 9 describes the signature of the primitive functions which have the obvious meanings. 3.3 Variables Variables are of course fundamental in Localizer, since they define the state of the com- putation. Variables are declared in the same ways as constants, except that they are not initialized. They are generally given an initial value in the Start section and modified in the Neighborhood and Restart sections. As should be clear from the examples, Localizer has an assignment operator, whose right-hand side is an expression. 3.4 Invariants Invariants are the key concept provided by Localizer to support the design of local search algorithms. Syntactically, invariants are simply generic data. However, invariants are not static since they may contain variables and/or other invariants. As a consequence, the values of invariants are state-dependent and Localizer is responsible to update them after each state transition. The following excerpt illustrates invariants from a graph-coloring model where C[i] refers to another invariant whose type is fintg (set of integers). constant literal ::= hexpri . hidentifieri hSet Bodyi ::= hexpri . hexpri j hexpri ::= hexpri . hexpri select hidentifieri from hrangei [ where hexpri select hidentifieri from hexpri [ where hexpri hrangei ::= hexpri . hexpri Figure 8: Expression syntax Arithmetic Set related Output min2(int,int)!int size(fTg)!int print(.)!void max2(int,int)!int random(fTg)!T println(.)!void floor(float)!int minof(fTg)!T ceil(float)!int insert(fTg,T)!void round(float)!int remove(fTg,T)!void exp(float)!float time()!int Figure 9: Primitive Functions select i from 1.n where size(C[i])=0g; in 1.n] of fEdgeg select i from C[k] select j from C[k] In addition to the standard expressions, invariants can also be defined in terms of distribute and dcount. They take as input a one-dimensional array of integers A and two sets I and O. I must be a subset of the index set of A. The result type of a distribute expression is one-dimensional array of set of integers B whose index set is O. fintg. The result type of a dcount expression is one-dimensional array of integers whose index set is O. Their meanings are given by the following equivalences: These expressions were introduced because of their ubiquity in practical applications. It is important to stress that the data declared by invariants cannot appear in left-hand sides of assignment statements. Operators The Operator section contains function definitions. Functions in Localizer are essentially similar to functions in C and use traditional assignment, conditional, and iterative statements, as well as recursion. In addition, Localizer provides some constructs which are useful for local search algorithms. The syntax of statements in Localizer is sketched in Figure 10. The forall instruction provides a convenient way of iterating over the elements of a set. The choose instruction can be used to select an element from a set, in a (don't care) nondeterministic way. It is possible to filter elements of the given set or to select an element optimizing a function. The following examples illustrate the various forms of choose statements: choose i from A; choose i from A minimizing D[i]; choose i from A such that gain[i] ? 0; choose i from A such that gain[i] ? 0 maximizing D[i]; The entire state, including constants, variables, and invariants, are accessible to functions. However, only variables and locals can be modified by assignments. Note that functions and imperative constructs can be used in the Start and restart sections, in the specification of the state transformation of the neighborhood, in the actions associated with specific acceptance criteria, and in the from instructions that appear in the where clause of the move instruction. Typical examples of functions were introduced in the models of Figures 5 and 6. while hexpri do hstatementi return hexpri choose hidentifieri from hexpri [ hoptClausei hoptClausei ::= minimizing hexpri maximizing hexpri ::= such that hexpri [minimizing hexpri [with hletBlocki ::= such that hexpri [maximizing hexpri [with hletBlocki Figure 10: The Syntax of Statements 3.6 Neighborhood The neighborhood is specified in Localizer with a move instruction of the form: move where from xn from Sn This instruction uses both declarative and procedural components. The first part of the statement specifies, with the imperative code the transformation of the current state into one of its neighbors. The second part starting with the where keyword specifies the objects used in the imperative code. The modeling effort is primarily devoted to the definition of the sets S i and the invariants they are based on. The syntax of move instructions is depicted in Figure 11. Once an element of the neighborhood has been selected, Localizer determines if it is an appropriate move. The acceptance criterion lists boolean conditions that are to be tried in sequence. As soon as the state satisfies one of the conditions, it is accepted and the state transformation is performed. The criterion is build according to the syntax where ::= hidentifieri from f hexpr 1 ::= hidentifieri from f hidentifieri : htypei j hselect 1 ::= hidentifieri from hlvaluei [ hoptClausei Figure 11: The Syntax of move Instructions ::= in resulting state hAcceptanceStatementi ::= in current state hAcceptanceStatementi improvement noDecrease ::= hAcceptanceConditioni and hAcceptanceConditioni ::= hAcceptanceConditioni or hAcceptanceConditioni ::= not hAcceptanceConditioni Acceptance criteria are, by default, evaluated in the new state. This new state must be constructed, which induces the update of all invariants. It is also possible to specify that the acceptance criterion should be evaluated in the current state by using the keywords in current state. Using the current state to evaluate moves may produce significant improvements in efficiency. 3.7 The Objective Function The objective function is stated in a separate section and is used to assess the "quality" of a given state. The objective function can in fact be viewed as an invariant which is maintained by Localizer and used to evaluate moves. Note that the objective function is optional and is only used when the acceptance criteria are evaluated in the resulting state. 3.8 Termination Criteria Termination is handled via several sections and depends on the nature of the problem. Each model starts with a keyword that is either solve or optimize to specify either a decision or an optimization problem. The Satisfiable Section The (optional) satisfiable section is used to specify whether a state is a solution. In decision problems, Localizer terminates whenever this is the case. In optimization problems, Localizer updates the best solution whenever the state is a solution which improves the best bound found so far. When the section is omitted, Localizer assumes that every state is a solution. The Local and Global Conditions These sections, which are optional, specify Boolean predicates which are used to control the innermost and outermost loops of the computation model. For instance, a simulated annealing model can use the local condition to implement a cutOff strategy that terminates the innermost loop and drops the temperature whenever the evaluation function has been updated sufficiently many times. The Parameter Section This section, also optional, is used to override the default values of some parameters of the system. Typical uses redefine the maxSearches and maxTrials parameters of the computation model. Implementation This section reviews the implementation of invariants which are the cornerstone of Local- izer. Informally speaking, invariants are implemented using a planning/execution model. The planning phase generates a specific order for propagating the invariants, while the execution phase actually performs the propagation. This model makes it possible to propagate only differences between two states and mimics, to a certain extent, the way specific local search algorithms are implemented. The planning/execution model imposes some restrictions on the invariants. These restrictions intuitively, makes sure that there is an order in which the invariants can be propagated so that a pair (variable,invariant) is considered at most once. Various such restrictions can be imposed. Static invariants can be ordered at compile time and are thus especially efficient. However, static invariants rule out some interesting models for scheduling and resource allocation problems. Dynamic invariants still make it possible to produce an ordering so that a pair (variable,invariant) is considered at most once. However, dynamic invariants require to interleave the planning and execution phases. Dynamic invariants seem to be a good compromise between efficiency and expressiveness. The rest of this section is organized as follows. The algorithms use normalized invariants and Section 4.1 reviews the normalization process. Section 4.2 describes static invariants and their implementation. This should give readers a preliminary understanding of the implementation. Section 4.3 describes dynamic invariants and their implementation. This section only considers arithmetic invariants but it is not difficult to generalize these results to invariants over sets. 4.1 Normalization The invariants of Localizer are rewritten into primitive invariants by flattening expressions and arrays. The primitive invariants are of the form: x := y \Phi z where c is a constant, x; are variables, \Phi is an arithmetic operator such as and or an arithmetic relation such as -; ? and =, and Q is an aggregate operator such as sum, prod, max, and min. Relations return 1 when true and 0 otherwise. An invariant assigns to x the element in position e in the list [x last invariant is useful for arrays which are indexed by expressions containing variables. At any given time, Localizer maintains a set of invariants I over variables V . Given an invariant I 2 I of the form x := e (where e is an expression), def(I) denotes x while exp(I) denotes e. Given a set of invariants I over V and x 2 V , invariants(x; I) returns the subset of invariants fI i ' I such that x occurs in exp(I i ). The set of variables V nfdef(I)jI ' Ig are the variables which are the parameters of the system of invariants. These variables only can be modified in the neighborhood definitions. Note also that a variable x can be defined by at most one invariant. i.e. there exist at most one I ' I such that 4.2 Static Invariants The basic assumption behind Localizer implementation is that invariants only change marginally when moving from one state to one of its neighbors. Consequently, the goal of the implementation is to run in time proportional to the amount of changes. More precisely, the implementation makes sure that a pair (variable; invariant) is considered at most once, i.e., when the variable is updated, the invariant is updated but it will never be reconsidered because of that variable. To achieve this goal, the implementation uses a planning/execution model where the planning phase determines an ordering for the updates and the execution phase actually performs them. The existence of a suitable ordering is guaranteed by the restrictions imposed on the invariants by the system. Note also that planning/execution models are often in graphical constraint systems (e.g., [1]). This section describes static invariants which impose a static restriction. Although this restriction may seem strong, it accommodates many models for applications such as satisfiability and graph coloring to name a few. The main practical limitation is that elements of arrays cannot depend on other elements in the same array. This restriction is lifted by dynamic invariants. Note, however, that static invariants have the nice property that the planning phase can be entirely performed at compile time. 4.2.1 The Planning Phase The basic idea behind static invariants is to require the existence of a topological ordering on the variables. This topological ordering is obtained by associating a topological number t(x) with each variable x. The topological number of an invariant I is simply t(def(I)). The topological numbers are obtained from constraints derived from the invariants. 1 The topological constraints of invariant I are defined as follows: Definition 2 The topological constraints of a set of invariants I denoted by tc(I) is I2I tc(I). Definition 3 A set of invariants I over t is static if there exists an assignment such that t satisfies tc(I). The planning phase for static invariants consists of finding the topological assignment. The planning phase can be performed at compile time since the topological constraints do not depend on the values of the variables in a given state. 4.2.2 The Execution Phase The execution phase is given a set of variables M which have been updated and a topological assignment t. It then propagates the changes according to the topological ordering. The algorithm uses a queue which contains pairs of the form hx; Ii. Intuitively, such a pair means that invariant I must be reconsidered because variable x has been updated. The main step of the algorithm consists of popping the pair hx; Ii with the smallest t(I) and to propagate the change, possibly adding new elements to the queue. The algorithm is shown in Figure 12. procedure execute(I,M,t) begin while do endwhile function POP(Q,t) pre: Q is not empty post: Figure 12: The Execution Phase for Static Invariants 4.2.3 Propagating the Invariants To complete the description of the implementation of static invariants, it remains to describe how to propagate the invariants themselves. The basic idea here is to associate two values x with each variable x. The value x represents the value of variable x at the beginning of the execution phase, while the value x c represents the current value of x. At the beginning of the execution phase, x of course. By keeping these two values, it is possible to compute how much a variable has changed and to update the invariants accordingly. For instance, the propagation of the invariant is performed by a procedure procedure propagate(x i ,x := sum(x 1 begin The procedure updates x c according to the change of x i . Note that, because of the topological has reached its final value. Note also that x c is not necessarily final after this update, because other pairs hx may need to be propagated. 4.3 Dynamic Invariants Static invariants are attractive since the planning phase can be performed entirely at compile time. However, there are interesting applications in the areas of scheduling and resource allocation where sets of invariants are not static. This section introduces dynamic invariants to broaden the class of invariants accepted by Localizer. Dynamic invariants are updated by a series of planning/execution phases where the planning phase takes place at execution time. 4.3.1 Motivation The main restriction of static invariants comes from the invariant The static topological constraint for this invariant is and it prevents Localizer from accepting expressions where some elements of an array may depend on some other elements of the same array. This constraint is strong, because the value of e is not known at compile time. In fact, it may not even be known before the start of the execution phase since some invariants may update it. However, there are many applications in scheduling or resource allocation where such invariants occur naturally. For instance, a scheduling application may be modeled in terms of an invariant start[3] := max(end[prec[3]]; end[disj[3]]); where start[i] represents the starting date of task i, prec[i] the predecessor of the task in the job and disj[i], the predecessor of task i in the disjunction. Variable disj[i] is typically updated during the local search and the above invariant is normalized into a set of the form: start 3 := max(p; d) Of course, such an application has also invariants of the form implying that the resulting set of invariants is not static. 4.3.2 Overview of the Approach The basic idea behind dynamic invariants is to evaluate the invariants by levels. Each invariant is associated with one level and, inside one level, the invariants are static. Once a level is completed, planning of the next level can take place using the values of the previous levels since lower levels are never reconsidered. With this computation model in mind, the topological constraint associated with an invariant can be reconsidered. The basic idea is to require that e be evaluated before x (i.e. the level of x is the level of e Once e is updated, then it is easy to find a weaker topological constraint since the value of e is known. The invariant can be simplified to The planning phase is thus divided in two steps. A first step, which can be carried out at compile time, partitions the invariants in levels. The second step, which is executed at runtime, topologically sorts the invariants within each level whenever the invariants at the lower level have been propagated. 4.3.3 Formalization The basic intuition is formalized in terms of two assignments l of two sets of constraints. Definition 4 The level constraints associated with an invariant I and denoted by lc(I) are defined as follows: ))g The level constraints are not strong except for the invariant element where the level of x is strictly greater than the level of e. Informally, it means that e must be evaluated in an earlier phase than x. Definition 5 The level constraints associated with a set of invariants I and denoted by l(I) is simply S I2I Definition 6 A set of invariants I is serializable if there exists an assignment l satisfying lc(I). A serializable set of invariants can be partitioned into a sequence ! I the invariants in I i have level i. This serialization can be performed at compile-time. The second step consists of ordering the invariants inside each partition. This ordering can only take place at runtime, since it is necessary to know the values of some invariants to simplify the element invariants. Definition 7 Let S be a computation state and let S(x) denote the value of x in S. The static constraints associated with an invariant I wrt to S, denoted sd(I ; S) is defined as follows: Definition 8 The static constraints associated with a set of invariants I wrt to a state S, denoted by sd(I; S) is simply S I2I Definition 9 A set of invariants I is static wrt to a state S if there exists an assignment The main novelty of course is in the invariant element where the topological constraint can ignore e since its value is known. In addition, since the final value of e is known, the topological constraints can be made precise since the element y e c that x depends upon is known. be a computation state. A set of invariants I is dynamic wrt S 0 if 1. I is serializable and can be partitioned into a sequence ! I 2. I i is static wrt S i where S i (i ? 0) is the state obtained by propagating the invariants I i\Gamma1 in S i\Gamma1 . Of course, dynamic invariants cannot be recognized at compile-time and may produce an execution error at runtime when Localizer is planning a level. 4.3.4 The Execution Algorithm The new execution algorithm is a simple generalization of the static algorithm and is shown in Figure 13. Note the planning step which is called for each level. procedure execute(I,M) begin do endfor Figure 13: The Execution Algorithm for Dynamic Invariants 5 Experimental results This section summarizes some preliminary results on the implementation of Localizer (about 35,000 lines of C++). The goal is not to report the final word on the implementation but rather to suggest that Localizer can be implemented with an efficiency comparable to specific local search algorithms. To demonstrate practicability, we experimented with Localizer on several problems: GSAT, graph coloring, graph partitioning, and job-shop scheduling. 5.1 GSAT GSAT is generally recognized as a fast and very well implemented system. The experimental results were carried out as specified in [8]. Table 1 gives the number of variables (V ), the number of clauses (C), and MaxTrials (I) for each size of benchmarks as well as the CPU times in seconds of Localizer (L), the CPU times in seconds of GSAT (G) as reported in [8], and the ratio L=G. The times of GSAT are given on a SGI Challenge with a 70 MHz MIPS R4400 processor. The times of Localizer were obtained on a SUN SPARC- scaled by a factor 1.5 to account for the speed difference between the two machines. Localizer times are for the incremental model presented in Section 2. Note that this comparison is not perfect (e.g., the randomization may be different) but it is sufficient for showing that Localizer can be implemented efficiently. 5.2 Graph Coloring Graph coloring was the object of an extensive experimental evaluation in [4] and this section reports on experimental results along the same lines. The experiments were conducted on graphs of densities 10, 50, and 90 and of sizes 125, 250, and 500. They were also conducted on so-called "cooked" graphs. Because of the nature of the experimental results reported in [4], it is not easy to compare the efficiency of Localizer to the efficiency of their algorithm. As a consequence, a very efficient C implementation of their algorithm was built from scratch by a graduate student who was closely supervised to obtain a very efficient incremental algorithm. As far as we can judge, the timings and the quality of this algorithm seem 5 200 860 2000 873.11 168.00 5.20 Av. 3.38 Table 1: GSAT: Experimental Results consistent with those in [4]. In addition, this algorithm is the most efficient implementation built by a graduate student in the combinatorial optimization class at Brown (CS-258) in the last three years (for the given model of course). In the following, we discuss the development time of the two implementations, the quality of the solutions obtained (to make sure that the algorithms are comparable in quality), and the efficiency. Development Time The C implementation of the algorithm is about 1500 lines long and required a full week. The Localizer model is about one page long. Quality of the Solutions Table 2 describes the quality of the coloring found by Lo- calizer. These results agree with those of the C implementation and with those reported in [4]. Each set of rows corresponds to a class of graphs and to 100 executions of Localizer on graphs from this class. The rows in each set report on the various values found by Localizer on these graphs and their frequencies. The columns report the number of vertices, the density of the graph, the size factor sf used in the experiments, the number of colors found by some solution, and the frequency of colorings of this quality. For instance, the first set of rows reports that, on graphs of 125 vertices and density of 50%, 92% of the executions led to a coloring with 19 colors and 8% of the executions led to a coloring with colors. The results are given both for random and cooked graphs and the frequencies are similar for both Localizer and the C implementation. Efficiency Table 3 compares the efficiency of Localizer with the C implementation on the same problems. Each row reports the average time of the two implementations for the 100 graphs in each class and computes the slowdown of Localizer. The experiments were performed on a SUN Sparc Ultra-1 running Solaris 5.5.1 and the standard C++ compiler. The average slowdown is 4.82, while minimum and maximum slowdowns are respectively 3.56 and 5.54. On these problems, the average slowdown is slightly higher than a machine generation but it remains reasonable given the preliminary nature of the implementation. This slowdown should also be contrasted with the substantial reduction in development time. Vertices Density Size Factor (sf) Colors Frequency random 125 50 3 19 92 random 250 50 4 43 34 44 random 500 50 4 random 125 90 1 - 44 7 random random 500 90 1 - 143 144 22 cooked 125 4 9 100 cooked cooked 500 2 25 71 Table 2: Graph Coloring: Quality of the Solutions Vertices Density Size Factor (sf) Localizer (L) C Implementation (C) L/C random 125 50 3 78.3 18.9 4.50 random 250 50 4 82.8 18.4 4.50 random 500 50 4 633.7 123.4 5.10 random 125 90 1 16.18 4.53 3.56 random random 500 90 1 162.7 29.6 4.88 cooked 125 4 22.09 4.18 5.28 cooked cooked 500 2 240.3 49.9 4.80 average 4.82 Table 3: Graph Coloring: Efficiency of Localizer 5.3 Graph Partitioning The problem has been studied experimentally in [3] and, once again, the experiments reported here are based on a similar setting. Table 4 depicts the experimental results of Localizer. The first row gives the setting of our parameters: T is the starting tempera- ture, TF is the percentage of reduction of the temperature, SF is the size factor and the remaining two were described previously. Table 5 compares Localizer with the results reported in [3]. 5.4 Job-Shop Scheduling To conclude, we report some preliminary results on job-shop scheduling. Localizer has been evaluated on a set of 28 classic benchmarks. The model used for these experiment implements the neighborhood, commonly referred to as N1, which considers the reversal of exactly one edge on the critical path. N1 has a number of nice properties: It preserves the satisfiability of the solution and the transition graph induced by the neighorhood is such that the optimal solution is reachable from all nodes of the graph. The experiments were conducted in a fashion similar to what is reported in [6]. The parameters were defined as ffl The maximal number of searches (maxSearches) is 1 ffl The maximal number of iterations for the inner loop (maxTrials) is 12000. ffl The tabu list has a varying length constrained in between 5 and 30. Moreover, the length is varied according to the rule of [6]. ffl Contrary to [6], the model does not use a restarting strategy. Graph Results Class V ertices Density SB F req: Localizer random 124 2 11-13 14-16 17-19 33 38 29 2.22 random 4 55-59 60-65 66-77 random 8 159-174 175-190 191- 23 53 24 3.24 random random random 2 92-106 107-121 122-131 12 42 36 5.10 random 4 324-343 344-363 364-380 38 43 19 6.50 random 8 828-877 878-927 928- 37 40 23 9.98 random 500 0.5 48-54 55-59 60-66 15 43 42 10.08 random 4 1661-1701 1702-1741 1742-1824 22 58 20 20.67 random 1000 0.25 90-103 104-118 119-126 41 52 7 19.99 random 0.5 439-455 456-475 476-503 36 43 21 22.62 geom. 500 5 4-13 14-23 24-37 7 58 geom. 20 148-246 247-346 347-450 41 44 15 11.40 geom. 1000 5 24-43 44-63 64-78 37 57 6 18.70 geom. 20 196-399 400-599 600-816 28 Table 4: Graph Partitioning: Experimental Results Graph Results Class V ertices Density L:Best L:Time J:Best J:T ime Ratio random 124 2 11 2.22 13 85.4 38.5 random random 500 0.5 47 10.08 52 379.8 37.7 random 1000 0.25 90 19.99 102 729.9 36.5 Table 5: Graph Partitioning: Comparison Results Table 6 reports the preliminary results. These results cannot be really compared with the results of [6], since the neighborhood used in their experiment is (RN1 [ RN2) while the model used here relies on N1 alone. The time reported correspond to the algorithm termination (once the 12000 iterations are spent), not the time required to produce the best or the optimal solution for the first time. Note also that the the value of the optimal solution was not used as a stopping criterion. If this condition was to be used, the running time would vary a lot more. For instance, LA01 usually produce the optimum solution in about 0:5 seconds. Interestingly enough, even the simple neighborhood N1 does quite well and finds the optimum solution for benchmarks (out of 28 benchmarks). The table also reports a coarse histogram that summarizes the frequencies of apparition for the solutions. These frequencies were obtained based on a serie of 100 experiment for each benchmark. In summary, the results seem to indicate that Localizer will also compare well on scheduling application. 6 Conclusion The main contribution of this paper is to show that local search can be supported by modeling languages to shorten the development time of these algorithms substantially, while preserving the efficiency of special-purpose algorithms. To substantiate this claim, we presented a progress report on the domain-specific language Localizer, introduced in [5]. Localizer statements are organized around the traditional concepts of local search and may exploit the special structure of the problem at hand. The main conceptual tool underly- Benchmark Results (MD=12000,MS=1,Neighborhood=N1) Name Job M: Opt Ranges F req: Loc Avg:sol: Table ing Localizer is the concept of invariants which make it possible to specify complex data structures declaratively. These data structures are maintained incrementally by Local- izer, automating one of the most tedious and error-prone parts of local search algorithms. Experimental results indicate that Localizer can be implemented to run with an efficiency comparable to specific implementations. Our current research focuses on building higher-level data structures to simplify the design of invariants which are the cornerstone of the language. Extending the strategies to accommodate dynamic k-opt [7], genetic algorithms, constraint techniques are also contem- plated. Longer term research will explore how Localizer can be turned into a programming language library to guarantee extensibility and wide applicability for expert users, while preserving the right level of abstraction. Acknowledgments This paper is dedicated to the memory of Paris C. Kanellakis who kept on gently pressuring us to pursue this topic. Thanks to D. McAllester and B. Selman for many discussions on this research and to Costas Bush for implementing the graph-coloring algorithm in C. This research was supported in part by an NSF NYI Award. --R The Programming Language Aspects of ThingLab AMPL: A Modeling Language for Mathematical Programming. Optimization by Simulated Annealing: An Experimental Evaluation Optimization by Simulated Annealing: An Experimental Evaluation Modeling Language for Local Search. Marco Trubian Mauro Dell'Amico. Combinatorial Optimization: Algorithms and Complexity. A New Method for Solving Hard Satisfiability Problems. Models for Using Stochastic Constraint Solvers in Constraint Logic Programming. Numerica: a Modeling Language for Global Optimization. --TR --CTR Simon de Givry , Laurent Jeannin, A unified framework for partial and hybrid search methods in constraint programming, Computers and Operations Research, v.33 n.10, p.2805-2833, October 2006 Pascal Hentenryck , Laurent Michel , Liyuan Liu, Contraint-Based Combinators for Local Search, Constraints, v.10 n.4, p.363-384, October 2005 Irit Katriel , Laurent Michel , Pascal Hentenryck, Maintaining Longest Paths Incrementally, Constraints, v.10 n.2, p.159-183, April 2005 Pascal Hentenryck , Laurent Michel, Control Abstractions for Local Search, Constraints, v.10 n.2, p.137-157, April 2005 Laurent Michel , Pascal Van Hentenryck, A constraint-based architecture for local search, ACM SIGPLAN Notices, v.37 n.11, November 2002 Marco Cadoli , Toni Mancini, Combining relational algebra, SQL, constraint modelling, and local search, Theory and Practice of Logic Programming, v.7 n.1-2, p.37-65, January 2007 Laurent Michel , Pascal Van Hentenryck, Comet in context, Proceedings of the Paris C. Kanellakis memorial workshop on Principles of computing & knowledge: Paris C. Kanellakis memorial workshop on the occasion of his 50th birthday, p.95-107, June 08-08, 2003, San Diego, California, USA
incremental algorithms;modeling language;local search
593313
Ordering Constraints over Feature Trees.
Feature trees are the formal basis for algorithms manipulating record like structures in constraint programming, computational linguistics and in concrete applications like software configuration management. Feature trees model records, and constraints over feature trees yield extensible and modular record descriptions. We introduce the constraint system FEAT of ordering constraints interpreted over feature trees. Under the view that feature trees the relation \leq corresponds to the information ordering (carries less information than). We present two algorithms in cubic time, one for the satisfiability problem and one for the entailment problem of FEAT. We show that FEAT has the independence property. We are thus able to handle negative conjuncts via entailment and obtain a cubic algorithm that decides the satisfiability of conjunctions of positive and negated ordering constraints over feature trees. Furthermore, we reduce the satisfiability problem of Drres weak subsumption constraints to the satisfiability problem of FEAT and improve the complexity bound for solving weak subsumption constraints from O(n^5) to O(n^3).
Introduction Feature constraints have been used for describing records in constraint programming [2, 24, 23] and record like structures in computational linguistics [13, 12, 20, 18, 19]. Following [3, 5, 4] we consider feature constraints as predicate logic formulae that are interpreted in the structure of feature trees. A feature tree is a possibly infinite tree with unordered labeled edges and with possibly labeled nodes. Edge labels are functional; i.e., the labels of the edges departing from the same node must be pairwise different. Under the view that feature trees represent symbolic information, the feature tree t 1 represents less information than the feature tree t 2 if t 1 has fewer edges and node labels than t 2 . The relation - that we define corresponds to the information ordering in precisely this sense. Algebraically, there is a homomorphic embed- wine red color - wine red 1997 color year ding from t 1 to t 2 (i.e., a mapping from nodes in t 1 to nodes in t 2 under which the node labeling is invariant). An example is given in the picture. We introduce the constraint system FT- of information ordering constraints over feature trees. The system FT- is obtained by adding ordering constraints to the constraint system FT [3]. The syntax of FT- constraints j is defined by where x and x 0 are variables and a is a label. The semantics of FT- is given by the interpretation over feature trees where the symbol - is interpreted as information ordering on feature trees. The semantics of x[a]y and a(x) are defined as in FT . For in- stance, both trees depicted above are possible values for x in solutions of the constraint It is clear that FT- is more expressive than FT since the information ordering is antisymmetric (i.e., As we show in the paper, FT- is strictly more expressive than FT . For instance, no constraint in FT can be equivalent to x-x 0 . Also, we do not know of any formula over FT (even with existential quantifiers) equivalent to 9x(x 1 -x- x 2 -x)-9x(x 1 -x- x 3 -x); this FT- formula expresses that x 1 is unifiable with both x 2 and x 3 (but does not imply unifiability of x 2 and x 3 ). We show that the satisfiability problem of conjunctions of positive and negative FT- constraints j-:j 1 -:j n is decidable in O(n 3 ). This result includes a decision procedure for the entailment problem of the form j 0 if and only if the formula j 0 -:j is unsatisfiable. To establish our result, we prove that FT- has the fundamental independence property (similar to its relatives RT [6], FT [3], and CFT [24]). We reduce the satisfiability problem of D-orre's weak-subsumption constraints [7] over feature algebras linearly to the one in FT- . Thereby, our algorithm improves on the best known satisfiability test for weak subsumption constraints which uses finite automata techniques and has an O(n 5 )-complexity bound [7]. Plan of the Paper. Section 2 surveys related work. Section 3 defines FT- . Section 4 presents the satisfiability test for FT- constraints. Section 8 contains the completeness proof. Section 5 presents the entailment test for FT- constraints, and proves the independence property of FT- . Section 6 defines weak subsumption constraints and reduces their satisfiability problem to the one of FT- constraints. Section 7 shows that FT- is strictly more expressive than FT . Related Work Ines Constraints. In previous work [17], we have introduced the constraint system INES of inclusion constraints over non-empty sets of trees and a cubic satisfiability test. The satisfiability test for FT- is inspired by and subsumes the one for INES. However, the entailment problems for FT- and INES constraints are different. The entailment problem of INES constraints is coNP-hard [16]. Intuitively, the entailment problem of FT- is less expressive than the one of INES because an FT- constraint j cannot uniquely describe a single feature tree (in absence of arity constraints); in contrast, INES constraints (which are inclusions between first-order terms with an implicit arity restriction) can uniquely describe a constructor tree as a singleton set. For instance, the INES constraint x'a describes the singleton fag. As a consequence, the entailment proposition x'a-a'y x'y holds in INES. No similar entailment phenomenon exists for FT- . Feature Constraints. The constraint system CFT [24] extends FT by arity constraints of the form xf f saying that the denotation of x has subtrees exactly at the features rational tree constraint system RT [6] but provides finer-grained constraints. The system EF [25] extends CFT by feature constraints x[y]z, providing for first-class features. Complete axiomatizations for FT and CFT have been given in [5] and [4], respectively. The satisfiability of EF constraints is shown NP-hard in [25]. The system FT- (sort) extends FT- by allowing a partial order on labels [15]. Subsumption Constraints. Subsumption is an ordering on the domain of feature alge- bras. Subsumption constraints have been considered in the context of unification-based grammars to model coordination phenomena in natural language [9, 7, 21]. There, one wants to express that two feature structures representing different parts of speech share common properties. For example, the analysis of "programming" and "linguistics" in the phrase Feature constraints for [ NP programming] and [ NP linguistics] should share (but might refine differently) the information common to all noun phrases. Since the satisfiability of subsumption constraints is undecidable [9], D-orre proposed subsumption as an decidable approximation of subsumption. As we show, the information ordering over feature trees (as investigated in this paper) coincides with the subsumption ordering interpreted over (the algebra of) feature trees. Independent Constraint Systems. A constraint system has the fundamental independence property if negated conjuncts are independent from each other, or: its constraints cannot express disjunctions (we will give a formal definition later). Apart from the mentioned tree constraint systems RT , FT , CFT [6, 1, 24, 3], constraint systems with the independence property include linear equations over the real numbers [14], or infinite boolean algebras with positive constraints [10]. 3 Syntax and Semantics of FT- The constraint system FT- is defined by a set of constraints together with an interpretation over feature trees. We assume an infinite set of variables ranged over by x;y;z, and an infinite set L of labels ranged over by a;b. Feature Trees. A path p is a finite sequence of labels. The empty path is denoted by e and the free-monoid concatenation of paths p and p 0 as pp 0 ; we have Given paths p and q, p 0 is called a prefix of p if domain is a non-empty prefix closed set of paths. A feature tree t is a pair (D; L) consisting of a tree domain D and partial labeling function L : D * L . Given a feature tree t, we write D t for its tree domain and L t for its labeling function. The set of all feature trees is denoted by F . A feature tree is called finite if its tree domain is finite, and infinite otherwise. Syntax. An FT- constraint j is defined by the following abstract syntax. An FT- constraint is a conjunction of basic constraints which are either inclusion constraints x-y, labeling constraints a(x), selection constraints x[a]y, or compatibility constraints x-y. Compatibility constraints are needed in our algorithm and can be expressed by first-order formulae over inclusion constraints (see Proposition 1). We identify to associativity and commutativity of conjunction, i.e., we view j as a multiset of inclusion, labeling, selection, and compatibility constraints. We write j in j 0 if all conjuncts in j are contained in j 0 . The size of a constraint j is defined as the number of label and variable occurrences in j. Semantics. We next define the structure F over feature trees in which we interpret constraints. The signature of F contains the binary relation symbols - and - and for every label a a unary relation symbol a() and a binary relation symbol [a]. In F these relation symbols are interpreted is a partial function (on D t 1 Let F denote first-order formulae built from FT- constraints with the usual first order connectives. We call F satisfiable (valid) if F is satisfiable (valid) in the structure F . We say that F entails F 0 , written F is valid, and that F is equivalent to F 0 if F 1 $F 2 is valid. We denote with V (F) the set of variables occurring free in F and with L(F) the set of labels occurring in F. Proposition 1. The formulae x-y and 9z(x-z- y-z) are equivalent in F . Proof. Let s be a variable assignment into F which also is a solution of the formula 9z(x-z- y-z). Since L s(x) [L s(y) ' L s(z) and L s(z) is a partial function, L s(x) [L s(y) is also a partial function. Hence s is a solution of x-y. Conversely, if s is a solution of x-y then L s(x) [L s(y) is a partial function. Thus, the pair is a feature tree and the variable assignment s 0 defined by s 0 for x 6= z is a solution of x-z- y-z. 2 4 Satisfiability Test We present a set of axioms valid for FT- and then interpret these axioms as an algorithm that solves the satisfiability problem of FT- . The axioms and the algorithm are inspired by the ones for INES constraints presented in [17]. Table contains five axiom schemes F1 - F5 that we regard as sets of axioms. The union of these sets of axioms is denoted by F, i.e., For instance, an axiom scheme x-x represents the infinite set of axioms obtained by instantiation of the meta variable x. An axiom is either a constraint j, an implication between constraints or an implication Proposition 2. The structure F is a model of the axioms in F. F1 x-x and x-y- y-z ! x-z false for a 6= b Table 1. Satisfiability of FT- Constraints. Proof. By a routine check. For illustration, we prove the statement for the second rule in F3, namely x-y- y-y . The following implications hold: The Algorithm F. The set of axioms F induces a fixed point algorithm F that, given an input constraint j, iteratively adds logical consequences of F [fjg to j. (Observe that actually only constraints of the form x-y and x-y are derived). More precisely, in every step F inputs a constraint j and terminates with false or outputs a constraint Termination with false takes place if there exists j 00 in j such that j 00 ! false 2 F. Output of j-j 0 is possible if j 0 2 F or there exists j 00 in j with Example 1. Inconsistency can be due to incompatible upper bounds. Consider: false for a 6= b We may add x-z by F3:1, then z-x via F3:3, then y-x with F3:2, and finally terminate with false via F5. Example 2. We need F4 for deriving the unsatisfiability of the constraint: Algorithm F may add x-y after several steps as shown in Example 1. Then it may proceed with x-y 0 via F4 and terminate with false via F5. Termination. The fixed point algorithm F terminates when reflexivity of inclusion x-x is restricted to variables x 2 V (j). Given a subset F of F, a constraint j is called F-closed if algorithm F under this restriction and w.r.t. the axioms in F cannot proceed on j. Note that false is not F-closed since it is not a constraint by definition. Example 3. Our control takes care of termination in presence of cycles like x[a]x. For instance, the following constraint is F-closed. In particular, F2 and F4 do not loop through the cycle x[a]x infinitely often. This example also illustrates why the fixed point algorithm would not be terminating if based in the axiom x[a]x Proposition 3. If j is a constraint with m variables then algorithm F with input j terminates under the above control in at most 2 Proof. Since F does not introduce new variables, it may add at most m 2 non-disjointness constraints x-y and m 2 inclusions x-y. 2 Proposition 4. Every F-closed constraint j is satisfiable over FT- . Proof. See Section 8. 2 Theorem 5. The satisfiability of FT- constraints can be decided in time O(n 3 ) (offline and online;see [11]) where n is the constraint size. Proof. Proposition 2 shows that j is unsatisfiable if F started with j terminates with false. Proposition 4 proves that j is satisfiable if F started with j terminates with a constraint. Since F terminates for all input constraints under the above control (Prop. this yields a effective decision procedure. The main idea of the complexity proof is that one needs at most O(n 2 ) steps (Prop. 3) each of which can be implemented in time O(n). The implementation can be organized incrementally by exploiting that algorithm F leaves the order unspecified in which the axioms are applied. Hence, we obtain that off-line and on-line complexity are the same. The implementation details and the complexity proof are omitted here, since they are similar to those presented in [17]. 2 5 Entailment, Independence, Negation In this section, we give a cubic algorithm testing entailment j 0 We then prove the independence property of FT- . Hence we can solve conjunctions of positive and negative FT-constraints j-:j A basic constraint - is a conjunction free constraint j, i.e., given by the following abstract syntax: The entailment j 0 is equivalent to the fact that the entailment j j=- holds for all basic constraints - in j. Next we characterize entailment problems j 0 syntactically. We say that a constraint syntactically contains -, written j ' -, if one of the following holds: exist x 0 , y 0 such that x 0 [a]y 0 in j; We say that a first-order formula F syntactically contains -, F ' -, if some j and F 0 such that j ' -. Lemma 6. Given a F-closed constraint j, we can compute a representation of j in linear time that allows to test syntactic containment j ' - for all - in time O(1). Proof. Simple. 2 It is easy to see that syntactic containment is semantically correct, i.e., j ' - implies -. For deciding entailment, we have to show that our notion of syntactic containment is semantically complete, i.e., if j 6' - then j 6j= - (Proposition 13). The idea is to construct a satisfiable extension of j (its saturation) which syntactically and simultaneously contradicts all - not syntactically contained by j (Lemma 12). Saturation is defined in terms of two operators G 1 and G 2 on constraints. The operator G 2 is such that G 2 (j) contradicts all - of the form x-y, x-y, and a(x) (i.e., no selection constraints) which are not syntactically contained in j (Lemma 10). The operator G 1 serves for contradicting selection constraints. For instance, consider j y-y. In this case, G 1 (j) enforces the existence of the feature a in the denotation of x by adding to j the constraint x[a]v xa for a fresh variable v xa . Now G 2 (G 1 (j)) is such that it contradicts either y-v xa or v xa -y. (see Example 4). In this sense, G 1 is a "preprocessor" for G 2 . Definition 7. Let j be a constraint, v 1 and v 2 distinct fresh variables, and l 1 and l 2 distinct labels. Furthermore, for every pair of variables x; y 2 V (j), and label every label a 2 L(j) let l x and l xy be fresh labels and v xa a fresh variable. We define G 1 (j) and G 2 (j) in dependence of Example 4. Consider the constraint which is F-closed up to trivial constraints and which does not entail x[a]y. In order to contradict x[a]y we compute the F-closure of G 1 (j 0 ) which is G 0 v xa -x-x-v xa -y-v xa and observe that it does not v xa -y. By definition of G 2 , G 2 (G 0 contradicts v xa -y. Hence, G 2 (G 0 x[a]y. Lemma 8. Let j be an F-closed (and hence satisfiable) constraint. Then G 1 (j) is satisfiable and its F closure G 0 satisfies the following two properties for all basic constraints -: 1. If j 6' - and V (-) 'V (j), then G 0 2. If j 6' x[a]y then G 0 Proof. The F-closure G 0 1 (j) of G 1 (j) has the following form up-to trivial constraints and symmetry of compatibility constraints. exist exist exist instance note that v xa -x 0 - x 0 -v xa in G 0 by clauses (4.2, 4.3) and reflexivity). All constraints in G 0 either belong to G 1 (j) or a derived from it by axioms in F. The F-closedness of G 0 1 (j) can be proved by a somewhat tedious case distinction. The same holds for the two additional properties of G 0 Lemma 9. If j is F-closed then G 2 (j) is satisfiable. Proof. It is not difficult to show that the constraint part of G 2 (j) is F-closed up to trivial constraints (x-x and x-x) and symmetric compatibility constraints. The critical bit is to check that the negated selection constraints added in clause (1) of G 2 (j) are consistent. Let :9y 0 (y[l x ]y 0 ) in G 2 (j). We must show that G 2 (j) 6j= 9y 0 (y[l x ]y 0 ). Assume the converse, G 2 (j) by Corollary 27 in Section 8, there exist z and z 0 such that G 0 By definition of G 2 (j) we know that z = x. However, (by definition of G 2 clause (1) does not apply. Thus :9y 0 (y[l x ]y 0 ) cannot be contained in G 2 (j), in contradiction to our assumption. 2 Lemma 10. Let j be an FT-constraint and let - be a basic constraint of the form x-y, x-y, or a(x) (i.e., not a selection constraint). Then G 2 (j) Proof. By inspection of the definition of G 2 (j). Clause (1) contradicts entailment of x-y by j by forcing x to have a feature l x which y must not have. Clause (2) contradicts x-y by forcing x and y to have a common feature l xy such that the subtrees of x and y at l xy are incompatible. Clause (3) contradicts a(x) for any label by forcing x to be unlabeled (i.e., compatible with at least two trees with distinct label). 2 Definition 11 Saturation. Let j be an F-closed constraint and G 0 1 (j) the F-closure of exists according to Lemma 8. The saturation of j is the formula Sat(j) given by Lemma 12. Let j be an F-closed constraint For all - such that V (-) ' V (j), j 6' - implies Proof. Let G 0 1 (j) the F-closure of G 1 (j) such that - is not a selection constraint, then G 2 (G 0 by Lemma 10. Otherwise, let Lemma 8.2. By Lemma 10, either G 2 (G 0 holds. In both cases, G 2 (G 0 Proposition 13. The notions of entailment and of syntactic containment coincide for basic constraints: If j is F-closed and - a basic constraint then j Proof. We assume j (The converse is correctness of syntactic con- tainment.) If V (-) 6' V (j) then - is of the form x-x or x-x such that j ' -. Otherwise, since Sat(j) contains j. Moreover, Sat(j) is satisfiable (Lemmas 8 and 9) such that Sat(j) 6j= :-. Hence, j ' - by Lemma 12. 2 Theorem 14 Entailment. Entailment problems of the form j 0 can be tested in cubic time. Proof. Let n be the size of j 0 -j. To decide j 0 first test whether j 0 is satisfiable. By Theorem 5 this can be done by computing the F-closure - this test fails then the entailment test is trivial. Otherwise, from Lemma 12 we obtain and hence that - j. There are O(n) such - and - 0 is of size O(n 2 ), hence, by Lemma 6, this is decidable in time O(n). The overall complexity sums up to O(n 3 Theorem 15 Independence. The constraint system FT- has the independence prop- erty; i.e., for every n - 1 and constraints j;j Proof. Assume unsatisfiable we are done. Also, if satisfiable for some j, then j is. Now let j and j-j i be satisfiable for all i and let j be F-closed (wlog. by Prop. 2). If there exists i with for all - syntactically contained by j i , then j and we are done. Otherwise, for all i there exists - i such that j 6' - i . Lemma 12 yields Sat(j) satisfiable (Lemma 8) and entails j, this contradicts our assumption that j Corollary Negation. The satisfiability of conjunctions of positive and negative FT- constraints j-:j 1 -:j k can be tested in time O(n 3 ) where n is the size of the given conjunction. Proof. If j is non-satisfiable then j- n trivially non-satisfiable. By Proposition 5, satisfiability of j is decidable in time O(n 3 ). Now assume j to be satisfiable. By the Independence Theorem 15, j- n nonsatisfiable if and only if j some i. By Lemma 12 this is equivalent to the existence of i such that for all - if -. Overall, there are O(n 2 ) candidates - to be tested for syntactic containment and O(n) possible can be tested in time O(1) such that the total complexity sums up to time O(n 3 6 Weak Subsumption Constraints We next introduce weak subsumption constraints that are used in computational linguistics We show that their satisfiability problem is subsumed by the one for FT- . Syntax. We assume given a set C of constants c and a set D of features d. We consider the set of labels subsumption constraint h is a FT- constraint of the following form. Note that compatibility constraints do not occur in [7]. We add them here to simplify our comparison. Semantics. We interpret weak subsumption constraints over the whole class of feature algebras with the induced weak subsumption ordering, which we will define below. A feature algebra A over C and D consists of a set dom A that is called the domain of A , a unary relation c() A on dom A for every constant c 2 C , and a binary relation [d] A on dom A for every feature d 2 D, which satisfy the following properties for all a;a 0 ; a 00 2 dom A , constants c; c 1: if a[d] A a 0 and a[d] A a 00 then a 2: if c 1 (a) A and c 2 (a) A then c In the literature [22, 7] a slightly different notion of feature algebras with constants has been considered. We will give a formal comparison between the two notions at the end of the present section. Proposition 17. The structure F over L is a feature algebra over C and D. Proof. The above properties follow from the axioms in F and the antisymmetry of the information ordering in FT- (x-y- y-x ! x=y). 2 Given a feature algebra A , we define the weak subsumption ordering - A as follows. A simulation for A is a binary relation D on the domain of A that satisfies the following properties for all elements a 1 , a 2 , a 0 2 of A's domain: 1: if a 1 2: if a 1 Da 2 ; a 1 [d] A a 0 2 then a 0 subsumption ordering - A of A is the greatest simulation relation for A . The subsumption relation on A induces a compatibility relation - A : exists a such that a 1 - A a and a 2 - A a A feature algebra A induces a structure with the same signature as F , in which - is interpreted as weak subsumption ordering - A , - as - A , c() as c() A , and [d] as [d] A . Proposition [8]. The structure F coincides with the structure induced by the feature algebra defined by F . Proof. It is sufficient to prove that the weak subsumption relation of the feature algebra defined by F coincides with the information ordering on F . The proof in the case for feature algebras with constants can be found in [8] on page 24 (Satz 6 and Satz 7). There the algebra of feature trees has been called algebra of path functions. A direct proof (additional 5 lines) is omitted for lack of space. 2 Theorem 19. A weak subsumption constraint h is satisfiable (over F ) if and only if h is satisfiable over the structure induced by some feature algebra A . Proof. If h is satisfiable then it is satisfiable over the structure induced by the feature algebra defined by F . Conversely, every structure induced by a feature algebra is a model of the axioms in F. Thus, if h is satisfiable over one such structure then it is equivalent to an F-closed constraint (and not false) and hence satisfiable over F . 2 Alternative Notions of Feature Algebras. In the literature [22, 7] a restricted notion of feature algebra has been considered that we call feature algebra with constants in the sequel. The focus on feature algebras with constants leads to a restricted satisfiability problem. This shows that the presented results properly extend the results in [7]. A feature algebra with constants is a feature algebra with the additional property that if c(a) A then not a[d] A a 0 (1) In order to handle the new property we consider the following mapping of weak subsumption constraints over C and D to weak subsumption constraints over C [ flabelg and D where label is a new constant not contained in C . Proposition 20. A constraint h is satisfiable in some feature algebra if and only if [ is satisfiable in some feature algebra with constants. Proof. If [ is satisfiable over a feature algebra A with constants C and features D [ flabelg then h is satisfiable over the feature algebra F with labels C [D. Given a solution s 0 of [ over A a solution s of h over F can be defined as follows: exists a in domain of A : s 0 (x)[p] A a and p 2 D g exists a in domain of A : s 0 (x)[plabel] A a and c(a) A g Conversely, let h be satisfiable in a feature algebra A . Then h is satisfiable in F by Theorem 19. We consider the following feature algebra with constants F and show that [ is satisfiable over F . The constants and features of F are C and D[flabelg, respectively. The domain of F contains all feature trees t without labeled internal nodes where a labeled internal node of t is a path p such that p 2 D t , exists c with not exists d with pd 2 D t . The selection and labeling relations of F are those of FT- restricted to trees without internal labels. Obviously, F satisfies all three axioms of a feature algebra with constants. Now let s be an A-solution of h. Then the variable assignment s 0 mapping x on s 0 (x) as given below is an F -solution of [ exists a 7 Expressiveness We show that FT- is strictly more expressive than FT but that FT- cannot express an constraint. An FT constraint h is of the form x=y, a(x), x[a]y, or h-h 0 , and an arity constraint of the form xfa g. An arity constraint xfa holds if x denotes a tree with subtrees at exactly a 1 through a n . Proposition 21. There is no FT- constraint which expresses that a variable x denotes the empty feature tree, i.e., if a 6= b then there is no constraint equivalent to Proof. If j were such a FT- constraint, then j as well as its finite F-closure would entail x-y for all variables y. This contradicts Proposition 13 for all those y such that (j)). Such a variable y exists since V (j) is finite. 2 Lemma 22. Let h be an FT constraint. Then h Proof. The FT constraint h is equivalent to the FT- constraint j obtained from h by replacing all equalities x=y by inequalities x-y- y-x. Hence, x-y in j iff y-x in j, and since algorithm F preserves this invariant it also holds for the F-closure of j. The claim follows from Proposition 13. 2 Proposition 23. If x 6= y then there is no FT constraint h equivalent to x-y. Proof. This follows immediately from Lemma 22 and Proposition 13. 2 8 Completeness of the Satisfiability Test Proposition 4. Every F-closed constraint j is satisfiable over FT- . The proof is based on the notion of path reachability and covers the rest of the sec- tion. We proceed as follows. We first define path reachability, then give two Lemmas, and finally compose the proof of Proposition 4 from these Lemmas. For all paths p and constraint j, we define a binary relation j "y is reachable from x over path p in j": a y if x[a]y in j; a meaning that "a can be reached from x over path p in j": For example, if j is the constraint x-y - a(y) - x[a]u - x[b]z - z[a]x - b(z) then the following reachability propositions hold: y j y, x j as x j Definition 24 Path Consistency. We call a constraint j path consistent if the following two conditions hold for all x, y, p, a, and b. 1. If x j 2. If x j Lemma 25. Every F1-F2-closed and path consistent constraint is satisfiable. Proof. Let j be F1-F2-closed and path consistent. We define the variable assignment trees as follows: ag The path consistency of j condition 1 implies that L min j (x) is a partial function. Thus is a feature tree. We now verify that min j is a solution of j. - Let x-y in j. For all x 0 , if y j by the definition of path reachability. Thus, D min j (y) 'D min j (x) . For all a if y j a by the definition of path reachability. Thus, L min j (y) ' L min j (x) , i.e., min j (y)-min j (x). Consider x[a]y in j. We have to prove for all p, z, and b the equivalences The first equivalence is equivalent to D min j and the second one to L min g. We start proving the first equivalence. If y j ap z since x[a]y in j. Suppose x j ap z. By definition of path reachability there exists x 0 and y 0 such that The F1-closedness of j and x j imply x-x 0 in j. The F2-closedness ensures y-y 0 in j such that y j holds. We now prove the second equivalence above. If ap b then there exists z such that x j ap z and b(z). The first equivalence implies z and thus y j b. The converse is simple. - Let a(x) in j. Reflexivity (F1:1-closedness) implies x-x in j. Thus x j e a such that (e; a) 2 L x . - Let x-y in j. We have to show that the set L min j (x) [L min j (y) is partial function. If a and y j b. The path consistency of Lemma 26. Every F3;F4;F5-closed constraint is path consistent. Proof. Let j be F3;F4;F5-closed. Condition 1 of Definition 24 follows from condition 2 of Definition 24 and F3:1-closedness. The proof of condition 2 is by induction on paths p. We assume x, y, a, and b such that x j in j, and x j e, then there exist n;m - 0, x F3-closedness implies that x n -ym in j (F3:2 yields x-y 1 in in j. Therefore y m-x in j by F3:3-closedness, and hence y m-x 1 in closedness.) Hence, F5-closedness implies a = b. In the case there exists there exist x 0 , y 0 , - y x in y in Since x-x 0 in j we have x 0 -y 0 in j by F3-closedness (as above). Thus, F4-closedness implies - x- y in j such that a=b holds by induction hypothesis. 2 Proof of Proposition 4. If j is F-closed then j is path consistent by Lemma 26 and thus satisfiable by Lemma 25. 2 Corollary 27. Let j be an F-closed constraint. Then j only if there are variables x 0 and y 0 such that j ' x 0 [a]y 0 and j ' x 0 -x. Proof. Assume holds for the minimal solution min j of an F-closed constraint that a 62 L min j (y) . Hence j 6j= 9y(x[a]y). 2. Acknowledgments . We would like to thank Jochen D-orre, Gert Smolka, and Ralf Treinen for discussions on the topic of this paper. We would also like to acknowledge many helpful remarks of the referees. The research reported in this paper has been supported by the the Esprit Working Group CCL II (EP 22457) the SFB 378 at the Universit?t des Saarlandes. --R A Complete Axiomatization of a Theory with Feature and Arity Constraints. A complete and recursive feature theory. In 2 nd Future Generation Computer Systems On Subsumption and Semiunification in Feature Algebras. Constraint logic programming: A survey. Functional Grammar. Applications of a Canonical Form for Generalized Linear Con- straints Ordering Constraints over Feature Trees with Ordered Sorts. Entailment for Set Constraints is not Feasible. Inclusion Constraints over Non-Empty Sets of Trees Feature Logics. An Introduction to Unification-based Approaches to Grammar Parsing and Type Inference for Natural and Computer Languages. Feature constraint logics for unification grammars. The Oz Programming Model. Records for Logic Programming. Feature constraints with first-class features --TR --CTR Franois Pottier, A versatile constraint-based type inference system, Nordic Journal of Computing, v.7 n.4, p.312-347, Winter 2000 Jens Palsberg , Tian Zhao , Trevor Jim, Automatic discovery of covariant read-only fields, ACM Transactions on Programming Languages and Systems (TOPLAS), v.27 n.1, p.126-162, January 2005 Ernst Althaus , Denys Duchier , Alexander Koller , Kurt Mehlhorn , Joachim Niehren , Sven Thiel, An efficient graph algorithm for dominance constraints, Journal of Algorithms, v.48 n.1, p.194-219, August Michael Benedikt , Leonid Libkin , Frank Neven, Logical definability and query languages over ranked and unranked trees, ACM Transactions on Computational Logic (TOCL), v.8 n.2, p.11-es, April 2007
weak subsumption;feature constraints;complexity;tree orderings;satisfiability;entailment
593316
Solving Various Weighted Matching Problems with Constraints.
This paper studies the resolution of (augmented) weighted matching problems within a constraint programming (CP) framework. The first contribution of the paper is a set of techniques that improves substantially the performance of branch-and-bound algorithms based on constraint propagation and the second contribution is the introduction of weighted matching as a global constraint ( WeightedMatching), that can be propagated using specialized incremental algorithms from Operations Research. We first compare programming techniques that use constraint propagation with specialized algorithms from Operations Research, such as the Busaker and Gowen flow algorithm or the Hungarian method. Although CP is shown not to be competitive with specialized polynomial algorithms for pure matching problems, the situation is different as soon as the problems are modified with additional constraints. Using the previously mentioned set of techniques, a simpler branch-and-bound algorithm based on constraint propagation can outperform a complex specialized algorithm. These techniques have been applied with success to the Traveling Salesman Problems [5], which can be seen as an augmented matching problem. We also show that an incremental version of the Hungarian method can be used to propagate a WeightedMatching constraint. This is an extension to the weighted case of the work of Rgin [19], which we show to bring significant improvements on a timetabling example.
Introduction Constraint Logic Programming (CLP) and more generally Constraint-based Programming (CP), has become an interesting approach for solving combinatorial optimization problems [VH 89]. For instance, on problems such as jobshop scheduling it has shown to be as efficient as more traditional approaches based on Operations Research (OR) [CL 94] while delivering more flexibility (i.e. it is easy to adapt the algorithm to a slightly customized problem with additional constraints). Our goal is to investigate the advantages and drawbacks of constraint programming compared to classical combinatorial optimization on the maximum weight matching problem in a bipartite graph and see how these techniques may be combined. Bipartite matching is interesting and important for three reasons. First, it is a well-understood problem, for which many polynomial algorithms ranging from simple to really elaborate (e.g., [GT 89]) are available. Therefore, the performance of the constraint satisfaction algorithm can be assessed against a rigorous gauge. Second, many real-life problems are either matching or "extended" matching with one or a few additional constraints. It is, therefore, a perfect domain to see where the boundary between flexibility and efficiency should lay. Moreover, many such matching problems are optimization problems with weights, and it may be more indicated to use weighted matching algorithms than to use generic optimization methods with non-weighted techniques. Third, matching is often a part of more complex problems and the efficient propagation of matching constraints is a basic building block of a modern constraint satisfaction system. By looking carefully into matching algorithms, we derive ideas that will be of use to designers of constraint solvers. The paper is organized as follows. Section 2 describes a "naive" approach to matching and a few improvements that can be implemented easily. We then move on to more complex heuristics and cutting rules and show how they can be used to handle larger problems. Section 3 compares the previous results with two well-known algorithms, the Hungarian (weighted) method and the Busaker & Gowen flow algorithm. We show that the constraint satisfaction approach does not really compare with these specialized algorithms when the problem is to find an optimal matching. On the other hand, it is a plausible approach for finding a set of "very good" solutions. This leads to Section 4, where we look at two extended matching problems: in the first case, we allow the substitution of a new cost matrix for a few (k) edges (machine-task allocation problem with renewal of k machines); in the second case, we study the resolution of two simultaneous matching problem (e.g., in a machine-task allocation problem, maximize production within a bounded energy consumption). For these two variations of the problem, the original algorithms are adapted and compared. Last, we apply in Section 5 the various techniques that were presented in this paper to a time-tabling problem. We show that we can use weighted matching as a global constraint, which captures hard and soft constraints at the same time, enabling a more efficient resolution. 2. Constraint based programming Formally, the matching problem can be described as follows : Let G=(V,E) be a bipartite graph, the vertex set V can be parted in two shores such that all edges link V 1 to V 2 . A matching is a set of edges M - E such that no two of them are adjacent. A perfect matching is a matching covering all vertices. We consider the case which a perfect matching corresponds to an assignment (such matchings frequently arise in production planning, for the assignment of jobs to machines or tasks to technicians). Moreover, we consider the case of a weighted graph and define the weight of a matching as the sum of the weights of its edges. The purpose is then to find, when it exists, a perfect matching of maximal weight (or of minimal weight if all weight w ij are replaced by A- w ij ), for some large constant A). 2.1. Naive CLP A natural way to see this problem as a constraint program is to think of the vertices in V 1 as variables and to those in V 2 as values. A vertex in V 1 is linked to the set of values in V 2 which form its domain. A matching corresponds to an assignment for the variables such that different variables are always assigned different values. The problem consists in finding an assignment of maximal weight. This constraint (AllDifferent) [R-94] is offered as a built-in in most CLP languages. Propagation is often performed by arc-consistency [Ma77], [MM 88]. when a value is assigned to a variable, it is taken out of the domains of the other variables. When a contradiction occurs (because a domain becomes empty), the system backtracks to the previous variable assigned and tries the next possible assignment. For optimization, the algorithm is a classical branch and bound scheme : a variable is selected and all possible values are recursively tried. Once a solution M of weight W has been found, the constraint weight > W is added to the system to help pruning subtrees of the search space using simple upper bounds for weight. At any time during the search, two upper approximations of the maximal matching are maintained : the sum s + for all variables of their best possible assignment (for each variable v, we define the value best(v) as the maximal value weight(v,x) for all possible values x) and reciprocally the sum s - for all values of their best possible inverted assignment. Instead of computing these sums explicitly for each node in the search tree, they are maintained incrementally. Each time a value x is removed from the domain of a variablev, if weight(x,v)=best(x) then best(x) is re-computed and the difference between the old and the new value is subtracted from s + . Similar updates are performed on best(v) and s - if weight(x,v)=best(v). Whenever one of the bounds s or s - becomes strictly smaller than the goal, one can predict that the upcoming solutions will have a smaller weight than the best one found so far, so this branch of the tree can be cut off. This propagation algorithm is widespread among the implementations of the alldifferent constraint. In a CLP system, the selection of the node on which to branch is usually done with the first-fail principle : one selects a variable with the smallest domain. Assignments are usually tried in decreasing order of weight to drive the search towards matchings with heavy weights (best first principle, greedy heuristic). As shown in table 1, this naive algorithm solves problems up to 2x20 nodes within seconds but requires unreasonable amounts of time for larger problems. Although it may not be the case for all implementations of the alldifferent constraint, taking both upper bounds is important. The program which just considers s + as upper bound is 10 times slower on 10x10 instances and cannot solve 20x20 problems within reasonable times. This fact gives us a hint of the progress that a symmetrical view of the problem can bring to a constraint approach. The next two sections (2.2 and 2.3) describe improvements that can be made to the propagation scheme and the heuristics (first fail, best first) in order to improve performance. 1 All the experiments reported in this paper have been done with the CLAIRE programming language 2 . 1 For programmers using a blackbox constraint system with no access to the propagation and search engine (as many CLP systems), these changes could only be encoded within the solver: this allows to state the problem with the single command alldifferent, but the changes can only be written by an expert of the constraint system. For programmers using an open constraint system with access to the solver and the propagation engine, the program is a little longer but any user can program these changes. 2 available at http://www.dmi.ens.fr/~laburthe/claire.html 2.2. Adding symmetry and regret Many improvements can be made on this algorithm to make it scale up for larger problems. The first idea suggested by the importance of the symmetrical upper bounds is to have a fully symmetrical description of the problem. The model that considers vertices of V 1 as variables and vertices of V 2 as values is an arbitrary breech to the symmetry of the problem (although perfectly valid). A way around this is to add the redundant model where V 2 is seen as the set of variables and V 1 as the set of values. Branching both on vertices from V 1 and V 2 implies some overhead (an inverse of the domain relation must be maintained), but it avoids missing evident decisions (for instance, when a value is in only one domain, it needs to be assigned to that variable). Hence, the matching is done from both sides at the same time. This is very similar to the use of symmetric redundant models for the n-queens problem [Jo 95], as would be expected since the n-queens is a customized matching problem (without weights). Moreover, the first-fail heuristic for the selection of the vertex on which to branch is not a panacea : The notion of regret gives a more acute description of the crucial vertices. The regret is usually defined as the difference between the optimal choice and the second best. In our case, if v is a vertex from V 1 and x is a value such that best(v)=weight(v,x), we define regret(x) as the difference between best(v) and the maximum of weight(v,x') for all values x'-x. Focusing on regret is justified by the fact that problem solutions do not change if a constant is added to all edges incident to one vertex (thus, only relative values matter, as opposed to absolute ones). The regret is actually the change that will be made to the upper bound if the best assignment for this vertex is discarded. Selecting the vertex with the largest regret as a branching point is an entropic choice in the sense that we try to maximize the impact of this decision. As described in [CL 94], entropic heuristics are a powerful tool for proofs of optimality since they limit the size of the tree by forcing much propagation to happen at each node. To avoid a real slow down of the algorithm at each node, the regret and the best assignment for each node are memorized and updated incrementally. Another idea would be to somehow re-break the symmetry. At any time, two upper bounds for the matching, s are available. As soon as one of them becomes less than the goal, the branch can be cut. Selecting vertices in V 1 with large regret decreases conversely. In cases when one of the bounds is clearly better than the other, it could seem worthwhile to encourage this dissymmetry in the choice of the branching node. We implemented this by giving a bonus (or penalty, depending of the situation) during the selection of the branching vertex, proportional to s to the vertices in V 1 . This heuristic is denoted as "balance" in table 1. The table below illustrates the effect of these techniques on randomly generated instances for 2- 20 and 2 - each vertex being typically connected to 5 to 8 edges, with weight a random number between 1 and 100. The entries of the table show the number of backtracks (b.) in the search process (1kb.=1000b.) and the running times (on a Pentium Pro 200) for finding an optimal solution and giving the proof of optimality. The naive approach (arc consistency and first fail) behaves poorly, the addition of regret and symmetry brings a gain of a factor 10. The idea of the balance does not really pay off. Naive Propagation 1.2 s. 52 kb. 500 s. 5 kb. 9 kb. 198 kb. b. 800 b. 3,5 kb. 800 b. 1.2 s. 9.5 s. 2,7 Mb. Basic 900 b. 3,5 kb. 900 b. 43 kb. 2.5 s. 1,1 Mb. Table programs 2.3 Finer tuning of the algorithm One of the aspects of the algorithm that can be sharpened is propagation. In addition to upper bound evaluation and domain reduction due to edge selection, we can also do some domain reduction because of the upper bound. As a matter of fact, we can dynamically remove all illegal edges. An edge uv is illegal as soon as (best(u) - weight(u,v)) is greater than the available slack (goal - s goal is the objective weight (goal is subsequently decreased to time an admissible solution M is found). We call this technique dynamic cut and report its behavior in table 2. The interest of dynamic cut is two- folds. On the one hand we detect some failures a little earlier (a rather small benefit). On the other hand, we maintain smaller and more relevant domains, which may improve the efficiency of the first-fail principle. Removing meaningless values produces more rapidly the domains with very small cardinals (1 to 3) that first-fail tells us to examine. If we want to re-introduce some of the first-fail behavior into our choice heuristic based on regret, we can use a simple trick (used in [CGL93]): select the vertex that will minimize the lexicographic pair where par is a fixed parameter. The effect is to select vertices by largest regret while there are still domains with cardinal more than par and to select them by first fail when all domains are smaller than par. This strategy yields some improvements, but only for small values of par (3). The search tree, as described now has many branches per node. Actually, it may be a waste of time to wonder what edge should be selected for a vertex, once we know that it will not be the best one, since these edges may well look alike and be numerous. The search tree can be constructed differently with a single alternative at each node : either the best edge is the selected one, or it is not. The tree becomes narrower (binary), maybe deeper, but in practice, it prevents the algorithm from wasting time in irrelevant choices. This new branching scheme is reported in table 2 as binary branching. The other tuning that can be done concerns the upper bound that is used for cutting branches of the search tree. Our formula that sums all best choices estimation of the best assignment. Indeed, if v -V 1 is the favorite choice of p vertices in V 2 , we know that we will have to take into account the regret for atleast p-1 of them (because they cannot all be matched to their common favorite vertex). Therefore, it is possible to tighten the bound by subtracting from s + a conservative estimate of the regrets that will occur (this is an application of the lookahead principle because we try to forecast the evolution of the bounds). For each vertex v -V 1 , we consider the previous set of vertices x from V 2 such that best(x). To this set of p vertices, we associate the sum of all regrets but the highest one, that we call look(v). s + . can be replaced by This strategy which estimates the difference between the simple bound and the actual choice is called lookahead in table 2. Note that the estimation s corresponds to a relaxation of the matching problem. This could suggest to use Lagrangean relaxation by affecting weights to nodes. In fact, this lookahead strategy uses similar ideas by affecting weights to nodes that are related to several others ones in the preference graph (these nodes which prevent the solution of the relaxed problem to be a matching). Moreover, the Hungarian method which associates weights to nodes can also be seen as related to Lagrangean relaxation. basic 1.2 s. 9.5 s. 2,7 Mb. dynamic cuts 24 kb. 1,5 Mb. 2.2 Mb. new branching 1.2 s. 95 kb. 6.5 s. lookahead 4,5 kb. 1.2 s. 6 s. Table A few remarks can be made from these results. First, the lookahead heuristic works well, although its benefits (obvious from the numbers of backtracks) only pay off when the size of the problem is large enough. On the other hand, both the dynamic cut and binary branching schemes produce unstable results and it is difficult to measure any significant improvement. These techniques (propagation, branching, bouding, heuristics) have been succesfully applied to the Traveling Salesman Problem (TSP) which can be represented as a weighted matching problems (trying to match each node with its direct successor in the cycle) coupled with a subtour elimination constraint. As described in [CL97], these improvements enable constraint programs to solve 30-city tours instead of 15-city tours (more complex techniques are also given for larger problems). 3. Traditional algorithmic approaches Matchings have been widely studied by the Operations Research community for years ([CoL 86], [Ge 94]). The two methods presented here do not pretend to be the best ones available today. We made a compromise by limiting ourselves to algorithms which were simple enough to be understood and implemented in a reasonable amount of time, not out of proportion with the implementation times of constraint-based programs . The algorithm of Busaker and Gowen (described in [GM 79], [VL 90]) embeds the matching problem into a flow problem by adding a source linked to all vertices in V 1 and a sink linked to all vertices in V 2 . In this case, the capacities on all edges are set to 1. It is an adaptation of the well-known Ford&Fulkerson maximum flow algorithm to the case of a network with weighted edges. It starts with a null flow and augments it by saturating it along augmenting paths until no augmenting path can be found, meaning that the flow is maximal. This polynomial algorithm runs in is the number of edges and N the number of nodes. However, unlike the Ford&Fulkerson it is not incremental in the sense that it cannot complete any partial flow into a minimal cost maximal flow (it requires the partial flow to be of maximal stream for its cost). . The Hungarian method is an algorithm of the class of primal-dual algorithms often associated to problems stated as linear programs [PS 82]. The primal program consists in finding a maximal matching of minimal weight. The dual program consists in finding positive weights p u associated to vertices u such that for all edges uv, and such that the sum of the p u is maximal. The algorithm constructs incrementally the weights p u , starting with all p u equal to 0, and constructing the graph G p formed by the edges uv verifying p u . The algorithm alternatively works in the dual model (updating the p u in order to add new edges to the graph G p ) and the primal model (finding augmenting paths in G p ). In the end, the matching in G p is a minimal weight maximal matching in G. This polynomial algorithm runs in O(N 3 ). Its major interest compared to the flow algorithm is its incrementality. When one wants to remove an edge uv from the graph and recompute the optimal solution, one just sets weight uv to a large positive constant and remove uv from G p . It takes then only one iteration to complete the optimal matching. 3.1. Comparison with CLP For finding an optimal solution to a pure matching program, CP is not competitive with these OR techniques. CP solves reasonably well problems with up to 2 - 40 or nodes, whereas the Hungarian method scales up to problems well over 2 - 100 nodes (on our examples, solved in 70 or 80 000 backtracks and approximately 200 s. by our best constraint programs, the flow algorithm takes 1 s. and the Hungarian method takes 0,3 s. However, if one no longer considers the problem of finding one optimal solution, but all optimal solutions (or rather all solutions within a given distance of the optimum), then CLP becomes a plausible competitor. Table 3 reports this experiment: four programs are compared: . The first one is just a straightforward adaptation of our previous constraint program basic. Instead of stopping at the first solution, we explore the whole tree. . The second is a similar branch and bound exploration where the flow algorithm is used to evaluate the optimal matching at each node instead of using an upper bound estimate. The flow algorithm is actually only triggered when one of the edges of the current "optimal" matching is removed from this dynamic graph. . The third program is similar with a different branching scheme. For each edge of the original optimal solution we explore two branches : either the edge is part of the matching or it is removed from the dynamic graph. . The fourth program is a branch and bound algorithm similar to the third one, except that the Hungarian method is used instead of the flow algorithm. We here take advantage of the incrementality of the Hungarian method (its ability to recompute the optimal matching in a single iteration when an edge is discarded at a node of the search tree). For this exploration, the bottleneck is the amount of computation performed at each node of the search tree. The incremental Hungarian method is the best algorithm since it explores very small search trees (all explored branches lead to an admissible solution), however, the constraint algorithm is a fair competitor because it performs much less work per node. 1% 4 sol. 2% 9 sol. 5%sol. 0,5%. 3 sol. 1% 8 sol. 2,5% sol. Basic 1249 b. b. 8000 b. 14 kb. 1.1 s. 38 kb. 2s. 140kb. Flow algorithm 137 b. 9 s. 224 b. 2000 b. 100 s. 100 b. 26 s. 172 b. 215 b. 48 s. 1,5kb. 200 s. Flow algorithm with new branching 57 b. 136 b. b. 34 b. b. 22 s. 195 b. Hungarian matching b. 58 b. b. b. b. b. b. Table finding all solutions Hence for the problem of finding all near-optimal perfect matchings, the constraint program behaves better than our basic Operations Research program (the adaptation of the flow algorithm) and not as well as our smarter O.R. program (incremental Hungarian method). The big advantage of the constraint program is it simplicity: it is indeed much easier to implement than both of the O.R. algorithms (naive and smart). It seems therefore to be good candidate for real-life matching problems involving a few additional constraints (and which thus require more exploration of the space of solutions than the pure problem). 4. Complex matching In this section, we address the case of two real-life variations of the matching problem. Ad-hoc algorithms could also probably be specially designed for these variations of the problem. However, from a methodological point of view, we decided to consider only adaptations of the original solutions that could be implemented in a reasonable amount of times (since the constraints programs are adapted in a matter of minutes). 4.1 Replacing k machines Suppose that in a plant, the manager has the funds to replace k machines among the n, by newer ones. We want to find an optimal assignment (maximum bringing maximal production) of tasks to machines, given that newer machines are more efficient than older ones. This means that the matching problem is extended with the additional choice of those k vertices in V 1 for which we prefer to use a different weight matrix (say weight' instead of weight). For the sake of comparison, we implemented three solutions for this problem, one using constraint propagation and two O.R. solutions, a basic adaptation of the flow algorithm and a smart adaptation of the Hungarian method: . The first one is a straightforward adaptation of our constraint-based program basic. The principle is to postpone the choice of the k "new" vertices as much as possible, but to take them into account with a modified upper bound estimate. Instead of taking the sum of the best choices for all nodes, we add to this sum s the sum of the k heaviest edges in the "difference graph" (defined by the gain in the production function between new and old machines, i.e. considering weight'(uv) - weight(uv) as the weight of the edge uv). Like the bounds s can be maintained incrementally. . The second program is based on the flow algorithm. It goes through a search tree where the nodes at depth i correspond to the choice of the i th machine to replace. All branches are explored. Whenever a decision of replacement is made, the production function is updated (i.e., the second weight function is used for the i th machine). At each node, the upper bound (computed for pruning) is the sum of the optimal solution for a matching in the original graph (where i of the k decisions have been made) and of the optimal solution for a matching of cardinal k-i in the difference graph. . The third program is based on the Hungarian method. The algorithm goes through the same search tree as the program looking for all near optimal solutions (cf section 3.3) : at each node a vertex v from V 1 selected and each branch matches one of the vertices x of by removing all edges vx' for x' - x. The first branch explored corresponds to the current assignment of v in the solution given by the Hungarian method. The upper bound is the sum of the best possible matching (given by the Hungarian method) with an estimate of a sub-matching of size k in the "difference graph" (same estimate as for the constraint algorithm) Constraints 266b 0,1s 433b 0.1s. 735b 0.1s. 19kb 2.1 s. 26kb 9.9 s. 4,4Mb 700 s. Flow 5 s. Hungarian Matching 20b. 22b. 31b. 30b. 30b. 34b. 40b. 71b. 107b. Table replacing k machines Figure 4 shows these three strategies on a few problems. The constraint-based approach performs better than the approach based on the flow algorithm but not as well as the algorithm using the Hungarian method. The constraint approach is much faster than this smart O.R. algorithm for n=10, a little faster for n=20, somewhat slower for n=30 and much slower for n=40 (which is anyway the upper limit for the constraint-based approaches on pure matching problems). Its main advantage is its simplicity (it took a few simple lines of code to modify the upper bound in the original constraint algorithm) compared to both O.R. algorithms. This makes the constraint approach a good solution for small and mid-sized problems (even for values of k up to 10), while the solution based on the incremental Hungarian method is well adapted to large problems. 4.2. Bi-matching This second problem addresses the case of the combination of two matching problems (with two distinct weight matrices), where one problem is used for satisfiability and the other one for optimization (in a task-to-machine assignment, this amounts to maximizing production while keeping the energy consumption below a certain level). The difficulty varies according to the role played by the passive matching problem : if the value chosen for the maximal energy consumption is low or high, the energy matching is dominating the problem or is marginal. On the other hand, intermediate values make the problem much harder, because the shape of the solution space is equally affected by both matching problems. This implies that the distribution of feasible solutions in the search tree is no longer concentrated in one or a few areas, but is more dispersed. From a linear programming perspective, a bi- matching problem is made of three components: the matching constraints (shared by the two matching problems), the energy consumption constraint and the cost function that represent the optimization on production. A classical approach in such a situation is to use Lagrangean relaxation to push the energy constraint into the cost function. Here again, we implemented several solutions that find the maximal production for a matching which consumption is less than a given constant . The first one is also a straightforward adaptation of our basic constraint program. The search is guided by the optimization on the production. The consumption constraint is only used for pruning: Two lower bounds of the consumption are estimated by the sum of the edges of least consumption for each vertex set. Moreover, instead of considering the regret for the production function, we consider the sum of the regrets for both the production and the energy function. Therefore, this program is obtained by a simple replication of a few lines from the original basic program, where energy is substituted to production. . The second algorithm is a branch and bound algorithm using twice the (incremental) Hungarian method. Two Hungarian matchings are constructed: one of maximal production and one of minimal consumption. Branching is based on the Hungarian matching of maximal solutions. Both these matchings are used to prune: a branch can be cut off when the minimal consumption is strictly greater than E or when the maximal production is less than the best one found so far. . The third one and fourth are based on Lagrangean relaxation with Busaker's flow algorithm and the Hungarian method. The general principle of this method is to put some constraints in the objective function with a certain coefficient, and to vary the value of this coefficient [Re 93], assuming that the relaxed problem (i.e., without the constraint that was pushed into the cost function) is simpler to solve. Here, If we "push" the energy constraint into the cost function, we obtain a simple matching problem that we solve with the flow algorithm. The objective function becomes the total weight of the matching with the following weight function on the edges: weight l (u, Let us call f(l) the value of the maximum matching for weight l By construction, f(l) (for l - 0) is an upper bound for the original problem since the energy consumption has to be smaller than E. Now, if we compute the value F which is the minimal value for f(l) when l varies (l - 0), F is also an upper bound (that can be found through a dichotomic search). This provides us with an upper bound of high quality, and also, while varying the value of l, a possible lower bound (if one of the l yields an admissible energy). These bounds are computed at each node of the search tree, and branching is made on the edges of the matching of the optimal flow, as was indicated in the third approach in the previous section. The third program is a straightforward implementation of the Lagrangean relaxation with the flow algorithm, the fourth one is an implementation of the Lagrangean relaxation with the Hungarian method, that tries to be as incremental as possible (trying to keep part of the dual solution for similar values of l) (inc. Hung. Match.) 2.4kb 2.6kb 7.1 s. 3.2 kb 140kb 500 s. 900 s 14kb. Lagragean Relaxation with the flow algorithm 20b. 31b. 102b. 22b. 107b. 31b. 72 s. 197b. 87b. Lagragean Relaxation with the Hung. Match. 14b. 5.1 s. 64b. 80b. 28b. 86b. 50 s. 28b. 180b. 296b. Constraints (Basic) 1kb. 28kb. 1.2 s. 30kb. 1.3 s. 2kb. 900kb 43 s. Table bi-matching Table 5 compares all four algorithms on two problems where the cost and weight functions are randomly generated with values ranging from 1 to 100. There is a balance between optimization on the cost and feasibility on energy : they both can cut the search tree. If E is small (close to the value of a minimal energy consumption matching), the energy is responsible for all the pruning and the program behaves like a simple matching one. Symmetrically, if E is large, the problem is almost unconstrained and looks like a simple maximal production matching program. The hardest situation is when energy and cost are both responsible for pruning the search tree. This situation corresponds to E@ 550 and E @ 900 in our examples. Lagrangean relaxation pays off for large problems, when no criterion dominates the other. However, it seems that for real life problems a criterion most often dominates the other. From this point of view, the constraint approach is much more efficient. Here again, the constraint program performs better than the simple O.R. algorithm (with the two incremental Hungarian matchings). But it also performs better (except for the balanced situation for n=30) than the complex LR program. In fact, the Hungarian method is only mildly incremental for the Lagrangean relaxation scheme. The dual solution cannot be easily repaired after a change of weight function. 5. Application In the previous section we have shown the strenghts of the constraint-propagation approach compared with more specialized algorithms. We shall now see that these techniques should be combined and not opposed. We consider a time-tabling problem which consists of filling a weekly schedule with a set of lessons (of duration 1 to 4 hours). Each lesson is given a set of possible start times and a set of preferred start times. The schedule is made of 10 half-days of 4 hours and lessons should not be interrupted. This section illustrates how weighted matching techniques can be applied to such a problem and provide significant improvement. We notice that if all lessons were of duration 1, this problem would indeed be a matching problem. This suggests a straightforward relaxation where we consider lessons to be interruptible. Second, preferences can be encoded using weights so that the minimum weight matching corresponds to the assignment using as many preferred start times as possible. In the rest of the paper, we consider m lessons l 1 . l m , and a schedule with n time slots of length 1 hour 4). For each lesson, we are given two lists of time slots (integers between 1 and n) that represent respectively the set of possible and preferrred start times. The goal is to find a start time for each lesson so that . no two lessons overlap. This is the ususal disjunctive scheduling constraint: for all pairs of lessons l,l' either start(l') - start(l) . each lesson fits into a half day. So within a half day (says the i th half-day, covering the time units {4i+1, 4i+2, 4i+3, 4i+4}), lessons of duration 4 can only start at the first unit (4i+1), lessons of size 3 can only start at units 4i+1 and 4i+2 , etc. . the number of defaults, defined by the number of lessons which are assigned a start time outside the set of preferred times, is minimal This problem is very dependent on the range of input data. The problem can be seen as the combination of a satisfiablility problem (filling the schedule) and an optimization problem (an assignment problem). When the lessons are large, if there are enough of them, the problem may look like a bin-packing, where the preference optimization is almost irrelevant. On the other hand, with smaller lessons, the packing is much easier and the optimisation problem dominates. Similarly, the difficulty of the preferences (i.e., the tightness of the preference sets) will augment the importance of the optimization component of the problem. We have picked three problems that are representative of different situations: Problem 1 uses rather short lessons (such as in a high-school schedule) with complex preferences (the optimum is 6 defaults). Problem 2 is similar (from a lesson size point of view) but has simpler preferences (the optimum is 1). The last problem uses larger lessons to explore how the algorithm would react to a problem where satisfiability is an issue. In order to exploit the matching relaxation, we decompose each lesson of duration k into k units of 1 hour. The total number U of units is, therefore, the sum of the durations of the lessons. The matching that we need to build associates its start time (a time slot) to each unit. For a unit u associated to a lesson l, and a time slot i, we give a weight to each edge (u,i) as follows: cannot happen at i because it would imply an illegal start time for the lesson l, if u can happen at i but this implies a start time outside the preferred set, happens at i when the lesson l starts at one of its preferred time. We can now define MinWeightAllDifferent({start(u 1 ), ., start(u U )}, w) as the minimum weight matching for this graph. Recall that this value can be computed efficiently with the Hungarian matching, and that it can also be maintained incrementally throughout the search procedure (section 3. described how to recompute this value when a few edges have been removed in the graph). Since the lesson l has k units, it is straightforward to check that an optimal solution of our time-tabling problem will correspond to a minimum weight matching. For each schedule with D defaults, the value of the matching is 1000 * U - 12 * D. Thus, if we search for a solution with less than D defautlts, we can use the redundent constraint : To solve the three problems, we have used a branch-and-bound algorithm that minimizes the number of defaults. Branching is done on the starting time of a lesson, trying all perefered values first. The lesson on which to branch is picked using the first-fail heuristics. Each start time decision is propagated as follows: . We discard values from other lessons that could cause an overlap with the lesson that has been assigned. . The start time of the units composing the lessons are set accordingly. . When the number of default reaches the upper bound, all edges that are not preferred are removed. . We use a redundent constraint to detect when it becomes impossible to place large lessons. For each value d of duration that is strictly larger than a half-day (for example 3 hours), we compute the set of lessons lessons(d) with this duration and the union of all possible half days for these lessons possible(d). we then check that possible(d) - lessons(d). Finally, we check that possible({4,3,2}) - 2. Problem 1 (24 lessons 39h) Problem 2 (26 lessons 40h) Problem 3 (14 lessons 39h) 1. simple propagation 189 kb. 1541 s. 286 kb. 1040 s. 43 kb. 246 s. 2. matching cut 31 kb. 177 s. 269 kb. 1008 s. 27 kb. 160 s. 3. Regin's filter 31 kb. 197 s. 269 kb. 1243 s. 27 kb. 160 s. 4. weighted matching cut 3511 b. 29 s. 234 b. 2.6 s. 17 kb. 120 s. 5. global consistency 1206 b. shaving - 382 b. 96 s. 183 b. 11 s. 5888 b. 314 s. Table Applying various weighted matching techniques to a time-tabling problem Table 6 gives the result obtained with the following approaches : 1. The first method only performs constraint propagation. 2. The second method checks that the underlying matching problems (for the units) is still feasible (i.e., there exists a perfect matching). 3. The third method removes all edges that do not belong to at least one perfect matching using Regin's algorithm [R-94]. 4. The fourth method uses our new global constraint MinWeightAllDifferent (1) which is propagated as explained earlier using the incremental Hungarian method. 5. The fifth method uses a look-ahead consistency check trying for each un-assigned lesson l to discard all prefererred (respectively non-preferred) edges. - Trying - here means to detect if the removal of the values would create a contradiction through the propagation of the constraints. If a contradiction is detected, we deduce that the lesson l must start at a preferred (resp. non-preferred) time. 6. The sixth method applies a global consistency technique which consists of trying each possible start time and removing those who produce a contradiction. This is similar to the shaving technique that is used for jobshop scheduling [CL96]. A few conclusions can be drawn from these experiments. . First, our global constraint MinWeightAllDifferent brings a serious improvement over standard propagation (local consistency). For problem 2, the number of backtracks is divided by a factor of 1000. Moreover, the overhead for keeping the value of the minimal weight matching up to date with the incremental Hungarian method is very reasonable, compared to the the cost of simple matching propagation and bounding for defaults. Indeed, with the addition of the MinWeightAllDifferent the average time per backtrack over the three instances goes from 4,1 ms. to 7,3 ms. Therefore, the gain in search tree sizes turns into a real speedup. . Second, the propagation of the unweighted matching constraint (the difference constraint), which can be done very efficiently with the filtering algorithm proposed in [R- 94], has a marginal impact on the resolution of the problem compared to the propagation of the weighted matching constraint. . Third, it seems a good idea to check global consistency by performing a limited breadth-first exploration of the search tree (limited to depth one). This seems to work better for branching decisions based on satisfaction of lessons (method 5) rather than for actual assignements of time slots to lessons (method 6). This is due to the fact that the complexity of method 5 is smaller than that of method 6. Moreover, these global consistency techniques work best for time-tables with large blocks (lessons with duration 3 and 4), rather than on problems with small lessons, where the matching relaxation is more accurate. 6. Conclusion A first conclusion might be that constraint propagation is not a competitive technique for pure weighted matching problems. Specialized algorithms are simple enough and should be used appropriately. On the other hand, constraint propagation is a more plausible approach to augmented matching problems. Depending on the complexity of additional constraints and the size of the problem, the performances obtained (easily) with a CP approach range from reasonable to excellent, compared to more specialized approaches. In the few cases where this is not good enough, the best approach is to incorporate some of the weighted matching techniques into a constraint based branch and bound search. The first contribution of this paper is a set of generic techniques that improve the performance of constraint-based branch-and-bound algorithms for such problems, namely a binary branching scheme based on regret, the use of the implicit symmetry and the look-ahead bounding functions. These techniques may be applied to any augmented matching problem and were actually applied with success to the Traveling Salesman Problem [CL97]. Our second contribution is to show that indeed a weighted matching can be considered as a global constraint (MinWeightAlldifferent), since we know how to propagate it efficiently using the incremental version of the Hungarian algorithm. This is a very useful global constraint, as we have shown with our time-tabling example. A constraint solver that offers this feature will be able to solve problems with hard and soft constraints at the same time, taking advantage of the weights, whereas a system that relies on simple matching only will implement soft constraints as a second layer, in a much less efficient way. We must also notice that this feature allows us to combine the best of both worlds for the two augmented problems that we studied in Part 4: we keep the simplicity and flexibility of the constraint-based approach, while ensuring the robustness of the underlying matching algorithm. Acknowledgments We would like to thank an anonymous referee for his helpful comments and suggestions on an earlier version of this paper. --R A Deductive and Object-Oriented Approach to a Complex Scheduling Problem Improved CLP Scheduling with Tasks Intervals. Cumulative Scheduling with Task Intervals. Solving small TSPs with Constraints. Introduction to Algorithms. Matching. in Handbook in Operations Research and Management Science (Networks) Faster Scaling algorithms for network problems. Concurrence et coop-ration de mod-les multiples Consistency in networks of relations. Running efficiently arc consistency Combinatorial Optimization. A Filtering Algorithm for Constraints of Difference in CSPs Modern Heuristic techniques for combinatorial problems. Constraint satisfaction in Logic Programming. Graph Algorithms. --TR
weighted matching;resource allocation;global constraint;hungarian method
593353
A Constraint-Based Interactive Train Rescheduling Tool.
In this paper, we report the design and implementation of a constraint-based interactive train rescheduling tool, a project in collaboration with the International Institute for Software Technology, United Nations University (UNU/IIST), Macau. We formulate train rescheduling as constraint satisfaction and describe a constraint propagation approach for tackling the problem. Algorithms for timetable verification and train rescheduling are designed under a coherent framework. Formal correctness properties of the rescheduling algorithm are established. We define two optimality criteria for rescheduling that correspond to minimizing the number of station visits affected and passenger delay respectively. Two heuristics are then proposed to speed up and direct the search towards optimal solutions. The feasibility of our proposed algorithms and heuristics are confirmed with experimentation using real-life data.
Introduction The PRaCoSy (People's Republic of China Railway Computing System) project [10] is undertaken by the International Institute for Software Technol- ogy, United Nations University (UNU/IIST). The aim of the project is to develop skills in software engineering for automation in the Chinese Railways. A specific goal of the project is the automation of the preparation and updating of the running map 1 , for dispatching trains along the 600 kilometer long railway line between Zhengzhou and Wuhan in the People's Republic of China. The Zhengzhou to Wuhan section has been chosen as a case study because it is along the busy Beijing-Guangzhou line, the arterial north-south railway in China. The rate of running trains, both goods and passengers, of this section is high and present management procedures are not adequate with the dramatic development of domestic economy. A running map [9] contains information regarding the topology of the railway, train number and classification, arrival and departure time of trains at each sta- tion, arrival and departure paths, etc. A computerized running map tool should read in stations and lines definition from a descriptor file, allow segments (sub- sets of all stations) and time intervals to be defined, allow train timetable to running map is a method of monitoring the movement of trains and rescheduling their arrivals and departures to satisfy operational constraints. be read, and finally display graphically the projection of the timetable against a given segment and a given interval. A sample running map is shown in figure 6. Train dispatchers, users of the tool, have to modify the timetable when trains in some sections cannot run according to the map, possibly due to accidents and/or train delays. The modification to the map should be performed in such a way that certain scheduling rules (laid down by the local railway bureau) are not violated. Therefore, a computer running map tool should check users' modifications against possible violation of scheduling rules, and warn users of such violations. In addition, the tool should also assist the user in repairing, either automatically or semi-automatically, an infeasible timetable so that the least train service disruption is made. We call this process rescheduling . Scheduling and rescheduling are different in two aspects. First, while scheduling creates a timetable from scratch, rescheduling assumes a feasible timetable and user modifications, which may introduce inconsistencies to the timetable, as input. Second, optimality criteria used in scheduling, such as minimum operating cost, are usually defined in the absolute sense. In rescheduling, however, the quality of the output is measured with respect to the original timetable. The PRaCoSy project has resulted in a running map tool capable of train timetable verification [7]. Our task at hand is to enhance the PRaCoSy tool to perform automatic rescheduling, which can be considered as constraint re- satisfaction. A major problem with the PRaCoSy implementation is that constraints are used only passively to test possible violation of scheduling rules. In view of this limitation, we decided to re-create the running map tool from scratch using a constraint programming approach. In this paper, we give algorithms for timetable verification and train rescheduling used in our tool, and show that constraint programming allows us to perform constraint checking and solving (or propagation) in a coherent framework. We study two notions of optimality for the rescheduled timetable with respect to the original timetable. These notions provide a measure of the quality of the rescheduling operation. We also present two heuristics that direct and speed up the search towards optimal rescheduled timetables. The rest of the paper is organized as follows. Section 2 defines basic terminology and discusses related work. Section 3 explains the timetable verification algorithm. In section 4, we show how to formulate train rescheduling as a constraint satisfaction problem and give an associated algorithm. We also discuss two heuristics that help to direct and speed up the search towards optimal solu- tions. In section 5, we describe our prototype implementation and sample runs of the tool. We summarize our contribution and shed light on future work in section 6. Preliminaries In the following, we provide informal definitions of necessary terminology according to [9] to facilitate subsequent discussions. Names not defined should be self-explanatory or clear from the context. Interested readers can refer to [9] for formal definitions of the same terminology. The topology of a railway system is defined by a collection of named stations and identified lines. We differentiate between the lines within a station and those connecting two stations by referring the former as lines and the latter as tracks. A train journey is a sequence of visits to connected stations. Each visit to a station is represented by the arrival time and the departure time of the train. A timetable is an association from trains to journeys to be made. Scheduling rules is a set of temporal constraints to restrict the arrival time and departure time of each visit in order to prevent such undesirable events as train crash. A timetable is valid (or feasible) if no scheduling rules are violated under the associations. Otherwise, the timetable is invalid (or infeasible). Given a feasible timetable with n station visits, which is represented by a set of assignments (or equality constraints) of the form AT where AT i and DT i denote the arrival and departure time respectively, for 1 - modifications that we can make to the timetable are to replace some assignments (or t d ng. In other words, we can only delay the arrival time or departure time of visits. Given an infeasible timetable, rescheduling is the process of modifying the timetable so as to make the timetable feasible. 2.1 Problem Statement There are six types of scheduling rules [9] in our railway system: the speed rule, the station occupancy rule, the station entry rule, the station exit rule, the line time rule, and the stopover rule. Let there be two trains 1 and 2 and two adjacent stations A and B. The variables ATXY and DTXY denote train Y 's arrival and departure time at/from station X respectively. The above scheduling rules can be formulated as the following types of scheduling constraints. The Speed Constraint The constant lng denotes the distance between station A and station B. This constraint enforces that the average train speed traveling between the two stations cannot exceed sp. The Station Occupancy Constraint This constraint enforces that there is at least ctr time units between two trains occupying a track. The Station Entry Constraint This constraint enforces that there is at least cen time units between two trains entering a station via a line. The Station Exit Constraint This constraint enforces that there is at least cex time units between two trains departing from a station via a line. The Line Time Constraints The line time rule is split into two constraints. The first constraint enforces that no train overtakes another train if they are traveling in the same direction on a line. The second constraint enforces that if there are two journeys on a line in opposing directions, the line must be unoccupied for at least cenx time units. The Stopover Constraint cst This constraint enforces that a train will stay in a station for at least cst time units. Given the topology of a railway system with a valid train timetable, due to unexpected events, the users of the running map tool may want to modify the timetable. Our work is to first check the feasibility of the modified timetable. If it is feasible, the previous timetable is replaced by the modified one. Other- wise, we reschedule the infeasible timetable to generate a new feasible timetable. Note that efficiency should be a critical concern in designing the verification and rescheduling algorithms since in real-life situations, rescheduling must be performed in a timely manner. The notion of "efficiency" may vary according to situations. Ten minutes, however, should be a tolerable bound in general [2]. Optimal solutions are not required usually. In most cases, it is impractical to generate optimal solutions within a given (usually small) time bound. Criteria for optimality, however, should be defined. Such definitions can serve as guidelines for designing various variable-ordering and value-ordering heuristics to generate "good" answers. A precise notion of optimality also enables us to measure the "quality" of the rescheduled timetable. In the following, we present two optimality criteria. A rescheduled timetable is minimum-changes optimal with respect to the original timetable if the least number of station visits are modified. This criterion can be satisfied easily in general since in most cases, we can simply delay the trains in question to the latest possible time. The resulting timetable, however, may introduce unreasonable long delay to some train visits. Thus this criterion should usually be applied with other criteria limiting the maximum delay. A rescheduled timetable is minimum-delay optimal with respect to the original timetable if the longest delay among all train visits is minimum. Let the denote the infeasible and the rescheduled timetable respectively. The goal of this criterion is to minimize the following expression: The aims of the two criteria could contradict one another and represent the extremes of a spectrum of other possible definitions of optimality. 2.2 Related Work Rescheduling is different from traditional scheduling in the sense that the possible solutions of rescheduling are restricted by the original schedule. Zweben et al [11] tackle this problem using constraint-based iterative repair with heuristics. The resultant GERRY scheduling and rescheduling system is applied to coordinate Space Shuttle Ground Processing. Our work is based on a propagation- based constraint solver. Somewhat related to our work is train scheduling. Komaya and Fukuda [5] propose a problem solving architecture for knowledge-based integration of simulation and scheduling. Two train scheduling systems are designed in this ar- chitecture. Fukumori et al [3] use the tree search and constraint propagation technique with the concepts of time belt in their scheduling system. This approach is claimed to be suitable for double-track line and continuous time unit. Recently, Chiang and Hau [1] attempt to combine repair heuristic with several search methods to tackle scheduling problems for general railway systems. There are two on-going projects that aim at automating train scheduling for real-life railway ministries. Our work is a direct outgrowth of the PRaCoSy project [10] at UNU/IIST. The latest PRaCoSy running map tool prototype uses constraints only passively to test for constraint violation in their verification engine. The Train Scheduling System (TSS) designed for Taiwan Railway Bureau (TRB) [6] is a knowledge-based interactive train scheduling system incorporating both an automatic and a manual schedulers. Users and the computer system are thus able to bring complementary skills to the scheduling tasks. 3 Timetable Verification In the following, we describe a timetable verification algorithm, which examines if a given timetable is valid with respect to a set of scheduling constraints. Violated scheduling rules (or constraints) in an invalid timetable will be located and displayed to the user. The algorithm, shown in figure 1, assumes the existence of a propagation-based constraint solver [8] propagate(). The timetable T can be viewed as a set of constraints for all variables in either the form AT i or DT is an arrival time and DT i is a departure time. After all variables become ground (line 5), we post the scheduling constraints one by one (lines 7-13). Having all variables ground, the propagate() engine performs essentially constraint checking. It is easy to check violated constraints */ 3 /* Initialization */ 4 I / fg Constraint store S is initialized to T */ 6 /* Constraint Verification */ 7 for each c 2 C 9 if inconsistency found I / I [ fcg Retract c from the constraint network */ endfor 14 end Fig. 1. The Timetable Verification Algorithm that a constraint is violated under the given timetable if inconsistency is found after the constraint is told to the store. Violated constraints are retracted so that the algorithm can proceed to check for other possible constraint violations. Again, the groundness of all variables allows us to retract a constraint by simply removing the undesirable constraint from the constraint store. 4 Rescheduling as Constraint Satisfaction Scheduling is an instance of constraint satisfaction problem. In the following, we show the same for rescheduling. Given a timetable T . Users modify T by adjusting its arrival and departure times, obtaining T 0 , which can be valid or invalid. If T 0 is invalid, the rescheduling process should attempts to repair T 0 to make it feasible. By repairing, we mean adjusting the value of the non-modified variables so that (1) the timetable becomes valid again and (2) the new timetable should be reasonably "close" to the original timetable T . By being close to T , we mean that the new timetable should create the least service disruptions. Example optimality criteria are given in section 2. Note that user modified variables must be kept fixed during the rescheduling process since the modifications represent dispatcher requirements. In order to formulate rescheduling as constraint satisfaction, we have to determine the variables of the problem, the domains associated with the variables, and the constraints of the problem. In rescheduling, the variables are the arrival and departure time of the timetable. Every variable share the integer domain . There are three types of constraints in the rescheduling problem. 1. Scheduling constraints: The scheduling constraints set forth in section 2. 2 There are 1440 minutes in 24 hours. 2. Modification constraints: For each arrival or departure time X which is modified by the user to new value t, we have the equality constraint t. This constraint enforces the user modifications to stay fixed during rescheduling. 3. Forward-labeling constraints: For each non-modified variable X with value t in the original timetable T , we have the constraint X - t. This constraint is necessary to ensure that we can only delay arrival or departure time. Rescheduling now becomes finding a solution to the above constraint satisfaction problem. A solution is optimal if the solution is "closest" to the original timetable. We are now ready to present the rescheduling algorithm, shown in figures 2 and 3. The algorithm can be divided into three phases. In phase one (line 4), we timetable */ 3 /* Initialization */ Save a copy of the constraint store after 5 propagating constraints in C */ 6 /* Rescheduling */ 7 while true 8 Read in user modifications U /* U is in the same form as T */ 9 modify(T , S0 , U , R) timetable */ 11 Prompt error messages else rescheduled timetable if necessary */ 14 endif Fig. 2. The Train Rescheduling Algorithm post and propagate all scheduling constraints to prune infeasible values in the variables. The pruned constraint network is then saved in S 0 . Since the scheduling constraints are the same for any timetable T and user modifications U , the same store S 0 will be reused in every rescheduling step. In real-life situation, rescheduling has to be performed repeatedly for different timetables and user modifications, this saving operation helps to avoid unnecessary invocations of constraint relaxation. Actual rescheduling takes place in the procedure modify() (lines 16-46). In the second phase (lines 19-35) of rescheduling, information is extracted from user modifications and the original timetable to post and propagate the modification constraints (lines 23-27) and the forward-labeling constraints (lines 28-35). If inconsistency is found, rescheduling is halted and failure is reported. In the third phase (lines 36-46), variables that are not modified by the users (extracted by the vars() function) are enumerated or labeled using some form of variable- procedure modify(in T , S0 , U , out R) previous feasible timetable, S0 : saved constraint network state, user modifications, R: rescheduled timetable */ 19 /* Initialization */ O / Assignment constraints associating with user-modified variables */ 22 /* Post all user modifications */ First and second type of constraints */ inconsistency found /* User modifications are inconsistent */ 26 return 28 /* Prune infeasible values: can only delay arrival & departure time */ 29 for constraint Third type of constraints */ inconsistency found return 34 endif 36 /* Rescheduling */ of variables for rescheduling */ Appropriate variable- and value-ordering should be used. 40 The function labeling() returns either fg or a set of equality constraints (or bindings) for each variable X 43 R / fail 44 else Fig. 3. The Train Rescheduling Algorithm (cont.) and value-ordering heuristics, which are embedded in the labeling() function, to speed up and direct the search towards a near-optimal solution. There are two situations in which user modifications lead to a timetable T 0 that is non-repairable. First, the user modifications are self conflicting. Since user modified variables must be kept fixed during rescheduling, it is impossible to repair other variables to make the timetable valid. Second, user modifications are not self conflicting but there is no room for other variables to adjust to make the timetable valid. Constraint propagation algorithms are well-known to be incomplete [8]. Thus, phase two of the rescheduling algorithm can detect some, but not all, of this kind of conflicts. Theoretically speaking, the enumerating procedure in phase three can guarantee to detect inconsistency but it would usually take impractically long to do so. In cases when the rescheduling algorithm fails to return an answer within a few minutes, users are advised to abort the current computation, re-adjust the modifications, and restart the rescheduling process. In the rest of this section, we present two variable labeling heuristics, which are designed to yield rescheduled timetables in the minimum-delay and the minimum-change optimal sense respectively. Smallest-First Principle Variables are ordered in the ascending order of the lower bound of their domains. Values in the variable domains are also enumerated in ascending order. This principle is founded on the assumption that a short delay on a train visit will cause short delay on the subsequent one, which means that delays propagate in a monotonic fashion. Experimental results confirm that, using actual timetables from PRaCoSy, this heuristic usually helps to generate solutions that are minimum-delay optimal efficiently. We construct below an unrealistic artificial example that defeats the heuristic. Figure 4 (a) shows a small segment of four journeys on a railway running (a) Before Rescheduling (b) After Rescheduling Fig. 4. A Non-Optimal Solution by Smallest-First Heuristic map. The three journeys A, B, and C share the same track in the Nanjing station, while the two journeys C and D take the same line in traveling from Nanjing to Longtan. The journeys A and D are fixed (indicated by thick lines) by the users. Suppose the station occupancy and the station exit rules enforces that at least ten minutes among each of the three points A,B, and C, and sixty minutes between the points C and D respectively. The modified timetable is thus infeasible due to the insufficient long distance between the points A and B. Figure 4 (b) shows the rescheduled timetable obtained using the smallest- first principle. The rescheduling starts from moving point B ahead in time to achieve the ten-minute requirement between points A and B. The movement in turn causes another conflict between points B and C. Point C is thus forced to move. However, there is no feasible location for point C to move between points B and D since point D is fixed by user. Therefore, we have to move point C one-hour ahead of point D. The maximum delay in this case is two hours. This solution is not optimal since a better solution can be obtained by simply moving point B twenty minutes ahead in time, as shown in figure 5. Fig. 5. A Smallest-Changes-Optimal Solution Consistent-Assignment-First Principle This heuristic suggests to instantiate first those variables that can be instantiated with its time value in the original timetable. We call these variables non-conflicting . The other variables are conflicting . Therefore, the labeling of the non-conflicting variables will be backtracked into last. Again values in the variable domains are enumerated in ascending order. The idea is to maintain as many variables with its original value as possible. This heuristic direct searching towards a minimum-change optimal solution. Note that a non-conflicting variable may be instantiated with a value other than its original value eventually if the conflicting variables have tried all possible combination of time values and no solution is found. For efficiency reason, we further classify the non-conflicting variables into two group. The first group contains variables that share journeys with one of conflicting variables. The second group contain the rest of the non-conflicting variables. Our heuristic suggest to label first the second group of non-conflicting variables. This ordering is essential since the labeling of variables sharing journeys with conflicting variables have a higher chance of being backtracked into. We apply this heuristic to reschedule the infeasible timetable in figure 4 (a). Recall that the points A and D are fixed by users. We classify the variables associated with point B and point C as conflicting and non-conflicting respec- tively. Thus point C is labeled first to retain its original position in the map and point B is forced to move until it reaches the location ten-minute ahead point C. In this specific case, the minimum-change optimal solution coincides with the minimum-delay optimal solution. This example also shows that, in general, the two heuristics give different first solution. 5 Prototype Implementation In order to demonstrate the feasibility of our algorithms, we have re-constructed and enhanced the PRaCoSy running map tool prototype [7] with rescheduling capability. The prototype consists of a constraint-based scheduler and a user- interface. The former is implemented in C++ with ILOG Solver library 2.0 [4] while the latter is built using Microsoft Visual Basic 3.0. In the following, an overview of the running map tool is presented. We then give a sample session of our tool using a segment (from Nanjingxi to Shanghai) of the China railway which amounts to 3005 constraints and 596 variables. The rescheduled timetables generated by the two heuristics are explained. We conclude this section by showing two examples to which our tool fails to respond in a timely manner. The running map display (figure consists of six columns (regions). Column one and column four show the abbreviated identifiers of stations. Column two shows the number of arrival and departure lines of stations. Column three is the main window for presenting the graphical representation of a timetable, with time and locations as the X and Y axes respectively. Column five shows the cumulative distance from the first station. Column six shows the distance between the current station and the previous station. In the cases where a timetable is too large to fit into the main window, two scrollbars will be enabled. Fig. 6. The Running Map Tool A user can click-and-drag any lines to modify the corresponding train visits on the map. The modified timetable will be validated using the verification algorithm when the "Check" button is pressed. If it is infeasible, a warning window, such as that shown in figure 7, will pop up to display all constraint violations. At this point, the user can either invoke the rescheduling algorithm by pressing the "Reschedule" button, correct the modifications manually, or restore the original feasible timetable. Figure 8 (a) shows a segment of a China railway timetable. Due to an acci- Fig. 7. A Warning Window (a) Original Feasible Timetable (b) Modified Infeasible Timetable Fig. 8. A Comparison of Two Heuristic dent, we have to delay the train departuring from Zhenjiang at 1:05 to 3:30, yielding the map shown in figure 8 (b). The user-modified departure time (pointed by an arrow) and all the subsequent train visits on the journey (the highlighted segment) are then fixed immediately by the running map tool. This user movement incurs seven constraint violations between the modified journey and its left adjacent journey. We reschedule the infeasible timetable with our two heuristics. Both of them succeed in generating a feasible timetable within a few seconds. Figure 9 (a) and figure 9 (b) show the rescheduled timetable generated using the smallest-first principle and the consistent-assignment-first principle respectively. (a) Smallest-First Principle (b) Consistent-Assignment-First Principle Fig. 9. A Comparison of Two Heuristic (cont.) Applying the smallest-first principle, we process all visits on the map from the left (earliest) to the right (latest). For each visit, if its associated arrival (or departure) time does not violate any scheduling constraints, we preserve the current value. Otherwise, we move the time ahead as little as possible to eliminate the inconsistencies. Thus some visits on a journey may be modified while others remain unchanged. This explains why, visually, a journey is not only shifted right horizontally, but can also be "bent" by the rescheduling process. The movement propagates in the above fashion from left to right. Whenever no further movements are possible, backtracking takes place. Instead of massaging several journeys to produce a feasible timetable, the consistent-assignment-first principle suggests to modify as few station visits as possible. This goal can be well approximated by first locating station visits that can remain unchanged with respect to the user modifications. These station visits will labeled first. The conflicting variables, having to change their original values, will be labeled last. Our experiments reveal that, in many cases, this heuristic produces timetable which is "almost identical" to the original timetable. As seen in figure 9 (b), most of the journeys retain their original locations. Even for the right-shifted journey, its shape is mostly preserved. Experimental results confirm that rescheduling can usually be completed within seconds. This is not always the case. Figure 10 (a) and figure 10 (b) provide two such examples. The infeasible timetable in figure 10 (a) can be rescheduled using the consistent-assignment-first principle in a few seconds, but the smallest-first principle fails to return an answer within five minutes. Figure (b) is simply a non-repairable timetable. Neither heuristics can return promptly to confirm the unsolvability of the problem. (a) Infeasible Timetable 1 (b) Infeasible Timetable 2 Fig. 10. Poor Performance Example 6 Concluding Remarks The contribution of this paper is three-fold. First, we define formally train rescheduling as a constraint satisfaction problem. Two algorithms for railway timetable verifications and rescheduling are then derived based on a propagation- based constraint solver. We define two optimality criteria, which are used to measure the "quality" of the rescheduled timetable. It is important to note that the optimality criteria are defined with respect to the original timetable. Sec- ond, based on the domain knowledge learned from domain analysis, we propose two heuristics to speed up and direct the search towards minimum-delay optimal and minimum-change optimal solutions respectively. The feasibility of our proposed algorithms and heuristics are confirmed with experimentation using real-life data. Third, we have re-constructed and enhanced the PRaCoSy running map tool prototype. It would be an interesting future work to study different stochastic methods for train rescheduling. Work is also in progress to experiment our rescheduling method on larger-scale real-life railway timetables. Acknowledgement We acknowledge, with pleasure, the interaction we have had with Fellows and Staff of UNU/IIST, the United Nations University, International Institute for Software Technology, Macau. In particular, we are indebted to Prof. Dines Bj-rner for inviting our participation in the PRaCoSy project. We also had numerous fruitful discussion and working sessions with S-ren Prehn, Chris George, Yulin Dong, Liansuo Liu, and Dong Yang. Last but not least, we thank the anonymous referees for their constructive comments, which help to improve the final version of the paper. --R Railway scheduling system using repair-based ap- proach The Zhengzhou Fundamental algorithm for train scheduling based on artificial intelligence. Solver Reference Manual Version 2.0 A knowledge-based approach for railway scheduling An interactive train scheduling workbench based on artificial intelligence. A simple running map display tool. Consistency in networks of relations. A railway running map design. An executive overview. Scheduling and rescheduling with iterative repair. --TR
search heuristics;constraint propagation;train rescheduling
593354
Fourier Elimination for Compiling Constraint Hierarchies.
Linear equality and inequality constraints arise naturally in specifying many aspects of user interfaces, such as requiring that one window be to the left of another, requiring that a pane occupy the leftmost 1/3 of a window, or preferring that an object be contained within a rectangle if possible. For interactive use, we need to solve similar constraint satisfaction problems repeatedly for each screen refresh, with each successive problem differing from the previous one only in the position of an input device and the previous state of the system. We present an algorithm for solving such systems of constraints using projection. The solution is compiled into very efficient, constraint-free code, which is parameterized by the new inputs. Producing straight-line, constraint-free code of this sort is important in a number of applications: for example, to provide predictable performance in real-time systems, to allow companies to ship products without including a runtime constraint solver, to compile Java applets that can be downloaded and run remotely (again without having to include a runtime solver), or for applications where runtime efficiency is particularly important. Even for less time-critical user interface applications, the smooth performance of the resulting code is more pleasing than that of code produced using other current techniques.
Introduction Constraints are a natural tool for user interface toolkits and other kinds of interactive graphical systems. Some important uses in this application area include specifying layout and other geometric information, maintaining consistency between application data and a view of that data, and maintaining consistency among multiple views. It is important to be able to express preferences as well as requirements in interactive constraint systems. One important use is to express a desire for stability when moving parts of an image: things should stay where they were unless there is some reason for them to move. A second use is to process potentially invalid user inputs in a graceful way. For example, if the user tries to move a gure outside of its bounding window, it is reasonable for the gure just to bump up against the side of the window and stop, rather than signalling an error. A third use is to balance con icting desires, for example in laying out a graph. The constraints needed to specify and maintain layout information are typically linear equalities and inequalities over the real numbers. 2 Inequality constraints in particular are needed to express relationships such as \inside," \above," \left-of," and \overlaps." For example, we can express the requirement that window1 be to the left of window2 as the constraint window1.rightSide window2.leftSide. Some of these layout constraints will be requirements, and others preferences. For interactive systems, a typical requirement is to re-satisfy the constraints repeatedly as the user moves some part of a gure|each time the screen is refreshed the constraints must be re-satised. Each of these constraint satisfaction problems diers from the previous ones only in the values of some of the constants in the constraints (for example, the mouse position). One strategy for achieving the required interactive response times is to compile a constraint satisfaction plan: a block of code that can be executed repeatedly to re-solve the constraints with different input parameters. (We can view this as a kind of partial evaluation of the constraint solving algorithm.) This has long been done for local propagation solvers (e.g. [1]), and more recently for simultaneous linear equations [4] and for acyclic sets of inequality constraints [2]. However, there have not been any systems that can compile plans for systems of constraints including a cyclic set of simultaneous equalities and inequalities. That lack is addressed here. The original motivation for this work was constraint solving for user interface toolkits and other kinds of graphical interactive systems. The technique is useful for two reasons: rst, it produces code that requires no runtime support for constraint solving; second, the code is very e-cient. The technique should also be useful in other applications where either of these considerations is a factor. The ability to get rid of runtime support is important in many real software tasks. One example of this sort is a constraint-based authoring environment for producing Java applets, where the behaviour of the applet is partially specied using constraints. After the applet is written and tested in the authoring environment, the constraint compiler is used to produce straight Java code that can be shipped over the net and run on a remote machine, without requiring a runtime constraint solver on the remote machine. Some preliminary work on such an application has been done [6], and we have used the constraint compiler to produce Java code for several applets. These include an interactive demonstration of a geometric theorem, and an abacus simulation, which were then included in web documents. As a second example, when building an embedded real-time engine controller, predictable performance is needed: compiled code can provide that, but calls to runtime constraint solvers generally cannot. Finally, in developing a product, a company might use constraints in developing the application and not want to ship a proprietary constraint solving package with that application [8]. The compiled code is also very e-cient|our measurements show speeds from 5 to 20 times faster than the same test cases performed using a runtime solver based on the simplex algorithm [7]. The runtime solver is reasonably e-cient as well, and in many applications the constraint solving time is dominated by the graphics refresh time for either technique. However, in some cases the simplex solver has markedly varying response times|much of the time it is extremely fast, but when a succession of pivots are required it slows down considerably, giving rise overall to a more jerky quality to the interaction. (People prefer uniform response times to varying ones.) The additional speed could be important in other cases as well, for example when the compiled code is to be run on a slower processor. In this paper we show how we can use Fourier elimination to compile heirarchical constraint solving. 3 In Section 2, we introduce the notion of solving constraints by projection, and show how it can be used to produce compiled code. We review the notion of constraint hierarchies for expressing preferential constraints in Section 3, before showing how to adapt the projection algorithm to handle hierarchical constraints in Section 4. In Section 5, we address the issue of redundancy management. In Section 6, we discuss some extensions to the algorithm. We examine some of the compile-time and run-time properties of our prototype implementation in Section 7, while in Section 8 we present some theoretical complexity results for the algorithm. Finally, in Section 9 we conclude. 2. Solving constraints using projection In this section we brie y illustrate how projection can be used to nd solutions to linear equality and inequality constraints. For the moment we will ignore preferential constraints, and also ignore issues of compilation. A primitive constraint in this context is a linear inequality or equality. A constraint is a set of primitive constraints. We assume every primitive constraint is written in a simplied form so that no variable appears twice in the same primitive constraint. Let vars(C) denote the set of variables appearing in constraint C, and similarly let vars(c) denote the set of variables appearing in equation or inequality c. We denote by 9W F the existential quantication of all variables in F except W , that is the formula 9v 1 9v 2 9v k F , where W is a set of variables and For each variable x 2 vars(C) we can partition C in the following way. Let C 0 x be the set of primitive constraints c 2 C where x 62 vars(c). Let x be the set of equations c 2 C where x 2 vars(c). Let C x be the set of inequalities c 2 C such that c is equivalent to an inequality of the form x e, where e is a linear expression not involving x. Finally, let C x be the set of inequalities c 2 C such that c is equivalent to an inequality of the form e x, again where e is a linear expression not involving x. The projection algorithm in Figure 1 eliminates a variable x from constraint C and returns constraint D $ 9xC using either Gaussian elimination or Fourier elimination. To solve constraints we project out each variable in turn using project until no variables remain. The resulting constraint C 0 is equivalent to 9 ; C. None of the primitive constraints in C 0 involve any variables and hence we can straightforwardly determine whether C 0 is satisable or not. This answers the satisability question if exists c 2 x where c with every occurrence of x replaced by e else for each c 2 C for each c 2 C x where c e x end for each end for each return D; Figure 1. Combined Fourier-Gaussian projection x l0 100 Figure 2. A simple constrained picture for C. We can then use the intermediate constraints C i produced after i 1 applications of project, as well as the variables x in their order of elimination, to build a solution to C. For 1 i n it is the case that vars(C i g. So Cn only contains variable xn and hence only contains constraints of the form satisable we can nd a valuation for xn which is a solution of Cn . We can extend a solution of C i+1 to a solution of C i as follows. Consider each of the constraints in C i that involve x i . If one is an equation equivalent to x e, where e is an expression only involving variables then we can let d contains inequalities containing x i . These can be written in the form x i e j or e k x i , where e j or e k is an expression only involving variables fx g. Using valuation we can determine the minimum u of the e j and maximum l of the e k . Any value for d i between l and u gives a valid solution of C i . The new solution is thus g. Figure 3. Constraint projection example We now give an example of constraint solving using projection, using the illustration in Figure 2. The constraints are as follows: xm is constrained to be the midpoint of the line from x l to x r , x l is constrained to be at least 10 to the left of x r , and all variables must lie in the range 0 to 100. We can represent this using the constraint C 1 shown in Figure 3. To solve this constraint using variable elimination, we start with C 1 , select a variable and project it out, and continue until no variables remain. Suppose we rst select x l . We can project it out using the equation yielding the constraint C 2 , and then project out x r , yielding the constraint C 3 (where we have eliminated some simple redundancy). Finally, eliminating xm we obtain C 4 , which is clearly satisable. We now show how to construct a solution from these constraints. By inspecting know we can pick any value between 5 and 95 for xm , say 50. Next we examine the constraints in C 2 involving x r . These are fxm 100g. Given 100g. We can thus pick any value for x r in the range [55::100], say 70. Finally, examining the constraints in C 1 , there is an equation involving x l , namely x so we can use this equation directly to set x l to 30. We now use this information to compile a sequence of statements that constructs a solution to the constraints, and returns it as a triple, illustrated in Figure 4. As it stands this sequence of statements isn't very interesting, since it only solves one problem. However, in Section 4 we show how to use a similar technique to compile code parameterized by appropriate inputs. 3. Constraint hierarchies As discussed earlier, for interactive graphics applications, it is important to be able to express preferences as well as requirements in the constraint system; in particular a desire for minimizing change as a gure is manipulated. We use constraint hierarchies [5] for specifying the desired solutions to a collection of required and preferential constraints independent of the particular algorithm involved. x l x u choose xm 2 [x l x l x u r := min f 2xm, 100 choose x r 2 [x l r ::x u r return Figure 4. Code to solve the constraints of Figure 2, obtained by Fourier elimination A labelled primitive constraint is a primitive constraint labelled with a strength, written sc, where s is a strength and c is a primitive constraint. Strengths are non-negative integers. For clarity, we give symbolic names to the dierent strengths of primitive constraints. Strength 0, with the symbolic name required, is always reserved for required constraints. For the purposes of this paper we shall use the following symbolic names for strengths: strong = 1, but in general the approach works for any number of strength levels. Note that a higher integer \strength" indicates a weaker constraint. A constraint hierarchy is a multiset of labelled primitive constraints. Given a constraint denotes the multiset of required primitive constraints in H, with their labels removed. In the same way, we dene the multisets H for strengths Since we are now dealing with non-required constraints, valuations of interest may not satisfy all constraints. Hence we need to be able to measure the degree to which a constraint is satised. We assume that for each primitive constraint c we have an error function e(c) that returns a non-negative real number indicating how nearly the primitive constraint c is satised for a valuation . In the work described here, the domain is the reals, and the error function returns a value that varies depending on how nearly the constraint is satised. For example, the error while the error for x y is max(0; x y). The set S of solutions to the hierarchy is dened as follows. S 0 is the set of valuations such that all the H 0 constraints hold; from this we form S by eliminating all potential valuations that are worse than some other potential valuation using the comparator predicate better. For now we assume the use of local comparators for better, which compare two solutions primitive constraint by primitive constraint. (In Section 6, we will look at how we can extend our approach to certain kinds of global comparators.) A valuation is locally-better than another valuation if, for each of the primitive (a) (b) (c) (d) Figure 5. Demonstrating a theorem about quadrilaterals constraints through to some strength level k 1, the error after applying is equal to that after applying , and at level k the error is strictly less for at least one primitive constraint and less than or equal for all the rest. locally-better A locally-better comparator with such an error function is known as locally-error- better. A traditional demonstration of constraint-based graphics is the Quadrilateral Theorem illustrated in Figure 5. The screen snapshots are taken from our Smalltalk implementation, which uses the code produced by the projection algorithm. Each side of the quadrilateral is bisected, and lines are drawn between the midpoints (these inner lines always form a parallelogram). This is expressed as a constraint on each midpoint that it lie halfway between the endpoints of its line. In addition, all points are constrained to be at least 10 pixels from the sides of the window, the north vertex is constrained to be at least 30 pixels above the south vertex, and the east vertex is constrained to be at least 30 pixels to the right of the west vertex (the latter two so that the quadrilateral cannot collapse to a point). Taken together, these constraints are too di-cult for most UI constraint solvers, since they involve cyclic equality and inequality constraints. In Figure 5(a) we have picked up one of the midpoints with the mouse and begun to move it by temporarily adding an edit constraint equating the position of the midpoint and the mouse. This constraint is strongly preferred but not required|we will violate it if necessary. The corner points are weakly constrained to stay where they are by stay constraints. The constraint heirarchy representing the quadilateral problem is shown in Figure 6, where (n x are the coordinates for the north, east, south and west vertices, respectively. The coordinates of the midpoints are (ne x ; ne y ), (se x ; se y ), (sw x ; sw y ) and (nw x ; nw y ). (Mouse x ; Mouse y ) are the current coordinates of the mouse, and old(n x ), old(n y ), required 2ne required 2ne required 2se required 2se required 2sw required 2sw required 2nw required 2nw required n y s y required e x w x required required strong se strong se Figure 6. Constraint heirarchy representing the quadrilateral problem etc. give the old coordinates of the points (where we want them to stay). The strong edit constraints ensure that the south east midpoint tries to follow the mouse, while the weak stay constraints try to maintain the corner points in their old positions. In Figure 5(b) the mouse has been moved to the right, and to keep the midpoint constraint satised the east vertex moved as well. We continue moving to the right. In Figure 5(c) the east vertex has run into the imaginary wall resulting from the constraint that it be at least 10 pixels from the window boundary, and can move no further. As a result, in order to maintain the midpoint constraint, the south vertex moving instead. Finally, in Figure 5(d) the mouse has moved beyond the permitted region for the midpoint. The midpoint has moved as close to the mouse as possible, thus causing the two endpoints of its line to be pressed against the boundary as well. Formally, the stay and edit constraints from this example are simply constraints of the form constant b. However, when we come to compile code for repeatedly solving a collection of constraints, it will be important to handle these constraints e-ciently, since the value of b will change for each solution: for the stay constraints, it will be the value of v on the previous step; for the edit constraints, it will come from some external input. The other constraints will remain unaltered for each step. 4. Compiling projection for hierarchies We can now describe the complete algorithm for compiling code that can repeatedly nd a locally-error-better solution to a constraint hierarchy given a series of input values. We assume the constraint hierarchy is in hierarchical normal form. In this form, the only kind of preferential constraints are ones of the form variable v and constant b; all other constraints are required (i.e. we must satisfy them in any solution). In practice the original problem will usually already be in this form. If not, it is easy to transform an arbitrary linear constraint hierarchy into this form by adding error variables. For example, strong(e b), where e is a linear expression, becomes required(e b e is a new error variable. Recall that a hierarchical constraint H can be separated into parts H i , each of which contain the primitive constraints of strength i. If H is in hierarchical normal form, then for i 1 H i will contain only equations of the form a variable and b is a constant. We can use this information to build a strength partitioning of the variables. Let V i be dened as follows: contains those variables whose strongest non-required primitive constraint is of strength i. For simplicity, we assume that every variable is involved in a non-required constraint. This will normally be the case in constraint-based graphics applications. (If not, we can add a very weak stay constraint|weaker than any of the existing constraints|to any variable not otherwise involved in a non-required constraint. Any locally-error-better solution to the new hierarchical constraint will also be a locally-error-better solution to the original constraint.) Consider solving the constraints for Figure 2 augmented with the preferential constraints strong(x and H g. The strength partitioning of the variables gives g. Now we can use the projection algorithm to build code for solving hierarchical normal form constraints, where each problem diers from the others only in the values of the constants b in the non-required primitive constraints. Each problem corresponds to a solution of the constraint heirarchy required to determine the positions of objects for a screen refresh during the manipulation of the diagram. We apply the projection algorithm to eliminate all variables in the set V j before eliminating any variables in V i where i < j. A total ordering x 1 x 2 xn on the variables x in a hierarchical normal form constraint H respects the hierarchy if Returning to our example of solving the constraints on Figure 2 augmented with l ), an ordering that respects the heirarchy is xm x r x l . The full algorithm for generating code takes an ordering of variables and the required constraint C. To solve the constraint C in conjunction with H 1 , H 2 , . , H k , for each variable x l we select a constraint that is, we arbitrarily choose one of the highest strength non-required constraints on x i to determine the value b i . Making such a choice is correct because of the locally-error-better comparator: minimising the error of one of the strongest non-required constraints will always give a locally-error-better so- for end for for i := n to 1 if exists c 2 C i where c x else minset := ;; maxset := ;; for each primitive constraint c 2 C i where x if c e x i then minset else if c x i e then maxset end for each emit( else if b i x l end for emit( return Figure 7. Code generation algorithm lution. (By choosing a dierent constraint of the same strength, we would compute a dierent but still valid locally-error-better solution.) The code produced by the algorithm in Figure 7 will either set x i to b i if this is a legitimate choice given the solution determined so far, or set x i to its lower or upper bound, whichever is closest to the value b i . Theorem 1 (Correctness of code generate) Let C be a constraint with variables g. Given a variable ordering x 1 x 2 xn for variables in constraint C, the solution (d returned by the code produced by code will be a locally-error-better solution for the constrain is the same as or weaker than strength s i ). Proof: The value for each variable is chosen from the available range of values for that variable as determined by Fourier-Gaussian elimination. Thus, by the correctness of Fourier-Gaussian elimination, the algorithm always produces solutions that satisfy the required constraints in C. It remains to show that the solution produced is also locally-error-better. Assume to the contrary that (d not a locally-error-better solution to the constraints. Then there exists another solution (d 0 n ) that is locally-error- better than (d be the largest index where d 0 We examine the selection of the value of the variable x j . By the correctness of Fourier-Gaussian elimination, 9 fx j ;:::;x ng C ng C) Given that the variables x take the values d respectively, then clearly the only solutions of x j in C j are in the range [x l . Hence the error for the equation is greater for the solution (d 0 means n ) is not locally-error-better than (d since it has a greater error in the equation x has the same error for every non-required equation of higher strength. Otherwise, if b j < x l j and so d 0 j must be in the range [x l ]). Again the error for the constraint x is greater for (d 0 than for (d which implies that (d 0 n ) is not locally-error-better than The remaining case of b j > x u j is similar. Continuing with our running example of compiling the constraints for Figure 2 augmented with strong(x l ), the resulting code is shown in Figure 8. Note the very high similarity of this code with that of Figure 4. Suppose the midpoint of the line is selected by the mouse to move to position 60, and the remaining points are constrained to stay where they are and x 70). This then imposes constraints strong(x which the code in Figure 8 generates the answer (60; 70; 50). If the mouse now moves to position 70, the edit and stay constraints translate as and the code generates the answer (70; 75; 65). If the mouse now moves to position 0, the code generates the answer (5; 10; 0). 5. Redundancy management One of the problems with Fourier elimination is that a large number of constraints can be produced (potentially an exponential number), many of which are redundant. For a constraint compiler based on Fourier elimination to succeed, it is crucial x l x u else xm := x l else xm := x u x l x u r := min f 2xm, 100 r ::x u r else if b r < x l r x r := x l else x r := x u return Figure 8. Code for the midpoint example constraints that the issue of redundant constraints be addressed. Detecting and eliminating all redundancy is not very practical, but there are a number of types of redundancy that are cheap to detect, while still being eective at keeping the number of redundant constraints down. One of these is quasi-syntactic redundancy. A primitive constraint is quasi-syntactically redundant with respect to or c 2 a 1 x this is the case c 1 can be dropped without aecting the result. This test is inexpensive, O(n log n) for testing n primitive constraints, yet allows us to get rid of a signicant number of redundant constraints. The other main class of redundancy which we detect and eliminate targets the output constraints. During the projection process, the best known (constant) bounds are maintained for each variable. Then when a variable (say x) is to be eliminated, the constraints that are to yield upper and lower bounds for x are examined. Any constraint that can be shown to be redundant with respect to another, by using the bounds information accumulated earlier, is discarded. Doing this redundancy elimination at this point can drastically reduce the size of the cross-product formed during the elimination of x, and also directly reduces the number of compiled constraints The bounds information is used by evaluating the minimum and maximum values an expression can take, subject to the known bounds. Best use is made of this by examining the expressions for, say, the upper bounds on x in a pairwise fashion, and evaluating the dierence between the expressions. If the minimum possible value of the dierence is non-negative, the rst expression is guaranteed to be no smaller than the second; if the maximum value is non-positive, it is guaranteed to be no larger. In both of these cases, one of the constraints may be discarded as redundant; otherwise, no redundancy information can be deduced. By evaluating the dierence of the constraints in this fashion, rather than each constraint individually, one is able to obtain much more redundancy information. For example, suppose we have bounds on x of x 2y, x y + 10, and we know that 5 y 5. For the rst bound, we can deduce that it lies between 10 and 10; for the second it is between 5 and 15. Since 5 < 10, we cannot guarantee that the rst constraint will always be tighter than the second. However, if we look at the dierence (y 10), we can work out that it lies between 15 and 5, and thus can guarantee that the rst constraint will always be tighter than the second, and so can discard the latter as redundant. The second main approach to managing redundancy is to try to minimise the number of redundant constraints generated in the rst place. This can be done by being clever about the order in which variables are eliminated. The algorithm as it stands leaves us reasonable freedom in the choice of which variable to eliminate (we can choose variables at the same strength in any order we like). Some simple heuristics about which variable to eliminate next can make a substantial dierence to the number of constraints generated. An obvious and useful heuristic is to choose a variable appearing in an equation, so that Gaussian elimination can be used rather than Fourier. If there are no such variables, choosing x that minimises x jjC x j can work well (this minimises for the next step the sum of the number of output constraints and the number of constraints remaining, prior to redundancy elimination). The actual heuristic we use counts the number of linear terms in the constraints, rather than the number of constraints, since this discourages the formation of constraints with many terms, which are particularly problematic for Fourier elimination. We have based a more reasoned approach to selecting which variable to eliminate next on the results of Theorem 4 (see Section 8). This often allows us to process relatively small groups of variables and constraints at any given time, which limits the scope for producing redundant constraints. In practice, this approach seems to work well. 6. Renements In this section we discuss some renements to the solver that would appear to increase the class of problems it can solve, but can be implemented simply by altering the input to the main algorithm. required required required x l 0 required xm 0 required xr 0 required x l 100 required xm 100 required xr 100 strong x Figure 9. Midpoint example|local comparator version 6.1. Optimising a function Rather than just returning any solution satisfying the input constraints, it is possible to optimise a linear objective function. This can be done by introducing a new variable representing the value of the objective function, plus a required constraint relating it to the variables it depends on. Then a constraint of an appropriate strength can try to set the value of the objective function to some suitably large or small (minimise) value. 4 Note that one can actually have as many objective functions at as many dierent strengths as one wants. For instance, one could optimise a particular function subject to only the required constraints, then optimise another subject to the rst remaining optimal, then add some non-required constraints on the values of some variables, and then apply a nal optimisation: all by assigning appropriate strengths to each. Note that normally an objective function would not share its strength with anything else. For example, if two objective functions were assigned the same strength, the solver would be free in its choice of which to optimise rst. Allowing such a choice would not usually make sense unless the objective functions were independent of each other. 6.2. Global comparators In Sections 3 and 4, we assumed that we were working with local comparators. However, it is possible to achieve an eect equivalent to using any linear global comparator. The way this is done is similar to the introduction of an objective above, except that this time the function is in terms of error variables. This may require the introduction of extra error variables which were not needed before, as well as some required non-negativity constraints on the error variables. required required required x l 0 required xm 0 required xr 0 required x l 100 required xm 100 required xr 100 strong x required x l l required xm required xr r required e l 0 required e l 0 required e required e m 0 required e required e r 0 required e l Figure 10. Midpoint example|global comparator version Consider the simple midpoint example again, where we have added weak stay constraints on x l , xm and x r , plus a strong constraint on x l . These constraints are shown in Figure 9. Suppose we wish to minimise the sums of the errors in the constraints, with medium strength. The error in the weak constraint x is jx r Rj. We cannot encode this error directly using a linear constraint due to the modulus. Instead, we introduce a pair of new variables to represent the error: one for when the dierence is negative, and one for when it is positive. If each of these variables is then constrained to be non-negative, their sum can be used to represent the error. 5 Thus we have x r r , with e r 0 and e r 0. Taking into consideration all the weak constraints, the global error is then l g. To minimise this, we also add, with medium strength, The nal result is shown in Figure 10. Note that we have kept the stay constraints, to ensure that the variables are set to particular values. As can be seen, and not too surprisingly, using global comparators comes at a cost, in terms of the size and number of the constraints to be compiled. 7. Computational results The projection-based compilation algorithm has been implemented and tested. Our prototype implementation is in Mercury [14], and includes a module that is easily adapted to generate code for dierent target languages. The current implementation produces Smalltalk code, which is stored in a le. Then, in the Smalltalk environment, the code is loaded and incorporated into a graphics application for execution. The advantages of using Smalltalk are that Smalltalk includes an extensive graphics library, making it easy to test interactive graphical programs, and also that we have a Smalltalk implementation of Cassowary, a simplex-based solver, which allows a head-to-head comparison of the run times of the two algorithms. We have investigated the performance of our algorithm for several medium-sized examples of constraints for interactive graphics. The boxcars benchmark has a number of boxes in a horizontal row. Each box is constrained to be to the right of the previous one, and all are constrained to stay within a specied horizontal range. The binary tree benchmark is a complete binary tree of a given height. Each pair of children are constrained to be at the same height, both must be at least some minimum distance below their parent, and they must be separated from each other by some minimum distance. All parent nodes must be centred over their children, and nally every node must lie within a certain bounding box. This formulation has some redundancy|we could have specied only that the left child be the minimum distance below the parent, rather than both the left and right children, since the children are at the same height. However, this redundancy arises naturally in the specication. The grid benchmark is an n n grid of points where every point is constrained to be on an imaginary vertical line through the points above and below it, and on a horizontal line through the points to the left and right. No point can be within a given distance of its neighbours, and all points lie within a given box. Again this specication leads to redundancy in the constraints. The compilation results are shown in Table 1. For each benchmark we give the following information: the number of variables in the original formulation; the number of primitive constraints in the original formulation; the typical number of primitive constraints in the compiled code (that is, the number of inequalities used for minset and maxset calculations, plus the equations used to emit calculation code, for a typical choice of edit variable(s)); and the typical CPU time required for the Mercury program to produce the Smalltalk code The compilations were done on a DEC AlphaServer 8400 with eight 300 MHz 21164 processors. The compiled versions did not suer from an exponential blow-up in the number of primitive constraints, and indeed all contain fewer constraints than were in the input, the reduction being due to the redundancy elimination performed during projection. The code produced is extremely e-cient, and also has predictable performance irrespective of the input values. Table 2 gives some measurements of the execution speed of the compiled code as compared with the execution speed of the Cassowary constraint solver [7], which uses an e-cient simplex-based algorithm specically adapted for repeatedly solving constraints arising in interactive graphics applica- tions. All timings were done using OTI Smalltalk Version 4.0, running on an IBM Thinkpad 760EL laptop computer. Table 1. Compilation statistics Problem size variables primitive compiled time constraints constraints (secs) boxcars 50 50 149 100 0.7 100 100 299 200 1.4 200 200 599 400 4.5 binary tree 5 62 199 111 0.5 9 1022 3319 1910 46.2 Table 2. Runtime statistics Problem Time (milliseconds) Avg. # of pivots Fourier Cassowary Graphics (Cassowary only) 50 boxcars 1.0 5.0 38.0 0.11 (infrequent collisions) 50 boxcars 1.0 16.0 38.0 0.51 (frequent collisions) depth 5 tree 5.3 68.6 46.0 1.55 (frequent collisions) depth 5 tree 5.3 124.5 46.0 3.77 (v. frequent collisions) The times shown are the average number of milliseconds to perform one update, i.e. to solve the constraint problem with a new value for the mouse position. The Fourier and Cassowary times are both without graphics refresh; the additional refresh time is shown in the \Graphics" column. Finally, for Cassowary, the last column gives the average number of simplex pivots per update. There are two dierent runs given for the boxcars example, one with the input varying slowly, so that collisions between the boxcars are relatively infrequent, and another with more rapidly varying input, causing more collisions. The time to run the Fourier code is the same for both cases; however, with frequent collisions the Cassowary time per update increases substantially. The reason is that a collision generally requires a pivot in the simplex tableau, which is expensive; when there is no collision, the tableau can be updated very e-ciently with no pivoting required. There are also two runs for the tree example. We used a tree with the node displays closely spaced, so that there were many collisions (a moving node bumping into a stationary one) in both cases; the second run involved moving the mouse more quickly to generate even more collisions. For these examples the Fourier code is approximately 5 to 20 times faster than the runtime simplex solver. For the relatively simple constraints of the boxcars example, the graphics refresh time is substantially more than the time required to satisfy the constraints in any of the tests. However, when using the simplex solver for the tree example, the constraint solving time becomes signicant compared to the graphics refresh time. In addition, the simplex solver has variable solving time, which is quite apparent when the mouse is moved very quickly on this example| often the update is extremely fast, but when numerous pivots are needed it slows down, giving a less pleasant jerky quality to the interaction. In fairness to the simplex approach, it should be possible to extrapolate the current direction of the mouse movement and pre-solve some of the pivots; but this has not been done in any of the current systems. In addition, for use with real-time systems the predictable response provided by the compiled Fourier code is essential. 8. Complexity analysis Most variable elimination algorithms have bad worst case complexities. Projecting one variable from a system of m linear inequalities produces O(m 2 ) inequalities in the worst case. Hence eliminating n variables from m primitive constraints can be O(m 2 n However, our empirical results so far have shown quite reasonable performance for practical problems. In this section we attempt to analyse the situation and point out reasons for this. One major factor in the reasonable performance of our algorithm is that in many practical problems each primitive constraint only involves a small number of vari- ables, and hence the worst case does not arise. There are also a number of restricted cases that do have much more reasonable worst case complexities. One such restricted case is when each constraint involves at most 2 variables. The grid benchmark above falls into this category. The following result is due to Nelson [12]. Theorem 2 Let C be a set of m inequalities involving n variables where each inequality involves at most 2 variables. Fourier elimination is O(mn d2 log ne+3 log n). Another example where the worst case phenomena cannot occur is when almost all of the constraints are equations. Theorem 3 Let C be a constraint involving n variables, m linearly independent equations, and l inequalities. Then there is a choice of variable elimination order where Fourier elimination is O(nm(m Proof: Use the equations to eliminate m variables, leaving l inequalities in n m variables. The result follows. It also appears that, in many practical constraint problems in interactive graph- ics, the constraints are not tightly connected. We examine a class of constraint graph which is not tightly connected. A constraint graph for a constraint C is an undirected graph constructed as follows. There are nodes for each variable in and each primitive constraint c 2 C. There is an edge between variable node v and primitive constraint node c if v 2 vars(c). Two nodes x and y in a graph G are bi-connected if there exist two node-distinct paths in G from x to y. A bi-connected component of a graph G is a maximal set of nodes N such that each bi-connected in G. A constraint graph G is k-bi-connected if there are no bi-connected components of G with more than k variable nodes. The binary tree benchmark is an example where the constraint graph is 3-bi-connected. Theorem 4 Let C be a constraint involving n variables and m primitive constraints whose constraint graph is k-bi-connected. Then there is a choice of variable elimination order where a slightly modied Fourier elimination algorithm with quasi-syntactic redundancy elimination is O(nm 2 k Proof: We start by forming each bi-connected component of the graph into a cluster. These clusters form a tree (or a forest, if the original graph was not fully connected). We proceed to eliminate the clusters one at a time, starting at the leaves. A leaf cluster will be connected to the rest of the tree via a single node. If that node represents a variable, say x, our task is easy. We simply project all the constraints in the cluster onto x. The only possible resultant constraints on x are bounds, and using quasi syntactic redundancy there can be at most two: l x and x u. These bounds are added to the parent cluster. If, on the other hand, the connecting node represents a constraint, we introduce a temporary variable t and split the constraint rst. Let the constraint be of the form where op is either or =, e 1 is a linear expression containing only variables included in the cluster, and e 2 is a linear expression containing no variables from the cluster. Then the constraint is split into e 1 the former being added to the current leaf cluster, the latter replacing the original constraint, and the variable t being added to the parent cluster. We then project all the constraints in the current leaf cluster onto t, and add the resulting bounds to the parent cluster, as before. Note that the introduction of temporary variables can increase the number of variable nodes in a cluster. Each temporary variable appears in exactly one equation and at most two bounds constraints, though we note that more than one temporary variable may appear in the same equation (if the original constraint was a cut node where several bi-connected components met). If we eliminate the temporary variables rst, it is easy to see that each equation involving temporary variables yields at most two inequalities involving the remaining variables, and that this elimination is O(l), where l temporary variables were involved. This then leaves a cluster with at most k variables to which the standard Fourier elimination will be applied. Eliminating the variables in a leaf cluster never adds more constraints to the parent cluster than were eliminated in the leaf; hence the maximum number of constraints in any cluster at any given time is m. The maximum number of variables eliminated in any cluster when the standard algorithm is applied is k. Hence processing one cluster with the standard algorithm is O(m 2 k ). There are at most n clusters in the tree, and thus at most n temporary variables introduced. Hence the complexity result holds. Theorem 4 in particular constrains the order in which variables can be elimi- nated. In interactive graphics problems there are typically two variables with edit constraints (the x and y coordinates of a point being moved). These variables must be eliminated last. This is compatible with the ordering required by the theorems if the x and y constraints are independent (as is the case in all our examples, and in many other cases as well). However, if the problem includes constraints relating the x and y coordinates (e.g. that a shape be a square), then we may be unable to use the complexity guarantees provided by the theorems. 9. Conclusions Fourier elimination can be used to generate very fast, constraint-free code to solve problems arising in interactive graphical applications. The approach is useful for applications such as real-time systems which need predictable performance, for smoothing the response time in an interactive system, for producing applications that can be run without employing a runtime constraint solver, and when execution speed is very important. The same approach of compiling constraint solving by projection can be applied to any constraint domain that has a variable elimination function that projects one variable out of a constraint, along with a method for nding solutions of a conjunction of constraints in one variable. Constraint domains that meet these requirements include Boolean constraints, unit two variable per inequality (integer) constraints [11], and partial order constraints. The current algorithm is a batch one. A direction for future research is the design of an incremental version, which would reuse part of a previous solution when accommodating changes in the constraint (beyond simply changing the constants in the Some previous work has involved hybrid constraint solving algorithms, which partition a set of constraints into regions that can then be turned over to an appropriate sub-solver for that class of constraint and constraint topology [3]. Compilation based on Fourier elimination is a promising candidate for use in this architecture, to handle cyclic collections of linear equality and inequality constraints. Theorem 4 is particularly interesting, because it shows how the structure of the constraints to be solved can be exploited to make the actual solving more e-cient. Sabin and Freuder [13] have also studied how the structure of the constraint graph can be exploited; in their case, they demonstrate substantial improvements in the solving time of random binary CSPs, by identifying (an approximation of) the cycle cutset of the constraint graph. This seems a very interesting area to explore further. Acknowledgements This project has been funded in part by the National Science Foundation Grant IIS-9975990, Australian Research Council Large Grant A10017012. Alan Borning's visit to the University of Melbourne was sponsored in part by a Fulbright award. Notes 1. Current address wh@icparc.ic.ac.uk, IC-Parc, William Penney Laboratory, Imperial College, Exhibition Road, London SW7 2AZ, United Kingdom 2. Note that while user interface objects will be mapped to particular pixel locations on the display, and thus their coordinates should, in theory, be integral, in practice this is usually not necessary (or desirable). For most applications, rounding real coordinates to the nearest integer point is quite appropriate, so locations on the screen are treated as reals to make the constraint solving easier. In this work we assume coordinates are reals. We note that non-linear constraints do arise in user interface applications for such attributes as areas and angles, and nite domain constraint also arise in handling fonts. In such cases more sophisticated techniques are required (e.g. [9]). 3. An earlier version of this paper appears as [10]. 4. MAXFLOAT or similar should be su-cient for most applications, but if not, it would be trivial to change the solver to implement this feature directly. 5. Strictly, their sum is only an accurate representation of the error when at least one of the variables is zero, but this will be guaranteed by minimising the global error. --R The programming language aspects of ThingLab Indigo: A local propagation algorithm for inequality constraints. The OTI constraint solver: A constraint library for constructing interactive graphical user interfaces. Ultraviolet: A constraint satisfaction algorithm for interactive graphics. Constraint hierarchies. Constraints for the web. Solving linear arithmetic constraints for user interface applications. Object Technology International Compiling constraint solving using projection. An n log n algorithm for the two-variable-per-constraint linear programming satis ability problem Understanding and improving the MAC algorithm. The execution algorithm of Mer- cury: an e-cient purely declarative logic programming language --TR --CTR Jakob Mauss , Frank Seelisch , Mugur Ttar, A constraint solver for model-based engineering, AI Communications, v.17 n.2, p.75-92, April 2004
linear arithmetic constraints;fourier elimination algorithm;user interfaces;hierarchical constraints;constraint compilation
593358
Consistency Techniques in Ordinary Differential Equations.
This paper takes a fresh look at the application of interval analysis to ordinary differential equations and studies how consistency techniques can help address the accuracy problems typically exhibited by these methods, while trying to preserve their efficiency. It proposes to generalize interval techniques into a two-step process: a forward process that computes an enclosure and a backward process that reduces this enclosure. Consistency techniques apply naturally to the backward (pruning) step but can also be applied to the forward phase. The paper describes the framework, studies the various steps in detail, proposes a number of novel techniques, and gives some preliminary experimental results to indicate the potential of this new research avenue.
Introduction Differential equations (DE) are important in many scientific applications in areas such as physics, chemistry, and mechanics to name only a few. In addition, computers play a fundamental role in obtaining solutions to these systems. The Problem A (first-order) ordinary differential equation (ODE) system O is a system of the form un In the following, we use the vector representation u simply, Given an initial condition assuming existence and uniqueness of solution, the solution of O is a function s O and the initial condition s . Note that differential equations of order p (i.e. f(t; u; u can always be transformed into an ODE by introduction of new variables. Although an ODE system can potentially be transformed into autonomous ODE the addition of a new function un+1 (t) (with u 0 prefer to keep the time variable explicit for a clearer presentation of some of our novel techniques. However, the autonomous form is more appropriate for some treatment such as automatic differentiation. There exist different mathematical methods for proving the existence and uniqueness of a solution of an ODE system with initial value. But, in practice, a system is generally required, not only to prove existence, but also to produce numerical values of the solution s (t) for different values of variable t. If, for some classes of ODE systems, the solution can be represented in closed form (i.e. combination of elementary functions), it is safe to say that most ODE systems cannot be solve explicitly [Hen62]. For instance, the innocent-looking equation cannot be solved in terms of elementary functions! Discrete variable methods aim to approximate the solution s (t) of any ODE system, not over a continuous range of t, but only at some points t methods include one-step methods (where is approximated from the approximation methods (where s ? Proceedings of the Fourth International Conference on Principles and Practice of Constraint Programming, LNCS, 1998. approximated from the approximation In general, these methods do not guarantee the existence of a solution within a given bound and may suffer from traditional numerical problems of floating-point systems. Interval Analysis in ODE Interval techniques for ODE systems were introduced by Moore [Moo66]. (See [BBCG96] for a description and a bibliography of the application of interval analysis to ODE systems.) These methods provide numerically reliable enclosures of the exact solution at points t To achieve this result, they typically apply a one-step Taylor interval method and make extensive use of automatic differentiation to obtain the Taylor coefficients [Moo79, Ral80, Ral81, Cor88, Abe88]. The major problem of interval methods on ODE systems is the explosion of the size of resulting boxes at points t There are mainly two reasons for this explosion. On the one hand, step methods have a tendency to accummulate errors from point to point. On the other, the approximation of an aribitrary region by a box, called the wrapping effect, may introduce considerable imprecision after a number of steps. One of the best systems in this area is Lohner's AWA [Loh87, Sta96]. It uses the Picard iteration to prove existence and uniqueness and to find a rough enclosure of the solution. This rough enclosure is then used to compute correct enclosures using a mean value method and the Taylor expansion on a variational equation on global errors. It also applies coordinate transformations to reduce the wrapping effect. Goal of the Paper This paper mainly serves two purposes. First, it provides a unifying framework to extend traditional numerical techniques to intervals. In particular, the paper shows how to extend explicit and implicit, one-step and multi-step, methods to intervals. Second, the paper attempts to take a fresh look at the traditional problems encountered by interval techniques and to study how consistency techniques may help. It proposes to generalize interval techniques into a two-step process: a forward process that computes an enclosure and a backward process that reduces this enclosure. In addition, the paper studies how consistency techniques may help in improving the forward process and the wrapping effect. The new techniques proposed in this paper should be viewed as defining an experimental agenda to be carried out in the coming years. The techniques are reasonably simple mathematically and algorithmically and were motivated by the same intuitions as the techniques at the core of the Numerica system [VHLD97]. In this respect, they should complement well existing methods. But, as it was the case for Numerica, only extensive experimental evaluation will determine which combinations of these techniques is useful in practice and which application areas they are best suited for. Very preliminary experimental results illustrate the potential benefits. The rest of this paper is organized as follows. Section 2 provides the necessary background and notations. Section 3 presents the generic algorithm that can be instantiated to produce the various methods. Section 4 describes how to find bounding box. Section 5 describes the step methods used in the forward phase. Section 6 describes the backward pruning based on box-consistency. Section 7 discusses the wrapping effect. Section presents some experimental results. Section 9 concludes the paper. Background and Definitions This paper uses rather standard notations of interval programming. F denotes the set of F-numbers, D the set of boxes ' ! n whose bounds are in F , I the set of intervals ' ! whose bounds are in F , and D (possibly subscripted) denotes a box in D. Given a real r and a subset A of ! n , - r denotes the smallest interval in I containing r and 2A the smallest box in D containing A. If g is a function, - and G denote interval extensions of g. We also use g i (x) and G i (D) to denote the i th component of g(x) and G(D). The solution of an ODE system can be formalized mathematically as follows. Solution of an ODE System with Initial Value. A solution of an ODE system O with initial conditions u(t 0 O and the initial conditions s In this paper, we restrict attention to ODE systems that have a unique solution for a given initial value. Techniques to verify this hypothesis numerically are given in the paper. Moreover, in practice, as mentioned, the objective is to produce (an approximation of) the values of the solution function s of the system O at different points t . It is thus useful to adapt the definition of a solution to account for this practical motivation. Solution of an ODE System. The solution of an ODE system O is a function such that s(t is the solution of O with initial conditions u(t 0 The solution of an ODE system O can be used to obtain the solution of O at any point for any initial value. It is useful to extend our definition to sets of values. Solution of an ODE System. Let s be the solution of an ODE system O. The set solution O at t 1 wrt t 0 and D is the set s(t The interval techniques presented in this paper aim at approximating set solutions as tightly as possible. The next definition introduces the concept of bounding box that is fundamental to prove the existence and the uniqueness of a solution to an ODE system over a box and to bound the errors. Bounding Box. Let s be the solution of an ODE system O. A box B is a bounding box of s in Informally speaking, a bounding box is thus an enclosure of the solution on the whole interval following theorem is an interesting topological property of solutions. Theorem 5. Let O be an ODE system u s be the solution of O (i.e. existence and uniqueness), and let F r be the frontier of D. Then, 1. closed set; 2. is the frontier of s(t Proof. (Sketch) Under the given hypothesis, s 2 C [Har64]. It can then be shown that a point in D n F r cannot belong to the frontier of s(t As a consequence, s(t can be computed by considering the frontier of D j only. 3 The Generic Algorithm The interval methods described in this paper can be viewed as instantiations of a generic algorithm. It is useful to present the generic algorithm first and to describe its components in detail in the rest of the paper. The generic algorithm is parametrized by three procedures: a procedure to compute a bounding box, since bounding-boxes are fundamental in obtaining enclosures, a step procedure to compute forward, and a procedure to prune by using step procedures backwards. Procedure BoundingBox computes a bounding box of an ODE system in an interval for a given box. Procedure Step computes a box approximating the value of s (t j ) given the approximations of s and the bounding boxes Procedure Prune prunes the boxes D j at t j using the box D j \Gamma1 at . The intuition underlying the basic step of the generic algorithm is illustrated in Figure 1. The next three sections review these three components. Note however that it is possible to use several step procedures, in which case the intersection of their results is also an enclosure. begin forall(j in 1::n) begin Fig. 1. Computing correct enclosures of the solution 4 The Bounding Box This section considers how to obtain a bounding box for an ODE system. As will become clear later on, bounding boxes are fundamental to obtain reliable solutions to ODE systems. The traditional interval techniques to obtain bounding boxes are based on Picard operator [Har64, Moo79]. Theorem6 (Picard Operator). Let D 0 and B be two boxes such that D 0 ' B, let O be an ODE system u continuous and has a Jacobian (i.e first-order partial derivatives) over \Phi be the transformation (Picard Operator) where F be an interval extension of f . If \Phi(B) ' B, then 1. The O system with initial value u(t has a unique solution s; 2. \Phi(B) is a bounding box of s in [t Theorem 6 can be used for proving existence and uniqueness of a solution and for providing a bounding box [Loh87, Cor95]. A typical algorithm starts from an approximation B applies Picard operator. If \Phi(B 0 ) 6' B, the algorithm widens B 0 into B 1 (e.g., by doubling its size) and iterates the process. The algorithm can also narrow the step size. Note that the existence of Jacobian(f) can be checked numerically by evaluating its interval extension over the box. Note also that the Picard operator uses a Taylor expansion of order 1. It can be generalized for higher orders, which is interesting to increase the step sizes. 5 The Step Methods This section describes the step methods. The step methods are presented in isolation. However, as mentioned previously, they can be used together, since the intersection of their results is also a step method. 5.1 Explicit One-Step Methods This section considers one-step methods: It first describes traditional numerical methods, moves to traditional interval methods, and proposes improvements which can be obtained from consistency techniques. Traditional Numerical Methods To understand traditional interval methods, it is useful to review traditional numerical methods. In explicit one-step methods, the solution s of an ODE system O is viewed as the summation of two functions: where the function sc can always be computed while the function e cannot. As a consequence, a traditional numerical method based on an explicit one-step method is an algorithm of the form forall(i in 1::n) This algorithm tries to approximate the solution s (t) for an initial value u(t Example 1 Taylor Method. The Taylor method is one of the best known explicit one-step methods where the function sc is given by the Taylor expansion of a given order p, i.e., Interval Methods The key idea underlying (explicit or implicit) one-step interval methods is to define an extension of the interval solution s. 3 Interval Solution of an ODE System. Let s be the solution of an ODE system O. An interval solution of O is an interval extension S of s, i.e. As a consequence, a traditional interval method based on an one-step method is an algorithm of the form forall(i in 1::n) This algorithm provides safe intervals for s (t 1 Direct Interval Extensions Traditionally, interval solutions are often constructed by considering an explicit one-step function s(t by taking an interval extension SC of sc and by using a bounding box to bound the error function e to obtain a function of the form S(t Direct Explicit One-Step Interval Extension. Let s be an explicit one-step solution of an ODE system O of the form A direct explicit one-step interval extension of s is an interval solution S of the form where SC is an interval extension of sc, B 0 is a bounding box of s in [t is an interval extension of e. 3 As usual, interval solutions could also be defined on particular subsets of F and D. Example 2 Taylor Interval Solution. The Taylor Interval Solution of order p of an ODE system O is defined as bounding box of s in [t and the interval functions F (j) are interval extensions of functions f (j) inductively defined as follows [Moo66] (f (j) is the "total j th derivative of f wrt (j \Gamma1) @t 1-m-n (j \Gamma1) @um More information on automatic generation of the value of these functions can be found in [Moo79, Ral80, Piecewise Interval Extensions Direct interval techniques propagate entire boxes through interval solu- tions. As a consequence, errors may tend to accummulate as computations proceed. This section investigates a variety of techniques inspired by, and using, consistency techniques that can be proposed to reduce the accummulation of errors. The main idea, which is used several times in this paper and was inspired by box-consistency, is to propagate small boxes as illustrated in Figure 2. u2 Fig. 2. A Piecewise Interval Solution s be a solution to an ODE system O of the form A piecewise explicit one-step interval extension of s is a function S(t defined as where SC is an interval extension of sc, B 0 is a bounding box of s in [t is an interval extension of e. Piecewise interval extensions of an ODE system are not only a theoretical concept: they can in fact also be computed. The basic idea here is to express piecewise interval extension as unconstrained optimization problems. Let s be a solution to an ODE system O of the form A piecewise explicit one-step interval extension of s is a function S(t defined as where SC is an interval extension of sc, B 0 is a bounding box of s in [t is an interval extension of e. Note that these minimization problems must be solved globally to guarantee reliable solutions. The implementation section discusses how a system like Numerica may be generalized to solve these problems. The efficiency of the system of course depends on the step size, on the size of D 0 , and on the desired accuracy. It is interesting to observe that the function SC does not depend on the error term and hence methods that are not normally considered in the interval community (e.g., Runge-Kutta method) may turn beneficial from a computational standpoint. It is of course possible to sacrifice accuracy for computation time by using projec- tions, the fundamental idea behind consistency techniques. For instance, interval methods are generally very fast on one-dimensional problems, which partly explains why consistency techniques have been successful to solve systems of nonlinear equations. s be a solution to an ODE system O of the form A box-piecewise explicit one-step interval extension of s wrt dimension i is a function S i defined as where SC is an interval extension of sc, B 0 is a bounding box of s in [t is an interval extension of e. The box-piecewise explicit one-step interval extension of s wrt E and B is the function Each of the interval solutions reduces to a one-dimensional (interval) unconstrained optimization problem. The following property is a direct consequence of interval extensions. Solution. The piecewise and box-piecewise one-step interval extensions are interval solutions. In essence, box-optimal solutions safely approximate a multi-dimensional problem by the intersection of many one-dimensional problems. Of course, it is possible, and probably desirable, to define notions such as box(k)-piecewise interval solutions where projections are performed on several variables. Finally, notice that optimal interval solutions were defined with respect to a given bounding box. More precise interval solutions could be obtained if local bounding boxes were considered in the above definitions. It is easy to generalize our definitions to integrate this idea. 5.2 Implicit One-Step Methods This section considers implicit one-step method. It first reviews traditional numerical methods and shows how they can be generalized to obtain interval methods. The presentation essentially follows the same lines as the previous section. Traditional Numerical Methods In implicit one-step methods, the solution of ODE O is viewed as the solution of an equation. solution to an ODE system O is a function of the form s(t is the solution of an equation Since the error term cannot be computed in general, the above equation is replaced in practice by its approximation As a result, an implicit one-step method is an algorithm of the form forall(i in 1::n) where solve(S) returns an element x in Solution(S), the set of solutions of S. Example 3 Trapezoid Method. The trapezoid method is an implicit-one step method that consists of solving, at each step, an equation of the form u Interval Methods We now show how to generalize implicit one-step methods to intervals. The basic idea is to replace the search for a solution to a system of equations by a search for the solutions of a set of interval equations. The resulting interval solution can then be used as in explicit methods. Direct Implicit One-Step Interval Extension. Let s(t is the solution of the equation be an implicit one-step interval solution of an ODE system O. Let SC be an interval extension of sc and E be an interval extension of e. A direct implicit one-step interval extension of s is an interval function S(t and B 0 is a bounding box of s in [t Note that this definition amounts to finding all solutions of an "interval equation" in a box. The definition uses the bounding box as the initial search space. However, any step method can be used instead to provide a smaller search space. Example 4 Trapezoid Interval Method. The trapezoid interval extension of the trapezoid method requires the solving of the system of interval-valued equations is an interval extension of f , and B 0 is a bounding box of s in [t It is possible to improve this result by incorporating the idea of piecewise interval extension proposed earlier. Coarser extensions can be defined in a similar way as well. is the solution of the equation be an implicit one-step interval solution of an ODE system O. Let SC be an interval extension of sc and E be an interval extension of e. A piecewise implicit one-step interval extension of s is an interval function S(t are canonical g and B 0 is a bounding box of s in [t 4 The interval relation - is an interval extension of equality, i.e., 5.3 Multi-Step Methods Some methods for solving ODE are multistep methods that compute the value at point from values at point 1). Obviously, the value at points t must be computed by some other method. In order to define interval extensions of such methods, we extend our definition of solution and interval solution. We use the following notations Multistep Solution of an ODE System. Let s be the solution of an ODE system O. The multistep solution of O at t k wrt ~ t; ~ D is the set Solution of an ODE System. Let ms be the multistep solution of an ODE system O. A multistep interval solution of O is an interval extension S of ms, i.e. Explicit Multi-Step Methods In explicit multi-step methods, the solution s of ODE O is decomposed as These methods can be generalized to intervals in a way similar to one-step methods. For brevity, we only give an example of such an interval method. Example 5 Adams-Bashforth Interval Solution (Order 4). Let h be t 4. The Adams-Bashforth interval solution of order 4 is the multi-step interval solution where and B is a bounding box of s in [t . Notice that F (4) ([t can be approximated by is a bounding box of s in Implicit Multistep Methods Implicit multi-step methods can be defined in a similar fashion. Let u The value of u k is the solution of the equation : Example 6 Adams-Moulton Implicit Multi-Step Interval Solution. Let 3. The Adams-Moulton implicit multi-step interval solution is the function defined as where & D is canonical & B is a bounding box of s in [t 5.4 Mean Value Form Step Methods It can be observed that, in all our explicit interval solutions S, indicating that the intervals are thus growing. Mean value forms have been proposed to use contraction characteristics of functions and may return smaller intervals. From Equation 1, we may apply the mean value theorem on variable u) to obtain for some - between u and m. As a consequence, any interval solution of s, may serve as a basis to define a new interval solution. Let D be a box hI be the center of I i , and EM be an interval solution of an ODE system O. The MVF solution of O in D wrt SCM , denoted by - is the interval function d In the above definiton, the interval function d can be evaluated by automatic differentiation, during the evaluation of SC(t The definition also generalizes to multistep interval extensions. Implementation Issues Several of the novel techniques proposed in this section can be reduced to unconstrained optimization prob- lems. In general, interval techniques for unconstrained optimization problems require the function to satisfy a stability requirement (i.e., the optimum is not on the frontier of the box defining the search space). This requirement is not guaranteed in this context since, by Theorem 5, we know that the minimum of function s is on the frontier of D 0 , and we minimize function sc, an approximation of s. min-stability. A function g is min-stable for box there exists with be a box and be a function to minimize in K. Here are some necessary conditions for a point d in K to be a minimum when the function is not min-stable. @g is in the interior of K i @g @g Traditional interval algorithms for unconstrained minimization can be generalized to include the interval meta-constraints d d d with d an interval extension of , and i. The search can also be restricted to the frontier by adding the redundant constraint 1-i-n and applying techniques such as constructive disjunction [VHSDar]. Note that combining these two necessary conditions require some extra care to preserve correctness. 6 Backwards Pruning: Box-Consistency for ODE This section proposes another technique to address the growth of intervals in the step methods. The fundamental intuition here is illustrated in Figure 3. We know that all the solutions at are in D j \Gamma1 . If, in D j , there is some box H such that then we know that the box H is not part of the solution at t j . In other words, it is possible to use the step methods backwards to determine whether pieces of the box can be pruned away. This section formalizes this idea in terms of box-consistency. Fig. 3. Pruning consistency aims at reducing a box D j at t j given that the solution at are known to be in D j \Gamma1 . Interval Projection of an ODE System. An interval projection ODE hS; ii is the association of an interval solution S and of an index i (1 - i - n). Consistency of an ODE System. Let S be an interval solution of an ODE system O. An interval projection ODE hS; ii is box-consistent at I where solution is box-consistent at its projections are box-consistent at Proposition22. Let D projection ODE hS; ii is box-consistent at and, when l Traditional propagation algorithms can now be defined to enforce box-consistency of ODE systems. 7 The Wrapping Effect The wrapping effect is the name given to the error resulting from the enclosure of a region (which is not a box) by a box. It only occurs for multidimensional function. In one dimension, a perfect interval extension of a continuous function g always yields the correct interval. However, a perfect interval extension of a multidimentional function g introduces overestimations in the resulting box, because the set Dg is not necessarily a box. This effect is especially important when the enclosure is used for finding a new region which is also enclosed by a box. The wrapping effect is thus central in interval methods for ODE. The following classical example, due to Moore [Moo66] and explained in [Cor95] , illustrates this problem : 0:9 1:1 The trajectories of individual point-valued solutions of this ODE are circles in the ((u) 1 ; (u) 2 )-phase space. The set of solution values is a rotated rectangle. Figure 4 shows that the resulting boxes at Moore shows that the width of the enclosures grow exponentially even if the stepwise converges to zero. The wrapping effect can be reduced by changing the coordinate system at each step of the computation process. The idea is to choose a coordinate system more appropriate to the shape of reducing the overestimation of the box representation of this set, as illustrated in Figure 5. Fig. 4. The Wrapping Effect (u)(u)Fig. 5. Reducing the Overestimation by Coordinate Transformation An appropriate coordinate system has to be chosen at each step. Assuming that such coordinate systems are given by mean of (invertible) matrices M j , a naive approach, based on an explicit one-step method, would consist of computing where D 0 j and D 0 are the boxes at t j and t j \Gamma1 in their local coordinate system. This approach is naive since it introduces three wrapping effects: in to restore the original coordinate system needed to compute S, in the computation of S, and in the computation of M \Gamma1 to produce the result in the new coordinate system. To remedy this limitation, more advanced techniques (see, for instance, [Loh87, Ste71, DS76]) have been proposed but they are all bound to a specific step procedure. For instance, Lohner merges the two naive steps together using a mean value form and use associativity in the matrix products to try eliminating the wrapping effect. More precisely, the key term to be evaluated in his step method is of the and the goal is to choose M \Gamma1 j so that M \Gamma1 is close to an identity matrix. Piecewise interval extensions, however, reduce the wrapping effect in the naive method substantially, as illustrated in Figure 6. The overestimations of on ffl-boxes introduce wrapping Fig. 6. Coordinate transformation on ffl-boxes Taylor Piecewise Taylor Exact solution 2.0 [-7.388637 , 7.388637] 5359.51% [-0.135783 , 0.135783] 0.33% [-0.135335 , 0.135335] 2.5 [-12.181631 , 12.181631] 14740.26% [-0.082424 , 0.082424] 0.41% [-0.082085 , 0.082085] Table 1. ODE u effects that are small compared to the overall size of the box and to the benefits of using piecewise interval extensions. In addition, this reduction of the wrapping effect is not tailored to a specific step method. The basic idea is thus (1) to find a linear approximation of to compute the matrix j from the linear relaxation; (3) to apply the naive method on ffl-boxes. Step (1) can be obtained by using, for instance, a Taylor extension, while Step (2) can use Lohner's method that consists of obtaining a QR factorization of the linear relaxation. Lohner's method has the benefit of being numerically stable. 8 Experimental Results This section compares some standard techniques with piecewise interval extensions. This goal is to show that consitency techniques can bring substantial gain in precision. The results were computed with Numerica with a precision of 1e-8, using optimal bounding boxes. Consider the ODE u for an initial box [-1,1] at t compares the results obtained by an interval Taylor method of order 4 with step size 0:5, the results obtained by the piecewise interval extension of the same method, and the exact solutions. Relative errors on the size of the boxes are also given. As can be seen, the intervals of the traditional Taylor method grow quickly, although this function is actually contracting. The piecewise interval extension, on the other hand, is close to the exact solutions and is able to exploit the contraction characteristics of the function. Consider now the ODE u for an initial box [0:1; 0:4] at t compares the results obtained by a mean value form of a Taylor method of order 4, the results obtained by the piecewise interval extension of the Taylor method of order 4, and the exact solutions. Once again, it can be seen that the standard method leads to an explosion of the size of the intervals, while the piecewise interval extension is close to the exact results. Note that the Taylor method of order 4 also behaves badly on this ODE. Taylor MVF Piecewise Taylor Exact solution 2.5 Table 2. ODE u Taylor Piecewise Taylor Exact solution Table 3. ODE u 0 Consider now the system ae u 0 for an initial box ([5.9,6.1],[3.9,4.1]) at t compares the results obtained by an interval Taylor method of order 4, the results obtained by the piecewise interval extension of the same method, and the exact solutions. Once again, similar results can be observed. 9 Conclusion This paper studied the application of interval analysis and consistency techniques to ordinary differential equations. It presented a unifying framework to extend traditional numerical techniques to intervals showing, in particular, how to extend explicit and implicit, one-step and multi-step, methods to intervals. The paper also took a fresh look at the traditional problems encountered by interval techniques and studied how consistency techniques may help. It proposed to generalize interval techniques into a two-step process: a forward process that computes an enclosure and a backward process that reduces this enclosures. In addition, the paper studied how consistency techniques may help in improving the forward process and the wrapping effect. Very preliminary results indicate the potential benefits of the approach. Our current work focuses on the full implementation and experimental evaluation of the techniques proposed in this paper in order to determine which combinations of these techniques will be effective in practice. Future work will also be devoted to the application of consistency techniques to ODE systems with boundary values, since interval analysis and consistency techniques are particularly well-adapted when compared to traditional methods (as observed by several members of the community). Acknowledgment Many thanks to Philippe Delsarte and Jean Mawhin for fruitfull discussions. This research is partially supported by the Actions de recherche concert'ees (ARC/95/00-187) of the Direction g'en'erale de la Recherche Scientifique - Communaut'e Fran-caise de Belgique and by an NSF NYI award. --R Precise Numerical Analysis. Computational Differen- tiation: Techniques Applications of differentiation arithmetic. Theory of Numerics in Ordinary and Partial Diffential Equations (W. Consistency Techniques in Ordinary Differential Equa- tions Guaranteed Error Boubds for the Initial Value problem Using Polytope Arithmetic. Ordinary Differential Equations. Discrete Variable Methods in Ordinary Differential Equations. Enclosing the solutions of ordinary initial and boundary value problems. Interval Analysis. Methods and Applications of Interval Analysis. Applications of software for automatic differentiation in numerical computation. Automatic Differentiation: Techniques and Applications Enclosing Solutions of Ordinary Differential Equations. A Heuristic to Reduce the Wrapping Effect in the Numerical Solution of ODE. Modeling Language for Global Optimization. The Design --TR --CTR Marco Kletting , Andreas Rauh , Harald Aschemann , Eberhard P. Hofer, Consistency tests in guaranteed simulation of nonlinear uncertain systems with application to an activated sludge process, Journal of Computational and Applied Mathematics, v.199 n.2, p.213-219, 15 February 2007
differential equations;consistency;continuous problems
593360
Excluding Symmetries in Constraint-Based Search.
We introduce a new method, called i>symmetry excluding search (SES), for excluding symmetries in constraint based search. To our knowledge, it is the first declarative method that can be applied to i>arbitrary symmetries. The SES-method is based on the notion of symmetric constraints, which are used in our modification of a general constraint based search algorithm. The method does not influence the search strategy. Furthermore, it can be used with either the full set of symmetries, or a subset of all symmetries.We proof correctness, completeness and symmetry exclusion properties of our method. Then, we show how to apply the SES-method in the special case of geometric symmetries (rotations and reflections) and permutation symmetries. Furthermore, we give results from practical applications.
Introduction In many search problems, one is faced with the existence of symmetries. Symmetries give rise to many dierent solutions found by the search procedure, which are all considered to be "similar". Often, one is not interested in also getting all the symmetric solutions to every found solution. Without exclusion of symme- tries, whenever a solution is found or proven to be inconsistent with the search problem, the search algorithm still considers all symmetric solutions. Those symmetries will then give rise to an (often exponential) amplication of the search space. Hence, symmetry exclusion promises to eciently prune the search tree. E.g., consider constraint problems, where nite domain variables have a geometric interpretation such as in the N-queens problem or the square-tiling problem (where set of squares of given sizes must t exactly into a xed square). A more complex, real-world problem is the lattice protein structure prediction problem. In [2], it is shown how to nd solutions for this problem using constraint programming techniques. In this case, dierent solutions are symmetric if they can be generated using re ections or rotations. In other problems, symmetric solutions can be generated by performing permutations on variables valuations (like in the map coloring (or graph coloring) problem). In the following, we consider search problems that stems from constraint satisfaction problems (CSP). A common approach is to transform (problem specif- ically) a CSP C 1 into a CSP C 2 that excludes at least some of the symmetries of the original problem C 1 . Unfortunately, symmetry exclusion was often not straightforward to imple- ment. Even then, it had to be redesigned for every special problem or enumeration strategy. This leaded to an in exibility in the program structure once it was introduced. Thus, the widespread usage of symmetry exclusion was strongly hindered by its complexity. Often symmetry exclusion was not done at all or only for a small set of symmetries. In other cases, where symmetry exclusion was implemented it distracted the programmers attention from more important tasks. In this paper, we will present a new approach for symmetry exclusion, which works by modifying the search algorithm. The technique of modifying the search algorithm to exclude symmetries was already used in the literature. In contrast to the known approaches, our approach is (to our knowledge) the rst declarative method that can be applied to arbitrary symmetries that can be dened declar- atively. Furthermore, it does not restrict the search strategy (i.e., the symmetry exclusion can be done independently of the search strategy). This is important since in many constraint problems, major part of the knowledge of the problem is encoded in the search strategy. We have implemented our method in the concurrent constraint programming language Oz [10] using the programmable search engine described in [9]. But the method can be implemented in any system that handles implication (via and allows to modify the search procedure such that additional constraints are added in one branch of a branching step. Related Work Previous work on symmetry exclusion by modifying the search algorithm handled only restricted forms of symmetries. In [7], the notion of interchangeable values is introduced. Two values a and b are interchangeable for a variable V i for every solution that maps V to a, the corresponding valuation that maps V to b is also a solution. The method in [3] handles only permutations of domain values in the case of binary constraints (thus, e.g., the all-distinct constraint has to be translated into binary inequalities, which implies that ecient propagation methods for all- distinct cannot be applied). Furthermore, it works by introducing a symmetry excluding form of constraint propagation, which is a modied form of domain propagation (thus, the method cannot make use of interval propagation). [1] and [4] consider only propositional calculus. The symmetries there are permutations of Boolean variables. These methods do not apply to general constraint satisfaction problems. The symmetry exclusion algorithm presented in [8] is essentially an implementation and application of our method to solve practical problems, which will be compared in greater detail in Section 3. An example of a symmetry exclusion method that works by a (not problem specic) transformation of the constraint satisfaction problem is [5]. They introduce symmetry breaking predicates, which are true only for the smallest solution within an equivalence class (where the ordering is xed in advance). Thus, if the search strategy encounters a non-minimal (according to the pre-xed ordering) solution rst, then it will be excluded by the symmetry breaking predicates. This implies that not every search strategy is possible given the pre-xed ordering. In contrast, there is no prexed ordering in which symmetrical solutions will be excluded by our method. This implies that we respect used-dened search strategies as much as possible. 1 Overview of the paper We start with an introductory example explaining the basic ideas of our method. After some preliminaries, we present the formal de- nition as wells as completeness and correctness proofs in Section 3. Additionally, we show that our method guarantees to exclude all specied symmetries. The proofs indicate that a subset of all symmetries can be enough to exclude the complete symmetry group. In Section 4, we treat geometric symmetries. We show how to dene symmetric constraints and present the results of applying the problem to a complex problem. In Section 5, we consider symmetries which are generated by permuting the values of the variables. E.g., such symmetries occur in the graph coloring problem. We proof, that in this case, it is sucient to exclude the subset of transposition in order to exclude all permutations. Fur- thermore, we present the results of applying the method to random instances of graph coloring. Introductory Example We start with a simple example to explain the main concepts. Consider the N-queens problem, where we have an array Q[1::N ], whose elements can take values from 1::N . states that there is a queen at position (i; j). Now consider the symmetry corresponding to the re ection at the axis 2 , which is parallel to the x-axis (in the sequel denoted by Rx). Clearly, for every solution of the N-queens, the the Rx-symmetric version is also a solution of N-queens. Hence, we want to exclude these symmetric solutions. To use our method, we have rst to introduce symmetric versions of the constraints which are inserted by the search procedure. One kind of constraint is j. If we have any distribution of queens on the board satisfying j, then the Rx-symmetric distribution will satisfy j. Hence, the constraint is the Rx-symmetric version of the constraint Similarly, we get that the Rx-symmetric constraint of Q[i] j is Q[i] N j. In the following, we write S Rx (c) to denote the Rx-symmetric constraint to c. Our method works by adding additional constraints at the right branches of the search tree, which exclude the symmetric solutions that are encountered in the right branch. Consider the following search tree for the 10-queens problem, where we indicate the constraints added by symmetry exclusion via frames: Clearly, if the search strategy depends on the constraint store, then any method of symmetry exclusion that prunes the search tree must in uence the way solutions are encountered (since the constraint store is changed by the symmetry exclusion). This can be interpreted intuitively as follows. In the topmost branch, if we search for a solution in the right branch satisfying Q[1] > 5, then the Rx-symmetric solution will satisfy Q[1] < 5. Hence, the symmetric solution was already encountered earlier in the left branch (under Q[1] 5), which is the reason that we can close the topmost right branch. For the second right branch labeled Q[2] 6= 3, we want again to exclude the Rx-symmetric solutions found in the right branch. Hence, we would like to add the constraint :S Rx to the right branch. But this would exclude too many solutions. The prerequisite of the Rx-symmetric solution to be found in the right branch is that both the solution and its Rx-symmetric version satises Q[1] 5. Now the only solutions that satisfying this conditions are solutions that satisfy 5. Hence, we can add under the condition that 5. But this is exactly the eect that we achieve by adding at the second right branch (since the constraint store contains already Q[1] 5). Preliminaries We x a rst-order signature including the equality : with a set of variables . Constraints are literals, and constraint formulae are quantier-free formulae over . We identify t. C denotes the set of all constraints. A set of constraints C C is interpreted as the conjunction of the constraints contained in C, and we will freely mix set notation and conjunction. The set of free variables of C is denoted by V(C). We x a standard interpretation A with domain D A , which describes our constraint theory. In the following, we assume a xed constraint set C Pr describing the problem to be solved. An assignment in A is a partial function We say that an (possible partial) assignment satises (short there is a total assignment 0 with A; 0 j= . We write for every we have . A constraint set C (syntactically) determines a set of variables X to an assignment i for all x 2 X exists a t such that x In many problems, one is interested in only a subset of all variables, whose valuations uniquely determine the valuation of the other (auxiliary) variables. This is captured by the denition of solution variables. Denition 1 (Solution Variables). Let X V be a set of variables. X is a solution variables set for C if where 9!() ('exists a unique satisfying ()') is short for solution variables set X , we say that is a X -solution for C if we denote the set of X -solutions of C. In the following, we x X . Hence, we use the term 'solution' as short for 'X -solution for C', and we write kCk as short for kCk X . Denition 2 (Symmetry). A symmetry s for C Pr is a bijective function A symmetry set S for C Pr is a set of symmetries operating on kC Pr k. A symmetry group S is a symmetry set for C Pr which is a group. Note that for every symmetry s for C Pr , also s 1 is a symmetry for C Pr . We denote the identity function on kC Pr k with id CPr (which is a symmetry by def- inition). Clearly, the set of all symmetries for C Pr is a group. But many cases, we do not want to consider all symmetries, since either there are too many of them, or some of them do not have an intuitive characterization. In this case, we consider only a subset of all symmetries, which may not be a group. 3 Symmetry Excluding Search Trees Now we have to dene the notion of a search tree as we need it for the symmetry exclusion. We describe the eects of applying a constraint solver to a constraint set C by a function P (C), which returns the closure of C under constraint propagation. Clearly, we have P (C) j=j C. Usually, a search tree is a labeled tree, where the nodes are labeled with the constraints in the constraint store, and the edges are labled by constraints or negated constraints. A single node v is of the form c C store The root node is labeled with P (C Pr ). Additionally, we need for the symmetry exclusion to store at a node v the positive and negative constraints that have been used along the path from the root to v to generate the node v. Denition 3 (Search Tree). Let t be a nite, binary, rooted and ordered tree, whose egdes are labeled by literals, and whose nodes are labeled by triples of constraint sets. The tree t is a search tree for C Pr if the following conditions are 1. The root node has the label (;; ;; P (C Pr )), 2. every binary node has the form store c :c store with C l store store Intuitively, the reason for distinguishing between positive and negative constraints is just that Cn describes the previously found solutions, i.e. Here, we denote with t the partial ordering of nodes induced by t. Note that we do not force C l store (resp. C r store ) to be equivalent to C store ^ c (resp. C store ^ :c). The reason is that we do have to add some additional constraints during search for excluding symmetries. Since C store describes all constraints valid at some node v, we set for a node v in t with label We now state general properties of search tree in the presence of symmetries. Denition 4 (Expanded, C Pr -Complete and S-Reduced Trees). The search tree t is completely expanded if every unary node either determines X to . A search tree t is C Pr -complete if for every 2 kC Pr k there is a leaf v in t with S be a symmetry set for C Pr . A search tree is C Pr -complete w.r.t. S if for every there is a leaf v for such that A search tree is S-reduced if for every leaf v with Note that (1) is equivalent to 8s 2 S 8v 0 S is closed under inversion. Before we can show how to produce a S-reduced search tree that is C Pr - complete w.r.t S, we rst have to dene how symmetries operate on constraints. In the following, we will assume that for every constraint c and every s, there is a constraint c 0 such that s s on sets of C Pr - solutions by s Ag. For every c, there are usually dierent constraints c 0 with s k. Hence, x a function s con on constraints such that ks con (c)k: In the next sections, we give examples of how to construct this function. Proposition 1. Let s con be dened as before. Then s con distributes over the Boolean operators, i.e., s (kc^c 0 ks con (c)^s con ks con (c)_ ks con (:c)k. Hence, we identify s con with its homomorphic extension to constraint sets and arbitrary formulae. Now we can state a general mechanism for generating a search tree that is S-reduced. Denition 5 (S-excluding). Let S be a symmetry set. A search tree for C Pr is S-excluding if every binary node v has the form store c :c store with C l Before we prove S-reducedness and C Pr -completeness w.r.t S of S-excluding search trees, we state a proposition to precise the eect of adding the implications in denition 5 to C r store . For convenience, we write C v store to access store in the label (C store ) of v. For a binary node v we refer to its left child as v l and to its right child as v r . Proposition 2. Let S be a symmetry set and t a S-excluding search tree. For every symmetry s 2 S and for every node v of t we have C v store s con (C v For notational convenience we introduce the notation C path C Pr . A good intuition is to think of C path (v) describing the constraint store of v in a simple not symmetry excluding search tree. Reducedness and completeness will be essentially corollaries to the following proposition. Lemma 1. Let S be a symmetry set. Every S-excluding search tree t for C Pr satises for every node v where A v Proof. We proof this by tree induction. For the root this is valid, since A v and (v)k. Assume that we have proven the claim for a binary node v: store c :c store For the left child v l the claim follows immediately from the induction hy- potheses, since A v l S . For v r , we have to show kv r S . Note that A vr subdivides into its two subsets A v S and )k)g. Further according to our denitions C path (v r ) is equivalent to C path (v) ^ :c. To show that kv r k kC path (v) ^ :ck A vr then we have to show for every v 0 t v that The rst case is v 0 t v. Let 2 kC path (v 0 )k. Then s() 62 kvk by induction hypotheses. Since kv r k kvk, this immediately implies s() 62 kv r k. The second case is (v 0 6 t v) and v 0 t v r , i.e., v 0 is a subnode of v l . Let k. From 2 kv l k we have s con (C p ^c). That's a contradiction, because from denition of S-excluding search tree s() It remains to be shown that kv r k kC path (v r )k A vr S Let 62 kv r k. We have to show 62 kC path (v r )k A vr We have the following cases: 1. 62 kC path (v r )k. Then (3) follows immediately. 2. 2 kC path (v r )k and 62 kvk. We have to show 2 A vr S . Now 2 kC path (v r )k implies 2 kC path (v)k and henceforth 2 A v by induction hypotheses. 3. 2 kC path (v r )k and 2 kvk. We will show that Since store , there is at least one s 2 S with 6j= (c) by the denition of an S-excluding tree. Fix one of these symmetries s. Then by Proposition 2, we get Hence, s 1 (), which exists by denition of symmetries, satises s 1 () that C path with we have a valuation such that 2 kC path (v l )k and which shows that is in the set dened by (4). Theorem 1. Let S be a symmetry set. Every S-excluding search tree t for C Pr is S-reduced. Proof. Fix a symmetry s 2 S, let v be a leaf of t with fg. We have to show for every node v 0 of t with v 0 t v that kv 0 v 0 to be a node of t with v 0 t v and kv It follows and from this the contradiction 62 kvk by Lemma 1. By now we understand that an S-excluding search tree may exclude more symmetries than actually declared by S directly. Hence, we have to investigate the completeness property of S-excluding search trees. Theorem 2. Let S be a symmetry set. Every S-excluding search tree t for C Pr is C Pr -complete w.r.t S 0 , where S 0 is the closure of S under composition and inversion. Proof. We have to show that t is C Pr -complete w.r.t. to S 0 , where S 0 is the closure under composition of S, i.e. for every 2 kC Pr k there is a leaf v with Hence x an 2 kC Pr k. Then there exists a leaf v 0 in t with 2 kC path (v 0 )k from denition 5. There are two cases. 1. 2 kv 0 k. This yields immediately kv and (5) is proven. 2. 62 kv 0 k. By Lemma 1 follows 2 A v 0 there is a s 1 2 S and a path (v 00 )k: By induction we get a sequence of symmetries s for a leaf v (n+1) of t. Since S 0 is the closure of S, claim (5) follows. We will given an example for symmetry excluding search trees in Secion 5, where the symmetrie group of permutation are excluded by the subset of all transpositions (as indicated by the above theorem). Now we are able to give a more detailed comparison with [8]. The SBDS- method therein is essential our symmetry exclusion restricted to the case where the search tree branches over constraints of the form V ar al. In this case, the constraint collected in C p is of the form V ar C p is a partial assignment A. For the restricted SBDS-method, consider the case that a symmetry S is dened by a permutation rst. Then S al n at some node v of the search tree, which branches over V ar al. Then the constraint added by our method for excluding the symmetry S is S (V ar which is by the denition of S the same as ar Now the antecedence is nothing else than a test whether S leaves C p (i.e., the partial assignment at the node v) unchanged. Thus, the entailment will be evaluated and removed directly at the right subnode of v. The restricted SBDS- method essentially replaces the entailment test by the test whether the symmetry leaves the assignment unchanged. Only those symmetries are excluded, where the test can be decided at the node the corresponding implication would be integrated by our method (even if the corresponding implication would delay this test to a later point). Note on implementation For the symmetry exclusion we have to add implications of the form s con (C v during the search procedure. Hence, one has to change the search procedure for implementing the symmetry exclusion. The antecedents of these implications can be computed during the search eciently, using the observation that the antecedences of the added implications of ancestor nodes are prexes of the ones added in their ospring. To be more precise, let v be a node, which branches over the constraint c, and let v l (resp. v r ) its left (resp. right) daughter. Now s con (C v l For Therefore the computation of the condition s con (C p ) can be done incrementally. We maintain Boolean variables for this reason in our implementation. We need one variable for each symmetry s, containing the values for s con (C p ). 4 Geometric Symmetries We will give now a concrete example for symmetries, namely geometry symme- tries. We will treat the special case where we have points in Z d , although our method is not restricted to this case. We will exemplify the symmetry constructions in two dimensions, and give an example for the exclusion of additional symmetries as indicated by Theorem 2. Denition of Geometric Symmetries There are many problems where one encodes the two-dimensional position of an object i using nite domain integer variables . Examples are the N-queens and the tiling problem. A more complex and realistic example (in three dimensions) is the lattice protein structure prediction problem [2]. The symmetries for Z 2 have exactly the same structure as for the general case Z d . They are dened by ane mappings that map Z d onto Z d . I.e., the matrix AS is an orthogonal matrix with the property that the set of columns fv of AS equals fe j e is a unit-vector of Z d g: E.g., for Z 2 , the matrix 0 1 denotes the rotation by 90 . For Z 2 , we have 8 symmetries consisting of the identity, the 4 re ections (at x- and y-axis, and the two diagonals) and the 3 rotations by 90 , 180 and 270 . For Z 3 , we have 48 symmetries including the identity. By now, the vector v S is not yet xed. There are two dierent approaches for xing the symmetry. We will consider Z 2 as an example. The methods work for all other dimensions as well. The rst case is that every possible solution lies within a xed square (in the general case, within a hypercube). 2 This is equivalent to the proposition that there are integers xmin such that for all 2 kC Pr k we have 2 The technique can be extended to the case that the hypercube is not xed in advance. Thus, the minimal square around the position of all objects is dened by the points in every solution. We call this the frame of the problem C Pr . In the N-queens problem, this is just the board. Now knowing the frame of a problem C Pr , we can x the vector v S for all symmetries. Consider as an example a problem whose frame is dened by (0; and (3; 2). Furthermore, consider the three symmetries re ection at the y-axis, rotation by 90 and rotation by 180 , which we will name S 1 , S 2 and S 3 in the following. The corresponding mappings are dened by S i where the matrices AS1 ; AS2 and AS3 are dened by The corresponding mappings of the frame are A symmetry S is compatible with the above dened frame if the frame is mapped to itself, i.e., if fv j (0; 2)g. For a given matrix AS , there exists a v S such that condition if and only if AS satises AS (3; For the matrices AS1 , AS2 and AS3 , we get ( 3; 2), ( 2; 3) and ( 3; 2), which excludes the symmetry characterized by AS2 . We nally get The second case is that we know a point should remain unchanged under the symmetries. In that case, we know that the symmetries are dened by S i (p). The remaining part is to dene symmetric constraints. We use a specic example where we leave the point (5; 5) x. Consider the two symmetries re ection at the y-axis and rotation by 90 . By what we have said above, the corresponding mappings are Now suppose that we have modeled points using variables and we want to dene S Ry con and S 90 con for the constraints of the other kind of constraints c, the denition S con (c) is analo- gous). Now the symmetric constraints S Ry con con must express the constraint valid for S Ry () (resp. S 90 for every possible with does not restrict the valuation of (Y i ), we know that S Ry con is con Analagously, we get that S Ry con is and that S 90 con Note that S Ry (c) has the same type as c (i.e., both have the same variable). This does not hold for c and S 90 (c). We will now turn to an example where we exclude only one symmetry, but the exclusion method excludes compositions of this symmetry. This shows that there are cases where the symmetry exclusion method excludes more than the symmetries contained in the symmetry set S (as indicated by Theorem 2). We will use an articial example to show the existence of such a case. Example 1. Let C Pr be a geometric constraint problem over the variables X;Y , where X and Y have the domain [0::10] associated. Furthermore, suppose that C Pr has (6; 6) as a solution that we can apply the rotational symmetries with x point (5; 5). This implies, that (4; 4), (4; are also solutions of C Pr . Now consider the following S-excluding search tree for (where the constraints added by the symmetry exclusion are in a framed box): Then the solution (4; (which is (6; rotated by 90 ) is excluded by the symmetry exclusion in t 1 . But furthermore, the solution (4; is excluded in t 2 by our symmetry exclusion method. This is (6; rotated by 180 . Hence, although contains only S 90 , the symmetry S 180 (which is just S 90 excluded by our method. Note that (6; 4) (which is (6; rotated by 270 ) is not excluded and may be found in t 2 . Clearly, in the case of geometric symmetries, the full symmetry group is usually small enough that we can include all symmetries. But we wanted to show the eects of excluding only a subset of the full symmetry group using the above simple example. In the next section on permutation, we will give an example where the exclusion of transpositions allows to exclude all permutations. Results We have applied the method to the lattice protein structure prediction [2], which is a hard combinatorical problem. The following table shows the number of solutions, search steps and runtimes for nding all minimal energy structures for 4 sequences, in the rst case without symmetry exclusions, and with symmetry exclusion in the second part: without sym. excl. with sym. excl. ratio seq. length #noex nnoex tnoex #ex nex tex #noex #ex nnoex nex ex We have added simple exclusion methods in both cases, which is the reason that we have only 16 (instead of 48) symmetries left in S1; S2 and S 6 . As one can see from the table, we have a nearly linear speedup in number of search steps, and a close to linear speedup in runtime. In S3, the simple exclusion does not apply, which gives rise to 48 symmetries. Furthermore, the optimal conformations and the symmetric ones are very similar, which implies that we have less speedup (since detecting the symmetric solution is harder). Permutation symmetries In the following, we consider nite domain integer problems, where the variables have the domain D N associated. In the following, we denote with Perm(D) the set of all permutations of D, and with Trans(D) the set of all transpositions of D (i.e., those permutations, which exchange just two elements of D). With Perm(D) , we denote the set of all symmetries that permute the values of the variables in X . Denition 6. Let X be the set of solution variables. The set of all value permutations of X is dened as the symmetry set The subset of all transposition symmetries S X is dened analogously. Now we consider search trees that branch over constraints having one single constraints of the forms The interesting part is that if a search tree branches over constraints that have one single Trans(D) -excluding tree is even S X reduced. Thus, we need to exclude only the quadratically many transpositions to exclude the exponentially many permutations in this case. Theorem 3. Let C Pr be a problem which has S X Perm(D) as a symmetry group. Let t be a S X Trans(D) -excluding search tree such that for every branch v, the constraint c used at this branch has one free variable out of X . Then t is S X Perm(D) -reduced. Proof (sketch). Consider a permutation and the corresponding symmetry S . Now consider any node l c(x) we know that ((x)) must be dierent from (x). Now let be the transposition such that let S be the corresponding symmetry. Then the exclusion of all transposition would exclude the solution if we can show that in the node considered above, we have that S 1 S () 2 kv l k. But this can be shown under the assumption that the search tree uses only constraints that have one Graph Coloring As noted above the graph coloring is an example for a problem with value permutations as symmetries. The graph coloring problem is as follows. Given a graph E) and a set of colors. An admissible coloring c of G, is a mapping of the vertices V to the set of colors, satisfying the constraint that for all (v What is the minimal set of colors, where we still can nd an admissible coloring. In this problem colorings are considered to be symmetric, if they are just permuted in colors to each other. We compare two implementations of a solver for this problem. The rst one is a naive implementation that uses a simple rst fail heuristic. This implementation does not have symmetry exclusion. The second implementation with full exclusion of symmetries is just a simple extension of the rst one, where we added our symmetry exclusion mechanism. Note that, due to the above theo- rem, it suces to handle only the set of transpositions instead of handling the group of all permutations to get the full symmetry exclusion. We give the results for some randomly generated problem instances of dier- ent problem sizes. problem size with symmetry excl. without symmetry excl. ratio verts edges colors cloned failed time cloned failed time cloned time 6 Conclusion and Future Work We have introduced the rst method that can exclude arbitrary symmetries in constraint-based search. In contrast to many other methods that can be found in the literature, the method does not restrict the search strategy. The method is based on a declarative description of the symmetries in the form of symmetric constraints, which usually can be obtained easily. We have given completeness and correctness proofs as well as the proof, that the method exclude all considered symmetries. In the case that the full symmetry group is too large, our method can handle arbitrary subsets of the full symmetry group. We have shown for the case of permutation symmetries that it is sucient to exclude the (quadratically many) transposition to exclude all (exponentially many) permutations. The method can be implemented in any (constraint-based) search machinery that handles implications and allows to introduce additional, user-dened constraints during search. This holds for the most modern constraint programming systems. Since the method is very general, we plan to investigate in which other (logic-based) search methods our symmetry exclusion can be used. Fur- thermore, we intend to investigate more general conditions under which a subset of the full symmetry group allows to exclude all symmetries (as it is the case for the permutations). --R Alfonso San Miguel Aguirre. Constraint techniques for solving the protein structure prediction problem. Study of symmetry in constraint satisfaction problems. Tractability through symmetries in propositional calculus. Symmetry breaking predicates for search problems. Thierry Boy de la Tour. Eliminating interchangeable values in constraint satisfaction prob- lems Symmetry breaking during search in constraint programming. The Oz programming model. --TR --CTR Y. C. Law , J. H. Lee, Symmetry Breaking Constraints for Value Symmetries in Constraint Satisfaction, Constraints, v.11 n.2-3, p.221-267, July 2006 Rolf Backofen , Sebastian Will, A Constraint-Based Approach to Fast and Exact Structure Prediction in Three-Dimensional Protein Models, Constraints, v.11 n.1, p.5-30, January 2006
geometric symmetries;permutation symmetries;symmetry excluding search SES;symmetry exclusion;constraint-based search
593397
Design of Extensible Component-Based Groupware.
Tailoring is identified as a key requirement for CSCW applications. One major tailoring mechanism is the extension of an application at run-time to change its behavior.This article shows how synchronous CSCW component-based applications can be designed to be extensible at run-time. We propose to split the act of tailoring into two steps: the design-time customization of new components in visual builder tools and their insertion into the running application. Thus the customization tool is not required to be part of the application.This article presents a new design pattern for extensibility and gives several examples based on that pattern. With the help of the pattern extensible application frameworks can be systematically created from a non-extensible application design. The different possibilities to place insertion points into the application design are discussed with respect to flexibility and ease of deployment. Finally, we present the advantages and limitations of this approach.
Introduction Human beings interact in different situations and their cooperative actions depend on the context. Rather than following a predefined schedule of events, people tend to act spontaneously in creative phases. In general, it is not foreseeable how people work together; therefore, it is not always possible to define in advance which artifacts are adequate to support their cooperative process. A CSCW system that reflects these observations must allow the creation and insertion of new cooperative modules and artifacts. Inserting new functionality into a running application is an act of tailoring. Tailoring is recognized in the CSCW literature as the key requirement for a system to adapt to different cooperative contexts (Malone et al., 1995; Trigg and B-dker, 1994). For Bentley and Dourish (1995), "support for customization is support for innovation". This article focuses on one important subset of tailoring: the ability to insert new functionality into an application and thus to change the behavior of the system. New functionality can be discovered by an extensible application at initialization time. It is harder to design applications that can be extended at run-time. Even harder is the design of c Publishers. Printed in the Netherlands. extensibility at run-time in distributed interactive applications, such as synchronous groupware. This article presents a general design pattern to solve the latter problem. Component-based frameworks are currently being investigated as a means to gain reusability on all layers and to adapt to change. Until recently, software engineering has focused on the development of code, which is reusable and extensible during the design phase. That focus has evolved towards the development of finished modules of code which can be reused and customized by the end-user. System modifications and extensions which were once strictly in the domain of the programmer are now being shifted into the domain of the end-user. This article applies the findings of framework research to the construction of tailorable CSCW systems, which allow the insertion of extensions on demand by the end-user. The code of the extensions is distributed on demand to every participant of the group; thus the extensions do not need to be pre-installed. Our approach provides the possibility to extend an application without terminating an ongoing cooperation. The goal of this article is to demonstrate how component technology supports an efficient way of constructing extensible CSCW applica- tions. Component models allow one to reason on different levels of abstraction depending on the composition level (Stiemerling and Cre- mers, 1998). We focus on the component model Java Beans and its supporting integrated development tools (IDE), which let the user create and customize components visually. The user may assemble components into larger composite components using the visual representation provided by an IDE rather than writing lines of code. Programmers can design the components and accompany them with special customizers to facilitate customization at design-time. In the extreme case, a new component can be assembled by only using drag and drop operations, so that even end-users can accomplish the task of creating new components. By using only the standard and widely used Java and Java Beans technologies for their realization, our concepts become applicable and usable by other groupware developers. If not otherwise stated, we will refer throughout the article to tailoring as the activities by the end-user to modify and extend the application at run-time. In contrast we use the term design-time customization to denote modifications at design-time. The tailoring support for extending a running program is split into two different support-systems, the customization and the insertion support Figure our approach. In the first step, an end-user uses an off-the-shelf visual builder tool to customize a component at design-time. This component is then, in a second step, inserted into the Running CSCW application Running CSCW application A Figure 1. The two-step approach to tailor a CSCW application. running distributed CSCW application. Decoupling the customization tool for the components from the actual CSCW application has the following advantages for the developer: \Gamma The product can be earlier delivered, because the tailoring functionality is not built into the product. \Gamma The developer can save resources, because a proprietary tailoring tool needs not to be developed. General off-the-shelf IDEs are continuously improved by third party vendors. The end-users profit from the decoupling as well. They can use their favorite builder tool for that component model and do not need to accustom to a new tool for every application. We will show that standard visual programming tools can be used efficiently by end-users to perform tailoring of CSCW applications. 2. Relevant Previous Work The work described in this article is based and influenced by the re-search in different domains. Relevant for this work are the publications about tailorability in general and its significance for CSCW; about framework design, component technology and design pattern evolving from object-oriented software engineering; but also about new organizational forms by the business science and coordination theory under the umbrella of "virtual organizations". All approaches have in common that they focus on evolving systems; they differ from their points of view. This paper tries to synthesize some of the findings. 2.1. Tailoring Tailoring is defined as the activity of modifying the appearance and behavior of an application at run-time by the end-user (Malone et al., 1995; M-rch, 1997; Trigg and B-dker, 1994). Tailoring support is normally built into the application, so that the user does not need a separate tool. In contrast to this definition, this article proposes a two-step approach, which uses customization at design-time to gain tailorability at run-time. Before explaining our approach in greater detail, we give a brief overview of tailoring. M-rch (1997) distinguishes three levels of tailor- ing. The levels are classified by the design distance which is experienced by the end-user during tailoring. The first level, customization at run- time, allows to modify the appearance of presentation objects and to change their attributes. The second level allows the integration of new components or commands by composition of existing functionality within the application. The third level allows the extension of an application by adding new code. Generally speaking, with an increasing level the tailoring possibilities for a user increase, but also become more complex. To overcome the design distance, M-rch (1995) proposes to use so called "application units". Application units consist of three parts: a presentation-object, which is the user-interface, a rationale that provides meta-information about the intended use, and the actual implementation. Tailoring beyond the first level involves end-user programming. End-user programming facilities can be offered by the application framework itself, such as in the "radically tailorable" tool for CSCW, Oval (Malone et al., 1995). The end-user can write small scripts, which are interpreted by the application, as in applications of the Microsoft Office suite (Solomon, 1995). But end-user programming can also be separated from the application and done in a general purpose language. Afterwards, the written functionality is inserted into the application at anticipated hooks. Since programming in a general purpose language is regarded as being hard for an end-user, she or he must be supported by high level tools for this task. Visual builder tools like IBM's Visual Age for Java offer an even higher abstraction than scripting languages and are thus usable by end-users (Weinreich, 1997). We will follow this approach. On the level of programming languages, the possibility to reflect and introspect code is viewed as enabling technology to write tailorable software; reflection in component models is used to support self-representation (Stiemerling and Cremers, 1998). Reflection can also be viewed as a design pattern (Buschmann et al., 1996). Dourish proposes a reflective model for collaborative systems (Dourish, 1995) and implemented the toolkit Prospero for CSCW (Dourish, 1996) using this model to express meta-information and to allow the change of the behavior of the underlying toolkit. Component models, such as JavaBeans and DCOM, offer reflection capabilities and meta-data on components (Krieger and Adler, 1998); our approach uses these capabilities to automate interface negotiation and to offer the end-user an easy-to-handle user-interface to tailor components within a visual builder tool. 2.2. Frameworks: Towards extensible applications A framework is a skeleton of cooperating classes that forms a reusable implementation. An application framework defines the overall architecture of the applications that are created by adapting the framework. Framework-based applications are adapted by extending the framework at explicit hooks also known as "hot spots" (Pree, 1994). Frameworks are currently successfully employed for general purpose software units, such as graphical user interfaces, system infrastruc- ture, and middleware integration frameworks; also application domain specific frameworks are emerging. Frameworks are distinguished into white-box and black-box frameworks and Schmidt, 1997). Object-oriented white-box frameworks use inheritance to offer the developer extension facilities. To insert extensions into white-box frameworks the developer must understand the class hierarchy and derive new classes which have to be relinked with the framework. Black-box frameworks use object composition and delegation instead. Black-box frameworks anticipate extensions by defining interfaces and providing hooks to insert new objects. Applications that can be extended at run-time need hooks like black-box frameworks. Unfortunately, designing frameworks - and especially black-box frameworks - is substantially harder than designing an ap- plication. However, the hot spots for a framework can be designed and implemented stepwise by a sequence of generalization transformations (Schmid, 1997). Since applications using a framework must conform to the framework's design and model of collaboration, the framework encourages developers to follow specific design patterns (Johnson, 1997). In the other direction, developers can use design patterns to generalize an object-oriented application into a framework (Schmid, 1995). 2.3. Component Technology In the field of software engineering, component based software development is seen as a major factor to facilitate reuse. Components can be purchased from third party vendors, customized and assembled within a component model. Examples for major component models are Microsoft's Distributed Component Object Model (DCOM) and SUN's component model for Java JavaBeans (JavaSoft, 1996). The component technology is predicted to acquire a significantly increasing importance (Kiely, 1998). Furthermore distributed component platforms are emerging, which allow interaction between components across system boundaries (Krieger and Adler, 1998). A component is an independent "unit of software that encapsulates its design and implementation and offers interfaces to the outside, by which it may be composed with other components to form a larger whole" (D'Souza and Wills, 1998). Frameworks provide a reusable context for components (Johnson, 1997). Components become most powerful within black-box frameworks, where they can be used to extend the hot spots. 2.3.1. Java Beans The examples in this article are implemented using JavaBeans, the component standard for Java. The specification for JavaBeans outlines that "a Java Bean is a reusable software component that can be manipulated visually in a builder tool" (JavaSoft, 1996). Beans are self-descriptive Java classes that follow design patterns that let builder tools or applications introspect a bean. Properties reflect the accessible state of a bean. The Java Beans component model uses an event mechanism to interconnect the beans. A bean sends an event to all beans that have registered their interest in that event. The standard distinguishes two extraordinary states in the life-cycle of a bean: A bean can be manipulated in an IDE at design-time or behave like an ordinary object during run-time. Properties and events can be manipulated within visual builder tools. The JavaBeans standard offers additional associated classes for each bean, which contain meta-information about the bean including special customizers and property editors to support a more intuitive interaction with the developer. The component-based approach together with visual integrated development environments (IDEs) directly support our goal to be able to customize an existing application at design-time and to be able to build new similar applications by reusing the components. Beans with associated customizers allow even non-programmers to customize applications in an intuitive way. The easy grasp is achieved by the use of graphical and form-based editors within the IDEs. 2.4. Design Patterns Design patterns help one to reason about recurring design problems. Object-oriented design patterns describe "communicating objects and classes that are customized to solve a general design problem in a particular context" (Gamma et al., 1994). Patterns abstract from the used programming language and provide a basis for reusable design building blocks: "Design patterns are the micro-architectural elements of frameworks" (Johnson, 1997). Design patterns are surprisingly useful to detect the hot-spots in an application design and to transform it into a domain-specific frame-work design (Schmid, 1995). Actually, the idea of hot spots was first introduced as a meta-pattern for framework design (Pree, 1994). In the domain of CSCW and user-interface design some patterns are well- known, such as the distributed versions of the Model-View-Controller and Presentation-Abstraction-Control patterns (Buschmann et al., 1996). Syri (1997) describes the use of the Mediator pattern to design tailorable cooperation support in CSCW systems. To design CSCW applications that are tailorable by extension, the hot spots must be discovered in the design phase and then implemented. To ease the implementation we will introduce a design pattern which can be used to insert those hooks into the application. The pattern focuses on the ability to insert new code at run-time that conforms to an interface. By applying this pattern, one thus designs a black-box framework for a specific CSCW problem. 2.5. Virtual Organizations This work is also influenced by recent publications about virtual organization (Mowshowitz, 1997; Turoff, 1997). The idea of virtual organization stems from virtual constructs, such as virtual memory and circuit routing, and generalizes their concepts toward an integrating theory. One common concept in the virtual constructs is that the mapping between an abstract requirement and a possible concrete satisfier is dynamic. The mapping has to adapt as well to evolving requirements as to changing satisfiers. The ability of virtual organized systems to dynamically adapt to environmental changes led us to think about how the dynamic exchange of software components could enhance CSCW systems. In this context, this paper provides a technical basis to insert new satisfiers into a running groupware application when a new requirement arises. Changes to requirements for CSCW applications can be a result of evolving cooperative work patterns, for example when users become more familiar with a CSCW product or the context and goal of a work group changes (Mark et al., 1997). 3. Enabling Technologies for Extending CSCW Applications This section introduces a design pattern, which is used to insert hot spots in the design of applications. Since CSCW applications are inherently distributed, the pattern is accompanied with components that allow the distribution of arbitrary events to a group. By using the event mechanism and encapsulating code within an event, we place an event receiver in the pattern to allow the simultaneous extension of synchronous CSCW applications at run-time. Finally, we investigate the applicability of inserting code at run-time. 3.1. Design Pattern for Extensibility In a component model, applications are developed by interconnecting and customizing components. The components themselves are composed of other, smaller components. The design pattern for extensi- bility, which will be introduced here, can be encapsulated into one component. The Extensibility pattern 1 is intended to be used to provide a default behavior, which can be changed at run-time. To change the behavior a new class can be inserted at a hook, which can either add new functionality or replace an existing class. The application sees only the specified behavior of a Proxy class. The structural representation of a pattern is given by the relationship between the used classes. Figure 2 shows the structure of the Extensibility pattern in the UML notation 2 . This pattern consists of a Proxy, which extends the interface of a Subject that may be inserted at run-time. 3 Inside the Proxy exists a Creator, which is responsible We follow the convention to use an initial upper-case letter to name patterns, to use bold names for classes and to use italic names for beans and events. Fowler (1997) provides a good overview on the UML notation. 3 In this pattern, the Proxy extends the Subject interface to be conform with the Proxy pattern. However, for the framework developer it is only important that Figure 2. Structure of the Extensibility pattern. to create a new object of an arbitrary class Real Subject conforming with the interface Subject. Actually this pattern is a combination of the Proxy and the Factory Method patterns from Gamma et al. (1994). Figure 3 shows the interaction between the objects. At initialization time, the Creator object passes a reference to a default Real Subject to the Proxy. Any event that the Proxy receives is delegated to the default Real Subject. When the Creator receives an event (how that happens will be discussed soon) to create a new Real Subject it instantiates the respective class and sets the reference in the Proxy to the newly created object. The Proxy now forwards all subsequent events to this object, unless the Creator changes the reference to a Real Subject again. A slight variation of the pattern allows to add instances of new classes instead of replacing the old objects. This can be easily accomplished by letting the Proxy store a set of all Real Subjects. All incoming events are then forwarded to all instantiated Real Sub- jects. This variation is useful if new functionality is added, which is independent in the application logic from the already existing objects. 3.2. Remote Events In an event based component model, events are the means to communicate state changes between components. The event mechanism the definition of the Proxy remains stable, while the developer of the pluggable components (the Real Subjects) relies on a stable definition of Subject. The Proxy class does not need to extend the same interface and may act as Adapter or Bridge. Proxy, Adapter, and Bridge are described in Gamma et al. (1994). C rea tor in it in it ins tan tia te de fau lt Sub je c t hand leEven hand leEven t ins tan tia te new Sub jec t hand leEven hand leEven t c rea te new Sub jec t Figure 3. Interaction diagram for the Extensibility pattern. follows the publisher-subscriber pattern (Buschmann et al., 1996). The Java Beans component model uses such an event mechanism. We will concentrate here on Java Beans, because it is the component model we have chosen for our implementation. The design of group communication components does not rely on this particular component model, since event mechanisms are a common property of component models. Since the Java Beans component model defines only the interaction between beans in the same virtual machine, we developed group communication beans, which act as access points for the distribution of events over process barriers (figure 4). The group communication beans follow the publisher-subscriber pattern for a distributed platform. The group communication beans expose the event model to the developer for remote event communication. Two types of beans are necessary: The GroupSender forwards events to all GroupReceivers , which are configured with the same group name. The group name is a property of the beans and can be easily set within visual builder tools for beans, and the events can be visually connected to and from these beans. Both beans can be specialized for any event by simple object-oriented subclassing and implementing the register and handler methods for that event type. Thus the group communication beans form a white-box framework for distributed event communication. Although the definition of new events is considered as a programming activity, which goes beyond the usual GS group communication GR GR GR group name group name event event event event components backbone Figure 4. Group communication beans: The GroupSender (GS) distributes an event to all subscribed GroupReceivers (GR). capabilities of an end-user, the implementation is automated within most IDEs, thus nearly no source coding is required by the developer. 3.3. Distributing and Inserting Components Readers who ask themselves how the Creator in the Extensibility pattern is triggered, will get their answers here. The Creator encapsulates a GroupReceiver that subscribes to a group on which events may arrive that carry the classes to be instantiated. Since Java provides a platform-independent byte-code, we can directly associate the classes with the events. In other implementations, objects would have to be called remotely (by using CORBA for instance). We decided to actually distribute code, since it is a more general solution than calling remote objects. For example, a remote object would have difficulties to access system dependent resources to show a new graphical user interface. New Class Event insert new class choose new class GS GR New Class Event Chooser Loader Figure 5. Distribution of a NewClass event. The distribution of a new component is handled by our design as the distribution of a NewClass event by the beans for group communication (figure 5). The bean, which acts as a Chooser , selects the class, which should be inserted in the distributed application. Often a Chooser is embedded in the user-interface to let the user decide, which class should be inserted. Eventually, the Chooser fires a NewClass event. The event is simply passed by the Chooser to a bean that is derived from a GroupSender , which publishes it to the configured group. The event is then received by all beans that extend a GroupReceiver for this event type and are subscribed on that group. The GroupReceiver passes the event to a Loader , which instantiates the class. The resulting object can then be used by the Creator to replace or add a new Real Subject. The combination of the Extensibility pattern with the group communication beans can be used to extend well specified hot spots in distributed applications; the specification is the interface Subject. If a hot spot defines a lot of methods, each component has to implement these methods, before it could be used to extend the hot spot. Some- times, however, it is not feasible to be constrained by an interface. In the case of truly independent components, such as applications, it would be needed to write an adapter (Gamma et al., 1994) to insert them. On the other side, even such components may use some of the available information by the loading component. Instead of using the static information provided by the interface, a variation of our pattern uses the reflection mechanisms of Java and Java Beans to connect to the available hooks. Figure 6 illustrates this concept. A NewClass event arrives at the loader (a). Upon arriving, the loader loads the class and instantiates it (b). Since the loader does not know at this time the features of the arrived bean, it uses introspection to discover the events, which can be fired by the new bean. For the events it is interested in, the new class event (a) (a) (b) (b) (c) (c) event listener event source loader loader loader Figure 6. A loader receives a new class (a) and instantiates it (b); then the loader and the new object can register mutually (c). loading bean adds its interest by calling the discovered registration methods (c). Now, the loading bean can receive events from the new bean. If the loading bean provides itself events and has discovered by introspection that the new bean implements the appropriate method to connect itself, it invokes that method. Then the loaded bean uses the same mechanism to subscribe itself to the events it is interested in. The NewClass event may additionally carry the name of a start method. If the new class is not a bean, no events are connected, but the start method will still be called. Thus it is possible to pass arbitrary Java programs and start them remotely. The newly loaded code can interact with the loading application by means of two mechanisms: by mutual registration for the provided events that are discovered during initialization and by the presented group communication beans. The latter are also used to communicate with other remote applications. 3.4. Applicability Extensibility of CSCW applications can be introduced on various levels of granularity, varying from the one extreme, where only new applications can be started, to the other extreme, where every component may be extended. The place and number of hot spots in the design determine the extensibility of the application framework. But, the number of hot spots does not only worsen the performance of the application, but it increases also the necessary effort of maintenance. Figure 7 shows how the level of extensibility relates with the granularity of components that can be inserted and the understandability and maintainability for the end-user. The MBone tools (Eriksson, 1994) may serve as an example for very small but successful extensibility: the user can click in the session directory (sdr) on a session, which starts the needed tools to join the audio and video session. The tools are stand-alone applications, which are started in a different process. Medium extensibility is granted by domain specific frameworks with Figure 7. Trade-off between application extensibility, component granularity, and understandability for the end-user. some hot spots; TeamWave (Roseman and Greenberg, 1997) is a groupware application, which uses a custom made component model on top of GroupKit (Roseman and Greenberg, 1996) to offer extensibility and tailoring support. The highest level of extensibility would be the usage of the Extensibility pattern for every component in a system. If extensibility is only provided by means of starting applications in new processes, the original and the new application must use a protocol to exchange data, which is normally different from local interaction. Therefore, inserting components into a running application has the advantage that they can be integrated seamlessly; the new components become part of the application. The components can interact locally and use same the interaction protocol of the component model. Our experiments with our pattern and component based CSCW applications suggest that most extensions of groupware applications happen at anticipated places. If the application uses design patterns, some hot spots can be found during the design phase (Schmid, 1997). However, it remains an art rather than pure engineering to design extensible applications. We will give some examples in the next sec- tion, how extensibility can be designed and implemented in CSCW applications. 4. Examples This section gives some examples, how the Extensibility pattern is used to design extensible CSCW applications. The first example presents a minimal CSCW component, which is used to distribute and start other cooperative components. We use a chat component as example to demonstrate the application of the Extensibility pattern. The insertion of a voting component during a chat session highlights the use of the Extensibility pattern to support unforeseen cooperation modes. Finally, we summarize our experience of using the Extensibility pattern in tele-teaching components. 4.1. Design of a minimally extensible CSCW application An example for a minimally extensible CSCW application is a loader that offers the functionality to distribute and insert coarse grained components, which are actually CSCW applications themselves. When the user selects a new component for insertion, the code is distributed to all participants of the group and started within their instances of the loader. Figure 8. The loader application in a visual IDE. Figure 8 shows the composition of the loader within a visual IDE 4 . The user interface consists of two beans to enter the participant and the group name and a button to insert a new component. When the user presses the button, a file dialog pops up, which lets the user select a component. After choosing the component an event is passed to a non-visual Controller bean, which generates a NewClass event and passes it to a GroupSender , which is configured with the group name. All loaders of the group members will eventually receive the NewClass event and start the associated component by the Creator . The Creator for the loader is configured to add every received component and to use the reflection capabilities to register for available events. The loaded component can query the properties of the loader via reflection - in this case it finds the participant and group name. In the presented form, the loader supports the insertion of symmetric CSCW applications, i.e. applications that are executed at each partic- ipant. For example, the loader can be used to insert the components of the next examples: a chat and a voting component. We have also developed a loader component for asymmetric groupware, which supports the local insertion of a server component, and distributes clients for this component to all other participants. 4.2. Design for functional extensions A well-known example for a synchronous CSCW application is a chat. A chat allows the exchange of textual messages between all members 4 This and all subsequent examples are built with IBM's Visual Age for Java. A puzzle piece denotes a non-visual bean, a puzzle piece in brackets a variable, an arrow a connection between an event and a method, and a dotted line a connection between two properties. Figure 9. The design of an extensible input component (right) for a chat tool (left). The this variable gives access to the methods and variables of the defining bean (here: the input component). of a group. This example will focus on the design of an extensible chat and present a component that can be inserted at run-time to support a simple floor control policy. Figure 9 shows a running chat application, and the component composition at design-time for the input part of the chat. A new message is distributed by a Chat event to all participants; the output part of the chat component eventually receives the event and shows it to the user. The reaction on user input is performed within the bean ChatInputControlProxy , which has access to the input field and some environment properties. Whenever the proxy generates a new Chat event, it is distributed by a GroupSender for this event (ChatGS) to all participants. In this example, the user-interface additionally offers a button to insert a new component into the running application. Figure 10. The design of the ChatInputControlProxy. Figure 11. User-interface for insertion of a new component and its implementation. Figure shows the internals of the proxy, which allows the replacement of the default strategy. The BeanCreator can receive new beans that implement the interface ChatInputControlI ; it takes as default the component ChatInputControl . The input field and the current instance of the input control are associated with a variable of the type ChatIn- putControlI . Depending on the actual input control bean, a Chat event is fired to the proxy, which forwards it to the GroupSender . This example implements a very simple mechanism to plug a new component into the running system (see figure 11). When the user clicks on the "Insert component" button of the chat application, a dialog box pops up and the user selects the hot spot to extend. Then the user chooses from a list of available components. The actual design and implementation for the selection uses the same components as the simple loader, which was previously described. As will be discussed later, a more sophisticated mechanism should be used in real-world applications. To add a floor control mechanism, the default implementation of ChatInputControl (figure 12, left) can be replaced by ChatInputFloor- Control (figure 12, right) during run-time. The new component displays an additional simple user-interface to request the token for input; the input field of the chat bean is only enabled, if the user has the token. It also uses GroupSenders to request and release 5 a token. The newly 5 This implementation implicitly releases the token after a user has sent a message. The server for the floor control is not shown here. Figure 12. The ChatInputControlProxy is configured by default with the bean Chat- InputControl (left); it can be replaced by ChatInputFloorControl (right) to support a token based floor control policy. inserted component interacts seamlessly with the existing components, since it implements the same interface ChatInputControlI . The design of the chat components follows a simplified Model-View- Controller pattern (Buschmann et al., 1996). To insert components, which provides new behavior, we designed the controller of the chat input component to be exchangeable; the design uses the presented Extensibility pattern. The other chat components are designed in a similar way. Another hot spot is designed in the chat output component; a possible extension would be to add a component to write a log file of the discussion. The chat example has shown the applicability of the Extensibility pattern to change the component's behavior at specially defined hot spots. It is thus classified in the medium level of granularity. 4.3. Design of a second application for insertion The loader can be used to start more than one cooperative tool for all group members. For example, the chat tool is inserted for a discussion in a meeting with remote participants. After a while, a decision must be made about the discussed topics. The chair decides to create a list of the topics, and each participant has to vote for one item on the list. So, the chair uses an IDE to customize a voting component to be inserted and distributed using the loader. The voting component is shown to each participant; after a participant has submitted his vote, a separate frame shows all arriving votes from the others. The design-time customization of a vote component is a very easy task: A question component is dropped on the vote panel within a visual Figure 13. A voting component (left) and the associated customizer (right). IDE (figure 13). The vote panel has an associated customizer to add new questions, to manipulate them, and to provide different language features. The customizer offers the user a graphical interface to hide all details of programming. The end-user only performs drag-and-drop operations and fills in text fields. The customizer constructs a new voting component with this information, which can then be inserted by the loader to be distributed to all participants. This example has shown the applicability of the Extensibility pattern for coarse grained components to support the insertion of new cooperation forms. It also has validated the approach to use off-the- shelf visual builder tools to let the end user build a new component by design-time customizing existing beans, which are then distributed and inserted into the running CSCW system. 4.4. Other examples We have redesigned some of our earlier developed remote education components to offer extensibility. As an example, we placed the Extensibility pattern in remote tutoring components (Hummes et al., 1998b) to allow the insertion of arbitrary components supporting cooperation among the students and tutors. The tutoring components allow students to contact a tutor, if they want assistance in a remote laboratory course. The tutor gives peer-to-peer advice by using cooperation beans. In the original implementation the components for cooperation could be changed only at design-time; the new implementation can use several cooperation forms by inserting them at run-time. The tutor can now also distribute questionnaires (Hummes et al., 1998a) to all students at the end of a laboratory course to monitor their learning progress. The tutor has prepared the questionnaire during the laboratory course based on the issues that have been discussed with the students. The creation of such a questionnaire is highly supported by customizers within a visual IDE. The presented customizer for the vote panel is actually a reused component for multiple choice questions from this tele-exam framework. 5. Discussion The examples have shown the applicability of the Extensibility pattern within component based CSCW applications. By using the pattern one actually designs domain specific application frameworks. These application frameworks can be extended at run-time by inserting new components. The new components can be created by the end-user outside the application within visual IDEs. The examples have used the Extensibility pattern to insert coarse and medium grained components. The placement of the hot spots with the pattern in the examples is based on the anticipation of possible extensions. This leads to the question whether a rule can be given where the hot spots should be located. The main problem is that a conflict exists between the level of extensibility and the level of understandability (cfg. section 3.4). If each component was made extensible, the design and implementation would become unnecessary complex. Even, if the performance affected by the added complexity could be improved (for example with the Flyweight pattern (Gamma et al., 1994)), the maintainability criteria still limits the amount of hot spots. On the other hand, too few hot spots limits the extensibility of the application. So, a compromise must be found depending on the domain of the application. We found that the design of cooperation offers a good starting-point to insert hot spots in CSCW applications. Components that are triggered by user actions and perform operations depending on these actions are candidates to be extended. If the design of an application that uses design patterns, the location of potential hot spots can be derived from the design (Schmid, 1997). In the case of CSCW appli- cations, patterns used for cooperation must be examined. In the often used Model-View-Controller pattern, a potential hot spot for extension in each application is located in the Controller, while the View would be a candidate for being extended only locally. Cooperations that can use different strategies, can change their strategies by placing the Extensibility pattern within the Strategy pattern (Gamma et al., 1994). A good example for adding a new strategy component would be a new algorithm for video encoding and decoding in conference systems. The Mediator pattern (Gamma et al., 1994) can be used to design tailorable 1CSCW systems by attaching cooperation enablers (Syri, 1997) to cooperative artifacts. By placing the Extensibility pattern within the Mediator new enablers could be introduced in the running system. While we have presented here a non exhaustive list of potential loca- tions, where hot spots could be useful, it is still up to the groupware designer to decide, where hot spots will eventually be placed. Her or his analysis will be oriented on the domain of the application. New components can be inserted on the demand of other components or on the demand of the end-user. In the latter case, the end-user must be supported by a user-interface to select the appropriate hot-spot and component to obtain a certain behavior by his extension. We have used a simple file chooser in our examples. A more sophisticated approach would present the user the potential plug-points and a list of available components that are available to extend each one. By introspecting the selected component, such a list can be created automatically. Additionally, the user should also get a description of the intent, effects and possible side-effects for each component. The presented implementation to insert components at run-time uses code distribution. To inform remote applications to insert a new component, we used group communication beans that can distribute arbitrary events. The needed information about the new components is encapsulated in an event. So, the implementation is coherent with the Java Beans event model. Thus it is supported by visual builder tools for Java Beans. The distribution of code has the advantage that components have access to the local system properties. Thus user-interface components can also be distributed. Another advantage lays in increased performance compared with remote object communication if the inserted component is often used. The biggest advantage in a cooperative environment is that the component which should be inserted in the running application needs not be installed at the remote machines before the application is started. The operation of loading and instantiating classes via the network opens severe security risks. Since Java is a network language, these risks are well-known and methods for protection exist. Java code can be signed. A signature authenticates the creator of the code. If code is manipulated after signing, this can be detected. Although signed code allows one to only accept code by trustworthy sources, the problem of who to trust remains. In a cooperative environment, this question is hard to answer. Even if all persons that are allowed to distribute new code are trustworthy, failures in the distributed code can cause damage (Zhang, 1997). The problem can be partly solved by giving explicit rights for customizing code (Stiemerling and Cremers, 1998). Another barrier can be inserted by granting new classes only the rights they need to function. If, however, a class claims to need full rights and is created by a person of full trust, the problem remains. This problem can not be generally solved. 6. Conclusion This article has focused on the insertion of new components into running synchronous CSCW applications to tailor their behavior. We have proposed to split the act of tailoring into the steps of the design-time customization of new components within visual IDEs and their insertion into the running application. This decoupling leads to a shorter development cycle of applications. Furthermore, the end-user needs only to accustom to one IDE to tailor different applications. When IDEs will be delivered as components, our approach can be taken to extend CSCW applications with those pluggable builder tools. We have presented a design pattern which is focused on modeling insertion points at hot spots in a general way. Since CSCW applications are inherently distributed, we have developed components that are used to distribute arbitrary events across process boundaries to a configurable group of receivers. These group communication components are used by our examples for all remote event communication. Extensions are implemented as Java beans and distributed through remote events. They are then automatically inserted at the provisioned hot spots. Once inserted, the new components are seamlessly integrated within the running application. Independent coarse grained components that function also without information about their environment can be inserted without conforming to a predefined interface. Nevertheless they can query their environment via reflection to register for events or to read and write properties. Thus arbitrary applications can be distributed and started remotely. The presented examples have shown the use of the Extensibility pattern to create hot spots within component based applications. This article has discussed the tension between extensibility and understandability in the design. Increasing the extensibility increases also the complexity of the design and thus decreases the understandability. This leads to the conclusion that a design is not reasonable where all components are extensible or exchangeable during run-time. So, we have revisited the design patterns, which were used in the examples, to find the locations where hot spots have been inserted. The located places have been compared with some design patterns in the literature. Thus we have shown, where the Extensibility pattern is most useful in the design of CSCW applications. However, it remains still a task for the application designers to identify the hot spots from their ex- pertise. Once potential insertion points are recognized, the developer can uniformly design the hot spots using the introduced Extensibility pattern. One problem, which should be addressed by further work, is how the extensibility can be presented to the end-user. The presentation should include the hot spots of an application and their possible extensions. Such a presentation must find means for an intuitive graphical user-interface to insert components at the right places. 7. Acknowledgments We want to thank David Turner for proof-reading earlier versions of this article. Arnd Kohrs has provided valuable Java hints in various discussions and without his help most of our components for remote tutoring would never have been implemented. Last, but not least, we want to acknowledge the detailed comments by the anonymous reviewers. The described work is part of the ACOST research project, which is funded by the research institute CNET Lannion of France Telecom. --R Communications of the ACM 40(10) UML distilled Design Patterns - Elements of Reusable Object-Oriented Software Kluwer Academic Publishers. http://www. Developing applications with Microsoft Office: strategies for de- signing --TR --CTR Yang , Du Li, Separating data and control: support for adaptable consistency protocols in collaborative systems, Proceedings of the 2004 ACM conference on Computer supported cooperative work, November 06-10, 2004, Chicago, Illinois, USA Jrgen P. Bansler , Erling Havn, Sensemaking in Technology-Use Mediation: Adapting Groupware Technology in Organizations, Computer Supported Cooperative Work, v.15 n.1, p.55-91, February 2006 Yang , Du Li, Supporting Adaptable Consistency Control in Structured Collaborative Workspaces, Computer Supported Cooperative Work, v.14 n.5, p.469-503, October 2005 P. Barthelmess , C. A. Ellis, The Neem Platform: An Evolvable Framework for Perceptual Collaborative Applications, Journal of Intelligent Information Systems, v.25 n.2, p.207-240, September 2005
java beans;tailoring;customization;design pattern;extensibility
593444
Mathematical Programming in Data Mining.
Mathematical programming approaches to three fundamental problems will be described: feature selection, clustering and robust representation. The feature selection problem considered is that of discriminating between two sets while recognizing irrelevant and redundant features and suppressing them. This creates a lean model that often generalizes better to new unseen data. Computational results on real data confirm improved generalization of leaner models. Clustering is exemplified by the unsupervised learning of patterns and clusters that may exist in a given database and is a useful tool for knowledge discovery in databases (KDD). A mathematical programming formulation of this problem is proposed that is theoretically justifiable and computationally implementable in a finite number of steps. A resulting k-Median Algorithm is utilized to discover very useful survival curves for breast cancer patients from a medical database. Robust representation is concerned with minimizing trained model degradation when applied to new problems. A novel approach is proposed that purposely tolerates a small error in the training process in order to avoid overfitting data that may contain errors. Examples of applications of these concepts are given.
Introduction Mathematical programming, that is optimization subject to constraints, is a broad discipline that has been applied to a great variety of theoretical and applied problems such as operations research [29, 54], network problems [60, 53], game theory and economics [71, 35], engineering mechanics [57, 37] and more recently to machine learning [3, 61, 23, 48, 46]. In this paper we describe three recent mathematical-programming-based developments that are relevant to data mining: feature selection [45, 10], clustering [11] and robust representation [67]. We note at the outset that we do not plan to survey either the fields of data mining or mathematical programming, but rather highlight some recent and highly effective applications of the latter to the former. We will, however, point out other approaches that are mostly not based on mathematical programming. The fundamental nonlinear programming problem [6, 44] consists of minimizing an objective function subject to inequality and equality constraints and is typically written as follows f(x) subject to g(x) - 0; where x is an n-dimensional vector of real variables, f is a real-valued function of x, g and h are finite dimensional vector functions of x. If all the functions f , g and h are linear then the problem simplifies to a linear program [15, 52, 69] which is the classical problem of mathematical Mathematical Programming Technical Report 96-05, August 1996 - Revised November 1996 & March 1997. This material is based on research supported by National Science Foundation Grant CCR-9322479. y Computer Sciences Department, University of Wisconsin, 1210 West Dayton Street, Madison, WI 53706, email: programming. If x is two-dimensional, a linear program can be thought of as the problem of finding a lowest point (not necessarily unique) on a tilted plane surrounded by a piecewise-linear fence. Extremely efficient algorithms exist for the solution of linear programs. Thus reducing a problem to a single or finite sequence of linear programs is tantamount to solving the problem. Another reason for emphasizing mathematical programming in this work is the very broad applicability of the optimization-under-constraints paradigm. A great variety of problems from many fields can be formulated and effectively solved as mathematical programs. According to the great eighteenth century mathematician Leonhard Euler: "Nothing happens in the universe that does not have a sense of either certain maximum or minimum" [68, p 1]. From the point of view of applicability to large-scale data mining problems, the proposed algorithms employ either linear programming (Sections 2 and 3) which is polynomial-time-solvable [36, 69], or convex quadratic programming (Section 4) which is also polynomial-time-solvable [69]. Extremely fast linear and quadratic programming codes [14] that are capable of solving linear programs with millions of variables [8, 40] and very large quadratic programs, make the proposed algorithms easily scalable and effective for solving a wide range of problems. One limitation however is that the problem features must be real numbers or easily mapped into real numbers. If some of the features are discrete and can be represented as integers, then the techniques of integer programming [24, 55, 14] can be employed. Integer programming approaches have been applied for example to clustering problems [64, 1], but will not be described here, principally because the combinatorial approach is fundamentally different than the analytical approach of optimization with real variables. Stochastic optimization methods based on simulated annealing have also been used in problems of inductive concept learning [50]. The problems considered in this paper are: 1. Feature Selection The feature selection problem treated is that of discriminating between two finite point sets in n-dimensional feature space by a separating plane that utilizes as few of the features as possible. The problem is formulated as a mathematical program with a parametric objective function and linear constraints [10]. A step function that appears in the objective function is approximated by a concave exponential on the nonnegative real line instead of the conventional sigmoid function of neural networks [28]. This leads to a very fast iterative linear-programming-based algorithm for solving the problem that terminates in a finite number of steps. On the Wisconsin Prognosis Breast Cancer (WPBC) [72, 51] database the proposed algorithm reduced cross-validation error on a cancer prognosis database by while reducing problem features from to 4. 2. Clustering The clustering problem considered in this paper is that of assigning m points in the n-dimensional real space R n to k clusters. The problem is formulated as that of determining k centers in R n such that the sum of distances of each point to the nearest center is minimized. Once the cluster centers are determined by a training set, a new point is assigned to the cluster with the nearest cluster center. If a polyhedral distance (such as the 1-norm distance) is used, the problem can be formulated as that of minimizing a piecewise-linear concave function on a polyhedral set which is shown to be equivalent to a bilinear program: minimizing the product of two linear functions on a set determined by satisfying a system of linear inequalities [11]. Although a bilinear program is a nonconvex optimization problem (i.e. minimizing a function that is not valley-like), a fast finite k-Median Algorithm consisting of solving few linear programs in closed form leads to a stationary point. Computational testing of this algorithm as a KDD tool [18] has been quite encouraging. On the Wisconsin Prognosis Breast Cancer Database (WPBC), distinct and clinically important survival curves were discovered from the database by the k-Median Algorithm, whereas the traditional k-Mean Algorithm [32, 64], which uses the square of the 2-norm distance, thus emphasizing outliers, failed to obtain such distinct survival curves for the same database. On four other publicly available databases each of the k-Median and k-Mean Algorithms did best on two of the databases. 3. Robust Representation This problem deals with modeling a system of relations within a database in a manner that preserves, to the extent possible, the validity of the representation when the data on which the model is based changes. This problem is closely related to the generalization problem of machine learning of how to train a system on a given training set so as to improve generalization on a new unseen testing set [39, 63, 73]. We use here a simple linear model [67] and will show that if a sufficiently small error - is purposely tolerated in constructing the model, then for a broad class of perturbations the model will be a more accurate representation than one obtained by a conventional zero error tolerance. A simple example demonstrates this result. 1.1 Notation We summarize below notation and background material used in this paper. ffl All vectors will be column vectors unless transposed to a row vector by a superscript T . ffl For a vector x in the n-dimensional real space R n , jxj will denote a vector of absolute values of components x of x. ffl The base of the natural logarithm will be denoted by ", and for y \Gammay will denote a vector in R m with component " \Gammay ffl For x 1, the norm kxk p will denote the p-norm, that is ( . ffl The notation A 2 R m\Thetan will signify a real m \Theta n matrix. For such a matrix, A T will denote the transpose, A i will denote row i and A I will denote those rows A i such that i 2 I ae for a given subset I of ffl A vector of ones in a real space of arbitrary dimension will be denoted by e. A vector of zeros in a real space of arbitrary dimension will be denoted by 0. ffl If x and y are vectors in R n , the notation minfx; yg will denote the vector in R n of componentwise minimum of x i and y i , ffl The notation arg min f(x) will denote the set of minimizers of f(x) on the set S. Similarly arg vertex min f(x) will denote the set of vertex minimizers of f(x) on a polyhedral set S. ffl A polyhedral set S in R n is the the intersection of a finite number of closed halfspaces in R n . ffl A vertex of a polyhedral set S is a boundary point of S that lies on the intersection of n linearly independent planes constituting the boundary. Typically is a vertex of S if for some subset I of A I I and A I 2 R n\Thetan is nonsingular. ffl A separating plane, with respect to two given point sets A and B in R n , is a plane that attempts to separate R n into two halfspaces such that each open halfspace contains points mostly of A or B: ffl Alternatively, a separating plane can be interpreted as a classical perceptron [62, 43] with a threshold determined by the distance of the plane to the origin, and the incoming arc weights of the perceptron determined by the components of the normal vector to the plane. Feature Selection Feature selection (or extraction) attempts to use the simplest model to describe the essence of a phenomenon. Hence it can be considered as an application of Occam's ``law of parsimony'', also known as Occam's Razor [65, 9], which states: "What can be done with fewer [assumptions] is done in vain with more". There are statistical [22], machine learning [39, 33] as well as mathematical programming [12, 45, 10] approaches to the feature selection problem. In this work we shall deal principally with the latter because of the novelty of the approach and it effectiveness. The problem that we shall address is the binary classification problem, i.e. the problem of discriminating between two given point sets A and B in the n-dimensional real space R n by using as few of the n dimensions of the space as possible. For example in the medical application described at the end of this section, we attempt to discriminate, with as few of the features as possible, between breast cancer patients that had a recurrence of the disease within two years of diagnosis and those who had not. The model that we shall adopt will be that of a perceptron or linear threshold unit (LTU) that employs as few of the features of the given problem as possible. Extensions to more complex models leading to compact and accurate decision trees have also been made [12]. Geometrically our approach corresponds to constructing a plane in R n defined by with normal w 2 R n and distance jflj to the origin, while suppressing as many of the components of w as possible. In addition, the set A must lie, to the extent possible, in the open halfspace and the set B in the open halfspace This corresponds to an LTU with a threshold fl and and incoming arc weights w 2 R n , with as many of the weights as possible set to zero. If we represent the set A by the matrix A 2 R m\Thetan and the set B by the matrix B 2 R k\Thetan , then the problem is to find fl 2 R and w 2 R n , with as many components equal to zero as possible, such that the following inequalities are satisfied in some best sense: where e is a vector of ones. Because linear programming cannot handle strict inequality constraints we rescale (5) as follows. We divide the variables (w; fl) by the positive quantity min and call the rescaled variables (with notational economy in mind and a slight abuse of notation) (w; fl) again, then (5) is equivalent to: e: (6) Since these inequalities may not have a solution in general, one resorts to satisfying them in some best approximate sense by minimizing an average sum of their violations. This leads to the following Robust Linear Programming formulation [4]: RLP min w;fl;y;z Robustness here refers to the fact that the useless null vector (w = 0) is naturally excluded as a solution of (7), which is not the case in other linear programming formulations of this problem [41, 66, 26, 25]. Note that because of the constraints of the problem, the variables y and z will satisfy the following conditions: Hence minimizing e T will force the satisfaction in some best sense of (6), and equivalently (5), by minimizing the average violations of (6):m which will be zero if and only if (6), or equivalently (5), is exactly satisfied. The linear programming formulation (7) which has a number of natural theoretical properties including robustness is also very effective computationally [4, 49]. However it does not address the problem of suppressing irrelevant features. In order suppress such features, the objective function of (7), which merely measures the average sum of the violations of the inequalities (6), is modified so as to also suppress as many of the components of the weight vector w as possible. This is achieved by weighting the original objective of (7) by weighting by - an exponential function approximation of the absolute value v of the weight vector w. This exponential function, e T approximates the 1-norm of a step function of v and leads to the following mathematical program with a concave objective function and linear constraints:(Feature Selection Concave) the problem degenerates to the robust (that is w 6= linear program of (7) which obtains a plane P (2) that separates the sets A and B in an optimal fashion without regard to feature suppression. When - ? 0, then in addition to the objective of separating A and B, we attempt to suppress as many of the components of w as possible by minimizing an exponential smoothing of the step function on the nonnegative real line for each where " is the base of the natural logarithms, and v is the absolute value of w. In most our applications a value of sufficient to make the exponential a good approximation of the step function to force suppression of unnecessary components of w. As described below in Algorithm 2.1 the parameter - is chosen to give the best cross-validated error. For small values of - it can be shown theoretically [47] that the minimization problem (8) picks that solution of the Robust Linear Program (7) that minimizes the exponential term of (8), and hence solves the RLP (7) while suppressing redundant components of w. Because the objective function of (8) is a concave function bounded below by zero on the nonempty polyhedral set of (8), it follows that it has a vertex solution if the feasible region does not contain lines extending to infinity in both directions [59, Corollaries 32.3.3, 32.3.4]. (Excluding such lines can be readily accomplished by a simple transformation of the variables (w; fl) into the nonnegative variables (w 1 using the standard transformation . For the sake of simplicity and because it is not needed computationally, we shall forgo this transformation here.) A fast, finitely-terminating successive linear programming algorithm has been proposed for solving this problem [10] as follows. Algorithm 2.1 Successive Linearization Algorithm (SLA) for FSV (8). Choose - 2 [0; 1). Start with a random (w Having (w determine the next iterate by solving the linear program: (w Stop when (v Comment: The parameter ff was set to 5. The parameter - was chosen in the set f0, 0.05, with the desired - being the one achieving the best cross-validated separation. It has been shown [45, Theorem 4.2] that this algorithm terminates in a finite number of steps, typically five or six, at a global solution or a stationary point satisfying a necessary optimality condition. This algorithm was tested on the 32-feature Wisconsin Prognostic Breast Cancer (WPBC) database [72, 51] which was collected from 28 patients for which cancer recurred within two years, and 119 patients for which cancer did not recur within two years. Thus in the terminology of our formulation 28 and 118. For this problem the separating plane obtained by the Successive Linearization Algorithm 2.1 with used only 4 features out of 32, while increasing tenfold cross-validation correctness by 35.4% [10]. If other values of the parameter - are used in the Successive Linearization Algorithm 2.1, then the number of features that determine the separating plane will vary between 1 and 32. The effect of using a different number of features is shown in Figure 1 which shows a plot of tenfold cross-validation correctness corresponding to the number of features used. As just indicated, Figure 1 shows that the best tenfold correctness occurs at four features. Number of Nonzero Elements of w Test Correctness Figure 1: Feature Selection in the Prognosis Problem: Tenfold cross-validation correctness versus number of features selected by the Successive Linearization Algorithm 2.1 3 Clustering via Mathematical Programming The unsupervised assignment of elements of a given set into groups or clusters of like points, is the objective of cluster analysis. There are many approaches to this problem, including statistical [32], machine learning [20], integer and mathematical programming approaches [64, 1, 58, 11]. We shall describe here the recent approach of [11] that utilizes a fast bilinear programming minimizing the product of two linear functions on a set defined by linear inequalities. A principal motivation behind our mathematical programming approach is a precise and concise statement of the clustering problem as a concave minimization problem (11) which has not been given before. We note that a concave minimization problem which involves minimizing a concave function (mountain- like function) on a polyhedral set S is more difficult than minimizing a convex function (valley-like function) on S because the former may have many local minima at vertices of S that are not global minima, whereas for the latter each local minimum is a global minimum. Nevertheless for our clustering application the concave formulation is very effective in discovering well separated survival curves by determining k cluster centers such that the sum of the 1-norm distances of each point in a given database to the nearest cluster center is minimized. A new point is then assigned to the cluster with center nearest to the point. This simple formulation can be restated as a bilinear program (12) that leads to a fast k-Median Algorithm 3.1. A reformulation of (11) using the square of the 2-norm instead of the 1-norm, leads to the k-Mean Algorithm [32, 64]. Although it is not the intention here to carry out a detailed comparative study of these two clustering algorithms, the k-Median Algorithm, as described below, does give well separated survival curves for breast cancer patients whereas the k-Mean algorithm does not. On four other publicly available databases, each of the k-Median and k-Mean Algorithms did best on two of the databases. This indicates the potential of the k-Median Algorithm as a KDD tool. We describe the mathematical programming approach now. For a given set A of m points in R n represented by the matrix A 2 R m\Thetan and a number k of desired clusters, we formulate the clustering problem as follows. Find cluster centers C such that the sum of the minima over ' 2 of the 1-norm distance between each point A and the cluster centers C specifically we need to solve the following mathematical program: minimize C;D min subject to \GammaD i' - A T Here D i' 2 R n , is a dummy variable that bounds the components of the difference A T between point A T i and center C ' , and e is an n \Theta 1 vector of ones in R n . Hence e T D i' bounds the 1-norm distance between A i and C ' . We note that just as in the case of robust regression [31],[27, pp 82-87], the use of the 1-norm here to measure the error criterion leads to insensitivity to outliers such as those resulting from distributions with pronounced tails. We also note that since the objective function of (11) is the minimum of k linear (and hence concave) functions, it is a piecewise-linear concave function [44, Corollary 4.1.14]. This is not the case for the 2-norm or p-norm, p 6= 1. Although (11) is NP-hard, it can be reformulated as the following bilinear program which can be solved effectively by using a k-Median Algorithm that consists of solving a succession of simple linear programs in closed form. We state the bilinear programming formulation and k-Median Algorithm for solving the clustering problem. Proposition 3.1 Clustering as a Bilinear Program The clustering problem (11) is equivalent to the following bilinear program: minimize subject to \GammaD i' - A T This essentially obvious result [11, Proposition 2.2] can be seen from the fact that, for a fixed i, setting all the components of T i' equal to zero except one corresponding to a smallest e T D i' , with respect to ', equal to 1, leads to the objective function of (11) from that of (12). Note that the constraints of (12) are uncoupled in the variables (C; D) and the variable T . Hence the Uncoupled Bilinear Program Algorithm UBPA [5, Algorithm 2.1] is applicable. Simply stated, this algorithm alternates between solving a linear program in the variable T and a linear program in the variables (C; D). The algorithm terminates in a finite number of iterations at a stationary point satisfying the minimum principle necessary optimality condition for problem (12) [5, Theorem 2.1]. We note however, because of the simple structure the bilinear program (12), the two linear programs can be solved explicitly in closed form. This leads to the following algorithmic implementation. Algorithm 3.1 k-Median Algorithm Given the cluster centers C j k at iteration j, compute k by the following two steps: (a) Cluster Assignment: For each A T determine '(i) such that C j '(i) is closest to A T i in the one norm. (b) Cluster Center Update: For choose C j+1 ' as a median of all A T i assigned to Stop when C j+1 ' . Assign each point to a cluster whose center is closest in the 1-norm to the point. Although the k-Median Algorithm is similar to the k-Mean Algorithm wherein the 2-norm distance is used [64, 22], it differs from it computationally, and theoretically. In fact, the underlying problem (12) of the k-Median Algorithm is a concave minimization on a polyhedral set while the corresponding problem for a two- or p-norm, p 6= 1, is: minimize C;D min subject to \GammaD i' - A T This is not a concave minimization on a polyhedral set, because the minimum of a set of convex functions is not in general concave. We also note that the k-Mean Algorithm finds a stationary point not of problem (13) with but of the same problem except that kD i' k 2 is replaced by kD i' k 2 2 and thus favoring outliers. Without this squared distance term, the subproblem of the k-Mean Algorithm becomes the considerably harder Weber problem [56, 13] which locates a center in R n closest in sum of Euclidean distances (not their squares!) to a finite set of given points. The Weber problem has no closed form solution. However, using the mean as a cluster center of points assigned to the cluster, as done in the k-Mean Algorithm, minimizes the sum of the squares of the distances from the cluster center to the points. Because there is no guaranteed way to ensure global optimality of the solution obtained by either the k-Median or k-Mean Algorithms, different starting points can be used to initiate the algorithm. Random starting cluster centers or some other heuristic can be used such as placing k initial centers along the coordinate axes at densest, second densest intervals on the axes. The latter heuristic was used in our computational results. To test the effectiveness of the k-Median Algorithm it was used as a KDD tool [18] to mine the Breast Cancer Database (WPBC) in order to discover medical knowledge. For such medical databases, extracting well-separated survival curves provides an essential prognostic tool. Survival curves [34, 38] give expected percent of surviving patients as a function of time. The k-Median Algorithm was applied to WPBC to extract such curves. Survival curves were constructed for 194 patients using two clinically available features for each patient: tumor size and number of cancerous lymph nodes excised. Using the k-Median Algorithm separated the points into 3 clusters. The survival curve for each cluster is depicted in Figure 2(a). The key observation to make here is that curves are well separated, and hence the clusters can be used as prognostic indicators to assign a survival curve to a patient depending on the cluster into which the patient falls. By contrast, the k-Mean Algorithm obtained poorly separated survival curves as shown in Figure 2(b), and hence are not useful for prognosis. Another comparison of the k-Median and k-Mean Algorithms was performed on databases with known classes. Correctness was measured by the ratio of the sum of the number of majority points in each cluster to the total number of points m in the data set. Table 1 shows results averaged over ten random starts for four databases from the Irvine repository of databases [51]. We note that for two of the databases the k-Median gave better correctness than the k-Mean and for the other two the k-Mean was better. Months Percent Disease-Free Survival Curves for 3 Clusters Using (Tumor,Lymph) (k-Median) (a) k-Median Percent Disease-Free Survival Curves for 3 Clusters Using (Tumor,Lymph) (k-Mean) (b) k-Mean Figure 2: Survival curves for the 3 clusters of 194 cancer patients obtained by the k-Median and k-Mean Algorithms Algorithm # Database ! WDBC Cleveland Votes Star/Galaxy-Bright Unsupervised k-Median 93.2% 80.6% 84.6% 87.6% Unsupervised k-Mean 91.1% 83.1% 85.5% 85.6% Table 1 Training set correctness using the unsupervised k-Median and k-Mean Algorithms and the supervised Robust LP on four databases Robust Representation We consider now the problem of how to generate a robust representation of a system so that the model remains valid under a class of data perturbation. This problem is closely related to the generalization problem of machine learning of how to train a system on a given training set so as to improve generalization on a new unseen testing set [39, 63, 73]. We shall concentrate on some recent results [67] obtained for a simple linear model and which make essential use of mathematical programming ideas. These ideas, although rigorously established for a simple linear model here, are likely to extend to more complex systems. In a somewhat related approach Vapnik has proposed a quadratic program for constructing a plane to obtain a smallest probability of error for separating two point sets [70, Section 5.4]. Bennett and Bredensteiner [2] have formulated a similar problem using linear programming. Vapnik [70, Section 5.9] also makes use of Huber's robust regression ideas [30] and extends the latter's robust regression loss function [70, p 152] by adding an ffl- insensitive zone to it. This ffl-insensitive zone is similar to our -tolerance zone (see (17) and (18) below) wherein errors are disregarded if they fall within the band [\Gamma- ]. Another similarity with Vapnik's work is the presence of a regularization term fflkxk 2 2 in our minimization (17) problem (introduced here in order to make the solution unique and to rigorously derive our generalization theorems) and Vapnik's bounding his weight vector in order to control the VC dimension [70, p 128, Theorem 5.1]. However, what our approach provides here that is novel, are precise deterministic conditions (Propositions 4.1 and 4.2 below) under which tolerating a -insensitivity zone will give better generalization results than the conventional zero-tolerance that is used in an ordinary least squares approach. The model that we shall consider here consists of the training set fA; ag where A is a given m \Theta n real matrix and a is a given m \Theta 1 real vector. A vector x in R n is to be "learned" such that the linear system which does not have an exact solution, is satisfied in some approximate fashion, and such that the error in satisfying for some unseen testing set (C; c) 2 R k\Thetan \Theta R k , is minimized. Of course, if we disregard the testing set error (15), the problem becomes the standard least-norm problem: min where k\Deltak is some norm on R m . However with an eye to possible perturbations in the given training set fA; ag, we pose the following motivational question: If the vector a of the training set is known only to an accuracy of - , where - is some small positive number, does it make sense to attempt to drive the error to zero as is done in (16), or is it not better to tolerate errors in the satisfaction of a up to a magnitude of -? In other words, instead of (14), we should try to satisfy the following system of inequalities, in some best sense: To do that, we solve the following regularized quadratic program for some nonnegative - and a small positive ffl: minimize x;y;z2 2subject to \Gammaz \Gamma e- Ax \Gamma a - e- Here y and z are the errors in satisfying the inequalities of (17) and ffl is a small fixed positive regularization constant that ensures the uniqueness of the x component of the solution. Although ffl was held fixed in our computational experiments, it is possible to optimize its value by cross-validation on a tuning set in order to obtain better generalization. We note immediately, that if degenerates to the regularized classical least squares problem: min x2R The key question to ask here, is this: Under what conditions does a solution x(-) of (18), for some - ? 0 give a smaller error than x(0) on a testing set? We are able to give an answer to this question and corroborate it computationally [67], by considering a general testing set (C; c) 2 R k\Thetan \Theta R k for the problem (15) as well as a simpler testing set, where only the right side of (14) is perturbed. We first restrict ourselves to the latter and simpler perturbation, that is: where p is some arbitrary fixed perturbation in R m , and consider the following associated error In particular we would like to know when is f(0) not a local minimum of f(-) on the set f- j - 0g. In fact we are only interested in the -interval [0; - ], where - - is defined by because the minimum value of (18) approaches zero, for - - , as ffl approaches zero. The following proposition gives a sufficient condition which ensures that solving (18), for some positive - , produces an x(-) that generalizes better on the system (20) than that obtained by solving a plain regularized least squares problem (19), that is f(- Proposition 4.1 Robust Representation of [67]. For a solution x(-) of (18) the testing set error function f(-) of (21) has a strict local maximum at 0 and a global minimum on [0; -], where - is defined by (22), at some - ? 0, whenever for some - 2 (0; ~ - ], for some sufficiently small ~ - . 410121416Tolerance Test Figure 3: A computational example of Proposition 4.1. The bottom curve depicts a decreasing test error f(-) of (21) for the perturbed system (23). The top curve demonstrates an increasing error effect of violating (23) for sufficiently small ffl. Computational results carried out in [67] have corroborated the improved generalization results of Proposition 4.1 above. We depict in Figure 3 a simple numerical example that uses the training model (14), where the vector x is learned with various error tolerances - (18) and is then tested on the perturbed system p. In the learning situation depicted by the bottom curve of Figure makes an acute angle with in the neighborhood of Hence, as - increases, the test error in satisfying decreases. In the upper curve, we show a perturbation for which the angle is reversed: the testing model is using the same perturbation p. Here the testing error goes up as - increases away from zero. We conclude this section by extending Proposition 4.1 to a more general testing model where C 2 R k\Thetan and c 2 R k are chosen arbitrarily. To do this we define a corresponding error function g(-) to (21) which measures the error in satisfying (24) by the x(-) learned by solving the regularized quadratic program (18). We thus have the error ck 2 For this very general formulation we are able to give the following result that tells us when tolerant training does lead to improved generalization. Proposition 4.2 Improved generalization with positive tolerance for testing model Let x(-) be defined by the tolerant training of Ax = a by the regularized quadratic program (18) with tolerance - 0. Let g(-) denote the error generated by x(-) in the testing model defined by (25). The zero-tolerance error g(0) is a local maximum of g(-) over the set f- 0g whenever where r(-) is the residual vector defined by Furthermore, the testing set error function g(-) of (25) has a global minimum on [0; - -], for - defined by (22), at some - ? 0, whenever condition (26) holds and - 5 Conclusion A number of ideas based on mathematical programming have been proposed for the solution of the fundamental problems of feature selection, clustering and robust representation. Examples of applications of these ideas have been given to show their effectiveness. We discuss now some issues associated with these approaches. All the methods here use real variables. Even though the class of problems falling in this category is quite broad, this requirement imposes a restriction on the type of problems that can be handled. Nevertheless the proposed methods can be applied to problems with discrete variables if one is willing to use the techniques of integer and mixed integer programming [55, 21] which are more difficult. In fact one of the proposed algorithms, the k-Median Algorithm, whose finite termination is established for problems with real variables, is directly applicable with no change to problems with ordered discrete variables such as integers. How well it performs on such problems would be an interesting problem to examine. Another important practical issue is scalability. As mentioned earlier, since linear programs with millions of variables can be solved with present day state-of-the-art methods, large-scale databases are amenable to the proposed linear-programming-based methods. In addition there is a large body of literature on the parallel solution and decomposition of large scale mathematical programs [7, 16, 17, 19] where for many of the algorithms only part of the data is loaded into memory at a time. Such parallel and decomposition algorithms extend further the applicability of the proposed methods to very large scale databases. From a numerical standpoint, mathematical programming codes, and especially linear and quadratic programming codes, are reliable and robust codes that have been in a constant state of improvement over last fifty years. The well-understood polynomial-time finite termination of linear and quadratic programming interior methods has led to very powerful and reliable commercial software such as CPLEX [14] that can easily and reliably implement all the proposed algorithms. Finally we point out that although a linear model was used for both the feature selection and robust representation models, nonlinear models that are linear in their parameters, e.g. quadratic surfaces, can be easily transformed into a linear system, as was done for example in [41] where quadratic separation was achieved by linear programming. However in some inherently nonlinear problems where for example the parameters of a separating surface appear nonlinearly, one may have to resort to nonlinear models and the theory and algorithms of nonlinear programming [42, 6]. This would again be a promising problem to pursue. We conclude with the hope that the problems solved demonstrate the theoretical and computational potential of mathematical programming as a versatile and effective tool for solving important problems in data mining and knowledge discovery in databases. --R A Tabu search approach to the clustering problem. Geometry in learning. Neural network training via linear programming. Robust linear programming discrimination of two linearly inseparable sets. Bilinear separation of two sets in n-space Nonlinear Programming. Parallel and Distributed Computation. Very large-scale linear programming: a case study combining interior point and simplex methods Occam's razor. Feature selection via mathematical pro- gramming Clustering via concave minimization. Feature minimization within decision trees. On the Fermat-Weber problem with convex cost functionals CPLEX Optimization Inc. Linear Programming and Extensions. Serial and parallel solution of large scale linear programs by augmented Lagrangian successive overrelaxation. The KDD process for extracting useful knowledge from volumes of data. Parallel variable distribution. Knowledge acquisition via incremental conceptual clustering. Fundamentals and Applications (Topics in Chemical Engineering. Statistical Pattern Recognition. Convergence properties of backpropagation for neural nets via theory of stochastic gradient methods. Integer Programming. Improved linear programming models for discriminant analysis. Mathematical methods for pattern classification. Fundamentals of Artificial Neural Networks. Introduction to the Theory of Neural Computation. Introduction to Operations Research. Robust estimation of location parameter. Robust Statistics. Algorithms for Clustering Data. Irrelevant features and the subset selection problem. Nonparametric estimation from incomplete observations. Mathematical Methods and Theory in Games A new polynomial time algorithm for linear programming. Mathematical programming in contact problems. Survival Analysis. Optimal brain damage. Linear and nonlinear separation of patterns by linear programming. Nonlinear Programming. Mathematical programming in neural networks. Nonlinear Programming. Machine learning via polyhedral concave minimization. Mathematical programming in machine learning. Nonlinear perturbation of linear programs. Serial and parallel backpropagation convergence via nonmonotone perturbed minimization. Breast cancer diagnosis and prognosis via linear programming. Combinatorial optimization in inductive concept learning. UCI repository of machine learning databases. Linear Programming. Network Programming. Operations Research. John Wiley A quadratically convergent method for minimizing a sum of euclidean norms. Inequality Problems in Mechanics and Applications. analysis and mathematical programming. Convex Analysis. Network Flows and Monotropic Optimization. A polynomial time algorithm for the construction and training of a class of multilayer perceptrons. Parallel Distributed Processing. Overfitting avoidance as bias. Readings in Machine Learning. Pattern classifier design by linear programming. Improved generalization via tolerant training. Foundations of the Theory of Learning Systems. Linear Programming: Foundations and Extensions. The Nature of Statistical Learning Theory. Theory of Games and Economic Behavior. WPBC: Wisconsin Prognostic Breast Cancer Database. The Mathematics of Generalization --TR --CTR Bernardete Ribeiro, Learning adaptive kernels for model diagnosis, Design and application of hybrid intelligent systems, IOS Press, Amsterdam, The Netherlands, M. A. Goberna , V. Jeyakumar , N. Dinh, Dual Characterizations of Set Containments with Strict Convex Inequalities, Journal of Global Optimization, v.34 n.1, p.33-54, January 2006 Sanjeev Arora , Prabhakar Raghavan , Satish Rao, Approximation schemes for Euclidean Carlotta Orsenigo , Carlo Vercellis, Accurately learning from few examples with a polyhedral classifier, Computational Optimization and Applications, v.38 n.2, p.235-247, November 2007 V. F. Demyanov , A. M. Bagirov , A. M. Rubinov, A method of truncated codifferential with application to some problems of cluster analysis, Journal of Global Optimization, v.23 n.1, p.63-80, May 2002 P. S. Bradley , O. L. Mangasarian , J. B. Rosen, Parsimonious Least Norm Approximation, Computational Optimization and Applications, v.11 n.1, p.5-21, Oct. 1998 scalable decision tree system and its application in pattern recognition and intrusion detection, Decision Support Systems, v.41 n.1, p.112-130, November 2005 Jong-Shi Pang, Guest Editorial, Computational Optimization and Applications, v.13 n.1-3, p.5-12, April 1999 Jong-Shi Pang, Guest Editorial, Computational Optimization and Applications, v.12 n.1-3, p.5-12, Jan. 1999 Tapas Kanungo , David M. Mount , Nathan S. Netanyahu , Christine Piatko , Ruth Silverman , Angela Y. Wu, The analysis of a simple Gentile, A new approximate maximal margin classification algorithm, The Journal of Machine Learning Research, 2, 3/1/2002 Tapas Kanungo , David M. Mount , Nathan S. Netanyahu , Christine D. Piatko , Ruth Silverman , Angela Y. Wu, An Efficient k-Means Clustering Algorithm: Analysis and Implementation, IEEE Transactions on Pattern Analysis and Machine Intelligence, v.24 n.7, p.881-892, July 2002 Gunnar Rtsch , Sebastian Mika , Bernhard Schlkopf , Klaus-Robert Mller, Constructing Boosting Algorithms from SVMs: An Application to One-Class Classification, IEEE Transactions on Pattern Analysis and Machine Intelligence, v.24 n.9, p.1184-1199, September 2002 Balaji Padmanabhan , Alexander Tuzhilin, On the Use of Optimization for Data Mining: Theoretical Interactions and eCRM Opportunities, Management Science, v.49 n.10, p.1327-1343, October Sudipto Guha , Adam Meyerson , Nina Mishra , Rajeev Motwani , Liadan O'Callaghan, Clustering Data Streams: Theory and Practice, IEEE Transactions on Knowledge and Data Engineering, v.15 n.3, p.515-528, March
robust representation;mathematical programming;clustering;feature selection
593449
Discovery of Frequent Episodes in Event Sequences.
Sequences of events describing the behavior and actions of users or systems can be collected in several domains. An episode is a collection of events that occur relatively close to each other in a given partial order. We consider the problem of discovering frequently occurring episodes in a sequence. Once such episodes are known, one can produce rules for describing or predicting the behavior of the sequence. We give efficient algorithms for the discovery of all frequent episodes from a given class of episodes, and present detailed experimental results. The methods are in use in telecommunication alarm management.
Introduction Most data mining and machine learning techniques are adapted towards the analysis of unordered collections of data. However, there are important application areas where the data to be analyzed consists of a sequence of events. Examples of such data are alarms in a telecommunication network, user interface actions, crimes committed by a person, occurrences of recurrent ill- nesses, etc. Recently, interest in knowledge discovery from sequential data has increased: see, e.g., [5, 8, 17, 19, 24]. Abstractly, such data can be viewed as a sequence of events, where each event has an associated time of occurrence. An example of an event sequence is represented in Figure 1. Here A; B;C;D;E; and F are event types, e.g., dioeerent types of alarms from a telecommunication network, or dioeerent types of user actions, and they have been marked on a time line. One basic problem in analyzing such a sequence is to -nd frequent epis- odes, i.e., collections of events occurring frequently together. For example, in the sequence of Figure 1, the episode iE is followed by Fj occurs several times, even when the sequence is viewed through a narrow window. Epis- odes, in general, are partially ordered sets of events. From the sequence in the -gure one can make, for instance, the observation that whenever A and occur (in either order), C occurs soon. When discovering episodes in a telecommunication network alarm log, the goal is to -nd relationships between alarms. Such relationships can then be used in an on-line analysis of the incoming alarm stream, e.g, to better explain the problems that cause alarms, to suppress redundant alarms, and to predict severe faults. In this paper we consider the following problem. Given a class of episodes and an input sequence of events, -nd all episodes that occur frequently in the event sequence. We describe the framework and formalize the know- time EDF A BCEF C D BAD C EFC BEAECF A D Figure 1: A sequence of events. ledge discovery task in Section 2. Algorithms for discovering all frequent episodes are given in Section 3. Thery are based on the idea of -rst -nding small frequent episodes, and then progressively looking for larger frequent episodes. Additionally, the algorithms use some simple pattern matching ideas to speed up the recognition of occurrences of single episodes. Section 4 outlines an alternative way of approaching the problem, based on locating minimal occurrences of episodes. Experimental results using both approaches and with various data sets are presented in Section 5. We discuss extensions and review related work in Section 6. Section 7 is a short conclusion. 2 Event sequences and episodes Our overall goal is to analyze sequences of events, and to discover recurrent combinations of events, which we call frequent episodes. We -rst formulate the concept of event sequence, and then look at episodes in more detail. 2.1 Event sequences We consider the input as a sequence of events, where each event has an associated time of occurrence. Given a set E of event types, an event is a is an event type and t is an integer, the (occurrence) time of the event. The event type can actually contain several attributes; for simplicity we consider here the event type as a single value. An event sequence s on E is a triple (s; T EDF A BCEF C D BAD C EFC BEAECF A D Figure 2: The example event sequence and two windows of width 5. is an ordered sequence of events such that A i 2 E for all are integers, T s is called the starting time and T e the ending time, and T s - Example 1 Figure 2 presents graphically the event sequence where Observations of the event sequence have been made from time 29 to just before time 68. For each event that occurred in the time interval [29; 68), the event type and the time of occurrence have been recorded. In the analysis of sequences we are interested in -nding all frequent episodes from a class of episodes. To be considered interesting, the events of an episode must occur close enough in time. The user de-nes how close is close enough by giving the width of the time window within which the episode must occur. We de-ne a window as a slice of an event sequence, and we then consider an event sequence as a sequence of partially overlapping windows. In addition to the width of the window, the user speci-es in how many windows an episode has to occur to be considered frequent. Formally, a window on event sequence is an event sequence consists of those pairs (A; t) from s where t s - . The time span t e \Gamma t s is called the width of the window w, and it is denoted width(w). Given an event sequence s and an integer win, we denote by W(s; win) the set of all windows w on s such that By the de-nition the -rst and last windows on a sequence extend outside the sequence, so that the -rst window contains only the -rst time point of the sequence, and the last window contains only the last time point. With this de-nition an event close to either end of a sequence is observed in equally many windows to an event in the middle of the sequence. Given an event sequence window width win, the number of windows in W(s; win) is Example 2 Figure 2 shows two windows of width 5 on the sequence s of the previous example. A window starting at time 35 is shown in solid line, and the immediately following window, starting at time 36, is depicted with a dashed line. The window starting at time 35 is Note that the event (F; 40) that occurred at the ending time is not in the window. The window starting at 36 is similar to this one; the dioeerence is that the -rst event (A; 35) is missing and there is a new event (F; 40) at the end. The set of the 43 partially overlapping windows of width 5 constitutes W(s; 5); the -rst window is (;; 25; 30), and the last is (h(D; Event (D; 67) occurs in 5 windows of width 5, as does, e.g., event (C; 50). 2.2 Episodes Informally, an episode is a partially ordered collection of events occurring together. Episodes can be described as directed acyclic graphs. Consider, for instance, episodes ff, fi, and fl in Figure 3. Episode ff is a serial episode: it occurs in a sequence only if there are events of types E and F that occur ff A A Figure 3: Episodes ff; fi, and fl. in this order in the sequence. In the sequence there can be other events occurring between these two. The alarm sequence, for instance, is merged from several sources, and therefore it is useful that episodes are insensitive to intervening events. Episode fi is a parallel episode: no constraints on the relative order of A and B are given. Episode fl is an example of non-serial and non-parallel episode: it occurs in a sequence if there are occurrences of A and B and these precede an occurrence of C; no constraints on the relative order of A and B are given. We mostly consider the discovery of serial and parallel episodes. We now de-ne episodes formally. An episode ff is a triple (V; -; g) where V is a set of nodes, - is a partial order on V , and mapping associating each node with an event type. The interpretation of an episode is that the events in g(V ) have to occur in the order described by -. The size of ff, denoted jffj, is jV j. Episode ff is parallel if the partial order - is trivial (i.e., x 6- y for all x; y 2 V such that x 6= y). Episode ff is serial if the relation - is a total order (i.e., x - y or y - x for all is injective if the mapping g is an injection, i.e., no event type occurs twice in the episode. Example 3 Consider episode Figure 3. The set V contains two nodes; say x and y. The mapping g labels these nodes with the event types that are seen in the -gure: An event of type is supposed to occur before an event of type F , i.e., x precedes y, and we have x - y. Episode ff is injective, since it does not contain duplicate event types; in a window where ff occurs there may, however, be multiple events of types E and F . We next de-ne when an episode is a subepisode of another; this relation is used extensively in the algorithms for discovering all frequent episodes. An there exists an injective mapping all ff is a superepisode of fi if and only if fi - ff. We write fi OE ff if fi - ff and ff 6- fi. Example 4 From Figure 3 we see that fi - fl since fi is a subgraph of fl. In terms of the de-nition, there is a mapping f that connects the nodes labeled A with each other and the nodes labeled B with each other, i.e., both nodes of fi have (disjoint) corresponding nodes in fl. Since the nodes in episode fi are not ordered, the corresponding nodes in fl do not need to be Consider now what it means that an episode occurs in a sequence. The nodes of the episode need to have corresponding events in the sequence such that the event types are the same and the partial order of the episode is respected. An occurs in an event sequence if there exists an injective mapping ng from nodes to events, such that Example 5 The window (w; 35; 40) of Figure 2 contains events A; B; C; and E. Episodes fi and fl of Figure 3 occur in the window, but ff does not. odes, a window width win, a frequency threshold min-fr, and a con-dence threshold min-conf. Output: The episode rules that hold in s with respect to win, min-fr, and min-conf. 1. /* Find frequent episodes (Algorithm 2): */ 2. compute F(s; win; min-fr); 3. /* Generate rules: */ 4. for all ff 2 F(s; win; min-fr) do 5. for all fi OE ff do 6. if fr(ff)=fr(fi) - min-conf then 7. output the rule fi ! ff and the con-dence fr(ff)=fr(fi); We de-ne the frequency of an episode as the fraction of windows in which the episode occurs. That is, given an event sequence s and a window width win, the frequency of an episode ff in s is occurs in wgj Given a frequency threshold min-fr, ff is frequent if fr(ff; s; win) - min-fr. The task we are interested in is to discover all frequent episodes from a given class E of episodes. The class could be, e.g., all parallel episodes or all serial episodes. We denote the collection of frequent episodes with respect to s, win and min-fr by F(s; win; min-fr). Once the frequent episodes are known, they can be used to obtain rules that describe connections between events in the given event sequence. For example, if we know that the episode fi of Figure 3 occurs in 4.2 % of the windows and that the superepisode fl occurs in 4.0 % of the windows, we can estimate that after seeing a window with A and B, there is a chance of about 0.95 that C follows in the same window. Such rules show the connections between events more clearly than frequent episodes alone. Algorithm 1 shows how rules and their con-dences can be computed from the frequencies of episodes, a window width win, and a frequency threshold min-fr. Output: The collection F(s; win; min-fr) of frequent episodes. 1. compute C 1 := fff 2. l := 3. while C l 6= ; do 4. /* Database pass (Algorithms 4 and 5): */ 5. compute F l := fff 2 C l j fr(ff; s; win) - min-frg; 7. /* Candidate generation (Algorithm 3): */ 8. compute C l := fff l and for all fi 2 E such that fi OE ff and 9. l we have fi 2 F jfij 10. for all l do output F l ; episodes. Note that indentation is used in the algorithms to specify the extent of loops and conditional statements. Algorithms Given all frequent episodes, the rule generation is straightforward. We now concentrate on the following discovery task: given an event sequence s, a set E of episodes, a window width win, and a frequency threshold min-fr, -nd F(s; win; min-fr): We give -rst a speci-cation of the algorithm and then exact methods for its subtasks. We call these methods collectively the algorithm. 3.1 Main algorithm Algorithm 2 computes the collection F(s; win; min-fr) of frequent episodes from a class E of episodes. The algorithm performs a levelwise (breadth-rst) search in the episode lattice spanned by the subepisode relation. The search starts from the most general episodes, i.e., episodes with only one event. On each level the algorithm -rst computes a collection of candidate episodes, and then checks their frequencies from the event sequence database. The crucial point in the candidate generation is given by the following immediate lemma. Lemma 6 If an episode ff is frequent in an event sequence s, then all subepis- odes fi - ff are frequent. The collection of candidates is speci-ed to consist of episodes such that all smaller subepisodes are frequent. This criterion safely prunes from consideration episodes that can not be frequent. More detailed methods for the candidate generation and database pass phases are given in the following subsections. 3.2 Generation of candidate episodes We present now a candidate generation method in detail. The method can be easily adapted to deal with the classes of parallel episodes, serial episodes, and injective parallel and serial episodes. Algorithm 3 computes candidates for parallel episodes. In the algorithm, an represented as a lexicographically sorted array of event types. The array is denoted by the name of the episode and the items in the array are referred to with the square bracket notation. For example, a parallel episode ff with events of types A; C; C; and F is represented as an array ff with . Collections of episodes are also represented as lexicographically sorted arrays, i.e., the ith episode of a collection F is denoted by F [i]. Since the episodes and episode collections are sorted, all episodes that share the same -rst event types are consecutive in the episode collection. In particular, if episodes F l [i] and F l [j] of size l share the -rst l \Gamma 1 events, then for all k with we have that F l [k] shares also the same Algorithm 3 Input: A sorted array F l of frequent parallel episodes of size l. Output: A sorted array of candidate parallel episodes of size l + 1. 1. C l+1 := ;; 2. k := 0; 3. if l = 1 then for h := 1 to jF l j do F l :block-start[h] := 4. for i := 1 to jF l j do 5. current-block-start 6. for (j := 7. /* F l [i] and F l [j] have l \Gamma 1 -rst event types in common, 8. build a potential candidate ff as their combination: */ 9. for x := 1 to l do ff[x] := F l [i][x]; 11. /* Build and test subepisodes fi that do not contain ff[y]: */ 12. for y := 1 to l \Gamma 1 do 13. for x := 1 to y \Gamma 1 do fi[x] := ff[x]; 14. for x := y to l do fi[x] := ff[x 15. if fi is not in F l then continue with the next j at line 6; 16. /* All subepisodes are in F l , store ff as candidate: */ 18. C l+1 [k] := ff; 19. C l+1 :block-start[k] := current-block-start; events. A maximal sequence of consecutive episodes of size l that share the -rst events is called a block. Potential candidates can be identi-ed by creating all combinations of two episodes in the same block. For the eOEcient identi-cation of blocks, we store in F l :block-start [j] for each episode F l [j] the i such that F l [i] is the -rst episode in the block. Algorithm 3 can be easily modi-ed to generate candidate serial episodes. Now the events in the array representing an episode are in the order imposed by a total order -. For instance, a serial episode fi with events of types C; A; F; and C, in that order, is represented as an array fi with replacing line 6 by for (j := F l :block-start[i]; F l Algorithm 3 generates candidates for serial episodes. There are further options with the algorithm. If the desired episode class consists of parallel or serial injective episodes, i.e., no episode should contain any event type more than once, simply insert line continue with the next j at line 6; after line 6. The time complexity of Algorithm 3 is polynomial in the size of the collection of frequent episodes and it is independent of the length of the event sequence. Theorem 1 Algorithm 3 (with any of the above variations) has time complexity O(l 2 jF l j 2 log jF l j). Proof The initialization (line j). The outer loop (line is iterated O(jF l j) times and the inner loop (line times. Within the loops, a potential candidate (lines 9 and 10) and l \Gamma 1 subcandidates (lines 12 to are built in time O(l importantly, the subsets need to be searched for in the collection F l (line 15). Since F l is sorted, each subcandidate can be located with binary search in time O(l log jF l j). The total time complexity is thus O(jF l j l log jF l In practical situations the time complexity is likely to be close to O(l 2 jF l j log jF l j), since the blocks are typically small. 3.3 Recognizing episodes in sequences Let us now consider the implementation of the database pass. We give algorithms which recognize episodes in sequences in an incremental fashion. For two windows 1), the sequences w and w 0 of events are similar to each other. We take advantage of this similarity: after recognizing episodes in w, we make incremental updates in our data structures to achieve the shift of the window to obtain w 0 . The algorithms start by considering the empty window just before the input sequence, and they end after considering the empty window just after the sequence. This way the incremental methods need no other special actions at the beginning or end. When computing the frequency of episodes, only the windows correctly on the input sequence are, of course, considered. 3.3.1 Parallel episodes Algorithm 4 recognizes candidate parallel episodes in an event sequence. The main ideas of the algorithm are the following. For each candidate parallel episode ff we maintain a counter ff:event-count that indicates how many events of ff are present in the window. When ff:event-count becomes equal to jffj, indicating that ff is entirely included in the window, we save the starting time of the window in ff:inwindow . When ff:event-count decreases again, indicating that ff is no longer entirely in the window, we increase the -eld ff:freq-count by the number of windows where ff remained entirely in the window. At the end, ff:freq-count contains the total number of windows where ff occurs. To access candidates eOEciently, they are indexed by the number of events of each type that they contain: all episodes that contain exactly a events of type A are in the list contains(A,a). When the window is shifted and the contents of the window change, the episodes that are aoeected are updated. If, for instance, there is one event of type A in the window and a second one comes in, all episodes in the list contains(A,2) are updated with the information that both events of type A they are expecting are now present. Algorithm 4 Input: A collection C of parallel episodes, an event sequence a window width win, and a frequency threshold min-fr. Output: The episodes of C that are frequent in s with respect to win and min-fr. 1. /* Initialization: */ 2. for each ff in C do 3. for each A in ff do 4. A:count := 0; 5. for i := 1 to jffj do contains(A; i) := ;; 6. for each ff in C do 7. for each A in ff do 8. a := number of events of type A in ff; 9. 10. ff.event-count := 0; 11. ff.freq-count := 0; 12. /* Recognition: */ 13. for start := T s do 14. /* Bring in new events to the window: */ 15. for all events (A; t) in s such that 16. A:count := A:count 17. for each ff 2 contains(A; A:count) do 18. ff:event-count := ff:event-count 19. if ff:event-count = jffj then ff:inwindow := start ; 20. /* Drop out old events from the window: */ 21. for all events (A; t) in s such that 22. for each ff 2 contains(A; A:count) do 23. if ff:event-count = jffj then 24. ff:freq-count := ff:freq-count 25. ff:event-count := ff:event-count \Gamma A:count ; 26. A:count := A:count \Gamma 27. /* Output: */ 28. for all episodes ff in C do 29. if ff:freq-count=(T 3.3.2 Serial episodes Serial candidate episodes are recognized in an event sequence by using state automata that accept the candidate episodes and ignore all other input. The idea is that there is an automaton for each serial episode ff, and that there can be several instances of each automaton at the same time, so that the active states reAEect the (disjoint) pre-xes of ff occurring in the window. Algorithm 5 implements this idea. We initialize a new instance of the automaton for a serial episode ff every time the -rst event of ff comes into the window; the automaton is removed when the same event leaves the window. When an automaton for ff reaches its accepting state, indicating that ff is entirely included in the window, and if there are no other automata for ff in the accepting state already, we save the starting time of the window in ff:inwindow . When an automaton in the accepting state is removed, and if there are no other automata for ff in the accepting state, we increase the -eld ff:freq-count by the number of windows where ff remained entirely in the window. It is useless to have multiple automata in the same state, as they would only make the same transitions and produce the same information. It suOEces to maintain the one that reached the common state last since it will be also removed last. There are thus at most jffj automata for an episode ff. For each automaton we need to know when it should be removed. We can thus represent all the automata for ff with one array of size jffj: the value of ff:initialized [i] is the latest initialization time of an automaton that has reached its ith state. Recall that ff itself is represented by an array containing its events; this array can be used to label the state transitions. To access and traverse the automata eOEciently they are organized in the following way. For each event type A 2 E, the automata that accept A are linked together to a list waits(A). The list contains entries of the form (ff; x) meaning that episode ff is waiting for its xth event. When an event enters the window during a shift, the list waits(A) is traversed. If an automaton reaches a common state i with another automaton, the earlier entry ff:initialized [i] is simply overwritten. The transitions made during one shift of the window are stored in a list transitions. They are represented in the form (ff; x; t) meaning that episode Algorithm 5 Input: A collection C of serial episodes, an event sequence window width win, and a frequency threshold min-fr. Output: The episodes of C that are frequent in s with respect to win and min-fr. 1. /* Initialization: */ 2. for each ff in C do 3. for i := 1 to jffj do 4. ff:initialized[i] := 0; 5. waits(ff[i]) := ;; 6. for each ff 2 C do 7. waits(ff[1]) := waits(ff[1]) [ f(ff; 1)g; 8. ff.freq-count := 0; 9. for t := 10. /* Recognition: */ 11. for start := T s do 12. /* Bring in new events to the window: */ 13. beginsat(start 14. transitions := ;; 15. for all events (A; t) in s such that 16. for all (ff; do 17. if 19. transitions := transitions [ f(ff; 20. else 21. transitions := transitions [ f(ff; 24. ff:initialized 25. 26. for all (ff; j; t) 2 transitions do 27. ff:initialized[j] := t; 28. beginsat(t) := beginsat(t) [ f(ff; j)g; 29. 30. /* Drop out old events from the window: */ 31. for all (ff; l) 2 beginsat(start \Gamma 1) do 33. else waits(ff[l 34. ff:initialized[l] := 0; 35. /* Output: */ 36. for all episodes ff in C do 37. if ff:freq-count=(T ff got its xth event, and the latest initialization time of the pre-x of length x is t. Updates regarding the old states of the automata are done immediately, but updates for the new states are done only after all transitions have been identi-ed, in order to not overwrite any useful information. For easy removal of automata when they go out of the window, the automata initialized at time t are stored in a list beginsat(t). 3.3.3 Analysis of time complexity For simplicity, suppose that the class of event types E is -xed, and assume that exactly one event takes place every time unit. Assume candidate episodes are all of size l, and let n be the length of the sequence. Theorem 2 The time complexity of Algorithm 4 is O((n Proof Initialization takes time O(jCj l 2 ). Consider now the number of the operations in the innermost loops, i.e., accesses to ff:event-count on lines and 25. In the recognition phase there are O(n) shifts of the window. In each shift, one new event comes into the window, and one old event leaves the window. Thus, for any episode ff, ff:event-count is accessed at most twice during one shift. The cost of the recognition phase is thus O(n jCj). In practice the size l of episodes is very small with respect to the size n of the sequence, and the time required for the initialization can be safely neglected. For injective episodes we have the following tighter result. Theorem 3 The time complexity of recognizing injective parallel episodes in Algorithm 4 (excluding initialization) is O( n win Proof Consider win successive shifts of one time unit. During such sequence of shifts, each of the jCj candidate episodes ff can undergo at most 2l changes: any event type A can have A:count increased to 1 and decreased to 0 at most once. This is due to the fact that after an event of type A has come into the window, A:count - 1 for the next win time units. Reading the input takes time n. This time bound should be contrasted with the time usage of a trivial non-incremental method where the sequence is pre-processed into windows, and then frequent sets are searched for. The time requirement for recognizing jCj candidate sets in n windows, plus the time required to read in n windows of size win, is O(n jCj l larger by a factor of win. Theorem 4 The time complexity of Algorithm 5 is O(n jCj l): Proof The initialization takes time O(jCj l +win). In the recognition phase, again, there are O(n) shifts, and in each shift one event comes into the window and one event leaves the window. In one shift, the eoeort per an ff depends on the number of automata accessed; there are a maximum of l automata for each episode. The worst-case time complexity is thus (note that win is O(n)). In the worst case the input sequence consists of events of only one event type, and the candidate serial episodes consist only of events of that particular type. Every shift of the window results now in an update in every automaton. This worst-case complexity is close to the complexity of the trivial non-incremental method O(n jCj l In practical situations, however, the time requirement is considerably smaller, and we approach the savings obtained in the case of injective parallel episodes. Theorem 5 The time complexity of recognizing injective serial episodes in Algorithm 5 (excluding initialization) is O(n jCj): Proof Each of the O(n) shifts can now aoeect at most two automata for each episode: when an event comes into the window there can be a state A A Figure 4: Recursive composition of a complex episode. transition in at most one automaton, and at most one automaton can be removed because the initializing event goes out of the window. 3.4 General partial orders So far we have only discussed serial and parallel episodes. We next discuss brieAEy the use of other partial orders in episodes. The recognition of an arbitrary episode can be reduced to the recognition of a hierarchical combination of serial and parallel episodes. For example, episode fl in Figure 4 is a serial combination of two episodes: a parallel episode ffi 0 consisting of A and B, and an episode ffi 00 consisting of C alone. The occurrence of an episode in a window can be tested using such hierarchical structure: to see whether episode fl occurs in a window one checks (using a method for serial whether the subepisodes ffi 0 and ffi 00 occur in this order; to check the occurrence of ffi 0 one uses a method for parallel episodes to verify whether A and B occur. There are, however, some complications one has to take into account. First, it is sometimes necessary to duplicate an event node to obtain a decomposition to serial and parallel episodes. Duplication works easily with injective episodes, but non-injective episodes need more complex methods. Another important aspect is that composite events have a duration, unlike the elementary events in E. A practical alternative is to handle all episodes basically like parallel episodes, and to check the correct partial ordering only when all events are in the window. Parallel episodes can be located eOEciently; after they have been found, checking the correct partial ordering is relatively fast. 4 An alternative occurrence 4.1 Outline of the approach In this section we describe an alternative approach to the discovery of epis- odes. Instead of looking at the windows and only considering whether an occurs in a window or not, we now look at the exact occurrences of episodes and the relationships between those occurrences. One of the advantages of this new approach is that focusing on the occurrences of episodes allows us to more easily -nd rules with two window widths, one for the left-hand side and one for the whole rule, such as iif A and B occur within 15 seconds, then C follows within The approach is based on minimal occurrences of episodes. Besides the new rule formulation, the use of minimal occurrences gives raise to the following new method, called Minepi, for the recognition of episodes in the input sequence. For each frequent episode we store information about the locations of its minimal occurrences. In the recognition phase we can then compute the locations of minimal occurrences of a candidate episode ff as a temporal join of the minimal occurrences of two subepisodes of ff. In addition to being simple and eOEcient, this formulation has the advantage that the con-dences and frequencies of rules with a large number of dioeerent window widths can be obtained quickly, i.e., there is no need to rerun the analysis if one only wants to modify the window widths. In the case of complicated episodes, the time needed for recognizing the occurrence of an episode can be signi-c- ant; the use of stored minimal occurrences of episodes eliminates unnecessary repetition of the recognition eoeort. We identify minimal occurrences with their time intervals in the following way. Given an episode ff and an event sequence s, we say that the interval occurrence of ff in s, if (1) ff occurs in the window e ) on s, and if (2) ff does not occur in any proper subwindow on i.e., not in any window w e ) on s such that t s - t 0 s e width(w). The set of (intervals of) minimal occurrences of an episode ff in a given event sequence is denoted by mo(ff): occurrence of ffg: Example 7 Consider the event sequence s in Figure 2 and the episodes in Figure 3. The parallel episode fi consisting of event types A and B has four minimal occurrences in s: The partially ordered episode fl has the following three minimal occurrences: An episode rule is an expression fi [win 1 are episodes such that fi - ff, and win 1 and win 2 are integers. The informal interpretation of the rule is that if episode fi has a minimal occurrence at interval occurs at interval [t s for some t 0 e such that t 0 e . Formally this can be expressed in the following way. Given win 1 and fi, denote mo win 1 and an interval [u s ; u e ), de-ne occ(ff; [u s ; u e true if and only if there exists a minimal occurrence [u 0 s and e - u e . The con-dence of an episode rule fi [win 1 is now Example 8 Continuing the previous example, we have, e.g., the following rules and con-dences. For the rule fi [3] ) fl [4] we have in the denominator and jf[35; 38)gj in the nu- merator, so the con-dence is 1=3. For the rule fi [3] ) fl [5] the con-dence is 1. Note that since fi is a subepisode of ff, the rule right-hand side ff contains information about the relative location of each event in it, so the inewj events in the rule right-hand side can actually be required to be positioned, e.g., between events in the left-hand side. There is also a number of possible de-nitions for the temporal relationship between the intervals. For instance, rules that point backwards in time can be de-ned in a similar way. For brevity, we only consider this one case. We de-ned the frequency of an episode as the fraction of windows that contain the episode. While frequency has a nice interpretation as the probability that a randomly chosen window contains the episode, the concept is not very useful with minimal occurrences: (1) there is no -xed window size, and (2) a window may contain several minimal occurrences of an episode. Instead of frequency, we use the concept of support, the number of minimal occurrences of an episode: the support of an episode ff in a given event sequence s is jmo(ff)j. Similarily to the a frequency threshold, we now use a threshold for the support: given a support threshold min-sup, an episode ff is frequent if jmo(ff)j - min-sup. The current episode rule discovery task can be stated as follows. Given an event sequence s, a class E of episodes, and a set W of time bounds, -nd all frequent episode rules of the form fi [win 1 4.2 Finding minimal occurrences of episodes In this section we describe informally the collection Minepi of algorithms that locate the minimal occurrences of frequent serial and parallel episodes. Let us start with some observations about the basic properties of episodes. Lemma 6 still holds: the subepisodes of a frequent episode are frequent. Thus we can use the main algorithm (Algorithm 2) and the candidate generation (Algorithm We have the following results about the minimal occurrences of an episode also containing minimal occurrences of its subepisodes. Lemma 9 Assume ff is an episode and fi - ff is its subepisode. If mo(ff), then fi occurs in [t s ; t e ) and hence there is an interval [u s ; such that t s - ff be a serial episode of size k, and let [t there are subepisodes ff 1 and ff 2 of ff of size for some t 1 Lemma 11 Let ff be a parallel episode of size k, and let [t Then there are subepisodes ff 1 and ff 2 of ff of size e s and t e g. The minimal occurrences of a candidate episode ff are located in the following way. In the -rst iteration of the main algorithm, mo(ff) is computed from the input sequence for all episodes ff of size 1. In the rest of the iterations, the minimal occurrences of a candidate ff are located by -rst selecting two suitable subepisodes ff 1 and ff 2 of ff, and then computing a temporal join between the minimal occurrences of ff 1 and ff 2 , in the spirit of To be more speci-c, for serial episodes the two subepisodes are selected so that ff 1 contains all events except the last one and ff 2 in turn contains all except the -rst one. The minimal occurrences of ff are then found with the following speci-cation: such that t s ! u For parallel episodes, the subepisodes ff 1 and ff 2 contain all events except one; the omitted events must be dioeerent. See Lemma 11 for the idea of how to compute the minimal occurrences of ff. The minimal occurrences of a candidate episode ff can be found in a linear pass over the minimal occurrences of the selected subepisodes ff 1 and ff 2 . The time required for one candidate is thus O(jmo(ff which is O(n), where n is the length of the event sequence. To optimize the running time, ff 1 and ff 2 can be selected so that jmo(ff 1 )j minimized. The space requirement of the algorithm can be expressed as jmo(ff)j, assuming the minimal occurrences of all frequent episodes are stored, or alternatively as jmo(ff)j), if only the current and next levels of minimal occurrences are stored. The size of P is bounded by n, the number of events in the input sequence, as each event in the sequence is a minimal occurrence of an episode of size 1. In the second iteration, an event in the input sequence can start at most jF 1 j minimal occurrences of episodes of size 2. The space complexity of the second iteration is thus O(jF 1 jn). While minimal occurrences of episodes can be located quite eOEciently, the size of the data structures can be even larger than the original database, especially in the -rst couple of iterations. A practical solution is to use in the beginning other pattern matching methods, e.g., similar to the ones given for in Section 3, to locate the minimal occurrences. Finally, note that Minepi can be used to solve the task of Winepi. Namely, a window contains an occurrence of an episode exactly when it contains a minimal occurrence. The frequency of an episode ff can thus be computed from mo(ff). 4.3 Finding con-dences of rules We now show how the information about minimal occurrences of frequent episodes can be used to obtain con-dences for various types of episode rules without looking at the data again. Recall that we de-ned an episode rule as an expression fi [win 1 ff [win 2 ]; where fi and ff are episodes such that fi - ff, and win 1 and win 2 are integers. To -nd such rules, -rst note that for the rule to be frequent, the episode ff has to be frequent. So rules of the above form can be enumerated by looking at all frequent episodes ff, and then looking at all subepisodes fi of ff. The evaluation of the con-dence of the rule fi [win 1 be done in one pass through the structures mo(fi) and mo(ff), as follows. For each locate the minimal occurrence of ff such that t s - u s and [u s ; u e ) is the -rst interval in mo(ff) with this property. Then check whether u The time complexity of the con-dence computation for a given episode and given time bounds win 1 and win 2 is O(jmo(fi)j jmo(ff)j). The con- -dences for all win in the set W of time bounds can be found, using a table of size jWj 2 , in time O(jmo(fi)j For reasons of brevity we omit the details. The set W of time bounds can be used to restrict the initial search of minimal occurrences of episodes. Given W, denote the maximum time bound by In episode rules, only occurrences of at most win units can be used; longer episode occurrences can thus be ignored already in the search of frequent episodes. We consider the support, too, to be computed with respect to a given win max . 5 Experiments We have run a series of experiments using Winepi and Minepi. The general performance of the methods, the eoeect of the various parameters, and the scalability of the methods are considered in this section. Consideration is also given to the applicability of the methods to various types of data sets. The experiments have been run on a PC with 166 MHz Pentium processor and main memory, under the Linux operating system. The sequences resided in a AEat text -le. 5.1 Performance overview For an experimental overview we discovered episodes and rules in a telecommunication network fault management database. The database is a sequence of 73679 alarms covering a time period of 7 weeks. There are 287 dioeerent types of alarms with very diverse frequencies and distributions. On the average there is an alarm every minute. However, the alarms tend to occur in bursts: in the extreme cases there are over 40 alarms in one second. We start by looking at the performance of the Winepi method described in Section 3. There are several performance characteristics that can be used to evaluate the method. The time required by the method and the number of episodes and rules found by the method, with respect to the frequency threshold or the window width, are possible performance measures. We present results for the two opposite extreme cases of the complexity: serial episodes and injective parallel episodes. Tables -nding frequent episodes in the alarm database with various frequency thresholds. The number of frequent episodes decreases rapidly as the frequency threshold increases. With a given frequency threshold, the numbers of serial and injective parallel episodes may be fairly similar, e.g., a frequency threshold of 0.002 results in Frequency Candidates Frequent Iterations Total threshold Episodes time Table 1: Performance characteristics for serial episodes with database, window width Frequency Candidates Frequent Iterations Total threshold Episodes time Table 2: Performance characteristics for injective parallel episodes with serial episodes or 93 parallel episodes. The actual episodes are, however, very dioeerent, as can be seen from the number of iterations: recall that each iteration l produces episodes of size l. For the frequency threshold of 0.002, the longest frequent serial episode consists of 43 events (all candidates of the last iteration were infrequent), while the longest frequent injective parallel episodes have 3 events. The number of iterations equals the number of candidate generation phases. The number of database passes equals the number of iterations, or is smaller by one when there were no candidates in the last iteration. Episodes Window width Figure 5: Number of frequent serial (solid line) and injective parallel (dotted line) episodes as a function of the window width; Winepi, alarm database, frequency threshold 0:002. The eoeect of the window width on the number of frequent episodes is represented in Figure 5. For each window width, there are considerably fewer frequent injective parallel episodes than frequent serial episodes. With the alarm data, the increase in the number of episodes is fairly even throughout the window widths that we considered. However, we will later show that this may depend heavily on the type of data we are using. Figure 6 represents the number of serial and injective parallel episodes found by the method, and Figure 7 the total processing time required, as the frequency threshold increases. Both curves decrease steeply with the increasing frequency threshold. The time requirement is much smaller for parallel episodes than for serial episodes with the same threshold. There are two reasons for this. The parallel episodes are considerably shorter (see Tables 1 and 2) and hence, fewer database passes are needed. The complexity of recognizing injective parallel episodes is also smaller. Episodes Frequency threshold Figure Number of frequent serial (solid line) and injective parallel (dotted line) episodes as a function of the frequency threshold with database, window width Time Frequency threshold Figure 7: Processing time for serial (solid line) and injective parallel (dot- ted line) episodes as a function of the frequency threshold; Winepi, alarm database, window width 5.2 Quality of candidate generation We now take a closer look at the candidates considered and frequent episodes found during the iterations of the procedure. As an example, let us look at what happens during the -rst iterations. Statistics of the -rst ten iterations Episode Episodes Candidates Frequent Match size episodes Table 3: Number of candidate and frequent serial episodes during the -rst ten iteration phases with Winepi; alarm database, frequency threshold 0:001, window width of a run with a frequency threshold of 0.001 and a window width of 60 s is shown in Table 3. The three -rst iterations dominate the behavior of the method. During these phases, the number of candidates is large, and only a small fraction (less than 20 per cent) of the candidates turns out to be frequent. After the third phase the candidate generation is eOEcient, few of the candidates are found infrequent, and although the total number of iteration phases is 45, the last 35 iterations involve only 1-3 candidates each. Thus we could safely combine several of the later iteration steps, to reduce the number of database passes. If we take a closer look at the frequent episodes, we observe that all frequent episodes longer than 7 events consist of repeating occurrences of two very frequent alarms. Each of these two alarms occurs in the database more than 12000 times (16 per cent of the events each). Support Candidates Frequent Iterations Total threshold Episodes time 50 12732 2735 83 28 500 813 138 1000 589 92 48 14 2000 405 64 Table 4: Performance characteristics for serial episodes with Minepi; alarm database, maximum time bound 60 s. Support Candidates Frequent Iterations Total threshold Episodes time 100 4376 1755 71 20 500 633 138 1000 480 89 48 12 2000 378 66 Table 5: Performance characteristics for parallel episodes with Minepi; alarm database, maximum time bound 60 s. 5.3 Comparison of algorithms Winepi and Minepi Tables 4 and 5 represent performance statistics for -nding frequent episodes with Minepi, the method using minimal occurences. Compared to the corresponding -gures for Winepi in Tables 1 and 2, we observe the same general tendency for a rapidly decreasing number of candidates and episodes, as the support threshold increases. The episodes found by Winepi and Minepi are not necessarily the same. If we compare the cases in Tables 1 and 4 with approximately the same number of frequent episodes, e.g., 151 serial episodes for Winepi and 138 for Time Support threshold Figure 8: Processing time for serial (solid line) and injective parallel (dotted line) episodes with Minepi; alarm database, maximum time bound 60 s. Minepi, we notice that they do not correspond to the same episodes. The sizes of the longest freuquent episodes are somewhat dioeerent (43 for the original, 48 for the minimal occurrence method). The frequency threshold corresponds, at the minimum, to about 150 instances of the episode, while the support threshold used for Minepi is 500. The dioeerence between the methods is very clear for small episodes. Consider an episode ff consisting of just one event A. Winepi considers a single event A to occur in windows of width 60 s, while Minepi sees only one minimal occurrence. On the other hand, two successive events of type A result in ff occuring in 61 windows, but the number of minimal occurrences is doubled from 1 to 2. Figure 8 shows the time requirement for -nding frequent episodes with Minepi. The processing time for Minepi reaches a plateau when the size of the maximal episodes no longer changes (in this case, at support threshold 500). The behavior is similar for serial and parallel episodes. The time requirements of Minepi should not be directly compared to Winepi: the episodes discovered are dioeerent, and our implementation of Minepi works entirely in the main memory. With very large databases this might not be Varying support threshold, four time bounds Support Distinct Rule gen. threshold rules time 50 50470 149 1000 1221 15 2000 1082 14 4000 1005 14 Varying number of time bounds, support threshold 1000 Number of All Rule gen. time bounds rules time Table Number of rules and rule generation time with Minepi; alarm data- base, serial episodes, support threshold 1000, maximum time bound 60 s, con-dence threshold 0. possible during the -rst iterations; either the minimal occurrences need to be stored on the disk, or other methods (e.g., variants of Algorithms 4 and must be used. 5.4 Rules The methods can easily produce very large amounts of rules. Recall that rules are constructed by considering all frequent episodes ff as the right-hand side and all subepisodes fi - ff as the left-hand side of the rule. Additionally, considers variations of these rules with all the time bounds in the given set W. Table 6 represents results with serial episodes. The initial episode generation with Minepi took around 14 s, and the total number of frequent episodes was 92. The table shows the number of rules obtained by Minepi with con- -dence threshold 0 and with maximum time bound 60 s. On the left, with a varying support threshold, rules that dioeer only in their time bounds are excluded from the -gures; the rule generation time is, however, obtained by generating rules with four dioeerent time bounds. Rules Confindence threshold Figure 9: Total number of distinct rules found by Minepi with various con- -dence thresholds; alarm database, maximum time bound 60 s, support threshold 100. The minimal occurrence method is particularly useful, if we are interested in -nding rules with several dioeerent time bounds. The right side of Table 6 represents performance results with a varying number of time bounds. The time requirement increases slowly as more time bounds are used, and the time increases slowlier than the number of rules. The amount of almost 80000 rules, obtained with may seem unnecessarily large and unjusti-ed. Remember, however, that there are only 1221 distinct rules. The rest of the rules present dioeerent combinations of time bounds, in this case down to the granularity of one second. For the cost of 43 s we thus obtain very -ne-grained rules from our frequent episodes. Dioeerent criteria can then be used to select the most interesting rules from these. Figure 9 represents the eoeect of the con-dence threshold to the number of distinct rules found by Minepi. Although the initial number of rules may be quite large, it decreases fairly rapidly if we require a reasonable con-dence. Data set Events Event Supp. Max Conf. Freq. Rules name types thr. time b. thr. epis. alarms 73679 287 100 protein Table 7: Characteristic parameter values for each of the data sets and the number of episodes and rules found by Minepi. 5.5 Results with dioeerent data sets In addition to the experiments on the alarm database, we have run Minepi on a variety of dioeerent data collections to get a better view of the usefulness of the method. The data collections that were used, some typical parameter values for them, and some results are presented in Table 7. The WWW data is part of the WWW server log from the Department of Computer Science at the University of Helsinki. The log contains requests to WWW pages at the department's server; such requests can be made by WWW browsers at any host in the Internet. We consider the WWW page fetched as the event type. The number of events in our data set is 116308, covering three weeks in February and March, 1996. In total, 7634 dioeerent pages are referred to. Requests for images have been excluded from consideration Suitable support thresholds vary a lot, depending on the number of events and the distribution of event types. A suitable maximum time bound for the device generated alarm data is one minute, while the slower pace of a human user requires using a larger time bound (two minutes or more) for the WWW log. By using a relatively small time bound we reduce the probability of unrelated requests contributing to the support. A low con-dence threshold for the WWW log is justi-ed since we are interested in all fairly usual patterns of usage, not only in the dominating ones. In the WWW server log we found, e.g., long often used paths of pages from the home page of the department to the pages of individual courses. Such behavior suggests that rather than using a bookmark directly to the home page of a course, many users quickly navigate there from the departmental home page. The two text data collections are modi-cations of the same English text. Each word is considered an event, and the words are indexed consecutively to give a itimej for each event. The end of each sentence causes a gap in the indexing scheme, to correspond to a longer distance between words in dioeerent sentences. We used text from GNU man pages (the gnu awk manual). The size of the original text (text1) is 5417 words, and the size of the condensed text -le (text2), where noninformative words such as articles, prepositions, and conjunctions, have been stripped ooe, is 2871 words. The number of dioeerent words in the original and the condensed text is 1102, resp. 905. For text analysis, there is no point in using large time bounds, since it is unlikely that there is any connection between words that are not fairly close to each other. This can be clearly seen in Figure 10 which represents the number of episodes found on various window widths using Winepi. This -gure reveals behavior that is distinctively dioeerent from the corresponding Figure 5 for the alarm database. We observe that for the text data, the window widths from 24 to 50 produce practically the same amount of serial episodes. The number of episodes will only increase with considerably larger window widths. For this data, the interesting frequent episodes are smaller than 24, while the episodes found with much larger window widths are noise. The same phenomenon can be observed for parallel episodes. Only few rules can be found in text using a simple analysis like this. The strongest rules in the original text involve either the word igawkj, or Episodes Window width Figure 10: Number of serial (solid line) and injective parallel (dotted line) episodes as a function of the window width; Winepi, compressed text data (text2), frequency threshold 0:02. common phrases such as the, value [2] ) of [3] (con-dence 0.90) meaning that in 90 % of the cases where the words ithe valuej are consec- utive, they are immediately followed by the preposition iofj. These rules were not found in the condensed text since all prepositions and articles have been stripped ooe. The few rules in the condensed text contain multiple occurrences of the word igawkj, or combinations of words occurring in the header of each man page, such as ifree softwarej. We performed scale-up tests with 5, 10, and 20 fold multiples of the compressed text -le, i.e., sequences of approximately 2900 to 58000 events. The results in Figure 11 show that the time requirement is roughly linear with respect to the length of the input sequence, as could be expected. Finally, we experimented with protein sequences. We used data in the database [1] of the ExPASy WWW molecular biology server of the Geneva University Hospital and the University of Geneva [11]. PROSITE contains biologically signi-cant DNA and protein patterns that help to Time Relative size of database r r r r Figure 11: Scale-up results for serial (solid line) and injective parallel (dotted line) episodes with Minepi; compressed text data, maximum time bound 60, threshold 10 for the smallest -le (n-fold for the larger -les). identify to which family of protein (if any) a new sequence belongs. The purpose of our experiment is to evaluate our algorithm against an external data collection and patterns that are known to exist, not to -nd patterns previously unknown to the biologists. We selected as our target a family of 7 sequences (iDNA mismatch repair proteins 1j, PROSITE entry PS00058). The sequences in the family are known to contain the string GFRGEAL of seven consequtive symbols. We transformed the data in a manner similar to the English text: symbols are indexed consecutively, and between the protein sequences we place a gap. The total length of this data set is 4941 events, with an alphabet of 22 event types. The method could be easily modi-ed to take several separage sequences as input, and to compute the support of an ff, e.g., as the number of input sequences that contain a (minimal) occurrence of ff of length at most the maximum time bound. The parameter values for the protein database are chosen on purpose to reveal the pattern that is known to be present in the database. The window width was selected to be 10, i.e., slightly larger than the length of the pattern that we were looking for, and the support threshold was set to 7, for the seven individual sequences in the original data. With this data, we are only interested in the longest episodes (of length 7 or longer). Of the more than 20000 episodes found, 17 episodes are of length 7 or 8. As expected, these contain the sequence GFRGEAL that was known to be in the database. The longer episodes are variants of this pattern with an eighth symbol fairly near, but not necessarily immediately subsequent to the pattern (e.g., GFRGEAL*S). These types of patterns belong to the pattern class used in PROSITE but, to our suprise, these longer patterns are not reported in the PROSITE database. 6 Extensions and related work The task of discovering frequent parallel episodes can be stated as a task of discovering all frequent sets, a central phase of discovering assocation rules [2], the rule generation methods are also basically the same for association rules and Winepi. The levelwise main algorithm has also been used successfully in the search of frequent sets [3]. Technical problems related to the recognition of episodes have been researched in several -elds. Taking advantage of the slowly changing contents of the group of recent events has been studied, e.g., in arti-cial intelligence, where a similar problem in spirit is the many pattern/many object pattern match problem in production system interpreters [9]. Also, comparable strategies using a sliding window have been used, e.g., to study the locality of reference in virtual memory [7]. Our setting dioeers from these in that our window is a queue with the special property that we know in advance when an event will leave the window; this knowledge is used by Winepi in the recognition of serial episodes. In Minepi, we take advantage of the fact that we know where subepisodes of candidates have occurred. The recent work on sequence data in databases (see [21]) provides interesting openings towards the use of database techniques in the processing of queries on sequences. A problem similar to the computation of frequencies occurs also in the area of active databases. There triggers can be speci-ed as composite events, somewhat similar to episodes. In [10] it is shown how -nite automata can be constructed from composite events to recognize when a trigger should be -red. This method is not practical for episodes since the deterministic automata could be very large. The methods for matching sets of episodes against a sequence have some similarities to the algorithms used in string matching (e.g., [12]). In par- ticular, recognizing serial episodes in a sequence can be seen as locating all occurrences of subsequences, or matches of patterns with variable length don't care symbols, where the length of the occurrences is limited by the window width. Learning from a set of sequences has received considerable interest in the -eld of bioinformatics, where an interesting problem is the discovery of patterns common to a set of related protein or amino acid se- quences. The classes of patterns dioeer from ours; they can be, e.g., substrings with -xed length don't care symbols [15]. Closer to our patterns are those considered in [24]. The described algorithm -nds patterns that are similar to serial episodes; however, the patterns have a given minimum length, and the occurrences can be within a given edit distance. Recent results on the pattern matching aspects of recognizing episodes can be found in [6]. The work most closely related to ours is perhaps [4]. There multiple sequences are searched for patterns that are similar to the serial episodes with some extra restrictions and an event taxonomy. Our methods can be extended with a taxonomy by a direct application of the similar extensions to association rules [13, 14, 22]. Also, our methods can be applied on analyzing several sequencies; there is actually a variety of choices for the de-nition of frequency of an episode in a set of sequencies. More recently, the pattern class of [4] has been extended with windowing, some extra time constraints, and an event taxonomy [23]. - For a survey on patterns in sequential data, see [17]. In stochastics, event sequence data is often called a marked point process [16]. It should be noted that traditional methods for analyzing marked point processes are ill suited for the cases where the number of event types is large. However, there exists an interesting combination of techniques: frequent episodes are discovered -rst, and then the phenomena they describe are analyzed in more detail with methods for marked point processes. There are also some interesting similarities between the discovery of frequent episodes and the work done on inductive logic programming (see, e.g., [20]); a noticeable dioeerence is caused by the sequentiality of the underlying data model, and the emphasis on time-limited occurrences. Similarly, the problem of looking for one occurrence of an episode can be viewed as a constraint satisfaction problem. The class of patterns discovered can be easily modi-ed in several direc- tions. Dioeerent windowing strategies could be used, e.g., considering only windows starting every win 0 time units for some win 0 , or windows starting from every event. Other types of patterns could also be searched for, e.g., substrings with -xed length don't care symbols; searching for episodes in several sequences is no problem. A more general framework for episode discovery has been presented in [18]. There episodes are de-ned as combinations of events satisfying certain user speci-ed unary of binary conditions. Conclusions We presented a framework for discovering frequent episodes in sequential data. The framework consists of de-ning episodes as partially ordered sets of events, and looking at windows on the sequence. We described an al- gorithm, Winepi, for -nding all episodes from a given class of episodes that are frequent enough. The algorithm was based on the discovery of episodes by only considering an episode when all its subepisodes are frequent, and on incremental checking of whether an episode occurs in a window. The implementation shows that the method is eOEcient. We have applied the method in the analysis of the alarm AEow from telecommunication networks, and discovered episodes have been embedded in alarm handling software. We also presented an alternative approach, Minepi, to the discovery of frequent episodes, based on minimal occurrences of episodes. This approach supplies more power for representing connections between events, as it produces rules with two time bounds. Both rule formalisms have their advantages. While the rules of Minepi are often more informative, the frequencies and con-dences of the rules of have nice interpretations as probabilities concerning randomly chosen windows. For a large part the algorithms are similar, there are signi-cant dioeerences only in the computation of the frequency or support. Roughly, a general tendency in the performance is that Winepi can be more eOEcient in the -rst phases of the discovery, mostly due to smaller space requirement. In the later iterations, Minepi is likely to outperform Winepi clearly. The methods can be modi-ed for cross-use, i.e., Winepi for -nding minimal occurrences and Minepi for counting windows, and for some large problems - whether the rule type of Winepi or Minepi - a mixture of the two methods could give better performance than either alone. Interesting extensions to the work presented here are facilities for rule querying and compilation, i.e., methods by which the user could specify the class in high-level language and the de-nition would automatically be compiled into a specialization of the algorithm that would take advantage of the restrictions on the episode class. Other open problems include the combination of episode techniques with marked point processes and intensity models. --R Mining association rules between sets of items in large databases. Fast discovery of association rules. Mining sequential patterns. Testing complex temporal relationships involving multiple granularities and its application to data mining. The working set model of program behavior. Situation recognition: Representation and algorithms. A fast algorithm for the many pattern/many object pattern match problem. Composite event speci-ca- tion in active databases Geneva University Hospital and University of Geneva Simple and eOEcient string matching with k mismatches. Discovery of multiple-level association rules from large databases A perspective on databases and data mining. Finding AEexible patterns in unaligned protein sequences. The Statistical Analysis of Failure Time Data. Identifying and using patterns in sequential data. Discovering generalized episodes using minimal occurrences. An algebraic formulation of temporal knowledge for reasoning about recurring events. Inductive Logic Programming. Mining generalized association rules. Mining sequential patterns: Generalizations and performance improvements. Combinatorial pattern discovery for scienti-c data: Some preliminary results --TR --CTR Srivatsan Laxman , P. S. Sastry , K. P. Unnikrishnan, Discovering Frequent Episodes and Learning Hidden Markov Models: A Formal Connection, IEEE Transactions on Knowledge and Data Engineering, v.17 n.11, p.1505-1517, November 2005 Jia-Wei Han , Jian Pei , Xi-Feng Yan, From sequential pattern mining to structured pattern mining: a pattern-growth approach, Journal of Computer Science and Technology, v.19 n.3, p.257-279, May 2004 Sandra de Amo , Daniel A. Furtado, First-order temporal pattern mining with regular expression constraints, Data & Knowledge Engineering, v.62 n.3, p.401-420, September, 2007 Tomoyuki Morita , Yasushi Hirano , Yasuyuki Sumi , Shoji Kajita , Kenji Mase, A pattern mining method for interpretation of interaction, Proceedings of the 7th international conference on Multimodal interfaces, October 04-06, 2005, Torento, Italy S. Parthasarathy , M. J. Zaki , M. Ogihara , S. Dwarkadas, Incremental and interactive sequence mining, Proceedings of the eighth international conference on Information and knowledge management, p.251-258, November 02-06, 1999, Kansas City, Missouri, United States Jian Pei , Jiawei Han , Wei Wang, Mining sequential patterns with constraints in large databases, Proceedings of the eleventh international conference on Information and knowledge management, November 04-09, 2002, McLean, Virginia, USA Jiawei Han , Jian Pei , Behzad Mortazavi-Asl , Qiming Chen , Umeshwar Dayal , Mei-Chun Hsu, FreeSpan: frequent pattern-projected sequential pattern mining, Proceedings of the sixth ACM SIGKDD international conference on Knowledge discovery and data mining, p.355-359, August 20-23, 2000, Boston, Massachusetts, United States Mohammed J. Zaki, Sequence mining in categorical domains: incorporating constraints, Proceedings of the ninth international conference on Information and knowledge management, p.422-429, November 06-11, 2000, McLean, Virginia, United States Ann Devitt , Joseph Duffin , Robert Moloney, Topographical proximity for mining network alarm data, Proceeding of the 2005 ACM SIGCOMM workshop on Mining network data, August 26-26, 2005, Philadelphia, Pennsylvania, USA Liping Ji , Kian-Lee Tan , Anthony K. H. Tung, Mining frequent closed cubes in 3D datasets, Proceedings of the 32nd international conference on Very large data bases, September 12-15, 2006, Seoul, Korea Zdenk Tronek, Episode directed acyclic subsequence graph, Nordic Journal of Computing, v.11 n.1, p.35-40, Spring 2004 Roy Villafane , Kien A. Hua , Duc Tran , Basab Maulik, Knowledge Discovery from Series of Interval Events, Journal of Intelligent Information Systems, v.15 n.1, p.71-89, JulyAug. 2000 Katharina Morik, Applications of knowledge discovery, Proceedings of the 18th international conference on Innovations in Applied Artificial Intelligence, p.1-5, June 22-24, 2005, Bari, Italy Ming-Yen Lin , Suh-Yin Lee, Interactive sequence discovery by incremental mining, Information SciencesInformatics and Computer Science: An International Journal, v.165 n.3-4, p.187-205, 19 October 2004 Kenji Yamanishi , Yuko Maruyama, Dynamic syslog mining for network failure monitoring, Proceeding of the eleventh ACM SIGKDD international conference on Knowledge discovery in data mining, August 21-24, 2005, Chicago, Illinois, USA Heikki Mannila , Christopher Meek, Global partial orders from sequential data, Proceedings of the sixth ACM SIGKDD international conference on Knowledge discovery and data mining, p.161-168, August 20-23, 2000, Boston, Massachusetts, United States Ming-Yen Lin , Suh-Yin Lee, Efficient mining of sequential patterns with time constraints by delimited pattern growth, Knowledge and Information Systems, v.7 n.4, p.499-514, May 2005 Anthony K.H. Tung , Hongjun Lu , Jiawei Han , Ling Feng, Breaking the barrier of transactions: mining inter-transaction association rules, Proceedings of the fifth ACM SIGKDD international conference on Knowledge discovery and data mining, p.297-301, August 15-18, 1999, San Diego, California, United States Helen Pinto , Jiawei Han , Jian Pei , Ke Wang , Qiming Chen , Umeshwar Dayal, Multi-dimensional sequential pattern mining, Proceedings of the tenth international conference on Information and knowledge management, October 05-10, 2001, Atlanta, Georgia, USA Minghua Zhang , Ben Kao , David W. Cheung , Kevin Y. Yip, Mining periodic patterns with gap requirement from sequences, Proceedings of the 2005 ACM SIGMOD international conference on Management of data, June 14-16, 2005, Baltimore, Maryland Ke Wang , Yabo Xu , Jeffrey Xu Yu, Scalable sequential pattern mining for biological sequences, Proceedings of the thirteenth ACM international conference on Information and knowledge management, November 08-13, 2004, Washington, D.C., USA Robert Gwadera , Mikhail J. Atallah , Wojciech Szpankowski, Reliable detection of episodes in event sequences, Knowledge and Information Systems, v.7 n.4, p.415-437, May 2005 Alberto Apostolico , Mikhail J. Atallah, Compact recognizers of episode sequences, Information and Computation, v.174 n.2, p.180-192, May 1 2002 Gong Chen , Xindong Wu , Xingquan Zhu, Mining sequential patterns across time sequences, New Generation Computing, v.26 n.1, p.75-96, January 2008 Foto Afrati , Aristides Gionis , Heikki Mannila, Approximating a collection of frequent sets, Proceedings of the tenth ACM SIGKDD international conference on Knowledge discovery and data mining, August 22-25, 2004, Seattle, WA, USA Kuo-Yu Huang , Chia-Hui Chang, SMCA: A General Model for Mining Asynchronous Periodic Patterns in Temporal Databases, IEEE Transactions on Knowledge and Data Engineering, v.17 n.6, p.774-785, June 2005 Ahmed Metwally , Divyakant Agrawal , Amr El Abbadi, Using association rules for fraud detection in web advertising networks, Proceedings of the 31st international conference on Very large data bases, August 30-September 02, 2005, Trondheim, Norway Mohammad El-Ramly , Eleni Stroulia , Paul Sorenson, Recovering software requirements from system-user interaction traces, Proceedings of the 14th international conference on Software engineering and knowledge engineering, July 15-19, 2002, Ischia, Italy Jian Pei , Jiawei Han , Wei Wang, Constraint-based sequential pattern mining: the pattern-growth methods, Journal of Intelligent Information Systems, v.28 n.2, p.133-160, April 2007 Chang-Shing Lee , Yuan-Fang Kao , Yau-Hwang Kuo , Mei-Hui Wang, Automated ontology construction for unstructured text documents, Data & Knowledge Engineering, v.60 n.3, p.547-566, March, 2007 R. B. V. Subramanyam , A. Goswami, A fuzzy data mining algorithm for incremental mining of quantitative sequential patterns, International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems, v.13 n.6, p.633-652, December 2005 Wei Wang , Jiong Yang , Philip S. Yu, Mining patterns in long sequential data with noise, ACM SIGKDD Explorations Newsletter, v.2 n.2, p.28-33, Dec. 2000 Jian Pei , Jiawei Han , Behzad Mortazavi-Asl , Jianyong Wang , Helen Pinto , Qiming Chen , Umeshwar Dayal , Mei-Chun Hsu, Mining Sequential Patterns by Pattern-Growth: The PrefixSpan Approach, IEEE Transactions on Knowledge and Data Engineering, v.16 n.11, p.1424-1440, November 2004 Andrzej Skowron , Piotr Synak, Reasoning in information maps, Fundamenta Informaticae, v.59 n.2-3, p.241-259, February 2004 Gregory Buehrer , Srinivasan Parthasarathy , Amol Ghoting, Out-of-core frequent pattern mining on a commodity PC, Proceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data mining, August 20-23, 2006, Philadelphia, PA, USA Hewijin Christine Jiau , Yi-Jen Su , Yeou-Min Lin , Shang-Rong Tsai, MPM: a hierarchical clustering algorithm using matrix partitioning method for non-numeric data, Journal of Intelligent Information Systems, v.26 n.2, p.185-207, March 2006 Andrzej Skowron , Piotr Synak, Reasoning in Information Maps, Fundamenta Informaticae, v.59 n.2-3, p.241-259, April 2004 Dong Xin , Jiawei Han , Xifeng Yan , Hong Cheng, On compressing frequent patterns, Data & Knowledge Engineering, v.60 n.1, p.5-29, January, 2007 Jiawei Han , Jian Pei , Yiwen Yin, Mining frequent patterns without candidate generation, ACM SIGMOD Record, v.29 n.2, p.1-12, June 2000 Jochen Hipp , Ulrich Gntzer , Gholamreza Nakhaeizadeh, Algorithms for association rule mining a general survey and comparison, ACM SIGKDD Explorations Newsletter, v.2 n.1, p.58-64, June, 2000 Jiawei Han , Jian Pei , Yiwen Yin , Runying Mao, Mining Frequent Patterns without Candidate Generation: A Frequent-Pattern Tree Approach, Data Mining and Knowledge Discovery, v.8 n.1, p.53-87, January 2004 Jian Pei , Jiawei Han , Laks V. S. Lakshmanan, Pushing Convertible Constraints in Frequent Itemset Mining, Data Mining and Knowledge Discovery, v.8 n.3, p.227-252, May 2004 Aristides Gionis , Teija Kujala , Heikki Mannila, Fragments of order, Proceedings of the ninth ACM SIGKDD international conference on Knowledge discovery and data mining, August 24-27, 2003, Washington, D.C. Dong Xin , Jiawei Han , Xifeng Yan , Hong Cheng, Mining compressed frequent-pattern sets, Proceedings of the 31st international conference on Very large data bases, August 30-September 02, 2005, Trondheim, Norway Valery Guralnik , Jaideep Srivastava, Event detection from time series data, Proceedings of the fifth ACM SIGKDD international conference on Knowledge discovery and data mining, p.33-42, August 15-18, 1999, San Diego, California, United States Gabriela Guimares , Lus Moniz Pereira, Inferring definite-clause grammars to express multivariate time series, Proceedings of the 18th international conference on Innovations in Applied Artificial Intelligence, p.332-341, June 22-24, 2005, Bari, Italy Yves Bastide , Rafik Taouil , Nicolas Pasquier , Gerd Stumme , Lotfi Lakhal, Mining frequent patterns with counting inference, ACM SIGKDD Explorations Newsletter, v.2 n.2, p.66-75, Dec. 2000 Yida Wang , Ee-Peng Lim , San-Yih Hwang, Efficient mining of group patterns from user movement data, Data & Knowledge Engineering, v.57 n.3, p.240-282, June 2006 Anthony K. H. Tung , Hongjun Lu , Jiawei Han , Ling Feng, Efficient Mining of Intertransaction Association Rules, IEEE Transactions on Knowledge and Data Engineering, v.15 n.1, p.43-56, January Tu-Bao Ho , Canh-Hao Nguyen , Saori Kawasaki , Si-Quang Le , Katsuhiko Takabayashi, Exploiting temporal relations in mining hepatitis data, New Generation Computing, v.25 n.3, p.247-262, January 2007 Chin-Chen Chang , Chih-Yang Lin , Henry Chou, Perfect hashing schemes for mining traversal patterns, Fundamenta Informaticae, v.70 n.3, p.185-202, April 2006 Ming-Yen Lin , Suh-Yin Lee, Incremental update on sequential patterns in large databases by implicit merging and efficient counting, Information Systems, v.29 n.5, p.385-404, July 2004 R. Anderson , Pedro Domingos , Daniel S. Weld, Personalizing web sites for mobile users, Proceedings of the 10th international conference on World Wide Web, p.565-575, May 01-05, 2001, Hong Kong, Hong Kong Richard Relue , Xindong Wu , Hao Huang, Efficient runtime generation of association rules, Proceedings of the tenth international conference on Information and knowledge management, October 05-10, 2001, Atlanta, Georgia, USA Mohammad El-Ramly , Eleni Stroulia , Paul Sorenson, From run-time behavior to usage scenarios: an interaction-pattern mining approach, Proceedings of the eighth ACM SIGKDD international conference on Knowledge discovery and data mining, July 23-26, 2002, Edmonton, Alberta, Canada Gregory Buehrer , Srinivasan Parthasarathy , Shirish Tatikonda , Tahsin Kurc , Joel Saltz, Toward terabyte pattern mining: an architecture-conscious solution, Proceedings of the 12th ACM SIGPLAN symposium on Principles and practice of parallel programming, March 14-17, 2007, San Jose, California, USA Akihiro Inokuchi , Takashi Washio , Hiroshi Motoda, Complete Mining of Frequent Patterns from Graphs: Mining Graph Data, Machine Learning, v.50 n.3, p.321-354, March Jiong Yang , Wei Wang , Philip S. Yu , Jiawei Han, Mining long sequential patterns in a noisy environment, Proceedings of the 2002 ACM SIGMOD international conference on Management of data, June 03-06, 2002, Madison, Wisconsin Jun Wook Lee , Ok Hyun Paek , Keun Ho Ryu, Temporal moving pattern mining for location-based service, Journal of Systems and Software, v.73 n.3, p.481-490, November-December 2004 Carl H. Mooney , John F. Roddick, Marking time in sequence mining, Proceedings of the fifth Australasian conference on Data mining and analystics, p.129-134, November 29-30, 2006, Sydney, Australia Wenke Lee , Salvatore J. Stolfo , Kui W. Mok, Algorithms for mining system audit data, Data mining, rough sets and granular computing, Physica-Verlag GmbH, Heidelberg, Germany, 2002 Jiong Yang , Wei Wang , Philip S. Yu, Mining Asynchronous Periodic Patterns in Time Series Data, IEEE Transactions on Knowledge and Data Engineering, v.15 n.3, p.613-628, March Jian Pei , Jiawei Han, Constrained frequent pattern mining: a pattern-growth view, ACM SIGKDD Explorations Newsletter, v.4 n.1, June 2002 Jiawei Han , Jian Pei, Mining frequent patterns by pattern-growth: methodology and implications, ACM SIGKDD Explorations Newsletter, v.2 n.2, p.14-20, Dec. 2000 Joo B. D. Cabrera , Lundy Lewis , Xinzhou Qin , Wenke Lee , Raman K. Mehra, Proactive Intrusion Detection and Distributed Denial of Service AttacksA Case Study in Security Management, Journal of Network and Systems Management, v.10 n.2, p.225-254, June 2002 Tamas Abraham, Event sequence mining to develop profiles for computer forensic investigation purposes, Proceedings of the 2006 Australasian workshops on Grid computing and e-research, p.145-153, January 16-19, 2006, Hobart, Tasmania, Australia Wei-Guang Teng , Ming-Syan Chen , Philip S. Yu, A regression-based temporal pattern mining scheme for data streams, Proceedings of the 29th international conference on Very large data bases, p.93-104, September 09-12, 2003, Berlin, Germany Gerhard Widmer , Simon Dixon , Werner Goebl , Elias Pampalk , Asmir Tobudic, In search of the Horowitz factor, AI Magazine, v.24 n.3, p.111-130, September Magnus Lie Hetland , Pl Strom, Evolutionary Rule Mining in Time Series Databases, Machine Learning, v.58 n.2-3, p.107-125, February 2005 Gosta Grahne , Jianfei Zhu, Fast Algorithms for Frequent Itemset Mining Using FP-Trees, IEEE Transactions on Knowledge and Data Engineering, v.17 n.10, p.1347-1362, October 2005 Theodore Johnson , Laks V. S. Lakshmanan , Raymond T. Ng, The 3W Model and Algebra for Unified Data Mining, Proceedings of the 26th International Conference on Very Large Data Bases, p.21-32, September 10-14, 2000 Jian Pei , Guozhu Dong , Wei Zou , Jiawei Han, Mining condensed frequent-pattern bases, Knowledge and Information Systems, v.6 n.5, p.570-594, September 2004 Gui-Rong Xue , Hua-Jun Zeng , Zheng Chen , Wei-Ying Ma , Hong-Jiang Zhang , Chao-Jun Lu, Implicit link analysis for small web search, Proceedings of the 26th annual international ACM SIGIR conference on Research and development in informaion retrieval, July 28-August 01, 2003, Toronto, Canada Yen-Liang Chen , Shih-Sheng Chen , Ping-Yu Hsu, Mining hybrid sequential patterns and sequential rules, Information Systems, v.27 n.5, p.345-362, July 2002 Mika Klemettinen , Heikki Mannila , Hannu Toivonen, Rule Discovery in Telecommunication AlarmData, Journal of Network and Systems Management, v.7 n.4, p.395-423, 7 Amol Ghoting , Gregory Buehrer , Srinivasan Parthasarathy , Daehyun Kim , Anthony Nguyen , Yen-Kuang Chen , Pradeep Dubey, Cache-conscious frequent pattern mining on a modern processor, Proceedings of the 31st international conference on Very large data bases, August 30-September 02, 2005, Trondheim, Norway Yen-Liang Chen , Ya-Han Hu, Constraint-based sequential pattern mining: the consideration of recency and compactness, Decision Support Systems, v.42 n.2, p.1203-1215, November 2006 Shao-Shin Hung , Ting-Chia Kuo , Damon Shing-Min Liu, An Efficient Mining and Clustering Algorithm for Interactive Walk-Through Traversal Patterns, Proceedings of the 2004 IEEE/WIC/ACM International Conference on Web Intelligence, p.356-362, September 20-24, 2004 Usama Fayyad, Knowledge discovery in databases: An overview, Relational Data Mining, Springer-Verlag New York, Inc., New York, NY, 2001 Yun Chi , Richard R. Muntz , Siegfried Nijssen , Joost N. Kok, Frequent Subtree Mining - An Overview, Fundamenta Informaticae, v.66 n.1-2, p.161-198, January 2005 E. Boros , V. Gurvich , L. Khachiyan , K. Makino, On Maximal Frequent and Minimal Infrequent Sets in Binary Matrices, Annals of Mathematics and Artificial Intelligence, v.39 n.3, p.211-221, November Alexandros Nanopoulos , Dimitrios Katsaros , Yannis Manolopoulos, A Data Mining Algorithm for Generalized Web Prefetching, IEEE Transactions on Knowledge and Data Engineering, v.15 n.5, p.1155-1169, September Taneli Mielikinen , Evimaria Terzi , Panayiotis Tsaparas, Aggregating time partitions, Proceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data mining, August 20-23, 2006, Philadelphia, PA, USA van der Aalst , Ton Weijters , Laura Maruster, Workflow Mining: Discovering Process Models from Event Logs, IEEE Transactions on Knowledge and Data Engineering, v.16 n.9, p.1128-1142, September 2004 Heikki Mannila , Hannu Toivonen, Levelwise Search and Borders of Theories in KnowledgeDiscovery, Data Mining and Knowledge Discovery, v.1 n.3, p.241-258, 1997 Jiong Yang , Wei Wang , Philip S. Yu, Mining Surprising Periodic Patterns, Data Mining and Knowledge Discovery, v.9 n.2, p.189-216, September 2004 Amol Ghoting , Gregory Buehrer , Srinivasan Parthasarathy , Daehyun Kim , Anthony Nguyen , Yen-Kuang Chen , Pradeep Dubey, Cache-conscious frequent pattern mining on modern and emerging processors, The VLDB Journal The International Journal on Very Large Data Bases, v.16 n.1, p.77-96, January 2007 Zhiping Zeng , Jianyong Wang , Lizhu Zhou , George Karypis, Out-of-core coherent closed quasi-clique mining from large dense graph databases, ACM Transactions on Database Systems (TODS), v.32 n.2, p.13-es, June 2007 David Al-Dabass , Evtim Peytchev , Mohamed Khalil , Manling Ren, Scalability issues in urban traffic systems, Proceedings of the 1st international conference on Scalable information systems, p.31-es, May 30-June 01, 2006, Hong Kong Klaus Julisch , Marc Dacier, Mining intrusion detection alarms for actionable knowledge, Proceedings of the eighth ACM SIGKDD international conference on Knowledge discovery and data mining, July 23-26, 2002, Edmonton, Alberta, Canada W. M. P. van der Aalst , A. J. M. M. Weijters, Process mining: a research agenda, Computers in Industry, v.53 n.3, p.231-244, April 2004 Ling Feng , Jeffrey Xu Yu , Hongjun Lu , Jiawei Han, A template model for multidimensional inter-transactional association rules, The VLDB Journal The International Journal on Very Large Data Bases, v.11 n.2, p.153-175, October 2002 Qing Li , Ling Feng , Allan Wong, From intra-transaction to generalized inter-transaction: landscaping multidimensional contexts in association rule mining, Information SciencesInformatics and Computer Science: An International Journal, v.172 n.3-4, p.361-395, 9 June 2005 Luc Dehaspe , Hannu Toivonen, Discovery of frequent DATALOG patterns, Data Mining and Knowledge Discovery, v.3 n.1, p.7-36, March 1999 Taneli Mielikinen, Frequency-based views to pattern collections, Discrete Applied Mathematics, v.154 n.7, p.1113-1139, 1 May 2006 Jos L. Balczar , Gemma C. Garriga, Horn axiomatizations for sequential data, Theoretical Computer Science, v.371 n.3, p.247-264, March, 2007 Xingquan Zhu , Xindong Wu , Ahmed K. Elmagarmid , Zhe Feng , Lide Wu, Video Data Mining: Semantic Indexing and Event Detection from the Association Perspective, IEEE Transactions on Knowledge and Data Engineering, v.17 n.5, p.665-677, May 2005 Jiawei Han , Laks V. S. Lakshmanan , Jian Pei, Scalable frequent-pattern mining methods: an overview, Tutorial notes of the seventh ACM SIGKDD international conference on Knowledge discovery and data mining, August 26-29, 2001, San Francisco, California Vipin Kumar , Mohammed Zaki, High performance data mining (tutorial PM-3), Tutorial notes of the sixth ACM SIGKDD international conference on Knowledge discovery and data mining, p.309-425, August 20-23, 2000, Boston, Massachusetts, United States
frequent episodes;event sequences;sequence analysis
593450
Adaptive Fraud Detection.
One method for detecting fraud is to check for suspicious changes in user behavior. This paper describes the automatic design of user profiling methods for the purpose of fraud detection, using a series of data mining techniques. Specifically, we use a rule-learning program to uncover indicators of fraudulent behavior from a large database of customer transactions. Then the indicators are used to create a set of monitors, which profile legitimate customer behavior and indicate anomalies. Finally, the outputs of the monitors are used as features in a system that learns to combine evidence to generate high-confidence alarms. The system has been applied to the problem of detecting cellular cloning fraud based on a database of call records. Experiments indicate that this automatic approach performs better than hand-crafted methods for detecting fraud. Furthermore, this approach can adapt to the changing conditions typical of fraud detection environments.
Introduction In the United States, cellular fraud costs the telecommunications industry hundreds of millions of dollars per year (Walters and Wilkinson 1994; Steward 1997). One kind of cellular fraud called cloning is particularly expensive and epidemic in major cities throughout the United States. Cloning fraud causes great inconvenience to customers and great expense to cellular service providers. Existing methods for detecting cloning fraud are ad hoc and their evaluation is virtually nonexistent. We have embarked on a program of systematic analysis of cellular call data for the purpose of designing and evaluating methods for detecting fraudulent behavior. Cloning fraud is one instance of superimposition fraud, in which fraudulent usage is superimposed upon (added to) the legitimate usage of an account. Other examples are credit card fraud, calling card fraud and some forms of computer in- trusion. Superimposition fraud typically occurs when a non-legitimate user gains illicit access to the account or service of a legitimate user. Superimposition fraud is detectable if the legitimate users have fairly regular behavior that is generally distinguishable from the fraudulent behavior. This paper presents a framework, and a corresponding system, for automatically generating detectors for superimposition fraud. We have applied the system in the domain of cellular cloning fraud. Under the framework, massive amounts of cellular call data are analyzed in order to determine general patterns of fraud. These patterns are then used to generate a set of monitors, each of which watches customers' behavior with respect to one discovered pattern. A monitor profiles each customer's typical behavior and, in use, measures the extent to which current behavior is abnormal with respect to the monitor's particular pattern. Each monitor's output is provided to a neural network, which weights the values and issues an alarm when the combined evidence for fraud is strong enough. This article is organized as follows. We first describe the problem of cellular cloning fraud and some existing strategies for detecting it. We then describe the framework in detail using examples from the implemented system. We present experimental results comparing the system against other known methods for detecting fraud. Finally, we discuss the evaluation and describe issues in the future of automatic fraud detection. 2. Cellular Communications and Cloning Fraud Whenever a cellular phone is on, it periodically transmits two unique identification numbers: its Mobile Identification Number (MIN) and its Electronic Serial Number (ESN). These two numbers together specify the customer's account. These numbers are broadcast unencrypted over the airwaves, and they can can be received, decoded and stored using special equipment that is relatively inexpensive. 2.1. Cloning Fraud Cloning occurs when a customer's MIN and ESN are programmed into a cellular telephone not belonging to the customer. When this second telephone is used, the network sees the customer's MIN and ESN and subsequently bills the usage to the customer. With the stolen MIN and ESN, a cloned phone user (whom we shall call a bandit) can make virtually unlimited calls, whose charges are billed to the customer. The attraction of free and untraceable communication makes cloned phones very popular in major metropolitan areas. If the fraudulent usage goes undetected, the customer's next bill will include the corresponding charges. Typically, the customer then calls the cellular service provider (the carrier) and denies the usage. The carrier and customer then determine which calls were made by the "bandit" and which were legitimate calls. The fraudulent charges are credited to the customer's account, and measures are taken to prohibit further fraudulent charges. In certain cases, the fraudulent call records will be referred to a law enforcement agency for prosecution. There are two primary motivations for cloning fraud. Obviously, cloning fraud allows low-cost communications. A bandit is not charged for calls, which are usually worth far more (in retail dollars) than the cost of the cloned phone. Less obviously, cloning fraud allows untraceable communications because the bandit's identity cannot be tied to the cloned account. This second aspect is very important to criminals (DeMaria and Gidari 1996). Cloning fraud is detrimental in many ways. First, fraudulent usage congests cell sites, causing service to be denied to legitimate customers. Second, most cellular calls are to non-cellular destinations, so fraud incurs land-line usage charges. Third, cellular carriers must pay costs to other carriers for usage outside the home territory. Because these are retail costs, they constitute a considerable financial burden to the customer's carrier. Fourth, the crediting process is costly to the carrier and inconvenient to the customer; the customer is more likely to switch to another carrier ("customer churn") if the other is perceived to be less susceptible to fraud. For these reasons, cellular carriers have a strong interest in reducing cloning fraud. 2.2. Strategies for dealing with cloning fraud There are two classes of methods for dealing with cloning fraud. Pre-call methods try to identify and block fraudulent calls as they are made. Post-call methods try to identify fraud that has already occurred on an account so that further fraudulent usage can be blocked. 2.2.1. Pre-call methods Pre-call detection methods involve validating the phone or its user when a call is placed. A common method is requiring that a Personal Number (PIN) be entered before every call. A PIN serves as a password that is validated by the switch prior to allowing the call into the network. PINs are in use throughout many metropolitan areas in the United States. Unfor- tunately, like MIN-ESN pairs, PINs are broadcast over the airwaves unencrypted. For technical reasons, PINs are more difficult to receive and decode, but with more sophisticated equipment PIN cracking is possible (Herzog 1995). Although PINs make cloning fraud more difficult, they do not prevent it. Other methods of prevention include RF Fingerprinting and Authentication (Red- den 1996). RF Fingerprinting is a method of identifying cellular phones by their transmission characteristics. Authentication is a reliable and secure private-key encryption method that imposes no inconvenience on the customer. It has been predicted that authentication will eliminate cloning fraud eventually. However, authentication requires changes in hardware: both phones and switches must be capable of processing authentication requests. Currently about thirty million non- authenticatable cell phones are in use in the United States alone, and their replacement will not be immediate (Steward 1997). In the meantime, cloning fraud will continue to be a problem and the industry will rely on post-call fraud detection methods. 2.2.2. Post-call methods Post-call methods periodically analyze call data on each account to determine whether cloning fraud has occurred. One such method, collision detection, involves analyzing call data for temporally overlapping calls. Since a MIN-ESN pair is licensed to only one legitimate user, any simultaneous usage is probably fraudulent. A closely related method, velocity checking (Davis and Goyal 1993), involves analyzing the locations and times of consecutive calls to determine whether a single user could have placed them while traveling at reasonable speeds. For example, if a call is made in Los Angeles 20 minutes after a call is made on the same account in New York, two different people are likely using the account. Date & Time Day Duration Origin Destination Fraud mins Brooklyn, NY Stamford,CT 1/05/95 14:53:27 Fri 5 mins Brooklyn, NY Greenwich,CT 1/08/95 09:42:01 Mon 3 mins Bronx, NY White Plains, NY 1/08/95 15:01:24 Mon 9 mins Brooklyn, NY Brooklyn, NY 1/09/95 15:06:09 Tue 5 mins Manhattan, NY Stamford, CT 1/09/95 16:28:50 Tue 53 sec Brooklyn, NY Brooklyn, NY 1/10/95 01:45:36 Wed 35 sec Boston, MA Chelsea, MA bandit 1/10/95 01:46:29 Wed 34 sec Boston, MA Yonkers, NY bandit 1/10/95 01:50:54 Wed 39 sec Boston, MA Chelsea, MA bandit 1/10/95 11:23:28 Wed 24 sec White Plains, NY Congers, NY 1/11/95 22:00:28 Thu 37 sec Boston, MA East Boston, MA bandit East Boston, MA bandit Figure 1. Call records of a sample frauded account Collisions and velocity checks are both believed to be accurate, but they share the disadvantage that their usefulness depends upon a moderate level of legitimate activity. Low-usage subscribers (for example, people who only use cellular phones in emergencies) will rarely cause collisions or velocity alarms with bandits. Another post-call method, dialed digit analysis, mines call data to build up a database of telephone numbers called by bandits during periods of fraudulent ac- tivity. For detection, this database is matched against the numbers called by cus- tomers, and alarms are produced when the number of hits is above a threshold ("dialed digit hits"). 2.2.3. User profiling User profiling methods constitute a special class of post-call methods. They involve analyzing calling behavior in order to detect usage anomalies suggestive of fraud. Profiling often works well with low-usage subscribers because unusual behavior is very prominent. For this reason, profiling is a good complement to collision and velocity checking because it covers cases the others might miss. Figure shows some chronological call data from an example (fabricated) frauded account (the fields shown are just a sample; our call data contain many more attributes than shown here). The column at the far right indicates whether the call is fraudulent or not; that is, whether the call was placed by the customer or the bandit. A fraud analyst looking at this account would quickly be able to recognize the two classes of calls: 1. The legitimate user calls from the metro New York City area, usually during working hours, and typically makes calls lasting a few minutes. 2. The bandit's calls originate from a different area (Boston, Massachusetts, about 200 miles away), are made in the evenings, and last less than a minute. Ideally, a fraud detection system should be able to learn such rules automatically and use them to catch fraud. This paper addresses the automatic design of user profiling methods. User profiling methods are attractive because they do not depend upon any special hardware capability, as authentication does, nor do they require that the customer replace or upgrade existing equipment. Moreover, the ability to generate such detectors is domain independent: such a system should be able to generate fraud detectors for any domain with superimposition fraud. 2.3. The Need to be Adaptive There are a number of commercially available expert systems for fraud detection that include user profiling. Fraud analysts or system administrators can tune the techniques by adjusting parameters, or by entering specific patterns that will trigger alarms. Unfortunately, determining which potential patterns will be useful is a time-consuming process of trial-and-error. Moreover, the patterns of fraud are dynamic. Bandits constantly change their strategies in response to new detection techniques or new cellular hardware capabilities. By the time a system is manually tuned, the fraudulent behavior may have changed significantly. The environment is dynamic in other ways as well. The level of fraud changes dramatically month-to-month because of modifications to work practices (both the carrier's and the bandits'). Also, the costs of missing fraud or of dealing with false alarms change with intercarrier contracts, and because of fraud analyst workforce issues. For all these reasons, it is important that a fraud detection system adapt easily to new conditions. It should be able to notice new patterns of fraud. It should also be able to modify its alarm generation behavior, for example, as the level of fraud or the cost of dealing with a false alarm changes. Such adaptability can be achieved by generating fraud detection systems automatically from data, using data mining techniques. 3. Automatic Construction of Profiling Fraud Detectors One approach to building a fraud detection system is to classify individual trans- actions, calls in our case, as being fraudulent or legitimate. Classification has been well explored, e.g., in machine learning and statistics, so this would seem to be a straightforward application of existing techniques. We have not had success using standard machine learning techniques to construct such a classifier. Some specific results are discussed in Section 6.3. In general, there are two problems that make simple classification approaches infeasible. call that would be unusual for one customer would be typical for another. For example, a call placed from Brooklyn is not unusual for a subscriber who lives there, but might be very strange for a Boston subscriber. 6 FAWCETT AND PROVOST Thus, it is necessary (i) to discover indicators corresponding to changes in behavior that are indicative of fraud, rather than absolute indicators of fraud, and (ii) to profile the behavior of individual customers to characterize their normal behavior. If there were available substantial information about an account's context, we could possibly ameliorate this problem. Context information would comprise behavior information such as what the phone is used for, what areas it is used in, what areas/numbers it normally calls, what times of day, and so on. Context information is not available, 1 so our solution is to derive it from historical data specific to each account. The discovery of context-sensitive fraud indicators and the profiling of individual accounts comprise two of the three major elements of the learning problem. At the level of the individual call, the variation in calling behavior is large, even for a particular user. Legitimate subscribers occasionally make calls that look suspicious. As far as we have been able to determine, it is not possible to achieve simultaneously the high degree of accuracy and high level of coverage necessary to classify individual calls effectively. Any classifier that fires on a significant number of defrauded accounts produces unacceptably large numbers of false alarms. Therefore, decisions to take corrective action cannot be made with confidence on the basis of individual calls. Instead, it is necessary to aggregate customer behavior, smoothing out the variation, and watch for coarser-grained changes that have better predictive power. This is the third major element of the learning problem; in the experiments we describe later, we aggregate customer behavior into account-days. In sum, the learning problem comprises three questions, each of which corresponds to a component of our framework. 1. Which call features are important? Which features or combinations of features are useful for distinguishing legitimate behavior from fraudulent behavior? 2. How should profiles be created? Given an important feature, how should we characterize/profile the behavior of a subscriber with respect to the feature, in order to notice important changes? 3. When should alarms be issued? Given the results of profiling behavior based on multiple criteria, how should they be combined to be effective in determining when fraud has occurred? Each of these issues corresponds to a component of our framework. 4. The Detector Constructor Framework Our Detector Constructor framework is illustrated in Figure 2. Under the frame- work, a system first learns rules that serve as indicators of fraudulent behavior. It Monitor Construction Monitor templates Rules Fraud Detector Call data Mn Training and Selection Rule Learning Classifier Profiling monitors Figure 2. The framework for automatically constructing fraud detectors. then uses these rules, along with a set of templates, to create profiling monitors These monitors profile the typical behavior of each account with respect to a rule and, in use, describe how far each account is from its typical be- havior. Finally, the system learns to weight the monitor outputs so as to maximize the effectiveness of the resulting fraud detector. Figure 3 shows how such a detector will be used. The monitors are provided with a single day's calls from a given account, and each monitor generates a number indicating how unusual that account-day looks for the account. The numeric outputs from the monitors are treated as evidence and are combined by the detector. When the detector has enough evidence of fraudulent activity on an account, based on the indications of the monitors, it generates an alarm. We now discuss each step of the framework in detail, illustrated by the particular choices made in our first implemented system, as applied to the problem of cloning fraud detection. The first Detector Constructor system is called DC-1. The call data used for detecting cloning fraud are chronological records of calls made by each subscriber, organized by account. These data describe individual calls using attributes such as DATE, FROM-STATE, DURATION and CELL-SITE. Account-Day Day Time Duration Origin Destination mins Bronx, NY Miami, FL Tue 10:05 3 mins Scarsdale, NY Bayonne, NJ Tue 11:23 24 sec Scarsdale, NY Congers, NY Tue 14:53 5 mins Tarrytown, NY Greenwich, CT Tue 15:06 5 mins Manhattan, NY Westport, CT Tue 16:28 53 sec Scarsdale, NY Congers, NY Tue 23:40 17 mins Bronx, NY Miami, FL BRONX at night # calls from exceeds daily threshold Airtime from BRONX at night SUNDAY airtime exceeds daily threshold1 0 Value normalization and weighting FRAUD ALARM Combining Evidence >= q Profiling Monitors Figure 3. A DC-1 fraud detector processing a single account-day of data. 4.1. Learning Fraud Rules The first stage of detector construction, rule learning, involves searching the call data for indicators of fraud. In the DC-1 system, the indicators are conjunctive rules discovered by a standard rule-learning program. As discussed above, an obvious way of mining fraud indicators is to create an example set consisting of all frauded calls and all legitimate calls, and apply a rule learning algorithm to the example set. However, this approach loses context information about the normal behavior of the account in which the fraud occurred. To illustrate the importance of context, consider a situation in which half the subscribers live in New York and half in Los Angeles. When cloned, the New York accounts are used in Los Angeles and the Los Angeles accounts are used in New York. Applying rule learning to the combined set of call records would uncover no fraud rules based on call origin; in other words, knowing that a call originated in New York says nothing about how likely it is to be fraud. In fact, this conclusion would be wrong: in this scenario a New York account with Los Angeles calls is much more likely to have been cloned than if it had only New York calls. This fact is missed when using a combined example set, because in combining examples all account context information is lost. In light of the need to maintain context information, rule learning is performed in two steps. Rules are first generated locally based on differences between fraudulent and normal behavior for each account, then they are combined in a rule selection step. 4.1.1. Rule generation DC-1 uses the RL program (Clearwater and Provost 1990; Provost and Aronis 1996) to generate indicators of fraud in the form of classification rules. Similar to other MetaDENDRAL-style rule learners (Buchanan and Mitchell 1978; Segal and Etzioni 1994; Webb 1995), RL performs a general- to-specific search of the space of conjunctive rules. This type of rule-space search is described in detail by Webb (Webb 1995). In DC-1, RL uses a beam search for rules with certainty factors above a user-defined threshold. The certainty factor we used for these runs was a simple frequency-based probability estimate, corrected for small samples (Quinlan 1987). In order to deal with the very large numbers of values for some of the attributes used to describe calls (more than 10,000 values in total), RL also used breadth-first marker propagation techniques, so that the algo- rithm's time complexity does not depend on the number of attribute values (Aronis and Provost 1997). (RL's time complexity is linear in the number of attributes and the number of examples.) The call data are organized by account, with each call record labeled as fraudulent or legitimate. When RL is applied to an account's calls it produces a set of rules that serve to distinguish, within that account, the fraudulent calls from the legitimate calls. As an example, the following rule would be a relatively good indicator of Certainty This rule denotes that a call placed at night from The Bronx (a Borough of New York City) is likely to be fraudulent. The Certainty means that, within this account, a call matching this rule has an 89% probability of being fraudulent. From each account, RL generates a "local" set of rules describing the fraud on that account. Each rule is recorded along with the account from which it was generated. The covering heuristic typically used by RL was disabled, so that all of the (maximally general) rules with probability estimates above threshold would be generated. This option was chosen because rule generation in DC-1 is local and decisions about coverage should not be made locally. The next step, rule selection, incorporates information about coverage and generality. 4.1.2. Rule Selection After all accounts have been processed, a rule selection step is performed. The purpose of this step is to derive a set of rules that will serve as fraud indicators. A rule selection step is necessary because the rule generation step typically generates tens of thousands of rules in total, most of which are specific only to single accounts. The system cannot know a priori how general each rule will be. For example, from one account RL may generate the rule: Given: Accts: set of all accounts Rules: set of all fraud rules generated from Accts Number of rules required to cover each account. Taccts : (parameter) Number of accounts in which a rule must have been found. Output: S: Set of selected rules. 1. /* Initialization */ 2. 3. for (a 2 Accts) do 4. for (r 2 Rules) do 5. Number of accounts in which r occurs */ of accounts generating r */ 7. end for 8. /* Set up Occur and AcctsGen */ 9. for (a 2 Accts) do set of rules generated from a; 11. for (r 2 Ra ) do 12. Occur[r] := Occur[r] 13. add a to AcctsGen[r]; 14. end for; end for 15. /* Cover Accts with Rules */ 16. for (a 2 Accts) do 17. list of rules generated from a; 18. sort Ra by Occur; 19. while (Cover[a] ! T rules ) do 20. r := highest-occurrence rule from Ra 21. Remove r from Ra 22. if 23. add r to 24. for (a2 2 AcctsGen[r]) do 25. 26. end for; end if 27. end while; end for Figure 4. Rule selection and covering algorithm used by DC-1 This rule is probably specific only to the account from which it was generated, but there is no a priori way to know a rule's generality in the generation step, which processes a single account at a time. If this rule is found in ("covers") many accounts, it is probably worth using; if it was only found in a single account, it is probably not a general indicator of fraudulent behavior. Even if the same account ADAPTIVE FRAUD DETECTION 11 is cloned again, it will not be defrauded in exactly the same way. Note that DC-1's notion of coverage is slightly different from the standard notion, because of the multiple levels of granularity; in particular, DC-1 selects a set of rules that covers the accounts, as opposed to typical classifier learning, in which a set of rules is selected that covers the examples (in this case, the calls). The rule selection algorithm is given in Figure 4. The algorithm identifies a small set of general rules that cover the accounts. Two parameters control the algorithm. T rules is a threshold on the number of rules required to "cover" each account. In the selection process, if an account has already been covered by T rules rules, it will not be examined. T accts is the number of accounts a rule must have been found in (i.e., mined from) in order to be selected at all. For each account, the list of rules generated by that account is sorted by the frequency of occurrence in the entire account set. The highest frequency unchosen rule is selected. An account is skipped if it is already sufficiently covered. The resulting set of rules is used in construction of monitors. 4.2. Constructing Profiling Monitors Rule learning produces a set of rules characterizing changes that commonly occur when an account is cloned. These rules are not universal: for a given account, we do not know to what extent that account's normal behavior already satisfies the rule. For example, the "Bronx-at-night" rule, mentioned above, may be very useful for someone living in Hartford, Connecticut, but it may cause many false alarms on a subscriber living in the Bronx. A fraud detection system should distinguish the two. In the latter case the system should inhibit the rule from firing, or at least require a much higher level of activation. Sensitivity to different users is accomplished by converting the rules into profiling monitors. Each monitor has a Profiling step and a Use step. Prior to being used on an account, each monitor profiles the account. In the Profiling step, the monitor is applied to a segment of an account's typical (non-fraud) usage in order to measure the account's normal activity. Statistics from this profiling period are saved with the account. In the monitor's Use phase, the monitor processes a single account-day at a time. The monitor references the normalcy measures calculated in Profiling, and generates a numeric value describing how abnormal 2 the current account-day is. Profiling monitors are created by the monitor constructor, which employs a set of templates. The templates are instantiated by rule conditions. Given a set of rules and a set of templates, the constructor generates a monitor from each rule-template pair. Two monitor templates are shown in Figure 5. At the top is a template that creates threshold monitors. Such a monitor yields a binary feature corresponding to whether the user's behavior was above threshold for the given day. The bottom of Figure 5 shows a template for a standard deviation monitor. In the Profiling period, such monitors measure the mean (-) and standard deviation (oe) of typical usage; in the Use period, they produce a continuous output representing how many standard deviations above the mean an account-day is. Threshold monitors ffl Given: Rule conditions from a fraud rule. ffl Profiling: On a daily basis, count the number of calls that satisfy rule con- ditions. Keep track of the maximum as daily threshold. ffl Use: Given an account-day, let C be the set of all calls on that day that satisfy rule conditions. daily threshold Standard deviation monitors ffl Given: Rule conditions from a fraud rule. ffl Profiling: On a daily basis, sum the airtime of all calls satisfying rule con- ditions. At the end of the training period, record the mean (-) and standard deviation (oe) of the samples. ffl Use: Given an account-day, let C be the set of all calls on that day that satisfy rule conditions. Let Call2C airtime(Call) Airtime if Airtime\Gamma- oe if Airtime ? - Figure 5. Two templates for creating monitors from rules. A threshold monitor learns a threshold on maximum use and outputs a 1 whenever daily usage exceeds the threshold. A standard deviation monitor outputs the number of standard deviations over the mean profiled usage. As an example, assume the Bronx-at-night rule mentioned earlier was used with the template shown in Figure 5b. Assume that, on some account, the subscriber called from the Bronx an average of five minutes per night with a standard deviation of two minutes. At the end of the Profiling step, the monitor would store the values (5,2) with that account. In Use on that account, if the monitor processed a day USED USED Std. Deviation Mean Std. Deviation Mean Use Profiling Use Profiling Figure 6. Using mean and standard deviation in profiling. containing three minutes of airtime from the Bronx at night, the monitor would emit a zero; if the monitor saw 15 minutes, it would emit (15 5. This value denotes that the account is five standard deviations above its average (profiled) usage level. Standard deviation monitors are sensitive both to the expected amount of activity on an account and to the expected daily variation of that activity. Figure 6 illustrates the difference, showing one monitor applied to two accounts. The account on the left has low variation in the Profiling period so its standard deviation is lower. Consequently the erratic behavior in its Use period will produce large values from the monitor. The account on the right has the same mean but exhibits much larger variation in Profiling period, so the standard deviation is higher. The variations in behavior during the Use period will not produce large values from the monitor. 4.3. Combining Evidence from the Monitors The third stage of detector construction learns how to combine evidence from the monitors generated by the previous stage. For this stage, the outputs of the monitors are used as features to a standard learning program. Training is done on account data, and monitors evaluate one entire account-day at a time. In training, the monitors' outputs are presented along with the desired output (the account- day's correct class: fraud or non-fraud). The evidence combination weights the monitor outputs and learns a threshold on the sum so that alarms may be issued with high confidence. Many training methods for evidence combining are possible. We chose a simple Linear Threshold Unit (LTU) (Nilsson 1965; Young 1984) for the experiments reported below. An LTU is simple and fast, and enables a good first-order judgment of the features' worth. 14 FAWCETT AND PROVOST A feature selection process is used to reduce the number of monitors in the final detector. Some of the rules do not perform well when used in monitors, and some monitors overlap in their fraud detection coverage. We therefore employ a sequential forward selection process (Kittler 1986) which chooses a small set of useful monitors. Empirically, this simplifies the final detector and increases its accuracy. The final output of DC-1 is a detector that profiles each user's behavior based on several indicators, and produces an alarm if there is sufficient evidence of fraudulent activity. Figure 3 shows an example of a simple detector evaluating an account-day. Before being used on an account, the monitors each profile the account. They are applied to a profiling segment (thirty days in our experiments) during which they measure unfrauded usage. In our study, these initial thirty account-days were guaranteed free of fraud, but were not otherwise guaranteed to be typical. From this initial profiling period, each monitor measures a characteristic level of activity. 4.4. Summary of the Detector Construction Process In sum, DC-1 begins by examining the call records of defrauded accounts. The call records are expressed in terms of a set of base level attributes with thousands of possible values. From these data, the system generates rules characterizing fraudulent calls within accounts, then selects a smaller set of general rules as indicators of fraudulent behavior. These rules are used as the basis from which to build a set of profiling monitors, each of which examines behavior based on one learned rule. A monitor learns the typical behavior of an account by scanning an initial sequence of the account's calls (its "profiling period") and saving some statistics. Subsequently, the monitor examines chunks of calls (in our experiments, account-days). The monitor subsequently examines each chunk (day) of each account's behavior and outputs a number indicating how far away from normal the behavior is. In order to construct a high-confidence detector, DC-1 must then learn to combine the outputs of the monitors effectively. To do this, it trains a classifier on a sample of account-days. Each account-day is a training instance expressed as a vector of monitor outputs for that day, and labelled as either containing fraud or not. After training, the system has a classifier that is able to combine the monitors effectively. The final output of the system is a set of monitors and a trained classifier for combining their outputs. In order to be applied to a new account, the monitors must see a profiling period of days from that account. The next sections describe our cellular call data and the experiments we have performed on the system. 5. The Data The call data used for this study are records of cellular calls placed over four months by users in the New York City area-an area with high levels of fraud. Each call is described by thirty-one attributes, such as the phone number of the caller, the duration of the call, the geographical origin and destination of the call, and any long-distance carrier used. Because of security considerations, we are unable to disclose all the features used in the system. To these thirty-one attributes are added several derived attributes that incorporate knowledge we judged to be potentially useful. One such attribute is a categorical TIME-OF-DAY variable representing the time segment of the day in which a call is placed. Its values are MORNING, AFTERNOON, TWILIGHT, EVENING and NIGHT. Another derived attribute is TO-PAYPHONE, a binary flag indicating whether the call terminated at a payphone. Note that any number of additional features could be added to encode relevant domain knowledge. Each call is also give a class label of legitimate or fraudulent. This is done by cross referencing a database of all calls that were credited as being fraudulent for the same time period. 5.1. Data Cleaning Like all real-world data, our cellular call data contain errors and noise from various sources. For example, calls are marked as fraudulent based on a process called block crediting. In this process, the customer and the carrier representative together establish the range of dates during which the fraud occurred, and the calls within the range are credited. The customer is usually not asked about each individual call. The block crediting process uses heuristics to discard obvious non-fraudulent calls from the credited block, but these heuristics are fallible. Also, if there is a disagreement about the fraud span, the customer service representative usually concedes, in the customer's favor, to a wider date span. Any erroneously credited calls constitute noise in our data. Because of these noise sources, we cleaned the data in several ways. ffl Each account's calls were scanned automatically to eliminate credited calls to numbers that had been called outside of the credited block. In other words, the program looked for credited calls made to a phone number that also had been called by the legitimate subscriber; in this case, the crediting may have been a mistake, so the call is discarded completely. ffl An account-day was classified as fraudulent only if five or more minutes of fraudulent usage occurred. Days including one to four minutes of fraudulent usage were discarded. This policy eliminated a small number of "gray area" account-days probably mislabelled due to small amounts of noise. For example, the database of credits due to fraud occasionally included credits for other reasons, such as wrong numbers. ffl Within any time period there will be fraud that has not yet been detected. We assumed that some genuinely fraudulent calls would not be marked as such because of this time lag. We attempted to minimize this noise by delaying the data retrieval by two weeks. ffl In preliminary experiments, rule learning uncovered some unusual attribute values (e.g., that seemed to be very strong indicators of fraud. Discussions with the database providers led us to conclude that these suspicious values were artifacts of the crediting process: in some circumstances, crediting would erase or replace certain fields. Because the values appeared primarily in credited records, data mining had extracted them as high-confidence fraud rules. We found five or six such misleading values and eliminated from the database all records containing them. In addition, the start times of calls had been recorded in local time with respect to the switch of origin. The calls were normalized to Greenwich Mean Time for chronological sorting. 5.2. Data Selection The call data were separated carefully into several partitions for rule learning, account profiling, and detector training and testing. Once the monitors are created and the accounts profiled, the system transforms raw call data into a series of account-days using the outputs of the monitors as features. Rule learning and selection used 879 accounts comprising over 500,000 calls. About 3600 accounts were selected for profiling, training, and testing. The only condition used to select these 3600 accounts was that they be guaranteed to have at least thirty fraud-free days of usage before any fraudulent usage. The initial thirty days of each account were used for profiling. The remaining days of usage were used to generate approximately 96,000 account-days. Using randomly selected accounts, we generated sets of 10,000 account-days for training and 5000 account-days for testing. Training and testing accounts were distinct, so their account-days were not mixed between training and testing. 3 Each set of account-days was chosen to comprise 20% fraud and 80% non-fraud days. 6. Experiments and Evaluation Rule learning generated 3630 rules, each of which applied to two or more accounts. The rule selection process, in which rules are chosen in order of maximum account coverage, yielded a smaller set of 99 rules sufficient to cover the accounts. Each of the 99 rules was used to instantiate two monitor templates, yielding 198 monitors. The final feature selection step reduced this to eleven monitors, with which the experiments were performed. 6.1. The Importance of Error Cost In this domain, different types of errors have different costs. A realistic evaluation should take misclassification costs into account. Classification accuracy, a standard metric within machine learning and data mining, is not sufficient. A false positive error (a false alarm) corresponds to wrongly deciding that a customer has been cloned. Based on the cost of a fraud analyst's time, we estimate the cost of a false positive error to be about $5. A false negative error corresponds Table 1. Accuracies and costs of various detectors. Detector Accuracy (%) Cost (US$) Accuracy at cost (%) Alarm on All 20 20000 20 Alarm on None 80 18111 \Sigma 961 80 Collisions Velocities High Usage 88 \Sigma :7 6938 \Sigma 470 85 \Sigma 1:7 State of the Art (SOTA) 90 \Sigma :4 6557 \Sigma 541 88 \Sigma :9 detector 92 \Sigma :5 5403 \Sigma 507 91 \Sigma :8 to letting a frauded account-day go undetected. Rather than using a uniform cost for all false negatives, we estimated a false negative to cost $.40 per minute of fraudulent airtime used on that account-day. This figure is based on the proportion of usage in local and non-local ("roaming") markets, and their corresponding costs. Because LTU training methods try to minimize errors but not error costs, we employed a second step in training. After training, the LTU's threshold is adjusted to yield minimum error cost on the training set. This adjustment is done by moving the decision threshold from -1 to +1 in increments of .01 and computing the resulting error cost. After the minimum cost on training data is found, the threshold is clamped and the testing data are evaluated. 6.2. DC-1 Compared with Alternative Detection Strategies Table 1 shows a summary of results of DC-1 compared against other detectors. The name of each detector is shown in the left-most column. Classification accuracy averages and standard deviations are shown in the second column. The third column shows the mean and standard deviations of test set costs. The right-most column, "Accuracy at cost," is the corresponding classification accuracy of the detector when the threshold is set to yield lowest-cost classifications. Each detector was run ten times on randomly selected training and testing ac- counts. For comparison, we evaluated DC-1 along with other detection strategies. ffl Alarm on All represents the policy of alarming on every account every day. The opposite strategy, Alarm on None, represents the policy of allowing fraud to go completely unchecked. The latter corresponds to the maximum likelihood accuracy classification. Note that the cost of Alarm on None does not take into account the inhibitory effect of fraud detection, without which fraud levels would likely continue to rise. ffl Collisions and Velocities is a detector using collision and velocity checks described in Section 2.2.2. DC-1 was used to learn a threshold on the number of collision and velocity alarms necessary to generate a fraud alarm. It is surprising that Collisions and Velocity Checks, commonly thought to be reliable indicators of cloning, performed poorly in our experiments. The performance of collisions and velocity checks was originally worse than reported here because of false alarms. Manual inspection of false alarms revealed a few synchronization problems; for example, some apparent collisions were caused when a call was dropped then quickly re-established in a neighboring cell whose clock did not agree with the first cell's. Some such conditions could be caught easily, so we patched the detection algorithms to check for them. The results in this paper are for the improved detectors. Investigation of confusion matrices revealed that the collision and velocity check detectors' errors were due almost entirely to false negatives. In other words, when the detectors fired they were accurate, but many fraud days never exhibited a collision or velocity check. ffl Some fraud analysts believe that cloning fraud is usually accompanied by large jumps in account usage, and sophisticated mining of fraud indicators is probably unnecessary since most fraud could be caught by looking for sudden increases in usage. We created the High Usage detector to test this hypothesis. It generates alarms based only on amount of usage. It is essentially a standard deviation monitor (see Figure 5) whose rule conditions are always satisfied. The threshold of this detector was found empirically from training data. Note that the evaluation of cost for the high usage detector may be overly optimistic, due to inadequacies in our cost model. In particular, a trained high usage detector learns to optimally "skim the cream," without regard to the fact that the errors it makes will involve annoying the best customers. In these cases, the cost of a false alarm may be much higher than the fixed cost we assigned. ffl The Best Individual DC-1 Monitor was used as an isolated detector. This experiment was done to determine the additional benefit of combining monitors. The best individual monitor was generated from the rule: Rule learning had discovered (in 119 accounts) that the sudden appearance of evening calls, in accounts that did not normally make them, was coincident with cloning fraud. The relatively high accuracy of this one monitor reveals that this is a valuable fraud indicator. Our TIME-OF-DAY attribute has five possible values: MORNING, AFTER- NOON, TWILIGHT, EVENING and NIGHT. Although EVENING is by far the most frequent value implicated in fraud, rule learning generated fraud rules involving each of these values. This suggests that any time-of-day change in a subscriber's normal behavior may be indicative of fraud, though the other shifts may not be predictive enough to use in a fraud monitor. ffl The DC-1 detector incorporates all the monitors chosen by feature selection. We used the weight learning method described earlier to determine the weights for evidence combining. ADAPTIVE FRAUD DETECTION 19 ffl The Of The Art") detector incorporates thirteen hand-crafted profiling methods that were the best individual detectors identified in a previous study. Each method profiles an account in a different way and produces a separate alarm. Weights for combining SOTA's alarms were determined by our weight-tuning algorithm. Details on the detectors comprising are given in Appendix A. The results in Table 1 demonstrate that DC-1 performs quite well. In fact, DC- outperforms in terms of both accuracy and cost. 4 In our experiments, lowest cost classification occurred at an accuracy somewhat lower than optimal. In other words, some classification accuracy can be sacrificed to decrease cost. More sophisticated methods could be used to produce cost sensitive classifiers, which would probably produce better results. Finally, the monitors of SOTA and DC-1 were combined into a hybrid detector. The resulting detector exhibits no increase in classification accuracy, but does show a slight improvement in fraud detection cost. In this work we have dealt with differing costs of false positive and false negative errors. However, we have still glossed over some complexity. For a given account, the only false negative fraud days that incur cost to the company are those prior to the first true positive alarm. After the fraud is detected, it is terminated. Thus, our analysis overestimates the costs slightly; a more thorough analysis would eliminate such days from the computation. 6.3. Fraudulent Call Classifiers Section 4.1 asserted that account context is important in the rule learning step: a global example set taken from all accounts would lose information about each account's normal behavior. In order to test this hypothesis, two such call classifiers were created from global example sets. Applying standard classification algorithms to the call data was difficult for several reasons. First, the description language is very detailed because many thousands of attribute values appear in the data. Because of this, the volume of data necessary was relatively large for desktop platforms; the use of fewer than 100,000 examples led to erratic classification behavior. Furthermore, in order to achieve high coverage of calls, massively disjunctive concept descriptions had to be learned- there were no simple classifiers that performed well. After trying many approaches, we chose two classifiers learned in the following manner. A set of 100,000 training examples was sampled from the accounts set aside for rule learning. The sample was random, but stratified to achieve a 50/50 class distribution. RL was applied to these data, with parameters set so that it learned massively disjunctive rule sets. The two classifiers, CC 1054 and CC 1861, comprise 1054 and 1861 rules, respectively. 5 Each of these rule sets covered around 60% of the calls in a 92212 example test set, with an accuracy of about 75% on the calls it covered. We observed a clear and graceful tradeoff between accuracy and coverage: as the coverage increased, the accuracy decreased. Table 2. A comparison of DC-1 to two global call classifiers Detector Accuracy (%) Cost (US$) Accuracy at cost (%) detector 92 \Sigma :5 5403 \Sigma 507 91 \Sigma :8 In order to achieve a competitive comparison, the call classifiers were then given the advantage of profiling and monitoring. A standard deviation airtime monitor was created from each. Specifically, instead of instantiating the monitor template with a single rule, the template was instantiated with the entire classifier. The resulting monitor profiled each account's normal behavior with respect to the clas- sifier's output. The call classifier monitor learns if a particular customer's legitimate behavior typically triggers a positive output. Furthermore, each call classifier monitor was inserted into the DC-1 weight-training framework in order to find an "optimal" output threshold for accuracy maximization or cost minimization. The results are shown in Table 2. The two call classifiers perform similarly, and outperforms both by a considerable margin. Indeed, we were surprised that the the call classifier monitors perform as well as they do. 6.4. Shifting Distributions of Fraud As discussed in Section 2.3, a fraud detection system should be able to adapt to shifting fraud distributions. For example, each month the relative amount of fraud changes slightly, and it is rarely possible to predict the level of fraud far into the future. Thus, unless it is adaptive, even a well-tuned detection system will begin to lose its edge. To illustrate this point, we simulated the effects of changing fraud distributions on detector performance. One DC-1 detector was trained on a fixed distribution of account-days (80% non-fraud, 20% fraud) and tested against several other distributions (ranging from 75% to 99% non-fraud account-days), to simulate a well-tuned but non-adaptive detection system. Another DC-1 detector was allowed to adapt to each distribution; its LTU threshold was re-trained for minimum predicted cost on a training set with the new distribution. The results are shown in Figure 7. The X-axis is the percentage of non-fraud account-days, and the Y-axis is the cost per account day. This figure shows that the second detector, which is allowed to adjust itself to each new distribution, is consistently more cost effective than the fixed detector. This difference increases as the testing distribution becomes more skewed from the distribution upon which the fixed detector was trained. We close by noting that these experiments illustrated changes in fraud detection performance with respect to fairly simple changes in fraud distribution (changing ADAPTIVE FRAUD DETECTION 210.20.611.4 Cost Percentage of non-fraud Trained for 80/20 Adapted to each distribution Figure 7. The effects of changing fraud distributions fraud volume). The patterns of fraud also change, particularly in reponse to detection methods. Thus the ability to use data mining to discover new patterns amplifies the benefit of adaptability. 6.5. Discussion It is difficult to evaluate DC-1 against existing expert systems for fraud detection. Fraud detection departments carefully protect information about how much fraud they have and how effective their detection strategies are. Likewise, vendors of fraud detection systems protect details of their systems' operation that may constitute trade secrets. Little performance data on fielded systems are available, and what data do exist are insufficient for careful evaluation. For these reasons, we evaluated DC-1 against individual known fraud detection techniques, as well as against a collection of techniques representing the state of the art as we understand it. Results in the previous sections show that the DC- detector performs better than the high-usage alarm and the collision/velocity alarm. DC-1 also out-performs the detector, consisting of a collection of the best fraud detection techniques known to us, trained by DC-1's evidence combining method. 22 FAWCETT AND PROVOST DC-1's framework has three main components, and is more complex than other approaches. Our experiments were designed not only to evaluate the overall performance of the system, but also to analyze the contribution of the individual components. In particular: ffl The High Usage detector profiles with respect to undifferentiated account usage. Comparison with DC-1's performance demonstrates the benefit of using rule learning to uncover specific indicators of fraudulent calls. ffl The Call Classifier detectors represent rule learning without the benefit of account context. Comparison with DC-1's performance demonstrates the value of DC-1's rule generation step, which does preserve account context. ffl Comparison of DC-1 with the single best individual DC-1 monitor demonstrates the benefit of combining evidence from multiple monitors. ffl Experiments with shifting fraud distributions indicate the benefit of making evidence combination sensitive to fraud distributions. In each of these cases, the composite DC-1 system out-performed the detector in which a significant piece was missing. These results suggest that each component contributes critically to the performance of the entire detector. Our system uses a Linear Threshold Unit to combine evidence from the monitors. Other methods of evidence combination are possible. We performed some experiments with multi-layer neural networks, but found that adding units to a hidden layer did not improve performance. These networks produced higher training accu- racies, but lower accuracies on the test sets; such behavior is symptomatic of data overfitting. Additional experimentation might yield better-performing networks, but we have not pursued this. It is possible that, because the neural network is applied far along in the fraud detection process as a means of combining evidence, non-linear combinations of the evidence contribute little to fraud detection performance By increasing the expressiveness of the language used for inductive learning, it may be possible to learn more general patterns of fraudulent behavior, reducing the need for highly disjunctive class descriptions. The caveats mentioned earlier about the inability to procure background knowledge for context notwithstanding, it may be possible to provide additional context by linking call and account data to geographic and demographic databases. Furthermore, it may be possible to learn context in one stage, and then apply relational learning approaches in a later stage. One such possibility is to make use of inductive learners that learn concept descriptions in first-order logic, such as FOIL (Quinlan 1990) or ILP methods (D-zeroski 1996). Given the appropriate context information, it is possible that more expressive methods could learn general relational rules such as the following, which indicates fraud when a user calls from an abnormal location. Another possibility is to use a learner that forms propositional rules, but can take advantage of relational background knowledge in the process (Aronis, Provost, and Buchanan 1996). We have not explored the use of relational learning to any great extent. We have linked the data to knowledge about the geographic locations of telephone numbers (Aronis and Provost 1997), which does produce useful generalizations of the areas to which calls are placed. Finally, it is important to point out additional limitations to the evaluation of learned classifiers on real-world problems. For the work described in this paper, we made use of techniques to deal with skewed class distributions, viz., stratified sam- pling, and to deal with nonuniform misclassification costs, viz., empirical threshold adjustment. We also ensured that our evaluation include cost effectiveness in addition to accuracy. However, because of the complexity and dynamics of real-world domains, determining precisely the target cost and class distributions is often im- possible. As noted above, levels of fraud and costs change monthly. It is important to be able to compare competing classification methods under imprecision in these distributions. The investigation and design of such techniques is an important area for future research (Provost and Fawcett 1997). 7. Related Work Fraud detection is related to intrusion detection, a field of computer security concerned with detecting attacks on computers and computer networks (Frank 1994; Sundaram 1996; Kumar 1995). Many forms of intrusion are instances of superimposition fraud, and thus candidates for systems built with our framework. Within the intrusion detection community, anomaly detection systems try to characterize behavior of individual users in order to detect intrusions on that user's account via anomalies in behavior. Existing anomaly detection systems typically examine audit trails of user activities, which fill the same roll as cellular call records in DC- 1. DC-1 would be considered a statistical anomaly detection system. Sundaram An open issue with statistical approaches in particular, and anomaly detection systems in general, is the selection of measures to monitor and the choice of metrics. It is not known exactly what the subset of all possible measures that accurately predicts intrusive activities is. DC-1's framework directly addresses this problem. Its rule learning step examines large numbers of fraud episodes in order to generate features (measures) that distinguish fraudulent from legitimate behavior. To the best of our knowledge, no published anomaly detection system does this. Calling card fraud and credit card fraud are other forms of superimposition fraud. A system built by Yuhas (1993, 1995) examines a set of records representing calling- card validation queries to identify queries corresponding to fraudulent card usage. Yuhas transformed the problem into a two-class discrimination task and trained several machine learning models on the data. All three models had comparable performance on the test sets. His system must be provided with appropriate fea- tures; it neither mines the data for fraud indicators nor measures typical customer usage. Stolfo et al. (1997) address credit card fraud detection. They also transform the problem into a two-class discrimination task, and do not use customer-specific information in detection. Specifically, they predict whether individual transactions are fraudulent. In our domain, we found that DC-1 significantly improves detection performance over systems that use transaction classification alone. It would be interesting to determine whether a system like DC-1 could improve performance on these other superimposition fraud tasks. Ezawa and Norton (1995, 1996) have addressed the problem of uncollectible debt in telecommunications services. They use a goal-directed Bayesian network for clas- sification, which distinguishes customers who are likely to default from those who are not. As with our work, Ezawa and Norton's work faces problems with unequal error costs and skewed class distributions. However, it does not face the problem of determining the typical behavior of individual customers so as to recognize superimposed fraudulent behavior. Mining the data to derive profiling features is not necessary. Because fraud happens over time, methods that deal with time series are relevant to this work. However, traditional time series analysis (Chatfield 1984; Farnum and Stanton 1989) in statistics strives either to characterize an entire time series or to forecast future events in the series. Neither ability is directly useful to fraud detection. Hidden Markov Models (Rabiner and Juang 1986; Smyth 1994) are concerned with distinguishing recurring sequences of states and the transitions between them. However, fraud detection usually only deals with two states (the "frauded" and "un-frauded" states) with a single transition between them. Yuhas (1995) mentions the possibility of recognizing "at home" and "travel" states in order to distinguish frauded states more effectively. This differentiation could be useful for reducing false alarms. We are aware of no work pursuing this idea. 8. Conclusion The detection of cellular cloning fraud is a relatively young field. Fraud behavior changes frequently as bandits adapt to detection techniques, and fraud detection systems should be adaptive as well. However, in order to build usage monitors we must know which aspects of customers' behavior to profile. Historically, determining such aspects has involved a good deal of manual work, hypothesizing useful features, building monitors and testing them. Determining how to combine them involves much trial-and-error as well. We have presented and demonstrated a framework that automates the process of generating fraud detectors. This framework is not specific to cloning fraud, but may be applied to superimposition fraud problems in any domain. Prime candidates are toll fraud, computer intrusion and credit-card fraud. For example, in credit-card fraud, data mining may identify locations that arise as new hot-beds of fraud. The constructor would then incorporate monitors that notice if customers begin to charge more than they usually do from these specific locations. Even with relatively simple components, DC-1 is able to exploit mined data to produce a detector whose performance exceeds that of the state-of-the-art. The system took several person-months to build; the DC-1 detector took several CPU-hours. Furthermore, DC-1 can be retrained at any time as necessitated by the changing environment. Such adaptability is beneficial in many ways. It can save effort in time-consuming manual feature identification and detector tuning. It can save on monetary losses that would occur during the manual identification and tuning process. It can save on less quantifiable damage done due to higher fraud, such as lower customer opinion (or even customer churn). Finally, it can act to prevent fraud; a system that quickly adapts to new patterns will be avoided by bandits in favor of easier prey. Acknowledgments This work was sponsored by NYNEX Science and Technology. The views and conclusions in this paper are those of the authors and do not represent official policy. We thank Nicholas Arcuri and the Fraud Control Department at Bell Atlantic NYNEX Mobile for many useful discussions about cellular fraud and its detection. We also thank Usama Fayyad, Andrea Danyluk and our anonymous referees for comments on drafts. In developing DC-1 we made extensive use of many freely-available software pack- ages. We wish to thank the generous authors, developers and maintainers of the following software: The Perl programming language and many of its user-contributed modules, Donald Tveter's Backprop program, numerous GNU packages including Emacs and GCC, the GMT geographic information system, Gnuplot and the processing system. Appendix State of the Art (SOTA) detector The Of The Art") detector incorporates thirteen profiling methods. Each method profiles an account in a different way and produces a separate alarm. Some of the monitors were designed by hand, but those that employ weights used DC-1's weight tuning methods. Specifically, SOTA contains the following monitors: ffl Two collision detectors, which scan for call collisions of greater than and ffl Two velocity detectors, using velocity thresholds of 400 and 600 miles per hour, respectively. ffl Three "dialed digits" monitors. We created a dialed digit database as follows. We scanned through accounts reserved for rule learning and recorded how many 26 FAWCETT AND PROVOST distinct accounts called a given number both legitimately and in a fraud period. A phone number was discarded if any legitimate subscriber called it; otherwise, a count was saved of the number of times it was called from a cloned phone. Because we did not know an ideal threshold on the number of "hits" required, we created three monitors each with a different threshold. ffl Two daily standard deviation usage monitors. One counted the number of calls on that account-day, one measured total airtime on that account-day. ffl Four "bad cellsite" indicators. It is commonly believed that certain cellsites are the locus for higher than average amounts of fraud, so calls originating from those cellsites might be suspicious. To test this strategy, we tallied the frauded accounts calling each of the cellsites in our region, then computed the percentage of frauded accounts using each cellsite. The twenty worst cellsites were extracted from this list. Using this cellsite list, we created four detectors that counted hits to these "bad cellsites" each in a different way. Notes 1. In fact, because many cellular phones belong to large corporate accounts, often even basic user information such as home town and work location is unavailable. 2. Technically, the numeric value only describes how much above normal the account is. Behavior levels below normal are not considered. 3. If account-days from a single account appear in both training and testing sets, the performance evaluation can be deceptively optimistic. Fraudulent behavior within a specific cloning episode is more similar than fraudulent behavior between episodes. When deployed, the monitors will be used to search for previously unseen cloning episodes. 4. Earlier work (Fawcett and Provost 1996) reported a higher accuracy for SOTA than is shown here. Further development of SOTA revealed that some of its component methods, developed in a prior study, had been built from account data that overlapped data used to test the methods. When a strict separation was enforced, performance declined slightly to the figures shown here. 5. To learn CC 1861 (1054) RL tried to cover the example set with rules each of which covered at least 50 (100) examples and had a Laplace estimate greater than or equal to 0.95, using a beam width of 5000. --R Increasing the efficiency of data mining algorithms with breadth-first marker propagation Exploiting background knowledge in automated discovery. The analysis of time series: An introduction (third edition). Inductive logic programming and knowledge discovery in databases. Quantitative forecasting methods. Combining data mining and machine learning for effective user profiling. Feature selection and extraction. edu/pub/COAST/kumar-phd-intdet Learning machines. Scaling up inductive learning with massive parallelism. Analysis and visualization of classifier performance: Comparison under imprecise class and cost distributions. Learning logical definitions from relations. Generating production rules from decision trees. IEEE ASSP Magazine 3 (1) Learning decision lists using homogeneous rules. Hidden markov models for fault detection in dynamic systems. JAM: Java agents for meta-learning over distributed databases OPUS: An efficient admissible algorithm for unordered search. Recursive estimation and time-series analysis Foster Provost has been a researcher in machine learning and data mining at NYNEX Science and Technology since --TR --CTR Tom E. Fawcett , Foster Provost, Fraud detection, Handbook of data mining and knowledge discovery, Oxford University Press, Inc., New York, NY, 2002 Jiawei Han , Russ B. Altman , Vipin Kumar , Heikki Mannila , Daryl Pregibon, Emerging scientific applications in data mining, Communications of the ACM, v.45 n.8, August 2002 Zhaohui Zheng , Xiaoyun Wu , Rohini Srihari, Feature selection for text categorization on imbalanced data, ACM SIGKDD Explorations Newsletter, v.6 n.1, June 2004 Taeho Jo , Nathalie Japkowicz, Class imbalances versus small disjuncts, ACM SIGKDD Explorations Newsletter, v.6 n.1, June 2004 Padhraic Smyth, Task and method selection: selection of tasks, Handbook of data mining and knowledge discovery, Oxford University Press, Inc., New York, NY, 2002 Signature-Based Methods for Data Streams, Data Mining and Knowledge Discovery, v.5 n.3, p.167-182, July 2001 Tom Fawcett , Peter A. Flach, A response to Webb and Ting's on the application of ROC analysis to predict classification performance under varying class distributions, Machine Learning, v.58 n.1, p.33-38, January 2005 Saharon Rosset , Uzi Murad , Einat Neumann , Yizhak Idan , Gadi Pinkas, Discovery of fraud rules for telecommunicationschallenges and solutions, Proceedings of the fifth ACM SIGKDD international conference on Knowledge discovery and data mining, p.409-413, August 15-18, 1999, San Diego, California, United States Tom E. Fawcett, Industry: adaptive fraud detection, Handbook of data mining and knowledge discovery, Oxford University Press, Inc., New York, NY, 2002 Kenji Yamanishi , Jun-ichi Takeuchi, Discovering outlier filtering rules from unlabeled data: combining a supervised learner with an unsupervised learner, Proceedings of the seventh ACM SIGKDD international conference on Knowledge discovery and data mining, p.389-394, August 26-29, 2001, San Francisco, California Wenke Lee , Salvatore J. Stolfo , Kui W. Mok, Mining in a data-flow environment: experience in network intrusion detection, Proceedings of the fifth ACM SIGKDD international conference on Knowledge discovery and data mining, p.114-124, August 15-18, 1999, San Diego, California, United States Chris Drummond , Robert C. Holte, Explicitly representing expected cost: an alternative to ROC representation, Proceedings of the sixth ACM SIGKDD international conference on Knowledge discovery and data mining, p.198-207, August 20-23, 2000, Boston, Massachusetts, United States Nilesh Dalvi , Pedro Domingos , Mausam , Sumit Sanghai , Deepak Verma, Adversarial classification, Proceedings of the tenth ACM SIGKDD international conference on Knowledge discovery and data mining, August 22-25, 2004, Seattle, WA, USA David Jensen , Matthew Rattigan , Hannah Blau, Information awareness: a prospective technical assessment, Proceedings of the ninth ACM SIGKDD international conference on Knowledge discovery and data mining, August 24-27, 2003, Washington, D.C. data mining approach for database intrusion detection, Proceedings of the 2004 ACM symposium on Applied computing, March 14-17, 2004, Nicosia, Cyprus Jennifer Neville , zgr imek , David Jensen , John Komoroske , Kelly Palmer , Henry Goldberg, Using relational knowledge discovery to prevent securities fraud, Proceeding of the eleventh ACM SIGKDD international conference on Knowledge discovery in data mining, August 21-24, 2005, Chicago, Illinois, USA Michael H. Cahill , Diane Lambert , Jos C. Pinheiro , Don X. Sun, Detecting fraud in the real world, Handbook of massive data sets, Kluwer Academic Publishers, Norwell, MA, 2002 Albert Orriols , Ester Bernad-Mansilla, The class imbalance problem in learning classifier systems: a preliminary study, Proceedings of the 2005 workshops on Genetic and evolutionary computation, June 25-26, 2005, Washington, D.C. Foster Provost , Ron Kohavi, Guest Editors Introduction: On Applied Research in MachineLearning, Machine Learning, v.30 n.2-3, p.127-132, Feb./ March, 1998 Hang , Honghua Dai, Applying both positive and negative selection to supervised learning for anomaly detection, Proceedings of the 2005 conference on Genetic and evolutionary computation, June 25-29, 2005, Washington DC, USA Tom Fawcett, ROC graphs with instance-varying costs, Pattern Recognition Letters, v.27 n.8, p.882-891, June 2006 F. Bonchi , F. Giannotti , G. Mainetto , D. Pedreschi, A classification-based methodology for planning audit strategies in fraud detection, Proceedings of the fifth ACM SIGKDD international conference on Knowledge discovery and data mining, p.175-184, August 15-18, 1999, San Diego, California, United States Tom Fawcett , Foster Provost, Activity monitoring: noticing interesting changes in behavior, Proceedings of the fifth ACM SIGKDD international conference on Knowledge discovery and data mining, p.53-62, August 15-18, 1999, San Diego, California, United States Mukund Narasimhan , Paul Viola , Michael Shilman, Online decoding of Markov models under latency constraints, Proceedings of the 23rd international conference on Machine learning, p.657-664, June 25-29, 2006, Pittsburgh, Pennsylvania Wenke Lee , Salvatore J. Stolfo , Kui W. Mok, Adaptive Intrusion Detection: A Data Mining Approach, Artificial Intelligence Review, v.14 n.6, p.533-567, December 1, 2000 Fabrizio Angiulli , Stefano Basta , Clara Pizzuti, Detection and prediction of distance-based outliers, Proceedings of the 2005 ACM symposium on Applied computing, March 13-17, 2005, Santa Fe, New Mexico Fisher , Daryl Pregibon , Anne Rogers, Hancock: a language for extracting signatures from data streams, Proceedings of the sixth ACM SIGKDD international conference on Knowledge discovery and data mining, p.9-17, August 20-23, 2000, Boston, Massachusetts, United States Tom Fawcett, An introduction to ROC analysis, Pattern Recognition Letters, v.27 n.8, p.861-874, June 2006 Information mining platforms: an infrastructure for KDD rapid deployment, Proceedings of the fifth ACM SIGKDD international conference on Knowledge discovery and data mining, p.327-331, August 15-18, 1999, San Diego, California, United States Markus M. Breunig , Hans-Peter Kriegel , Raymond T. Ng , Jrg Sander, LOF: identifying density-based local outliers, ACM SIGMOD Record, v.29 n.2, p.93-104, June 2000 Rakesh Agrawal , Sridhar Rajagopalan , Ramakrishnan Srikant , Yirong Xu, Mining newsgroups using networks arising from social behavior, Proceedings of the 12th international conference on World Wide Web, May 20-24, 2003, Budapest, Hungary P. Janeja , Vijayalakshmi Atluri , Ahmed Gomaa , Nabil Adam , Christof Bornhoevd , Tao Lin, DM-AMS: employing data mining techniques for alert management, Proceedings of the 2005 national conference on Digital government research, May 15-18, 2005, Atlanta, Georgia Perlich , Foster Provost, Distribution-based aggregation for relational learning with identifier attributes, Machine Learning, v.62 n.1-2, p.65-105, February 2006 Jennifer Neville , David Jensen, Relational Dependency Networks, The Journal of Machine Learning Research, 8, p.653-692, 5/1/2007 Gang Wu , Edward Y. Chang, KBA: Kernel Boundary Alignment Considering Imbalanced Data Distribution, IEEE Transactions on Knowledge and Data Engineering, v.17 n.6, p.786-795, June 2005 Tamas Abraham, Event sequence mining to develop profiles for computer forensic investigation purposes, Proceedings of the 2006 Australasian workshops on Grid computing and e-research, p.145-153, January 16-19, 2006, Hobart, Tasmania, Australia Miroslav Kubat , Robert C. Holte , Stan Matwin, Machine Learning for the Detection of Oil Spills in Satellite Radar Images, Machine Learning, v.30 n.2-3, p.195-215, Feb./ March, 1998 Weng-Keen Wong , Andrew Moore , Gregory Cooper , Michael Wagner, What's Strange About Recent Events (WSARE): An Algorithm for the Early Detection of Disease Outbreaks, The Journal of Machine Learning Research, 6, p.1961-1998, 12/1/2005 Fisher , Daryl Pregibon , Anne Rogers , Frederick Smith, Hancock: A language for analyzing transactional data streams, ACM Transactions on Programming Languages and Systems (TOPLAS), v.26 n.2, p.301-338, March 2004 Klaus Julisch , Marc Dacier, Mining intrusion detection alarms for actionable knowledge, Proceedings of the eighth ACM SIGKDD international conference on Knowledge discovery and data mining, July 23-26, 2002, Edmonton, Alberta, Canada Ting Liu , Andrew W. Moore , Alexander Gray, New Algorithms for Efficient High-Dimensional Nonparametric Classification, The Journal of Machine Learning Research, 7, p.1135-1158, 12/1/2006 Chris Drummond , Robert C. Holte, Cost curves: An improved method for visualizing classifier performance, Machine Learning, v.65 n.1, p.95-130, October 2006 Clifton Phua , Damminda Alahakoon , Vincent Lee, Minority report in fraud detection: classification of skewed data, ACM SIGKDD Explorations Newsletter, v.6 n.1, June 2004 Maarten Van Someren , Tanja Urbani, Applications of machine learning: matching problems to tasks and methods, The Knowledge Engineering Review, v.20 n.4, p.363-402, December 2005 Foster Provost , Tom Fawcett, Robust Classification for Imprecise Environments, Machine Learning, v.42 n.3, p.203-231, March 2001 Sofus A. Macskassy , Foster Provost, Classification in Networked Data: A Toolkit and a Univariate Case Study, The Journal of Machine Learning Research, 8, p.935-983, 5/1/2007 Gediminas Adomavicius , Alexander Tuzhilin, Expert-Driven Validation of Rule-Based User Models in Personalization Applications, Data Mining and Knowledge Discovery, v.5 n.1-2, p.33-58, January-April 2001 Yongwon Lee , Bruce G. Buchanan , John M. Aronis, Knowledge-Based Learning in Exploratory Science: Learning Rules to Predict Rodent Carcinogenicity, Machine Learning, v.30 n.2-3, p.217-240, Feb./ March, 1998 Klaus Julisch, Clustering intrusion detection alarms to support root cause analysis, ACM Transactions on Information and System Security (TISSEC), v.6 n.4, p.443-471, November Foster Provost , Venkateswarlu Kolluri, Data mining tasks and methods: scalability, Handbook of data mining and knowledge discovery, Oxford University Press, Inc., New York, NY, 2002 Predrag Radivojac , Nitesh V. Chawla , A. Keith Dunker , Zoran Obradovic, Classification and knowledge discovery in protein databases, Journal of Biomedical Informatics, v.37 n.4, p.224-239, August 2004 M. A. Maloof , P. Langley , T. O. Binford , R. Nevatia , S. Sage, Improved Rooftop Detection in Aerial Images with Machine Learning, Machine Learning, v.53 n.1-2, p.157-191, October-November Jianping Fan , Hangzai Luo , Jing Xiao , Lide Wu, Semantic video classification and feature subset selection under context and concept uncertainty, Proceedings of the 4th ACM/IEEE-CS joint conference on Digital libraries, June 07-11, 2004, Tuscon, AZ, USA Foster Provost , Venkateswarlu Kolluri, A Survey of Methods for Scaling Up Inductive Algorithms, Data Mining and Knowledge Discovery, v.3 n.2, p.131-169, June 1999 Darse Billings , Lourdes Pea , Jonathan Schaeffer , Duane Szafron, Learning to play strong poker, Machines that learn to play games, Nova Science Publishers, Inc., Commack, NY, 2001
rule learning;constructive induction;intrusion detection;profiling;fraud detection;applications
593453
Parallel Algorithms for Discovery of Association Rules.
Discovery of association rules is an important data mining task. Several parallel and sequential algorithms have been proposed in the literature to solve this problem. Almost all of these algorithms make repeated passes over the database to determine the set of frequent itemsets (a subset of database items), thus incurring high I/O overhead. In the parallel case, most algorithms perform a sum-reduction at the end of each pass to construct the global counts, also incurring high synchronization cost.In this paper we describe new parallel association mining algorithms. The algorithms use novel itemset clustering techniques to approximate the set of potentially maximal frequent itemsets. Once this set has been identified, the algorithms make use of efficient traversal techniques to generate the frequent itemsets contained in each cluster. We propose two clustering schemes based on equivalence classes and maximal hypergraph cliques, and study two lattice traversal techniques based on bottom-up and hybrid search. We use a vertical database layout to cluster related transactions together. The database is also selectively replicated so that the portion of the database needed for the computation of associations is local to each processor. After the initial set-up phase, the algorithms do not need any further communication or synchronization. The algorithms minimize I/O overheads by scanning the local database portion only twice. Once in the set-up phase, and once when processing the itemset clusters. Unlike previous parallel approaches, the algorithms use simple intersection operations to compute frequent itemsets and do not have to maintain or search complex hash structures.Our experimental testbed is a 32-processor DEC Alpha cluster inter-connected by the Memory Channel network. We present results on the performance of our algorithms on various databases, and compare it against a well known parallel algorithm. The best new algorithm outperforms it by an order of magnitude.
Introduction With recent progress in automated data gathering, and the availability of cheap storage, a lot of businesses have routinely started collecting massive amounts of data on various facets of the organization. The eventual goal of this data gathering is to be able to use this information to gain a competitive edge, by discovering previously unknown patterns in the data, which can guide the decision making. Such high-level inference process may provide a host of useful information on customer groups, buying patterns, stock trends, etc. This process of automatic information inferencing is commonly known as Knowledge Discovery and Data mining (KDD). We look at one of the central KDD tasks - mining for associations. Discovery of association rules is an important problem in database mining. The prototypical application is the analysis of sales or basket data (Agrawal et al. 1996). Basket data consists of items bought by a customer along with the transaction identifier. Besides the retail sales example, association rules have been shown to be useful in domains such as decision support, telecommunications alarm diagnosis and predic- tion, university enrollments, etc. 1.1. Problem Statement The problem of mining associations over basket data was introduced in (Agrawal, Imielinski, & Swami 1993). It can be formally stated as: Let I be a set of m distinct attributes, also called items. Each transaction T in the database D of transactions, has a unique identifier, and contains a set of items, called itemset, such that T ' I, i.e. each transaction is of the itemset with k items is called a k-itemset. A subset of length k is called a k-subset. An itemset is said to have a support s if s% of the transactions in D contain the itemset. An association rule is an expression A ) B, where itemsets A; B ae I, and ;. The confidence of the association rule, given as B)=support(A), is simply the conditional probability that a transaction contains B, given that it contains A. The data mining task for association rules can be broken into two steps. The first step consists of finding all frequent itemsets, i.e., itemsets that occur in the database with a certain user-specified frequency, called minimum support. The second step consists of forming conditional implication rules among the frequent itemsets (Agrawal & Srikant 1994). The second step is relatively straightforward. Once the support of frequent itemsets is known, rules of the form are generated for all frequent itemsets , provided the rules meet a desired confidence. On the other hand the problem of identifying all frequent itemsets is hard. Given m items, there are potentially frequent itemsets, which form a lattice of subsets over I. However, only a small fraction of the whole lattice space is frequent. Discovering the frequent itemsets requires a lot of computation power, memory and disk I/O, which can only be provided by parallel computers. Efficient parallel methods are needed to discover the relevant itemsets, and this is the focus of our paper. 1.2. Related Work Sequential Algorithms Several algorithms for mining associations have been proposed in the literature. The Apriori algorithm (Mannila, Toivonen, & Verkamo 1994; Agrawal & Srikant 1994; Agrawal et al. 1996) was shown to have superior performance to earlier approaches (Agrawal, Imielinski, & Swami 1993; Park, Chen, et al. 1995; Houtsma & Swami 1995) and forms the core of almost all of the current algorithms. Apriori uses the downward closure property of itemset support to prune the itemset lattice - the property that all subsets of a PARALLEL ASSOCIATION RULES 3 frequent itemset must themselves be frequent. Thus only the frequent k-itemsets are used to construct candidate 1)-itemsets. A pass over the database is made at each level to find the frequent itemsets among the candidates. For very large disk resident databases, these algorithms incur high I/O overhead for scanning it in each iteration. The Partition algorithm (Savasere, Omiecinski, & Navathe 1995) minimizes I/O by scanning the database only twice. It partitions the database into small chunks which can be handled in memory. In the first pass it generates the set of all potentially frequent itemsets (any itemset locally frequent in a partition), and in the second pass their global support is obtained. Another way to minimize the I/O overhead is to work with only a small random sample of the database (Toivonen 1996; Zaki et al. 1997a). We recently proposed new algorithms (Zaki et al. 1997b) which scan the database only once, generating all frequent itemsets. These new algorithms were shown to outperform previous Apriori based approaches by more than an order of magnitude (Zaki et al. 1997b). The performance gains are obtained by using effective itemset clustering and lattice traversal techniques. This paper presents efficient parallel implementations of these new algorithms. Parallel Algorithms There has been relatively less work in parallel mining of associations. Three different parallelizations of Apriori on a distributed-memory machine (IBM SP2) were presented in (Agrawal & Shafer 1996). The Count Distribution algorithm is a straight-forward parallelization of Apriori. Each processor generates the partial support of all candidate itemsets from its local database parti- tion. At the end of each iteration the global supports are generated by exchanging the partial supports among all the processors. The Data Distribution algorithm partitions the candidates into disjoint sets, which are assigned to different proces- sors. However to generate the global support each processor must scan the entire database (its local partition, and all the remote partitions) in all iterations. It thus suffers from huge communication overhead. The Candidate Distribution algorithm also partitions the candidates, but it selectively replicates the database, so that each processor proceeds independently. The local database portion is still scanned in every iteration. Count Distribution was shown to have superior performance among these three algorithms (Agrawal & Shafer 1996). Other parallel algorithms improving upon these ideas in terms of communication efficiency, or aggregate memory utilization have also been proposed (Cheung et al. 1996b; 1996a; Han, Karypis, & Kumar 1997). The PDM algorithm (Park, Chen, & Yu 1995b) presents a parallelization of the DHP algorithm (Park, Chen, & Yu 1995a). How- ever, PDM performs worse than Count Distribution (Agrawal & Shafer 1996). In recent work we presented the CCPD parallel algorithm for shared-memory machines (Zaki et al. 1996). It is similar in spirit to Count Distribution. The candidate itemsets are generated in parallel and are stored in a hash structure which is shared among all the processors. Each processor then scans its logical partition of the database and atomically updates the counts of candidates in the shared hash tree. CCPD uses additional optimization such as candidate balancing, hash-tree balancing and short-circuited subset counting to speed up performance (Zaki et al. 1996). We also presented a new parallel algorithm Eclat (Zaki, Parthasarathy, & Li 1997) on a DEC Alpha Cluster. Eclat uses the equivalence class itemset clustering scheme along with a bottom-up lattice traversal. It was shown to outperform Count Distribution by more than an order of magnitude. This paper will present parallelization results on new clustering and traversal techniques. 1.3. Contribution The main limitation of all the current parallel algorithms (Park, Chen, & Yu 1995b; Zaki et al. 1996; Agrawal & Shafer 1996; Cheung et al. 1996b; 1996a) is that they make repeated passes over the disk-resident database partition, incurring high I/O overhead. Furthermore, the schemes involve exchanging either the counts of candidates or the remote database partitions during each iteration. This results in high communication and synchronization overhead. The previous algorithms also use complicated hash structures which entails additional overhead in maintaining and searching them, and typically suffer from poor cache locality (Parthasarathy, Our work contrasts to these approaches in several ways. We present new parallel algorithms for fast discovery of association rules based on our ideas in (Zaki, Parthasarathy, & Li 1997; Zaki et al. 1997b). The new parallel algorithms are characterized in terms of the clustering information used to group related itemsets, and in terms of the lattice traversal schemes used to search for frequent itemsets. We propose two clustering schemes based on equivalence classes and maximal uniform hypergraph cliques, and we utilize two lattice traversal schemes, based on bottom-up and hybrid top-down/bottom-up search. The algorithms also use a different database layout which clusters related transactions together, and the work is distributed among the processors in such a way that each processor can compute the frequent itemsets independently, using simple intersection operations. An interesting benefit of using simple intersections is that the algorithms we propose can be implemented directly on general purpose database systems (Holsheimer et al. 1995; Houtsma & Swami 1995). These techniques eliminate the need for synchronization after the initial set-up phase, and enable us to scan the database only two times, drastically cutting down the I/O overhead. Our experimental testbed is a 32- processor DEC Alpha SMP cluster (8 hosts, 4 processors/host) inter-connected by the Memory Channel (Gillett 1996) network. The new parallel algorithms are also novel in that they utilize this machine configuration information, i.e., they assume a distributed-memory model across the 8 cluster hosts, but assume a shared-memory model for the 4 processors on each host. We experimentally show that our new algorithms outperform the well known Count Distribution algorithm. We also present extensive performance results on their speedup, sizeup, communication cost and memory usage. The rest of the paper is organized as follows. We begin by providing more details on the sequential Apriori algorithm. Section 3 describes some of the previous Apriori based parallel algorithms. We then present the main ideas behind our new algorithms - the itemset and transaction clustering, and the lattice traversal techniques, in section 4. Section 5 describes the design and implementation of the PARALLEL ASSOCIATION RULES 5 new parallel algorithms. Our experimental study is presented in section 6, and our conclusions in section 7. 2. Sequential Apriori Algorithm In this section we will briefly describe the Apriori algorithm (Agrawal et al. 1996), since it forms the core of all parallel algorithms (Agrawal & Shafer 1996; Cheung et al. 1996b; 1996a; Han, Karypis, & Kumar 1997; Park, Chen, & Yu 1995b; Zaki et al. 1996). Apriori uses the downward closure property of itemset support that any subset of a frequent itemset must also be frequent. Thus during each iteration of the algorithm only the itemsets found to be frequent in the previous iteration are used to generate a new candidate set. A pruning step eliminates any candidate at least one of whose subsets is not frequent. The complete algorithm is shown in table 1. It has three main steps. The candidates for the k-th pass are generated by joining L k\Gamma1 with itself, which can be expressed as: C and denotes the i-th item, and X [i : j] denotes items at index i through j in itemset X . For example, let L BC, BD, BE, DEg. Then C BDEg. Table 1. The Apriori Algorithm 1. ffrequent 1-itemsets 2. for 3. 4. for all transactions t 2 D 5. for all k-subsets s of t 7. 8. Set of all frequent itemsets = Before inserting an itemset into C k , Apriori tests whether all its are frequent. This pruning step can eliminate a lot of unnecessary candidates. The candidates, C k , are stored in a hash tree to facilitate fast support counting. An internal node of the hash tree at depth d contains a hash table whose cells point to nodes at depth d + 1. All the itemsets are stored in the leaves. The insertion procedure starts at the root, and hashing on successive items, inserts a candidate in a leaf. For counting C k , for each transaction in the database, all k-subsets of the transaction are generated in lexicographical order. Each subset is searched in the hash tree, and the count of the candidate incremented if it matches the subset. This is the most compute intensive step of the algorithm. The last step forms L k by selecting itemsets meeting the minimum support criterion. For details on the 6 ZAKI, PARTHASARATHY, OGIHARA AND LI performance characteristics of Apriori we refer the reader to (Agrawal & Srikant 1994). 3. Apriori-based Parallel Algorithms In this section we will look at some previous parallel algorithms. These algorithms assume that the database is partitioned among all the processors in equal-sized blocks, which reside on the local disk of each processor. The Count Distribution algorithm (Agrawal & Shafer 1996) is a simple parallelization of Apriori. All processors generate the entire candidate hash tree from Each processor can thus independently get partial supports of the candidates from its local database partition. This is followed by a sum-reduction to obtain the global counts. Note that only the partial counts need to be communicated, rather than merging different hash trees, since each processor has a copy of the entire tree. Once the global L k has been determined each processor builds C k+1 in par- allel, and repeats the process until all frequent itemsets are found. This simple algorithm minimizes communication since only the counts are exchanged among the processors. However, since the entire hash tree is replicated on each processor, it doesn't utilize the aggregate memory efficiently. The implementation of Count Distribution used for comparison in our experiments differs slightly from the above description and is optimized for our testbed configuration. Only one copy of the hash tree resides on each of the 8 hosts in our cluster. All the 4 processors on each host share this hash tree. Each processor still has its own local database portion and uses a local array to gather the local candidate support. The sum-reduction is accomplished in two steps. The first step performs the reduction only among the local processors on each host. The second step performs the reduction among the hosts. We also utilize some optimization techniques such as hash-tree balancing and short-circuited subset counting (Zaki et al. 1996) to further improve the performance of Count Distribution. The Data Distribution algorithm (Agrawal & Shafer 1996) was designed to utilize the total system memory by generating disjoint candidate sets on each processor. However to generate the global support each processor must scan the entire database (its local partition, and all the remote partitions) in all iterations. It thus suffers from high communication overhead, and performs very poorly when compared to Count Distribution (Agrawal & Shafer 1996). The Candidate Distribution algorithm (Agrawal & Shafer 1996) uses a property of frequent itemsets (Agrawal & Shafer 1996; Zaki et al. 1996) to partition the candidates during iteration l, so that each processor can generate disjoint candidates independent of other processors. At the same time the database is selectively replicated so that a processor can generate global counts independently. The choice of the redistribution pass involves a trade-off between decoupling processor dependence as soon as possible and waiting until sufficient load balance can be achieved. In their experiments the repartitioning was done in the fourth pass. After this the only dependence a processor has on other processors is for pruning the candidates. Each processor asynchronously broadcasts the local frequent set to other processors PARALLEL ASSOCIATION RULES 7 during each iteration. This pruning information is used if it arrives in time, otherwise it is used in the next iteration. Note that each processor must still scan its local data once per iteration. Even though it uses problem-specific information, it performs worse than Count Distribution (Agrawal & Shafer 1996). Candidate Distribution pays the cost of redistributing the database, and it then scans the local database partition repeatedly, which will usually be larger than jjDjj=P . 4. Efficient Clustering and Traversal Techniques In this section we present our techniques to cluster related frequent itemsets together using equivalence classes and maximal uniform hypergraph cliques. We then describe the bottom-up and hybrid itemset lattice traversal techniques. We also present a technique to cluster related transactions together by using the vertical database layout. This layout is able to better exploit the proposed clustering and traversal schemes. It also facilitates fast itemset support counting using simple intersections, rather than maintaining and searching complex data structures. 4.1. Itemset Clustering Lattice of Subsets of {1,2,3,4,5} Border of Frequent Itemsets Lattice of Subsets of {1,2,3,4} Lattice of Subsets of {3,4,5} Sublattices Induced by Maximal Itemsets Figure 1. Lattice of Subsets and Maximal Itemset Induced Sub-lattices We will motivate the need for itemset clustering by means of an example. Consider the lattice of subsets of the set f1; 2; 3; 4; 5g, shown in figure 1 (the empty set has been omitted in all figures). The frequent itemsets are shown with dashed circles and the two maximal frequent itemsets (a frequent itemset is maximal if it is not a proper subset of any other frequent itemset) are shown with the bold circles. Due to the downward closure property of itemset support - the fact that all subsets of a frequent itemset must be frequent - the frequent itemsets form a border, such that all frequent itemsets lie below the border, while all infrequent itemsets lie above it. The border of frequent itemsets is shown with a bold line in figure 1. An optimal association mining algorithm will only enumerate and test the frequent itemsets, i.e., the algorithm must efficiently determine the structure of the border. This structure is precisely determined by the maximal frequent itemsets. The border corresponds to the sub-lattices induced by the maximal frequent itemsets. These sub-lattices are shown in figure 1. Given the knowledge of the maximal frequent itemsets we can design an efficient algorithm that simply gathers their support and the support of all their subsets in just a single database pass. In general we cannot precisely determine the maximal itemsets in the intermediate steps of the algorithm. However we can approximate this set. Our itemset clustering techniques are designed to group items together so that we obtain supersets of the maximal frequent itemsets - the potential maximal frequent itemsets. Below we present two schemes to generate the set of potential maximal itemsets based on equivalence classes and maximal uniform hypergraph cliques. These two techniques represent a trade-off in the precision of the potential maximal itemsets generated, and the computation cost. The hypergraph clique approach gives more precise information at higher computation cost, while the equivalence class approach sacrifices quality for a lower computation cost. 4.1.1. Equivalence Class Clustering Let's reconsider the candidate generation step of Apriori. Let L f ABC, ABD, ABE, ACD, ACE, ADE, BCD, BCE, BDEg. Assuming that L k\Gamma1 is lexicographically sorted, we can partition the itemsets in L k\Gamma1 into equivalence classes based on their common prefixes, i.e., the equivalence class a 2 L k\Gamma2 , is given as: Candidate k-itemsets can simply be generated from itemsets within a class by joining all pairs, with the the class identifier as the prefix. For our example we obtain the equivalence classes: fEg. We observe that itemsets produced by the equivalence class [A], namely those in the set fABC, ABD, ABE, ACD, ACE, ADEg, are independent of those produced by the class [B] (the set fBCD, BCE, BDEg). Any class with only 1 member can be eliminated since no candidates can be generated from it. Thus we can discard the class [D]. This idea of partitioning L into equivalence classes was independently proposed in (Agrawal & Shafer 1996; Zaki et al. 1996). The equivalence partitioning was used in (Zaki et al. 1996) to parallelize the candidate generation step in CCPD. It was also used in Candidate Distribution (Agrawal & Shafer 1996) to partition the candidates into disjoint sets. PARALLEL ASSOCIATION RULES 9 At any intermediate step of the algorithm when the set of frequent itemsets, L k for - 2, has been determined we can generate the set of potential maximal frequent itemsets from L k . Note that for with the entire item universe as the maximal itemset. However, For any k - 2, we can extract more precise knowledge about the association among the items. The larger the value of k the more precise the clustering. For example, figure 2 shows the equivalence classes obtained for the instance where 2. Each equivalence class is a potential maximal frequent itemset. For example, the class [1], generates the maximal itemset 12345678. 4.1.2. Maximal Uniform Hypergraph Clique Clustering Let the set of items I denote the vertex set. A hypergraph (Berge 1989) on I is a family of edges or subsets of I, such that E i 6= ;, and [ n simple hypergraph is a hypergraph such simple graph is a simple hyper-graph each of whose edges has cardinality 2. The maximum edge cardinality is called the rank, j. If all edges have the same cardinality, then H is called a uniform hypergraph. A simple uniform hypergraph of rank r is called a r-uniform hypergraph. For a subset X ae I, the sub-hypergraph induced by X is given as, ng. A r-uniform complete hypergraph with m vertices, denoted as K r , consists of all the r-subsets of I. A r-uniform complete sub-hypergraph is called a r-uniform hypergraph clique. A hypergraph clique is maximal if it is not contained in any other clique. For hypergraphs of rank 2, this corresponds to the familiar concept of maximal cliques in a graph. Given the set of frequent itemsets L k , it is possible to further refine the clustering process producing a smaller set of potentially maximal frequent itemsets. The observation used is that given any frequent m-itemset, for m ? k, all its k-subsets must be frequent. In graph-theoretic terms, if each item is a vertex in the hypergraph, and each k-subset an edge, then the frequent m-itemset must form a k-uniform hypergraph clique. Furthermore, the set of maximal hypergraph cliques represents an approximation or upper-bound on the set of maximal potential frequent itemsets. All the "true" maximal frequent itemsets are contained in the vertex set of the maximal cliques, as stated formally in the lemma below. HLk be the k-uniform hypergraph with vertex set I, and edge set L k . Let C be the set of maximal hypergraph cliques in H, i.e., let M be the set of vertex sets of the cliques in C. Then for all maximal frequent itemsets f , 9t 2 M , such that f ' t. An example of uniform hypergraph clique clustering is given in figure 2. The example is for the case of L 2 , and thus corresponds to an instance of the general clustering technique, which reduces to the case of finding maximal cliques in regular graphs. The figure shows all the equivalence classes, and the maximal cliques within them. It also shows the graph for class 1, and the maximal cliques in it. It can be seen immediately the the clique clustering is more accurate than equivalence class clustering. For example, while equivalence class clustering produced the potential maximal frequent itemset 12345678, the hypergraph clique 12, 13, 14, 15, 16, 17, 18, 23, 25, 27, 28, 34, 35, 36, 45, 46, 56, 58, 68, 78 Frequent 2-Itemsets Equivalence Class Graph Maximal Cliques For Equivalence Class 1 Maximal Cliques per Class Equivalence Classes Figure 2. Equivalence Class and Uniform Hypergraph Clique Clustering clustering produces a more refined set f1235; 1258; 1278; 13456; 1568g for equivalence class [1]. The maximal cliques are discovered using a dynamic programming algorithm. For a class [x], and y 2[x], y is said to cover the subset of [x], given by For each class C, we first identify its covering set, given as fy 2 Cjcov(y) 6= ;; and cov(y) 6' cov(z); for any z 2 C; z ! yg. We recursively generate the maximal cliques for elements in the covering set for each class. Each maximal clique from the covering set is prefixed with the class identifier (eliminat- ing any duplicates) to obtain the maximal cliques for the current class (see (Zaki et al. 1997c) for details). For general graphs the maximal clique decision problem is NP-Complete (Garey & Johnson 1979). However, the equivalence class graph is usually sparse and the maximal cliques can be enumerated efficiently. As the edge density increases the clique based approaches may suffer. Some of the factors affecting the edge density include decreasing support and increasing transaction size. The effect of these parameters was presented in (Zaki et al. 1997b). PARALLEL ASSOCIATION RULES 11 4.2. Lattice Traversal The equivalence class and uniform hypergraph clique clustering schemes generate the set of potential maximal frequent itemsets. Each such potential maximal item-set induces a sub-lattice of the lattice of subsets of database items I. We now have to traverse each of these sub-lattices to determine the "true" frequent item- sets. Our goal is to devise efficient schemes to precisely determine the structure of the border of frequent itemsets. Different ways of expanding the frequent itemset border in the lattice space are possible. Below we present two schemes to traverse the sublattices. One is a purely bottom-up approach, while the other is a hybrid top-down/bottom-up scheme.16 13 14 12 Support Potential Maximal Frequent Itemset (123456) Sort Itemsets by Support Top-Down Phase True Maximal Frequent Itemsets: 1235, 13456 HYBRID TRAVERSAL Figure 3. Bottom-up and Hybrid Lattice Traversal 4.2.1. Bottom-up Lattice Traversal Consider the example shown in figure 3. It shows a particular instance of the clustering schemes which uses L 2 to generate the set of potential maximal itemsets. Let's assume that for equivalence class [1], there is only one potential maximal itemset, 123456, while 1235 and 13456 are "true" maximal frequent itemsets. The supports of 2-itemsets in this class are also shown. Like figure 1, the dashed circles represent the frequent sets, the bold circles the maximal such itemsets, and the boxes denote equivalence classes. The potential maximal itemset 123456 forms a lattice over the elements of equivalence class 16g. We need to traverse this lattice to determine the "true" frequent itemsets. A pure bottom-up lattice traversal proceeds in a breadth-first manner generating frequent itemsets of length k, before generating itemsets of level k 1, i.e., at each intermediate step we determine the border of frequent k-itemsets. For example, all pairs of elements of [1] are joined to produce new equivalence classes of frequent 3-itemsets, namely (producing the maximal itemset 1235), 6g. The next step yields the frequent class, producing the maximal itemset 13456). Most current algorithms use this approach. For example, the process of generating C k from L k\Gamma1 used in Apriori (Agrawal et al. 1996), and related algorithms (Savasere, Omiecinski, & Navathe 1995; Park, 1995a), is a pure bottom-up exploration of the lattice space. Since this is a bottom-up approach all the frequent subsets of the maximal frequent itemsets are generated in intermediate steps of the traversal. 4.2.2. Hybrid Top-down/Bottom-up Search The bottom-up approach doesn't make full use of the clustering information. While it uses the cluster to restrict the search space, it may generate spurious candidates in the intermediate steps, since the fact that all subsets of an itemset are frequent doesn't guarantee that the itemset is frequent. For example, the itemsets 124 and 126 in figure 3 are infre- quent, even though 12, 14, and 16 are frequent. We can envision other traversal techniques which quickly identify the set of true maximal frequent itemsets. Once this set is known we can either choose to stop at this point if we are interested in only the maximal itemsets, or we can gather the support of all their subsets as well (all subsets are known to be frequent by definition). In this paper we will restrict our attention to only identifying the maximal frequent subsets. One possible approach is to perform a pure top-down traversal on each cluster or sublattice. This scheme may be thought of as trying to determine the border of infrequent itemsets, by starting at the top element of the lattice and working our way down. For example, consider the potential maximal frequent itemset 123456 in figure 3. If it turns out to be frequent we are done. But in this case it is not frequent, so we then have to check whether each of its 5-subsets is frequent. At any step, if a k-subset turns out to be frequent, we need not check any of its subsets. This approach doesn't work well in practice, since the clusters are only an approximation of the maximal frequent itemsets, and a lot of infrequent supersets of the "true" maximal frequent itemsets may be generated. In our example we would generate 10 infrequent itemsets - 123456, 12345, 12346, 12356, 12456, 1234, 1245, 1236, 1246, and 1256 - using the pure top-down scheme, instead of only two infrequent itemsets generated in the pure bottom-up approach - 124, and 126. We therefore propose a hybrid top-down and bottom-up approach that works well in practice. PARALLEL ASSOCIATION RULES 13 The basic idea behind the hybrid approach is to quickly determine the "true" maximal itemsets, by starting with a single element from a cluster of frequent k- itemsets, and extending this by one more itemset till we generate an infrequent itemset. This comprises the top-down phase. In the bottom-up phase, the remaining elements are combined with the elements in the first set to generate all the additional frequent itemsets. An important consideration in the top-down phase is to determine which elements of the cluster should be combined. In our approach we first sort the itemsets in the cluster in descending order of their support. We start with the element with maximum support, and extend it with the next element in the sorted order. This approach is based on the intuition that the larger the support the more the likely is the itemset to be a part of a larger itemset. Figure 3 shows an example of the hybrid scheme on a cluster of 2-itemsets. We sort the 2-itemsets in decreasing order of support, intersecting 16 and 15 to produce 156. This is extended to 1356 by joining 156 with 13, and then to 13456, and finally we find that 123456 is infrequent. The only remaining element is 12. We simply join this with each of the other elements producing the frequent itemset class [12], which generates the other maximal itemset 1235. The bottom-up and hybrid approaches are contrasted in figure 3, and the pseudo-code for both schemes is shown in table 3. 4.3. Transaction Clustering: Database Layout The KDD process consists of various steps (Fayyad, Piatetsky-Shapiro, & Smyth 1996). The initial step consists of creating the target dataset by focusing on certain attributes or via data samples. The database creation may require removing unnecessary information and supplying missing data, and transformation techniques for data reduction and projection. The user must then determine the data mining task and choose a suitable algorithm, for example, the discovery of association rules. The next step involves interpreting the discovered associations, possibly looping back to any of the previous steps, to discover more understandable patterns. An important consideration in the data preprocessing step is the final representation or data layout of the dataset. Another issue is whether some preliminary invariant information can be gleaned during this process. There are two possible layouts of the target dataset for association mining - the horizontal and the vertical layout. 4.3.1. Horizontal Data Layout This is the format standardly used in the literature (see e.g., (Agrawal & Srikant 1994; Mannila, Toivonen, & Verkamo 1994; Agrawal et al. 1996)). Here a dataset consists of a list of transactions. Each transaction has a transaction identifier (TID) followed by a list of items in that transaction. This format imposes some computation overhead during the support counting step. In particular, for each transaction of average length l, during iteration k, we have to generate and test whether all \Gamma l k-subsets of the transaction are contained in C k . To perform fast subset checking the candidates are stored in a complex hash-tree data structure. Searching for the relevant candidates thus adds additional computation overhead. Furthermore, the horizontal layout forces us to 14 ZAKI, PARTHASARATHY, OGIHARA AND LI ITEMS TIDS ITEMS TIDS Horizontal Layout Vertical Layout Figure 4. Horizontal and Vertical Database Layout scan the entire database or the local partition once in each iteration. Both Count and Candidate Distribution must pay the extra overhead entailed by using the horizontal layout. Furthermore, the horizontal layout seems suitable only for the bottom-up exploration of the frequent border. It appears to be extremely complicated to implement the hybrid approach using the horizontal format. An alternative approach is to store all the potential maximal itemsets and all their subsets in a data structure with fast look-up, (e.g., hash-trees (Agrawal et al. 1996)). We can then gather their support in a single database scan. We plan to explore this in a later paper. 4.3.2. Vertical Data Layout In the vertical (or inverted) layout (also called the decomposed storage structure (Holsheimer et al. 1995)), a dataset consists of a list of items, with each item followed by its tid-list - the list of all the transactions identifiers containing the item. An example of successful use of this layout can be found in (Holsheimer et al. 1995; Savasere, Omiecinski, & Navathe 1995; Zaki, Parthasarathy, & Li 1997; Zaki et al. 1997b). The vertical layout doesn't suffer from any of the overheads described for the horizontal layout above due to the following three reasons: First, if the tid-list is sorted in increasing order, then the support of a candidate k-itemset can be computed by simply intersecting the tid-lists of any two data structures need to be maintained. We don't have to generate all the k-subsets of a transaction or perform the search operations on the hash tree. Second, the tid-lists contain all relevant information about an itemset, and enable us to avoid scanning the whole database to compute the support count of an itemset. This layout can therefore take advantage of the principle of locality. All frequent itemsets from a cluster of itemsets can be generated, before PARALLEL ASSOCIATION RULES 15 moving on to the next cluster. Third, the larger the itemset, the shorter the tid- lists, which is practically always true. This results in faster intersections. For example, consider figure 4, which contrasts the horizontal and the vertical layout (for simplicity, we have shown the null elements, while in reality sparse storage is used). The tid-list of A, is given as 4g. Then the tid-list of AB is simply, T 4g. We can immediately determine the support by counting the number of elements in the tid-list. If it meets the minimum support criterion, we insert AB in L 2 . The intersections among the tid-lists can be performed faster by utilizing the minimum support value. For example let's assume that the minimum support is 100, and we are intersecting two itemsets - AB with support 119 and AC with support 200. We can stop the intersection the moment we have 20 mismatches in AB, since the support of ABC is bounded above by 119. We use this optimization, called short-circuited intersection, for fast joins. The inverted layout, however, has a drawback. Examination of small itemsets tends to be costlier than when the horizontal layout is employed. This is because tid-lists of small itemsets provide little information about the association among items. In particular, no such information is present in the tid-lists for 1-itemsets. For example, a database with 1,000,000 (1M) transactions, 1,000 frequent items, and an average of 10 items per transaction has tid-lists of average size 10,000. To find frequent 2-itemsets we have to intersect each pair of items, which requires operations. On the other hand, in the horizontal format we simply need to form all pairs of the items appearing in a transaction and increment their count, requiring only \Gamma There are a number of possible solutions to this problem: 1. To use a preprocessing step to gather the occurrence count of all 2-itemsets. Since this information is invariant, it has to be performed once during the lifetime of the database, and the cost can be amortized over the number of times the data is mined. This information can also be incrementally updated as the database changes over time. 2. To store the counts of only those 2-itemsets with support greater than a user specified lower bound, thus requiring less storage than the first approach. 3. To use a small sample that would fit in memory, and determine a superset of the frequent 2-itemsets, L 2 , by lowering the minimum support, and using simple intersections on the sampled tid-lists. Sampling experiments (Toivonen 1996; Zaki et al. 1997a) indicate that this is a feasible approach. Once the superset has been determined we can easily verify the "true" frequent itemsets among them. Our current implementation uses the pre-processing approach due to its simplicity. We plan to implement the sampling approach in a later paper. The solutions represent different trade-offs. The sampling approach generates L 2 on-the-fly with an extra database pass, while the pre-processing approach requires extra storage. For m items, count storage requires O(m 2 ) disk space, which can be quite large for large values of m. However, for used in our experiments this adds only Memory Channel Address Receive Transmit Transmit Receive Receive Figure 5. Memory Channel space. The lined region is mapped for both transmit and receive on node 1 and for receive on node 2. The gray region is mapped for receive on node 1 and for transmit on node 2. a very small extra storage overhead. Using the second approach can further reduce the storage requirements, but may require an extra scan if the lower bound on support is changed. Note also that the database itself requires the same amount of memory in both the horizontal and vertical formats (this is obvious from figure 4). 5. New Parallel Algorithms: Design and Implementation 5.1. The DEC Memory Channel Digital's Memory Channel (MC) network (Gillett 1996) provides applications with a global address space using memory mapped regions. A region can be mapped into a process' address space for transmit, receive, or both. Virtual addresses for regions map into physical addresses located in I/O space on the MC's PCI adapter. Virtual addresses for receive regions map into physical RAM. Writes into transmit regions are collected by the source MC adapter, forwarded to destination MC adapters through a hub, and transferred via DMA to receive regions with the same global identifier (see figure 5). Regions within a node can be shared across different processors on that node. Writes originating on a given node will be sent to receive regions on that same node only if loop-back has been enabled for the region. We do not use the loop-back feature. We use write-doubling instead, where each processor writes to its receive region and then to its transmit region, so that processes on a host can see modification made by other processes on the same host. Though we pay the cost of double writing, we reduce the amount of messages to the hub. In our system unicast and multicast process-to-process writes have a latency of 5.2 -s, with per-link transfer bandwidths of MB/s. MC peak aggregate bandwidth is also about 32 MB/s. Memory Channel guarantees write ordering and local cache PARALLEL ASSOCIATION RULES 17 coherence. Two writes issued to the same transmit region (even on different nodes) will appear in the same order in every receive region. When a write appears in a receive region it invalidates any locally cached copies of its line. 5.2. Initial Database Partitioning We assume that the database is in the vertical format, and that we have the support counts of all 2-itemsets available locally on each host. We further assume that the database of tid-lists is initially partitioned among all the hosts. This partitioning is done off-line, similar to the assumption made in Count Distribution (Agrawal & Shafer 1996). The tid-lists are partitioned so that the total length of all tid-lists in the local portions on each host are roughly equal. This is achieved using a greedy algorithm. The items are sorted on their support, and the next item is assigned to the least loaded host. Note that the entire tid-list for an item resides on a host. Figure 6 shows the original database, and the resultant initial partition on two processors. 5.3. New Parallel Algorithms We present four new parallel algorithms, depending on the clustering and lattice traversal scheme used: ffl Par-Eclat: uses equivalence class clustering and bottom-up lattice traversal. ffl Par-MaxEclat: uses equivalence class clustering and hybrid traversal. ffl Par-Clique: uses maximal uniform hypergraph clique clustering and bottom-up lattice traversal. ffl Par-MaxClique: uses maximal uniform hypergraph clique clustering and hybrid traversal. The algorithms using the bottom-up lattice traversal, namely Par-Eclat and Par- Clique, generate all frequent itemsets, while those using the hybrid traversal, namely Par-MaxEclat and Par-MaxClique, generate only the maximal frequent itemsets. As noted earlier, it is trivial to modify the hybrid traversal algorithms to generate all frequent itemsets. But here we are interested in examining the benefits of quickly identifying the maximal elements for the hybrid scheme. Below we present the parallel design and implementation issues, which are applicable to all four algorithms. 5.4. Parallel Design and Implementation The new algorithms overcome the shortcomings of the Count and Candidate Distribution algorithms. They utilize the aggregate memory of the system by partitioning the itemset clusters into disjoint sets, which are assigned to different processors. The dependence among the processors is decoupled right in the beginning so that the Equivalence Classes Equivalence Class Weights [1]: 6 Equivalence Class Assignment P1: [2], [3] Partitioned Database Original Database: Tid-Lists After Tid-List Communication Sublattices for P1: [2], [3] Sublattice for P0: [1] Sublattice Induced by L2 Figure 6. Database Partitioning and Cluster Scheduling PARALLEL ASSOCIATION RULES 19 Table 2. Pseudo-code for the New Parallel Algorithms 1. Begin ParAssociation: 2. /* Initialization Phase*/ 3. Form L2 from 2-itemset support counts 4. Generate Clusters from L2 using: 5. Equivalence Classes or Uniform Hypergraph Cliques 6. Partition Clusters among the processors P 7. Scan local database partition 8. Transmit relevant tid-lists to other processors 9. Receive tid-lists from other processors 10. /* Asynchronous Phase */ 11. for each assigned Cluster, C2 12. Compute Frequent Itemsets: Bottom-Up(C2) or Hybrid(C2) 13. /* Final Reduction Phase*/ 14. Aggregate Results and Output Associations 15. End ParAssociation redistribution cost can be amortized by the later iterations. Since each processor can proceed independently, there is no costly synchronization at the end of each iteration. Furthermore the new algorithms use the vertical database layout which clusters all relevant information in an itemset's tid-list. Each processor computes all the frequent itemsets from one cluster before proceeding to the next. The local database partition is scanned only once. In contrast Candidate Distribution must scan it once in each iteration. These algorithms don't pay the extra computation overhead of building or searching complex data structures, nor do they have to generate all the subsets of each transaction. As the intersection is performed an itemset can immediately be inserted in L k . Notice that the tid-lists also automatically prune irrelevant transactions. As the itemset size increases, the size of the tid-list decreases, resulting in very fast intersections. There are two distinct phases in the algorithms. The initialization phase, responsible for communicating the tid- lists among the processors, and the asynchronous phase, which generates frequent itemsets. The pseudo-code for the new algorithms is shown in table 2. 5.4.1. Initialization Phase The initialization step consists of three sub-steps. First, the support counts for 2-itemsets from the preprocessing step are read, and the frequent ones are inserted into L 2 . Second, applying one of the two clustering schemes to L 2 - the equivalence class or maximal hypergraph clique clustering - the set of potential maximal frequent itemsets is generated. These potential maximal clusters are then partitioned among all the processors so that a suitable level of load-balancing can be achieved. Third, the database is repartitioned so that each processor has on its local disk the tid-lists of all 1-itemsets in any cluster assigned to it. Scheduling We first partition the L 2 into equivalence classes using the common prefix as described above. If we are using equivalence class clustering then we already have the potential maximal itemsets. However, if we are using the clique clustering, we generate the maximal cliques within each class (see section 4). We next generate a schedule of the equivalence classes on the different processors in a manner minimizing the load imbalance and minimizing the inter-process commu- nication. Note that it may be necessary to sacrifice some amount of load balancing for a better communication efficiency. For this reason, whole equivalence classes, including all the maximal cliques within them, are assigned to the same processor. Load balancing is achieved by assigning a weighting factor to each equivalence class based on the number of elements in the class. Since we have to consider all pairs for the next iteration, we assign the weight to a class with s elements. Once the weights are assigned we generate a schedule using a greedy heuristic. We sort the classes on the weights, and assign each class in turn to the least loaded processor, i.e., one having the least total weight at that point. Ties are broken by selecting the processor with the smaller identifier. These steps are done concurrently on all the processors since all of them have access to the global L 2 . Figure 6 shows an example along with the equivalence classes, their weights, and the assignment of the classes on two processors. Notice how an entire sublattice induced by a given class is assigned to a single processor. This leads to better load balancing, even though the partitioning may introduce extra computation. For example, if 234 were not frequent, then 1234 cannot be frequent either. But since these belong to different equivalence classes assigned to different processors, this information is not used. Although the size of a class gives a good indication of the amount of work, better heuristics for generating the weights are possible. For example, if we could better estimate the number of frequent itemsets that could be derived from an equivalence class we could use this estimation as our weight. We believe that de-coupling processor performance right in the beginning holds promise, even though it may cause some load imbalance, since the repartitioning cost can be amortized over later iterations. Deriving better heuristics for scheduling the clusters, which minimize the load imbalance as well as communication, is part of ongoing research. Tid-list Communication Once the clusters have been partitioned among the processors each processor has to exchange information with every other processor to read the non-local tid-lists over the Memory Channel network. To minimize communication, and being aware of the fact that in our configuration there is only one local disk per host (recall that our cluster has 8 hosts, with 4 processors per host), only the hosts take part in the tid-list exchange. Additional processes on each of the 8 hosts are spawned only in the asynchronous phase. To accomplish the inter-process tid-list communication, each processor scans the item tid-lists in its local database partition and writes it to a transmit region which is mapped for receive on other processors. The other processors extract the tid-list from the PARALLEL ASSOCIATION RULES 21 receive region if it belongs to any cluster assigned to them. For example, figure 6 shows the initial local database on two hosts, and the final local database after the tid-list communication. Table 3. Pseudo-code for Bottom-up and Hybrid Traversal 1. Input: C equivalence 2. class or maximal clique 3. clustering of k-itemsets. 4. Output: Frequent itemsets 2 C k 5. 6. for all I i 2 C k do 7. C 8. for all I do 9. 10. if N.sup - minsup then 11. C 12. end; 13. if C k+1 14. Bottom-Up(C k+1 ); 15. end; 1. 2. /* Top-Down Phase */ 3. 4. for all I 5. 6. if N.sup - minsup then 7. 8. else break; 9. end; 11. /* Bottom-Up Phase */ 12. for all I i 2 S2 ,do 13. 15. if C3 16. end; 5.4.2. Asynchronous Phase At the end of the initialization step, the relevant tid-lists are available locally on each host, thus each processor can independently generate the frequent itemsets from its assigned maximal clusters, eliminating the need for synchronization with other processors. Each cluster is processed in its before moving on to the next cluster. This step involves scanning the local database partition only once. We thus benefit from huge I/O savings. Since each cluster induces a sublattice, depending on the algorithm, we either use a bottom-up traversal to generate all frequent itemsets, or we use the hybrid traversal to generate only the maximal frequent itemsets. The pseudo-code of the two lattice traversal schemes is shown in table 3. Note that initially we only have the tid-lists for 1-itemsets stored locally on disk. Using these, the tid-lists for the 2-itemset clusters are generated, and since these clusters are generally small the resulting tid-lists can be kept in memory. In the bottom-up approach, the tid-lists for 2-itemsets clusters are intersected to generate 3-itemsets. If the cardinality of the resulting tid-list exceeds the minimum support, the new itemset is inserted in L 3 . Then we split the resulting frequent 3-itemsets, equivalence classes based on common prefixes of length 2. All pairs of 3-itemsets within an equivalence class are intersected to determine L 4 , and so on 22 ZAKI, PARTHASARATHY, OGIHARA AND LI till all frequent itemsets are found. Once L k has been determined, we can delete We thus need main memory space only for the itemsets in L k\Gamma1 within one maximal cluster. For the top-down phase of the hybrid traversal only the maximal element seen so far needs to be memory-resident, along with the itemsets not yet seen. The new algorithms are therefore main memory space efficient. Experimental results on the memory usage of these algorithms are presented in the next section. Pruning Candidates Recall that both Count and Candidate Distribution use a pruning step to eliminate unnecessary candidates. This step is essential in those algorithms to reduce the size of the hash tree. Smaller trees lead to faster support counting, since each subset of a transaction is tested against the tree. However, with the vertical database layout we found the pruning step to be of little or no help. This can be attributed to several factors. First, there is additional space and computation overhead in constructing and searching hash tables. This is also likely to degrade locality. Second, there is extra overhead in generating all the subsets of a candidate. Third, there is extra communication overhead in communicating the frequent itemsets in each iteration, even though it may happen asynchronously. Fourth, because the average size of tid-lists decreases as the itemsets size increases, intersections can be performed very quickly with the short-circuit mechanism. At the end of the asynchronous phase we accumulate all the results from each processor and print them out. 5.5. Salient Features of the New Algorithms In this section we will recapitulate the salient features of our proposed algorithms, contrasting them against Count and Candidate Distribution. Our algorithms differ in the following respect: ffl Unlike Count Distribution, they utilize the aggregate memory of the parallel system by partitioning the candidate itemsets among the processors using the itemset clustering schemes. ffl They decouple the processors right in the beginning by repartitioning the database, so that each processor can compute the frequent itemsets independently. This eliminates the need for communicating the frequent itemsets at the end of each iteration. ffl They use the vertical database layout which clusters the transactions containing an itemset into tid-lists. Using this layout enables our algorithms to scan the local database partition only two times on each processor. The first scan for communicating the tid-lists, and the second for obtaining the frequent itemsets. In contrast, both Count and Candidate Distribution scan the database multiple times - once during each iteration. ffl To compute frequent itemsets, they performs simple intersections on two tid- lists. There is no extra overhead associated with building and searching complex hash tree data structures. Such complicated hash structures also suffer from poor cache locality (Parthasarathy, Zaki, & Li 1997). In our algorithms, all the PARALLEL ASSOCIATION RULES 23 available memory is utilized to keep tid-lists in memory which results in good locality. As larger itemsets are generated the size of tid-lists decreases, resulting in very fast intersections. Short-circuiting the join based on minimum support is also used to speed this step. ffl Our algorithms avoid the overhead of generating all the subsets of a transaction and checking them against the candidate hash tree during support counting. 6. Experimental Evaluation Table 4. Database properties Database T I D1 D1 Size D4 D4 Size All the experiments were performed on a 32-processor (8 hosts, 4 processors/host) Digital Alpha cluster inter-connected via the Memory Channel network (Gillett 1996). In our system unicast and multicast process-to-process writes have a latency of 5.2 -s, with per-link transfer bandwidths of 30MB/s. Each Alpha processor runs at 233MHz. There's a total of 256MB of main memory per host (shared among the 4 processors on that host). Each host also has a 2GB local disk attached to it, out of which less than 500MB was available to us. All the partitioned databases reside on the local disks of each processor. We used different synthetic databases, generated using the procedure described in (Agrawal & Srikant 1994). These have been used as benchmark databases for many association rules algorithms (Agrawal & Srikant 1994; Holsheimer et al. 1995; Park, Chen, & Yu 1995a; Savasere, Omiecinski, & Navathe 1995; Agrawal et al. 1996). Table 4 shows the databases used and their properties. The number of transactions is denoted as D r , where r is the replication factor. For all the databases are roughly 90MB in size. Except for the sizeup experiments, all results shown are on databases with a replication factor of We could not go beyond a replication factor of 6 (used in sizeup experiments) since the repartitioned database would become too large to fit on disk. The average transaction size is denoted as jT j, and the average maximal potentially frequent itemset size as jI j. The number of maximal potentially frequent itemsets 2000, and the number of items We refer the reader to (Agrawal Srikant 1994) for more detail on the database generation. All the experiments were performed with a minimum support value of 0.25%. For a fair comparison, all algorithms discover frequent k-itemsets for k - 3, using the supports for the 2-itemsets from the preprocessing step. Execution Time (sec) Count Distribution Par-Eclat Execution Time (sec) Par-Eclat Par-Clique Par-MaxEclat Par-MaxClique Figure 7. Parallel Performance on T10.I4.D2048K 6.1. Performance Comparison In this section we will compare the performance of our new algorithms with Count Distribution (henceforth referred to as CD), which was shown to be superior to both Data and Candidate Distribution (Agrawal & Shafer 1996). In all the figures the different parallel configurations are represented as Hx:Py:T z, where denotes the number of hosts, the number of processors per host, and z, the total number of processors used in the experiments. Figures 7, 8, and 9 show the total execution time for the different databases and on different parallel configurations. The configurations have been arranged in increasing order of T . Configurations with the same T are arranged in increasing order of H . The first column compares Par-Eclat with CD, and the second column compares the new algorithms, so that the differences among them are more apparent. It can be clearly seen that Par-Eclat out-performs CD for almost all configurations on all the databases, with improvements as high as a factor of 5. If we look at the best new algorithm from the second column, we see an improvement of about an order of magnitude. Even more dramatic improvements are possible for lower minimum support (Zaki, Parthasarathy, & Li 1997). An interesting trend in the figures is that the performance gap seems to decrease at larger configurations, with CD actually performing better at H8.P4.T32 for T10.I4.D2084K and T15.I4.D1471K. To see why consider figure 10 a, which shows the total number of frequent itemsets of different sizes for the different databases. Also from figure 11, which shows the initial database repartitioning and tid-list communication cost as a percentage of PARALLEL ASSOCIATION RULES 25 the total execution time of Par-Eclat, it becomes clear that there is not enough work for these two databases, to sufficiently offset the communication cost, consequently more than 70% of the time is spent in the initialization phase. For T20.I6.D1137K, which has more work, Par-Eclat is still about twice as fast as CD. The basic argument falls on the classic computation versus communication trade-off in parallel computing. Whenever this ratio is high we expect Par-Eclat to out-perform CD. We also expect the relative improvements of Par-Eclat over CD to be better for larger databases. Unfortunately due to disk space constraints we were not able to test the algorithms on larger databases. In all except the configurations, the local database partition is less than available memory. For CD, the entire database would be cached after the first scan. The performance of CD is thus a best case scenario for it since the results do not include the "real" hit CD would have taken from multiple disk scans. As mentioned in section 5.5, Par-Eclat was designed to scan the database only once during frequent itemset computation. Execution Time (sec) Count Distribution Par-Eclat Total Execution Time (sec) Par-Eclat Par-Clique Par-MaxEclat Par-MaxClique Figure 8. Parallel Performance on T15.I4.D1471K The second column in figures 7, 8, and 9 shows the differences among the new algorithms for different databases and parallel configurations. There are several parameters affecting their performance. It can be seen that in general Par-Clique and Par-MaxClique, perform better than Par-Eclat and Par-MaxEclat, respectively. This is because they use the maximal hypergraph clique approach, which generates more precise clusters. On the other axis, in general Par-MaxClique, and Par- MaxEclat, out-perform Par-Clique and Par-Eclat, respectively. This is because the hybrid lattice traversal scheme only generates maximal frequent itemsets, saving on the number of intersections. The results are also dependent on the number 26 ZAKI, PARTHASARATHY, OGIHARA AND LI Execution Time (sec) Count Distribution Par-Eclat Execution Time (sec) Par-Eclat Par-Clique Par-MaxEclat Par-MaxClique Figure 9. Parallel Performance on T20.I6.D1137K of frequent itemsets. The larger the number of frequent itemsets, the more the opportunity for the hybrid approach to save on the joins. For example, consider which shows the total number of tid-list intersections performed for the four algorithms on the three databases. For T20.I6.D1137K, which has the largest number of frequent itemsets (see figure 10 a), Par-MaxClique cuts down the number of intersections by more than 60% over Par-Eclat. The reduction was about 20% for Par-MaxEclat, and 35% for Par-Clique. These factors are responsible for the trends indicated above. The winner in terms of the total execution time is clearly Par-MaxClique, with improvements over Par-Eclat as high as 40%. 6.2. Memory Usage Figure 12 shows the total memory usage of the Par-Eclat algorithm as the computation of frequent itemsets progresses. The mean memory usage for the tid-lists is less than 0.7MB for all databases, even though the database itself is over 360MB. The figure only shows the cases where the memory usage was more than twice the mean. The peaks in the graph are usually due to the initial construction of all the 2-itemset tid-lists within each cluster. Since the equivalence class clusters can be large, we observe a maximum usage of 35MB for Par-Eclat, which is still less than 10% of the database. For the other algorithms, we expect these peaks to be lower, since the maximal clique clustering is more precise, resulting in smaller clusters, and the hybrid traversal doesn't need the entire cluster 2-itemsets initially. PARALLEL ASSOCIATION RULES 271000300050007000 Number of Frequent Itemsets Frequent Itemset Size (k) Frequent Itemsets at Number of Intersections Par-Eclat Par-MaxEclat Par-Clique Par-MaxClique Figure 10. a) Number of Frequent k-Itemsets; b) Number of Intersections Communication Communication Communication Figure 11. Communication Cost in Par-Eclat 6.3. Sensitivity Analysis Figures 13, 14, and 15 (first column) show the speedup on the different databases and parallel configurations. Due to disk constraints we used a replication 28 ZAKI, PARTHASARATHY, OGIHARA AND LI246810Memory Usage 2Mean Time T10.I4.D2084K: Memory Usage (Mean=0.55MB) Par-Eclat12345Memory Usage 2Mean Time T15.I4.D1471K: Memory Usage (Mean=0.43MB) Par-Eclat5152535Memory Usage 2Mean Time T20.I6.D1137K: Memory Usage (Mean=0.69MB) Par-Eclat Figure 12. Memory Usage in Par-Eclat (H1.P1.T1)1.52.53.5H1.P1.T1 H1.P2.T2 H2.P1.T2 H1.P4.T4 H2.P2.T4 H4.P1.T4 H2.P4.T8 H4.P2.T8 H8.P1.T8 H4.P4.T16 H8.P2.T16 H8.P4.T32 Relative T10.I4.D2084K: Speedup Par-Eclat Par-MaxEclat Par-Clique Total Execution Time (sec) Replication Factor T10.I4.D2084K: Sizeup Par-Eclat Par-MaxEclat Par-Clique Par-MaxClique Figure 13. T10.I4.D2048K: Speedup and Sizeup(H4.P1.T4) factor of 4, for database sizes of approximately 360MB. The speedup numbers are PARALLEL ASSOCIATION RULES 291.52.53.54.5 Relative T15.I4.D1471K: Speedup Par-Eclat Par-MaxEclat Par-Clique Par-MaxClique50150250350 Total Execution Time (sec) Replication Factor T15.I4.D1471K: Sizeup Par-Eclat Par-MaxEclat Par-Clique Par-MaxClique Figure 14. T15.I4.D1471K: Speedup and Sizeup(H4.P1.T4) not as impressive at first glance. However, this is not surprising. For example, on the largest configuration H8.P4.T32, there's only about 11MB of data per processor. Combined with the fact that the amount of computation is quite small (see figure a), and that about 50% to 70% of the time is spent in tid-list communication (see figure 11), we see a maximum speedup of about 5. Another reason is that the communication involves only the 8 hosts. Additional processes on a host are only spawned after the initialization phase, which thus represents a partially-parallel phase, limiting the speedups. If we take out the communication costs we see a maximum speedup of 12 to 16. An interesting trend is the step-effect seen in the speedup graphs. For the configurations which have the same number of total processors, the ones with more hosts perform better. Also, for configurations with more total processors, with 4, the configurations immediate preceding it, with only 1 processor per host, performs better. In both the cases, the reason is that increasing the number of processors on a given host, causes increased memory contention (bus traffic), and increased disk contention, as each processor tries to access the database from the local disk at the same time. Sizeup: For the sizeup experiments we fixed the parallel configuration to H4.P1.T4, and varied the database replication factor from 1 to 6, with the total database size ranging from about 90MB to 540MB. Figures 13, 14, and 15 (second column) show the sizeup for the four algorithms on the different databases. The figures indicate an almost linear sizeup. The slightly upward bend is due to the relative computation versus communication cost. The larger the database the more the time spent in communication, while the intersection cost doesn't increase at the Relative Par-Eclat Par-MaxEclat Par-Clique Total Execution Time (sec) Replication Factor Par-Eclat Par-MaxEclat Par-Clique Par-MaxClique Figure 15. T20.I6.D1137K: Speedup and Sizeup(H4.P1.T4) same pace. Moreover, the number of frequent itemsets remains constant (since we use percentages for minimum support, as opposed to absolute counts) for all replication factors. 7. Conclusions In this paper we proposed new parallel algorithms for the discovery of association rules. The algorithms use novel itemset clustering techniques to approximate the set of potentially maximal frequent itemsets. Once this set has been identified, the algorithms make use of efficient traversal techniques to generate the frequent itemsets contained in each cluster. We propose two clustering schemes based on equivalence classes and maximal hypergraph cliques, and study two lattice traversal techniques based on bottom-up and hybrid search. We also use the vertical database layout to cluster related transactions together. The database is also selectively replicated so that the portion of the database needed for the computation of associations is local to each processor. After the initial set-up phase, the algorithms do not need any further communication or synchronization. The algorithms minimize I/O overheads by scanning the local database portion only two times. Once in the set-up phase, and once when processing all the itemset clusters. The algorithms further use only simple intersection operations to compute frequent itemsets and don't have to maintain or search complex hash structures. An added benefit of using simple intersections is that the algorithms we propose can be im- PARALLEL ASSOCIATION RULES 31 plemented directly on general purpose database systems (Holsheimer et al. 1995; Using the above techniques we presented four new algorithms. The Par-Eclat (equivalence class, bottom-up search) and Par-Clique (maximal clique, bottom-up algorithms, discover all frequent itemsets, while the Par-MaxEclat (equiv- alence class, hybrid search) and Par-MaxClique (maximal clique, hybrid search) discover the maximal frequent itemsets. We implemented the algorithms on a processor DEC cluster interconnected with the DEC Memory Channel network, and compared it against a well known parallel algorithm Count Distribution (Agrawal 1996). Experimental results indicate that a substantial performance improvement is obtained using our techniques. Acknowledgments This work was supported in part by an NSF Research Initiation Award (CCR- 9409120) and ARPA contract F19628-94-C-0057. --R Parallel mining of association rules. Fast algorithms for mining association rules. Fast discovery of association rules. Mining association rules between sets of items in large databases. Hypergraphs: Combinatorics of Finite Sets. A fast distributed algorithm for mining association rules. Efficient mining of association rules in distributed databases. The KDD process for extracting useful knowledge from volumes of data. Computers and Intractability: A Guide to the Theory of NP-Completeness Memory channel: An optimized cluster interconnect. Scalable parallel data mining for association rules. In ACM SIGMOD Conf. A perspective on databases and data mining. In 11th Intl. Efficient algorithms for discovering association rules. An effective hash based algorithm for mining association rules. Efficient parallel data mining for association rules. In ACM Intl. An efficient algorithm for mining association rules in large databases. Sampling large databases for association rules. Parallel data mining for association rules on shared-memory multi-processors Evaluation of sampling for data mining of association rules. New algorithms for fast discovery of association rules. New algorithms for fast discovery of association rules. A localized algorithm for parallel association mining. Srinivasan Parthasarathy is currently a doctoral student at the University of Rochester. Wei Li received his Ph. --TR --CTR Jennifer Seitzer , James P. Buckley , Yi Pan, INDED: A Distributed Knowledge-Based Learning System, IEEE Intelligent Systems, v.15 n.5, p.38-46, September 2000 Sukomal Pal , Aditya Bagchi, Association against dissociation: some pragmatic considerations for frequent itemset generation under fixed and variable thresholds, ACM SIGKDD Explorations Newsletter, v.7 n.2, p.151-159, December 2005 A fast algorithm for mining sequential patterns from large databases, Journal of Computer Science and Technology, v.16 n.4, p.359-370, 7/1/2001 Eui-Hong (Sam) Han , George Karypis , Vipin Kumar, Scalable Parallel Data Mining for Association Rules, IEEE Transactions on Knowledge and Data Engineering, v.12 n.3, p.337-352, May 2000 Mohammed J. Zaki, Parallel and Distributed Association Mining: A Survey, IEEE Concurrency, v.7 n.4, p.14-25, October 1999 Congnan Luo , Anil L. Pereira , Soon M. Chung, Distributed Mining of Maximal Frequent Itemsets on a Data Grid System, The Journal of Supercomputing, v.37 n.1, p.71-90, July 2006 Jiawei Han , Laks V. S. Lakshmanan , Jian Pei, Scalable frequent-pattern mining methods: an overview, Tutorial notes of the seventh ACM SIGKDD international conference on Knowledge discovery and data mining, August 26-29, 2001, San Francisco, California Chenyong Hu , Benyu Zhang , Yongji Wang , Shuicheng Yan , Zheng Chen , Qing Wang , Qiang Yang, Learning quantifiable associations via principal sparse non-negative matrix factorization, Intelligent Data Analysis, v.9 n.6, p.603-620, November 2005 Yang , Srinivasan Parthasarathy , Sameep Mehta, A generalized framework for mining spatio-temporal patterns in scientific data, Proceeding of the eleventh ACM SIGKDD international conference on Knowledge discovery in data mining, August 21-24, 2005, Chicago, Illinois, USA Gregory Buehrer , Srinivasan Parthasarathy , Shirish Tatikonda , Tahsin Kurc , Joel Saltz, Toward terabyte pattern mining: an architecture-conscious solution, Proceedings of the 12th ACM SIGPLAN symposium on Principles and practice of parallel programming, March 14-17, 2007, San Jose, California, USA Assaf Schuster , Ran Wolff , Dan Trock, A high-performance distributed algorithm for mining association rules, Knowledge and Information Systems, v.7 n.4, p.458-475, May 2005 Shengnan Cong , Jiawei Han , Jay Hoeflinger , David Padua, A sampling-based framework for parallel data mining, Proceedings of the tenth ACM SIGPLAN symposium on Principles and practice of parallel programming, June 15-17, 2005, Chicago, IL, USA Mohammed J. Zaki, Scalable Algorithms for Association Mining, IEEE Transactions on Knowledge and Data Engineering, v.12 n.3, p.372-390, May 2000 Aleksandar Lazarevic , Zoran Obradovic, Boosting Algorithms for Parallel and Distributed Learning, Distributed and Parallel Databases, v.11 n.2, p.203-229, March 2002 Mohammed J. Zaki , Neal Lesh , Mitsunori Ogihara, PlanMine: Predicting Plan Failures Using Sequence Mining, Artificial Intelligence Review, v.14 n.6, p.421-446, December 1, 2000 John D. Holt , Soon M. Chung, Parallel mining of association rules from text databases, The Journal of Supercomputing, v.39 n.3, p.273-299, March 2007 Vipin Kumar , Mohammed Zaki, High performance data mining (tutorial PM-3), Tutorial notes of the sixth ACM SIGKDD international conference on Knowledge discovery and data mining, p.309-425, August 20-23, 2000, Boston, Massachusetts, United States Petr Hjek , Martin Holena, Formal logics of discovery and hypothesis formation by machine, Theoretical Computer Science, v.292 n.2, p.345-357, 27 January Andrzej Skowron , Jan Komorowski , Zdzislaw Pawlak , Lech Polkowski, Rough sets perspective on data and knowledge, Handbook of data mining and knowledge discovery, Oxford University Press, Inc., New York, NY, 2002 Andrzej Skowron, Rough sets and boolean reasoning, Granular computing: an emerging paradigm, Physica-Verlag GmbH, Heidelberg, Germany, 2001 A. Maniatty , Mohammed J. Zaki, Systems support for scalable data mining, ACM SIGKDD Explorations Newsletter, v.2 n.2, p.56-65, Dec. 2000
lattice traversal;parallel data mining;maximal hypergraph cliques;association rules
593455
High Performance OLAP and Data Mining on Parallel Computers.
On-Line Analytical Processing (OLAP) techniques are increasingly being used in decision support systems to provide analysis of data. Queries posed on such systems are quite complex and require different views of data. Analytical models need to capture the multidimensionality of the underlying data, a task for which multidimensional databases are well suited. Multidimensional OLAP systems store data in multidimensional arrays on which analytical operations are performed. Knowledge discovery and data mining requires complex operations on the underlying data which can be very expensive in terms of computation time. High performance parallel systems can reduce this analysis time.Precomputed aggregate calculations in a Data Cube can provide efficient query processing for OLAP applications. In this article, we present algorithms for construction of data cubes on distributed-memory parallel computers. Data is loaded from a relational database into a multidimensional array. We present two methods, sort-based and hash-based for loading the base cube and compare their performances. Data cubes are used to perform consolidation queries used in roll-up operations using dimension hierarchies. Finally, we show how data cubes are used for data mining using Attribute Focusing techniques. We present results for these on the IBM-SP2 parallel machine. Results show that our algorithms and techniques for OLAP and data mining on parallel systems are scalable to a large number of processors, providing a high performance platform for such applications.
Introduction On-line Analytical Processing (OLAP) [1] systems enable analysts and managers to gain insight into the performance of an enterprise through a wide variety of views of data organized to reflect the multidimensional nature of the enterprise data. OLAP gives insight into data through fast, consistent, interactive access to a wide variety of possible views of information. In contrast to traditional databases, it answers questions like "what if ?" and "why ?" in addition to "who ?" and "what ?". OLAP is used to build decision support systems which help in extracting knowledge from data. OLAP is used to summarize, consolidate, view, apply formulae to, and synthesize data according to multiple dimensions. Queries posed on such systems are quite complex and require different views of data. Traditionally, a relational approach (relational OLAP) has been taken to build such systems. Relational databases are used to build and query these systems. A complex analytical query is cumbersome to express in SQL and it might not be efficient to execute. More recently, multidimensional database techniques (multidimensional OLAP) have been applied to decision-support applications. Data is stored in multidimensional arrays which is a more natural way to express the multi-dimensionality of the enterprise data and is more suited for analysis. A "cell" in multidimensional space represents a tuple, with the attributes of the tuple identifying the location of the tuple in the multidimensional space and the measure values represent the content of the cell. Data mining can be viewed as an automated application of algorithms to detect patterns and extract knowledge from data [2]. An algorithm that enumerates patterns from, or fits models to, data is a data mining algorithm. Data mining is a step in the overall concept of knowledge discovery in databases (KDD). Large data sets are analyzed for searching patterns and discovering rules. Automated techniques of data mining can make OLAP more useful and easier to apply in the overall scheme of decision support systems. Data mining techniques like Associations, Classification, Clustering and Trend analysis [2] can be used together with OLAP to discover knowledge from data. An approach to data mining called Attribute Focusing targets the consumer by using algorithms that lead the user through the analysis of data. Attribute Focusing has been successfully applied in discovering interesting patterns in the NBA [5] and other applications. Earlier applications of this technique were to software process engineering [4]. Since data cubes have aggregation values on combinations of attributes already calculated, the computations of attribute focusing are greatly facilitated by data cubes. We present a parallel algorithm to calculate the "interestingness" function used in attribute focusing on the data cube. Typically, large amounts of data are analyzed for OLAP and data mining applications. Ad-hoc analytical queries are posed by analysts who expect the system to provide real-time performance. Parallel computers can be used in such a scenario for a variety of reasons. Scalable solutions can build a highly accurate data mining model quicker. Mining large databases and constructing complex models take a large amount of computing power which can take hours of valuable time. Scalable parallel systems can reduce this wait time to minutes or even seconds, thus increasing productivity and better understanding of the knowledge discovery process. The use of many processors enables the use of more memory and a larger database can be handled in the main memory attached to the processors. We have currently considered main-memory databases. Extensions to disks using parallel I/O will be future work. In this article we present scalable parallel algorithm and techniques for OLAP in multidimensional databases. Parallel construction and maintenance of data cubes and their use for OLAP queries is shown. Results show that our implementations are scalable to a large number of processors. Consolidation queries make use of hi- Category Examples Distributive Sum(), Count(), Minimum(), Maximum() Algebraic Average(), Standard Deviation(), MaxN()(N largest values), smallest values), Center of Mass() Holistic Median(), MostFrequent()(i.e. Mode()), Rank() Table 1: Categories of aggregate functions erarchies on dimensions to make OLAP queries possible at different levels of detail. We show the performance of these algorithms on the OLAP council benchmark [15] which models a real OLAP environment on a IBM parallel machine. The IBM SP-2 is a network of RS/6000 workstations connected together on a high speed communication switch, which is fast getting popular as a parallel computing platform. We show the use of data cubes to perform data mining by using attribute focusing to find two-way associations between attributes. These can easily be extended to n-way associations. To the best of our knowledge, this is the first effort on high performance parallel computation of data cubes for MOLAP and data mining using them. The rest of the paper is organized as follows. Section 2 gives an overview of the data cube and its operators. The model of parallel computation we have used is given in Section 3. Section 4 presents the issues in data cube construction. Section 5 presents parallel data cube construction for MOLAP. Section 6 gives results for a OLAP council benchmark suite on the IBM-SP2 parallel machine. Section 7 presents consolidation queries that use hierarchies defined on the various dimensions. Section 8 describes data mining on data cubes with results on the IBM-SP2. Section 9 concludes the article. 2 Data Cube The Data cube operator was introduced recently by Gray et. al [7] to support multiple aggregates. Data cube computes aggregates along all possible combinations of dimensions. This operation is useful for answering OLAP queries which use aggregation on different combinations of attributes. Data can be organized into a data cube by calculating all possible combinations of GROUP-BYs. For a data set with k attributes this would lead to 2 k GROUP-BY calculations. In this article we present algorithms for calculating the data cube using multidimensional arrays on a distributed memory parallel computer. The cube treats each of the k aggregation attributes as a dimension in k-space. An aggregate of a particular set of attribute values is a point in this space. The set of points form a k-dimensional cube. Aggregate functions are classified into three categories as shown in Table 1. Distributive functions allow partitions of the input set to be aggregated separately and later combined. Algebraic functions can be expressed in terms of other distributive functions, e.g. average() can be expressed as the ratio of sum() and count(). Holistic functions, such as median() cannot be computed in parts and combined. 2.1 Operations on the Data Cube Data Cube operators generalize the histogram, cross-tabulation, roll-up, drill-down and sub-total constructs required by financial databases. The following operations can be defined on the data cube. ffl Pivoting: This is also called rotating and involves rotating the cube to change the dimensional orientation of a report or page on display. It may consist of swapping the two dimensions (row and column in a 2D-cube) or introducing another dimension instead of some dimension already present in the cube. Slicing-dicing : This operation involves selecting some subset of the cube. For a fixed attribute value in a given dimension, it reports all the values for all the other dimensions. It can be visualized as slice of the data in a 3D-cube. dimensions have a hierarchy defined on them. Aggregations can be done at different levels of hierarchy. Going up the hierarchy to higher levels of generalization is known as roll-up. For example, aggregating the dimension up the hierarchy (day ! month ! quarter::) is a roll-up operation. ffl Drill-down: This operation traverses the hierarchy from lower to higher levels of detail. Drill-down displays detail information for each aggregated point. ffl Trend analysis over sequential time periods Product Date Supplier d2 p3 d5 d4 d3 Figure 1: Example database of sales data Figure 1 shows a multidimensional database with product, date, supplier as dimensions and sales as a measure. Partitioning a data set into dimensions and measures is a design choice. Dimensions usually have a hierarchy associated with them, which specify aggregation levels and the granularity of viewing data. For example day ! month ! quarter ! year is a hierarchy on date. An aggregate of an attribute is represented by introducing a new value ALL. The right hand side table in Figure 2 illustrates a data cube for the set of tuples given in the left hand side [7]. 2.2 OLAP Alternatives Traditionally, OLAP systems have been build on top of a relational database system. These are referred to as relational OLAP (ROLAP). An OLAP engine is built on top of a relational database. This generates analytical queries in SQL which may become cumbersome for complex queries and affect performance. Relational systems have to embed a multidimensional view on the underlying data. Alternatively, multidimensional OLAP (MOLAP) systems have appeared only recently. These use multidimensional databases modeled as Model Year Color Sales Chevy 1990 Red 5 Chevy 1990 Blue 87 Ford 1990 Green 64 Ford 1990 Blue 99 Ford 1991 Red 8 Ford 1991 Blue 7 Model Year Color Sales Chevy 1990 Blue 87 Chevy 1990 Red 5 Chevy 1990 ALL 92 Chevy ALL Blue 87 Chevy ALL Red 5 Chevy ALL ALL 92 Ford 1990 Blue 99 Ford 1990 Green 64 Ford 1990 ALL 163 Ford 1991 Blue 7 Ford 1991 Red 8 Ford 1991 ALL 15 Ford ALL Blue 106 Ford ALL Green 64 Ford ALL Red 8 ALL 1990 Blue 186 ALL 1990 Green 64 ALL 1991 Blue 7 ALL 1991 Red 8 Ford ALL ALL 178 ALL 1990 ALL 255 ALL 1991 ALL 15 ALL ALL Blue 193 ALL ALL Green 64 ALL ALL Red 13 ALL ALL ALL 270 Figure 2: Data cube Illustration using a relation multidimensional arrays for OLAP operations. An intuitive view of data provides a sophisticated analytical functionality and support for complex queries. Data relationships are modeled more naturally and intuitively. Spatial databases [8] represent and model geometric objects (points, lines, polygons etc.) in multidimensional space. MOLAP data can be considered as points in the multidimensional space of attributes and benefit from some spatial database techniques. The operations in OLAP although different from the operations, such as overlap, containment etc., used in spatial databases can benefit from the spatial indexing structures developed for spatial databases. 3 Communication Costs Distributed Memory Parallel Computers (shared-nothing machines) consist of a set of processors (tens to a few hundred) connected through an interconnection network. The memory is physically distributed across the processors. Interaction between processors is through message passing. Popular interconnection topologies are 2D meshes (Paragon, Delta), 3D meshes (Cray T3D), hypercubes (nCUBE), fat tree (CM5) and multistage networks (IBM-SP2). Most of these machines have cut-through routed networks which will be used for modeling the communication cost of our algorithms. For a lightly loaded network, a message of size m traversing d hops of a cut-through (CT) routed network incurs a communication delay given by represents the handshaking costs, t h represents the signal propagation and switching delays and t w represents the inverse bandwidth of the communication network. The startup time t s is often large, and can be several Primitive Running time on p processors Broadcast O((ts Reduce O((ts Parallel Prefix O((ts Gather O(ts log All-to-All Communication O((ts log p) Table 2: Running times of various parallel primitives on a multi-staged network hundred machine cycles or more. The per-word transfer time t w is determined by the link bandwidth. t w is often higher (an order to two orders of magnitude is typical) than t c , the time to do a unit computation on data available in the cache. Parallelization of applications requires distributing some or all of the data structures among the processors. Each processor needs to access all the non-local data required for its local computation. This generates aggregate or collective communication structures. Several algorithms have been described in the literature for these primitives and are part of standard textbooks [12]. The use of collective communication provides a level of architecture independence in the algorithm design. It also allows for precise analysis of an algorithm by replacing the cost of the primitive for the targeted architecture. Table 2 provides a complexity analysis of these operations on a multi-staged network. These costs are used in the analysis of the algorithms presented in the next section. 1. Broadcast. In a Broadcast operation, one processor has a message of size m to be broadcast to all other processors. 2. Reduce. Given a vector of size m on each processor and a binary associative operation, the Reduce operation computes a resultant vector of size m and stores it on every processor. The i th element of the resultant vector is the result of combining the i th element of the vectors stored on all the processors using the binary associative operation. 3. Parallel Prefix. Suppose that x are p data elements with processor P i containing x i . Let \Omega be a binary associative operation. The Parallel Prefix operation stores the value of x on processor P i . 4. Gather. Given a vector of size m on each processor, the Gather operation collects all the data and stores the resulting vector of size mp in one of the processors. 5. All-to-All Communication. In this operation each processor sends a distinct message of size m to every processor. 4 Data Cube Construction The data cube operator is the n-dimensional generalization of the group-by operator. Consider the following query which uses the cube operator. SELECT Model, Year, Color, SUM(sales) AS Sales FROM Sales WHERE Model in 'Ford', 'Chevy' AND Year BETWEEN 1990 AND 1992 GROUP BY CUBE(Model, Year, Color); aggregate calculations are needed for a N-dimensional data cube. For example, are calculated for the above query: fModel, Year, Colorg, fModel, Yearg, fModel, Colorg, fYear, Colorg, fModelg, fYearg, fColorg and ALL, as shown in Figure 2. ABCD ALL # of GROUP-BYs44431 Figure 3: Lattice for cube operator Several optimizations can be done over the naive method of calculating each aggregate separately from the initial data [7]. Optimizations 1 and 2 are normally considered for a uniprocessor model. Optimization 3 is an added and important consideration for a parallel implementation to reduce the overheads from communication costs. 1. Smallest Parent: For computing a group-by this selects the smallest of the previously computed group-bys from which it is possible to compute the group-by. Consider a four attribute cube (ABCD). Group-by AB can be calculated from ABCD, ABD and ABC. Clearly sizes of ABC and ABD are smaller than that of ABCD and are better candidates. The actual choice will be made by picking the smaller of ABC and ABD. 2. Effective use of Cache: This aims at using the cache effectively by computing the group-bys in such an order that the next group-by calculation can benefit from the cached results of the previous calculation. This can be extended to disk based data cubes by reducing disk I/O and caching in main memory. For example, after computing ABC from ABCD we compute AB followed by A. In MOLAP systems the sizes of the intermediate levels are fixed and the order can be determined easily. 3. Minimize inter-processor Communication: Communication is involved among the processors to calculate the aggregates. The order of computation should minimize the communication among the processors because inter-processor communication costs are typically higher than computation costs. For example, have a higher communication cost to first aggregate along B and then divide C among the processors in comparison to CD ! C where a local aggregation on each processor along D will be sufficient. A lattice framework to represent the hierarchy of the group-bys was introduced in [11]. This is an elegant model for representing the dependencies in the calculations and also to model costs of the aggregate calculations. A scheduling algorithm can be applied to this framework substituting the appropriate costs of computation and communication. A lattice for the group-by calculations for a four-dimensional cube (ABCD) is shown in Figure 3. Each node represents an aggregate and an arrow represents a possible aggregate calculation which is also used to represent the cost of the calculation. Calculation of the order in which the GROUP-BYs are created depends on the cost of deriving a lower order (one with a lower number of attributes) group-by from a higher order (also called the parent) group-by. For example, between ABD ! BD and BCD ! BD one needs to select the one with the lower cost. Cost estimation of the aggregation operations can be done by establishing a cost model. This is described later in the section on aggregation. We assume that the total available memory on the processors is large enough to hold the data sets in memory. This is a reasonable assumption since most parallel machines these days have 128-256 MB main memory per node. With 16 nodes we can handle databases of size 2GB and larger data sets can be handled by increasing the number of processors. Hence, it is important to develop scalable algorithms to handle larger databases. In this article we develop in-memory algorithms to calculate the data cube. External algorithms are also being explored as part of this research. 5 Parallel Data Cube Construction for MOLAP We assume that data is provided as a set of tuples stored in a file and the number of distinct elements are given for each attribute. For illustration purposes, let A; B; C and D be the attributes in a data set with D a ; D b ; D c and D d as the number of their distinct values, respectively. We assume that the number of distinct values in each dimension is known. However, the values are determined from the database by the algorithm. Without loss of generality, let D a D b D c D d . If this is not the case, it can be made true by a simple renaming of the attributes. Let p be the number of processors, numbered N be the number of tuples. Figure 4 shows the various steps in the data cube construction algorithm. Each step is explained in the next few subsections. Algorithm 1 Parallel data cube construction and operations Step 1. Partition tuples between processors (Partitioning) Step 2. Load tuples into multidimensional array (Loading) Step 3. Generate schedule for order of group-by calculations Step 4. Perform aggregation calculations (Aggregation) Step 5. Redistribute/Assign sub-cubes to processors for query processing Step 6. Define local and distributed hierarchies on all dimensions Figure 4: Steps in parallel data cube construction First, the tuples are partitioned on p processors in a partitioning step. The partitioning phase is followed by a loading phase in which a multidimensional array is loaded on each processor from the tuples acquired after the partitioning phase. This creates the base cube. Loading can either be done by a hash-based method or a sort-based method. We have implemented both and have compared their scalability properties. This is followed by a aggregation phase which calculates the various aggregate sub-cubes. We describe each phase in the next few subsections. 5.1 Partitioning A sample based partitioning algorithm is used to partition the tuples among the processors. Attribute A is used in this partitioning. This is done to ensure the partitioning of data at the coarsest grain possible. This divides A nearly equally onto the processors and also establishes an order on A. If A x 2 P i and A y 2 P j then A x A y for In fact, in the partitioning scheme used here, tuples end up being sorted locally on each processor. The algorithm is illustrated in Figure 5 with communication steps shown in bold. Algorithm 2 Sample based partitioning - Number of processors. Step 1. Sort tuples on attribute A locally on each processor Step 2. Pick up a sample of size p of attribute A from each processor. Gather the sample on processor 0 and sort it locally. Step 3. Pick up splitters from this list on processor 0 and Broadcast them. Step 4. Using the splitters each processor divides it's tuples into p sorted partitions with all values of attribute A in a lower partition being lower than a higher numbered partition. Step 5. Using All-to-All communication each partition is sent to its destination processor. Step 6. Each processor merges the sorted partitions obtained from other processors resulting in a set of tuples sorted globally on attribute A. Figure 5: Sample based partitioning 5.2 Loading Once tuples are partitioned on processors, they are loaded into a multidimensional array (md-array). The size of the md-array in each dimension is the same as the number of unique values for the attribute represented in that dimension. A tuple is represented as a cell in the md-array indexed by the values of each of the attributes. Hence, each measure needs to be loaded in the md-array from the tuples. We describe two methods to perform this task, a hash based method and a sort-based method. 5.2.1 Hash-based method Figure 6 describes the hash-based cube loading algorithm. Each attribute is hashed to get a hash table of unique values for it. A sort on the attribute's hash table will index the dimension of the base cube corresponding to that attribute. These hash tables are then probed to fill in the measure values in the base cube. Hashing techniques are known to provide good performance on the average, though it heavily depends on the choice of a good hash function. 5.2.2 Sort-based method Figure 7 describes the sort-based cube loading algorithm. This method provides regularity of access over the sorted hash tables since the attributes probing them are sorted, unlike the hash-based method, where Algorithm 3 Hash-based data cube loading Algorithm number of attributes(dimensions) Step 1. For each tuple, hash each of the attributes into a separate hash table, one for each attribute. k hash tables are created as a result of this. (hash phase) Step 2. Compress and sort each hash table. Step 3. Update the index of the hash table with its order in the corresponding sorted list. Step 4a. Pick up each tuple and probe the hash tables for each of its attributes to obtain the indices in each dimension. (probe phase) Step 4b. Update the cell at the index in the md-array with the measure values of the tuple. Figure Hash-based Algorithm for multidimensional data cube Loading accesses in the hash table have no order. The sorted hash-tables are scanned only in one direction for all the tuples which have the same value for all the attributes in dimensions which come earlier, i.e have more unique values, since the order in which the attributes are sorted is with respect to their number of unique values. For example, for two consecutive records (a table for D is scanned from the current position to get the index. However, for (a tables for both C and D need to be scanned from the beginning. Algorithm 4 Sort-based data cube loading Algorithm - Number of attributes(dimensions) Step 1. Sort tuples locally on each processor in such a way that B is sorted for each unique value of A, C is sorted for each unique value of B and so on. Note that A is already sorted on each processor. Step 2. For each tuple hash each of the attributes into a separate hash table, one for each attribute. k hash tables are created as a result of this. (hash phase) Step 3. Compress and sort each hash table. Step 4. Pick up each tuple and calculate the indices of each attribute in the md-array in the following way. (Assume that the sorting order is A, B and C) Step 4a. Pick up the attribute at the innermost level of the sort and linearly scan the sorted hashed list for it to pick up the index. Step 4b. As long as the values at higher levels are the same, scan past the point where the scan stopped last, to find the index. If a attribute value changes, scan the sorted hash lists from the start for all inner attributes. Step 4c. Update the cell at the index in the md-array with the measure values of the tuple. Figure 7: Sort-based Algorithm for multidimensional data cube Loading 5.3 Aggregation An example of a base cube is shown in Figure 8 with 2. Hence each processor has Da portion of A. We illustrate the costs of calculating the various GROUP-BYs from the base cube for a three attribute (or 3D) cube in Table 3. Let t op be the cost of an addition, t copy the cost of a copying a byte. Communication is modeled by collective communication operations as described in the previous section. These costs are then used by a scheduling algorithm, which generates a schedule for calculating the various group-bys. Some calculations are local and some are non-local and need multiple processors to exchange data leading to communication among processors. An example of a local aggregate calculation, ABCD ! ABD, is shown in Figure 9. Even with local calculations the costs can differ depending on how the calculation is made. Calculating AC from ABC requires summing on the B dimension and calculating AC from ACD requires aggregation on D. Depending on how the multidimensional array is stored these costs could be different since the stride of the access can affect the cache performance. From the cost calculations shown in Table 3, we see that the cost of calculating aggregates from a parent are lower if the order of the attributes in the aggregate is a prefix of the parent. Calculating ABC ! AB ! A is a local calculation on each node. Each cube is distributed across the processors since dimension A is distributed. The intermediate cubes, resulting from aggregation of parent cubes are also distributed among the processors. This results in good load balancing among the processors. Calculations involving multiple cubes can also be distributed across processors as a result of this. The first attribute in the aggregate cube is always distributed among processors. As a result, A is distributed in ABCD, ABC, AB and A, B is distributed in BCD, BC and B, and C is distributed in CD and C and D is distributed. Figure 8 shows A as distributed in ABCD. Source Target Cost Table 3: Calculation of GROUP-BYs for a three attribute data cube (D a \Theta D b \Theta D c ), on p processors Figures and 11 illustrate a global aggregate calculation of calculating ABCD ! BCD. First a local aggregation is done along dimension A (Figure 10). This is followed by a Reduce operation on BCD (Figure 11). Each processor then keeps the corresponding portion of BCD, distributing B evenly among processors. Clearly, as can be observed from the aggregation calculations shown in the table above, there are multiple ways of calculating a GROUP BY. For calculating the aggregate at a level where aggregates can be calculated from multiple parents, we need to pick up an assignment such that the cost is minimized. In [13], this is solved by posing it as a graph problem and using minimum cost matching in a bipartite graph. We have augmented it by our cost model and the optimizations needed. Again, refer to Figure 3. Suppose we want to calculate aggregates at a level k from the parent at level k + 1. A bipartite graph E) is defined as follows. A group of nodes X is the nodes at level k. Clearly jX Another group of nodes jY j is the nodes at level k . The edges connecting the nodes at level k and k belong to E. The edge weights are the costs of calculating the particular aggregate at level k from the parent at level k + 1. Costs described in Table 3 are used here. A node at level k possibly calculate aggregates at level k. Hence the nodes are replicated these many times at level k these are added to Y . We seek a match d2 A: 1,2,3,4,5,6,7,8 B: b1,b2,b3,b4 C: c1,c2 Figure 8: Basecube for 4 attributes on 4 processors A: 1,2 B: b1,b2,b3,b4 C: c1,c2 d2 d2 ABCD ABD op op Figure 9: Local aggregation calculation, ABCD ! ABD on processor P0 ABCD BCD A: 1,2,3,4,5,6,7,8 B: b1,b2,b3,b4 C: c1,c2 d2 Figure 10: Global aggregation calculation, local phase, ABCD ! BCD A: 1,2,3,4,5,6,7,8 B: b1,b2,b3,b4 C: c1,c2 ABCD BCD Reduce Operation d2 d2 d2 d2 d2 Figure 11: Global aggregation calculation, global phase, ABCD ! BCD between nodes from level k level k which minimizes the total edge costs. Figure 12 shows an example of calculating level 2 from level 3 of the lattice of Figure 3. This is done for each level and the resulting order of calculations at each level is picked up. This creates a directed acyclic graph which is then traversed, from the base cube as the root, to each child in a depth first search manner, calculating each aggregation. This results in preserving all of the three optimizations for multiple group-bys described in an earlier section. AD AC BD CD ABD -> AB 15 BD AD AD AC BD CD BD Figure 12: A minimum cost matching for calculation of Level 2 aggregates from Level 3 6 Results We have implemented the algorithms presented in the previous section using 'C' and message passing using the Message Passing Interface (MPI) on the IBM SP-2 parallel machine. Each node of the SP-2 is a RS/6000 processor, with 128MB memory. The interconnection network is a multistage network connected through a high speed switch. The use of C and MPI makes the programs portable across a wide variety of parallel platforms with little effort. 6.1 Data sets We have used the OLAP council benchmark [15] which simulates a realistic On-Line Analytical Processing business situation. The data sets have attributes taken from the following set of attributes: Product (9000), Customer (900), Time (24), Channel (9) and Scenario (2). The values in brackets following each attribute are the number of unique values for that attribute. These are taken to be the size of the dimensions in the resulting multi-dimensional arrays in the sub-cubes of the data cube. For the History Sales data, we have picked out 1000, 101, 100 and 10 distinct values for its 4 attributes from the benchmark data and generated the records by picking out each attribute randomly from the set of distinct values for that attribute. Product, Customer and Channel are character strings with 12 characters each. Time is an integer depicting year and month (e.g. 199704 is April 1997) and Scenario is a 6 character string showing if it is an "Actual" or a "Budget" value. The measure stored in each cell of the array is a float value depicting the sales figure for the particular combination of the attributes. This can potentially contain many more values. We have used the Sum() function to compute the aggregates in the results presented here. The characteristics of the data sets are summarized in Table 4. Data Set Dimensions Number of records Base cube dimensions Data size Shipping Cost Customer, Time, Scenario 27000 900 \Theta 24 \Theta 2 1MB Production Cost Product, Time, Scenario 270,000 9000 \Theta 24 \Theta 2 10MB Current Sales Product, Customer, Channel 81000 and 608866 9000 \Theta 900 \Theta 9 320MB Budget Product, Customer, Time 951,720 9000 \Theta 900 \Theta 12 420MB History Sales Product, Customer, Time, Channel 1,010,000 1000 \Theta 101 \Theta 100 \Theta 10 500 MB Table 4: OLAP benchmark data sets used for data cube construction 6.2 Data cube Construction We present results for the different phases of data cube construction for all the data sets described above. Figure 13 shows the performance of hash-based and sort-based methods for the Shipping Cost data with 27000 records. We observe that the algorithm scales well for this small data size. Each individual component scales well also. The two methods take almost the same time for the data cube construction. Processors0.030.080.120.18Time (sec) Shipping Cost (Customer, Scenario, Time) Partition Hash Load Aggregate Total Processors0.030.080.120.18Time (sec) Shipping Cost (Customer, Scenario, Time) Partition Hash Load Aggregate Total Figure 13: Various phases of cube construction using (a) hash-based (b) sort-based method for Shipping Cost data Production Cost data contains 270,000 records and the performance of the two methods on this data set is shown in Figure 14. Again, as we increase the number of processor, each of the component in the construction algorithm scales well resulting in good overall scalability for both methods. Figure 15 shows the performance of hash-based and sort-based methods for the Current Sales data with 81000 records. This data has a density of 0.11%. The aggregation time is the main component in the total time because of the large cube size for this data set. Figure shows the performance of hash-based and sort-based methods for the Current Sales data with records. This data has a density of 0.84%. All the individual components have increased except the aggregation component. The number of tuples has increased in this data set which affects the components till the loading phase. Aggregation times are dependent on the cube size, which has not changed. Figure 17 gives the performance on Budget data with nearly a million records for 8, 16, 32, 64 and 128 processors. It can be observed that for large data sets the methods perform well as we increase the number of processors. Processors0.20.81.21.82.2 Time (sec) Production Cost (Product, Scenario, Time) Partition Hash Load Aggregate Total Processors0.20.81.21.82.2 Time (sec) Production Cost (Product, Scenario, Time) Partition Hash Load Aggregate Total Figure 14: Various phases of cube construction using (a) hash-based (b) sort-based method for Production Cost data (N= 270,000 records, Processors0.51.52.53.54.55.56.57.5Time (sec) hash-based Current Sales (Customer, Product, Channel) Partition Hash Load Aggregate Total Processors1.03.05.07.09.0 Time (sec) sort-based Current Sales (Customer, Product, Channel) Partition Hash Load Aggregate Total Figure 15: Various phases of cube construction using (a) hash-based (b) sort-based method for Current Sale data Processors1.03.05.07.09.011.013.0 Time (sec) hash-based (Customer, Product, Channel) Partition Hash Load Aggregate Total Processors1.03.05.07.09.011.013.0Time (sec) sort-based (Customer, Product, Channel) Partition Hash Load Aggregate Total Figure Various phases of cube construction using (a) hash-based (b) sort-based method for Current Sale data Processors1.03.05.07.09.0 Time (sec) Budget (Product, Customer, Time) Partition Hash Load Aggregate Total Processors1.03.05.07.09.0 Time (sec) Budget (Product, Customer, Time) Partition Hash Load Aggregate Total Figure 17: Various phases of cube construction using (a) hash-based (b) sort-based method for Budget data Figure gives the performance on History Sales data containing 4 attributes (4 dimensional base cube) with more than a million records for 8, 16, 32 and 64 processors. The density of the cube is 1%. The hash-based method performs better than the sort-based method for this case. The addition of a dimension increases the stride factor for accesses for the sort-based loading algorithm, deteriorating the cache performance. The hash-based method has no particular order of reference in the multidimensional array and benefits from a better cache performance. It can be observed that for large data each component in cube construction scales well as we increase the number of processors. Processors1.03.05.07.09.011.0 Time (sec) HistSale(Product, Customer, Time, Channel) Partition Hash Load Aggregate Total Processors1.03.05.07.09.011.0 Time (sec) HistSale(Product, Customer, Time, Channel) Partition Hash Load Aggregate Total Figure Various phases of cube construction using (a) hash-based (b) sort-based method for History Sales data Next, we compare the effect of density of the data sets on the data cube construction algorithm. The size of the multi-dimensional array is determined by the number of unique values in each of the dimensions. Hence the aggregation costs, which use the multi-dimensional array are not affected by the change in density as can be seen from Figure 19 and Figure 20. The other phases of the data cube construction deal with the tuples and the number of tuples increase as the density of the data set increases. The costs for partition, sort, hash and the load phases increase because the number of tuples on each processor increases. The scalability for both densities, and both hash-based and sort-based methods is good as observed from the figures. Comparing the sort-based methods and the hash-based methods, we observe that the hash-based method performs slightly better for 3D cubes but a lot better for the 4D case. We have used a multi-dimensional Partition Hash Load Aggregate Total1.03.05.07.09.011.013.0 Time (sec) Current Sales, (Customer, Product, Channel) Partition Hash Load Aggregate Total1.03.05.07.0 Time (sec) Current Sales, (Customer, Product, Channel) Partition Hash Load Aggregate Total0.51.52.53.5Time (sec) Current Sales, (Customer, Product, Channel) Figure 19: Comparison of cube construction components for two different densities for the hash based method Partition Sort Hash Load Aggregate Total1.03.05.07.09.011.013.015.0 Time (sec) Current Sales, (Customer, Product, Channel) Partition Sort Hash Load Aggregate Total0.51.52.53.54.55.56.57.5Time (sec) Current Sales, (Customer, Product, Channel) Partition Sort Hash Load Aggregate Total0.51.52.53.54.5 Time (sec) Current Sales, (Customer, Product, Channel) Figure 20: Comparison of cube construction components for two different densities for the sort based method array and the outermost dimension of the array (the one changing the slowest) does not match the inner-most sorted-dimension. The inner-most dimension is used for the attribute with the largest number of unique values to make the aggregate calculations efficient. The sort-based loading algorithm starts with the inner-most attribute in the sorted order which is also the smallest dimension. This also happens to be the outermost dimension of the array having the maximum stride. This is done to make the accesses during the aggregate calculation regular in the largest dimension, which is the innermost dimension. Chunking of arrays [14] can make the memory accesses for the sort-based method more regular since it does not favor any particular dimension for contiguous storage. 7 Aggregate Query Processing Most data dimensions have hierarchies defined on them. Typical OLAP queries probe summaries of data at different levels of the hierarchy. Consolidation is a widely used method to provide roll-up and drill-down functions in OLAP systems. Hierarchies can be used for data mining at multiple levels in attribute oriented approaches as discussed in a later section. Attribute-oriented induction [9] is performed for association rule discovery by ascending a predefined concept hierarchy. Each dimension in a cube can potentially have a hierarchy defined on it. This hierarchy can be applied to all the sub-cubes of the data cube. We discuss each of these in the following subsections. 7.1 Hierarchies Figure shows an example of a hierarchy for a 3D base cube with (Location, Time, Person) as the attributes (dimensions). Assume that hierarchies (City, Quarter, Group) are defined on these dimensions. Each dimension represents an attribute and each value in that dimension is a distinct value for that attribute. A hierarchy describing the grouping of some of the attribute values is defined in a dimension hierarchy. A hierarchy can have several levels. Figure 21 defines a first level hierarchy on each dimension where (a (b on. The distributed dimension can have its hierarchy distributed over a processor boundary. A global index of the hierarchy is maintained at each processor. For example, cities in C3 are distributed across P 0 and P 1 in the figure, containing the indices for (a 7 ; a 8 ; a 9 ). A consolidation operation would need a inter-processor communication if the consolidated value of C3 is required. Hierarchy at higher levels can be defined similarly. Support for roll up operations using consolidation can then be provided along each dimension. Drill downs can use them to index into the appropriate dimension index to provide a higher level of detail. A hierarchy is defined over a dimension in the base cube. Hence the sub-cubes of the data cube can use the same hierarchy, since the dimensions in the sub-cubes are a subset of the dimensions in the base cube. Additionally, more dimensions can get distributed over the processors, and a distributed hierarchy needs to be maintained for these dimensions as well. In our example, in the cube (Time, Person) a 2D cube calculated from (Location, Time, Person), Time is a distributed dimension on which a distributed hierarchy needs to be maintained. As a result, each dimension has two hierarchies for it, one local and one distributed. Queries will choose the appropriate hierarchies when performing operations on these cubes. a a a a a Person a a a c c c c a a a a a a a c c c c C3 C4 C5 Q3 Q3 Location Time Figure 21: An example hierarchy defined on the dimensions for a two processor case 7.2 Consolidation Consolidation provides an aggregate for the hierarchies defined on a dimension. For example, if it is desired to report the value for group G1 for a city C1 in the quarter Q1, then the corresponding values are aggregated by using the dimension hierarchy defined in the previous section. Some consolidation operations are local, as in they can be performed at a single processor, like C4 in the figure. However, some operations are non-local, like the values for city C3 in any quarter will need the aggregates values from both the processors as shown in Figure 22. A inter-processor communication is needed to further aggregate the data to get the correct result. local Q3 Q3 Location Time Person non-local Figure 22: Consolidation for a level 1 hierarchy defined on the cube in Figure 21 7.3 Results We again use the OLAP council benchmarks described earlier and define some sample hierarchies on each dimension. Table 5 defines a sample hierarchy for the first two levels. For our experiments we have divided the Dimension Hierarchy Level 1 Level 2 Customer 9 4 Product 8 3 Channel Time 8 2 Table 5: Sample level 1 and level 2 number of groups for hierarchies defined on the dimensions size of a dimension equally among the number of groups defined for the hierarchy at that level. The allocation of actual attribute values to the groups at a higher level is contiguous for these experiments. However, these can be arbitrarily assigned, if desired. Figure 23 shows the times for performing consolidation operation using the hierarchies on the dimensions. For Current Sales we have used the level 1 hierarchy on Product and Customer. For History Sales data, the hierarchy is used for all the dimensions, i.e. Product, Customer, Time and Channel. The figure for Budget also includes hierarchies along all dimensions, Product, Customer and Time. We see good scaling properties for each data set as the number of processors is increased. Comparing the consolidation time for Current Sales and Budget, we observe that the time for Budget data is lower even though their base cubes are similar in size. The presence of a hierarchy along the Time dimension, which is the outermost dimension in the multi-dimensional array, improves the cache performance. Chunks of size d 12 9 \Theta 9000 8 , from outermost to innermost dimension, are consolidated for Budget, whereas chunks of size 9 \Theta 900 9 \Theta 9000 8 are consolidated for Current Sales. However, both scale well as the number of processors is increased. Time (sec) CurrSale Consolidation Query Time (sec) HistSale Consolidation Query Processors0.120.380.620.881.12Time (sec) Budget Consolidation Query Figure 23: Time for consolidation query using the hierarchies defined in Table 5 for Current Sales, History Sales and Budget data sets 8 Data Mining on Data Cubes Data mining techniques are used to discover patterns and relationships from data which can enhance our understanding of the underlying domain. A data mining tool can be understood along the following components data, what is analyzed ; statistics, the kind of patterns that are searched for by the tool; filtering, how many of those patterns are to be presented to the user and the selection of those patterns; visualization, the visual representation used to present the patterns to the user; interpretation, what should the user think about when studying the pattern; exploration, in what order should the patterns be considered by the user; process support, how does one support data mining as a process as opposed to an exercise done only once. Attribute focusing(AF) is a data mining method that has most of these but in particular relies on exploration and interpretation [5]. Discovery of quantitative rules is associated with quantitative information from the database. The data cube represents quantitative summary information for a subset of the attributes. Attribute-oriented approaches [5] [9] [10] to data mining are data-driven and can use this summary information to discover association rules. Transaction data can be used to mine association rules by associating support and confidence for each rule. Support of a pattern A in a set S is the ratio of the number of transactions containing A and the total number of transactions in S. Confidence of a rule A ! B is the probability that pattern B occurs in S when pattern A occurs in S and can be defined as the ratio of the support of AB and support of A. The rule is then described as A ! B[support; confidence] and a strong association rule has a support greater than a pre-determined minimum support and a confidence greater than a pre-determined minimum confidence. This can also be taken as the measure of "interestingness" of the rule. Calculation of support and confidence for the rule A ! B involve the aggregates from the cube AB, A and ALL. Additionally, dimension hierarchies can be utilized to provide multiple level data mining by progressive generalization (roll-up) and deepening (drill-down). This is useful for data mining at multiple concept levels and interesting information can potentially be obtained at different levels. Attribute Focusing calculates associations between attributes by using the notion of percentages and sub- populations. The overall distribution of an attribute is compared with its distribution in various subsets of data. If a certain subset of data has a characteristically different distribution for the focus attribute, then that combination of attributes are marked as interesting. An event, E is a string in which x j is a possible value for some attribute and x k is a value for a different attribute of the underlying data. E is interesting to that x j 's occurrence depends on the other x i 's occurrence. The ``interestingness'' measure is the size I j (E) of the difference between: (a) the probability of E among all such events in the data set and, (b) the probability that x 1 occurred independently. The condition of interestingness can then be defined as I j (E) ? ffi, where ffi is some fixed threshold. Another condition of interestingness, in attribute focusing depends on finding the optimal number of attribute values, n formally described as I j (En AF seeks to eliminate all but the most interesting events by keeping E only if the number of attribute values, n, is just right. Eliminate one or more x i 's and I j decreases, include one ore more new x i 's to the string and I j gets no larger. The convergence to n removes patterns like En\Gamma1 and En+1 which are less interesting than En and have information already contained by En . Hence, as a result of this the user does not have to drill down or roll up from a highlighted pattern, since the event descriptions returned are at their most interesting level. 8.1 Attribute Focusing on data cubes In this section we present an algorithm to compute the first measure of interestingness of the attribute focusing method using data cubes. Figure 24 shows the algorithm. Algorithm 5 Calculation of interestingness using data cubes Step 1. Replicate each single attribute sub-cubes on all processors using a Gather followed by a Broadcast. Step 2. Perform a Reduce operation of ALL (0D cube) followed by a Broadcast to get the correct value of ALL on all processors. Step 3. Take the ratio of each element of the AB sub-cube and ALL to get P (AB). Similarly calculate P (A) and P (B) using the replicated sub-cubes A and B. Step 4. For each element i in AB calculate jP compare it with a threshold ffi, setting AB[i] to 1 if it is greater, else set it to 0. Figure 24: Algorithm for calculating interestingness between attributes A and B with data cubes Consider a 3 attribute data cube with attributes A, B and C, defining showing 2-way associations, we will calculate the interestingness function between A and B, A and C and finally between B and C. When calculating associations between A and B, the probability of E, denoted by P (AB) is the ratio of the aggregation values in the sub-cube AB and ALL. Similarly the independent probability of A, P (A) is obtained from the values in the sub-cube A, dividing them by ALL. P (B) is similarly calculated from B. The calculation threshold ffi, is performed in parallel. Since the cubes AB and A are distributed along the A dimension no replication of A is needed. However, since B is distributed in sub-cube B, and B is local on each processor in AB, B needs to be replicated on all processors. AB and A cubes are distributed, but B is replicated on all processors. Figure 25 shows a sample calculation of P(AB), P(A) and P(B) on three processors. A sample calculation is, highlighted in the figure, which is greater than ffi values from 0.001 to 0.01, and the corresponding attribute values are associated within that threshold. A A A Figure 25: Use of aggregate sub-cubes AB, A and B for calculating "interestingness" on three processors 8.2 Results We have used a few data sets from the OLAP council benchmark. In particular, Current Sales (3D), Bud- get(3D) and History Sales (4D)(as described in previous sections). We perform 2-way association calculations by performing attribute focusing for all combinations of two attributes. For example, for a 3D cube, ABC, interestingness measure will be calculated between A and B, A and C and between B and C. Typically, a few different ffi values are used for analysis of data, to vary the degree of association between the attributes. We run each calculation for 20 different ffi values, ranging from 0.001 to 0.02 in steps of 0.01. Figure 26 shows the time for these calculations on Current Sales data on 4, 8, 16 and 32 processors. We observe good speedups as the number of processors is increased. Also, communication time increases as the number of processors is increased. Replication of the 1D cubes involve communication, and with larger number of processors, the terms involving p in the complexity of the collective communication operations (refer to Table 2) increase. However, since communication time is marginal compared to computation times, the increase does not affect the overall scalability. Processors1.03.05.07.09.011.0Time (sec) CurrSale Attribute Focusing Processors2.06.010.014.0 Time (msec) CurrSale Communication time for AF Figure 26: Time for (a) attribute focusing time for 20 different ffi values for Current Sales data, (b) associated communication cost (in milliseconds) on 4, 8, 16 and processors Figure shows the attribute focusing calculation time for the History Sales data. The associated communication time is slightly higher in this case since 4 1D arrays are replicated, instead of 3 in the 3D case. The communication time (which is shown in milliseconds) is still at least an order of magnitude smaller than the computation time. The size of the cubes is smaller than the Current Sales data and hence the magnitudes of time are much smaller. Hence, the speedups are more modest as we increase the number of processors. Processors0.030.080.120.18Time (sec) HistSale Attribute Focusing Processors2.06.010.014.018.0Time (msec) HistSale Communication Time for AF Figure 27: Time for (a) attribute focusing for 20 different ffi values for History Sales data, (b) associated communication cost (in milliseconds) on 8, 16 and processors Figure 28 shows the speedups we observe for the Budget data for attribute focusing calculations. The communication costs are similar to the ones observed in the Current Sales data since they involve similar sized arrays. Processors0.501.502.503.504.505.506.50Time (sec) Budget Attribute Focusing Time Processors1.03.05.07.09.011.013.015.017.019.021.0 Time (msec) Budget Communication Time for AF Figure 28: Time for (a) attribute focusing for 20 different ffi values for Budget data, (b) associated communication cost (in milliseconds) on 8, 16, 32 and 64 processors 9 Conclusions On-Line Analytical Processing is fast gaining importance for business data analysis using large amounts of data. High performance computing is required to provide efficient analytical query processing in OLAP systems. Aggregations are an important function of OLAP queries and can benefit from the data cube operator introduced in [7]. Multidimensional databases have recently gained importance since they model the multi-dimensionality of data intuitively and are easy to visualize. They provide support for complex analytical queries and are amenable to parallelization. Most data to be analyzed comes from relational data warehouses. Hence multidimensional databases need to interface with relational systems. In this article, we presented algorithms and techniques for constructing multidimensional data cubes on distributed memory parallel computers and perform OLAP and data mining operations on them. The two methods for base cube loading, sort-based and hash-based, perform equally well on small data sets but the hash-based method is better for larger data sets and with increase in number of dimensions. We would expect the sort-based method to outperform the hash-based method for large amounts of data due to the regularity of accesses that cube loading from sorted data provides. The mismatch in the order of the dimensions when using a multidimensional array does not make the memory usage of the sort-method very efficient. Sort-based method is expected to be better for external memory algorithms because it will result in reduced disk I/O over the hash-based method. Since OLAP operations frequently require information at varying degree of detail, we have developed methods to perform consolidation queries by defining levels of hierarchies on each dimension. Our results show that these techniques are scalable for a large number of processors. Finally, we have shown a implementation of scalable data mining algorithm using attribute focusing. Our techniques are platform independent and are portable across a wide variety of parallel platforms without much effort. We are currently investigating efficient sparse methods for multidimensional sparse arrays, disk-based algorithms and techniques that apply parallel I/O optimizations to perform OLAP and data mining. Acknowledgments We would like to thank Inderpal Bhandari of IBM T.J Watson Research Center for discussions and helpful comments. --R "Providing OLAP to user-analysts : An IT mandate" "From data mining to knowledge discovery: An overview" "Advanced Scout: Data Mining and Knowledge Discovery in NBA Data" "A case study of software process improvement during development" "Attribute Focusing: Data mining for the layman" "Parallel Data Cube Construction for High Performance On-Line Analytical Process- ing" "Data Cube: A Relational Aggregation Operator Generalizing Group-By, Cross-Tab, and Sub-Totals" "An Introduction to Spatial Databases" Discovery of Multiple-Level Association Rules from Large Databases "Implementing Data Cubes Efficiently" "Introduction to Parallel Computing: Design and Analysis of Algorithms" "On Computing the Data Cube" "Efficient Organization of Large Multidimensional Arrays" "OLAP Council Benchmark" "On the Performance of an Array-Based ADT for OLAP Workloads" --TR --CTR Sanjay Goil , Alok Choudhary, High performance multidimensional analysis of large datasets, Proceedings of the 1st ACM international workshop on Data warehousing and OLAP, p.34-39, November 02-07, 1998, Washington, D.C., United States Sanjay Goil , Alok Choudhary, High performance multidimensional analysis and data mining, Proceedings of the 1998 ACM/IEEE conference on Supercomputing (CDROM), p.1-2, November 07-13, 1998, San Jose, CA Raymond T. Ng , Alan Wagner , Yu Yin, Iceberg-cube computation with PC clusters, ACM SIGMOD Record, v.30 n.2, p.25-36, June 2001 Muto , Masaru Kitsuregawa, A dynamic load balancing strategy for parallel datacube computation, Proceedings of the 2nd ACM international workshop on Data warehousing and OLAP, p.67-72, November 02-06, 1999, Kansas City, Missouri, United States Ying Chen , Frank Dehne , Todd Eavis , Andrew Rau-Chaplin, Parallel ROLAP Data Cube Construction on Shared-Nothing Multiprocessors, Distributed and Parallel Databases, v.15 n.3, p.219-236, May 2004 Lixin Fu , Joachim Hammer, CubiST: a new algorithm for improving the performance of ad-hoc OLAP queries, Proceedings of the 3rd ACM international workshop on Data warehousing and OLAP, p.72-79, November 06-11, 2000, McLean, Virginia, United States Ying Chen , Frank Dehne , Todd Eavis , Andrew Rau-Chaplin, PnP: sequential, external memory, and parallel iceberg cube computation, Distributed and Parallel Databases, v.23 n.2, p.99-126, April 2008 Frank Dehne , Todd Eavis , Susanne Hambrusch , Andrew Rau-Chaplin, Parallelizing the Data Cube, Distributed and Parallel Databases, v.11 n.2, p.181-201, March 2002 Hammer , Lixin Fu, CubiST++: Evaluating Ad-Hoc CUBE Queries Using Statistics Trees, Distributed and Parallel Databases, v.14 n.3, p.221-254, November Frank Dehne , Todd Eavis , Andrew Rau-Chaplin, The cgmCUBE project: Optimizing parallel data cube generation for ROLAP, Distributed and Parallel Databases, v.19 n.1, p.29-62, January 2006 Alex A. Freitas, Understanding the Crucial Role of AttributeInteraction in Data Mining, Artificial Intelligence Review, v.16 n.3, p.177-199, November, 2001
high performance;attribute focusing;data mining;parallel computing;data cube
593470
A Microeconomic View of Data Mining.
We present a rigorous framework, based on optimization, for evaluating data mining operations such as associations and clustering, in terms of their utility in decision-making. This framework leads quickly to some interesting computational problems related to sensitivity analysis, segmentation and the theory of games.
Introduction Data mining is about extracting interesting patterns from raw data. There is some agreement in the literature on what qualifies as a "pattern" (association rules and correlations [1, 2, 3, 5, 6, 12, 20, 21] as well as clustering of the data points [9], are some common classes of patterns but only disjointed discussion of what "interesting" means. Most work on data mining studies how patterns are to be extracted automatically, presumably for subsequent human evaluation of the extent in which they are interesting. Automatically focusing on the "interesting" patterns has received very limited formal treatment. Patterns are often deemed "interesting" on the basis of their confidence and support [1], information content [19], and unexpectedness [14, 18]. The more promising concept of actionability -the ability of the pattern to suggest concrete and profitable action by the decision-makers [15, 17, 18], and on the sound of it very close to our concerns in this paper- has not been defined rigorously or elaborated on in the data mining literature. We want to develop a theory of the value of extracted patterns. We believe that the question can only be addressed in a microeconomic framework. A pattern in the data is interesting only to the extent in which it can be used in the decision-making process of the enterprise to increase utility. 1 Any enterprise faces an optimization problem, which can generally be stated as where D is the domain of all possible decisions (production plans, marketing strategies, etc.), and f(x) is the utility or value of decision x 2 D. Such optimization problems are the object of study in mathematical programming and microeconomics. 2 The feasible region D and the objective f(x) are both comparably complex components of the problem -and classical optimization theory often treats them in a unified way via Lagrange multipliers and penalty functions [7]. However, from our point of view there is a major difference between the two: We assume that the feasible region D is basically endogenous to the enterprise, while the objective function f(x) is a function that reflects the enterprise's interaction with a multitude of other agents in the market (customers, suppliers, To quote [8], "merely finding the patterns is not enough. You must be able to respond to the patterns, to act on them, ultimately turning the data into information, the information into action, and the action into value." 2 There is such an optimization problem associated with virtually every enterprise; however in real life such problems are so involved and complex, that often nobody knows exactly their detailed formulation. The decision-makers of the enterprise base their decisions on a very rough, approximate, and heuristic understanding of the nature and behavior of the underlying optimization problem.The fact that the details of the optimization problem being solved are nebulous and unknown to the decision-makers does not make the problem less real -or its mathematical study less useful. In fact, economic theory during this century has flourished on models such as these, in which the precise nature of the functions involved is essentially unknowable; the mathematical insights derived from the abstract problem are still valuable as heuristic guides in decision-making. employees, competitors, the rest of the world). That is, at a first approximation the objective function can be rewritten as where C is a set of agents or other factors influencing the utility of the enterprise. We shall be calling elements of C "customers." We shall be deliberately vague on what they are. There are two different possible views here: On a concrete level, we can think of them as profiles of customers and other relevant agents, about whom we have gathered relevant information by a first stage of data mining; it is this first stage that our point of view seeks to influence and direct. A more abstract, but potentially equally useful, point of view is that, alternatively, we can also think of the elements of C as rows of the raw table being mined -customers, transactions, shipments, and so on. What makes this relevant to data mining is the following crucial assumption: We assume that the contribution of customer i to the utility of the enterprise under decision x, f i (x), is in fact a complicated function of the data we have on customer i. Let y i denote the data we have on customer i (the ith row of the table); then f i (x) is just g(x; y i ), some fixed function of the decision and the data. Hence, our problem is to The conventional practice in studying such problems is to replace y is some aggregate value 3 of the customers' data (aggregate demand of a product, aggregate consumer utility function, etc. Such aggregation is well-known to be inaccurate, resulting in suboptimal decisions, because of non-linearities (non-zero second partial deriva- tives) in the function g(x; y i ). Aggregation had been tolerated in traditional microeconomics because (1) the computational requirements otherwise would be enormous, and (2) it is difficult to obtain the data y i . The point in data mining, in our view, is that we now have the computational power and wealth of data necessary to attack the unaggregated optimization problem, to study the intricate ways in which correlations and clusters in the data affect the enterprise's optimal decisions. Our goal in this paper is to study certain aspects of data mining from these perspectives data mining in the context of economically motivated optimization problems, with a large volume of unaggregated data. The framework and models that we develop from these perspectives touch on a range of fundamental issues in combinatorial optimization, linear programming, and game theory; we feel that they suggest some of the first steps in a research agenda aimed at assessing quantitatively the utility of data mining operations. 3 We use "aggregate" in its microeconomics usage - summary of a parameter over a large population - which is related but not identical to its technical meaning in databases. Structure of the rest of the paper In Section 2 we present three examples which illustrate our point of view, and identify and explore its various aspects: We show by a simple example how nonlinearity is an essential aspect of interestingness; we point out that the important operations of clustering and market segmentation are affected (and in fact defined) by microeconomic considerations; and we indicate ways in which such considerations can affect the relational semantics of the mined database. Motivated by the second of these examples, in Section 3 we introduce a novel and interesting genre of problems, called segmentation problems, which capture in a crisp and stylized form the clustering aspect of data mining. One can define a segmentation problem (in fact, several versions) for any conventional optimization problem; we focus on a few natural ones with obvious data-mining flavor and interest. We show that even some of the simplest possible segmentation problems are NP-complete; however, they can be solved in time linear in the number of customers. In the Section 4 we show how optimization theory (in particular, linear programming sensitivity analysis) can be employed to develop tangible criteria of "interestingness" for data mining operations. Up to this point in the paper, we consider a single enterprise interested in mining its data to derive value. In Section 5, we turn to the problem of two competing enterprises each trying to segment a common market, adopting a set of policies to the segments they target. Building on the classical setting of game theory, we develop a notion of segmented matrix games to model this setting. This quickly leads to a number of novel (and largely unsolved) issues in computational game theory. In related work [13] in the area of discrete algorithms and complexity theory 4 , we have studied approximation algorithms for some of the most basic segmentation problems that arise from our framework. This leads to interesting connections with classical problems of combinatorial optimization - such as facility location and the maximization of submodular functions - and to settings in which one can concretely analyze the power of methods such as random sampling and greedy iterative-improvement algorithms. We refer the reader to [13] for further details. Three examples We pointed out above that aggregation is especially unsatisfactory and inaccurate when the cost function g(x; y i ) is nonlinear in y i . The next two anecdote-based examples illustrate certain interesting and common kinds of nonlinearities. The third example, based on the first, illustrates some of the more subtle ways in which the semantics of the underlying database can affect our objective in searching for correlations and nonlinearities. Example 1: Beer and Diapers. 5 Suppose that a retailer stocks two products in quantities 4 Or, as we might say, targeted at the complexity segment of the readership! 5 The correlation between the amount of beer and the amount of diapers bought by consumers is one of the amounts to be stocked are the only decision variables, bounded above by capacity: c. The profit margins in the two products are m 1 and m 2 . We have a table with 0-1 values (y 1;i ; y i;2 ) for each customer indicating whether the customer will buy a unit of each of the products. That is, in this toy example demand is known deterministically. In the first scenario, customers arrive in random order, and buy whatever part of their desired basket is available. The revenue of the enterprise is in this case a function of x 1 , x 2 , and the aggregate demands Y \Deltay i;2 . Aggregation would not distort the optimal decision, and data mining is moot. But suppose that the customers arrive in random order, and buy their desired basket in an all-or-nothing fashion; if not all items are available, the customer buys nothing. The expected profit for the enterprise from customer i is now of the form . Because of the nonlinear term, associations between the y 1;i and the y i;2 columns are now important, and the optimum decision by the enterprise depends critically on them. Aggregation will no longer do the trick, and data mining is desirable, even necessary. We propose that associations and correlations between attributes in a table are interesting if they correspond to nonlinear terms in the objective function of the enterprise, that is, for the cost function g and some attributes y This sheds an interesting and novel light on data mining activities, and begs the development of a quantitative mathematical theory of "interestingness", based on mathematical programming. We start on this path in Section 4. Example 2: Market Segmentation. Telephone companies in the U.S.A. have divided their customers into two clusters: residence and business customers; they offer different terms and prices to the two. How good is this segmentation? And are there automatic ways to discover such profitable segmentations of a market? Suppose that an enterprise has decided to subdivide its market into two segments, and apply a different marketing strategy to each segment. 6 For the purpose of the present discussion, it is not necessary to determine in detail what such a strategy consists of - the enterprise may offer different terms and prices to each segment, or send a different catalog to each segment. Thus, the decision space of the enterprise is now D 2 , where D is the set of all possible strategies. For each customer i and decision x 2 D, the enterprise reaps a profit of c i \Delta x; i.e. for simplicity, we are assuming the profit is linear. For each pair of strategies adopted, the enterprise subdivides the set C of its customers into those i for which whom the enterprise will apply strategy x 1 ), and the rest. That is, the function f i (x) is now the delightful nuggets of data mining lore. 6 This can of course generalized to k segments, or to an undetermined number of segments but with a fixed cost associated with the introduction of each segment; see Section 3. which is an interesting form of non-linearity. The enterprise adopts the pair of policies Instead of experimenting with arbitrary plausible clusterings of the data to determine if any of them are interesting and profitable, the enterprise's data miners arrive at the optimum clustering of C into two sets - those presented with strategy x 1 and those presented with x 2 - in a principled and systematic way, based on their understanding of the enter- prise's microeconomic situation, and the data available on the customers. We further explore segmentation, and the computational complexity of the novel problems that it suggests in Section 3, and from the standpoint of approximability in [13]. Example 3: Beer and Diapers, Revisited. 7 Let us now formulate a more elaborate data-mining situation related to that of Example 1. Suppose that a retailer has a database of past transactions over a period of time and over many outlets, involving the sales of several items. Suppose further that the database is organized as a relation with these attributes: transaction(location, dd, mm, yy, tt, item1, location is the particular outlet where the sale occurred, dd, mm, yy, tt records the day and time the sale occurred, and itemi is the amount of item i in the transaction. We wish to data mine this relation with an eye towards discovering correlations that will allow us to jointly promote items. Analyzing correlations between columns over the whole table is a central current problem in data mining (see, for example, [12]). However, in this example we focus on a more subtle issue: Correlations in horizontal partitions of the table (i.e., restrictions of the relation, subsets of the rows). In a certain rigorous sense, mining correlations in subsets of the rows is ill-advised: there are so many subsets of rows that it is very likely that we can find a subset exhibiting strong correlations between any two items we choose! Obviously, we need to restrict the subsets of rows for which correlations are meaningful. We posit that defining the right restrictions must take explicitly into account the ways in which we plan to generate revenue by exploiting the mined correlations. For example, suppose that the actions we contemplate are joint promotions of items at particular stores. Then the only restrictions that are legitimate are unions of ones of the form transaction[location = 'Palo Alto']. If, in addition, it is possible to make to target promotions at particular times of the day, then restrictions of the form transaction[location = 'Palo Alto' and 12 !tt] are legiti- mate. If we can target promotions by day of the week, then even more complex restrictions such as transaction[location= 'Palo Alto' and 'Monday'] may be allowed. The point is that the sets of rows on which it is meaningful to mine correlations -the targetable restrictions of the relation- depend very explicitly on actionability considerations. 7 This example and the research issues it suggests are the subject of on-going joint work with Rakesh Agrawal. Furthermore, the actions necessary for exploiting such correlations may conflict with each other. For example, it may be impossible to jointly promote two overlapping pairs of items, or we may have a realistic upper bound on the number of joint promotions we can have in each store. Or we may have a measure of the expected revenue associated with each correlation we discover, and these estimates may in fact interact in complex ways if multiple actions on correlations are discovered. We wish to find a set of actions that generates maximum revenue. This point of view leads to interesting and novel optimization problems worthy of further study in both complexity and data mining. 3 Market Segmentation Consider any optimization problem with linear objective function c \Delta x: Almost all combinatorial optimization problems, such as the minimum spanning tree prob- lem, the traveling salesman problem, linear programming, knapsack, etc., can be formulated in this way. Suppose that we have a very large set C of customers, each with his/her own version c i of the objective vector. We wish to partition C into k parts C so that we maximize the sum of the optima Problem (3) captures the situation in which an enterprise wishes to segment its customers into clusters, so that it can target a different marketing strategy - e.g. a different advertising campaign, or a different price scheme - on each cluster. It seems a very hard problem, since it requires optimization over all partitions of C, and therefore computation exponential in n. Consider now the problem in which we wish to come up with k solutions x so as to maximize the quantity In contrast to problem (3), problem (4) can be solved exhaustively in time O(nm k ), where m is the number of solutions; despite its exponential dependence on k, and its dependence on m which is presumably substantial, it is linear in n, which is assumed here to be the truly large quantity. As we shall see in Section 3.2 the exponential dependence on k and the dependence on m seems inherent, even when the underlying optimization problem is extremely simple. It is easy to see that problems (3) and (4) are equivalent. The intuitive reason is that they are max-max problems, and therefore the maximization operators commute. That is, in order to divide the customers in k segments, all we have to do is come up with k solutions, and then classify each customer to one of k segments, depending on which solution is maximum for this customer. The computational implications of this observation are very favorable, since O(m k n), the time naively needed for (4), is much better than O(m2 nk ). There is another variant of the general segmentation problem, arguably more realistic, in which we are seeking to choose k solutions x of our choice, to minimize where fl is the cost of adding another solution and segment. Like problem (4), problem (5) can be solved exhaustively in time that is linear in n, with a larger dependence on m and k. Problems (3) (or (4)) and (5) constitute a novel genre of problems, which we call segmentation problems. These problems are extremely diverse (we can define one for each classical optimization problem). We believe that they are interesting because they capture the value of clustering as a data mining operation. Clustering is an important problem area of algorithmic research that is also of significant interest to data mining -which it predates. One of the main motivations of clustering has been the hope that, by clustering the data in meaningfully distinct clusters, we can then proceed to make independent decisions for each cluster. To our knowledge, this is the first formalism of clustering that explicitly embodies this motivation. 3.1 Specific problems There is no end to the problems we can define in this way: The minimum spanning tree segmentation problem, the TSP segmentation problem, the linear programming segmentation problem, and so on -and at least three variants of each. There are a few of these problems, however, that seem especially natural and compelling, in view of the data-mining motivation (we only give the fixed k version of hypercube segmentation: Given n vectors in c and an integer k, find a set of k vectors x to maximize the sum This problem captures the situation in which we know the preferences of n customers on d components of a product, for which there is a binary choice for each component. We wish to develop k semi-customized versions of the product, so as to maximize the total number of customer-component pairs for which the customer likes the component of the variant he or she chooses. Another interesting segmentation problem is catalog segmentation: Given n vectors in c find a set of k vectors x each with exactly r ones, to maximize the sum In this case, we know the interests of each customer, and we wish to mail k customized catalogs, each with r pages, to maximize total hit (i.e. the total number of pages of interest to customers). 3.2 Complexity Even the most trivial optimization problems (e.g., maximizing a linear function over the d-dimensional whose ordinary version can be solved by aligning the solution with the cost vector) become NP-complete in their segmentation versions. We summarize the complexity results below: Theorem 3.1 The segmentation problems (all three versions) corresponding to the following feasible sets D are NP-complete: (1) The d-dimensional unit ball, even with d-dimensional unit L 1 ball; (3) the r-slice of the d-dimensional hypercube (the catalog segmentation problem), even with the d-dimensional hypercube, even with the set of all spanning trees of a graph G, even with 2. Notice that the optimization problems underlying these problems are extremely easy: The one underlying (1) can be solved by aligning the solution with the cost vector, the one for (2) has only 2d vertices, the one for (3) can be solved by choosing the r most popular elements, and the one for (4) by simply picking the vertex that coordinate-wise agrees in sign with the cost vector. Since (2) has 2d vertices it can be solved in O((2d) k n) time, which is polynomial when k is bounded. The NP-completeness reductions are surprisingly diverse: (1) is proved by a reduction from max cut, (2) from hitting set, (3) from bipartite clique, and (4) from maximum satisfiability with clauses that are equations modulo two. Finally, for spanning tree segmentation we use a reduction from hypercube segmentation (the latter problem is essentially a special case of the former, in which the graph is a path with two parallel edges between each pair of consecutive nodes). Here we sketch only the proof of (1). Suppose that we have a graph direct its edges arbitrarily, and consider the node-edge incidence matrix of G (the jV j \Theta jEj matrix with the (i; th entry equal to 1 if the j th edge enters the i th node, \Gamma1 if the j th edge leaves the i th node, and 0 otherwise). Let the jV j rows of this matrix define the cost vectors fv of the segmentation problem. Thus, we seek to divide these jV j vectors into two sets, S 1 and choose an optimal solution for each set. Let oe i denote the sum of the vectors in the set S i , for 2. Since D is the unit ball, an optimal solution for S i is simply the unit vector in the direction of oe i , and hence the value of the solution associated with simply the sum of the Euclidean norms, koe However, it is easy to see that for any partition of the vertices, koe is twice the square root of the number of edges in the cut (S 1 (because in the two sums the only entries that do not cancel out are the ones that correspond to edges in the cut); hence, solving the segmentation problem is the same as finding the maximum cut of the graph. 2 When the problem dimension is fixed, most of these problems be solved in polynomial time: Theorem 3.2 Segmentation problems (2-5) in the previous theorem can be solved in linear time when the number of dimensions is fixed. Problem (1) (the unit ball) can be solved in time O(n 2 in two dimensions, and is NP-complete (for variable k) in three dimensions. When the number of dimensions is a fixed constant d, the number of extreme solutions in each problem (2-5) is constant (2d, r Thus the number of all possible sets of k solutions is also a bounded constant, call it c; obviously, such problems can be solved in time proportional to cn. For (1), the 2-dimensional algorithm is based on dynamic programming, while the NP-completeness proof for is by a reduction from a facility location problem. 2 4 Data Mining as Sensitivity Analysis The field of sensitivity analysis in optimization seeks to develop principles and methodologies for determining how the optimum decision changes as the data change. In this section we give an extensive example suggesting that many common data mining activities can be fruitfully analyzed within this framework. To fix ideas, we shall consider the case in which the optimization problem facing the enterprise is a linear program [10, 16], that is, we have an m \Theta n matrix A (the constraint n), an m-vector b (the resource bounds), and an n-row vector c (the objective function coefficients), and we seek to c \Delta x: (1) The columns of A -the components of x- are called activities, and the rows of A are called constraints. Extensions to non-linear inequalities and objectives are possible, with the Kuhn-Tucker conditions [7] replacing the sensitivity analysis below. We assume, as postulated in the introduction, that the entries of A and b are fixed and given (they represent the endogenous constraints of the enterprise), while the coefficients of c depend in complex ways on a relation, which we denote by C (we make no distinction between the relation C, and its set of rows, called customers). The ith tuple of C -the ith customer- is denoted y i , and we assume that c j is just is a function mapping the product of the domains of C to the reals. We noted in the introduction that the desirability of data mining depends on whether the functions f j are "non-linear;" we shall next formalize what exactly we mean by that. Suppose that the function f j (y linear, and thus appropriate single-attribute aggregates (in particular, averages) of the table fy i g will accurately capture the true value of c j . For example, if f all we need in order to compute c j is the average values of the first two columns of C. If then we say that f j is nonlinear. Assume that all attributes of the relation C are real numbers in the range [0; 1], and that f j depends on two attributes, call them k j and ' j (extending to more general situations is straightforward, but would make our notation and development less concrete and more cumbersome). We also assume that we have an estimate D j - 0 on the absolute value of the derivative . means that f j is nonlinear. We investigate under what circumstances it is worthwhile to measure the correlations of the pair of attributes corresponding to the coefficient c j -that is to say, to data mine the two attributes related to the jth activity. Without data mining, the coefficient c j will be approximated by the function f j of the aggregate values of the attributes k j and ' j . Suppose that we have solved the linear program (1) of the enterprise, based on the aggregate estimation of the c j 's. This means that we have chosen a subset of m out the n activities, set all other activities at zero level, and selected as the optimum level of the kth chosen activity the kth component of the vector B here by B we denote the square nonsingular submatrix of A corresponding to the m chosen activities, and by c B the vector c restricted to the m chosen activities. It is the fundamental result in the theory of linear programming that, if a square submatrix B is nonsingular and satisfies these two inequalities, and only is the optimum. Matrix B the simplex tableau, a matrix maintained by the simplex algorithm. The important question for us is the following: Under what changes in the c j 's will the chosen level of activities continue to be optimal? The theory of sensitivity analysis of linear programming [10, 16] answers this question precisely: If the c j 's are changed to new values c 0 , then the optimal solution remains the same if and only if the condition c 0 \Gammac 0 A is the simplex tableau. That is, if each coefficient c j is changed from its value c j , calculated based on aggregated data, to its true value c 0 based on raw data, the optimum decision of the enterprise remains the same under the above conditions. This suggests the following quantitative measure of "interestingness" vis 'a vis data mining of the jth activity: Definition 4.1 For each activity j, define its interestingness I j as follows: I where the c i 's are defined as above. Notice that both terms in I j may be infinite; our convention is that, if D I j is zero. The larger I j is, the more likely it is that mining correlations in the attributes k j and ' j will reveal that the true optimum decision of the enterprise is different from the one arrived at by aggregation. To put it qualitatively: An activity j is interesting if the function f j has a highly non-linear cross-term, and either c j is small, or the jth row of the tableau has large positive coefficients at columns with small 's. As the above analysis suggests, our point of view, combined with classical linear programming sensitivity analysis, can be the basis of a quantitative theory for determining when data mining can affect decisions, ultimately a theory for predicting the value of data mining operations. 5 Segmentation in a Model of Competition So far we have considered the data mining problems faced by a single enterprise trying to optimize its utility. But it is also natural to consider data mining in a setting that involves competition among several enterprises; to indicate some of the issues that arise, we consider the classical framework of two-player games. Recall that in classical game theory, such a game involves two players: I with m strategies and II with n strategies. The game is defined in terms of two m \Theta n matrices A; B, where A ij 2 R is the revenue of Player I in the case that I chooses strategy i and II chooses strategy j; the matrix B is defined similarly in terms of Player II. Such games have been studied and analyzed in tremendous depth in the area of game theory [4]. To add a data mining twist to the situation, suppose that two corporations I and II each have a fixed set of m and n marketing strategies, respectively, for attracting consumers. Each combination of these strategies has a different effect on each consumer, and we assume that both corporations know this effect. Thus, the set C of customers can be thought of as a set of , where the (i; th entry of A k is the amount won by I if I chooses strategy i and II chooses strategy j with respect to customer k; similarly for B k and II. The two corporations are going to come up with segmented strategies. Player I partitions the set of customers C into k sets, and chooses a row-strategy for each set; similarly, Player II partitions C into ' sets, and chooses a column-strategy for each set. Thus, the strategy space of the players is the set of all partitions of the N customers into k and ' sets, respectively. The classical theory of games tells us that there is in general a mixed equilibrium, in which each player selects a linear combination (which can be viewed as a probability distribution) over all possible segmentations of the market. Under these choices, each player is doing "the best he can": neither has an incentive to alter the mix of strategies he has adopted. This can be viewed as an existential result, a definition of rational behavior, rather than a concrete and efficient computational procedure for determining what each player should do. Example 4: Catalog Wars. Suppose that the strategy space of Player I consists of all possible catalogs with p pages to be mailed, and similarly for Player II, where each corporation has a different set of alternatives for each page of the catalog. Each corporation knows which alternatives each given customer is going to like, and a corporation attracts a customer if the corporation's catalog contains more pages the customer likes than the competitor's catalog For each customer, we will say that the corporation that attracts the customer has a payoff of +1, while the competitor has a payoff of \Gamma1; the payoff will be zero in the case of a tie. Thus, in this hypothetical case, each individual customer is a zero-sum game between the corporations. Each corporation is going to mail out a fixed number k of versions of the catalog, with each customer getting one version from each. The overall zero sum game has as strategy space for each player the set of all possible k-tuples of catalogs. The payoff for Player I of each pair of k-tuples, in the set-up above, is the number of customers who like more pages from some of the k catalogs of Player I than from any of the catalogs of Player II, minus the corresponding number with the roles of Players I and II interchanged. There are many game-theoretic and computational issues suggested by such market segmentation games. Under what conditions is the existence of a pure Nash equilibrium guar- anteed? What is the computational complexity of finding a mixed equilibrium? 6 Conclusions and Open Problems We are interested in a rigorous framework for the automatic evaluation of data mining operations. To this end, we have proposed a set of principles and ideas, and developed frameworks of four distinct styles. Our framework of optimization problems with coefficients depending nonlinearly on data, and our definition of interestingness within this framework (Section 4), should be seen as an example of the kinds of theories, methodologies, and tools that can be developed; we have refrained from stating and proving actual results in that section exactly because the potential range of interesting theorems that are straightforward applications of these ideas is too broad. The segmentation problems we study in Section 3 are also meant as stylized and abstract versions of the kinds of computational tasks that emerge in the wake of our point of view. The more database-theoretic concept of targetable restrictions of databases introduced in Example 3 needs further exploration. Finally, we have only pointed out the interesting models and problems that arise if segmentation is seen in a context involving competition. The technical and model-building open problems suggested by this work are too many to list exhaustively here. For example: What interesting theorems can be stated and proved within the sensitivity analysis framework of Section 4? Are there interesting and realistic segmentation problems that are tractable, or for which fast and effective heuristics can be developed? How does one model temporal issues within this framework (we can view problem faced by the enterprise as a Markov Decision Process [11]), and what insights result? More importantly, what new ideas are needed in order to apply our framework in realistic situations, in which the precise nature of the enterprise's decision problem is murky, and the data on customers incomplete, unreliable, and only very implicitly containing information on the parameters of interest (such as revenue resulting from each possible marketing strategy)? Finally, we have focused on data mining as an activity by a revenue-maximizing enterprise examining ways to exploit information it has on its customers. There are of course many important applications of data mining - for example, in scientific and engineering contexts - to which our framework does not seem to apply directly. However, even in these applications it is possible that insight can be gained by identifying, articulating quantitatively, and taking into account the goals and objectives of the data mining activity. --R "Mining association rules between sets of items in a large database." "Fast discovery of association rules." Fast algorithms for mining association rules. Handbook of Game Theory "Dynamic itemset counting and implication rules for market basket data" "Beyond Market Baskets: Generalizing Association Rules to Correlations." Nonlinear Programming: Analysis and Methods. Data Mining Techniques. "Data mining: An overview from a database perspective." Finite State Markov Decision Processes. "Data mining, hypergraph transversals, and machine learning" "Segmentation problems," "Post-analysis of learned rules." "A comparison of approaches for maximizing business payoff of prediction models" Combinatorial Optimization: Algorithms and Complexity (second edition). "The interestingness of deviations." "What makes patterns interesting in knowledge discovery systems ." "Rule induction using information theory." Mining generalized association rules. Sampling large databases for finding association rules. --TR --CTR Martin Ester , Rong Ge , Wen Jin , Zengjian Hu, A microeconomic data mining problem: customer-oriented catalog segmentation, Proceedings of the tenth ACM SIGKDD international conference on Knowledge discovery and data mining, August 22-25, 2004, Seattle, WA, USA Tom Brijs , Bart Goethals , Gilbert Swinnen , Koen Vanhoof , Geert Wets, A data mining framework for optimal product selection in retail supermarket data: the generalized PROFSET model, Proceedings of the sixth ACM SIGKDD international conference on Knowledge discovery and data mining, p.300-304, August 20-23, 2000, Boston, Massachusetts, United States Ali Amiri, Customer-oriented catalog segmentation: effective solution approaches, Decision Support Systems, v.42 n.3, p.1860-1871, December 2006 Jon Kleinberg , Christos Papadimitriou , Prabhakar Raghavan, Segmentation problems, Journal of the ACM (JACM), v.51 n.2, p.263-280, March 2004 Jianying Hu , Aleksandra Mojsilovic, High-utility pattern mining: A method for discovery of high-utility item sets, Pattern Recognition, v.40 n.11, p.3317-3324, November, 2007 Hong Yao , Howard J. Hamilton, Mining itemset utilities from transaction databases, Data & Knowledge Engineering, v.59 n.3, p.603-626, December 2006 Heikki Mannila, Theoretical frameworks for data mining, ACM SIGKDD Explorations Newsletter, v.1 n.2, January 2000 Tom Brijs , Gilbert Swinnen , Koen Vanhoof , Geert Wets, Building an Association Rules Framework to Improve Product Assortment Decisions, Data Mining and Knowledge Discovery, v.8 n.1, p.7-23, January 2004 Tom Brijs , Gilbert Swinnen , Koen Vanhoof , Geert Wets, Using association rules for product assortment decisions: a case study, Proceedings of the fifth ACM SIGKDD international conference on Knowledge discovery and data mining, p.254-260, August 15-18, 1999, San Diego, California, United States Dharmendra S. Modha , W. Scott Spangler, Feature Weighting in k-Means Clustering, Machine Learning, v.52 n.3, p.217-237, September Cuiping Li , Beng Chin Ooi , Anthony K. H. Tung , Shan Wang, DADA: a data cube for dominant relationship analysis, Proceedings of the 2006 ACM SIGMOD international conference on Management of data, June 27-29, 2006, Chicago, IL, USA Raymond Chi-Wing Wong , Ada Wai-Chee Fu , Ke Wang, Data Mining for Inventory Item Selection with Cross-Selling Considerations, Data Mining and Knowledge Discovery, v.11 n.1, p.81-112, July 2005 Inderjit S. Dhillon , Dharmendra S. Modha, Concept Decompositions for Large Sparse Text Data Using Clustering, Machine Learning, v.42 n.1-2, p.143-175, January-February 2001 Ioannis N. Kouris , Christos H. Makris , Athanasios K. Tsakalidis, Using information retrieval techniques for supporting data mining, Data & Knowledge Engineering, v.52 n.3, p.353-383, March 2005
clustering;optimization;market segmentation
593484
Parallel Formulations of Decision-Tree Classification Algorithms.
Classification decision tree algorithms are used extensively for data mining in many domains such as retail target marketing, fraud detection, etc. Highly parallel algorithms for constructing classification decision trees are desirable for dealing with large data sets in reasonable amount of time. Algorithms for building classification decision trees have a natural concurrency, but are difficult to parallelize due to the inherent dynamic nature of the computation. In this paper, we present parallel formulations of classification decision tree learning algorithm based on induction. We describe two basic parallel formulations. One is based on Synchronous Tree Construction Approach and the other is based on Partitioned Tree Construction Approach. We discuss the advantages and disadvantages of using these methods and propose a hybrid method that employs the good features of these methods. We also provide the analysis of the cost of computation and communication of the proposed hybrid method. Moreover, experimental results on an IBM SP-2 demonstrate excellent speedups and scalability.
Introduction Classification is an important data mining problem. A classification problem has an input dataset called the training set which consists of a number of examples each having a number of attributes. The attributes are either continuous, when the attribute values are ordered, or categorical, when the attribute values are un- ordered. One of the categorical attributes is called the class label or the classifying attribute. The objective is to use the training dataset to build a model of the class label based on the other attributes such that the model can be used to classify new data not from the training dataset. Application domains include retail target marketing, fraud detection, and design of telecommunication service plans. Several classification models like neural networks [17], genetic algorithms [11], and decision trees [20] have been proposed. Decision trees are probably the most popular since they obtain reasonable accuracy [9] and they are relatively inexpensive to compute. Most current classification algorithms such as C4.5 [20], and SLIQ [18] are based on the ID3 classification decision tree algorithm [20]. In the data mining domain, the data to be processed tends to be very large. Hence, it is highly desirable to design computationally efficient as well as scalable algorithms. One way to reduce the computational complexity of building a decision tree classifier using large training datasets is to use only a small sample of the training data. Such methods do not yield the same classification accuracy as a decision tree classifier that uses the entire data set [24, 5, 6, 7]. In order to get reasonable accuracy in a reasonable amount of time, parallel algorithms may be required. Classification decision tree construction algorithms have natural concurrency, as once a node is generated, all of its children in the classification tree can be generated concurrently. Furthermore, the computation for generating successors of a classification tree node can also be decomposed by performing data decomposition on the training data. Nevertheless, parallelization of the algorithms for construction the classification tree is challenging for the following reasons. First, the shape of the tree is highly irregular and is determined only at runtime. Furthermore, the amount of work associated with each node also varies, and is data dependent. Hence any static allocation scheme is likely to suffer from major load imbalance. Second, even though the successors of a node can be processed concurrently, they all use the training data associated with the parent node. If this data is dynamically partitioned and allocated to different processors that perform computation for different nodes, then there is a high cost for data movements. If the data is not partitioned appropriately, then performance can be bad due to the loss of locality. In this paper, we present parallel formulations of classification decision tree learning algorithm based on induction. We describe two basic parallel formulations. One is based on Synchronous Tree Construction Approach and the other is based on Partitioned Tree Construction Approach. We discuss the advantages and disadvantages of using these methods and propose a hybrid method that employs the good features of these methods. We also provide the analysis of the cost of computation and communication of the proposed hybrid method. Moreover, experimental results on an IBM SP-2 demonstrate excellent speedups and scalability. 2. Related Work 2.1. Sequential Decision-Tree Classification Algorithms Most of the existing induction-based algorithms like C4.5 [20], CDP [1], SLIQ [18], and SPRINT [21] use Hunt's method [20] as the basic algorithm. Here is a recursive description of Hunt's method for constructing a decision tree from a set T of training cases with classes denoted fC 1 g. Case 1 T contains cases all belonging to a single class C j . The decision tree for T is a leaf identifying class C j . Case 2 T contains cases that belong to a mixture of classes. A test is chosen, based on a single attribute, that has one or more mutually exclusive outcomes On g. Note that in many implementations, n is chosen to be 2 and this leads to a binary decision tree. T is partitioned into subsets T contains all the cases in T that have outcome O i of the chosen test. The decision tree for T consists of a decision node identifying the test, and one Outlook Temp(F) Humidity(%) Windy? Class sunny sunny sunny 85 85 false Don't Play sunny 72 95 false Don't Play sunny 69 70 false Play overcast 72 90 true Play overcast 83 78 false Play overcast overcast 81 75 false Play rain 71 80 true Don't Play rain rain rain 68 80 false Play rain 70 96 false Play Table 1. A small training data set [Qui93] Play Dont' Play Play Play Play Dont' Play Dont' Play Play sunny overcast rain Outlook sunny overcast rain Play Outlook Humidity non-leaf node expandable leaf node non-expandable leaf node (a) Initial Classification Tree (b) Intermediate Classification Tree (c) Final Classification Tree Figure 1. Demonstration of Hunt's Method branch for each possible outcome. The same tree building machinery is applied recursively to each subset of training cases. Case 3 T contains no cases. The decision tree for T is a leaf, but the class to be associated with the leaf must be determined from information other than T . For example, C4.5 chooses this to be the most frequent class at the parent of this node. Attribute Value Class Play Don't Play overcast 4 0 Table 2. Class Distribution Information of Attribute Outlook Attribute Value Binary Test Class Play Don't Play Table 3. Class Distribution Information of Attribute Humidity Table 1 shows a training data set with four data attributes and two classes. Figure 1 shows how Hunt's method works with the training data set. In case 2 of Hunt's method, a test based on a single attribute is chosen for expanding the current node. The choice of an attribute is normally based on the entropy gains of the attributes. The entropy of an attribute is calculated from class distribution information. For a discrete attribute, class distribution information of each value of the attribute is required. Table 2 shows the class distribution information of data attribute Outlook at the root of the decision tree shown in Figure 1. For a continuous attribute, binary tests involving all the distinct values of the attribute are considered. Table 3 shows the class distribution information of data attribute Humidity. Once the class distribution information of all the attributes are gath- ered, each attribute is evaluated in terms of either entropy [20] or Gini Index [4]. The best attribute is selected as a test for the node expansion. The C4.5 algorithm generates a classification-decision tree for the given training data set by recursively partitioning the data. The decision tree is grown using depth-first strategy. The algorithm considers all the possible tests that can split the data set and selects a test that gives the best information gain. For each discrete attribute, one test with outcomes as many as the number of distinct values of the attribute is considered. For each continuous attribute, binary tests involving every distinct value of the attribute are considered. In order to gather the entropy gain of all these binary tests efficiently, the training data set belonging to the node in consideration is sorted for the values of the continuous attribute and the entropy gains of the binary cut based on each distinct values are calculated in one scan of the sorted data. This process is repeated for each continuous attribute. Recently proposed classification algorithms SLIQ [18] and SPRINT [21] avoid costly sorting at each node by pre-sorting continuous attributes once in the begin- ning. In SPRINT, each continuous attribute is maintained in a sorted attribute list. In this list, each entry contains a value of the attribute and its corresponding record id. Once the best attribute to split a node in a classification tree is determined, each attribute list has to be split according to the split decision. A hash table, of the same order as the number of training cases, has the mapping between record ids and where each record belongs according to the split decision. Each entry in the attribute list is moved to a classification tree node according to the information retrieved by probing the hash table. The sorted order is maintained as the entries are moved in pre-sorted order. Decision trees are usually built in two steps. First, an initial tree is built till the leaf nodes belong to a single class only. Second, pruning is done to remove any overfitting to the training data. Typically, the time spent on pruning for a large dataset is a small fraction, less than 1% of the initial tree generation. Therefore, in this paper, we focus on the initial tree generation only and not on the pruning part of the computation. 2.2. Parallel Decision-Tree Classification Algorithms Several parallel formulations of classification rule learning have been proposed re- cently. Pearson presented an approach that combines node-based decomposition and attribute-based decomposition [19]. It is shown that the node-based decomposition (task parallelism) alone has several probelms. One problem is that only a few processors are utilized in the beginning due to the small number of expanded tree nodes. Another problem is that many processors become idle in the later stage due to the load imbalance. The attribute-based decomposition is used to remedy the first problem. When the number of expanded nodes is smaller than the available number of processors, multiple processors are assigned to a node and attributes are distributed among these processors. This approach is related in nature to the partitioned tree construction approach discussed in this paper. In the partitioned tree construction approach, actual data samples are partitioned (horizontal parti- tioning) whereas in this approach attributes are partitioned (vertical partitioning). In [8], a few general approaches for parallelizing C4.5 are discussed. In the Dynamic Task Distribution (DTD) scheme, a master processor allocates a subtree of the decision tree to an idle slave processor. This scheme does not require communication among processors, but suffers from the load imbalance. DTD becomes similar to the partitioned tree construction approach discussed in this paper once the number of available nodes in the decision tree exceeds the number of processors. The DP-rec scheme distributes the data set evenly and builds decision tree one node at a time. This scheme is identical to the synchronous tree construction approach discussed in this paper and suffers from the high communication overhead. The DP-att scheme distributes the attributes. This scheme has the advantages of being both load-balanced and requiring minimal communications. However, this scheme does not scale well with increasing number of processors. The results in [8] show that the effectiveness of different parallelization schemes varies significantly with data sets being used. Kufrin proposed an approach called Parallel Decision Trees (PDT) in [15]. This approach is similar to the DP-rec scheme [8] and synchronous tree construction approach discussed in this paper, as the data sets are partitioned among proces- sors. The PDT approach designate one processor as the "host" processor and the remaining processors as "worker" processors. The host processor does not have any data sets, but only receives frequency statistics or gain calculations from the worker processors. The host processor determines the split based on the collected statistics and notify the split decision to the worker processors. The worker processors collect the statistics of local data following the instruction from the host processor. The PDT approach suffers from the high communication overhead, just like DP-rec scheme and synchronous tree construction approach. The PDT approach has an additional communication bottleneck, as every worker processor sends the collected statistics to the host processor at the roughly same time and the host processor sends out the split decision to all working processors at the same time. The parallel implementation of SPRINT [21] and ScalParC [13] use methods for partitioning work that is identical to the one used in the synchronous tree construction approach discussed in this paper. Serial SPRINT [21] sorts the continuous attributes only once in the beginning and keeps a separate attribute list with record identifiers. The splitting phase of a decision tree node maintains this sorted order without requiring to sort the records again. In order to split the attribute lists according to the splitting decision, SPRINT creates a hash table that records a mapping between a record identifier and the node to which it goes to based on the splitting decision. In the parallel implementation of SPRINT, the attribute lists are split evenly among processors and the split point for a node in the decision tree is found in parallel. However, in order to split the attribute lists, the full size hash table is required on all the processors. In order to construct the hash table, all-to-all broadcast [16] is performed, that makes this algorithm unscalable with respect to runtime and memory requirements. The reason is that each processor requires O(N) memory to store the hash table and O(N) communication overhead for all-to-all broadcast, where N is the number of records in the data set. The recently proposed ScalParC [13] improves upon the SPRINT by employing a distributed 7hash table to efficiently implement the splitting phase of the SPRINT. In ScalParC, the hash table is split among the processors, and an efficient personalized communication is used to update the hash table, making it scalable with respect to memory and runtime requirements. Goil, Aluru, and Ranka proposed the Concatenated Parallelism strategy for efficient parallel solution of divide and conquer problems [10]. In this strategy, the mix of data parallelism and task parallelism is used as a solution to the parallel divide and conquer algorithm. Data parallelism is used until there are enough sub-tasks are genearted, and then task parallelism is used, i.e., each processor works on independent subtasks. This strategy is similar in principle to the partitioned tree construction approach discussed in this paper. The Concatenated Parallelism strategy is useful for problems where the workload can be determined based on the size of subtasks when the task parallelism is employed. However, in the problem of classificatoin decision tree, the workload cannot be determined based on the size of data at a particular node of the tree. Hence, one time load balancing used in this strategy is not well suited for this particular divide and conquer problem. 3. Parallel Formulations In this section, we give two basic parallel formulations for the classification decision tree construction and a hybrid scheme that combines good features of both of these approaches. We focus our presentation for discrete attributes only. The handling of continuous attributes is discussed in Section 3.4. In all parallel formulations, we assume that N training cases are randomly distributed to P processors initially such that each processor has N=P cases. 3.1. Synchronous Tree Construction Approach In this approach, all processors construct a decision tree synchronously by sending and receiving class distribution information of local data. Major steps for the approach are shown below: 1. Select a node to expand according to a decision tree expansion strategy (eg. Depth-First or Breadth-First), and call that node as the current node. At the beginning, root node is selected as the current node. 2. For each data attribute, collect class distribution information of the local data at the current node. 3. Exchange the local class distribution information using global reduction [16] among processors. 4. Simultaneously compute the entropy gains of each attribute at each processor and select the best attribute for child node expansion. 5. Depending on the branching factor of the tree desired, create child nodes for the same number of partitions of attribute values, and split training cases accordingly Class Distribution Information Class Distribution Information Figure 2. Synchronous Tree Construction Approach with Depth-First Expansion Strategy 6. Repeat above steps (1-5) until no more nodes are available for the expansion. Figure 2 shows the overall picture. The root node has already been expanded and the current node is the leftmost child of the root (as shown in the top part of the figure). All the four processors cooperate to expand this node to have two child nodes. Next, the leftmost node of these child nodes is selected as the current node (in the bottom of the figure) and all four processors again cooperate to expand the node. The advantage of this approach is that it does not require any movement of the training data items. However, this algorithm suffers from high communication cost and load imbalance. For each node in the decision tree, after collecting the class distribution information, all the processors need to synchronize and exchange the distribution information. At the nodes of shallow depth, the communication overhead is relatively small, because the number of training data items to be processed is relatively large. But as the decision tree grows and deepens, the number of training set items at the nodes decreases and as a consequence, the computation of the class distribution information for each of the nodes decreases. If the average branching factor of the decision tree is k, then the number of data items in a child node is on the average 1 th of the number of data items in the parent. However, the size of communication does not decrease as much, as the number of attributes to be considered goes down only by one. Hence, as the tree deepens, the communication overhead dominates the overall processing time. The other problem is due to load imbalance. Even though each processor started out with the same number of the training data items, the number of items belonging to the same node of the decision tree can vary substantially among processors. For example, processor 1 might have all the data items on leaf node A and none on leaf node B, while processor 2 might have all the data items on node B and none on node A. When node A is selected as the current node, processor 2 does not have any work to do and similarly when node B is selected as the current node, processor 1 has no work to do. This load imbalance can be reduced if all the nodes on the frontier are expanded simultaneously, i.e. one pass of all the data at each processor is used to compute the class distribution information for all nodes on the frontier. Note that this improvement also reduces the number of times communications are done and reduces the message start-up overhead, but it does not reduce the overall volume of communications. In the rest of the paper, we will assume that in the synchronous tree construction algorithm, the classification tree is expanded breadth-first manner and all the nodes at a level will be processed at the same time. 3.2. Partitioned Tree Construction Approach In this approach, whenever feasible, different processors work on different parts of the classification tree. In particular, if more than one processors cooperate to expand a node, then these processors are partitioned to expand the successors of this node. Consider the case in which a group of processors Pn cooperate to expand node n. The algorithm consists of following steps: Processors in Pn cooperate to expand node n using the method described in Section 3.1. Once the node n is expanded in to successor nodes, the processor group Pn is also partitioned, and the successor nodes are assigned to processors as follows: Case 1: If the number of successor nodes is greater than jP n j, 1. Partition the successor nodes into jP n j groups such that the total number of training cases corresponding to each node group is roughly equal. Assign each processor to one node group. 2. Shuffle the training data such that each processor has data items that belong to the nodes it is responsible for. 3. Now the expansion of the subtrees rooted at a node group proceeds completely independently at each processor as in the serial algorithm. Case 2: Otherwise (if the number of successor nodes is less than jP n j), 1. Assign a subset of processors to each node such that number of processors assigned to a node is proportional to the number of the training cases corresponding to the node. Data Item Data Item Figure 3. Partitioned Tree Construction Approach 2. Shuffle the training cases such that each subset of processors has training cases that belong to the nodes it is responsible for. 3. Processor subsets assigned to different nodes develop subtrees inde- pendently. Processor subsets that contain only one processor use the sequential algorithm to expand the part of the classification tree rooted at the node assigned to them. Processor subsets that contain more than one processor proceed by following the above steps recursively. At the beginning, all processors work together to expand the root node of the classification tree. At the end, the whole classification tree is constructed by combining subtrees of each processor. Figure 3 shows an example. First (at the top of the figure), all four processors cooperate to expand the root node just like they do in the synchronous tree construction 1approach. Next (in the middle of the figure), the set of four processors is partitioned in three parts. The leftmost child is assigned to processors 0 and 1, while the other nodes are assigned to processors 2 and 3, respectively. Now these sets of processors proceed independently to expand these assigned nodes. In par- ticular, processors 2 and processor 3 proceed to expand their part of the tree using the serial algorithm. The group containing processors 0 and 1 splits the leftmost child node into three nodes. These three new nodes are partitioned in two parts (shown in the bottom of the figure); the leftmost node is assigned to processor 0, while the other two are assigned to processor 1. From now on, processors 0 and 1 also independently work on their respective subtrees. The advantage of this approach is that once a processor becomes solely responsible for a node, it can develop a subtree of the classification tree independently without any communication overhead. However, there are a number of disadvantages of this approach. The first disadvantage is that it requires data movement after each node expansion until one processor becomes responsible for an entire subtree. The communication cost is particularly expensive in the expansion of the upper part of the classification tree. (Note that once the number of nodes in the frontier exceeds the number of processors, then the communication cost becomes zero.) The second disadvantage is poor load balancing inherent in the algorithm. Assignment of nodes to processors is done based on the number of training cases in the successor nodes. However, the number of training cases associated with a node does not necessarily correspond to the amount of work needed to process the subtree rooted at the node. For example, if all training cases associated with a node happen to have the same class label, then no further expansion is needed. 3.3. Hybrid Parallel Formulation Our hybrid parallel formulation has elements of both schemes. The Synchronous Tree Construction Approach in Section 3.1 incurs high communication overhead as the frontier gets larger. The Partitioned Tree Construction Approach of Section 3.2 incurs cost of load balancing after each step. The hybrid scheme keeps continuing with the first approach as long as the communication cost incurred by the first formulation is not too high. Once this cost becomes high, the processors as well as the current frontier of the classification tree are partitioned into two parts. Our description assumes that the number of processors is a power of 2, and that these processors are connected in a hypercube configuration. The algorithm can be appropriately modified if P is not a power of 2. Also this algorithm can be mapped on to any parallel architecture by simply embedding a virtual hypercube in the architecture. More precisely the hybrid formulation works as follows. ffl The database of training cases is split equally among P processors. Thus, if N is the total number of training cases, each processor has N=P training cases locally. At the beginning, all processors are assigned to one partition. The root node of the classification tree is allocated to the partition. ffl All the nodes at the frontier of the tree that belong to one partition are processed together using the synchronous tree construction approach of Section 3.1. ffl As the depth of the tree within a partition increases, the volume of statistics gathered at each level also increases as discussed in Section 3.1. At some point, a level is reached when communication cost become prohibitive. At this point, the processors in the partition are divided into two partitions, and the current set of frontier nodes are split and allocated to these partitions in such a way that the number of training cases in each partition is roughly equal. This load balancing is done as described as follows: - On a hypercube, each of the two partitions naturally correspond to a sub- cube. First, corresponding processors within the two sub-cubes exchange relevant training cases to be transferred to the other sub-cube. After this exchange, processors within each sub-cube collectively have all the training cases for their partition, but the number of training cases at each processor can vary between 0 to 2\LambdaN . Now, a load balancing step is done within each sub-cube so that each processor has an equal number of data items. ffl Now, further processing within each partition proceeds asynchronously. The above steps are now repeated in each one of these partitions for the particular subtrees. This process is repeated until a complete classification tree is grown. ffl If a group of processors in a partition become idle, then this partition joins up with any other partition that has work and has the same number of processors. This can be done by simply giving half of the training cases located at each processor in the donor partition to a processor in the receiving partition. Computation Frontier at depth 3 Figure 4. The computation frontier during computation phase A key element of the algorithm is the criterion that triggers the partitioning of the current set of processors (and the corresponding frontier of the classification tree If partitioning is done too frequently, then the hybrid scheme will approximate the partitioned tree construction approach, and thus will incur too much data movement cost. If the partitioning is done too late, then it will suffer from high cost for communicating statistics generated for each node of the frontier, like the synchronized tree construction approach. One possibility is to do splitting when the accumulated cost of communication becomes equal to the cost of moving records around in the splitting phase. More precisely, splitting is done when Partition 1 Partition 2 Figure 5. Binary partitioning of the tree to reduce communication costs (Communication MovingCost LoadBalancing As an example of the hybrid algorithm, Figure 4 shows a classification tree frontier at depth 3. So far, no partitioning has been done and all processors are working cooperatively on each node of the frontier. At the next frontier at depth 4, partitioning is triggered, and the nodes and processors are partitioned into two partitions as shown in Figure 5. A detailed analysis of the hybrid algorithm is presented in Section 4. 3.4. Handling Continuous Attributes Note that handling continuous attributes requires sorting. If each processor contains N=P training cases, then one approach for handling continuous attributes is to perform a parallel sorting step for each such attribute at each node of the decision tree being constructed. Once this parallel sorting is completed, each processor can compute the best local value for the split, and then a simple global communication among all processors can determine the globally best splitting value. However, the step of parallel sorting would require substantial data exchange among processors. The exchange of this information is of similar nature as the exchange of class distribution information, except that it is of much higher volume. Hence even in this case, it will be useful to use a scheme similar to the hybrid approach discussed in Section 3.3. A more efficient way of handling continuous attributes without incurring the high cost of repeated sorting is to use the pre-sorting technique used in algorithms SLIQ [18], SPRINT [21], and ScalParC [13]. These algorithms require only one pre-sorting step, but need to construct a hash table at each level of the classification tree. In the parallel formulations of these algorithms, the content of this hash table needs to be available globally, requiring communication among processors. definition Total number of training samples Total Number of processors Number of processors cooperatively working on tree expansion A d Number of categorical attributes C Number of classes Average number of distinct values in the discrete attributes Present level of a decision tree t s Start up time of communication latency [KGGK94] t w Per-word transfer time of communication latency [KGGK94] Table 4. Symbols used in the analysis. Existing parallel formulations of these schemes [21, 13] perform communication that is similar in nature to that of our synchronous tree construction approach discussed in Section 3.1. Once again, communication in these formulations [21, 13] can be reduced using the hybrid scheme of Section 3.3. Another completely different way of handling continuous attributes is to discretize them once as a preprocessing step [12]. In this case, the parallel formulations as presented in the previous subsections are directly applicable without any modification Another approach towards discretization is to discretize at every node in the tree. There are two examples of this approach. The first example can be found in [3] where quantiles [2] are used to discretize continuous attributes. The second example of this approach to discretize at each node is SPEC [23] where a clustering technique is used. SPEC has been shown to be very efficient in terms of runtime and has also been shown to perform essentially identical to several other widely used tree classifiers in terms of classification accuracy [23]. Parallelization of the discretization at every node of the tree is similar in nature to the parallelization of the computation of entropy gain for discrete attributes, because both of these methods of discretization require some global communication among all the processors that are responsible for a node. In particular, parallel formulations of the clustering step in SPEC is essentially identical to the parallel formulations for the discrete case discussed in the previous subsections [23]. 4. Analysis of the Hybrid Algorithm In this section, we provide the analysis of the hybrid algorithm proposed in Section 3.3. Here we give a detailed analysis for the case when only discrete attributes are present. The analysis for the case with continuous attributes can be found in [23]. The detailed study of the communication patterns used in this analysis can be found in [16]. Table 4 describes the symbols used in this section. 4.1. Assumptions ffl The processors are connected in a hypercube topology. Complexity measures for other topologies can be easily derived by using the communication complexity expressions for other topologies given in [16]. ffl The expression for communication and computation are written for a full binary tree with 2 L leaves at depth L. The expressions can be suitably modified when the tree is not a full binary tree without affecting the scalability of the algorithm. ffl The size of the classification tree is asymptotically independent of N for a particular data set. We assume that a tree represents all the knowledge that can be extracted from a particular training data set and any increase in the training set size beyond a point does not lead to a larger decision tree. 4.2. Computation and Communication Cost For each leaf of a level, there are A d class histogram tables that need to be com- municated. The size of each of these tables is the product of number of classes and the mean number of attribute values. Thus size of class histogram table at each processor for each leaf is: Class histogram size for each The number of leaves at level L is 2 L . Thus the total size of the tables is: Combined class histogram tables for a processor = C A d M 2 L At level L, the local computation cost involves I/O scan of the training set, initialization and update of all the class histogram tables for each attribute: Local Computation cost = '( A d N where t c is the unit of computation cost. At the end of local computation at each processor, a synchronization involves a global reduction of class histogram values. The communication cost 1 is : Per level Communication cost (2) When a processor partition is split into two, each leaf is assigned to one of the partitions in such a way that number of training data items in the two partitions is approximately the same. In order for the two partitions to work independently of each other, the training set has to be moved around so that all training cases for a leaf are in the assigned processor partition. For a load balanced system, each processor in a partition must have N training data items. This movement is done in two steps. First, each processor in the first partition sends the relevant training data items to the corresponding processor in the second partition. This is referred to as the "moving" phase. Each processor can send or receive a maximum of N data to the corresponding processor in the other partition. Cost for moving phase 2 After this, an internal load balancing phase inside a partition takes place so that every processor has an equal number of training data items. After the moving phase and before the load balancing phase starts, each processor has training data item count varying from 0 to 2\LambdaN . Each processor can send or receive a maximum of N training data items. Assuming no congestion in the interconnection network, cost for load balancing is: Cost for load balancing phase 2 A detailed derivation of Equation 4 above is given in [23]. Also, the cost for load balancing assumes that there is no network congestion. This is a reasonable assumption for networks that are bandwidth-rich as is the case with most commercial systems. Without assuming anything about network congestion, load balancing phase can be done using transportation primitive [22] in time 2 N Splitting is done when the accumulated cost of communication becomes equal to the cost of moving records around in the splitting phase [14]. So splitting is done (Communication Cost) Moving Cost Balancing This criterion for splitting ensures that the communication cost for this scheme will be within twice the communication cost for an optimal scheme [14]. The splitting is recursive and is applied as many times as required. Once splitting is done, the above computations are applied to each partition. When a partition of processors starts to idle, then it sends a request to a busy partition about its idle state. This request is sent to a partition of processors of roughly the same size as the idle partition. During the next round of splitting the idle partition is included as a part of the busy partition and the computation proceeds as described above. 4.3. Scalability Analysis Isoefficiency metric has been found to be a very useful metric of scalability for a large number of problems on a large class of commercial parallel computers [16]. It is defined as follows. Let P be the number of processors and W the problem size (in total time taken for the best sequential algorithm). If W needs to grow as to maintain an efficiency E, then fE (P ) is defined to be the isoefficiency function for efficiency E and the plot of fE (P ) with respect to P is defined to be the isoefficiency curve for efficiency E. We assume that the data to be classified has a tree of depth L 1 . This depth remains constant irrespective of the size of data since the data "fits" this particular classification tree. Total cost for creating new processor sub-partitions is the product of total number of partition splits and cost for each partition split (='( N using Equations 3 and 4. The number of partition splits that a processor participates in is less than or equal to L 1 - the depth of the tree. Cost for creating new processors partitions L 1 '( Communication cost at each level is given by Equation 2 (= '(log P )). The combined communication cost is the product of the number of levels and the communication cost at each level. Combined communication cost for processing attributes L 1 '(log P The total communication cost is the sum of cost for creating new processor partitions and communication cost for processing class histogram tables, the sum of Equations 5 and 6. Total Communication cost Computation cost given by Equation 1 is: Total computation Total parallel run time (Sum of Equations 7 and 8)= Communication time Computation time. Parallel run In the serial case, the whole dataset is scanned once for each level. So the serial time is To get the isoefficiency function, we equate P times total parallel run time using Equation 9 to serial computation time. Therefore, the isoefficiency function is Isoefficiency is '(P log P ) assuming no network congestion during load balancing phase. When the transportation primitive is used for load balancing, the isoefficiency is O(P 3 ). 5. Experimental Results We have implemented the three parallel formulations using the MPI programming library. We use binary splitting at each decision tree node and grow the tree in breadth first manner. For generating large datasets, we have used the widely used synthetic dataset proposed in the SLIQ paper [18] for all our experiments. Ten classification functions were also proposed in [18] for these datasets. We have used the function 2 dataset for our algorithms. In this dataset, there are two class labels and each record consists of 9 attributes having 3 categoric and 6 continuous attributes. The same dataset was also used by the SPRINT algorithm [21] for evaluating its performance. Experiments were done on an IBM SP2. The results for comparing speedup of the three parallel formulations are reported for parallel runs on 1, 2, 4, 8, and 16 processors. More experiments for the hybrid approach are reported for up to 128 processors. Each processor has a clock speed of 66.7 MHz with 256 MB real memory. The operating system is AIX version 4 and the processors communicate through a high performance switch (hps). In our implementation, we keep the "attribute lists" on disk and use the memory only for storing program specific data structures, the class histograms and the clustering structures. First, we present results of our schemes in the context of discrete attributes only. We compare the performance of the three parallel formulations on up to 16 processor IBM SP2. For these results, we discretized 6 continuous attributes uniformly. Specifically, we discretized the continuous attribute salary to have 13, commission to have 14, age to have 6, hvalue to have 11, hyears to have 10, and loan to have equal intervals. For measuring the speedups, we worked with different sized datasets of 0.8 million training cases and 1.6 million training cases. We increased the processors from 1 to 16. The results in Figure 6 show the speedup comparison of the three parallel algorithms proposed in this paper. The graph on the left shows the speedup with 0.8 million examples in the training set and the other graph shows the speedup with 1.6 million examples. The results show that the synchronous tree construction approach has a good speedup for 2 processors, but it has a very poor speedup for 4 or more processors. There are two reasons for this. First, the synchronous tree construction approach incurs high communication cost, while processing lower levels of the tree. Second, a synchronization has to be done among different processors as soon as their communication buffer fills up. The communication buffer has the histograms of all the discrete variables for each node. Thus, the contribution of each node is independent of its tuples count, the tuple count at a node being proportional to the computation to process that node. While processing lower levels of the tree, this synchronization is done many times at each level (after every 100 nodes for our experiments). The distribution of tuples for each decision tree node becomes quite different lower down in the tree. Therefore, the processors wait for each other during synchronization, and thus, contribute to poor speedups. The partitioned tree construction approach has a better speedup than the synchronous tree construction approach. However, its efficiency decreases as the number of processors increases to 8 and 16. The partitioned tree construction approach Processors Speedup for 0.8 million examples, partitioned -", hybrid -o-", synchronous -x-" Processors Speedup for 1.6 million examples, partitioned -", hybrid -o-", synchronous -x-" Figure 6. Speedup comparison of the three parallel algorithms. suffers from load imbalance. Even though nodes are partitioned so that each processor gets equal number of tuples, there is no simple way of predicting the size of the subtree for that particular node. This load imbalance leads to the runtime being determined by the most heavily loaded processor. The partitioned tree construction approach also suffers from the high data movement during each partitioning phase, the partitioning phase taking place at higher levels of the tree. As more processors are involved, it takes longer to reach the point where all the processors work on their local data only. We have observed in our experiments that load imbalance and higher communication, in that order, are the major cause for the poor performance of the partitioned tree construction approach as the number of processors increase. The hybrid approach has a superior speedup compared to the partitioned tree approach as its speedup keeps increasing with increasing number of processors. As discussed in Section 3.3 and analyzed in Section 4, the hybrid controls the communication cost and data movement cost by adopting the advantages of the two basic parallel formulations. The hybrid strategy also waits long enough for splitting, until there are large number of decision tree nodes for splitting among processors. Due to the allocation of decision tree nodes to each processor being randomized to a large extent, good load balancing is possible. The results confirmed that the proposed hybrid approach based on these two basic parallel formulations is effective. We have also performed experiments to verify our splitting criterion of the hybrid algorithm is correct. Figure 7 shows the runtime of the hybrid algorithm with different ratio of communication cost and the sum of moving cost and load balancing cost, i.e., (Communication Cost) Moving Cost log2(Splitting Criteria Ratio), x=0 -> ratio=1 Runtimes Runtimes for splitting at different values of ratio, 8 processors, 0.8 million examples 10100300500log2(Splitting Criteria Ratio), x=0 -> ratio=1 Runtimes Runtimes for splitting at different values of ratio, 16 processors, 1.6 million examples Figure 7. Splitting criterion verification in the hybrid algorithm. The graph on the left shows the result with 0.8 million examples on 8 processors and the other graph shows the result with 1.6 million examples on processors. We proposed that splitting when this ratio is 1.0 would be the optimal time. The results verified our hypothesis as the runtime is the lowest when the ratio is around 1.0. The graph on the right with 1.6 million examples shows more clearly why the splitting choice is critical for obtaining a good performance. As the splitting decision is made farther away from the optimal point proposed, the runtime increases significantly. The experiments on 16 processors clearly demonstrated that the hybrid approach gives a much better performance and the splitting criterion used in the hybrid approach is close to optimal. We then performed experiments of running the hybrid approach on more number of processors with different sized datasets to study the speedup and scalability. For these experiments, we used the original data set with continuous attributes and used a clustering technique to discretize continuous attributes at each decision tree node [23]. Note that the parallel formulation gives almost identical performance as the serial algorithm in terms of accuracy and classification tree size [23]. The results in Figure 8 show the speedup of the hybrid approach. The results confirm that the hybrid approach is indeed very effective. To study the scaleup behavior, we kept the dataset size at each processor constant at 50,000 examples and increased the number of processors. Figure 9 shows the runtime on increasing number of processors. This curve is very close to the ideal case of a horizontal line. The deviation from the ideal case is due to the fact that the isoefficiency function is O(P log P ) not O(P ). Current experimental data is consistent with the derived isoefficiency function but we intend to conduct additional validation experiments. Number of processors Speedup curves for different sized datasets million examples 1.6 million examples 3.2 million examples 6.4 million examples 12.8 million examples 25.6 million examples Figure 8. Speedup of the hybrid approach with different size datasets. Number of Processors Total Run Time Runtimes of our algorithm for 50K examples at each processor Figure 9. Scaleup of our algorithm 6. Concluding Remarks In this paper, we proposed three parallel formulations of inductive-classification learning algorithm. The Synchronous Tree Construction Approach performs well if the classification tree remains skinny, having few nodes at any level, throughout. For such trees, there are relatively large number of training cases at the nodes at any level; and thus the communication overhead is relatively small. Load imbalance is avoided by processing all nodes at a level, before synchronization among the processors. However, as the tree becomes bushy, having a large number of nodes at a level, the number of training data items at each node decrease. Frequent synchronization is done due to limited communication buffer size, which forces communication after processing a fixed number of nodes. These nodes at lower depths of the tree, which have few tuples assigned to them, may have highly variable distribution of tuples over the processors, leading to load imbalance. Hence, this approach suffers from high communication overhead and load imbalance for bushy trees. The Partitioned Tree Construction Approach works better than Synchronous Tree Construction Approach if the tree is bushy. But this approach pays a big communication overhead in the higher levels of the tree as it has to shuffle lots of training data items to different processors. Once every node is solely assigned to a single processor, each processor can construct the partial classification tree independently without any communication with other processors. However, the load imbalance problem is still present after the shuffling of the training data items, since the partitioning of the data was done statically. The hybrid approach combines the good features of these two approaches to reduce communication overhead and load imbalance. This approach uses the Synchronous Tree Construction Approach for the upper parts of the classification tree. Since there are few nodes and relatively large number of the training cases associated with the nodes in the upper part of the tree, the communication overhead is small. As soon as the accumulated communication overhead is greater than the cost of partitioning of data and load balancing, this approach shifts to the Partitioned Tree Construction Approach incrementally. The partitioning takes place when a reasonable number of nodes are present at a level. This partitioning is gradual and performs randomized allocation of classification tree nodes, resulting in a better load balance. Any load imbalance at the lower levels of the tree, when a processor group has finished processing its assigned subtree, is handled by allowing an idle processor group to join busy processor groups. The size and shape of the classification tree varies a lot depending on the application domain and training data set. Some classification trees might be shallow and the others might be deep. Some classification trees could be skinny others could be bushy. Some classification trees might be uniform in depth while other trees might be skewed in one part of the tree. The hybrid approach adapts well to all types of classification trees. If the decision tree is skinny, the hybrid approach will just stay with the Synchronous Tree Construction Approach. On the other hand, it will shift to the Partitioned Tree Construction Approach as soon as the tree becomes bushy. If the tree has a big variance in depth, the hybrid approach will perform dynamic load balancing with processor groups to reduce processor idling. Acknowledgments A significant part of this work was done while Anurag Srivastava and Vineet Singh were at IBM TJ Watson Research Center. This work was supported by NSF grant ASC-9634719, Army Research Office contract DA/DAAH04-95-1-0538, Cray Re-search Inc. Fellowship, and IBM partnership award, the content of which does not necessarily reflect the policy of the government, and no official endorsement should be inferred. Access to computing facilities was provided by AHPCRC, Minnesota Supercomputer Institute, Cray Research Inc., and NSF grant CDA-9414015. Notes 1. If the message size is large, by routing message in parts, this communication step can be done constant k0 . Refer to [16] section 3.7 for details. --R Database mining: A performance perspective. CLOUDS: Classification for large or out-of-core datasets Megainduction: Machine Learning on Very Large Databases. Experiments on multistrategy learning by met- alearning Metalearning for multistrategy learning and parallel learning. Concatenated parallelism: A technique for efficient parallel divide and conquer. Genetic Algorithms in Search Use of contextual information for feature ranking and discretization. A new scalable and efficient parallel classification algorithm for mining large datasets. Unstructured tree search on simd parallel computers. Decision trees on parallel processors. Introduction to Parallel Computing: Algorithm Design and Analysis. An introduction to computing with neural nets. SLIQ: A fast scalable classifier for data mining. A coarse grained parallel induction heuristic. SPRINT: A scalable parallel classifier for data mining. Experiments on the costs and benefits of windowing in ID3. --TR --CTR Amir Bar-Or , Daniel Keren , Assaf Schuster , Ran Wolff, Hierarchical Decision Tree Induction in Distributed Genomic Databases, IEEE Transactions on Knowledge and Data Engineering, v.17 n.8, p.1138-1151, August 2005 Robert Grossman , Yike Guo, Data mining tasks and methods: parallel methods for scaling data mining algorithms to large data sets, Handbook of data mining and knowledge discovery, Oxford University Press, Inc., New York, NY, 2002 Doina Caragea , Adrian Silvescu , Vasant G. Honavar, Analysis and synthesis of agents that learn from distributed dynamic data sources, Emergent neural computational architectures based on neuroscience: towards neuroscience-inspired computing, Springer-Verlag New York, Inc., New York, NY, 2001 Olcay Taner Yldz , Onur Dikmen, Parallel univariate decision trees, Pattern Recognition Letters, v.28 n.7, p.825-832, May, 2007 Vipin Kumar , Mohammed Zaki, High performance data mining (tutorial PM-3), Tutorial notes of the sixth ACM SIGKDD international conference on Knowledge discovery and data mining, p.309-425, August 20-23, 2000, Boston, Massachusetts, United States Raymond T. Ng , Alan Wagner , Yu Yin, Iceberg-cube computation with PC clusters, ACM SIGMOD Record, v.30 n.2, p.25-36, June 2001 S. Ruggieri, Efficient C4.5, IEEE Transactions on Knowledge and Data Engineering, v.14 n.2, p.438-444, March 2002 Massimo Coppola , Marco Vanneschi, Parallel and distributed data mining through parallel skeletons and distributed objects, Data mining: opportunities and challenges, Idea Group Publishing, Hershey, PA, Aleksandar Lazarevic , Zoran Obradovic, Boosting Algorithms for Parallel and Distributed Learning, Distributed and Parallel Databases, v.11 n.2, p.203-229, March 2002 Vipin Kumar , Mohammed Zaki, High performance data mining (tutorial PM-3), Tutorial notes of the sixth ACM SIGKDD international conference on Knowledge discovery and data mining, p.309-425, August 20-23, 2000, Boston, Massachusetts, United States Massimo Coppola , Marco Vanneschi, High-performance data mining with skeleton-based structured parallel programming, Parallel Computing, v.28 n.5, p.793-813, May 2002 Jack Dongarra , Ian Foster , Geoffrey Fox , William Gropp , Ken Kennedy , Linda Torczon , Andy White, References, Sourcebook of parallel computing, Morgan Kaufmann Publishers Inc., San Francisco, CA,
data mining;classification;scalability;decision trees;parallel processing
593486
Effect of Data Distribution in Parallel Mining of Associations.
Association rule mining is an important new problem in data mining. It has crucial applications in decision support and marketing strategy. We proposed an efficient parallel algorithm for mining association rules on a distributed share-nothing parallel system. Its efficiency is attributed to the incorporation of two powerful candidate set pruning techniques. The two techniques, distributed and global prunings, are sensitive to two data distribution characteristics: data skewness and workload balance. The prunings are very effective when both the skewness and balance are high. We have implemented FPM on an IBM SP2 parallel system. The performance studies show that FPM outperforms CD consistently, which is a parallel version of the representative Apriori algorithm (Agrawal and Srikant, 1994). Also, the results have validated our observation on the effectiveness of the two pruning techniques with respect to the data distribution characteristics. Furthermore, it shows that FPM has nice scalability and parallelism, which can be tuned for different business applications.
Introduction Association rule discovery has attracted a lot of attention from the research and business communities [1, 2, 4]. An association rule is a rule which implies certain association relationships among a set of objects (such as "occur together" or "one implies the other") in a database. The intuitive meaning of an association Y are set of items, is that transactions of the database which contain X tend to contain Y . A classical example is that 98% of customers that purchase tires and automobile accessories in a department store also have automotive services carried out. This example is a typical association in a basket database which sounds like common sense knowledge; however, there could be a lot of associations among the data which may not be able to deduce from common knowledge. Therefore, efficient automated technique to discover this type of rules is a very important area of research in data mining [2, 7, 8, 15, 17]. Applications of association rule mining range from decision support to product marketing and consumer behavior prediction. Previous studies examined efficient mining of association rules from many different angles. An influential association rule mining algorithm, Apriori [2], has been developed for rule mining in large transaction databases. The scope of the study has also been extended to efficient mining of sequential patterns [20], generalized association rules [19], multiple-level association rules [9], quantitative association rules [21], constrainted association rules [14] etc. Although these studies are on sequential data mining techniques, algorithms for parallel mining of association rules have also been proposed [3, 16, 22, 23]. The development of parallel systems for mining of association rules has its unique importance - databases or data warehouses [18] have been used more often to store a huge amount of data; data mining in such databases require substantial processing power, and parallel system is a possible solution. This observation motivates us to study efficient parallel algorithms for mining association rules in large databases. In this work, we study the problem on parallel system with distributed share-nothing memory such as the IBM SP2 [11]. In this model, the database is partitioned and distributed across the local disks of the processors; and the processors communicate via a fast network. It has been well studied that the major cost of mining association rules is the computation of the set of large itemsets (i.e., frequently occurring sets of items, see Section 2.1) in the database [1, 2]. An itemset (a set of items) is large if the percentage of transactions that containing all these items is greater than a given threshold. The most representative parallel algorithm for mining association rules is the CD algorithm (Count Distribution), which is designed for share-nothing parallel systems [3]. It extends directly the basic technique of Aprori to parallel system. Our proposed algorithm, FPM (Fast Parallel Mining), has the following distinct feature in comparison with CD : FPM has explored an important property between locally large itemsets (those that are large with respect to the partition of a processor) and globally large itemsets (those that are large with respect to the entire database) to develop two powerful pruning techniques, distributed pruning and global pruning, which can reduce the number of candidate sets at each individual processor. Since the number of candidate sets is a dominant parameter of the computation cost, with a substantially smaller candidate sets, FPM performs much better than CD. Another contribution of this work is the discovery that the effectiveness of the two aforementioned pruning techniques, and hence the performance of the parallel mining, depends on the data distribution characteristics in the database partitioning. We have captured the distribution characteristics in two factors : data skewness and workload balance. These two factors are orthogonal properties. Intuitively, a partitioned database has high data skewness if most globally large itemsets are locally large only at a very few partitions. On the other hand, a partitioned database has a high workload balance if all the processors have similar number of locally large itemsets in their partitions. (More precise definitions of skewness and workload balance will be given in Sections 3 & 4.) We have defined metrics to measure data skewness and workload balance. We found out that both the distributed and global prunings have super performance in the best case of high data skewness and high workload balance. The combination of high balance with moderate skewness is the second best case. On the other hand, the high skewness, moderate balance combination only provide moderate improvement over CD, while the combination of low skewness and low balance is the worst case in which only marginal improvement can be found. We have implemented FPM on an IBM SP2 parallel machine with processors. Extensive performance studies have been carried out. The results confirm our observation on the relationship between pruning effectiveness and data distribution. The rest of this paper is organized as follows. Section 2 overviews the parallel mining of association rules. The techniques of distributed and global prunings, together with the FPM algorithm are described in Section 3. In the same section, we have also investigated the relationship between the effectiveness of the prunings and the two data distribution characteristics of data skewness and workload balance. In section 4, we define metrics to measure data skewness and workload balance of a data partition. Section 5 reports the result of an extensive performance study. In Section 6, we discuss a few issues including possible extensions of FPM to enhance its scalability. Section 7 is the conclusion. 2 Parallel Mining of Association Rules 2.1 Sequential Algorithm for Mining Association Rules be a set of items. Let D be a database of transactions, where each transaction T consists of a set of items such that T ' I . Given an itemset X ' I , a transaction T contains X if and only if X ' T . An association rule is an implication of the form X [2]. The association rule holds in D with confidence c if the probability of a transaction in D which contains X also contains Y is c. The association rule X ) Y has support s in D if the probability of a transaction in D contains both X and Y is s. The task of mining association rules is to find all the association rules whose support is larger than a minimum support threshold and whose confidence is larger than a minimum confidence threshold. For an itemset X , its support is the percentage of transactions in D which contains X , and its support count, denoted by X :sup , is the number of transactions in D containing X . An itemset X is large (or more precisely, frequently occurring) if its support is no less than the minimum support threshold. An itemset of size k is called a k-itemset. It has been shown that the problem of mining association rules can be reduced to two subproblems [1]: (1) find all large itemsets for a given minimum support threshold, and (2) generate the association rules from the large itemsets found. Since (1) dominates the overall cost of mining association rules, the research has been focused on developing efficient methods to solve the first subproblem [2]. An interesting serial algorithm, Apriori [2], has been proposed for computing large itemsets at mining association rules in a transaction database, which is outlined as follows [2]. The large itemsets are computed through iterations. At each iteration, Apriori scans the database once and finds all the large itemsets of the same size. At the k-th iteration, Apriori creates the set of candidate sets C (k) by applying the candidate set generating function Apriori gen on L (k\Gamma1) , where L (k\Gamma1) is the set of all large 1)-itemsets found at the (k \Gamma 1)-st iteration, and Apriori gen generates only those k-itemset candidates whose every 1)-itemset subset is in L (k\Gamma1) . 2.2 Count Distribution Algorithm for Parallel Mining CD (Count Distribution) is a parallel version of Apriori. It is one of the earliest proposed and representative parallel algorithms for mining of association rules [3]. We describe here briefly its steps for comparison purpose. The database D is partitioned into distributed across n processors. The program fragment of CD at processor k-th iteration is outlined in Figure 1. (For convenience, we use X :sup(i) to represent the local support count of an itemset X in partition D i .) In step 1, every processor computes the same candidate set C k by applying the aprior gen function on L k\Gamma1 , which is the set of large itemsets found at the (k \Gamma 1)-th iteration. In step 2, local support counts (support in D i ) of candidates in C k are found. In steps 3 & 4, local support counts are exchanged with all other processors to get global support counts (support in D) and globally large itemsets (large with respect to D) L k are computed independently by each processor. CD repeats steps 1 - 4 until no more candidate is found. We have implemented CD on an IBM SP2 using the MPI (Message Passing Interface) [13]. scan partition D i to find the local support count X :sup(i) for all with all other processors to get global support counts X :sup , for all Figure 1: Count Distribution Algorithm Pruning Techniques and the FPM Algorithm 3.1 Distributed Pruning It is important to observe some interesting properties related to large itemsets in a parallel environments since such properties may substantially reduce the number of candidate sets. (The preliminary form of the results in this section have been developed in [6] and extended here.) First, there is an important relationship between large itemsets and the processors in the database: every globally large itemsets must be locally large at some processor(s). If an itemset X is both globally large and locally large at a processor p i , X is called gl-large at processor p i . The set of gl-large itemsets at a processor will form a basis for the processor to generate its own candidate sets. Second, a gl-large itemset at a processor has the following monotonic subset relationship property: if an itemset is gl-large at a processor p i , all of its subsets are also gl-large at p i . Combining these two properties, we have the following results. Lemma 1 If an itemset X is globally large, there exists a processor n), such that X and all its subsets are gl-large at processor p i . 1 We use GL i to denote the set of gl-large itemsets at processor p i , and GL i(k) to denote the set of gl-large k-itemsets at processor p i . It follows from Lemma 1 that if X 2 L (k) , then there exists a processor p i , such that all its are gl-large at processor p i , i.e., they belong to GL i(k\Gamma1) . In a straightforward adaptation of Apriori, the set of candidate sets at the k-th iteration, denoted by CA (k) , which stands for size-k candidate sets from Apriori, would be generated by applying the Apriori gen function on L (k\Gamma1) . That is, At each processor p i , let CG i(k) be the set of candidates sets generated by applying Apriori gen on GL i(k\Gamma1) , i.e., where CG stands for candidate sets generated from gl-large itemsets. Hence CG i(k) is generated from GL i(k\Gamma1) . Since GL i(k\Gamma1) ' L (k\Gamma1) , CG i(k) is a subset of CA (k) . In the following, we use CG (k) to denote the set [ n Theorem 1 For every k ? 1, the set of all large k-itemsets L (k) is a subset of CG This result is stronger than that in [17] - the result there states that a globally large itemset is locally large in some partition; while states that all its subsets must be locally large together at the same partition. Proof. Let It follows from Lemma 1 that there exists a processor n), such that all the subsets of X are gl-large at processor p i . Hence X 2 CG i(k) . Therefore, indicates that CG (k) , which is a subset of CA (k) and could be much smaller than CA (k) , can be taken as the set of candidate sets for the size-k large itemsets. In effect, the set of candidates in CA (k) has been pruned down to those in CG (k) - we called this technique distributed pruning. This result forms a basis for the reduction of the set of candidate sets in the algorithm FPM. First the set of candidate sets CG i(k) can be generated locally at each processor p i at the k-th iteration. After the exchange of support counts, the gl-large itemsets GL i(k) in CG i(k) can be found at the end of that iteration. Based on GL i(k) , the candidate sets at processor p i for the 1)-st iteration can then be generated according to Theorem 1. According to our performance studies, the number of candidate sets generated by distributed pruning can be substantially reduced to about of that generated in CD. Example 1 illustrates the effectiveness of the reduction of candidate sets using distributed pruning. Example 1 Assuming there are 3 processors in a parallel system in which the database D has been partitioned into Suppose the set of large 1-itemsets (computed at the first iteration) L G; Hg, in which A; B, and C are locally large at processor p 1 , B; C, and D are locally large at processor p 2 , and E; F; G, and H are locally large at processor p 3 . Therefore, GL G; Hg. Based on Theorem 1, the set of size-2 candidate sets at processor p 1 is CG fAB;BC;ACg. Similarly, CG EG;EH;FG;FH;GHg. Hence, the set of candidate sets for large 2-itemsets is CG candidates. However, if Apriori gen is applied to L (1) , the set of candidate sets CA would have 28 candidates. This shows that it is very effective to apply distributed pruning to reduce the candidate sets. 2 3.2 Global Pruning As a result of the count exchange, the local support counts X :sup(i) , for all processor are also available at every processor. With this information, another powerful pruning technique called global pruning can be developed. Let X be a candidate k-itemset. At each partition D i , X :sup(i) - Y :sup(i) , if Y ae X . Therefore the local support count of X , X :sup(i) , is bounded by the value minfY 1g. Since the global support count of X , X :sup , is the sum of its local support count at all the processors, the value is an upper bound of X :sup . If X :maxsup ! minsup \Theta jDj, then X can be pruned away. This technique is called global pruning. Note that global pruning requires no additional information except the local support counts resulted from count exchange in the previous iteration. Table 1 gives an example to show that global pruning can pruning away candidates which cannot be pruned by distributed pruning. Suppose the global support count threshold is 15 and the local support count threshold at each local support at processor 1 13 local support at processor 2 3 3 12 34 1 4 local support at processor 3 global support gl-large at processor 1 \Theta \Theta gl-large at processor 2 \Theta \Theta gl-large at processor 3 \Theta \Theta \Theta \Theta Table 1: High data skewness and high workload balance case processor is 5. Distributed pruning cannot prune away CD, as C and D are both gl-large at processor 2. Whereas global pruning can prune away CD, as CD can also be pruned, because EF would survive global pruning. From this example, it is clear that global pruning is more effective than distributed pruning, i.e., what can pruned away by distributed pruning will be pruned away by global pruning. The three pruning techniques, the one in apriori gen, the distributed and global prunings, have increasing pruning power, and the latter ones subsume the previous one. 3.3 Fast Parallel Mining Algorithm (FPM) We present the FPM algorithm in this section. It is an enhancement of CD. The simple support counts exchange scheme in CD is retained in FPM. The main difference is the incorporation of both the distributed and global prunings in FPM to reduce the candidate set size. The first iteration of FPM is the same as CD. Each processor scans its partition to find out local support counts of all size-1 itemsets and use one round of count exchange to compute the global support counts. At the end of the 1-st iteration, in addition to L 1 , each processor also finds out the gl-large itemsets GL 1(i) , for For the k-th iteration of FPM, k ? 1, the program fragment executed at processor i, 1 - i - n, is described in Figure 2. compute candidate sets CG (distributed pruning) prune candidates in CG k by global pruning; scan partition D i to find the local support count X :sup(i) for all remaining candidates with all other processors to get global support counts X :sup , for all return Figure 2: The FPM Algorithm Similar to CD, FPM is also implemented by collective communication operations of MPI on the SP2. In order to compare the effects of distributed and global pruning, we have also implemented a variant FNG (FPM with no global pruning) of FPM. FNG does not perform the global pruning, i.e., it's procedure is the same as that in Figure 2, except step 2 is removed. 4 Data Skewness and Workload Balance In a database partition, two data distribution characteristics, data skewness and workload balance, have orthogonal effects on prunings and hence performance of FPM. Intuitively, the data skewness of a partitioned database is high if the supports of most large itemsets are clustered in a few partitions. It is low if the supports of most large itemsets are distributed evenly across the processors. In Table 1, it is clear that all the itemsets have high skewness. For a partition with high skewness, even though the support of each large itemset is clustered at a few processors, the clusterings of different large itemsets may be distributed evenly across the processors or concentrated on a few of them. In the first case, the clusterings of the large itemsets are distributed evenly among the processors; hence, each processor would have similar number of locally large itemsets. We characterise this case as high workload balance. In the second case, the clusterings would be concentrated on a few processors; hence some processors would have much more locally large itemsets than the others. This is the low workload balance case. For example, the itemsets in Table 1 not only have high skewness, it also has a good workload balance; because A, B are locally large at processor 1, and C, D at processor 2, whereas E, F at processor 3. It follows from our discussion of the pruning techniques that high data skewness would increase the chance of candidate sets pruning; however, it is not the only factor, workload is another critical factor. In the following, we will see that given a good data skewness, if the distribution of the clusterings amount the processors are not even, then the pruning effects would be reduced significantly, and, to aggravate the problem more, the work of computing the large itemsets would be concentrated on a few processors which is a very troublesome issue for parallel computation. Example 2 As explained above, Table 1 is a case of high data skewness and high workload balance. The supports of each itemset are distributed mostly in one partition; hence, the skewness is high. On the other hand, every partition has the same number of locally large itemsets; therefore, the workload balance is also high. In this case, CD will generate candidates in the second iteration. Whereas, the distributed pruning will generate only three candidates AB, CD and EF , which shows that the pruning has good effect for this distribution. local support at processor 1 13 33 12 34 2 1 local support at processor 2 local support at processor 3 global support gl-large at processor 1 gl-large at processor 2 \Theta \Theta \Theta \Theta \Theta \Theta gl-large at processor 3 \Theta \Theta \Theta \Theta Table 2: High data skewness and low workload balance case Table 2 is an example of high data skewness but low workload balance. The thresholds are the same as that in Table 1, i.e., the global support threshold is 15 and the local support threshold at each processor is 5. The support count distribution of each item is the same as that in Table 1 except that items A; B; C and D are now locally large together at processor 1 instead of distributed between processors 1 and 2. In this lower workload balance case, distributed pruning will generate 7 size-2 candidates, namely AB, AC, AD, BC, BD, CD and EF , while CD will still have 15 candidates. Thus, the distributed pruning remains to be very effective, but not as good as that in the high workload balance case (Table 1). local support at processor 1 6 12 4 13 5 12 local support at processor 2 6 12 5 12 4 13 local support at processor 3 global support gl-large at processor 1 \Theta gl-large at processor 2 \Theta gl-large at processor 3 \Theta Table 3: Low data skewness and high workload balance case Table 3 is an example of low data skewness and high workload balance. The support counts of the items A, B, C, D, E and F are almost equally distributed over the 3 processors. Hence, the data skewness is low. On the other hand, the workload balance is high, because the number of locally large itemsets in each processor is almost the same. In this case, both CD and distributed pruning generate the same 15 candidate sets; hence, if we restrict pruning to the distributed pruning, then it has no advantage over CD in this case. However, global pruning can prune away the candidates AC, AE and CE. In other words, FPM still has a 20% of improvement over CD in this pathological case of low skewness and high balance. 2 Following Example 2, it is observed that global pruning is more effective than distributed pruning and can perform significant candidates reduction even in the moderate data skewness or low workload balance cases. As a note, low skewness and low balance cannot occur together. Also, according to our analysis, distributed pruning can prune away almost n (n is the number of partitions) of all the size-2 candidates generated by CD in the high data skewness and high workload balance case. In summary, distributed pruning is very effective when a database is partitioned with high skewness and high balance. On the other hand, in the worst cases of high skewness with low balance or high balance with low skewness, the effect of distributed pruning is degraded to the level in CD, however, global pruning may still perform better than CD. To strengthen our studies, we investigated the problem of defining metrics to measure skewness and balance. 4.1 Data Skewness Metric We have developed a skewness metric based on the well established notion of entropy [5]. Given a random variable entropy is a measurement on how even or uneven its probability distribution is over its values. If a database is partitioned over n processors, the value pX X:sup can be regarded as the probability of occurrence of an itemset X in partition D i , n). The entropy log(p X (i))) is a measurement of the distribution of the local supports of X over the partitions. For example, if X is skewed completely into a single partition D k , (1 - k - n), i.e., it only occurs in D k , then pX k. The value of is the minimal in this case. On the other hand, if X is evenly distributed among all the partitions, then pX the value of log(n) is the maximal in this case. Therefore the following metric can be used to measure the skewness of a data partition. Table local count at processor 1 13 local count at processor 2 1 3 12 34 1 4 0.348 high local count at processor 3 data skewness S(X) 0.452 0.633 0.429 0.697 0.429 0.586 high TS(D) 0.494 workload balance TB(D) 0.999 Table local count at processor 1 13 33 12 34 2 1 0.601 local count at processor 2 high local count at processor 3 data skewness S(X) 0.452 0.633 0.429 0.697 0.429 0.586 low TS(D) 0.494 workload balance TB(D) 0.789 Table 3 local count at processor 1 6 12 4 13 5 12 0.329 local count at processor 2 6 12 5 12 4 13 0.329 low local count at processor 3 data skewness S(X) 0.015 0.001 0.012 0.001 0.012 0.001 high TS(D) 0.005 workload balance TB(D) 0.999 Table 4: Data Skewness and Workload Balance Given a database partition D i , (1 - i - n), the skewness S(X) of an itemset is defined by log(n). The skewness S(X) has the following properties: the skewness is at its lowest value when X is distributed evenly in all partitions. such that pX the skewness is at its highest value when X occurs only in one partition. in all the other cases. It follows from the property of entropy that S(X) increases with respect to the skewness of X ; hence, it is a suitable metric for the skewness of an individual itemset. Table 4 shows the skewness of the large itemsets in Tables 1, 2 and 3. In addition to measuring the skewness of an itemset, we also need a metric to measure the skewness of the database partition. We define the skewness of a database partition as a weighted sum of the skewness of all the large itemsets. In other words, the skewness of a partition is a measurement of the total skewness of all the large itemsets. Given a database partition D i , (1 - i - n), the skewness TS(D) of the partition is defined by X2LS S(X) \Theta w(X), where LS is the set of all the large itemsets, Y 2I S Y:sup is the weight of the support of X over all the large itemsets in LS , and S(X) is the skewness of X. TS(D) has some properties similar to S(X). when the skewness of all the itemsets are at its minimal value. when the skewness of all the itemsets are at its maximal value. in all the other cases. In Table 4, the skewness TS(D) of the partitions for the three situations have computed. (For illustration purpose, we only have computed TS(D) with respect to the skewness of all the size-1 large itemsets.) Note that we have ignored the small itemsets in the computation of the skewness of a partition. Since the purpose of our task is to investigate the effect of data skewness on candidate sets pruning, and this only involves large itemsets, this restriction would in fact make the metric more relevant to candidate set pruning. 4.2 Workload Balance Metric Workload balance is a measurement of the distribution of the support clusterings of the large itemsets over the partitions at the processors. Based on the definition of w(X) in Definition 2 and that of pX (i) in Definition 1, we define X2Ls w(X) \Theta pX (i) to be the itemset workload in a partition D i , where L s is the set of all the large itemsets. Intuitively, the workload W i in partition D i is the ratio of the total supports of the large itemsets in D i over all the partitions. Note that A partition has high workload balance if W i are the same for all partitions D i , On the other hand, if distribution of W i over the partitions are very uneven, then the workload balance is low. As has been pointed out, the workload balance has important bearing on the pruning and performance of parallel mining. In parallel to the metric for data skewness, we also define a metric workload balance factor to measure the workload balance of a partition, which is based also on entropy. Definition 3 For a database partition D i , database D, the workload balance factor TB(D) of the partition is given by log(n) . The metric TB(D) has the following properties: when the workload across all processors are the same; when the workload is concentrated on one processor; in all the other cases. In Table 4, the workload W i of the first and last cases (Tables 1 and 3) have a high balance, and the values of are almost equal to 1. In the second case (Table 2), the workload at processor 2 has been shifted to processor 1, and hence created an unbalance case; the value of TB(D) thus has been reduced to 0.789, which indicates a moderate workload balance. The data skewness metric and workload balance factor are not independent. Theoretically, each one of them could have values range from 0 and 1. However, some combinations of their values are not admissable. First, let us consider some boundary cases. Theorem 2 Let be a partition of a database D. 1. If then the admissable values of TB(D) range from 0 to 1. If 2. If then the admissable values of TS(D) range from 0 to 1. If Proof: 1. By definition 0 - TB(D) - 1. What we need to prove is that the boundary cases are admissable when implies that large itemsets X. Therefore, each large itemset is large at one and only one processor. If all the large itemsets are large at the same processor i, (1 - i - n), then On the other hand, if each processor has the same number of large itemsets, then W then large itemsets X . This implies that W i are the same for all 1 - i - n. Hence 2. It follows from the first result of this theorem that both are admissable when 1. Therefore the first part is proved. Furthermore, if there exists a partition D i , This implies that all large itemsets are locally large at only D i . Hence shown in Theorem 2 that not all possible combinations are admissable. In general, the admissable combinations will be a subset of the unit square such as the one in Figure 3. It always contains the two segments Figure 3) and Figure 3), but not the After defining the metrics and studying their characteristics, we can validate our observation on the relationship between data skewness, workload balance and candidates pruning effect in our performance studies. Performance Studies In order to confirm our analysis that the proposed FPM is an efficient algorithm for mining associations in a parallel system, we have implemented all the algorithms on an IBM SP2 and carried out a substantial performance evaluation and comparison. We have the following three goals in our studies: (1) to verify that FPM is faster than the representative algorithm CD, and confirm that the major performance gain is from the two pruning techniques; (2) to confirm the observation that both data skewness and workload balance are two critical factors in the performance of FPM; (3) to demonstrate that FPM has good parallelism and scalability as a parallel algorithm. We implemented FPM, its variant FNG, and CD. The IBM SP2 parallel system we used has processors (66.7MHz) with 64 MB main memory, running the AIX operating system. Communication between Figure 3: Admissable Combinations of Skewness(S) and Balance(B) processors are through a high performance switch with an aggregated peak bandwidth of 40 MBps and a latency of about 40 microseconds. Data was allocated to the local disk in each processor, and the database partition on each node is about 100MB in size. In order to be able to control the experiments to test different data distributions and scenarios, many works [2, 3, 9, 15, 16] in mining association rules have adopted the standard technique introduced in [2] to generate the database. We have enhanced the technique for the generation of database partitions and introduced parameters to control the skewness and workload balance. Table 5 is a list of the parameters used in our synthetic databases. Details of the data generation technique is in the appendix. D number of transactions in each partition average size of the transactions I average size of the maximal potentially large itemsets L number of maximal potentially large itemsets N number of items partition skewness n number of partitions Table 5: Synthetic Database Parameters 5.1 Relative Performance In order to compare the performance between FMP, FNG, and CD, a databases and twenty data sets have been generated. The data sets generated and their skewness and balance factors are listed in Table 6. The number of partitions in each case is and the size of each partition is about 100MB. The name of a partition is denoted by Dx.Ty.Iz.Sr.Bl, where x is the number of transactions in each partitions, y is the average size of the transactions, z is the average size of the itemsets. The two parameters Sr and Bl are two important parameters used to control the skewness and workload balance in the data generation. (The Bl values are listed separately in the table.) In Table 6, we have also computed the measured skewness TS(D) and the balance factor TB(D) of the partitions generated. It is important to note that these measured skewness and workload are very close to the values of the controlled parameters, i.e., the values of S and B are good approximations of values of TS(D) and TB(D). In addition, they cover a wide range of skewness and balanace. Thus, our synthesized data partitions are good simulation of data partitions of various distribution characteristics. We believe this is technically valuable because even real data may not be general enough to cover all possible distributions. Name Table Attributes of Synthetic Databases We ran FPM, FNG and CD on the database partitions in Table 6. The minimum support threshold is 0.5%. The improvement of FPM and FNG over CD in response time are recorded in Table 7, and the result is very encouraging. FPM and FNG are consistently faster than CD in all cases. In the following, we analyze the performance gain of FPM and FNG in three aspects: (1) improvement when the workload balance is high, and the skewness varies from high to moderate; (2) improvement when the skewness is high, and the workload balance varies from high to moderate; (3) desirable and undesirable combinations of skewness and workload balance values. Response Time FPM/CD FNG/CD Table 7: Performance Improvement of FPM and FNG over CD Figure 4 is the relative performance between FPM, FNG and CD on partitions with different skewness and a high balance value performs much better than CD when the skewness is relatively high (s ? 0:5). On the Response time(sec.) Relative Performance-D3278K.T5.I2.S?.B100(n=16, minsup=0.5%) CD Figure 4: Relative Performance on Databases with High Balance and Different Skewness other hand, FPM outperforms CD significantly even when the skewness is in the moderate range, (0:1 - s - 0:5). When B=90, the result in Table 7 shows that FPM is again much faster than CD. This demonstrates that FPM outperforms CD consistently given a high workload balance and at least a moderate skewness.1000300050000 0.1 0.3 0.5 0.7 0.9 1 Response time(sec.) Relative Performance-D3278K.T5.I2.S90.B?(n=16, minsup=0.5%) CD Figure 5: Relative Performance on Databases with High Skewness and Different Workload Balance Figure 5 is the relative performance given a high skewness different workload balance. Both FPM and FNG perform much better than CD when the workload balance is relatively high (B ? 0:5); however, the improvement in the range of moderate balance, (0:1 - B - 0:5), is marginal. This confirms our observation that workload balance is an essential requirement. It shows that a high skewness has to accompany by a high workload balance. The effect of a high skewness with a moderate balance may not be as good as that of a high balance with a moderate skewness. In Figure 6, we vary both the skewness and balance together from a low value range to a high value range. The Response time(sec.) Skewness and Balance(S,B) Relative Performance-D3278K.T5.I2.S?.B?(n=16, minsup=0.5%) CD Figure Relative Performance on Databases when both Skewness and Balance change trend shows that the improvement of FPM over CD increases faster when both values approach the high value range. Combining the observations in the above three cases together with the results in Table 7, we can divide the area covering all the admissable combinations of skewness and balance in our experiments into four performance regions as shown in Figure 7. Region A is the most favorable region in which the balance is high and the skewness varies from high to moderate. FPM in general is 50% to 100% faster than CD. In region B, the workload balance value has degraded moderately and the skewness remains high; in this case, the gain in FPM over CD is around 50%. Region covers combinations that have very low workload balance; the gain in FPM falls into a moderate range of about 30%. Region D contains the most undesirable combinations; FPM only has marginal performance gain. Figure 8 provides us another view to understand the candidates pruning effects. It shows the ratio on the number of candidate sets between FPM (FNG) and CD for the same experiments in Figure 4. The reduction ratios for the runs in the database D3278K.T5.I2.Sr.B100, are in the first graph. When the skewness is high, distributed pruning has a 79.2% of reduction in candidate sets comparing with CD, and global pruning has a 93.9% reduction. When the skewness is low, distributed pruning only has a 6.6% reduction, but global pruning has a 30.7% reduction. This confirm our observation on the effect of high balance combined with high or moderate skewness. 5.2 Scalability and Parallelism : Speedup and Scaleup In order to study the efficiency of FPM as a parallel algorithm, we investigate its speedup and scaleup against CD. Speedup is the reduction in response time vs the number of processors, given that the total size of the database remains unchanged. The more processors are used, the faster the computation should be. The ideal speedup is a linear function on the number of processors. Scaleup is the performance vs the number of processors when the database size is scaled up proportional to the number of proccesors. If the algorithm has high efficiency and low overhead, its performance would maintain uniform when both the number of processors and the size of the database scaled up proportionally. In the speedup experiment, we execute the algorithms on a fixed size database with various number of processors * A Figure 7: Performance regions (FPM/CD) in the admissible combinations of skewness and workload on number of candidate sets Pruning Effect-D3278K.T5.I2.S?.B100(n=16, minsup=0.5%) Apriori_gen(no pruning) Distributed Pruning/Apriori_gen Global Pruning/Apriori_gen Figure 8: Pruning Effect on Databases in Figure 4 Response time(sec.) Number of processors CD FPM14122028 Number of processors Relative Speedup-D3278K.T5.I2.S90.B100 CD Figure 9: Speedup on a Database and partitions. We selected the database with high skewness and balance as a presentative to perform the study. The database is listed in Table 8. It has a total size of 1.6GB, and was first divided into 16 partitions. Subsequently, we combined the partitions to form databases with 8, 4, 2, and zero partitions. Figures 9 is the execution times and speedups of FPM, FNG, and CD on the databases. The speedups are also shown in Table 8. FNG had a linear speedup and FPM achieved a remarkable superlinear speedup. The reason behind FPM's superlinear speedup is the increase in the skewness when the number of partitions increases. In the scaleup experiment, both the database size and the number of processors are scaled up proportionally. The number of processors involved were increased from 1 to 16, and the sizes of the databases were increased correspondingly from 100MB to 1.6GB. Each database were partitioned according to the number of processors such that every partition is maintained at the size of 100MB. We performed the experiment based on the database D3278K.T5.I2.S90.B100, i.e., the databases are generated with the same parameters. Figure 10 shows the result of the experiment. Surprisingly, both FPM and FNG not only can maintain the performance, their response time in Databases Speedup of FPM Speedup of FNG Speedup of CD Table 8: Speedup on five databases with different distribution characteristics fact had gone down when the database was scaled up. The prime reason for this is the increase in pruing capability when the number of partitions increases. 6 Discussion To restrict the search of large itemsets in a small set of candidates is essential to the performance of mining association rules. The pruning techniques we proposed are theoretically interesting, because they have effect only in the parallel case but not the serial case. Both distributed and global pruning provide significant amount of pruning power, in particular, when the data distribution is in a favorable situation, i.e. when workload balance is high and the skewness is at least at a moderate level. It is important to study partition techniques that can deliver a good data distribution. Random approaches in general will deliver partitions which have good balance. However, the skewness would be difficult to guarantee together with good workload balance. Clustering technique such as the k-means algorithm [12] will give good skewness. It remains an open problem how to modify clustering technique to generate partitions which have good skewness and also good workload balance. 7 Conclusion A parallel algorithm FPM for mining association rules has been proposed. A performance study carried out on an IBM SP2 shared-nothing memory parallel system shows that FPM consistently outperforms CD. It also has nice scalability in terms of speedup and scaleup. The gain in performance in FPM is due mainly to the pruning techniques incorporated. It was discovered that the effectiveness of the pruning techniques depend highly on the data distribution characteristics, which can be measured by two data skewness and workload balance. Our analysis and experiment results show that the pruning techniques are very sensitive to workload balance, though good skewness will also have important positive effect. The techniques are very effective in the best case of high balance and high skewness. The combination of high balance and moderate skewness is the second best case. In the worst case of low balance and low skewness, FPM can only deliver the performance close to that of CD. Since mining associations has many interesting applications, important future works would include fine tuning of the proposed parallel techniques on real business cases. --R Mining Association Rules between Sets of Items in Large Databases. Fast algorithms for mining association rules. Parallel mining of association rules: Design Dynamic itemset counting and implication rules for market basket data. Elements of information theory. A fast distributed algorithm for mining association rules. Maintenance of Discovered Association Rules in Large Databases: An Incremental Updating Technique. Advances in Knowledge Discovery and Data Mining. Discovery of Multiple-level association rules from large databases Scalable parallel data mining for association rules. Scalable POWERparallel Systems Some methods for classification and analysis of multivariate observations Message Passing Interface Forum. Exploratory Mining and Pruning Optimizations of Constrainted Association Rules. An effective hash-based algorithm for mining association rules Efficient Parallel Data Mining for Association Rules. An efficient algorithm for mining association rules in large databases. Database Achievements and Opportunities Into the 21st Century. Mining generalized association rules. Mining sequential patterns: Generalizations and performance improvements. Mining quantitative association rules in large relational tables. Hash based parallel algorithms for mining association rules. Parallel data mining for association rules on shared-memory multi-processors --TR --CTR Frans Coenen , Paul Leng, Partitioning strategies for distributed association rule mining, The Knowledge Engineering Review, v.21 n.1, p.25-47, March 2006 Vipin Kumar , Mohammed Zaki, High performance data mining (tutorial PM-3), Tutorial notes of the sixth ACM SIGKDD international conference on Knowledge discovery and data mining, p.309-425, August 20-23, 2000, Boston, Massachusetts, United States
data skewness;data mining;workload balance;parallel mining;association rules;parallel computing
593499
Scalable Techniques for Mining Causal Structures.
Mining for association rules in market basket data has proved a fruitful area of research. Measures such as conditional probability (confidence) and correlation have been used to infer rules of the form the existence of item A implies the existence of item B. However, such rules indicate only a statistical relationship between A and B. They do not specify the nature of the relationship: whether the presence of A causes the presence of B, or the converse, or some other attribute or phenomenon causes both to appear together. In applications, knowing such causal relationships is extremely useful for enhancing understanding and effecting change. While distinguishing causality from correlation is a truly difficult problem, recent work in statistics and Bayesian learning provide some avenues of attack. In these fields, the goal has generally been to learn complete causal models, which are essentially impossible to learn in large-scale data mining applications with a large number of variables.In this paper, we consider the problem of determining casual relationships, instead of mere associations, when mining market basket data. We identify some problems with the direct application of Bayesian learning ideas to mining large databases, concerning both the scalability of algorithms and the appropriateness of the statistical techniques, and introduce some initial ideas for dealing with these problems. We present experimental results from applying our algorithms on several large, real-world data sets. The results indicate that the approach proposed here is both computationally feasible and successful in identifying interesting causal structures. An interesting outcome is that it is perhaps easier to infer the lack of causality than to infer causality, information that is useful in preventing erroneous decision making.
Introduction In this paper we consider the problem of determining casual relationships, instead of mere as- sociations, when mining market basket data. We discuss ongoing research in Bayesian learning where techniques are being developed to infer casual relationships from observational data, and we identify one line of research in that community which appears to hold promise for large-scale data mining. We identify some problems with the direct application of Bayesian learning ideas to mining large databases, concerning both the issue of scalability of algorithms and the appropriateness of the statistical techniques, and introduce some ideas for dealing with these problems. We present experimental results from applying our algorithms on several large, real-world data sets. The results indicate that the approach proposed here is both feasible and successful in identifying interesting causal structures. A significant outcome is that it appears easier to infer the lack of causality, information that is useful in preventing erroneous decision making. We conclude that the notion of causal data mining is likely to be a fruitful area of research for the database community at large, and we discuss some possibilities for future work. Let us begin by briefly reviewing the past work involving the market basket problem, which involves a number of baskets, each of which contains a subset of some universe of items. An alternative interpretation is that each item has a boolean variable representing the presence or absence of that item. In this view, a basket is simply a boolean vector of values assigned to these variables. The market basket problem is to find "interesting" patterns in this data. The bulk of past research has concentrated on patterns that are called association rules, of the type: "item Y is very likely to be present in baskets containing items A major issue in mining association rules has been that of finding appropriate definitions of "interest" for specific applications. An early approach, due to Agrawal, Imielinski, and Swami [2], was to find a pair of items that occur together often (that is, have high support), and also have the property that one item often occurs in baskets containing the other item (that is, have high confidence). 1 In effect, this framework chooses conditional probability as the measure of interest. Many variants of this interest measure have been considered in the literature, but they all have a flavor similar to conditional probability. These measures were critiqued by Brin, Motwani, and Silverstein [9], who proposed statistical correlation as being a more appropriate interest measure for capturing the intuition behind association rules. In all previous work in association rules and market basket mining, the rules being inferred, such as "the existence of item A in a basket implies that item B is also likely to be present in that basket," often denoted as A ) B, indicate only the existence of a statistical relationship between items A and B. They do not, however, specify the nature of the relationship: whether the presence of A causes the presence of B, or the converse, or some other phenomenon causes both to appear together. The knowledge of such causal relationships is likely to be useful for enhancing understanding and effecting change; in fact, even the knowledge of the lack of a casual relationship will aid decision making based on data mining. We illustrate these points in the 1 The support-confidence definition easily extends beyond pairs of items to incorporate sets of arbitrary size. following hypothetical example. Consider a supermarket manager who notes that his meat-buying customers have the following purchasing pattern: buy hamburgers 33% of the time, buy hot dogs 33% of the time, and buy both hamburgers and hot dogs 33% of the time; moreover, they buy barbecue sauce if and only if they buy hamburgers. Under these assumptions, 66% of the baskets contain hot dogs and 50% of the baskets with hot dogs also contain barbecue sauce. The manager will find that the association rule hot- dogs ) barbecue-sauce has both high support and confidence. (Of course, the rule hamburger ) barbecue-sauce has even better confidence, but that is an obvious association.) A manager who has a deal on hot dogs may choose to sell them at a large discount, hoping to increase profit by simultaneously raising the price of barbecue sauce. However, the correct causal model (that the purchase of hamburgers causes the purchase of barbecue sauce) tells us that this approach is not going to work. In fact, the sales of both hamburgers and barbecue sauce are likely to plummet in this scenario, as the customers buy more hot dogs and fewer hamburgers, leading to a reduction in sales of barbecue sauce. If the manager could infer the correct causal model, or even infer that "hot-dogs causes barbecue-sauce" is not part of any possible causal model, he could avoid a pricing fiasco. A basic tenet of classical statistics ([6], [20]) is that correlation does not imply causation. Thus, it appears impossible to infer causal relationships from mere observational data available for data mining, since we can only infer correlations from such data. In fact, it would seem that to infer causal relationships it is essential to collect experimental data, in which some of the variables are controlled explicitly. This experimental method is neither desirable nor possible in most applications of data mining. Fortunately, recent research in statistics and Bayesian learning communities provide some avenues of attack. Two classes of technique have arisen: Bayesian causal discovery, which focuses on learning complete causal models for small data sets [8, 12, 14, 15, 16, 17, 21, 22, 23, 25, 26, 27], and an offshoot of the Bayesian learning method called constraint-based causal discovery, which use the data to limit - sometimes severely - the possible causal models [11, 26, 24]. While techniques in the first class are still not practical on very large data sets, a limited version of the constraint-based approach is linear in the database size and thus practical on even gigabytes of data. We present a more flexible constraint-based algorithm, which is linear in the number of records (baskets) in the database, though it is cubic in the number of items in each record. Despite the cubic time bound, the algorithm proves to be practical for databases with thousands of items. In this paper, we explore the applicability of a constraint-based causal discovery to discovering causal relationships in market basket data. Particularly, we build on ideas presented by Cooper [11], using local tests to find a subset of the causal relationships. In the rest of this section, we discuss causality for data mining in the context of research into causal learning. We begin, in Section 2 with a particular constraint-based algorithm, due to Cooper [11], upon which we build the algorithms This example is borrowed from a talk given by Heckerman. presented in this paper. We then enhance the algorithm so that for the first time causality can be inferred in large-scale market-basket problems. ffl Section 2 introduces "CCU" inferences, a form of causal structure not used by [11]. ffl Section 3 discusses weaknesses of the Cooper algorithm, notably a susceptibility to statistical error, and how power statistics such as correlation can be used to mitigate these problems. ffl In Section 4 we describe in detail the algorithms we developed for discovering causal relation- ships, and we also discuss discovery of noncausal relationships, an important technique that filters many statistically unlikely inferences of causality. ffl For the first time, we are able to run causality tests on real, large-scale data. In Section 5 we test this algorithm on a variety of real-world data sets, including census data and text data. In the former data sets we discover causal relationships (and nonrelationships) between census categories such as gender and income. In the text data set we discover relationships between words. ffl Finally, in Section 6, we discuss possible directions for future research. 1.1 Previous Research in Causality As we have mentioned, there has been significant work in discovering causal relationships using Bayesian analysis. A Bayesian network is a combination of a probability distribution and a structural model that is a directed acyclic graph in which the nodes represent the variables (attributes) and the arcs represent probabilistic dependence. In effect, a Bayesian network is a specification of a joint probability distribution that is believed to have generated the observed data. A causal Bayesian network is a Bayesian network in which the predecessors of a node are interpreted as directly causing the variable associated with that node. In Bayesian learning techniques, the user typically specifies a prior probability distribution over the space of possible Bayesian networks. These algorithms then search for that network which maximizes the posterior probability of the data provided. In general, they try to balance the complexity of the network with its fit to the data. The possible number of causal networks is severely exponential in the number of variables, so practical algorithms must use heuristics to limit the space of networks. This process is helped by having a quality prior distribution, but often the prior distribution is unknown or tedious to specify, particularly if the number of variables (i.e., items) is large. In this case, an uninformative prior is used. Even when informative priors are available, the goal of finding a full causal model is aggressive, and Bayesian algorithms can be computationally expensive. While improved heuristics, and the use of sampling, may make Bayesian algorithms practical, this has yet to be demonstrated for data sets with many variables. In our view, inferring complete causal models (i.e., causal Bayesian networks) is essentially impossible in large-scale data mining applications with thousands of variables. For our class of applications, the so-called constraint-based causal discovery method [24, 26] appears to be more useful. The basic insight here, as articulated by Cooper [11], is that information about statistical independence and dependence relationships among a set of variables can be used to constrain (sometimes significantly) the possible causal relationships among a subset of the variables. A simple example of such a constraint is that if attributes A and B are independent, then it is clear that there is no causal relationship between them. It has been shown that, under some reasonable set of assumptions about the data (to be discussed later), a whole array of valid constraints can be derived on the causal relationships between the variables. Constraint-based methods provide an alternative to Bayesian methods. The PC and FCI algorithms [26] use observational data to constrain the possible causal relationships between variables. They allow claims to be made such as "X causes Y ," "X is not caused by Y ," "X and Y have a common cause," and so on. For some pairs, they may not be able to state the causal relationship. These constraint-based algorithms, like the Bayesian algorithms, attempt to form a complete causal model and therefore can take exponential time. (Due to the complexity of their causal tests, they may also be less reliable than simpler algorithms.) Cooper [11] has described an algorithm called LCD that is a special case of the PC and FCI algorithms and runs in polynomial time. Since our algorithm is based on Cooper's, we discuss it in some detail in Section 2. 1.2 Causality for Market Basket Analysis Finding causality in the context of data mining is particularly difficult because data sets tend to be large: on the order of megabytes of data and thousands of variables. While this large size poses a challenge, it also allows for some specializations and optimizations of causal algorithms, holding out promise that algorithms crafted particularly for data mining applications may yield useful results despite the large amount of data to be processed. This promise holds particularly true for market basket data, which is an even more specialized application. Below, we note some issues involved with finding causality in the market basket case. ffl Market basket data is boolean. This may allow for added efficiency over algorithms that work with discrete or continuous data. As we shall see in Section 3.1, some statistical tests that are essential to constraint-based causal discovery have pleasant properties in the boolean case. ffl The traditional market basket problem assumes there is no missing data: that is, for a given item and basket it is known whether or not the item is in the basket. This assumption obviates the need for complex algorithms to estimate missing data values. ffl Because market basket data is usually voluminous, algorithms that need a large amount of data to develop a causal theory are well suited to this application. ffl With thousands of items, there are likely to be hundreds of thousands of causal relation- ships. While an optimal algorithm might find all these relationships and output only the "interesting" ones (undoubtedly using another data mining algorithm as a subroutine!), it is acceptable to find and output only a small number of causal relationships. Selection may occur either due to pruning or due to an algorithm that finds only causal relationships of a certain form. Algorithms that output a small number (possibly arbitrarily decided) of causal relationships may not be useful outside a data mining context. Data mining, however, is used for exploratory analysis, not hypothesis testing. Obviously, a technique for finding all causal relationships and somehow picking "interesting" ones is superior to one that chooses relationships arbitrarily, but even techniques in the latter category are valuable in a data mining setting. ffl Again, for market basket applications with thousands of items, finding a complete causal model is not only expensive, it is also difficult to interpret. We believe isolated causal rela- tionships, involving only pairs or small sets of items, are easier to interpret. ffl For many market basket problems, discovering that two items are not causally related, or at least not directly causally related (that is, one may cause the other but only through the influence of a third factor), may be as useful as finding out that two items are causally related. While complete causal models illuminate lack of causality as easily as they illuminate causality, algorithms that produce partial models are more useful in the market basket setting if they can discover some noncausal relationships as well as causal relationships. These aspects of discovering causality for market basket data drove the development of the algorithms we present in Section 4. Many of these issues point to constraint-based methods as being well suited to market basket analysis, while others indicate that tailoring constraint-based methods - for instance by providing error analysis predicated on boolean data and discovering lack of causality - can yield sizable advantages over using generic constraint-based techniques. 2 The LCD Algorithm The LCD algorithm [11] is a polynomial time, constraint-based algorithm. It uses tests of variable dependence, independence, and conditional independence to restrict the possible causal relationships between variables. The crux of this technique is the Markov condition [26]. A be a node in a causal Bayesian network, and let B be any node that is not a descendant of A in the causal network. Then the Markov condition holds if A and B are independent, conditioned on the parents of A. The intuition of this condition is as follows: If A and B are dependent, then B must either be a (possibly indirect) cause of A or (possibly indirectly) caused by A. In the second case, B is a descendant of A, while in the first B is an ancestor of A and has no effect on A once A's immediate parents are fixed. Assuming the Markov condition, we can make causal claims based on independence data. For instance, suppose we know, possibly through a priori knowledge, that A has no causes. Then if B is dependent on A, B must be caused by A, though possibly indirectly. If we have a third variable C dependent on both A and B, then the three variables lie along a causal chain. Variable A, since it has no causes, is at the head of the chain, but we don't know whether B causes C or vice versa. If, however, A and C become independent conditioned on B, then we conclude by the Markov condition that B causes C. In the discussion that follows, we denote by B ! C the claim that B causes C. Note that, contrary to normal use in the Bayesian network literature, we do not use this to mean that B is a direct cause of C. Because we have restricted our attention to only three variables, we cannot in fact say with assurance that B is a direct cause of C; there may be a confounding variable D, or some hidden variable, that mediates between B and C. A confounding variable is a variable that interacts causally with the items tested, but was not discovered because it was not included in the tests performed. A hidden variable represents an effect that interacts causally with the items tested, but is not captured by any variable in the data set. Even with hidden and confounding variables, we can say with assurance that A is a cause of B and B a cause of C. We can also say with assurance that A is not a direct cause of C, since its causality is mediated by B, at the minimum. If we drop the assumption that A has no causes, then other models besides A consistent with the data. In particular, it may be that A / C. In this case, it is impossible, without other knowledge, to make causality judgments, but we can still say that A is not a direct cause of C, though we do not know if it is an indirect cause, or even caused by C instead of causing C. We summarize these observations in the CCC rule, so named since it holds when A, B, and C are all pairwise correlated. 3 Rule 1 (CCC causality) Suppose that A, B, and C are three variables that are pairwise depen- dent, and that A and C become independent when conditioned on B. Then we may infer that one of the following causal relations exists between A, B, and C: Now suppose two variables B and C are independent, but each is correlated with A. Then B and C are not on a causal path, but A is on a causal path with both of them, implying either both are ancestors of A or both are descendants of A. If B and C become dependent when conditioned on A, then by the Markov condition they cannot be descendants of A, so we can conclude that B and C are causes of A. This observation gives rise to the CCU rule, so named since two variable pairs are correlated and one is uncorrelated. Rule 2 (CCU causality) Suppose A, B, and C are three variables such that A and B are corre- lated, A and C are correlated, and B and C are uncorrelated, and that B and C become correlated when conditioned on A. Then we may infer that B and C cause A. Again we cannot say whether hidden or confounding variables mediate this causality. 3 Here correlation indicates dependence and not a specific value of the correlation coefficient. While use of the term "correlation" we intend should be clear from context. Algorithm: LCD Input: A set V of variables. w, a variable known to have no causes. A data set D. Tests for dependence, independence, and conditional independence (D, I, and CI respectively). Output: A list of causal relationships supported by the data. For all variables x 6= w If D(x; y) For all variables y 62 fx; wg If D(x; y) and D(y; w) and CI(x; output 'x causes y' Table 1: The LCD algorithm. The LCD algorithm uses the CCC rule (but not the CCU rule) to determine causal relation- ships. It looks at triples of items, where one item is known a priori to have no cause. In this way it can disambiguate the possible causal models. The algorithm assumes tests for dependence and conditional independence. It is shown in Table 1. The LCD algorithm depends on the correctness of the statistical tests given as input. If one test wrongly indicates dependence or conditional independence, the results will be invalid, with both false positives and false negatives. An additional assumption, as has already been stated, is in the applicability of the Markov condition. We list some other assumptions, as described by Cooper [11], and their validity for market basket data. Database Completeness The value of every variable is known for every database record. This is commonly assumed for market basket applications. Discrete Variables Every variable has a finite number of possible values. The market basket problem has boolean variables. Causal Faithfulness If two variables are causally related, they are not independent. This is a reasonable assumption except for extraordinary data sets, where, for instance, positive and negative correlations exactly cancel. Markov Condition This condition is reasonable if the data can actually be represented by a Bayesian network, which in turn is reasonable if there is no feedback between variables. Bias The probability distribution over the data set is equal to the probability distribution over the underlying causal network. The reasonableness of this assumption depends on the specific problem: if we only collect supermarket data for customers who use a special discount card, there is likely to be selection bias. If we collect data for all customers, or random customers, selection bias is unlikely to be a problem. Valid Statistical Testing If two variables are independent, then the test of independence will say so. If they are dependent, the test of dependence will say so. This assumption is unreasonable, since all tests have a probability of error. When many tests are done, as is the case for the LCD algorithm, this error is an even bigger concern (see Section 3.1). A criticism of the LCD algorithm is that it finds only causal relationships that are embedded in CCC triples, presumably a small subset of all possible causal relationships. Furthermore, this pruning is not performed on the basis of a goodness function, but rather because of the exigencies of the algorithm: these are the causal relationships that can be discovered quickly. While this trait of the LCD algorithm is limiting in general, it is not as problematic in the context of data mining. As we mentioned in Section 1.2, data mining is used for exploratory analysis, in which case it is not necessary to find all, or even a small number of specified, causal relationships. While not ideal, finding only a small number of causal relationships is acceptable for data mining. 3 Determining Dependence and Independence Cooper [11] uses tests for dependence and independence as primitives in the LCD algorithm, and also proposes Bayesian statistics for these tests. In our approach, we use instead the much simpler refer to [9] for a discussion on using the chi-squared tests in market basket applications. The necessary fact is that if two boolean variables are independent, the - 2 value is likely to exceed the threshold value - 2 ff with probability at most ff. There are tables holding - 2 ff for various values of ff. 4 We say that if the - 2 value is greater than - 2 ff , then the variables are correlated with probability We extend this definition to the market basket problem by adding the concept of support, which is the proportion of baskets that a set of items occurs in. be a support threshold and c 2 (0; 1) be a confidence threshold. An itemset S ' I is said to be (s; c)-correlated (hereafter, merely correlated) if the following two conditions are met: 1. The value of support(S) exceeds s. 2. The - 2 value for the set of items S exceeds the - 2 value at significance level c. Typical values for the two threshold parameters are then we would expect that, for 5% of the pairs that are actually uncorrelated, we would claim (incorrectly) they are correlated. Support is not strictly necessary; we use it both to increase the effectiveness of the chi-squared test and to eliminate rules involving infrequent items. Intimately tied to the notion of correlation is that of uncorrelation, or independence. Typically, uncorrelation is defined as the opposite of correlation: an itemset with adequate support is uncorrelated if the - 2 value does not support correlation. In effect, the chi-squared test is being applied as a one-tailed test. 4 In the boolean case the appropriate row of the table is the one for 1 degree of freedom. This definition is clearly problematic: with sets with a - 2 value just below the cutoff will be judged uncorrelated, even though we judge there is almost a 95% chance the items are actually correlated. We propose, instead, a two-tailed test, which says there is evidence of dependence if ff and evidence of independence if ff 0 . The following definition is based on this revised test. be a support threshold and c 2 (0; 1) be a confidence threshold. An itemset S ' I is said to be (s; c)-uncorrelated (hereafter, merely uncorrelated) if the following two conditions are met: 1. The value of support(S) exceeds s. 2. The - 2 value for the set of items S does not exceed the - 2 value at significance level c. then we would expect that, for 5% of the pairs that are actually uncorrelated, we would fail to say they are uncorrelated. Note that we would not necessarily say they are correlated: a pair of items may be neither correlated nor uncorrelated. Such a pair cannot be part of either CCC causality or CCU causality. We can use the chi-squared test not only for dependence and independence, but also for conditional dependence and conditional independence. Variables A and B are independent conditioned on C if p(AB j C). The chi-squared test for conditional independence looks at the statistic - 2 (AB j is the chi-squared value for the pair A; B limited to data where As with standard correlation, we use a two-tailed chi-squared test, using different thresholds when testing for conditional dependence as opposed to conditional independence. For example, suppose everybody over both drives and votes, but nobody under does either. Then driving and voting are dependent - quite powerfully, since driving is an excellent predictor of voting, and vice versa - but they are independent conditioned on age. To see this, note that if we know somebody's age, knowing whether they drive yields no extra insight in predicting whether they vote. Note that both the correlation and the uncorrelation tests bound the probability of incorrectly labeling uncorrelated data, but do not estimate the probability of incorrectly labeling correlated pairs. This is a basic problem in statistical analysis of correlation: while rejecting the null hypothesis of independence requires only one test, namely that the correlation is unlikely to actually be 0, rejecting the null hypothesis of dependence requires an infinite number of tests: that the correlation is not 0:5, that the correlation is not 0:3, and so on. Obviously, if the observed correlation is 0:1, it is likelier that the actual correlation is 0:3 than that it is 0:5, giving two different probabilities. It is unclear what number would capture the concept that the pair is "correlated." One solution to this problem is to define correlation as "the correlation coefficient is higher than a cutoff value." For boolean data, this is equivalent to testing the chi-squared value as we do above; see Section 3.1 and Appendix A for details. 3.1 Coefficient of Correlation The LCD algorithm can perform the tests for dependence and independence tens of thousands of times on data sets with many items. Though the individual tests may have only a small probability of error, their frequent use means there will be hundreds of errors in the final result. This problem is exacerbated by the fact a single erroneous dependence judgment could form the basis of up to rules. This problem is usually handled in the statistical community by lowering the tolerance value for each individual test, so that the total error rate is low. In general, with thousands of tests the error rate will have to be set intolerably low. However, for boolean data even a very low tolerance value is acceptable because of a connection between the probability of error and the strength of correlation, presented in the following theorem. This proof of this theorem, along with the concept of correlation coefficient at its heart, can be found in Appendix A. Theorem 1 Let X and Y be boolean variables in a data set of size n, with correlation coefficient ae. will fail to be judged correlated only if the confidence level for the correlation test is below that for - 2 Because of this relationship, by discarding rules that are more likely to be erroneous, we are at the same time discarding rules with only a weak correlation. Weak rules are less likely to be interesting in a data mining context, so we are at the same time reducing the probability of error and improving the quality of results. 4 Algorithms for Causal Discovery In the following discussion we shall use the following terminology: A pair of items constitutes a C-edge if they are correlated according to the correlation test, and they constitute a U-edge if they are uncorrelated according to the uncorrelation test. (Note that an item pair may be neither a C-edge nor a U-edge.) We denote the number of items by m, the number of baskets by n, and the degree of node A - that is, the number of C- and U - edges involving item A - by \Delta A . When necessary, we shall also refer to \Delta C A and \Delta U A , which are the degree of A when restricted to C- and U-edges, respectively. Let \Delta be the maximum degree, that is, maxA are defined similarly. We consider the performance of algorithms with respect to three factors: memory use, running time, and number of passes required over the database. Since our techniques look at triples of items, memory is enough to store all the count information needed for our algorithms. Because of this, machines with O(m 3 ) memory require only one pass over the database in all cases. The algorithms below assume that m is on the order of thousands of items, so caching the required database information in this way is not feasible. However, we consider O(m memory to be available. Situations where less memory is available will have to rely on the naive algorithm, which requires only O(1) memory. 4.1 The Naive Algorithm Consider first the brute force search algorithm for determining all valid causal relations from market basket data. Effectively, we iterate over all triples of items, checking if the given triple satisfies the conditions for either CCC or CCU causality. This requires a conditional independence test, which requires the count nABC . Thus, the brute force algorithm requires O(m 3 ) passes over the database, and this alone takes time O(nm 3 ). However, the algorithm requires only O(1) memory. If M words of memory are available, we can bundle count requests to reduce the number of database passes to O(m 3 =M ). 4.2 The CC-path Algorithm The naive algorithm can be speeded up easily if O((\Delta C memory is available: Consider each item A in turn, determine all items connected to A via C-edges, and for each pair B and C of these C-neighbors check if either causality rule applies to ABC. This approach requires that we examine instead of O(m 3 ). More importantly, it requires only n passes over the database, since in the pass for item A, we can use the space to store counts for all ABC in which B and C are connected to A by a C-edge. The running time of the resulting algorithm is accurately, the running time is O(n accurately still, it is O(n A m)), since it takes O(nm) time just to calculate which neighbors of A are correlated. In the future, we ignore such fine distinctions.) This algorithm has the same worst-case running time as the naive algorithm, but unless \Delta C A is very large it is faster than performing the naive algorithm and bundling n 2 count requests. 4.3 The CU-path Algorithm The CC-path algorithm is so named because it looks at (with A as the joint) and then checks for the existence of the third, uncorrelated, edge. Another approach, appropriate only for finding CCU causality, is to look for C \Gamma U paths and check if the third edge is correlated. This algorithm is superior when \Delta U A for most A. Let \Delta A \Delta U A g (in general, this will be less than \Delta C \Delta U ). The CU-path algorithm requires O(\Delta CU ) memory, O(m) passes over the database. Again, there is the tighter but more complicated time bound of O(n A \Delta U A ). 4.4 The CU-path Algorithm with Heuristic The CU-path algorithm allows for a heuristic that is not available for the CC-path algorithm. It follows from the fact every CCU triple has two C \Gamma U paths but only one Therefore, for every U edge there is a choice, when looking for C \Gamma U paths, of whether to look at one endpoint or the other. From a computational point of view it makes sense to pick the endpoint that abuts fewer C-edges. As a result there will be fewer C \Gamma U paths to process. (One way to think of it is this: there are C \Gamma U paths that are part of CCU triples, and those that are not. The former Theoretical Theoretical clari Theoretical clari Algorithm Space Time Time DB passes DB passes naive O(1) O(nm 3 CC-path CU-path CU-path with heuristic Table 2: Summary of running time and space for finding CCU causal relationships, in both theory and practice on the clari data set (Section 5.4). This data set has Time is in seconds of user time. To improve running time, the algorithm grouped items together to use the maximum memory available on the machine. Thus, a comparison of memory use is not helpful. The naive algorithm was not run on this data set. we must always look at, but the latter we can try to avoid.) This heuristic has proven extremely successful, particularly when the number of C-edges is large. For the clari data set (Section 5.4), the CU-path heuristic cut the running time in half. The improvement in the smaller clari.world data set (Section 5.2 was percent. Optimizations are possible. For instance, the algorithms described above check twice whether a pair of items share an edge, once for each item in the pair. It would be faster, memory permitting, to determine all correlated and uncorrelated edges once as a pre-processing step and store them in a hash table. Even better would be to store edges in an adjacency list as well as in the hash table, to serve as a ready-made list of all C- and U-edges abutting the "joint" item. In experiments, this improvement in data structures halved the running time. Caching as many triple counts as will fit in main memory will also improve the running time by a constant factor. 4.5 Comparison of Performance Table 2 holds a summary of the algorithms we consider and their efficiencies. Note that the number of database passes is not the same for all algorithms. This is because if an item lacks a correlated (or uncorrelated) neighbor, we need not perform a database pass for that item. For the clari data set (Section 5.4), there are 316295 C-edges but only 5417 U-edges. This explains the superior performance of the CU-path algorithm, both in terms of time and database passes. When the data is very large, we expect the I/O cost - the cost of moving data between main and secondary memory - to dominate, rather than processing cost. However, there is really no difference, for these algorithms, between processor and I/O costs. In justification, for a fixed amount of main memory, the number of passes over the data that we need is proportional to the number of DB passes times the space required in each pass. The reason is that we may process in parallel the passes for just as many items as can have their needed space in main memory simultaneously. If we look at Table 2, we see that the time required for each algorithm is n times the product of the space and DB passes. Thus, "time" is proportional to I/O cost for the algorithms we consider, moved-last-5yrs male support A never-married employed car/cab ? \Gamma0:0497 \Gamma0:0138 0:2672 householder 20-40K native-amer A / B / C 0:2205 \Gamma0:0111 \Gamma0:0537 military pay govt ? 0:1350 \Gamma0:0795 \Gamma0:5892 Table 3: Some of the 25 causal CCC relationships found in census data. The causal relationship is given when it can be disambiguated using a priori information. ae is the coefficient of correlation between the pair of items, and is positive when the two items are found often together, and negative when they are rarely found together. A and B each cause C ae AC ae BC black and nograd-HS each cause car/cab \Gamma0:0207 \Gamma0:1563 asian and laborer each cause !20K 0:0294 \Gamma0:0259 asian and laborer each cause 20-40K \Gamma0:0188 0:0641 employed and military each cause under-43 \Gamma0:0393 \Gamma0:2104 employed and military each cause never-married \Gamma0:0497 \Gamma0:0711 sales and householder each cause nograd-HS \Gamma0:0470 \Gamma0:0334 Table 4: Some of the 36 causal CCU relationships found in census data. The causal relationship is uniquely determined. ae is the coefficient of correlation between the pair of items. It is not given for the AB pair because this pair is uncorrelated. and henceforth we shall only consider processor time in our comparisons. 5 Experimental Results We use two data sets for our analysis, similar to the ones used by Brin, Motwani, and Silverstein [9]. One holds boolean census data (Section 5.1). The other data set is a collection of text data from UPI and Reuters newswires (Section 5.2). We actually study two newsgroup corpora, one of which is significantly larger than the other. In the experiments below, we used a chi-squared cutoff edges. We use the definition of support given by Brin, Motwani, and Silverstein [9]. All experiments were performed on a Pentium Pro with a 166 MHz processor running Solaris x86 2.5.1, with 96 Meg. of main memory. All algorithms were written in C and compiled using gcc 2.7.2.2 with the -O6 compilation option. 5.1 Census Data The census data set consists of binary items; it is a 5% random sample of the data collected in Washington state in the 1990 census. The items are listed in Appendix B. To simplify the pinterpretation of many items, we discarded responses of those under and over 60 years old. 5 Census data has categorical data that we divided into a number of boolean variables. Thus, for marital status we have several items: married, divorced, separated, widowed, never-married. Every individual has true for one of these variables and false for the rest. The test for CCU causality took 3 seconds of user CPU time to complete, while the test for CCC causality took 35 seconds of user CPU time. This indicates the census data has many more C edges than U edges, which is not surprising since all variables derived from the same census question are of necessity correlated. In Table 3 we show some of the results of finding CCC causality. Since several variables (such as male and under-43) cannot have causes, census data fits well into Cooper's LCD framework, and it is often possible to determine the direction of causation. Because of the possibility of confounding and hidden variables however, we cannot determine direct causality. The CCC test, however, allows us to rule out direct causality, and this in itself yields interesting results. Intuitively, an apparent relationship between two variables is explained by a third variable. For example, being in a support position is correlated with having moved in the past five years. This may lead one to believe that support personnel are unusually unlikely to move around. However, when we condition on being white, the apparent relationship goes away. From this, we can guess that being white causes one to move frequently, and also causes one not to be in support. Notice that in any case the correlation between support and moving is very weak, indicating this rule is not powerful. Another CCC rule shows that if people who are never married are less likely to drive to work, it is only because they are less likely to be employed. (The conditional is used here because we cannot be sure of the causal relationship: are the unmarried less likely to have jobs, or are the unemployed less likely to get married?) Table 4 shows some of the CCU causal relationships discovered on census data. While the causal relationship is uniquely determined, confounding and hidden variables keep us from determining if causality is direct. For instance, in the first row of Table 4, we can say not having graduated high school causes one not to drive to work, but we do not know if this is mediated by the fact high school dropouts are less likely to have jobs. A causal rule such as nograd-HS ! employed may exist, but if so neither the CCC nor CCU causality tests found it. As we see, these algorithms are better at exploratory analysis than hypothesis testing. Note that both CCC and CCU causality tests discovered a causal relationship between being employed and never having been married. The CCU result can be used to disambiguate among the possible causal relationships found from the CCC test. A danger of this if that if the CCU result is inaccurate, due to statistical error, using it to disambiguate the CCC result propagates the error. As it is, an improper uncorrelation judgment can cause many erroneous causal inferences. For instance, the uncorrelated edge sales-householder is the base of 10 CCU judgments, causing 5 For instance, one choice for marital status is, "Never married or under 15 years old," while the question regarding means of transportation to work conflates being unemployed and being under 16. causal inferences. If this edge was marked uncorrelated incorrectly, all 20 causal inferences are unjustified. In fact, our a priori knowledge would lead us to believe there is a correlation between being in sales and being the head of a household. Causal inferences based on this U-edge, such as the last entry in Table 4 are clearly false - dropping out of high school is temporally prior to getting a job or a house, and thus cannot be caused by them - leading us to question all causal inferences involving the sales-householder edge. 5.2 Text Data We analyzed 3056 news articles from the clari.world news hierarchy, gathered on 13 September 1996, comprising megabytes of text. For the text experiments, we considered each article to be a basket, and each word 6 to be an item. These transformations result in a data set that looks remarkably different from the census data: there are many more items than baskets, and each basket is sparse. To keep the number of items to a reasonable level, we considered only words that occurred in at least 10 articles. We also removed commonly occurring "stop words" such as "the," "you," and "much." We were left with 6723 distinct words. Since we have no a priori knowledge to distinguish between the possible causal models returned by the CCC algorithm, we ran only the CCU algorithm on the text data. The algorithm returned 73074 causal relationships. To study these, we sorted them by (the absolute value of) their correlation coefficient. We would expect the very top pairs to be obvious causal relationships, and indeed we see from Figure 5 that this is the case. To explore more interesting causal relationships, we also show some results from 5% down the list of correlations. Even the first set of causal relationships, along with its obvious relationships such as "united" causing "states," has some surprises. One is in the relationships "quoted" causes "saying," probably part of the set phrase, ". was quoted as saying ." Though this may not illuminate the content of the corpus, it does lend insight into the writing style of the news agency. Another interesting property is the frequency of causal relationships along with their converse. For instance, "prime" causes "minister" and "minister" causes "prime." The probable reason is that these words are usually found in a phrase and there is therefore a deterministic relationship between the words; that is, one is unlikely to occur in an article without the other. When words are strongly correlated but not part of a phrase - "iraqi" and "iraq" are an example - then we only see the causal relationship in one direction. This observation suggests a somewhat surprising use of causality for phrase detection. If words that always occur together do so only as part of a phrase, then we can detect phrases even without using word location information, just by looking for two-way causality. Presumably, incorporating this strategy along with conventional methods of phrase detection would only improve the quality of phrase identification. The causal relationships at the 5% level are as intriguing as those at the top of the list. The causal relationship "infiltration" ! "iraqi" points to an issue that may bear further study. Other 6 A word was defined to be a series of alphanumeric characters. However, we allowed a word to have a single internal punctuation mark, which we removed. We also lower-cased all words. Thus the sentence "It'll cost you one-no, two-pence for that pince-nez" has the words itll, cost, you, one, no, two, pence, for, that, and pincenez. Causal relationships - 2 value ae Causal relationships - 2 value ae united ! states 1691:0389 0:7439 forces ! company 70:5456 \Gamma0:1519 states ! united 1691:0389 0:7439 company ! forces 70:5456 \Gamma0:1519 prime prime 1288:8601 0:6494 commitment ! peace 70:2756 0:1516 quoted ! saying 866:6014 0:5325 british ! perry 70:2082 0:1516 news ! agency 718:1454 0:4848 support ! states 70:1291 0:1515 agency ! news 718:1454 0:4848 states ! support 70:1291 0:1515 Table 5: Causal relationships from the top and 5% mark of the list of causal relationships for words in the clari.world news hierarchy. The list is sorted by (the absolute value of the) correlation coefficient (last column). The - 2 value measures the confidence that there is a causal relationship; all these - 2 values indicate a probability of error of less than 0:0001. The ae value measures the power of the causality. relationships, such as "saturday" ! "state," seem merely bizarre. Note also how quickly the correlation coefficients drop for this data set; only 2:3% of the causal relationships have ae ? 0:2. A final note about the causal relationships is they seem to disproportionately concern Iraq. While Iraq was much in the news during September 1996, the news hierarchy includes articles concerning all parts of the world, from Latin America to Africa to Oceania. However, there were a diffuse set of issues relating to these areas, each with its own vocabulary. Only Iraq seemed to have the large number of cohesive articles necessary to pass the support and confidence thresholds. Whether this is a feature of the causal algorithm 7 or a bug depends on the individual application. 5.3 Comparing Causality with Correlation A question naturally arises: what is the advantage of causal discovery over merely ranking correlated item pairs? In Figure 6 we show the top 10 correlations, as measured by the correlation coefficient. These results are directly comparable to the top portion of Figure 5. Two difference are immediately noticeable. One is the new item pairs. Some of them, like "iraq" and "warplanes," seem like significant additions. Others, like "hussein" and "northern," have plausible explanations (the U.S. flight zone is restricted to the Northern and Southern bands of the Iraq) but do not seem to belong so high on the list; other, more perspicuous, causal relationships have a lower correlation coefficient. The other noticeable difference is that, since correlation is symmetric, there is no case of a pair and its converse both occurring. Insofar as asymmetric causalities yield extra understanding of the data set, identifying causal relationships yields an advantage over identifying correlations. A third difference, not noticeable in the figure, is that there are many more correlation rules 7 Actually, since support and confidence are concepts used to identify C- and U-edges, this is a concern not only for causal discovery but for any methodology based on a similar definition of correlation. reuter - upi 2467:2895 \Gamma0:8985 states - united 1691:0389 0:7439 minister - prime 1288:8601 0:6494 quoted - saying 866:6014 0:5325 agency - news 718:1454 0:4848 hussein - northern 678:5580 0:4712 Table Correlations from the top of the list of correlations, sorted by (the absolute value of the) correlation coefficient (last column). This list is a superset of the list of top causal relationships Figure 5). The - 2 value measures the confidence that there is a correlation; all these - 2 values indicate a probability of error of less than 0:0001. The ae value measures the power of the causality. than causal rules. While there are around 70 thousand causal relationships in this data set, there are 200 thousand correlated pairs. 5.4 Performance on a Large Text Data Set While the clari.world hierarchy is somewhat homogeneous, which as we have seen is an advantage when pruning, at megabytes it is a small dataset. We therefore repeated the text experiments on the entire clari hierarchy, a larger, more heterogeneous news hierarchy that covers sports, business, and technology along with regional, national, and international news. This data set was gathered on 5 September 1997. While clari.world is logically a subtree of the clari hierarchy, the clari.world database is not a subset of the clari database since the articles were collected on different days. The clari data set consists of 27803 articles and 186 megabytes of text, and is thus ten times larger than the clari.world data set. However, the number of items was kept about the same - the larger data set, at 6303 items, actually has fewer items than the clari.world data set - due to the pruning of infrequent words. In both cases, words were pruned that occurred in fewer than 0:3% of all documents; for the clari data set this worked out to an 84 document minimum. As in the smaller data set, the terms with the highest correlation come from a coherent subset of documents from the collection. Unfortunately, the coherent subset of the clari collection is a large mass of Government postings soliciting bids for automotive supplies, and most of the words used in these posts are abbreviations of technical terms. Thus, the top causalities are "recd" causes "solnbr" and "desc" causes "solnbr"; we must go down 25 lines before we find causality involving two English words. Even then, the relationship - "office" causes "contact" - involves words from the procurement articles. The causal relationships found 5% down the list are a little more interesting. Some are shown Causal relationships - 2 value ae cause ! company 558:2142 0:1417 constitutes ! number 557:8370 0:1416 modification today people ! update 556:6686 0:1415 28 556:1250 0:1414 Table 7: Causal relationships from the list of causal relationships for words in the clari news hierarchy, starting from 5% down the list when the list is sorted by (the absolute value of the) correlation coefficient (last column). The - 2 value measures the confidence that there is a causal all these - 2 values indicate a probability of error of less than 0:0001. The ae value measures the power of the causality. in Figure 7. Note that, as in the smaller text data set, the correlation coefficient rapidly becomes rather low. 6 Conclusion and Further Research In data mining context, constraint-based approaches promise to find causal relationships with the efficiency needed for the large data sets involved. The size of the data mining data sets mitigate some of the weaknesses of constraint-based approaches, namely that they sometimes need large amounts of data in order to make causal judgments, and instead of finding all causal relationships, they only find a subset of these relationships. For data mining, which seeks to explore data rather than to test a hypothesis, finding only a portion of the causal relationships is acceptable. Furthermore, with large data sets, the number of actual causal relationships is likely to be so large that algorithms that return only a small portion of the relationships are likely to yield interesting results. Another weakness of constraint-based algorithms, the error inherent in repeated use of statistical tests, is mitigated in boolean data by using a power statistic to reduce the probability of error without discarding powerful causal relationships. We developed a series of algorithms, based on techniques used in Cooper's LCD algorithm, that run in time linear in the size of the database and cubic in the number of variables. For large data sets with thousands of variables, these algorithms proved feasible and returned a large number of causal relationships and, equally interesting, not-directly-causal relationships. This feasibility came from heuristics and algorithmic choices that improved on both the time and memory requirements of the naive cubic-time algorithm. Finding causal relationships is useful for a variety of reasons. One is that it can help in visualizing relationships among variables. Another is that, unlike correlation, causation is an asymmetric concept. In contexts where it is possible to intervene on the variables (for instance, when a manager can choose not to stock certain food items) causality can help predict the effect of the intervention, whereas a correlation analysis cannot. In the context of text analysis, causation can help identify phrases. There are still a variety of unresolved and unexplored issues in the area of mining for causal relationships. We briefly list some of them below. Choosing Thresholds Is there a way to determine optimal values for the correlation and uncor- relation cutoffs for a given data set? Better yet, is it possible to replace cutoff values with estimates of the probability of correlation efficiently? Disambiguation We mentioned in Section 5.1 that, at the risk of propagating error, we can use known causal rules to disambiguate the CCC causality rule. However, we have seen that both may occur in data. Is there a principled way to resolve bidirectional causality for disambiguation? Can we then devise efficient algorithms for using incremental causal information to perform disambiguation? Hidden Variables Bidirectional causality may indicate deterministic relationships (as in text phrases), error in statistical tests, or the presence of hidden variables. Under what situations can we be confident hidden variables are the cause? What other techniques can we use to discover hidden variables? Heuristics for Efficiency How can we make the above algorithm even more efficient? The largest speedups could be obtained by avoiding the need to check all triples. Can we determine when the conditional independence test will fail on triples without explicitly testing them? Can we reduce the number of items, perhaps by collapsing items with similar distributions? Acknowledgments We would like to thank the members of the Stanford Data Mining research group, particularly Lise Getoor, for their useful comments and suggestions. We would also like to thank Greg Cooper and David Heckerman for fruitful discussions. --R Inferring structure in semistructured data. Mining association rules between sets of items in large databases. Database mining: a performance perspective. Fast discovery of association rules. Fast algorithms for mining association rules in large databases. Categorical Data Analysis. A Survey of exact inference for contingency tables. Probabilistic evaluation of counterfactual queries. Beyond market baskets: Generalizing association rules to correlations. Dynamic itemset counting and implication rules for market basket data. A simple constraint-based algorithm for efficiently mining observational databases for causal relationships A Bayesian method for the Induction of Probabilistic Networks from Data. Advances in Knowledge Discovery and Data Mining. A Bayesian approach to learning causal networks. Bayesian networks for data mining. Learning Bayesian networks: The combination of knowledge and statistical data. A Bayesian approach to causal discovery. A definition and graphical representation of causality. Mathematical Statistics with Applications. From Bayesian networks to causal networks. Causal diagrams for empirical research. Graphical models for probabilistic and causal reasoning. A theory of inferred causation. An algorithm for fast recovery of sparse causal graphs. Causal inference in the presence of latent variables and selection bias. Mining generalized association rules. Sampling large databases for finding association rules. Correlation and causation. --TR --CTR Stefano Ceri , Francesco Di Giunta , Pier Luca Lanzi, Mining constraint violations, ACM Transactions on Database Systems (TODS), v.32 n.1, p.6-es, March 2007 Pedro Domingos , Matt Richardson, Mining the network value of customers, Proceedings of the seventh ACM SIGKDD international conference on Knowledge discovery and data mining, p.57-66, August 26-29, 2001, San Francisco, California Zhiping Zeng , Jianyong Wang , Lizhu Zhou , George Karypis, Out-of-core coherent closed quasi-clique mining from large dense graph databases, ACM Transactions on Database Systems (TODS), v.32 n.2, p.13-es, June 2007 Man Leung Wong , Shing Yan Lee , Kwong Sak Leung, Data mining of Bayesian networks using cooperative coevolution, Decision Support Systems, v.38 n.3, p.451-472, December 2004 Ling Feng , Jeffrey Xu Yu , Hongjun Lu , Jiawei Han, A template model for multidimensional inter-transactional association rules, The VLDB Journal The International Journal on Very Large Data Bases, v.11 n.2, p.153-175, October 2002 Qing Li , Ling Feng , Allan Wong, From intra-transaction to generalized inter-transaction: landscaping multidimensional contexts in association rule mining, Information SciencesInformatics and Computer Science: An International Journal, v.172 n.3-4, p.361-395, 9 June 2005 Ioannis Tsamardinos , Laura E. Brown , Constantin F. Aliferis, The max-min hill-climbing Bayesian network structure learning algorithm, Machine Learning, v.65 n.1, p.31-78, October 2006
data mining;association rules;text mining;causality;market basket
593514
Scaling Kernel-Based Systems to Large Data Sets.
In the form of the support vector machine and Gaussian processes, kernel-based systems are currently very popular approaches to supervised learning. Unfortunately, the computational load for training kernel-based systems increases drastically with the size of the training data set, such that these systems are not ideal candidates for applications with large data sets. Nevertheless, research in this direction is very active. In this paper, I review some of the current approaches toward scaling kernel-based systems to large data sets.
Introduction Kernel-based systems such as the support vector machine (SVM) and Gaussian processes (GP) are powerful and currently very popular approaches to supervised learning. Kernel-based systems have demonstrated very competitive performance on several applications and data sets. Kernel-based systems also have great potential for KDD appli- cations, since their degrees of freedom grow with training data size, and they are therefore capable of modeling an increasing amount of detail with increasing data set size. Unfortunately, there are at least three problems when one tries to scale up these systems to large data sets. First, training time increases dramatically with the size of the training data set; second, memory requirements increase with data; and third, prediction time is proportional to the number of kernels, which is equal to (or at least increases with) the size of the training data set. A number of researchers have approached these issues and have developed solutions for scaling kernel-based systems to large data sets. I will review these approaches in this paper, where focus will be on the computational complexity of training. The paper is organized as follows. Section 2 is a brief introduction to kernel-based systems and implementation issues. Section 3 presents various approaches to scaling kernel-based systems to large data sets. First, in Section 3.1, I will present approaches based on the idea of com- 2001 Kluwer Academic Publishers. Printed in the Netherlands. Volker Tresp mittee machines. Section 3.2 covers data compression approaches. Here, the kernel systems are applied to compressed or sub-sampled versions of the original training data set. Section 3.3 discusses e-cient methods for the linear SVM based on modications of the optimization problem to be solved. Some training regimes for kernel-based systems require the solution of a system of linear equations the size of the training data set. Section 3.4 presents various approaches to e-ciently solving this system of equations. In Section 3.5, projection methods are discussed. The solutions here are based on a nite-dimensional approximation to the kernel system. Section 4 provides conclusions. 2. Kernel-Based Systems for Supervised Learning 2.1. Kernel-based System In kernel-based systems, the response ^ f(x) to an input x is calculated as a superposition of kernel functions K(x; Here, w i is the weight on the i-th kernel. Kernels are either dened for all input patterns D or for a subset of the training data, as in the case of the SVM. Sometimes a bias term b is included, and In regression, ^ corresponds to an estimate of the regression function, and in (binary) SVM classication, the estimated class label is the sign of ^ In some cases, the kernel functions K(x; are required to be positive denite func- tions. Typical examples are linear kernels K(x; x positive integer q, and Gaussian kernels are a special case, since here the equivalent linear representation with weight vector computationally more e-cient if the input dimension is smaller than the size of the training data set. The rows of the design matrix A are the input vectors of the training data set. Kernel-based systems dier in how the weights w i are determined based on the training data. Scaling Kernel-Based Systems 3 2.2. Gaussian Process Regression (GPR) In GPR, one assumes a priori that a function f(x) is generated from an innite-dimensional Gaussian distribution with zero mean and covariance dened at input points x i and x j . Furthermore, we assume a set of training data where targets are generated according to where i is independent additive Gaussian noise with variance 2 . The optimal regression function ^ f(x) takes on the form of Equation 1, where the kernel is determined by the covariance function. Based on our assumptions, the maximum a posterior (MAP) solution for the cost function2 w (w y) 0 (w y) (4) where , with is the N N Gram ma- trix. 1 The optimal weight vector is the solution to a system of linear equations, which in matrix form becomes Here is the vector of targets and I is the N N-dimensional unit matrix. MacKay (1997) provides an excellent introduction to Gaussian processes. 2.3. Generalized Gaussian Process Regression (GGPR) Gaussian processes can nd a wider range of applications by permitting a more exible measurement process. In particular, we might assume that y i 2 f0; 1g is the class label of the i-th pattern and that is the posterior probability for Class 1. The corresponding cost function log be the the values of f at the location of the training data. Using this identity, the cost function can also be written as2 (f m 4 Volker Tresp In this case, the optimal weights are found iteratively using an iterated re-weighted least squares algorithm. The approach can be generalized to any density from the exponential family of distributions. Discussion of GGPR can be found in Tresp (2000b), Williams and Barber (1998), and Fahrmeir and Tutz (1994). 2.4. Support Vector Machine (SVM) The SVM classies a pattern according to the sign of Equation 2. The most commonly used SVM cost function is2 w Here, D is an NN-dimensional diagonal matrix with ones and minus ones on the diagonal (corresponding to the class label of the respective patterns), e is an N-dimensional vector of ones, C 0 is a constant negative components of the vector within the bracket to zero. Note the similarity between the SVM and the GGPR cost function. The dierence is that the measurement process for the SVM is not derived from the exponential family of distributions as it is for the GGPR. More common is the following equivalent denition with a dieren- tiable cost function and constraints. The weights in the SVM minimize the cost function 1 are the slack variables with i 0 subject to the constraint that Here and in the rest of the paper is meant to be applied component- wise. By introducing an N-dimensional vector of Lagrange multipliers for enforcing the constraints, one obtains the dual problem: Maximize with the constraints The weights w and the Lagrange multipliers are related via The optimal weight vector is sparse | that is the weight w i for many kernels is zero after training. The remaining kernels with nonzero weights dene the support vectors. It can be shown that the solution minimizes a bound on the generalization error (Vapnik, 1998). Vapnik Scaling Kernel-Based Systems 5 (1998), Scholkopf, Burges and Smola (1999), Christianini and Shawe-Taylor (2000), and Muller, Mika, Ratsch, Tsuda, and Scholkopf (2001) are excellent sources of information about the SVM. 2.5. Training Solving the system of linear equations (Equation 5) for GPR requires operations. The iterated re-weighted least squares algorithm used for training the GGPR typically requires that systems of similar linear equations must be solved repeatedly. For the support vector machine, the dual optimization problem requires the solution of a quadratic programming (QP) problem with linear constraints. Due to the large dimension of the optimization problem, general QP routines are unsuitable even for relatively small problems (a few hundred data points): the common interior point methods require repeated solution of linear equations with the dimensionality of the number of variables. To solve the high-dimensional QP problems for the SVM, special algorithms have been developed. The most important ones are chunk- ing, decomposition, and sequential minimal optimization (SMO). All methods iteratively solve smaller QP problems. Both chunking and decomposition require a QP solver in their inner loops. Decomposition scales better with the size of the training data set since the dimensionality of the QP problem is xed, whereas for the chunking algorithm, this dimensionality increases until it reaches the number of support vectors. Chunking was one of the earliest approaches used for optimizing the SVM machine. Decomposition was introduced by Osuna, Freund, and Girosi (1997) and further developed by Joachims (1998). The SMO is an extreme case of a decomposition algorithm since only two variables (Lagrange multipliers) are optimized at a time. This can be done ana- lytically, so that the SMO does not require a QP optimizer in its inner loop. Decomposition and the SMO are in general faster than chunking. Experimental evidence suggests that the computational complexity for both decomposition and the SMO scales approximately to the square of size of the training data set. As an alternative to QP, several gradient optimization routines have been developed. In addition, Perez-Cruz, Alarcon-Diana, Navia-Vazquez, and Artes-Rodr iguez (2001) used a fast iterated re-weighted least squares procedure recently to train the SVM. 6 Volker Tresp 2.6. More Kernels: Regularization Networks, Smoothing Splines, the Relevance Vector Machine and Kernel Fisher Discriminant It is not possible to discuss all currently popular kernel-based learning systems in detail in the limited space allowed here. The regularization networks of Poggio and Girosi (1990) are essentially identical to GPR. Here, the kernels are Green's functions derived from the appropriate regularization problem. Similarly, smoothing splines are closely related (Wahba, 1990). The relevance vector machine (Tipping, 2000) achieves sparseness by pruning away dimensions of the weight vector using an evidence framework. Finally, the kernel Fisher discriminant is the well-known linear Fisher discriminant approach transformed into a high-dimensional feature space by means of kernel functions (Muller et al., 2001). 3. Scaling Kernel-Based Systems to Large Data Sets This section is the heart of the paper. In it, I discuss various approaches toward scaling kernel-based systems to large data sets. 3.1. Committee Machines 3.1.1. Introduction In committee machines, dierent data sets are assigned to dierent committee members (i.e., kernel systems) for training, and the predictions of the committee members are combined to form the prediction of the committee. Here, I want to discuss two committee approaches that are particularly suitable for (and have been applied to) kernel-based systems. 3.1.2. A Bayesian Committee Machine (BCM) In the BCM approach (Tresp, 2000a), the data are partitioned into M data sets (of approximately same size) and M learning systems are trained on their respective training data set. The BCM calculates the unknown responses at a number of test points at the same time. Let f NQ ) be the vector of these response variables at the NQ test points. The underlying assumption of the BCM is that Under this assumption, data sets are independent given f q . This is a good assumption if f q contains many points, since these points dene Scaling Kernel-Based Systems 7 the map and make all data independent. The approximation also improves if the number of data in each set D i is large, which increases their independence from the other data sets on average. Based on the assumption, one obtains A practical algorithm is obtained if we assume that all probability distributions are approximately Gaussian. Then P (f q jD) is also approximately Gaussian with E(f q and Here, qq is the NQ NQ prior covariance matrix at the test points. Equation 8 has the form of a committee machine where the predictions of the committee members at all NQ inputs are used to form the prediction of the committee at those inputs. The prediction of each module i is weighted by the inverse covariance of its prediction. An intuitive appealing eect is that by weighting the predictions of the committee members by the inverse covariance, modules that are uncertain about their predictions are automatically weighted less than modules that are certain about their predictions. I applied the BCM to GPR in Tresp (2000). For GPR, the mean and the covariance of the posterior Gaussian densities are readily computed. Subsequently, the BCM was applied to GGPR (Tresp, 2000b) and to the SVM (Schwaighofer and Tresp, 2001). For both the GGPR and the SVM, the posterior distributions are only approximately Gaussian. In Tresp (2000), it was shown that NQ , the dimension of f q , should be at least as large as the eective number of parameters. 2 If this is the case, the BCM and its generalizations give excellent results. furthermore, it is possible to derive online Kalman lter versions of the BCM, which only require one pass through the data set and the storage of a matrix of the dimension of the number of test points (Tresp, 2000a). After training, the prediction at additional test points requires resources dependent on the number of test points but is independent 2 Since the latter is closely related to the kernel bandwidth, there is also a close connection between the kernel bandwidth and NQ . 8 Volker Tresp of the size of the training data set. On several data sets I found no dierence in performance between the BCM approximation and the optimal GPR prediction based on the inversion of the full covariance matrix. The BCM was applied to training data sets of size where the full inversion is clearly unfeasible. 3.1.3. Boosting In boosting, committee members are trained sequentially and the training of a particular committee member is dependent on the training and the performance of previously trained members. Boosting can reduce both variance and bias in the prediction. The reason is that in training a committee member, more weight is put on data that are misclassied by previously trained committee members. Schapire (1990) developed the original boosting approach, boosting by ltering. Here, three learning systems are used and the existence of an oracle that can produce an arbitrary quantity of training data is assumed. The rst learning system is trained on K training data. Then the second learning system is trained on the same quantity of data but these training data are generated so that half of them are classied correctly and half of them are classied incorrectly by the rst learning system. The third learning system is trained only on data about which learning systems one and two disagreed. A majority vote of the three learning systems determines the classication. Note that the second learning system obtains 50% of patterns for training that are di-cult for the rst learning system, and that the third learning system only obtains critical patterns in the sense that learning the rst and second learning systems disagree on those patterns. This original boosting algorithm is particularly useful for large data sets, since a large number of training data are ltered out and are not directly used for training. In recent years, the focus of interest has shifted toward boosting algorithms that can also be applied to smaller data sets: boosting by resampling and boosting by reweighting. Recently, Pavlov, Mao, and Dom (2000) used a form of boosting by resampling, Boost- SMO, in the context of training the SVM with large data sets. They used a small fraction of data (2 { 4%) to train a committee member using the SMO algorithm. To train a subsequent committee member, they chose those data with a higher probability which were di-cult for the previous committee member to classify. In this way, only a portion of the complete training data set is used for training. The overall prediction is then a weighted combination of the predictions of the committee members. In some experiments, Boost-SMO was faster by a factor of 10 than SMO while providing essentially the same prediction accuracy. In their implementation, the probability of the data are reweighted prior Scaling Kernel-Based Systems 9 to the training of a new committee member, so multiple passes through the data are required. Nevertheless, online versions of this approach are also conceivable. 3.2. Data Compression Data compression is a generally applicable solution to dealing with large data sets. The idea is to train the learning system on a smaller data set that is either generated by subsampling or by preclustering the data. In the latter case, the cluster centers are used as training patterns. The idea of preclustering data has been extended in an interesting direction by applying the concept of squashing to training a linear SVM (Pavlov, Chudova, and Smyth, 2000). For training, the SMO algorithm is used. Clustering is performed using a metric derived from the likelihood prole of the data. First, a small percentage of the original training data set are randomly chosen as cluster centers. Then, for a linear SVM, (typically 50{100) random weights v and an oset b are generated following their prior distributions (a probabilistic version of the SVM is used). For each data point the log-likelihood for each weight vector is calculated, producing an L-dimensional vector (likelihood prole) for each data point. A data point is then assigned to the cluster center with the closest likelihood prole. Finally, the weight for each cluster center is proportional to the number of data assigned to a cluster. This procedure leads to a considerable reduction in training data by taking into account the statistical properties of the data. The training time using the SMO with squashing was comparable to the training time of Boost-SMO (see previous section) by providing a comparable prediction accuracy. 3.3. Fast Algorithms for Linear SVMs The SVM was originally formulated as a linear classier; kernels were introduced in order to be able to obtain nonlinear classication bound- aries. Training the linear SVM is considerably faster than training the kernel SVM. The following section discusses approaches that lead to even faster training algorithms for the linear SVM by modifying the cost function. 3.3.1. Active Support Vector Machine (ASVM) The ASVM was developed by Mangasarian and Musicant (2001). Here, the optimization problem of the SVM is reformulated. The modied cost function is2 (v with constraints where A is the design matrix. In the design matrix, the input vectors of the training data set are represented as rows. Note that the cost function contains the square of the bias b such that the margins with respect to both orientation v and location relative to the origin b are maximized. furthermore, the 2-norm of the slack vector is minimized. Based on these modications, a dual problem is formulated that contains non-negativity constraints but no equality constraints. Due to these modications, a very simple iterative optimization algorithm was derived. At each iteration step, a system of linear equations the size of the input dimension plus one needs to be solved. The number of iteration steps is nite. As an example, a data set with 7 million points required only 5 iterations and needed 95 CPU minutes. 3.3.2. Lagrange Support Vector Machine (LSVM) A variation on the ASVM is the LSVM (Mangasarian and Musicant, 2000). It is based on the same reformulation of the optimization problem and leads to the same dual problem. The dierence is that the LSVM works directly with the Karush-Kuhn-Tucker necessary and su-cient optimality condition for the dual problem. The algorithm re- quires, prior to the optimization iterations, the inversion of one matrix Q of the size of the input dimension plus one. The iteration has the simple form where i is the vector of Lagrange multipliers at step i, and is a positive constant. The LSVM and the ASVM are comparable in speed although the ASVM is faster on some problems. The great advantage of the LSVM is denitely the simplicity of the algorithm. The LSVM can also be applied to nonlinear kernels, but a matrix of the size of the number of data points needs to be inverted. For more examples of fast linear SVM-variants, see Mangasarian and Musicant (1999). 3.4. Approximate Solutions to Systems of Linear Equations Gaussian processes and some variants of the SVM (see Section 3.3.2) require the solution of a large system of linear equations. Scaling Kernel-Based Systems 11 3.4.1. Method by Skilling For Gaussian processes and also for some variants of the SVM, it is necessary to solve a linear system of equation of the form with some matrix M . The solution to this linear set of equations is identical to the minimum of This cost function is minimized iteratively using the conjugate gradient method in O(k N 2 ) steps where k is the number of iterations of the conjugate gradient procedure. Often k can be set to be much smaller than N without a signicant loss in performance, particularly when M has a large number of small eigenvalues. This approach is due to Skilling and was one of the earliest approaches to speeding up the training of GPR systems (Gibbs and MacKay, 1997). Note that the computational complexity of this approach is quadratic in the size of training data set and this approach is therefore not well suited for massive data sets. 3.4.2. The Nystrom Method The Nystrom method was introduced by Williams and Seeger (2001), and it is applicable in particular to GPR. Let's assume a decomposition of the Gram matrix of the form = UU 0 where is a diagonal matrix and U is an N m matrix with typically m << N . In this case, we can solve the system of linear equations (see Equation 5) using the Woodbury formula (Press, Teukolsky, Vetterling, and Flannery, 1992) and obtain In this form, only a matrix of size mm needs to be inverted, and w still has the dimension of the number of training data N . An example of an appropriate decomposition of the Gram matrix would be an eigenvalue decomposition. The computational complexity of a full eigenvalue decomposition scales as O(N 3 ), so not much is gained unless the Gram matrix has a large number of small eigenvalues. A particularly interesting approximation was introduced by (Williams and Seeger, 2001). They performed an eigendecomposition of a (ran- domly chosen) m m submatrix of . Based on p eigenvectors and eigenvalues of the decomposition of the smaller matrix, the corresponding decomposition of the larger matrix can be approximated using the Nystrom method. The Nystrom method is a method for numerically solving integral equations. The computational complexity of this approach is O(m 2 N) such that this approach scales linear in N . The authors veried the excellent quality of this approximate method using a training set size of 7291 and obtained very good results with The approach is applicable to both regression and classication. 3.5. Projection Methods Leading to a Finite-Dimensional Representation In general, the degrees of freedom of a kernel system grow with the number of kernels (i.e., data points). The approaches discussed in this section project the basically innite-dimensional problem on a nite- dimensional representation. The problem then reduces to estimating the parameters in this nite-dimensional problem. In all approaches, the solution assumes the form of a system with xed basis functions. The BCM approach in Section 3.1.2 can also be considered to be a specic projection approach in which the basis functions are the kernels dened by the test points. 3.5.1. Optimal Projections The problem formulation is: Given that both the input data distributions P (x) and the size of training data N are known, what is the best linear combination of an m-dimensional set of basis functions that provides the best approximation to a GPR kernel-based system. In projected Bayes regression (Zhu, Williams, Rohwer, Morciniec, 1998), this problem is solved based on an innite-dimensional principle component analysis for N !1. The result is that one should use the rst m eigenfunctions of the covariance function describing the Gaussian process asymptotically. The computational complexity is O(Nm 2 ). Trecate, Williams, and Opper (1998) described an improved variational approximation for nite data size. Their approach has a smaller test set error if compared to projected Bayes regression, and the computational complexity scales as O(N 2 m). This approach is quadratic in N , whereas projected Bayes regression is linear in N , because the representation increases with data size. Csato and Opper (2001) presented an online variant based on these ideas, which leads to a sparse representation by limiting the representation to a small number of well chosen kernel functions. 3.5.2. Projection on a Subset of Kernels Let's select a subset of the training data set of size m < N and let m;m denote the corresponding kernel matrix. We can now approximate Scaling Kernel-Based Systems 13 is the vector of covariances between the functional values at x i and the data in the subset. This approximation is an equality if either x i or x j are elements of the subset and is an approximation otherwise. With this approximation, the regression function is a superposition of only m kernel functions and the optimal weight vector minimizes the cost where N;m contains the covariance terms between all N training data and the subset of data chosen to be kernel functions. Although we only used a xed number of kernels, all training data contributed to determine the weight vector w. The methods described in the following subsections are based on this decomposition. The connection between the decomposition of the Gram matrix in this section and the BCM approximation is discussed in the Appendix. 3.5.3. Reduced Support Vector Machines (RSVM) The RSVM (Lee and Mangasarian, 2000) uses a nonstandard SVM cost function of the form2 (w If we compare this equation with the original SVM cost function of Equation 6, we notice that the cost term for the weights is simplied. As in the previous section, w denotes the kernel weights for m randomly selected kernels and, as in the case of the ASVM (Section 3.3.1), the square of the bias b is included. Here, the 2-norm of a transformed slack variable is included with where N;m is dened as in the previous subsection and where log(1 +exp( x)) is applied component-wise. The function g() is a \soft" twice-dierentiable version of is typically a large positive number. The advantages of these modications are that (a) no constraints are needed (as in the formulation in Equation 6); (b) due to the modications of the cost function, a quadratically converging Newton algorithm can be used for training; and (c) due to the projection on the nite number of kernels, the time complexity of the optimization routine scales linearly in the number of data points. In an 14 Volker Tresp experiment using the adult data set 3 with N=32600 training data and the RSVM needed 16 minutes, whereas the SMO algorithm needed more than two hours. Surprisingly, on some smaller data sets the RSVM performed even better than the full SVM. This was explained by a smaller tendency toward overtting of the RSVM. Experimental results showed that the random selection of the training data did not signicantly increase the variance in the prediction. 3.5.4. Sparse Greedy Matrix Approximation In the previous two subsections an expansion in terms of a nite number of kernels dened at a random subset of the training data was sought. In contrast, in the BCM approximation the kernels are dened by the test points. In Smola and Scholkopf (2000), the expansion is based on a sub-set of the training data, but here, the kernels are not selected randomly: The goal is to nd the m kernels that can best represent the N kernels of all training data where proximity is dened in the Reproducing Kernel Hilbert Space (i.e., the feature space). This simplies calculations drastically since the inner product of two kernels dened for x i and x j is simply equal to K(x The authors describe a \greedy" algorithm in which at each step the best kernel out of randomly selected candidate kernels is added to the set of already selected kernels. The computational cost of this version is O(L m N 2 ), where N is the total number of training data and m is the size of the selected subset of training data. In Smola and Scholkopf's paper, motivated by theoretical considerations. The authors showed that the approximation based on m kernels is very close to the approximation based on the optimal m basis vectors. The optimal m should also be closely related to the eective number of parameters, as discussed in Section 3.1.2, and is therefore dependent on the kernel bandwidth. A variant of this approach applicable to GPR is described in Smola and Bartlett (2001). In their paper, they derived a stopping criterion and bounds on the approximation error. Using a training data set of size the authors demonstrated that, with less than 10% of the training data used as kernels, they obtained statistically insignicant dierences in performance between GPR based on the inversion of the full covariance matrix and their approximation. 4. Conclusions I have summarized the most important approaches for scaling up kernel-based systems to large data sets. For nonlinear kernels, various authors 3 Retrievable from: http://www.ics.uci.edu/ mlearn. Scaling Kernel-Based Systems 15 have achieved a considerable reduction in training time which makes nonlinear kernel systems applicable to data sets of maybe a few 100,000 data points. All approaches assume that a representation based on a nite subset of the training data (respectively the kernels) is su-cient which is a reasonable assumption if the kernel bandwidth is low. Nev- ertheless, the goal stated in the introduction, of being able to model an increasing amount of detail when su-cient data become available would require that the kernel bandwidth be scaled up with increasing data size, thus increasing the degrees of freedom appropriately. This is only possible to a limited degree with the methods presented. Combining kernel-based systems with a hierarchical partitioning of the map or local algorithms might be an interesting direction for future research. Finally, for kernel systems using linear kernels, training time is considerably faster than for kernel systems using nonlinear kernels. Here, systems with more than a few million data points have been trained witin a reasonable time. Appendix Let's assume that one is interested in the prediction at a set of query points, dened as a subset of the training data as in Section 3.5.2 or as the test data, as in Section 3.1.2. Let f q denote the unknown functional values at those query points. Let f w be the expansion in terms of the kernel functions dened at the query points. Here, q;q is the covariance matrix dened for this subset of points. Then, the weight vector which leads to the optimal predictions at the subset of points Here, is the covariance of the training data given f q . Also, N;q is the covariance between the training data and the query points and N;N is the covariance matrix dened for the training data. Note that the inverse of cov(yjf q ) needs to be calculated which is an N N matrix. The approximation used in Section 3.5.2 simply sets cov(yjf q I where I is the unit matrix. The BCM uses a block diagonal approximation of cov(yjf q ), and the calculation of the optimal weight vector w requires the inversion of matrices of the block size. The approximation improves if few blocks are used (then a smaller number of elements are set zero) and when the dimension of f q is large, Volker Tresp since then the two terms on the right side of Equation 11 cancel and more detailed discussion can be found in Tresp and Schwaighofer (2001). Acknowledgements Salvatore Ingrassia from the Universita della Calabria and Stefano Ricci from the Universita di Pavia provided me with extensive comments on an earlier version of this paper. Their comments helped improve the paper considerably. In addition, valuable discussions with Alex Smola, Chris Williams, John Platt, Lehel Csato, and Anton Schwaighofer are gratefully acknowledged. --R Support vector machines. Multivariate statistical modeling based on generalized linear models. Making large-scale support vector machine learning practical RSVM: reduced support vector ma- chines Introduction to Gaussian processes. Massive support vector regression. Lagrangian support vector machine. Active support vector machine classi Scaling support vector machines using boosting algorithm. Towards scalable support vector machines using squashing. Fast training of support vector classi Fast training of support vector machines using sequential minimal optimization. Networks for approximation and learning. Numerical recipes in C. The strength of weak learnability. The Bayesian committee support vector machine. Sparse greedy Gaussian process regression. The relevance vector machine. Scalable kernel systems. Statistical learning theory. Spline models for observational data. Bayesian classi Gaussian regression and optimal --TR --CTR Paul Bradley , Johannes Gehrke , Raghu Ramakrishnan , Ramakrishnan Srikant, Scaling mining algorithms to large databases, Communications of the ACM, v.45 n.8, August 2002 Navneet Panda , Edward Y. Chang , Gang Wu, Concept boundary detection for speeding up SVMs, Proceedings of the 23rd international conference on Machine learning, p.681-688, June 25-29, 2006, Pittsburgh, Pennsylvania Daniel Schneega , Steffen Udluft , Thomas Martinetz, Kernel rewards regression: an information efficient batch policy iteration approach, Proceedings of the 24th IASTED international conference on Artificial intelligence and applications, p.428-433, February 13-16, 2006, Innsbruck, Austria Jian-Tao Sun , Ben-Yu Zhang , Zheng Chen , Yu-Chang Lu , Chun-Yi Shi , Wei-Ying Ma, GE-CKO: A Method to Optimize Composite Kernels for Web Page Classification, Proceedings of the 2004 IEEE/WIC/ACM International Conference on Web Intelligence, p.299-305, September 20-24, 2004 Ying Lu , Jiawei Han, Cancer classification using gene expression data, Information Systems, v.28 n.4, p.243-268, June
gaussian processes;support vector machine;massive data sets;kernel-based systems;committee machines
593526
Adaptive Sampling Methods for Scaling Up Knowledge Discovery Algorithms.
Scalability is a key requirement for any KDD and data mining algorithm, and one of the biggest research challenges is to develop methods that allow to use large amounts of data. One possible approach for dealing with huge amounts of data is to take a random sample and do data mining on it, since for many data mining applications approximate answers are acceptable. However, as argued by several researchers, random sampling is difficult to use due to the difficulty of determining an appropriate sample size. In this paper, we take a sequential sampling approach for solving this difficulty, and propose an adaptive sampling method that solves a general problem covering many actual problems arising in applications of discovery science. An algorithm following this method obtains examples sequentially in an on-line fashion, and it determines from the obtained examples whether it has already seen a large enough number of examples. Thus, sample size is not fixed a priori&semi; instead, it i>adaptively depends on the situation. Due to this adaptiveness, if we are not in a worst case situation as fortunately happens in many practical applications, then we can solve the problem with a number of examples much smaller than required in the worst case. We prove the correctness of our method and estimates its efficiency theoretically. For illustrating its usefulness, we consider one concrete task requiring sampling, provide an algorithm based on our method, and show its efficiency experimentally.
Introduction Scalability is a key requirement for any knowledge discovery and data mining algorithm. It has been previously observed that many well known machine learning algorithms do not scale well. Therefore, one of the biggest research challenges is to develop new methods that allow to use machine learning techniques with large amount of data. Once we are facing with the problem of having a huge input data set, there are typically two possible ways to address it. One way could be to redesign known algorithms so that, while almost maintaining its performance, can be run e-ciently with much larger input data sets. The second possible approach is random sampling. For most of the data mining applications, approximate answers are acceptable. Thus, we could take a random sample of the instance space and do data mining on it. However, as argued by some researchers (see, for instance, [17]) this approach is less recommendable due to the di-culty of determining appropriate sample size needed. In this paper, we advocate for this second approach of reducing the dimensionality of the data through random sampling. For this, we propose a general problem that covers many data mining problems and a general sampling algorithm for solving it. A typical task of knowledge discovery and data mining is to nd out some \rule" or \law" explaining a huge set of examples well. It is often the case that the size of possible candidates for such rules is still manageable. Then the task is simply to select a rule among all candidates that has certain \utility" on the dataset. This is the problem we discuss in the paper, and we call it General Rule Selection. More specically, we are given an input data set X of examples, a set H of rules, and a utility function U that measures \usefulness" of each rule on X. The problem is to nd a nearly best rule h, more precisely, an h satisfying U(h) (1 )U(h ? ), where h ? is the best rule and is a given accuracy parameter. Though simple, this problem covers several crucial topics in knowledge discovery and data mining as shown in Section 4. We would like to solve the General Rule Selection problem by random sampling. From a statistical point of view, this problem can be solved by taking rst a random sample S from the domain X and then selecting h 2 H with the largest U(h) on S. If we choose enough number of examples from X randomly, then we can guarantee that the selected h is nearly best within a certain condence level. We will refer this simple method as a batch sampling approach. One of the most important issues when doing random sampling is choosing proper sample size, i.e., the number of examples. Any sampling method must take into account problem parameters, an accuracy parameter, and a condence parameter to determine appropriate sample size needed to solve the desired problem. Non theoretically sound sampling methods, like taking a xed fraction of the data set, while widely used in empirical machine learning research, are not appropriate for data mining where the amount of data available is huge. Moreover, these methods do not take into account any accuracy or condence considerations and therefore we do not discuss them here. Widely used and theoretically sound tools to determine appropriate sample size for given accuracy and condence parameters are the so called concentration bounds or large deviation bounds like the Cherno or the Hoeding bounds. They are commonly used in most of the theoretical learning research (see [7] for some examples) as well as in many other branches of computer science [13]. For some examples of sample size calculated with concentration bounds for data mining problems, see, e.g.,[8], [15] and [18]. While these bounds usually allow us to calculate sample size needed in many situations, it is usually the case that resulting sample size is immense to obtain a reasonable good accuracy and condence. Moreover, in most of the situations, to apply these bounds, we need to assume the knowledge of certain problem parameters that are unknown in practical applications. It is important to notice that, in the batch sampling approach, the sample size is calculated a priori and thus, it must be big enough so that it will work well in all the situations we might encounter. In other words, the sample size provided by the above theoretical bounds for the batch sampling approach should be the worst case sample size and thus, it is overestimated for most of the situations. This is one of the main reasons why researchers have found that, in practice, these bounds are overestimating necessary sample size for many non worst-case situations; see, e.g., the discussion of Toivonen for sampling for association rules discovery [15]. For overcoming this problem, we propose in this paper to do the sampling in an on-line sequential fashion instead of batch. That is, an algorithm obtains the examples sequentially one by one, and it determines from those obtained examples whether it has already received enough examples for issuing the currently best rule as nearly the best with high condence. Thus, we do not x sample size a priori. Instead sample size will depend adaptively in the situation at hand. Due to this adaptiveness, if we are not in a worst case situation as fortunately happens in most of the practical cases, we may be able to use signicantly much less examples than in the worst case. Following this approach, we propose a general algorithm | AdaSelect | for solving the General Rule Selection problem, which provides us with e-cient tools for many knowledge discovery applications. This general algorithm evolves from our preliminary works on on-line adaptive sampling for more specic problems related to model selection and association rules done in [3, 4]. The idea of adaptive sampling is quite natural, and various methods for implementing this idea have been proposed in the literature. In statistics, in particular, these methods have been studied in depth under the name \sequential test" or \sequential analysis" [16]. However, their main goal has been to test statistical hypotheses. Thus, even though some of their methods are applicable for some instances of the General Rule Selection problem, as far as the authors know, there has been no method that is as reliable and e-cient as AdaSelect for the General Rule Selection problem. By using a simple example, we explain in the next section the dierences and advantages of AdaSelect over related algorithms including the one following the batch sampling approach. The paper is organized as follows. In the next section, we explain still at some intuitive level the advantage of our algorithm over other random sampling methods. In Section 3, we present our problem and algorithm, and prove theorems concerning its reliability and complexity. In Section 4, we describe several applications of our algorithm to particular data mining problems. Further improvements that might be useful to study particular situations are described in Section 5. We conclude in Section 6 highlighting future work. Related Work The idea of determining the number of necessary examples adaptively and thereby reducing the average number of examples is quite natural, and methods for implementing this idea have been proposed in the literature. In statistics these methods are called \sequential test" or \sequential analysis". Since A. Wald published a pioneer 1 textbook on sequential analysis [16], many researchers have studied sequential analysis methods in depth. Their main goal has been, however, to test statistical hypotheses. Thus, even though some of their methods are applicable for our purpose, that is, for evaluating and/or comparing the value of a given function on a given huge dataset, there has been no method, as far as the authors know, that is as reliable, e-cient and general as our algorithm AdaSelect. To illustrate the dierence between AdaSelect and other methods, let us consider the following very simple problem, for which we only need a very simple utility function. The problem is to test whether the probability p ? that a given condition C holds in a 1 According to the Wald's textbook, the idea of a sequential test procedure goes back to H.F. Dodge and H.G. Romig [2]. database X is more than 1=2 or not. If C held exactly in 1=2 of the transactions in X, then we would have to check all the database to notice it; thus, let us suppose that Intuitively, if the condition holds in many or very few transactions (for instance, 90% or 10% of the database) then just sampling a small number of transactions should be enough to gure out that the answer is positive or negative. On the other hand, the closer to 50% of the transactions that the condition holds the more di-cult it should be to test through sampling what is the correct answer. We can use AdaSelect to solve this problem as follows. We keep two functions, h 1 and h 2 that map transactions into 0=1 values. Given one transaction x, h 1 (x) outputs 1 if the transaction x satises the condition C; otherwise, h 1 (x) outputs 0. On the other hand, h 2 is the negation of h 1 . For the utility function, we use U(h that is, the dierence (positive or negative) from half of the proportion of transactions. For our purpose, the algorithm needs to detect whether U(h 1 ) is positive (in which case negative). Thus, we just need to x to any number less than 1. If we run our algorithm under this setting, we would be able to determine, with high condence, whether the condition C occurs with more than 50% of the transactions in X. Notice that the above application of AdaSelect always solves the problem, no matter how small is ? as far as it is not 0. On the other hand, its sample complexity (the number of necessary examples) depends on ? , and hence, the algorithm might require a lot of transactions when ? is small. More specically, the number of transactions need to be sampled from the database is O((1=? ) ln(1= (ignoring the factor depending on the condence parameter). Therefore, the closer the value of ? to 0 (and thus, p ? to 1=2), the larger the number of examples needed will be. A batch sampling approach is also applicable for this problem. For any , by using an appropriate large deviation bound, e.g., the Cherno bound, we can determine the number of examples su-cient to detect, with high condence, whether the condition holds more than of X. Notice here that we have to choose appropriate since the high condence of this algorithm is guaranteed only if . On the other hand, the number of examples needed is O(1= Therefore, we have to use that is smaller than ? , but if we underestimate ? , then the sample complexity becomes large. In other words, this batch sampling approach has to be always big enough to cover all possible values of bigger than and thus, it depends on since in the worst case ? might be equal to . On the other hand, AdaSelect depends on the case at hand and thus, its sample size depends on ? . Moreover, for practical applications it might be unrealistic to assume that we can assume the knowledge of a lower bound on the quantity ? that we want to estimate. So we can clearly see here the advantage of our algorithm over the simple batch sampling. A more sophisticated way to solve the above problem could be the following. Instead of using the Cherno bound (that provides an estimate that is within a multiplicative error from the probability being estimated and thus adequate for our purposes, but which cannot be used unless a lower bound on ? is known) we can use the Hoeding bound, also usually referred as the additive version of the Cherno bound. The sample size provided by this bound is independent of the probability being estimated and thus, we do not need to assume the knowledge of any lower bound on ? . On the other hand, the estimate is only guaranteed to be within an additive error from the value being estimated. Therefore, in general, it is impossible to use just one single sample to guarantee that our problem is solved. Instead, we can do the following. Suppose that we run a batch sampling calculating the appropriate size with the Hoeding bound for we obtain an estimate p 1 for p ? . Then with high probability, p 1 should satisfy that 1=4. Thus, if p 1 is greater than 3=4, then we can determine (with high condence) that p ? is greater than 1=2; on the other hand, if p 1 is less than 1=4, then we can again determine that p ? is less than 1=2. However, if p 1 is not in that range, that is, 1=4 p 1 3=4, then we cannot conclude anything about whether p ? is above or below 1=2. This last situation does not occur if we choose small enough, but then the sample complexity becomes large if we underestimate . One way to solve this problem is to iterate an execution of the algorithm with a smaller accuracy parameter when the obtained estimate is not in a safe range. In the above case, for example, we can run the algorithm again with . Continue this process until the obtained estimate after the ith iteration is not in the range [1=2 2 (i+1) ; 1=2+2 (i+1) ]. Suppose that the algorithm always gives an i -close estimate of p ? . Then routine calculations show that the algorithm terminates after log(1= iterations. This approach uses O(1= 2 ) examples for obtaining an -close estimate (ignoring the factor depending on the condence parameter) at every step. Hence, altogether we need O((1=? which gives the same order as AdaSelect. In fact, very roughly speaking, this is the idea of AdaSelect. On the other hand, in AdaSelect the iteration phase is naturally incorporated, and thus, AdaSelect has better sample complexity. Furthermore, it is now easy to generalize the technique for estimating and/or comparing more general function values. This method of sampling just described is referred in classical statistics as multiple sampling [1], and it was one of the earlier works of sequential sampling methods [16]. More recent work on adaptive sampling comes from the database community and it is due to Lipton et.al. [9, 10]. There the following problem is discussed 2 . Given a database and a query over the databases (for instance, a selection or a joint), we want to estimate the query size, that is, the number of transactions associated with the query in the database, up to a certain error and condence level. They designed algorithms for this task. Below we refer their algorithms as Adaptive Estimator. More specically, for a given database X, a condition C, an accuracy parameter , and a condence parameter -, Adaptive Estimator estimates (through random sampling) the probability p ? that transactions in up to a multiplicative error . That is, with probability at least 1 -, the algorithm yields estimate p of p ? such that (1 )p ? p (1 . For achieving this task, Adaptive Estimator collects examples from the database sequentially at random while checking whether collected examples are su-cient for terminating the execution with the current estimate. The number of examples used by the algorithm depends on -. The correctness of the algorithm is proved by using large deviation tools, in particular, using the central limit theorem. Thus, in other words, Adaptive Estimator is a sequential version of the multiplicative Cherno bound. However, it is important to notice that Adaptive Estimator only works for estimating the total probability. In our example, it provides us with a multiplicative estimate for p ? but not for ? . Our approach is quite similar to the Adaptive Estimator. The algorithm AdaSelect we propose also collects examples from a given database sequentially at random while checking whether collected examples are su-cient. Furthermore, we also use one of the large deviation tools in order to guarantee the correctness, i.e., the reliability of the output of the algorithm. Thus, one may think supercially that AdaSelect diers from Adaptive Estimator only on the point that it enable us to estimate a more general value, i.e., the value of a user dened utility function. In fact, this is why we could estimate the advantage ? instead of p ? in the above example. Note that this dierence is essential. For our example problem, Adaptive Estimator has exactly the same problem as the batch sampling method with the Hoeding bound and thus, it cannot be used only once in general. It must be used iteratively as we described before. Also we remark that the choice of the utility function U is important. If we use U(h for the above problem, then the execution of AdaSelect becomes essentially the same as Adaptive Estimator. Technically, due to our generalization, the stopping condition of the sampling algorithm is not monotone, which requires more careful probability analysis. 2 In order to keep the explanation at an intuitive level we explain with some typical algorithm and omit the details and the dierences of their algorithms. See the papers cited above for more details. The algorithm AdaSelect presented in this paper evolves from our previous work on hypothesis selection [3] and sampling for association rules [4]. In fact, the two cases treated there are very simple cases of the more general utility function introduced in this paper. The algorithm in the present paper greatly generalizes the adaptive algorithms provided there so that it can be use to attack several other problems. 3 The Adaptive Sampling Algorithm In this section we will formally describe the problem we would like to solve; then we present our algorithm and investigate its reliability and complexity. We begin introducing some notation. Let be a (large) set of examples and let be a (nite, not too large) set of n functions such that can be thought as a function that can be evaluated on an example x producing a real value y as a result. Intuitively, each h 2 H corresponds to a \rule" or \law" explaining examples, which we will call below a rule, and y measures the \goodness" of the rule on x. (In the following, we identify h with its corresponding rule, and we usually call h a rule.) For example, if the task is to predict a particular Boolean feature of example x in terms of its other features, then we could set y the feature is predicted correctly by h, and y predicted incorrectly. We also assume that there is some xed real-valued and nonnegative utility function U(h), measuring some global \goodness" of the rule (corresponding to) h on the set X. More specically, for any S X, U(h; S) is dened as where F is some function IR 7! IR and avg( ) denotes taking the arithmetic average, i.e., i2I a i )=jIj. Then U(h) is simply dened as U(h; X). In Section 4 we will describe several applications of our framework and how U is instantiated to specic functions; then its meaning will become more clear. Now we are ready to state our problem. General Rule Selection Given: X, H, and , 0 < < 1. Goal: Find h 2 H such that U(h) (1 be the rule with maximum value of U(h ? ). Remark 1: (Accuracy Parameter ) Intuitively, our task is to nd some h 2 H whose utility is reasonably high compared with the maximum U(h ? ), where the accuracy of U(h) to U(h ? ) is specied by the parameter . Certainly, the closer U(h) is to U(h ? ) the better. However, depending on the choice of U , the accuracy is not essential in some cases, and we may be able to use a large . The advantage of our algorithm becomes clear in such cases. (Recall the example discussed in Section 2. See also the discussion at the end of this section.) Remark 2: (Condence Parameter -) We want to achieve the goal above by \random sampling", i.e., by using examples randomly selected from X. Then there must be some chance of selecting bad examples that make our algorithm to yield unsatisfactory h 2 H. Thus, we introduce one more parameter specifying condence and require that the probability of such error is bounded by -. Remark 3: (Condition on H) In order to simplify our discussion, we assume in the following that the value of y h;x (= h(x)) is in [0; d] for some constant d > 0. (From now on, d will denote this constant.) Remark 4: (Condition on U) Our goal does not make sense if U(h ? ) is negative. Thus, we assume that U(h ? ) is positive. Also in order for (any sort of) random sampling to work, it cannot happen that a single example changes drastically the value of U ; otherwise, we would be forced to look at all examples of X to even approximate the value of U(h). Thus, we require that the function F that denes U is smooth. Formally, F need to be c-Lipschitz for some constant c 0, as dened below. (From now on, c will denote the Lipschitz constant of F .) Denition. 1 Function F : IR 7! IR is c-Lipschitz if for all x; y it holds jF (x) F (y)j c jx yj. The Lipschitz constant of F is the minimum c 0 such that F is c-Lipschitz (if there is any). Observe that all Lipschitz functions are continuous, and that all dierentiable functions with a bounded derivative are Lipschitz. In fact, if F is dierentiable, then by Mean Value Theorem, the Lipschitz constant of F is max x jF 0 (x)j. As we will see in Section 4 all natural functions used in the applications we describe satisfy this condition for some c. Also note that from the above conditions, we have cd U(h) cd for any h 2 H. Remark 5: (Minimization Problem) In some situations, the primary goal might not be to maximize some utility function over the data but to minimize some penalty function P . That is, we want to nd some h such that P (h) (1 can solve the General Hypothesis Selection Problem by an algorithm and analysis very similar to the one we present here. (End Remarks.) One trivial way to solve our problem is evaluating all functions h in H over all examples x in X, hence computing U(h) for all h, and then nding the h that maximizes this value. Obviously, if X is large, this method might be extremely ine-cient. We want to solve this task much more e-ciently by random sampling. That is, we want to look only at a fairly small, randomly drawn subset S X, nd the h that maximizes U(h; S), and still be sure with probability 1 - that this h we output satises U(h) (1 ) U(h ? ). One can easily think of the following simple batch sampling approach. Obtain a random sample S from X of a priori xed size m and output the function from H that has the highest utility in S. There are several statistical bounds to calculate an appropriate number m of examples. In this paper, we choose the Hoeding bound, which has been widely used in computer science (see, e.g., [7, 13]). One can use any reasonable bound here, and choosing which one to use should be determined by considering reliability and e-ciency. The reason that we choose the Hoeding bound is that basically no assumption is necessary 3 for using this bound to estimate the error probability and calculate the sample size. On the other hand, other bounds, for example, the central limit theorem might be appropriate for some practical situations, since it behaves better although being slighlty less reliable. It is easy to modify our algorithm for analyze it with these alternative bounds. Roughly, for any sample size and error value , the Hoeding bound provides us with an upper bound on the probability that an estimate calculated from a randomly drawn sample is apart from its real value by . Thus, by using this bound, we can determine sample size m that guarantees that the batch sampling yields a rule satisfying the requirement of the problem with probability at least 1 -. While the batch sampling solves the problem, its e-ciency is not satisfactory because it has to choose sample size for the worst case. For overcoming this ine-ciency, we take a sequential sampling approach. Instead of statically decide the sample size, our new algorithm obtains examples sequentially one by one and it stops according to some condition based on the number of examples seen and the values of the functions on the examples seen so far. That is, the algorithm adapts to the situation at hand, and thus if we are not in the worst case, the algorithm would be able to realize of that and stop 3 The only assumption is that we can obtain samples that are independently obtained from the same distribution, a natural assumption that holds for all the problems considered here. Algorithm repeat x randomly drawn example from X; Where < 1 is some constant close to 1 (see the proof of Theorem 3). output h 2 H with the largest U(h; S t ); Figure 1: Pseudo-code of our on-line sampling AdaSelect. Figure 1 shows a pseudo-code of the algorithm we propose, called AdaSelect, for solving the General Rule Selection problem. Now we provide two theorems discussing the reliability and the complexity of the algorithm AdaSelect. In the proofs we will make use of the following lemma. Lemma. 2 Let S X be a set of size t obtained by independently drawing t elements from X at random. For any h 2 H and 0, Proof. Let G be the value avg(y hx : x 2 X), and g be the random variable avg(y S). Observe that using the fact that F is c-Lipschitz, Prf jg E(g)j =c g: But g is the average of t independent random variables with range bounded by d. By the Hoeding bound, this probability is less than 2 exp , as claimed. (End Proof) We rst prove the reliability of Algorithm AdaSelect. Theorem. 3 With probability 1 -, AdaSelect(X; H; ; -) outputs a function h 2 H such that U(h) (1 )U(h ? ). Proof. For a xed > 0, dene H g. We show that the function output by AdaSelect(X; H; ; -) is in H bad with probability less than -. That is, we want to bound the following error probability P error by -. Here we regard one repeat-loop iteration as a basic step of the algorithm and measure the algorithm's running time in terms of the number of repeat-loop iterations. We use t to denote the number of executed repeat-loop iterations. In particular, let t 0 be the integer such that the following inequalities hold. Note that t is strictly decreasing as a function of t; hence, t 0 is uniquely determined. As we will see below, the algorithm terminates at the t 0 th step (i.e., just after the t 0 th repeat-loop iteration) with high probability. For deriving the bound, we consider the following two cases: (Case 1) Some h 2 H bad satises the stopping condition of the repeat-loop before the t 0 th step, and (Case does not satisfy the stopping condition during the rst t 0 steps. Clearly, whenever the algorithm makes an error, one of the above cases certainly occurs. Thus, by bounding the probability that either (Case 1) or (Case 2) occurs, we can bound the error probability P error of the algorithm. First we bound that the probability of (Case 1). Let h bad be the rule in H bad with the largest utility. (Case 1) holds g Prf h satises the stopping condition of AdaSelect at the tth step g the stopping condition of AdaSelect at the tth step g Here g. For any t, 1 t t 0 , this P 1 (t) is bounded as follows. where we have used the fact that for any t t 0 , t > U(h ? )(=2) and that U(h ? )(1 U(h bad ). Then by Lemma 2, we have Thus, we estimate (Case 1) holds g 2n exp Next consider (Case 2). Clearly, (Case 2) implies U(h ? (2= 1). Thus, the probability of (Case 2) is bounded by the following For bounding P 2 , we rst estimate t 0 where we let b . Here by using the assumption that t 0 +1 U(h ? )(=2) and the fact that U(h) cd, we Hence, we may consider that b 1. Therefore, if b , then P 2 is bounded as follows. 2nt In summary, if b then the probability that either (Case 1) or (Case holds is bounded by -. (End Proof) Next we estimate the running time of the algorithm. As above we regard one repeat- loop iteration as a basic step of the algorithm and measure the algorithm's running time in terms of the number of repeat-loop iterations that is exactly the number of required examples. In the above proof, we have already showed that the probability that the algorithm does not terminate within t 0 steps, that is, (Case 2) occurs, is at most -. Thus, the following theorem is immediate from the above proof. Theorem. 4 With probability 1 -, AdaSelect(X; H; ; -) halts within t 0 steps (in other words, AdaSelect(X; H; ; -) needs at most t 0 examples), where t 0 is the largest integer such that t 0 Let us express t 0 in a more convenient form. Recall that 1. Then, since approximately if we approximate x y ln x by x y ln y., we have cd Let us discuss the meaning of this formula. Since both n and - are within the log func- tion, their in uence to the complexity is small. In other words, we can handle a relative large number of rules and require a very high condence without increasing too much the sample size needed. The main terms in the above formula are 1= and (cd)=U(h ? ). (Recall that U(h) cd for any h 2 H; hence, both 1= and (cd)=U(h ? ) are at least 1.) Depending on the choice of U , in some cases we may assume that (cd)=U(h ? ) is not so large; or, in some other cases, we may not need small , and thus, 1= is not so large. AdaSelect performs well in the latter case. More specically, AdaSelect shows its advantage most when U is chosen so that (1) relatively large is su-cient, and (2) though (cd)=U(h ? ) is not bounded in general, it is not so large in lucky cases, which happen more often than the bad cases. As we will see in Section 4, a clever choice of U might allow us to choose a large so that the overall number of examples might not be very large. 4 Examples of Applications In this section we will describe two domains where an instance of our algorithm can be used to solve a particular problem. The two domains studied are model or hypothesis selection and induction of decision trees. Due to space limitations we cannot further describe other possible applications but for problems like association rules mining (which we have already studied in [4]) or subgroup discovery, where batch sampling methods based on large deviation methods have already been proposed to speed up the computational process (see for instance, [15]and [18]) our algorithm can also be applied. In order to instantiate our framework to a particular problem we need to specify what is the meaning of the functions class H, what is the utility function U , and how conditions the conditions on U are satised. We will do so for the following problems. Model or hypothesis selection.- This is the most typical application of our framework. The class of functions H can be seen as a xed set of hypotheses or models. This set could be obtained, for instance, from dierent runs of rival learning algorithms, or the same learning algorithm but with dierent input parameters or architecture, or it could contain several dierent memory-based hypothesis or be just xed to a certain restricted model space as a consequence of a design decision. Notice that this later the case is the case of algorithms that try to select a hypothesis from a simple and small class of hypotheses (for instance, decision stumps) and then amplify its precision using voting methods like boosting [5]. Thus, in the rest of the discussion, we will assume that the class H is xed, nite and of tractable size. The utility function should capture the criterion of \goodness" of the hypothesis, the typical one being the prediction error. In order to keep the discussion at simple level we will assume that the hypotheses are binary and that the \goodness" criteria is the prediction error. For this case, one might naively set the utility function to be just the identity function, that is, U(h; Notice, however, that the worst possible prediction error is 1/2, which is the same as answering by ipping a random coin. Hence, more precisely speaking, for the \goodness" of hypothesis h, we should measure its advantage from from random guessing. Thus, we want our algorithm to output a hypothesis whose advantage from random guessing is close to the best possible advantage in the class. In fact, this setting ts particularly well with voting methods like AdaBoost [5] where typically one just needs to obtain a hypothesis that is better than random guessing at every step. For this purpose, we should set U(h; In particular, for selecting a hypothesis for voting methods, the choice of may not be so important; we can set it to any constant smaller than 1. A similar setting was previously studied by Maron and Moore in [11] where they proposed an algorithm called Hoeding races to accelerate model selection. Their idea was to discard hypotheses when there are clearly not going to be among the best ones. They criteria for discarding them was based on inverting the Hoeding bound, we refer the reader to their paper for details. We can clearly add this feature to our algorithm without compromising its reliability and complexity and possibly accelerating the total running time. Notice that combining Hoeding races with our algorithm does not reduce much the number of examples needed since it depends logarithmically on the number of models n while it might greatly reduce the computation time that depends linearly on n. In a follow up research, Moore and Lee [12] developed a more e-cient version of Hoeding races based on a Bayesian approach under the assumption that the the models accuracies over the dataset are normally distributed. Furthermore, they also introduce a modication for discarding models that are almost indistinguishable to others and thus, allowing the race to output a model that is not the best in the class. Notice that this is also what we are doing through our accuracy parameter . Our framework allows one to use more complicated utility functions that could incorporate size or smoothness considerations together with prediction error or for the case of real-valued functions we could also consider for instance mean square error. Decision tree induction.- Algorithms for decision tree induction typically work by choosing a test for the root from certain node function class (and subsequently, for the root of each subtree) by exploring the training data and choosing the one that is the best according to certain splitting criteria like the entropy measure or the Gini index. In a large dataset it could be possible to reduce the training time by choosing the split based only on a subsample of the whole data. Musick, Catlett, and Russell [14] described an algorithm that implements this idea and chooses the sample based on how di-cult is the decision at each node. Typically, their algorithm uses a small sample at the root node and it enlarges it progressively as the tree grows. Here, we propose an alternative way to select the appropriate sample size needed at every node by using an instantiation of our algorithm that we describe in the following. We will follow the notation from [6]. For simplicity, we assume that we want to construct a decision tree that approximates and unknown function using a training data set S X. Furthermore, we assume that the class of node functions F is xed a priori and it is nite and small, for instance just the input variables and its negations 4 . In fact, this is the class commonly used by standard software packages as C4.5 or CART. Finally, we denote by G : [0; [0; 1] the splitting criteria used by the top-down decision tree induction algorithm. A typical example of this function is the binary entropy Let T be any decision tree whose internal nodes are labeled by functions in F and its nodes labeled by values f0; 1g and let l the leaf of T where we want to make a new internal node substituting leaf l by a function h from F . Our goal is to choose a function h such that the value of G in the original tree T (that is, the sum of the values of G in every leaf weighted by the probability of reaching that leaf) decreases by substituting l by h in 4 This assumption makes sense if all the attributes are discrete or have been discretized a priori. T and labeling the new labels l 0 and l 1 according to the majority class of the instances reaching l 0 and l 1 respectively. More formally, let ^ S S be the set of instances that reach l. For any x in ^ S and any h in F , we denote by p 1; the probability that f(x) is 1, by p ;1 the probability that h(x) is 1, by p 1;1 the probability that both h(x) and f(x) are 1, and by p 0;1 the probability that h(x) is 0 but f(x) is 1. Notice that these probabilities are taken from the distribution induced on ^ S from the initial distribution on the training set S which is typically assumed to be the uniform distribution. Thus, for given T and leaf l from T , and for a given S, our goal is to nd the function h that has the maximum value (T; l; h), where (T; l; denoted by h ? this function. As we mentioned, if we have a large dataset, a more e-cient way to attack this problem is the following. For a given tree T and leaf l, take a sample from X and output the function h that has the highest value (T; l; h) in the sample. If the sample size is chosen appropriately, the value of (T; l; h), where h is the function output by this algorithm, is close to (T; l; h ? ). We can apply our algorithm to use the appropriate amount of data as follows. We use F for H and (T; l; h) for U(h; X) 5 . It only remains to determine the constant c for the Lipschitz condition on U . Notice that U is just the addition of three G functions over dierent inputs. Thus, if G is Lipschitz for certain constant c so it is U . Here we just state how to obtain the Lipschitz constant for G and leave to the reader the calculation of the appropriate constant for the whole function U . If G(q) is the Gini index, then its derivative is G 0 4. Thus, by the Mean Value Theorem, the Lipschitz constant is 4. If G(q) is the binary entropy H(q) or the improved splitting criterion presented in [6], its derivatives are not bounded in the [0; 1] range and therefore it cannot be a xed constant that works for all the possible values. However, suppose that we ignore the input values very close to 0 or 1 and we consider, for instance, the interval [0:05; 0:95] then both functions have Lipschitz constant 5 in that interval. Notice that the inputs for the function G are estimated in the sample. Thus, the above assumption makes sense since we have to admit certain error for those values. Thus, we can now run our algorithm with inputs H, U , d as discussed above and the desired accuracy and condence level -; then the algorithm outputs, with probability larger than 1 -, a node function h such that (T; l; h) (1 )(T; that again, the crucial point is in the choice 5 Precisely speaking, (T ; l; h) does not t to our denition of U(h; X), but it is easy to see that our algorithm also works for this situation. of U . Moreover, according to the results in [6], any h such that (T; l; h) (T; should su-ce to reduce the overall error of the whole tree after adding the new internal node; thus, we can x to be just 1=2. We now informally described another possible application of our algorithm to a problem in decision tree induction, the problem of nding a good cutting point for a continuous attribute in order to discretize it. In this case, every function in H should represent a possible cutting point. Even though the total number of cutting points is innite, notice that we only need to consider cutting points that appears in the dataset set considered; thus, there are at most k of them, where k is the size of X. Although X might be very big, notice that the bound given in Theorem 4 depends only logarithmically in the size of H. Thus, for this particular application, the size of X gives us an upper bound on the size of H. Furthermore, notice that for running our algorithm we do not need to have all the functions in H a priori. So every time that a new example reveals a new possible cutting point, we can just add that new function to the class and calculate its utility function on the examples seen so far. The utility function in this case is the splitting criterion used to compare splitting functions as described above. Thus, our algorithm for this problem would choose an almost optimal, with respect to some splitting criterion, cutting point for a particular continuous attribute. Once this is done we can use that attribute as a discrete attribute to decide which attribute is used to label the node we are building. 5 Further Improvements In this section we describe some improvements that can be incorporated into the main algorithm AdaSelect. Some of these have a completely rigorous justication; others seem intuitively clear but harder to capture in a form of theorem. Use (multiplicative) Cherno bounds instead of (additive) Hoeding bounds.- We have developed algorithm AdaSelect using the Hoeding bound, in which the deviation from the average value appears additively. In some cases, it is also possible to use Cherno's bound, in which the deviation appears multiplicatively. The proof is slightly more involved, because applying Cherno seems to require previous knowledge of the expected value of the utility function, which is precisely what we re trying to estimate, but some careful, also adaptive, application lets us circumvent this problem for some utility functions. In particular, this can be done for the case where F is the identity function plus or minus a constant in [0; 1], as in the problem of model or hypothesis selection. In this case, using Cherno has a clear advantage for the case of small probabilities. More precisely, we can then bound the number of examples by O(1= the dependence on 1=U(h ? ) to linear instead of quadratic). Although this seems asymptotically better, the hidden constant is bigger (precisely, we had 1=2 and now we have and thus, it happens to be better for U(h ? ) < 1=6. Considering the variance.- In some cases, it may not be true that the values of y hx always are in [0; d] for a moderate d, just so with large probability. There are versions of Hoeding, the so called Berstein bounds, that, instead of this assumption, incorporate the variance of the y hx into the bound. When estimates of these variances are known, plugging them into AdaSelect may give quite an advantage over using a worst-case constant d. Adapting to the local Lipschitz constant.- Our bound depends on the constant c, which is the global Lipschitz constant of F . However, in many regions a much smaller constant may be enough to bound the growth rate of F . Consider for example the entropy- style functions discussed for decision-tree induction. We saw there that an appropriate choice for c is around 4 or 5, for probabilities near 0 and 1. However, for probabilities near 1=2, a quite common case, these functions are almost at, i.e., a c much closer to 0 should su-ce. In practice, one should probably use at each step the worst-case Lipschitz constant within the uncertainty interval of U(h), i.e., [U(h; S t It is di-cult to give a rigorous analysis of this new version, but it is intuitively clear that this may result in a signicant improvement while essentially not aecting reliability. 6 Concluding Remarks We have presented a new methodology for sampling that, while keeping all the theoretical guarantees of previous ones, it is applicable in a wider setting and moreover, it is very likely that it is useful in practice. The key point was that, rather than deciding a priori the sample size and obtain it in batch, our algorithm performs sampling sequentially and maintains a stopping condition that depends on the problem at hand. In future work, we should verify the advantage of our approach experimentally. Although the theorem provided in this paper suggests that our algorithm might be e-ciently applicable in several domains, there is still plenty of work to test whether this assertion is true or not. On one hand, our algorithm will take prot of non being in a worst case situation, and therefore, it is very likely that it will outperform the usual batch sampling approach. On the other hand, asymptotics alone do not guarantee practicality since a huge constant might spoil all the advantage of sampling over using all data, even if the most e-cient sampling algorithm is used. We have tested this second point in our preliminary work [3, 4] using synthetic data so we could test a wide range of values and the results are very promising. The sample size used by our method greatly outperform the one used by the batch approach. Moreover, the sample size was, in many cases, reasonable small so that it suggested that reducing the data size through sampling with our algorithm might allows us to improve the overall running time. The experimental results in [10], although in a dierent context, are also very encouraging. In any case, it remains to do experiments using real world data to test the real practicality of our algorithm. Furthermore, we have pointed our several improvements that one might perform when applying the algorithm for a particular situation. Due to the generality of our approach, every application deserves an individual and deeper study to see the possible enhancements that one can do for a particular setting. Acknowledgments We would like to thank Heikki Mannila for pointing us the work in sampling for database query estimation and for encouraging us to follow up our previous research on adaptive sampling. We would also like to thank Chris Watkins for telling us about Hoeding races and Pedro Domingos for pointing us to several related machine learning papers. --R Multiple Sampling with Constant Probability. A Method of Sampling Inspection. Ricard Gavald Ricard Gavald A decision-theoretic generalization of on-line learning and an application to boosting On the boosting ability of top-down decision tree learning algorithms An Introduction to Computational Learning Theory. The power of sampling in knowledge discovery. and Je Randomized Algorithms. Decision Theoretic Subsampling for Induction on Large Databases. Sampling large databases for association rules. Sequential Analysis. Bala Iyer and Je An algorithm for multi-relational discovery of subgroups --TR --CTR Geoff Hulten , Pedro Domingos, Mining complex models from arbitrarily large databases in constant time, Proceedings of the eighth ACM SIGKDD international conference on Knowledge discovery and data mining, July 23-26, 2002, Edmonton, Alberta, Canada Szymon Jaroszewicz , Tobias Scheffer, Fast discovery of unexpected patterns in data, relative to a Bayesian network, Proceeding of the eleventh ACM SIGKDD international conference on Knowledge discovery in data mining, August 21-24, 2005, Chicago, Illinois, USA Osamu Watanabe, Sequential sampling techniques for algorithmic learning theory, Theoretical Computer Science, v.348 n.1, p.3-14, 2 December 2005 Huan Liu , Hiroshi Motoda, On Issues of Instance Selection, Data Mining and Knowledge Discovery, v.6 n.2, p.115-130, April 2002 Jaekyung Yang , Sigurdur Olafsson, Optimization-based feature selection with adaptive instance sampling, Computers and Operations Research, v.33 n.11, p.3088-3106, November 2006 Pierre-Alain Laur , Richard Nock , Jean-Emile Symphor , Pascal Poncelet, Mining evolving data streams for frequent patterns, Pattern Recognition, v.40 n.2, p.492-503, February, 2007
data mining;concentration bounds;scalability;adaptive sampling;knowledge discovery
593551
A Discrete Event Systems Approach for Protocol Conversion.
A protocol mismatch occurs when heterogeneous networks try to communicate with each other. Such mismatches are inevitable due to the proliferation of a multitude of networking architectures, hardware, and software on one hand, and the need for global connectivity on the other hand. In order to circumvent this problem the solution of protocol conversion has been proposed. In this paper we present a systematic approach to protocol conversion using the theory of supervisory control of discrete event systems, which was partially first addressed by Inan. We study the problem of designing a converter for a given mismatched pair of protocols, using their specifications, and the specifications for the channel and the user services. We introduce the notion of converter languages and use it to obtain a necessary and sufficient condition for the existence of protocol converter and present an effective algorithm for computing it whenever it exists.
Introduction There is a growing need for global communication over networks of computers. However, the heterogeneity of existing networks does not allow direct and consistent communication and results in mismatch of protocols. Such mismatches are inevitable due to the proliferation of differing hardware, software and networking standards, and the desired urgency towards global connectivity. While a possible solution to such differing standards or protocol mis-matches would be to standardize networking protocols, such a move may not be practical, and in any case will take years to be agreed upon worldwide, resulting in connectivity problems for the present. As a result, the alternative method of protocol conversion has been proposed. These alternative measures of developing protocol converters will be the only form of solution available until such time when everybody adheres to a global standard. Green [21] argues that protocol conversion is a necessity that cannot be ignored citing reasons that it is a little too late to standardize architectures. An established base of DECnet, ARPAnet, IBM SNA, TCP/IP and X.25 users will find it difficult to switch to an open and global standard simply because of the sheer effort involved, as well as rendering obsolete many existing hardware and software solutions by numerous vendors. Different protocols and architectures also tend to serve different communities of users, with the need to maintain contact with the external world still being of great importance. In Figure consists of the sending end protocol P 0 and the receiving end protocol P 1 . Similarly, the protocol Q is composed of Q 0 and Q 1 . A protocol mismatch occurs when the sending end protocol P 0 of P tries to communicate with the receiving end protocol Q 1 of Q, and similarly also when Q 0 tries to communicate with P 1 . Protocol mismatch Protocol mismatch Figure 1: Protocol configuration A practical solution to such a protocol mismatch is to interpose a translator or a converter between the two protocols so that it traps all messages being sent from one system to the other and translates the messages of the sender system in such a manner that the receiver system can understand them without loss in consistency. This is depicted in Figure 2, where denotes one of the protocols, Q 1 the other of the mismatched protocols, and C denotes the converter. The resulting protocol conversion system has to adhere to the user service specification, which defines the properties of the protocol system as needed by the users. user Figure 2: Interposing a protocol converter Tanenbaum [28] provides a classification for such converters depending on the layer of the protocol stack at which they are used. Repeaters are used at the physical layer; bridges are used at the link layer; routers or gateways are used at the network layer; and finally protocol converters are used at the transport layer. The events that occur at the user interface are called the external events; the remaining events are called the internal events. Let G denote the composition of P 0 and Q 1 , and K denote the user service specification. Then G is a system that evolves over external as well as internal events, whereas K is a formal language defined only over the external events. The composition of G with a converter C, denoted GkC, implements the service specification K if certain safety and progress properties hold. Safety captures the notion that nothing illegal should happen, i.e., the event traces of GkC should correspond to those allowed by K. Progress captures the notion that the composition should not block the occurrence of an external event whenever it is feasible in the specification K, i.e., whenever the specification does not "block" it. The role of the converter C is to restrict the behavior of G so that GkC implements K. However, it must do so under its limited control and observation capabilities. In other words, certain events are controllable-their enablement/disablement can be controlled by C (the remaining are uncontrollable); similarly, certain events are observable-their occurrence can be sensed by C (the remaining are unobservable). Thus the design of protocol converters can be studied in the framework of supervisory control of discrete event systems pioneered by Ramadge and Wonham [23] and subsequently extended by other researchers (refer to the survey articles [24, 29], and the book [12]). This is the motivation for the work presented here. The problem of protocol conversion has been studied by some researchers and the paper by Calvert and Lam [2] provides a nice survey. One of the first approaches to protocol conversion is the bottom-up approach taken by Okumura [20] and by Lam [17]. The bottom-up approaches are heuristic in nature and may not be able to determine a converter even when one exists. The top-down approach of Calvert and Lam [2] is algorithmic in nature. However, the converter they design is a state machine that evolves over the set of internal events, which implies that the set of internal events is treated as the set of controllable as well as the set of observable events for the converter. This is unrealistic, since the converter can observe only those events that occur at its interface, i.e., its "input" and "output" events, and it can control only its output events. So in general these event sets are different. This is further illustrated in the example below. The top-down approach or the algorithmic approach is also known as the quotient approach since C can be viewed as the "quotient of G and K". The protocol conversion problem was first treated by Inan in the supervisory control framework as an important application example in [8], in further depth in [9], and the theoretical foundation on which this application is based was given in [10]. However, these work of Inan only address the safety constraint of the protocol conversion problem which requires that the projected language of the supervised system should equal the given specification language. The additional progress constraint, which requires that the supervised system never block an external event that is not blocked by the specification itself, was also considered as part of the protocol conversion problem in the work of Calvert-Lam [2]. We treat both the safety and progress constraints in the work presented here. The existence of converter requires controllability and observability conditions together with extra safety and progress conditions implying that the work presented here also requires the generalization of the conventional supervisory control. Rudie-Wonham [25] have also applied the supervisory control techniques for deriving a missing portion (either sender or receiver portion) of the protocol assuming some knowledge of it. This problem is quite similar to that of protocol conversion but again considers safety specification only. In a recent work, Kumar-Fabian [11] show that the progress constraint may be expressed using a type of nonblocking property with respect to a family of marked languages-one for each external event. This nonblocking property differs from the conventional one used in supervisory control [3] in two ways: first, nonblocking with respect to a family of markings (as opposed to a single marking) is needed, and second, each type of marking depends, in a very specific manner, on the entire closed-loop behavior (as opposed to being just those traces of the closed-loop behavior that are marked in the open-loop behavior). In this paper we derive a converter using the formal techniques from supervisory control of discrete event systems. The supervisory control framework provides necessary and sufficient conditions for the existence of supervisors (in the present setting converters) for a given plant (in the present setting composition of the two mismatched protocols) so that the closed-loop system (in the present setting composition of converter and the mismatched protocols) meets a desired specification (in the present setting the user service specification). However, since the user service specification is a partial specification, i.e., it is defined only on the subset consisting of the external events (as opposed to the entire event set which is customary in the supervisory control framework), the supervisory control theory results cannot be applied directly, and appropriate extensions have been obtained in the paper. We introduce the notion of converter languages and show that the existence of a converter is equivalent to the existence of a converter language contained in the language of G. The set of converter languages is closed with respect to union. So a supremal converter language exists, and we provide an effective algorithm for computing it. The test for the existence of a converter reduces to the test of non-emptiness of the supremal converter language. Moreover, the generator for the supremal converter language serves as a choice for the converter. No converter exists if the supremal converter language is empty, in which case the given service specification may be minimally altered so that a converter exists. This problem has recently been addressed by Takai-Takae-Kodama [27] and Kumar-Fabian [11]. We illustrate our work by the example of two incompatible protocols namely the Alternating Bit protocol and the Nonsequenced protocol [28]. These two are obviously incompatible because the alternating bit protocol labels all data packets with sequence numbers 0 or 1, while the nonsequenced protocol does not work with labeled data packets. Abridged versions of the work presented here first appeared in [14, 15]. Motivating Example For the example we assume that the converter is collocated with the receiving end. So, as shown in Figure 3, the sending end P 0 is the composition of the sender protocol P s and the sender's channel P c , whereas the receiving end Q 1 consists of only the receiver protocol Q r . In s -A Qls tm del acc Figure 3: A typical protocol conversion system general Q 1 is also a composition of Q r and the receiver's channel. The mismatched protocol components for the example are the Alternating Bit protocol sender (P s ), Alternating Bit channel (P c ), and the Nonsequenced protocol receiver (Q r ). The state machines for each of these along with that for the service specification are presented in this section. Thus in this example The events occurring at various interfaces for the present example are indicated in Figure 3. The external event set consists of the accept event (acc) and the deliver event (del). Lower (respectively, upper) case letters are used for internal events occurring at the sender (respectively, receiver) end. An event label having a negative (respectively, positive) sign as its prefix represents a transmit (respectively, receipt) event of a protocol. However, the sign convention for the channel is the opposite since a receipt event of the channel is a transmit event of the adjoining protocol and vice-versa. So the receipt events of the channel are prefixed with negative signs, whereas the transmit events of the channel are prefixed with positive sign. Since the converter is interposed between the channel and the receiver protocol, this fixes the events that occur at the converter interface. Thus, for instance, -di represents a transmit event of a data packet with label i (where at the sender protocol (and a receipt event of the same data packet at the channel), whereas -A represents a transmit event of an acknowledgment at the receiver protocol (and a receipt event of the same acknowledgment at the converter). Other events include the timeout event (tm) and the channel loss event (ls). Thus in the above example the event set \Sigma consists of the following: A subset of \Sigma consisting of the events occur at the converter interface. These constitute the set of observable events, whereas all the remaining events are unobservable. Part of the observable events are the output events for the converter, and their occurrence can be controlled. So the converter output events constitute the set of controllable events, which are: All the other events are uncontrollable to the converter. Note that the set of controllable events for the converter is contained in the set of its observable events. We exploit this property when we design the converter. Also, note that the set of controllable events and the set of observable events are both different from the set of internal events. This distinction is not noted in the work of Calvert and Lam [2]. The Alternating Bit sender depicted in Figure 4 has six states. The initial state is the state 0, where it is in a position to accept data from the user. The data is transmitted with label 0. The next data is accepted from the user after an acknowledgment with the correct label 0 is received. Otherwise, when either the wrong acknowledgment with label 1 is received or the sender times out (due to loss of data or acknowledgment), the data is retransmitted with label 0. This procedure is repeated after each accept event, except that the label used alternates between 0 and 1. acc Figure 4: Alternating Bit sender The Alternating Bit channel shown in Figure 5 has six states. The channel initially receives a data packet (events -di), which it may either lose (event ls) sending it back to its initial state, or may successfully transmit (events +di) sending it to the state where ls ls Figure 5: Alternating Bit channel it receives an acknowledgment packet (events -ai). Acknowledgments may either get lost (event ls) or it may get successfully transmitted (events +ai) sending the channel back to its initial state in either case. The Nonsequenced receiver is shown in Figure 6. This is a very simple state machine, which on receiving a data delivers it to the user and sends an acknowledgment to the sender. Since no labels are present data packets labeled by the Alternating Bit sender cannot be interpreted consistently by the Nonsequenced receiver resulting in a protocol mismatch.0+D -A del acc del Figure Nonsequenced receiver and Service specification Finally, the protocol system should provide the service of loss free transmission over the lossy channel which is accomplished by requiring that the accept and deliver events should alternate. This service specification is depicted in Figure 6. Weaker service specifications of the type "the order of accepted and delivered message sequences be identical" can also be considered. However, more complex protocols will be needed to offer such a service. 3 Notation and Preliminaries We use \Sigma to denote the universe of events. \Sigma denotes the set of all finite length event sequences, called traces, including the zero length trace, denoted ffl. A subset of \Sigma , i.e., a collection of traces, is called a language. For a language H, the notation pr(H), called the prefix closure of H, is the set of all prefixes of traces in H. H is said to be prefix closed if Given a trace s 2 \Sigma and a subset of events - the projection of s on - denoted s" - \Sigma, is the trace obtained by erasing the events not belonging to - \Sigma from s. State machines [7] are used for representing untimed behavior of discrete event systems (such as protocol and channel systems) as well as for representing qualitative or logical specifications (such as user service specifications). Formally, a state machine P is a quadruple denotes the set of states, \Sigma denotes the finite set of events, is the partial transition function, and x is the initial state. A triple (x; oe; x 0 is called a transition of P if x 0 2 ff(x; oe); it is said to be an epsilon-transition if denotes the set of states reachable by execution of zero or more events in - \Sigma from state x in and the notation - denotes the set of events in - \Sigma that are executable at x in . The transition function is extended from events to traces ff in an obvious way. The generated language of P is the set of all traces that it can execute starting from its initial state, i.e., P is called a deterministic state machine if the transition function is a partial function of the X. The completion of a deterministic state machine P , denoted is the state machine obtained by "completing" the transition function of P by adding a new state, a "dump" state, x d , and adding transitions from each state x of P to the dump state on those events that are not defined at x in P . Formally, defined x d otherwise Note that L(P Synchronous composition [6] of state machines is used to represent concurrent operation of component systems. Given two deterministic state machines P := (X composition of P and Q denoted PkQ := (X; \Sigma; ff; x 0 ), is defined as: undefined otherwise Thus when P and Q are composed, their common events occur synchronously, while the other events occur asynchronously. The generated language of the composition is given by: Note that when \Sigma since all events must occur synchronously. Also note that although the state set for PkQ is X P \Theta XQ , many of the states remain unreachable in PkQ. We adopt the convention that by writing PkQ we mean its reachable or trim component [7]. In supervisory control of discrete events systems, synchronous composition of an uncontrolled plant, modeled as a state machine G, and a supervisor, modeled as a state machine S having an identical event set as the plant, is used as a control mechanism. A certain sublanguage H ' L(G) represents the desired behavior of the plant, and the control objective is to design a supervisor S such that the controlled plant behavior L(GkS) equals H. The supervisor to be designed has limited control and observation capabilities in the sense that (i) it cannot prevent the occurrence of certain uncontrollable events, and (ii) it can only observe the occurrence of certain observable events. Letting \Sigma u ' \Sigma denote the set of uncontrollable events and \Sigma denote the set of observable events, the following result from supervisory control theory states a necessary and sufficient condition for the existence of such a supervisor: Theorem 1 [23, 18] Given a plant G, a desired behavior H ' L(G), the set of uncontrollable events \Sigma u , and the set of observable events \Sigma o , there exists a supervisor S (compatible with control and observation capabilities) such that only if Prefix closure and Non-emptiness: Controllability: pr(H )\Sigma The controllability condition requires that extension of any prefix of H by an uncontrollable event that is feasible in the plant should also be a prefix of H. This is because the occurrence of uncontrollable events cannot be prevented. A pair of traces is called indistinguishable if each trace in the pair has identical projection on the set of observable events. The observability condition requires that extensions of a pair of indistinguishable prefixes of H by a common feasible event should either both or neither be prefixes of H. This is because identical control action must be taken following an indistinguishable pair of traces. Note that H is controllable (respectively, observable) if and only if pr(H) is controllable (respectively, observable). Tests for controllability and observability conditions are known when G has finitely many states and H is a regular language so that it admits a finite state machine representation. In fact if G has m states and H has a state machine representation with n states, then controllability can be tested in O(mn) time, whereas the observability can be tested in O(mn 2 In case the desired behavior H fails to satisfy any of the required conditions, a maximally permissive supervisor is designed that achieves a maximal sublanguage of H satisfying the required conditions. It is known that controllability is preserved under union so that a unique maximal controllable sublanguage, called supremal controllable sublanguage of a given language exists [22]; however, observability is not preserved under union, so maximal observable sublanguages are not unique [18]. Hence sometimes normal sublanguages instead of observable sublanguages are considered [18]: Normality requires that traces of G that are indistinguishable from a prefix of H must themselves be prefixes of H. Normality is preserved under union so that the supremal normal sublanguage of a given language exists [18, 1, 13] and it can be tested in O(mn 2 ) time [12, p. 103]. Moreover, normality implies observability, and the converese holds in the presence of controllability when the controllable events are also observable: Theorem 2 [9, Proposition 4.2] Given G and H ' L(G), if controllable and observable if and only if it is controllable and normal. A similar result also appeared in [12, Theorem 4.3] and in the timed setting in [19, Proposition 4]. 4 Existence and Computation of Converter In this paper we are interested in solving a slightly different supervisory control problem, where the objective is to obtain a supervisor, which we refer to as a converter in this context, so that the closed-loop system implements a given service specification defined on the subset of external events. Given P := (X; \Sigma; ff; x 0 ), a set of external events \Sigma e ' \Sigma and a service specification K ' \Sigma e , P implements K if the following hold: Safety: L(P )"\Sigma Progress: Safety requires that each generated trace of P should correspond to a prefix of the specifi- cation, i.e., no "illegal" traces should occur in P . Since K is a partial specification (defined only on the external event set), there may exist more than one trace of P that correspond to the same prefix of K. Progress requires that if an external event is possible after such a prefix of K, then it should also be possible "eventually", i.e., after occurrence of zero or more internal events following each corresponding trace of P . Note that safety only guarantees that such an external event is eventually possible following at least one (and not all) of the corresponding traces of P . Remark 1 The definition of "implements" given above is equivalent to that given in [2] but is stated differently for simplicity. For example the definition of safety given in [2] uses containment instead of equality. However, containment can be replaced by equality since the reverse containment follows from progress. Similarly, the definition of progress given in [2] uses a "state characterization" instead of a "language characterization". This is because in [2] P is represented as a nondeterministic state machine over only external events by replacing each transition on an internal event by an epsilon-transition. So a language based characterization of progress is not possible and a state based characterization is used. In our case P is a state machine over both external and internal events. In the following theorem we provide a necessary and sufficient condition for the existence of a converter for a given pair of mismatched protocols and a given service specification. We first introduce the notion of converter languages. As described above the notation G is used to denote the composition of the mismatched protocols, and K is used to denote the service specification. Given a pair of mismatched protocols G, a service specification K ' \Sigma e , a set of uncontrollable events \Sigma u , and a set of observable events \Sigma called a converter language if the following hold: Controllability: pr(H )\Sigma Safety: pr(H)"\Sigma Progress: Note that H ' L(G) is a converter language if and only if its prefix closure is also a converter language. Using the result of Theorem 1 we next show that a necessary and sufficient condition for the existence of a converter is the existence of a nonempty converter language. Theorem 3 Given a pair of mismatched protocols G, a service specification K ' \Sigma e , a set of uncontrollable events \Sigma u , and a set of observable events \Sigma o , there exists a converter C (compatible with control and observation capabilities) such that GkC implements K if and only if there exists a nonempty converter language. Proof: We first prove the necessity. Suppose there exists a converter C such that GkC implements K. We claim that H := L(GkC) is the the required converter language. Since C is control and observation compatible, from the necessity part of Theorem 1 it follows that H is nonempty, prefix closed, controllable and observable. Furthermore since GkC implements K, it follows from Definition 1 that the safety and progress properties of Definition 2. Thus H is a nonempty converter language. In order to see the sufficiency, suppose there exists a nonempty converter language H ' L(G). Then pr(H) is nonempty, controllable and observable. So from the sufficiency of Theorem 1 it follows that there exists C which is control and observation compatible such that pr(H). Furthermore, since H is a converter language, it follows from Definition 2 that it satisfies the safety and progress properties, which implies GkC implements K as desired. Since the set of controllable events for a converter is a contained in the set of its observable events, i.e., it follows from Theorems 2 and 3 that a language is a converter language if and only if it satisfies the properties of controllability, normality, safety, and progress. This fact can be used to obtain the following corollary which states that the set of converter languages is closed under union so that a supremal one exists. Define the following set of converter sublanguages of L(G) which implement the specification K: Conv(G; K) := fH ' L(G) j H is a converter languageg: e , the supremal converter language sup Conv(G; K) exists. Proof: First note that since ; is a converter sublanguage, Conv(G; K) 6= ;. Let be an indexing set such that for each - 2 , H - ' L(G) is a converter language. We claim that -2 H - is also a converter language, i.e., it satisfies the conditions of controllability, normal- ity, safety, and progress. The first two properties follow from the fact that controllability and normality are preserved under union. Safety follows from the following series of equalities: where we have used the fact that prefix closure and projection operations commute with arbitrary union, and each H - satisfies safety. Finally, to see progress, pick s 2 pr( pr(K). Then there exists - 2 such that Also, since H- is a converter language, it satisfies progress. So there exists The following theorem provides a concrete condition for the existence of a converter and forms a basis for the test developed in this paper. It also specifies a choice for a prototype converter. Theorem 4 Let G; K; \Sigma e ; \Sigma be as in Theorem 3. Then there exists a converter C such that GkC implements K if and only if sup Conv(G; K) is nonempty, in which case C can be chosen to be a generator of sup Conv(G; K). Proof: In order to see the necessity suppose there exists a converter. Then from the necessity part of Theorem 3 there exists a nonempty converter language H 2 Conv(G; which implies sup Conv(G; K) is nonempty. Sufficiency follows from the sufficiency part of Theorem 3 since sup Conv(G; K) is a converter language and it is given to be nonempty. Finally, let C be any generator of sup Conv(G; K), i.e., where the last equality follows from the fact that sup Conv(G; K) ' L(G). Consequently, satisfies the safety and progress properties of Definition 2, which implies GkC implements K. From Theorem 4 the task of checking the existence of a converter as well as that of designing one when it exists reduces to the task of computing the supremal converter language and verifying its non-emptiness. We next present an algorithm for computing sup Conv(G; K) assuming that G has finitely many states, say m, and K is a regular language so that it admits a finite state machine representation say S states. For the motivating example, the state machine representation for K consists of only two states as shown in Figure 6. As with the computation of the supremal controllable sublanguage given in [13], the algorithm for the computation of sup Conv(G; K) has two steps: In the first step it constructs a sufficiently "refined" version of the state machine G; 1 and in the next step, it removes certain "bad" states from this refined state machine. Initially certain states which correspond to traces that violate safety are marked "bad", i.e., these are the states that are reachable by execution of traces whose projection on external event set are not prefixes of K. If there are no such bad states, then sup Conv(G; K) equals L(G). Otherwise, a converter must be designed to restrict the behavior of G so that only those "good" states remain reachable which correspond to traces that also satisfy progress, controllability, and normality conditions. Progress requires that the set of external events that can be executed following the execution of zero or more internal events from a certain good state should contain the set of external events executable at the corresponding point in K. If a good state fails to satisfy the progress, it is marked bad. Controllability requires that no bad state should be reachable from a good state on an uncontrollable event, since execution of an uncontrollable event cannot be prevented and the system can uncontrollably reach a bad state from a good state. So in order to compute sup Conv(G; K) if there exists an uncontrollable transition from a good state to a bad one, then that good state is marked bad. Finally, normality requires that the states corresponding to a set of indistinguishable traces should either be all good or all bad. So if a good state and a bad state can be reached by the execution of pair of indistinguishable traces, then that good state is marked bad. The algorithm terminates when there are no more additional bad states to be marked. Then sup Conv(G; K) consists of traces corresponding to the remaining good states, and it is nonempty if and only if the set of good states is nonempty. It is clear that the state machine representation of G needs to be sufficiently refined so that the states corresponding to the traces violating either of the conditions can be unambiguously identified. First in order to deal with safety, progress and controllability conditions we refine the machine G by composing it with the state machine S obtained by completing the transition function of S. For the motivating example, the state machine S is shown in Figure 7, in which the dump state is explicitly depicted. Note that e as expected, and given a trace s 2 \Sigma only if its execution does not result in the dump state in S. Let G 1 := SkG, then since \Sigma e ' \Sigma, we have where the last equality follows from the fact that e . For notational simplicity let fy d g) \Theta X, where y d is the dump state of S, denotes the state Given two state machines G i is said to be a refined version of G 2 if there exists a function and oe 2 \Sigma. dump del acc del acc acc, del Figure 7: Completion of service specification set of G 1 , Z is the transition function, and z is the initial state. Note that given a trace s 2 its execution leads to a state z = (y d ; x) in G 1 if and only if s"\Sigma e 62 pr(K). Next in order to deal with the normality condition we further refine G 1 . First we obtain a nondeterministic state machine G 2 that generates all traces that are indistinguishable from the traces of G 1 . Since a trace remains indistinguishable when unobservable events are either inserted or erased, the following construction yields the desired G Algorithm 1 Given G 1 := SkG, add transitions in G 1 to obtain G 2 as follows: 1. For each z 2 Z and oe 2 2. For each transition (z; oe; z 0 ) of G 1 such that z 6= z 0 and oe 2 add an epsilon- transition (z; ffl; z 0 ). in the algorithm has the effect of inserting (respectively, erasing) unobservable events. Remark 2 Note that if z is reachable by execution of a trace s in G 1 , then z is also reachable by execution of all traces that are indistinguishable from s in G 2 . So if s trace indistinguishable from s and if z 0 is reachable by execution of s 0 in G 1 , then z 0 is also reachable by execution of s in G 2 (since s is indistinguishable from s 0 ). In fact the set of states reachable by execution of s in G 2 is the set of states that are reachable in G 1 by execution of those traces in G 1 that are indistinguishable from s. Next using the power-set construction [7] we obtain a deterministic state machine G 3 with the same language as L(G 2 ). Finally we construct the machine G 4 := G 1 kG 3 . Since show below that G 4 is a sufficiently refined version of G. We first outline the construction of G 4 in the following algorithm: Algorithm 2 Given G := (X; \Sigma; ff; x 0 ) and a deterministic generator S := (Y; \Sigma e ; fi; y 0 ) of 1. Obtain S by adding a dump state and completing the transition function of S. (Then e .) 2. Obtain G 1 := SkG. (Then L(G 1 and the state set of G 1 is denoted Z.) 3. Obtain the nondeterministic state machine G 2 by adding transitions in G 1 as described in Algorithm 1. (Then and the state set of G 2 is Z.) 4. Obtain G 3 by "determinizing" G 1 using the power set construction. (Then L(G 3 and the state set of G 3 is 2 Z .) 5. Obtain G 4 as G 1 kG 3 . (Then L(G 4 and the state set of G 4 is Z \Theta 2 Z .) Clearly, G 4 is a refined version of G. For notational simplicity, let G 4 := (R; \Sigma; ffi; r 0 ), Z is the state set for G 4 . Note that each state r in G 4 is of the form Z), where z 2 Z and - denotes the transition function of G 4 . The initial state of G 4 is r In the following lemma we list some of the properties of G 4 which demonstrate that it is a sufficiently refined version of G. We first define the concept of a matching pair of states. Definition 3 A pair of states r are said to be a matching Lemma 1 The following hold for G 4 which is the state machine constructed in Algorithm 2: 1. 2. Consider s 2 L(G 4 ); then s"\Sigma e 2 pr(K) if and only if is such that y 6= y d . 3. If is a state of G 4 , then z 2 - Z. 4. Consider a matching pair of states r Proof: The first part follows from the construction and is proved above. In order to see the second part, consider s execution of s"\Sigma e in S results in the state y of S. Since the assertion follows from the fact that only if y 6= y d . In order to see the third part, let s 2 L(G 4 ) be such that ffi(r Z). Then execution of s results in the state z in G 1 and state - Z in G 3 . Since G 3 is obtained by "determinizing" the state machine G 2 , this implies that execution of s results in each state Z in G 2 . (Recall that G 2 is nondeterministic.) Since G 2 is obtained by adding transitions in G 1 , z is one such state. Hence z 2 - Z. Finally in order to see the fourth part, consider a matching pair of states r Z), 2. From the third part we have that z 1 Z. Consider s 1 2 L(G 4 ) such that execution of s 1 results in state z 1 in G 1 , and each state - Z in G 2 (including the states z 1 and z 2 ). Then from Remark 2 concerning G 2 , there exists a trace indistinguishable from s 1 , i.e., s 2 "\Sigma such that its execution results in the state z 2 in G 1 . (If no such trace exists, then z 2 cannot be reached by execution of s 1 in .) Finally, since s 2 is indistinguishable from s 1 , its execution in G 2 also results in the set of states - Z. So the execution of s 2 results in the state r Z) in G 4 as desired. We are now ready to present the algorithm that iteratively marks bad states in G 4 , and upon termination yields a generator for sup Conv(G; K). The notation R k b ' R is used to denote the set of bad states at the kth iteration. Algorithm 3 Consider G 4 := (R; \Sigma; ffi; r 0 ) obtained in Algorithm 2. 1. Initialization step: 2. Iteration step: R k+1 b := R k r 3. Termination step: If R k+1 and go to step 2. The algorithm initially sets the iteration counter marks a state R to be a bad state if its first coordinate is the dump state, i.e., if . This is because if any trace such that ffi(r then from the second part of Lemma 1, violates the safety condition. The set of bad states at the kth iteration is denoted by R k b . In the kth iteration step, a good state b is marked bad if any of the following hold: (i) If there exists an uncontrollable event from r to a bad state. This is because if s 2 L(G 4 any trace such that ffi(r then s violates the controllability condition. (ii) If there exists a matching bad state r 0 2 R k b . This is because if any trace such that ffi(r then from the fourth part of Lemma 1, there exists a trace s 0 indistinguishable from s such that ffi(r violates the normality condition. (iii) If the set of external events that are executable at state y in S is not contained in the set of external events that are executable from those good states that are reached by the execution of zero or more internal events from r in G 4 . This is because if any trace such that ffi(r then s violates the progress condition. The algorithm terminates if no additional bad states are marked in the kth iteration; else, the iteration counter is incremented by one and the iteration step is repeated. Note that since G 4 has finitely many states, the algorithm is guaranteed to terminate in a finite number of iterations. Also, since the algorithm marks a state to be a bad state if and only if all traces that lead to it violate either safety, progress, controllability, or normality, the state machine obtained by the removal of the terminal set of bad states (and all transitions leading towards or away from them) from G 4 generates the language sup Conv(G; K). Hence we obtain the following result stating the correctness of the algorithm. Theorem 5 Given a finite state machine G and a regular language K ' \Sigma e , Algorithm 3 terminates in a finite number of steps, and the state machine obtained by removal of the states R b ' R from G 4 generates sup Conv(G; K), where R b denotes the set of bad states at the termination of the algorithm. be the number of states in G and n be the number of states in the state machine representation of K. Then the computational complexity of Algorithm 3 is since the number of states in G 4 is O(mn2 mn ), which implies that there are O(mn2 mn ) number of iterations with equally many computations in each iteration. Also, no algorithm of polynomial complexity exists. This follows from the fact that in the special case when \Sigma the converter design problem reduces to the standard supervisory control problem under partial observation with the desired behavior constraint specified as a range of languages, which is known to be an NP-complete problem [30]. 5 Implementation Issues and Example Converters Since the computation of sup Conv(G; K) is intractable, a computationally tractable heuristic approach to converter design is desirable. We propose two heuristics and utilize each to design a converter for the motivating example. Since the computational intractability arises due to the presence of partial observa- tion, one possibility is to first compute the supremal sublanguage of L(G) satisfying safety, progress, and controllability, and next verify whether it is also normal, each of which can be done in polynomial time. In case the language is also normal, then it equals the desired supremal converter language and we are successful in computing it in polynomial time. We next outline a polynomial time computation of the supremal sublanguage of L(G) satisfying safety, progress, and controllability. Clearly, this language equals sup Conv(G; K) when all events are observable (so that normality trivially holds). Note that when all events are observable, G 2 equals G 1 , i.e., no transitions are added when Algorithm 1 is invoked, so G 4 also equals G 1 . Hence the following modification of Algorithm 3 computes the desired supremal sublanguage of L(G) satisfying safety, progress, and controllability. Algorithm 4 Given G := (X; \Sigma; ff; x 0 ) and a deterministic generator S := (Y; \Sigma e ; fi; y 0 ) of 1. Initialization step: 2. Iteration step: 3. Termination step: If Z k+1 and go to step 2. Using the fact that G 1 has O(mn) states, the computational complexity of Algorithm 4 can be determined to be O(m 2 n 2 ). The algorithm computes the supremal sublanguage of L(G) that satisfies safety, progress, and controllability (but may violate normality). The test whether this language is also normal (with respect to the given set of observable events) can be performed in O(m(mn) 2 In case the test for normality fails, then as in the work of Cho-Marcus on iterative computation of supremal controllable and normal sublanguage [4], we can iterate between the supremal normal sublanguage computation and the computation of the supremal language that meets safety, progress, and controllability until a fixed point is reached, which however will result in an exponential computational complexity. This is outlined in the following algorithm: Algorithm 5 Given G and K ' \Sigma e , compute the sup Conv(G; K) as follows: 2. Compute the supremal sublanguage - H k of H k satisfying safety, progress, and controllability using Algorithm 4. If - H k is normal, then sup Conv(G; go to step 3. 3. Compute the supremal normal sublanguage H k+1 of - H k . If H k+1 satisfies safety and progress (Algorithm 6 given below provides a test for safety and progress), then sup Conv(G; replace G by the generator of H k+1 , set and go to step 2. Remark 4 Note that in the above algorithm we do not need to check the controllability of H k+1 in step 3 since it is known that the supremal normal computation preserves controllability [5, Proposition 3.9]. Also, in step 3 if H k+1 does not satisfy safety and progress, then we need to compute its supremal sublanguage satisfying safety and progress. We can use Algorithm 4 for doing this (although in this case the controllability trivially holds). How- ever, we need to replace G by the generator of H k+1 since Algorithm 4 only computes the supremal sublanguage of L(G) satisfying safety, progress, and controllability. We have written a C-program for Algorithm 5 (and the associated Algorithm 4) that utilizes a finite state machine library originally developed for supervisory control [26]. Using the program for Algorithm 4 we first computed the supremal sublanguage of L(G) satisfying controllability, safety, and progress for the example. The state machine G, which is the composition of Alternating Bit sender, Alternating Bit channel, and Nonsequenced receiver contains 66 states. The composition of G with S (which has three states) contains 198 states. The initial iteration of the safety check disqualified 66 states. The next iteration which performs controllability and progress checks disqualified additional 34 states. No additional states were disqualified in the next iteration of controllability and progress checks. This resulted in a test converter with states. The test converter also passed the normality test; thus qualifying it as a valid converter. A second possible heuristic is to guess a test converter C and verify its correctness by checking whether L(GkC) is a converter language, which as we show below can be done in polynomial time. First note that it is not difficult to guess a converter; a simple possibility is to design a system that emulates the missing portions of the mismatched protocols, i.e., the receiver protocol P r of P and the sending protocol Q s of Q. For the motivating example the "guess converter" we consider below emulates the functions of the Alternating Bit receiver and the Nonsequenced sender. In order to test whether L(GkC) is a converter language, we must check whether it satisfies controllability, normality, safety, and progress. Polynomial time tests for controllability and normality can be found in [12, pp. 75, 103]. Here we present a polynomial time test for safety and progress. For notational simplicity let G 5 := GkC and let its state set be \Phi. We first refine G 5 by composing it with S. For notational simplicity, let G 6 := SkG 5 and let its state set be \Theta := (Y [ fy d g) \Theta \Phi. Note that since \Sigma e ' \Sigma, where the last equality follows from the fact that e . Also, given a trace s 2 L(G 6 execution leads to a state to satisfy safety, no state ' of G 6 should be of the form (y d ; OE). Also for L(G 5 ) to satisfy progress, the set of external events that can be executed following the execution of zero or more internal events from each state should contain the set of external events executable at the corresponding state y in S. So we have the following algorithm for checking safety and progress: Algorithm 6 Consider the composition of mismatched protocols G, a test converter C, and a deterministic generator S of pr(K). 1. Construct G 5 := GkC, and denote its state set by \Phi. 2. Construct G 6 := SkG 5 , and denote its state set by \Theta. 3. Then L(GkC) satisfies safety and progress if and only if If the number of states in the converter is p, then the computational complexity of the algorithm is O(m has O(mnp) states and equally many reachability computations need to be performed. A test converter for the motivating example is shown in Figure 8. The converter adopts +D -A -A Figure 8: A test converter the following simple conversion strategy. Initially when one or more data packets with label arrive (event +d0), it removes the label and forwards a single data packet to the receiver (event +D). No action is taken at this point if more copies of the same data packet arrive (due to a sender timeout). When the receiver transmits an acknowledgment (event -A), the converter attaches the label 0 and puts it onto the sender's channel (event -a0). The procedure is repeated when data packets with a different label arrive (except for the difference in the label used). However, if another data packet with the same label arrives (due to a sender timeout or a loss of acknowledgment in the channel), then the same acknowledgment is retransmitted. We have verified that L(GkC) (where lability, and normality, i.e., it is a converter language. Since L(GkC) is obviously nonempty, it follows from Theorem 3 that C is indeed a valid converter. 6 Conclusion In this paper we have studied the problem of designing protocol converters that need to be interposed between a pair of mismatched protocols. Our approach is systematic and is based on the recent theory of supervisory control of discrete event systems. The work presented here provides a new framework for protocol converter designers on one hand, and serves as an application for the supervisory control theorists on the other hand. The basic concepts of controllability, observability, normality, and computation of supremal languages from supervisory control, and safety, and progress from protocol design play important role in the protocol conversion problem. The converter that we derive is maximally permissive in the sense that any other converter will further restrict the behavior of the entire system. However, the maximally permissive converter may not have a minimal number of states. Design of such minimal converters is an interesting problem for future research. --R Formulas for calculating supremal controllable and normal sublanguages. Formal methods of protocol conversion. Dealing with blocking in supervisory control of discrete event systems. On supremal languages of class of sublanguages that arise in supervisor synthesis problems with partial observations. Supervisory control of discrete event processes with partial observation. Communicating Sequential Processes. Introduction to Automata Theory Supervisory control and formal methods for distributed systems. Supervisory control: Theory and application to the gateway synthesis prob- lem Nondeterministic supervision under partial observations. On supervisory control of partial specification arising in protocol conversion. Modeling and Control of Logical Discrete Event Systems. On controllability and normality of discrete event dynamical systems. Design of protocol converters: A discrete event systems approach. Protocol conversion using supervisory control techniques. Protocol conversion. On observability of discrete-event systems Supervisory control of timed discrete event systems under partial observation. A formal protocol conversion method. Protocol conversion. On the supremal controllable sublanguage of a given language. Supervisory control of a class of discrete event processes. The control of discrete event systems. Supervisory control of communicating processes. A software library for discrete event systems and other finite state machine based applications. The extremal languages arising in supervisory control for service specifications. Computer Networks. Logical aspects of control of discrete event systems: a survey of tools and techniques. On the control of discrete event dynamical systems. --TR --CTR Nina Yevtushenko , Tiziano Villa , Robert K. Brayton , Alex Petrenko , Alberto L. Sangiovanni-Vincentelli, Solution of parallel language equations for logic synthesis, Proceedings of the 2001 IEEE/ACM international conference on Computer-aided design, November 04-08, 2001, San Jose, California Young Moo Lee , Kyu Ho Park, A protocol converter for nonblocking protocols, Integration, the VLSI Journal, v.33 n.1, p.71-88, December 2002 Khaled El-Fakih , Nina Yevtushenko , Sergey Buffalov , Gregor V. Bochmann, Progressive solutions to a parallel automata equation, Theoretical Computer Science, v.362 n.1, p.17-32, 11 October 2006 Nina Yevtushenko , Tiziano Villa , Robert K. Brayton , Alex Petrenko , Alberto L. Sangiovanni-Vincentelli, Compositionally Progressive Solutions of Synchronous FSM Equations, Discrete Event Dynamic Systems, v.18 n.1, p.51-89, March 2008
observability;controllability;protocol conversion;discrete event systems;progress;normality;supervisory control;synchronous composition;safety
593691
Structuring Acyclic Petri Nets for Reachability Analysis and Control.
The incidence matricesfrom places to transitions and vice versaof an acyclic Petri net can obtain a block-triangular structure by reordering their rows and columns. This allows the efficient solution of some reachability problems for acyclic Petri nets. This result is further used in supervisory control of Petri nets; supervisors for Petri nets with uncontrollable transitions are constructed by extending the method of Yamalidou et al. (1996) to Petri nets where transitions can be executed simultaneously. A large class of Petri nets with uncontrollable transitions is given for which the maximally permissive supervisor can be realized by a Petri net. The original specification is algorithmically transformedby using the results for acyclic Petri netsinto a new specification to take the presence of uncontrollable transitions into account. The supervisor is obtained by simple matrix multiplications and no linear integer programs need to be solved. Furthermore, a class of Petri nets is given for which the supervisor can be realized by extending the enabling rule with OR-logic.
on-line computations for implementing the feedback controller are reduced signicantly and properties of the closed-loop system can be studied using standard techniques for Petri nets. For Petri nets in which all transitions are controllable and observable, and legal markings have to satisfy a conjunction of linear inequalities, the controller design problem was solved by Yamalidou et al. (1996) using place invariants (Giua et al. (1992) obtained the same result using a dierent methodology). The controller is found by simple matrix multiplication. When some transitions in the Petri net are uncontrollable, the construction of maximally permissive Petri net supervisors becomes much more dicult. This problem was studied by Giua et al. (1992) and Giua et al. (1993). A class of systems for which a maximally permissive supervisor can be realised by a Petri net was obtained by Li and Wonham (1994) based on a linear integer programming approach. In (Moody and Antsaklis, 1998) eorts were made to extend the supervisor design of Yamalidou et al. (1996) to Petri nets with uncontrollable transitions. However, they do not specify a class of Petri nets and of legal sets for which a maximally permissive Petri net supervisor can be constructed. The goal of this paper is two-fold. In the rst part we study acyclic Petri nets and in the second part we construct maximally permissive Petri net supervisors for a large class of Petri nets with uncontrollable transitions. Our motivation for studying acyclic Petri nets |which form by themselves a very restricted class of Petri nets| is that a lot of Petri nets have an acyclic uncontrollable part. We partition the set of places and the set of transitions of an acyclic Petri net. This partition indicates how tokens ow through the net and is used to order rows and columns such that both the incidence matrix from places to transitions and the incidence matrix from transitions to places obtain a block-triangular structure. This allows us to give a direct proof for the necessary and sucient condition for point to point reachability which was given by Ichikawa and Hiraishi (1987) and by Li and Wonham (1994). It also shows that in an acyclic Petri net every reachable marking can be reached in as many steps as there are sets in the partition of the transitions set. Moreover, it overcomes the problem studied by Kostin (1997) of not knowing in which order transitions should be executed when one already has a solution of the Petri net state equation. Also, a new algorithm for checking point to point reachability in acyclic Petri nets is presented. In the second part of the paper we construct supervisors for Petri nets. By using a dierent technique we show that the result of Yamalidou et al. (1996) also holds for Petri nets in which several transitions can be executed simultaneously and one transition can be executed more than once at the same time. A class of Petri nets is specied for which a maximally permissive supervisor can be realised by a Petri net even when uncontrollable transitions are present in the Petri net. This class is more general than the one of Li and Wonham (1994). The Petri nets obtained by applying the supervisory control reduction result of Stremersch and Boel (1999) to Petri nets belonging to this class have an acyclic uncontrollable part. The block-triangular structure of its incidence matrices |obtained in the rst part of the paper| is used to obtain an algorithm transforming the legal set into a specication which takes the presence of uncontrollable transitions into account without loss of maximal permissiveness, i.e. without becoming more restrictive than necessary. Then we apply the supervisor design method of Yamalidou et al. (1996) to the new specication. The Petri net supervisor is obtained by simple matrix multiplication and no linear integer programmes need to be solved. Finally, a class of Petri nets is given for which the supervisor can be realised by extending the standard enabling rule for Petri nets with OR-logic. The outline of the paper is as follows. Section 2 introduces Petri net denitions and notations. In Sections 3 and 4 the incidence matrices of an acyclic Petri net are structured and some reachability issues for acyclic Petri nets are studied respectively. Section 5 generalises the result of Yamalidou et al. (1996), discusses conditions under which maximally permissive Petri net supervisors can be constructed when not all transitions are controllable and reviews the reduction result of Stremersch and Boel (1999). A class of Petri nets for which a maximally permissive supervisor can be realised by a Petri net is given in Section 6. This class is made more general in Section 7. Supervisors with OR-logic in their enabling condition are studied in Section 8. Final conclusions are given in Section 9. denitions A Petri net is a tuple are disjoint, nite sets of places and transitions respectively. Further, are the incidence functions from transitions to places and from places to transitions respectively. Each transition corresponds to an event in the discrete event system. For each is the set of input (output) places of t: Similarly, denotes the set of input (output) transitions of p 2 P: Analogous denitions are used for sets of transitions and sets of places. Dene the incidence matrices F (from transitions to places) and F (from places to transitions) as Further dene F A marking is a function M be the set of all markings. It can be identied with bag for short| is a function as the set of all transition bags. A bag 2 is state-enabled under The set of state-enabled bags under is denoted by m : The marking of N evolves in the following way. If m is the marking, then a bag 2 m can be executed and the marking changes to m when is executed. It is said that m 0 is reached from m by the execution of : This is denoted by m Note that this notation implies that From the denition of it follows that transitions are allowed to be executed simultaneously and that the same transition is allowed to be executed more than once at the same time. This assumption is referred to as the transition bag assumption (see also (Peterson, 1981) and (Ushio, 1990)). Denition 1 introduces for each k 2 N and m 2 M the set R k (m) M: It contains the markings reachable from m in at most k successive executions of transition bags. Denition The innite sequence (R i 0 is non-decreasing. Consequently, R1 (m) is well dened for each m: For subsets M 0 of M dene R k A path in N is a nite, nonempty sequence of places (2 P ) and transitions (2 T ) such that for each two consecutive elements a and b in ; b 2 a (Desel and Esparza, 1995). Let be the set of all paths in N . A path 2 from a 2 P [T to b 2 P [T is a cycle if no element occurs more than once in and a 2 b : A Petri net is called acyclic if it does not contain any cycles. 3 Structural analysis of acyclic Petri nets In this section it is shown in a constructive manner that by properly ordering the rows and columns of the incidence matrices F + and F of an acyclic Petri net these matrices obtain a block-triangular structure. 1 For a set X; ]X denotes its cardinality. Both markings and transition bags can also be written as column matrices. Both notations will be used interchangeably. Except for column and row matrices, matrices are set bold. The notation is used for the subset and for the proper subset relation. Denition 2 For an acyclic Petri net recursively dene for all k 2 N n f0g the mutually disjoint sets of transitions T k T and sets of places P k The set T 1 is the set of source transitions |transitions without input places (Murata, 1989)| of N : The set P 1 is the set of places of which the set of input transitions is a subset of In particular, all places without input transitions are in Lemma 1 For each acyclic Petri net it holds that Proof. (i) Assume that T Denition 2 this implies that or, equivalently, that As holds that (Denition 2) or that This argument can be repeated for t 1 (see (3)). As [ k 1 in this way one can construct an arbitrarily long path Consequently contains at least ]T +1 transitions and thus at least two of them will be equal. Hence, and N contain a cycle. This contradicts N being acyclic. (ii) The proof is analogous to that of part (i). 2 For the result of Lemma 1(i) is that, if T 1 6= T then T 2 6= ;: For the innite sequence mutually disjoint sets of transitions one can dene s.t. T k and can be empty. If T 1 6= ;; resp. sets. For the result of Lemma 1(ii) is that, if P For the innite sequence mutually disjoint sets of places one can dene partitioned into 0 sets. Proposition 1 For an acyclic Petri net N it holds that Proof. From Denition 2 it follows that for each k 2 N Together with the denition of and 0 this nishes the proof. Corollary 1 For an acyclic Petri net it holds for each k 2 that and for each k 0 that Corollary 2 Every path in an acyclic Petri net, with and ; is such that Similar properties for paths beginning or ending in a transition follow immediately. Note that for every pair of places (transitions) belonging to the same set P i (T i ); there exists no path which connects these two places (transitions). Next the consequences of the partitioning of P and T for the matrices F (from transitions to places) and F (from places to transitions) are studied. For simplicity of notation assume that T 1 and that rows (columns) in F + and F are ordered such that places (transitions) in P 1 places (transitions) in P 2 (T 2 ) and so forth, then F + and F have the following Figure 1: Petri net of Example 1. block-triangular structure (see Corollary 1): 1;1 O 1;2 O . O O 1;1 F 1;2 O 2;1 O 2;2 F O . F O The matrices F i;j and F i;j are in N ]P i ]T j and the matrices O i;j are zero matrices of dimension ]P i The i-th row of block matrices (i corresponds to places in i-th column of block matrices (i corresponds to transitions in T The denitions of the sets P i and T i imply that each row in F i;i with i > 1 contains at least one nonzero element (if not the corresponding place would belong to a set P i 0 with i 0 < i) and each column in F i 1;i with i > 1 contains at least one nonzero element (if not the corresponding transition would belong to a set T i 0 with then the last row (rst column) of F + and F has to be left out. Example 1. For the acyclic Petri net of Figure 1 in which all arcs have weight 1, When bringing rows and columns in the order the Reachability issues for acyclic Petri nets Reachability issues for acyclic Petri nets are studied in this section. Theorem 1 which gives a necessary and sucient condition for a marking to be reachable from a given marking in an acyclic Petri net can be found in (Ichikawa and Hiraishi, 1987). Here a constructive proof is given based on the results of Section 3. In (Li and Wonham, 1994) this theorem was also proven. The present proof is in the same spirit, but is written out in detail because it allows to formulate useful reachability properties for acyclic Petri nets and to construct Petri net supervisors in Sections 6, 7 and 8. For such that 0: (7) Theorem 1 Given an acyclic Petri net It holds that there exists a 2 such that Proof. The 'only if'-part follows from the state evolution of a Petri net. For the 'if'-part, (8) implies that Partition P and T as in Section 3. We prove that (see (6)-(7)) is a valid sequence of transition bags starting from the initial marking This is equivalent to which in turn follows from We prove (11) by induction on i: Because 1 holds that for each p 2 or that induction hypothesis is that To prove that one has to check that (see (6) and (7)) Equation (12) implies that (see (6) and (7) for the second equality) From Corollary 1 we have that ( T i Consequently, If this is substituted in (13), it is left to check that This follows from (9). As for each t 2 T (see (6) and (7)) it holds that m (10). For each T 0 T dene R T 0 k (m) as the set of markings which are reachable from m in at most k successive executions of transition bags with a zero image on T n T only transitions in T 0 are executed. It is clear that R T From the proof of Theorem 1 one immediately obtains the following important result. Corollary 3 For an acyclic Petri net holds that (R (R Starting from a given marking m 0 in an acyclic Petri net, every reachable marking m f can be reached in at most successive executions of transition bags in the order T that only depends on the structure of the Petri net, i.e. only on F + and F ; and is independent of both the initial and the nal marking. Corollary 3 solves the problem studied by Kostin (1997) in which, starting from a solution of (8), an algorithm is constructed for deciding in which order transitions have to be executed to reach m f from one can do one step better than the result in Corollary 3, i.e. every reachable marking can be reached in at most 1 successive executions of transition bags. Proposition 2 Given an acyclic Petri net marking (R T3 (R T1[T2 Proof. By using the notation of (6)-(7) and of the proof of Theorem 1 we prove that if Because of using (6)-(7) for suces to show that From (10) we have that It follows that As ( the second term in the left hand side of (15) is equal to zero. This implies (14). Note that if T i.e. if there are no source transitions, then For the Petri net in Figure 1 it holds that P each reachable marking can be reached in at most two executions of transition bags. The results presented so far are for the transition bag assumption (see Section 2). More restrictive assumptions were dened by Holloway et al. (1997). Under the concurrency assumption multiple transitions are allowed to be executed simultaneously but each transition not more than once, i.e. is the set of binary valued functions on Under the no concurrency assumption only one transition at a time can be executed, i.e. is the set of binary functions on T for which at most one transition has image 1. It is clear that Theorem 1 also holds in case these assumptions are used by decomposing the transition bags in (10) in components which satisfy these limitations. Proof. As 1 it is clear that 1 From its denition, 2 2 ~ ~ From it follows that ~ Consequently, we have together with If satises (8), then under the concurrency assumption m f can be reached in steps and under the no concurrency assumption in t2T (t) steps (see also (Li and Wonham, 1994)). Note that both these numbers |in contrast with the result of Corollary 3| depend on and thus also on m 0 and on m f : Proposition 3 gives a necessary and sucient condition for the set of reachable markings in an acyclic Petri net to be nite. Proposition 3 In an acyclic Petri net N in which every source transition has at least one output place (to rule out trivial cases), the set of reachable markings R1 (m) is nite for each m 2 M; i there are no source transitions in N ; i.e. i Proof. 'If'-part: As T we have that (R T 2(m)) ) (see Proposition 2) and that the set of transition bags which are zero on T n T 2 and state-enabled under m is nite. This also holds for all sets of transition bags which are zero on T n T i+1 and state-enabled under m i with (R T 2(m)) Together with the fact that is nite this nishes the proof. 'Only if'-part: By contraposition. Assume T 1 6= ;: As each source transition has at least one output place the marking of these places can grow unbounded. Thus, if there are no source transitions in an acyclic Petri net, then the set of reachable markings is nite and reachability questions can be answered via enumeration. The enumeration process can be structured by using the result of Proposition 2. The results presented so far only solve the problem of ordering events for a given solution of (8). Propositions 4 and 5 give an alternative for checking if a system of linear equalities has a nonnegative integer valued solution to decide whether m f is reachable from m 0 in an acyclic Petri net. Proposition 4 Consider an acyclic Petri net. If ~ (R and there exists a place (R T i+1 m)) Proof. As each place p in [ i its input transitions in [ i holds for each marking m in R T (R T i+1 m)) ) that m(p) ~ Thus, if there exists a place p in [ i then we have for this p that m(p) < m f (p); from which the result follows. Intuitively this result follows from the fact that when executing transition bags as in (10) the marking of a place increase or remain constant until the execution of i and thereafter will decrease or remain constant. Dene for each the natural number All input and output transitions of a place p in P i belong to [ !(p) Proposition 5 Consider an acyclic Petri net. If ~ (R T 1(m 0 and if there exists a place (R T i+1( ~ m)) Proof. If (R T i+1 m)) ) it holds that ~ Together with ~ this nishes the proof. Propositions 4 and 5 can be used in a branch and bound algorithm for checking wether m f is reachable from m 0 in the case First construct the set R T 2 which are reachable from m 0 by the execution of one transition bag with transitions in T 2 and which are not rejected by Propositions 4 and 5. Analogously, the set R T 3 (R T 2 procedure is repeated until m f is found, until for the set R T j (R T 2 empty or until m f is not found in R T (R T 1 (R T 2 The usefulness of this approach will depend on the values of !(p) i with Proposition 5) and the number of choice places. An example where this approach is benecial is when all but the matrices F i;i and F i 1;i with are zero. The generalised tree structure of Li and Wonham (1994) belongs to this class of acyclic Petri nets. This tree structure is dened with the additional constraint that none of its places has more than one output transition. 5 Petri net supervisors In the second part of this paper we design maximally permissive supervisors for a class of Petri nets containing uncontrollable transitions. Consider a Petri net which we refer to as the Petri net plant. The control goal is that |by dynamically disabling controllable transitions| the marking of the Petri net plant can never leave a given legal set A M; which is dened by an inequality with integer coecients: 2 Intersections of sets of this type are handled by successively applying the method of this paper. Assume that A is a nonempty proper subset of M: Legal sets which are unions of sets of this type are discussed in Section 8. When all transitions in the Petri net plant are controllable and observable this problem was solved by Yamalidou et al. (1996) by a method based on place invariants. They used the no concurrency assumption: only one transition can be executed at a time. By using a dierent technique, it is shown in x5.1 that their result also holds under the more general transition bag assumption. In x5.2 we discuss Petri net supervisors for Petri nets with uncontrollable transitions and in x5.3 the reduction result of Stremersch and Boel (2000) for supervisory control of Petri nets is reviewed. 5.1 All transitions controllable and observable Consider the legal set A as in (17) and an initial marking m 0 2 A; i.e. Am 0 b: For the supervisor to allow the execution of 2 under m 0 ; not only A(m+F) b should hold but also A(m+F 0 for each 0 2 with This is because the supervisors considered here cannot force events to they can only disable them. Thus 2 can be allowed if and only if Dene the row matrix B 2 Z 1]T as: 3 Because of 0 0 and 0 ; (19) is equivalent to As b Am 0 0 and B 0 this condition is in the standard form of the Petri net enabling condition. Thus, to guarantee that (18) holds the condition B b Am 0 is added to F This can be seen as an extra place p c with marking b Am 0 and outgoing arcs As the above reasoning can be repeated for m 1 the marking of p c evolves to: when the plant's marking changes from m 0 to contains the outgoing arcs of Consequently, the incidence matrices of the closed-loop Petri net (plant and supervisor) are@ F The supervisor is obtained via a simple matrix multiplication. Because (18) and (21) are equivalent, this supervisor only disables transition bags if this is necessary in order to achieve the control goal A; i.e. it is maximally permissive. It is clear that supervisors for legal sets dened by the intersection of a nite number of sets as in (17) can be designed by successively applying this method. In x5.2 we 3 For is the element of X which corresponds to t 2 T: discuss how this result can be generalised to Petri net plants with uncontrollable transitions. 5.2 Some transitions uncontrollable as the set of uncontrollable (controllable) transitions. If uncontrollable transitions are present in the Petri net plant and the marking has always to be in some set A (not necessarily dened as in (17)), the control specication has to be changed from A to its subset (see (Holloway and Krogh, 1990)) A := in order to avoid that a marking outside A could be reached via the execution of a sequence of uncontrollable transition bags. As dened above, the set R T u 1 (m) is the set of markings which can be reached from m by any nite sequence of bags containing only uncontrollable transitions. As shown in x5.1, when A can be described by the intersection of linear inequalities then the supervisor for A can be realised by a Petri net. In general, even if A is as in (17), this will not be the case. This problem was studied in (Giua et al., 1992) and (Giua et al., 1993). In (Li and Wonham, 1994) Petri net supervisors were constructed in the presence of uncontrollable transitions by using linear integer programmes. In Sections 6 and 7 we will dene a class of Petri nets |more general than the one in (Li and Wonham, 1994)| for which the supervisor for A can be realised by a Petri net. To construct this supervisor no linear integer programmes need to be solved. 5.3 Reduction result for supervisory control In this section we review the supervisory control reduction result of Stremersch and Boel (1999) and Stremersch and Boel (2000). Note that the following results are valid for general legal sets A: Introduce for every p in P the order relation p in M: Denition in M and p in For every p in partially ordered set. Let (X; ) be a partially ordered set and X 0 X: Then X 0 is an up-set (down-set) in (X; ) if (Davey and Priestley, 1990). Dene O " (X; the set of all up-sets (down-sets) in (X; Denition 4 For A M dene the sets of places The sets of Denition 4 are mutually disjoint and their union is equal to P: A path in N is uncontrollable if all transitions in are uncontrollable, i.e. if all transitions in it belong to T as the set of all uncontrollable paths in N : For all the rst (last) element of : Dene for each p 2 P the sets (p) := fp In words, (p) is the set of places belonging to uncontrollable paths ending in p; (p) is the set of uncontrollable input transitions of (p) and (p) is the set of input places of uncontrollable output transitions of p together with the singleton fpg (fpg is added to guarantee p 2 (p) in case p \T ;). For each holds that p 2 (p) (the sequence which only consists of p is an uncontrollable path). Next, we dene for each of the sets set of places and a set of uncontrollable transitions. Denition 5 Dene for each A M the sets I I d := [ U d := [ p2P d Proposition 6 8A Dene the sets I := I # [ I " [ I d and Theorem 2 formulates the reduction result. Theorem 2 8A M;8m Thus, we get from (22) and Theorem 2 that U If then it is easier to construct A with (27) than with (22). To check whether R only transitions in U are needed. To decide whether transitions in U are state-enabled the marking of places in U is needed and to check whether the marking of places in needed. The union of these two sets is I (see Proposition 6). The set of places I; the set of transitions U and the functions f A U I and f A := f j I U form the Petri net NA := ( I ; which is a sub-Petri net of N : Note that all transitions in NA are uncontrollable. The structure of NA will determine the class of Petri nets with uncontrollable transitions for which a Petri net supervisor can be obtained. To avoid confusion in the remainder of the paper all sets of input and output places (transitions) of transitions (places) in U I) are taken in the original net N and not in the subnet NA : 6 Design of a Petri net supervisor In this section we will dene a class of Petri nets (x6.1) for which the maximally permissive supervisor can be realised by a Petri net. It is constructed in x6.2. 6.1 A class of Petri nets with uncontrollable transitions In this section, A is dened by a linear inequality with integer coecients as in (17). Proposition 7 Consider A as in (17). If A 6= ; and A 6= M; then we have for the sets of Denition 4 that Proof. See Appendix A. Thus I Introduce the following three hypotheses: (H1) NA is acyclic, When hypotheses (H1) to (H3) hold, we will construct in x6.2 for A as in (17) a row matrix A 2 Z 1]P and a number b 2 Z such that A Thus, the supervisor for the Petri net plant N with set of uncontrollable transitions T u can then be realised by a Petri net (see x5.2). From it follows that 0 This implies that I (25) and (26)). In Section 7 conditions (H2) and (H3) will be further weakened. Reorder rows and columns in F + and F such that places in I and transitions in U come rst. Consequently, F + and F can be written as where both F A ; which contains arcs from transitions in U to places in I; and F A ; which contains arcs from places in I to transitions in are in N ] I ] The matrix F 2 ; which contains arcs from places in P n I to transitions in U ; is zero because of I (see Proposition 6). Because of (H1) we can |following the procedure of Section 3| partition the set of places I in nonempty sets and the set of transitions U in nonempty sets T implies that there are no source transitions in NA ). Assume rst that It will be shown in Remark 1 that the conclusions also hold for A and F A such that places (transitions) in P 1 places (transitions) in P 2 (T 3 ) and so forth. Because of the partitions of I and U and because of (H2), F A and F A obtain the following block structure: O 1;2 O O O 1;C C C C C C C A F 1;2 O O 2;2 F O As was pointed out in Section 3, the i-th row (column) of block matrices with A and F A corresponds to places (transitions) in P i (T i+1 ). The denitions of the sets P i and imply that each row in F i;i with 2 i 1 contains at least one nonzero element, and that each column in F i;i+1 with 1 i 1 contains at least one nonzero element. p3 c p6 c Figure 2: The Petri net of Example 2. Moreover, because of (H2), (H3) and the denition of NA in x5.3, each column in F A is zero, apart from one element in F i 1;i which is equal to 1. This implies that the (unique) input place of t 2 is an element of P i 1 and that all output transitions of which belong to U are in T Example 2. Consider the Petri net in Figure 2. Controllable and uncontrollable transitions have a label 'c' and 'u' respectively. The legal set is Hence, I The sub-net NA is indicated in Figure 2. Since NA is acyclic, one can partition into sets explained in Section 3: and into sets T By bringing rows and columns respectively in the order we obtain from (31) that Compare the structure of these matrices with (28)-(30). 4 Lemma 3 will be used in Theorem 3. Lemma 3 Given integers c positive integers d and a nonnegative integer c: If 4 d for each i 2 i2f1;:::;kg Proof. As the set described by the linear inequalities 5 is a polytope in R k and each of its vertices has integer valued coordinates. Therefore, The left hand side of (34) is a linear programme over a polytope. If c > 0; then this polytope vertices. With each of them corresponds a candidate optimal solution X 4 If a nonzero integer a divides an integer b; this is denoted as a j b: On the other hand, a does not divide b is denoted as a - b: 5 R is the set of real numbers. c Consequently, c c i2f1;:::;kg both the left and the right hand side of (32) are equal to zero. Corollary 4 is a direct generalisation of the necessary and sucient condition for reachability in acyclic Petri nets of Theorem 1. Corollary 4 Given a Petri net If the sub-Petri net N a 2 with j TnT 6.2 Construction of a Petri net supervisor From (H1) and a direct generalisation of Proposition 2, we can rewrite A in (27) by using the partitions of I and U as: (R T 1 (R Introduce the set A recursively dene the sets A i as A g: (36) It is immediately clear that and that A When hypotheses (H1) to (H3) hold, we will prove by induction on i that there exists for each holds because A I : The induction hypothesis is that there exists a matrix A Based on A construct a matrix A such that A From (36), a marking m belongs to A Using the induction hypothesis this is equivalent to A As ) is an acyclic sub-Petri net of NA (see Proposition 6 and (H1)), (38) is equivalent to (see Corollary 4): By using the structure of F + and F (see (28)-(30)), it holds for each 2 with j TnT i denote the elements of j T i Then, (39) is equivalent to (recall that A for each I A I A For I A Because of (H3), (41) is equivalent to I A Because of (H2), it holds for dierent places p 1 and p 2 in I that p \ p \ Therefore, (42) is equivalent to I A Applying Lemma 3 for each we see that (43) is equivalent to I A b; (44) I A A b: This can be written as A A where A I n P A Consequently, we have obtained a matrix A such that A Bringing the results for all holds for A Theorem 3 states the main result of this section. Theorem 3 If conditions (H1) to (H3) hold, then the row matrix A 2 Z 1]P with A U I A (p U is such that A Remark 1. The result in Theorem 3 also covers the case Remark 2. If the legal set A is equal to then the presented theory can be applied successively if for each of the subnets NA i (as dened in hypotheses (H1)-(H3) hold. Example 2 (continued). Consider again the Petri net in Figure 2. Using the result of Theorem 3 one obtains: A (p 7 A (p 8 A (p 5 A (p 6 A (p 2 A (p 3 A (p 1 Therefore, and the supervisor is The closed-loop system can be found in Figure 3. The supervisor consists of one control place p c and the arcs connected to it. Note that none of the uncontrollable transitions has an incoming arc from the control place. 4 Because legal sets which are down-sets in (M;) form an important class of specication sets, we formulate Proposition 8. c c Figure 3: Closed-loop Petri net of Example 2. Proposition 8 If 0 A and A 6= 0; then 0 < A j Proof. As 0 A it is true that Proposition 7) and thus As A 6= 0 it is true that P # 6= ; and thus also that I I: From the denition of I # and of (:) it follows that for each p 2 I we have p \ I 6= ;: Since NA is acyclic, one can construct an uncontrollable path in N such that \ It is clear that will be less than ]P: As p 2 I and p \ holds that A (p Next, we prove by induction that A (p then we prove A (p i ) > 0: From Theorem 3, A (p i U I A (p I A (p 7 Weaker hypotheses for the construction of a Petri net supervisor 7.1 Hypothesis (H2): source transitions in NA In this section condition (H2) is replaced by the weaker condition (H2') In contrast to (H2) source transitions, these are transitions with an empty set of input places, are now allowed in as the set of source transitions in N ; i.e. T s := can be applied to A to obtain from (27) that (R Dene the set A ] as A By using Theorem 3 a matrix A can be constructed such that A From (46) and (47) it follows that A It follows immediately that If there exists a t 2 U \ T s such that 0 < In particular, this is the case if Proposition 8). On the other hand, if for each 7.2 Hypothesis (H3) In this section it is shown that it is possible to weaken hypothesis (H3). In (43), we used Lemma 3 with all d i 's, i.e. all f (p; t) with I and t 2 p \ equal to 1. Next, we study conditions under which Lemma 3 can be used when more general arc weights occur. Going through the proof of Theorem 3 again, the summand of the second term in the left hand side of (43) is then I A To be able to apply Lemma 3 we should have that U U Figure 4: Diagram for I when NA is acyclic. U We introduce hypothesis (H3b) because (H3a) should also hold for each m which can be reached from m by the closed-loop system, i.e. the original system synchronised with the supervisor. The condition (H3b) is not only sucient but also necessary if each output and each input transition of p is executed at least once in one of the possible trajectories of the closed-loop system. U in (H3b), then f because NA is acyclic and (H3b) implies that For I with p \ U as the integer which is equal to f (p; t); 8t 2 p \ U . The conjunction of conditions (48) and (49) is equivalent to (H3b): I Consider the diagram for I in Figure 4. It holds that p \ p \ T This is because for each I; it holds that p\T u U or p \T u U (see (23)-(26)). Thus, p\p \T u would contradict NA being acyclic. Therefore, together with the denition of f p ; (49) is equivalent to the conjunction of the following three conditions There are two cases to consider: Case 1. If in (52) f imply that Consequently the discrete event system modeled by N with initial marking m can be modeled by the simpler controlled Petri net f ) with initial marking m; where and Hypothesis (H3) holds for the new model Consequently, the results of Section 6 can be used to construct a supervisor for it. Case 2. If in (52) f a generalisation of hypothesis (H3) is obtained: it can be replaced by the conjunction of (H3a) and (H3b). The construction of the row matrix A is done in a similar way as above. 8 Petri net supervisors with OR-logic 8.1 Unions of linear inequalities as a legal set For Petri nets with all transitions controllable and observable it was discussed in x5.1 how for legal sets described by the conjunction of linear inequalities the corresponding maximally permissive supervisor can be realised by a Petri net. However if the legal set is described by the disjunction of linear inequalities, it is necessary to extend the Petri net enabling condition with OR-logic as was pointed out by Li and Wonham (1994). Consider a legal set where for each i 2 Consider an initial marking m 0 2 A; i.e. it satises As discussed in x5.1, for the supervisor to allow the execution of 2 under m 0 ; not only should hold but also should hold for each 0 2 with Transforming this condition to a simpler one |as was done in x5.1| is in general very dicult or even impossible under the transition bag assumption. Hence, we adopt the no concurrency assumption in this section: we only consider 's in noc Moreover, we assume that the Petri net does not contain any self-loops. Then, the enabling condition F m is equivalent to To allow 2 noc it should hold that _ Because of the denition of noc and because m 0 2 A; this is equivalent to _ Condition (61) should be added to the enabling condition 0 m+ F: This supervisor is maximally permissive because (60) and (61) are equivalent. The closed-loop system is no longer a Petri net. 8.2 Not all transition controllable In this section not all transitions are controllable. Here, the legal set is where for each i 2 and any of the sets A i is redundant. (Note that, in contrast to (58)-(59), 0 A i and 0 b i :) For these legal sets we dene a class of Petri nets with uncontrollable transitions for which a closed-form description of A is given. This extends the result of Li and Wonham (1994) to Petri nets of which the (acyclic) uncontrollable part does not have a generalised tree structure. Dene for each i 2 i as those of Denition 4 for A i in (63). From (63) and Propositions 12 and 13 in Appendix B it follows that A i Together with Proposition 7 it follows that P # Proposition 9 Consider A as in (62)-(63). If for each i 2 holds for the sets of Denition 4 that P Proof. See Appendix B. For the sets P the sub-Petri net NA can be built as in x5.3. Theorem 4 gives conditions under which A can be written in a form as in (62)-(63). Theorem 4 Given a legal set A as in (62)-(63). If (i) NA is acyclic, then the following proof species an algorithm for constructing row matrices A numbers b In contrast to the conditions (H1)-(H3) of Section 6 now there can be synchronisations in NA : Hypothesis (H4) is related to the precedence path condition in (Holloway and Krogh, 1990) and (Holloway et al., 1996). Proof. Assume that NA does not contain any source transitions, i.e. ;: The case T 1 6= ; is treated as in x7.1. Because the sets of reachable markings under the transition bag assumption and under the no concurrency assumption are equal, we can use the transition bag assumption to construct A : Because of (i), one can partition U in sets T Introduce the notation such that t 0 k (m) as shorthand for R ftg k (m): For each m 2 M; it follows directly from Corollary 3 and Lemma 2 that R U (R t 1(m)) Recursively dene for the sets of markings A A where A (see (63)). For each we will construct by induction matrices A and numbers b such that A A I A From (62),(63) and Proposition 9, it is clear that the assertion holds for 0: The induction hypothesis is that this is true for i: From (66), m 2 A By the induction hypothesis, this is equivalent to _ A Because is an acyclic sub-Petri net of NA (see (i)) and because of (iv) and Corollary 4, (70) is equivalent to (F(t i ) is the column of F which corresponds to t i ) _ A Next, we prove that A Because of (69) a sucient condition is that for each p 2 As NA is acyclic, it is true that F(p; t i it is true that t i I (see Proposition 6(i)), thus p 2 I: Because of (iii), it holds that t k 1 This implies that p 62 [ also have that p 62 I thus A Consequently, (71) is equivalent to _ A And this can be rewritten as: _ A where A I A The proof suggests a more direct method for constructing the set A : For each marking m in M a marking mm can be dened in the following way with for each The validity of the sequence in (74) follows from the proof of Theorem 4. For +)-polynomial 6 in the marking m of places in I (see also (Boel et al., 1998)). Moreover, m 2 A From which one obtains the result of the theorem. From the invariance property (this follows directly from (74)-(76)) U it follows that transitions in U do not have to be observed by the supervisor. Example 3. Consider the Petri net in Figure 5 with legal set From Proposition 9 it follows that P We have that and g: The sub-Petri net NA is acyclic with the marking mm is constructed as 6 In (min; +)-algebra ordinary addition is replaced by the minimum operation, and ordinary multiplication is replaced by the ordinary addition. c c p6 c p3 c c Figure 5: Petri net of Example 3. in (74)-(76): with other elements in 2 and 3 are zero. One obtains This also results from mm j inserting mm in (77) it follows that A Remark 3. For legal sets of the form rst Petri net supervisors with OR-logic as above can be constructed for each of the specications These controllers are then to be combined: the enabling rule for the controller part is described by the conjunction of the enabling rules for each of the A i -supervisors. 9 Conclusion We have structured the incidence matrices of acyclic Petri nets. This has resulted in a new proof for the well-known necessary and sucient condition for reachability in these Petri nets, and in the solution of other reachability issues. This result was further used to construct maximally permissive Petri net supervisors which guarantee that the marking of the original Petri net never leaves a given set. For a large class of Petri nets the method based on place invariants wass extended to Petri nets in which not all transitions are controllable. The original specication was algorithmically transformed into a new specication which takes the presence of uncontrollable transitions into account. Moreover, a class of systems was dened for which the maximally permissive supervisor can be realised by extending the standard Petri net enabling condition with OR-logic. Acknowledgement This paper presents research results of the Belgian Programme on Interuniversity Poles of Attraction, initiated by the Belgian State, Prime Minister's Oce for Science, Technology and Culture. The scientic responsibility rests with its authors. This work was partially supported by the European Community Esprit-LTR Project 26270 VHS (Verication of Hybrid Systems). R.K. Boel is supported by the Flemish Foundation for Scientic Research as Senior Research Associate. --R Introduction to lattices and order. Free choice Petri nets. Generalized mutual exclusion constraints on nets with uncontrollable transitions. Petri net supervisors for generalized mutual exclusion constraints. A survey of Petri net methods for controlled discrete event systems. A generalization of state avoidance policies for controlled Petri nets. Analysis and control of discrete event systems represented by Petri nets. The novel algorithm for determining the reachability in acyclic Petri nets. Control of vector discrete-event systems|Part 2: controller syn- thesis Petri net supervisors for DES with uncontrollable and unobservable transitions. Petri nets: properties Petri Net Theory and the Modelling of Systems. Controlled Petri nets and general legal sets. Reduction of the supervisory control problem for Petri nets. ieee Transactions on Automatic Control. Maximally permissive feedback and modular control synthesis in Petri nets with external input places. Feedback control of Petri nets based on place invariants. --TR
uncontrollable transitions;acyclic Petri net;petri net;supervisory control
593697
Regular Ordering and Applications in Control Policies.
In this paper we introduce the notion of regular ordering for periodic sequences based on the gaps between the entries. We define the notion of regular preserving functions using Schur convexity. This is used to extend some optimization results in queuing control problems. In particular, we show that the maximal traveling time in a stochastic event graph as well as the transmission times in a channel with redundancy, decrease (in the stochastic sense) when the input sequence becomes more regular.
Introduction A classical problem in the control of queues is to compute the optimal routing policy of a stream of incoming customers into several queues in parallel. Often, the goal is to minimize the average waiting time of the customers or the average load of the system. When the number of queues is not larger than two, then the problem has been solved in several cases. In the Markovian case, the optimal policy is a threshold policy and can be computed using dynamic programming (see [20]). When the system is not Markovian, the optimal policy is a Sturmian sequence as shown in [2] and te exact computation of the optimal policy has been done in the deterministic case in [17]. However, when the number of queues in parallel is larger than two, then the problem becomes more dicult. One of the reasons for this \three queue hiatus" is that Sturmian sequences in dimension three or more do not exist in most cases (see [2]). The only case where this problem has been solved so far is when the system is fully loaded and deterministic (see [22]). Here, we will deal with the problem of routing customers in several queues in parallel using a softer approach. We will not try to compute the optimal policy, which seems to be a very hard task. Instead, we will introduce a partial order on the routing sequences called the regular ordering. The main result of the paper is the following: If the routing sequence s is more regular than the routing sequence s 0 , then the maximal waiting times under s are smaller than the maximal waiting times under s 0 , for the stochastic order. The work of this author was partially supported by the special year on networking at Dimacs, Piscataway, NJ. This statement deserves several comments. First, the notion of regularity is close the the notion of balance in the sense that a balanced sequence is the most regular sequence possible (see Section 2.2). An integer sequence is balanced if its partial sums over two arbitrary windows of the same length diers by at most one. They have been extensively investigated in the past [24, 19] and lead to fruitful results in several elds such as theoretical physics [25], combinatorics [10, 9] as well as control theory [21, 16]. In a recent paper, [3], it is shown under rather general assumptions that balanced sequences minimize several cost functions in discrete event systems. This result relies on the following theorem. are multimodular functions, then the Cezaro limit lim N!1N is minimized by a balanced sequence. Second, majorization and Schur convex functions seem to be the right notions to use when comparing the dispersion of nite sequences [23]. However, majorization does not take into account the initial positions in the sequences to be compared since it is left invariant by the group of all permutations. It is often desirable to take the initial positions into account. For many systems, the input sequence certainly induces a dierent behavior than the input sequence (1; 0; which looks more balanced. However, these two sequences cannot be compared by using the classical majorization technique because the former is a permutation of the latter. To overcome this dif- culty, we introduce the notion of gap sequences (similar to partial sums, already used in [13]) that takes into account the initial positions in the sequence. This will narrow the gap between the notion of Schur convexity and multimodularity as shown in Appendix B. Finally, the usual performance measure in queuing networks is the average waiting time. However, in communication models where real time constraints are important (such as voice and video traf- c) the maximal waiting time (or sojourn time) are more important than the average. In Section 3, we show that for FIFO stochastic event graphs, the maximal waiting time (or sojourn time) is regular preserving. As for computational issues, nding good allocation patterns (i.e. very regular sequences), is possible via a mathematical programming problem with convex objective functions as in corollary B.3. A similar procedure for minimizing the average waiting time has been used by Combe and Boxma [15] and in [14] for dierent performance measures. The rest of the paper is organized as follows. In section 2, we introduce the notion of gap sequences and balanced sequences and the regular ordering. In section 3, we introduce a model of a controlled queue with periodic inter-arrival sequences and stationary service times. We show that the maximum waiting time is regular preserving w.r.t. the inter-arrival time as well as w.r.t the service times. Section 3.6 extends the result to a routing problem between several queues. Finally we show in Section 4, that the transmission rate on a link with redundancy is also regular preserving. Preliminaries This section is devoted to the denition of gap sequences, balanced sequences and to the introduction of the notion of regularity. 2.1 Gap Sequences We use the notation for the set of all non-negative integers and we consider the set P(T ; n) of all non-negative integer sequences size T which sum exactly to n. Let s 2 P(T ; n). We call (p s the positions in s of all its partial sums. We dene the vector d i (s) of the gaps of order i > 1 in the sequence s, by: When the sequence s is in f0; 1g T , then the gaps are never null and measure the size of the number of elements in the sequence s between the ones. This is illustrated in Figure 1. Figure 1: Gaps of order one and two for a sequence s in P(9; 3). 2.2 Balanced Sequences Denition 2.1. A sequence s is balanced if for all j, k and n, 6 1: Theorem 2.2 ([24]). If a sequence s in P(T ; n) is balanced, then there exists 2 [0; 1] such that for all j > 0, Proof. The proof comes from [24] adapted to the nite case. If s is a balanced sequence then the associated gap sequences have the following property. Lemma 2.3. If s is a balanced sequence in P(T ; n), then for all j; i, d i eg. Proof. By denition, d i By denition of p j , we have: jg. Therefore, This implies that Similarly, we have: As for d i j , we get by subtracting the two previous equations The following lemma in the converse of Lemma 2.3. Lemma 2.4. Let s be a sequence such that for all i, all the gaps of order i, d i dier by at most one. Then, s is balanced. Proof. Assume that s is not balanced. There exist two intervals of the same length, l. w 1 and w 2 such that the sum in w 1 , denoted n 1 is larger than the sum in w 2 , denoted n 2 , plus two: 2. This means that there exists j 1 and j 2 such that d n1 j1 6 l and d n1 2. Lemma 2.5. A sequence s 2 P(T ; n) is balanced if and only if d 1 (s) is balanced. Proof. The \only if" assertion: we consider the sequence d 1 (s). Let w kg be two windows of length k. Then the sums over w 1 of the sequence d 1 are of the form d 1 and the sums over w 2 are d 1 By lemma 2.3, then these two quantities dier by at most one. The \if" assertion: d 1 (s) is balanced means that there exists such that for all j, d 1 This means that Therefore, for any order i, the gaps of order i dier by at most one. By applying Lemma 2.4, s is balanced. Remark 2.6. Note however, that the gaps of order higher than one in a balanced sequence may not be balanced. For example, if d 1 (which is balanced), then d 2 (5; 6; 5; 5; 6; 6; 5) which is not balanced because it contains two consecutive 5 (which sum up to 10) and also two consecutive 6 (which sum up to 12). 2.3 Regularity and Schur Convexity If x 2 R n , then we denote by the components of x arranged in decreasing order. For x; y 2 R n , we say that x (resp. strictly) majorizes y (denoted x y) (resp. x y) if y [i] and (resp. with strict inequality for at least one m). A function f : R n ! R is (resp. strictly) Schur convex if x y f(x) > f(y)). For more details on the theory of majorization, see [23]. Denition 2.7. If s; s 0 2 P(T ; n), then s 0 is more regular than s (denoted s Lemma 2.8. The balanced sequences are the minimal sequences in P(T ; n) for the regularity order. Moreover, if s is balanced, then for all s Proof. This result is a consequence of Lemma 2.4. Indeed, if n divides iT , then for a balanced sequence , d i n , which is minimal. If n does not divide iT , then for the balanced sequence, c; d iT eg and is also minimal. Balanced sequences in P(T ; n) will also be called the \most regular" sequences in P(T ; n) in the following. Example 2.9. Let us compare three sequences in P(6; 3): s and s using the regular ordering. The respective gap sequences are: Using majorization, it is not dicult to check that According to the denition of the regular ordering, this means that s 3 Actually, s 3 is balanced and is the most regular sequence in P(6; 3). Denition 2.10 (regular-preserving functions). A function f : N T ! R is regular-preserving (r.p.) if x Lemma 2.11. A function f : N T ! R is regular-preserving if and only if it can be written under the form: is increasing in all coordinates and H are Schur convex. Proof. The only if assertion is a direct consequence of Lemma 2.8. As for the if assertion, then f is necessarily Schur convex in all d i . Now, the combination of all these arguments for all i shows that f is of the form given in the lemma. Note that if f is regular preserving on the set P(T ; n) then it reaches its minimal value for the balanced sequences. Remark 2.12. (Regularity without using the gap sequences) The denition of regularity, as well as regular preserving functions, can be done in a similar fashion by using the partial sum sequences of the original sequence s instead of the partial sums of its order one gap sequence. All the results given in this paper are still true in this new framework. The passage from one framework to the other is simply done through Equations (1) and (A.4). Application 1: Maximal Waiting Times in Networks In this part, we show that the maximum waiting time in a queuing system with general stationary inter-arrival and service times is a regular preserving function. 3.1 The D/D/1 Model First, we consider a slotted D/D/1 queue where the arrival sequence s is periodic with n customers arriving every T time slots (each of unit length). Therefore, s 2 P(T ; n) ! . By denition, the number of customers arrivibg at time i is s i . The i-th customer arrives at time p i (s) and brings a workload of in the system. We will also denote by C load brought by n consecutive customers. We nally require that the queue is initially empty and that the customers are served in a FIFO order (this last assumption can be relaxed to a non-idling server assumption in some cases). let us consider the waiting time W i of the ith customer that enters the system. We have f We dene M(s; Theorem 3.1. The function M(s) is regular preserving. Proof. First, let us remark that if C > T , then W In the rest of the proof, we will only consider the case where C 6 T . In that case, note that after a transient period of length n, the waiting times are periodic: W Also note that since the system is initially empty, the maximum waiting time is not reached during the initial transient period. So we have f Equation (2) follows from the fact that C 6 T . Equation (3) comes from Equation (18). Changing the order of the two max operators yields Equation (4). Now, the function d j 7! max n i Tg is Schur convex for each j and the function M is increasing in d j for all j, and thus satises the conditions of Lemma 2.11. This means that M is regular preserving. Corollary 3.2. Among all inter-arrival sequences in P(T ; n), the balanced sequence minimizes the maximum waiting time in the system. Moreover, for two sequences s; s Remark 3.3. The function M(s) is multimodular (see Appendix B). This is a direct consequence of Corollary B.3. 3.2 Characterization using regular preserving functions The function M when plays an important role. Indeed, we have the following result giving yet another characterization of balanced sequences. Theorem 3.4. If balanced. Proof. When Note that jT=n is the average value of d j (s). Using Lemmas 2.3 and 2.4, s is balanced is equivalent to the fact that M(s) < 1 3.3 The average waiting time is not regular preserving In this section, we will show that unlike for the maximal waiting time, the average waiting time in a stable D/D/1 queue is not regular preserving through an example. Consider in P(16; 5) the sequences s and s 0 with the respective one-order gaps, It is immediate to check that s D s 0 . However, if one computes the average waiting time in both cases, when the load brought by each customer is equal to 3, one gets: Remark 3.5. The function W is multimodular (see [1]). Therefore, it is minimized by a balanced sequence in P(16; 5): W if u is a sequence such that d 1 3.4 The G/G/1 Model Here, we generalize the previous model. The deterministic arrival process is replaced by a stochastic process with stationary interarrival times: f i g i2N . The load carried by the k-th customer is denoted by k . The sequence f k g k2N , is stationary and independent of the arrival times. We denote by i (s) the time elapsed between the arrival of customer i and customer i +1 for sequence s, ps (i+1) j=ps (i)+1 Using Lindley's equation, the waiting time of the i-th customer satises f If we consider the maximal waiting time within on period, starting with customer m+ 1, we dene In a similar fashion as with the D/D/1 case, we dene f Intuitively, n r;m is constructed by going back in time for r customers only, while n goes back to time Note that if the queue has emptied during the last r arrivals, then both quantities coincide. We assume that the queuing system is stable, that is TE( 1 generalizes the previous assumption, C 6 T to the stochastic case). Under this stability condition, the queue empties almost surely in nite time. Therefore, taking r large enough will ensure that n r;m and will coincide. More precisely, n an;an an Now, we have, an;an k=i an k=i an Lemma 3.6. Let us assume that s and s 0 are two sequences in P(T ; n) such that s E s 0 . Then, an;an st n an;an Proof. For a given pair (i; ng f1; ang, we know by denition of the regular ordering that there exits ng such that d j Now, we dene a coupling of the service sequence as well as of the arrival sequence such that N. Note that the coupled sequences have the same probability distributions as the original ones because of stationarity and independence between them and because the sequences and are independent of s and s 0 . Using this coupling, and considering Equation (8), we have directly i+an X Since during a periods, the total number of arrivals under s and s 0 is the same and using the coupling, we have i+an 1 ps (i+an) X p=ps (i)+1 ps (i)+Ta X p=ps (i)+1 Finally, using the fact that d j ps (i+j) p=ps (i)+1 ps (i)+d j p=ps (i)+1 Using these three relations, an i+an X i+an 1 This is true for all i, and therefore also for the maximum over all possible i from 1 to an. Using the fact that under the coupling, the sequences 0 and 0 are in a \typical" situation, we get an;an st n an;an Now, the stationary version of the queue is considered. Using a backward coupling argument and Lemma 3.6, we obtain the following theorem. Theorem 3.7. Let W i be the stationary waiting time in a stable G/G/1 queue. Then for two sequences s; s 3.5 The Event Graph Model We further generalize the previous model in several ways. Networks assumptions. We replace the single queue by a network made of Q servers (tran- sitions) and P buers (places) forming a FIFO stochastic event graph with a single entry. (see [7] for more details on this class of networks. In particular, this class includes all queues in tandem with general blocking. Stochastic assumptions. The service time at node j for the k-th customer is denoted by j k . The sequence f j k g k2N;j2Q is stationary and independent of the arrival sequence. However no independence assumption among the service times is required. Initial state assumptions. We assume that initially, the event graph is empty, in the sense that there is not any customer in any buer. (this is a generalization of the empty queue assumption). Without loss of generality, we can further assume that each buer contains at most one customer initially. To any stochastic event graph, we can associate a family of matrices, A('), of size QQ. The entry (i; in matrix A(') is the maximum over all the sums of the service times of the '-th customer, ' , in all the nodes k on any path from node j to node i with no initial tokens, except at node j, and is 1 otherwise. In the following, we will be using the (max; +) notation, where stands for max and for +. Matrix operations are dened accordingly. For more details on (max,+) notations and its use in event graphs, see for example [7]. The matrices A(') exist and are nite whenever the network cannot reach a deadlock state (see [7]). In that case, let X q (n) be the epoch when the n-th customer completes its service in server q. Then the vector the following recurrence equation: Let Wn be a vector , with its q-th component equal can be seen as the traveling time for customer n between its entrance in the system and its passage in server q. Lemma 3.8 ([1]). Wn satises the following equation: where D(h) is the diagonal matrix with h on the diagonal and 1 everywhere else, and B(n) is a vector which describes the input connection. Equation 9 can be seen as a vectorial form of the Lindley's equation. Its solution is similar to the solution of the scalar case. O If we dene as previously, n O We get by a similar technique as in the G/G/1 case (with a vectorial notation), n an;an an provided the event graph is stable (see [6]). The quantity n an;an can be transformed using the fact that diagonal matrices commute with everything. an;an O O O an O B(an j i+an 1 O an O B(an j i+an 1 O O Lemma 3.9. Let us assume that s and s 0 are two sequences in P(T ; n) ! such that s E s 0 . Then, an;an st n an;an Proof. The proof is similar to the single queue case, using a coupling argument and the denition of the regular ordering. For a given pair (i; ng f1; ang, we know by denition of the regular ordering that there exits ng such that d j Now, we dene a coupling of the service sequence such that for each queue q, 0 q k+i . As for the interarrival sequence, we set 0 Note that the coupled sequences have the same probability distributions as the original ones because of stationarity and independence between them and because the sequences and are independent of s and s 0 . Using this coupling, we have by denition of A(k), i+an 1 O B(an O B(an During a consecutive periods, the total number of arrivals with s and s 0 is the same. Therefore, we have using the fact that D( k (s)) are diagonal matrices, i+an 1 O ps (i+an) O p=ps (i)+1 ps (i)+Ta O p=ps (i)+1 O O As for the last term, O ps (i+j) O p=ps (i)+1 ps (i)+d j O p=ps (i)+1 O O These three relations are true for all i and j, therefore, considering Equation (14), n an;an st an;an Finally, we obtain the following extension of Theorem 3.7. Theorem 3.10. In a stochastic event graph, the maximum traveling time M(s) over one period of the admission sequence is a regular preserving function (in the stochastic sense). 3.6 Routing Problem Now, we consider a network of K identical queues in parallel. Customers enter a system composed by K queues in parallel. The routing of the customers is controlled by a sequence of vectors fS i g, with means that the i-th customer is routed to queue j. Note that S is a feasible admission sequence as long as Customers arrive in the system at each time unit and each customer brings a constant load of . Figure 2 displays the model we are considering. 3.6.1 The Constrained Case We consider periodic routing policies S satisfying the following constraint: are xed with In other words, S We denote this ste of policies by Figure 2: Routing of customers in a multiple queue system Now, we dene I(S) to be the maximum waiting of any customer in the system, for S 2 is the maximum waiting time in queue j. We say that the vector (n 1 ; ; nK ) is balanceable if there exists S 2 A(n 1 ; ; nK ) such that for all j, S j is balanced. See [2] for a more detailed discussion on balanceable vectors. From Lemma 2.8 and 3.1, we get, Lemma 3.11. If (n 1 ; ; nK ) is balanceable then I(S) is minimized by a routing policy S such that each S j is balanced. Proof. This result is a direct consequence of Lemmas 2.8 and 3.1. This result is similar to some extend to the results proved in [2] where multimodularity of the waiting times in each queue is used. However, we get a new result here which follows directly from the previous lemma and which holds for all (n necessarily balanceable). Theorem 3.12. If such that for all j, S j D U j , then I(S) > I(U). 3.6.2 The Unconstrained Case Now, we do not x the number of customers sent to each queue. We want to consider all admission sequences S in the set Lemma 3.13. The function I(S) is minimized on A for some balanced S Proof. First, note that if n (see [2] for a proof). Let b be a vector with such a property. Second, note that if not all n j 2 dT=Keg, then at least one of then is larger than dT=Ke. Finally, note that if s is a balanced sequence in P(T Now, the proof goes like this. By Theorem 3.12, I is minimized on A(b) by a balanced sequence S(b). Moreover, there is some j such that dT=Ke. For any U which is not in A(b), then there exists k such that n k > dT=Ke. This nishes the proof 3.7 Computational Problems For the constrained problem, we can compute the best sequence using a procedure similar to the one used in [13] with quadratic programming techniques. As for the unconstrained case, from Lemma 3.13, it is easy to see that when T=K 2 N, then the best routing policy (which minimizes the function I) is the round robin policy. Similar results can be found in [2]. However, when T=K 62 N, then little was know before. Here, we can come up with the optimal routing sequence using the following procedure. First, we compute the optimal sequence using Lemma 3.13. This sequence is unique up to a permutation. Note that by symmetry of the cost function, all those permutation will perform the same. Moreover, since all balanced sequences with the same number of appearance in the routing sequence are shifts of each other, then for all j such that n dT=Ke, all the waiting times are equal. The construction of S j (b) for any j can be done by any known method to construct balanced sequence, as for example, by using the Beaty construction (see [11]) or the iterative formula given in Theorem 2.2. 3.8 Development of an example Let us assume that we want to route customers into three parallel event graphs dierent service times. For some reason, the proportions of customers sent to must be 1=2; 1=3 and 1=6 respectively. This is a case where the routing policy is constrained to stay in the set A(3; 2; 1) with a period equal to The proportions 1=2; 1=3; 1=6 are not balanceable. Therefore, it is not possible to nd a routing sequence which is balanced for each event graph. However, we can use the regular ordering to compare several policies. If we consider the sequence of customers sent in E 1 we get S In example 2.9, we showed that S 0 If we consider the sequence of customers sent in E 2 we get S (0; 0; 2; 0; 0; 2). The sequence S 0 2 is balanced. Therefore, S 0 Finally, if we consider the sequence of customers sent in E 2 we get S which are equal. Since the sequence S 0 is more regular than S for all the event graphs, then the maximal waiting time under S 0 is smaller than the maximal waiting time under S (in the stochastic sense). This is true for all the distributions of the service times in the three event graphs. However, if we consider S 1 but S 0 . The routing sequences are not comparable for the regularity ordering. For some distributions of the service times in the three event graphs then S 0 will be better than S 00 and for some other distributions, S 00 will be better than S 0 . Application 2: Assignment to one-buer queues with re- dundancy In this section, we present an example of optimal control of admission into a single buer queue with redundancy. { Transmission opportunities occur at time T At each transmission opportunity, a controller can decide to actually transmit a packet, or not. If transmission occurs, we assume that it is instantaneous. Dene { At a down link node in the network a packet that has been transmitted is buered in a single buer till its service there is completed. The duration of the service is exponentially distributed and it has expectation 1 . We assume that the delay between the transmitter and the down link node is zero (the results of this paper do not change if we take this time to be any other constant period). { If a packet is transmitted before the previous packet has completed its service time then it is lost. A model that can be used to solve this problem was introduced and studied in [5]. Here, we further focus on the following problem arising from higher layer network considerations. We consider the problem of redundancy of packet transmission: each packet is retransmitted at K consecutive transmission opportunities, which we call a frame, so as to decrease the probability of losses. Thus, only if all the K packets in a frame are lost then the frame, or equivalently the original information packet, is considered to be lost. Our goal, roughly speaking, is to obtain a control with two objectives: on one hand it maximizes the average throughput, and on the other it minimizes the loss probabilities of frames. Since the throughput is the acceptance rate minus the losses, the problem can be formulated as maximization of the average number of accepted packets, on one hand, and minimizing the loss probabilities of frames, on the other. We now formulate more precisely our problem. We dene a control policy as a sequence fs i g, a transmission is scheduled at the ith transmission opportunity, and it is 0 otherwise. We consider here the case where the controller has no information on the state of the buer (nor on which packet is the beginning of a frame). . The actual kth transmission occurs at times p s (k) (see Denition 1). We dene the process X k to be the number of packets in the buer just prior to time p s (k). Dene as where is a negative constant. We consider the problem of minimizing G over all sequences s. Note that D(N)=N is the acceptance rate, whereas the second term is the average number of frames lost among the rst N ones. The cost thus contains a term responsible for maximizing the throughput and a term for minimizing the loss rates. The parameter can be viewed as a Lagrange multiplier related to a constrained problem; indeed, in [5] a similar problem (for the case studied where the loss rate is to be minimized subject to a constraint on the rate of acceptance of packets. Next, we focus on the cost G(s). We shall restrict to policies s that are periodic. (Recent techniques in [12] can be used to show that restricting to periodic policies is without loss of optimality.) Our main result, stated below, is that G(s) is regular-preserving. (This allows us to conclude, in particular, that there exists an optimal sequence which is balanced.) Theorem 4.1. Assume that k are stationary and that the service times are i.i.d. exponentially distributed, and independent of the sequence k . Let s 2 P(T ; n) ! . Then (i) G can be rewritten as where d k i is dened in (1). (ii) G(s) is regular-preserving for n > K . Proof. (i) Let n denote the length of the time interval that started when the (n K+1)th packet was transmitted, and ended when the nth packet is transmitted. fn only if there has been no service completion during n , and is otherwise 1. Thus, the expectation of fn (s) equals the probability that an exponential random variable with parameter is greater than or equal to n . Using the periodicity of s, this yields (i). (ii) This follows directly from Theorem B.2. Note that since g 0 is convex in x, A Appendix: properties of the gap sequences This appendix gives several properties of the gap sequences which have some interest by their own. The two following lemmas are straightforward consequences of the denition of the gaps. Lemma A.1. The gap sequences satisfy the following properties: Proof. The proof is a straightforward consequence of the denition of the gaps. Lemma A.2. Let s and s 0 be two sequences in P(T ; n). If s 0 is a shift of s or a mirror of s (i.e. j+k or s Proof. The proof also follows directly from the denition of the gaps. Remark A.3. The converse of Lemma A.2 is false in general, as shown by the following counter example choose s and s 0 such that d 1 s 0 is not a mirror nor a shift of s. However, for all order i, we have fd i )g. We will prove in the following (see Remark A.6) that the converse of Lemma A.2 is true in the special case of balanced sequences. Lemma A.4. If s 2 P(T ; n), then d 1 is a shift of s. Proof. By using Equation (17), then d 1 be a shift of s such that and such that d 1 We will show that d 1 (d 1 ig. Now we compute We can construct the dierent gaps by computing in the set Z[X ]=(X n 1). Indeed, from the set (s), we can dene the polynomial Note that by denition, we have P i Lemma A.5. Let i > 1. The original sequence s can be retrieved from d i (s) if and only if i and are relatively prime. Proof. There is a one to one correspondence between s and d 1 (s). Therefore, we rather show that the knowledge of d i enables one to compute d 1 (s) if and only if i and n are relatively prime. If i and n are relatively prime, then the polynomials X n 1 and X are relatively prime. Using the Bezout equality, there are two polynomials u(X) and v(X) in Z[X] and a non-zero integer k such that 1). Therefore, This counter example was provided to the authors by Jer^ome Galtier by multiplying by P 1 s (X), we get kP 1 s (X)u(X) mod (X n 1). The knowledge of P 1 induces the knowledge of d 1 (s), which in turn gives all gaps by using Equation (16). On the other hand, if i and n are not relatively prime, then gcd((X and if we denote by q(X) 4 1). Therefore, two dierent gap sequences of order one are compatible with the gap sequence of order i. Remark A.6. Note that if for a sequence s, all the gaps of all order i dier by at most one and are balanced by lemma 2.4. On the other hand, all balanced sequences are shift of each other (see Theorem 2.2) (It is interesting to see that in this case, all mirror transformations are also shifts). This is an answer to the question asked in Remark A.3 in the case where the gap sequences are all formed by two consecutive numbers. relations between regularity and multimodu- larity Multimodularity is a notion of discrete convexity which has been introduced in [21]. It has been used extensively in the context of the admission control of queues [3, 1, 2, 4, 18]. Therefore, it is interesting to see if regular preserving functions have some kind of discrete convexity properties. In this appendix we will show that all regular preserving functions are multimodular in some sense whereas the reverse is not true. denote the vector with all components equal to 0 except the i-th component which is equal to 1. Dene b g. Since we deal with sequences which sums are xed to n, we use an adapted version of the denition of multimodularity. Denition B.1. A function f on P(T ; n) is multimodular with respect to F if for all This is not the classical denition given by Hajek in [21], which considers arbitrary sequences in N T . However, both denitions are closely related. One can show that the projection of a multimodular function (in the sense of Hajek) on P(T ; n) is multimodular according to Denition B.1. (see [3] as well as [8] for a deeper insight on the denition given here.) Theorem B.2. Let f(s) be a function that only depends on d i (s), which is also denoted by can be put under the from is a convex function, then it is is multimodular on P(T ; n). Proof. According to the denition of multimodularity, we need to check the inequality (19) for all First note that for all 1 6 k 6 T , there exists j k such that: Now, for any 1 6 k; l 6 pn+i such that j k 6= j l i, then b l and b k modify dierent gaps. Therefore, For This is positive since g is convex. The following corollary covers the case when several gap orders are used. Corollary B.3. If for all are convex functions, then dene the function where c i are non negative constants. Then f is multimodular. Proof. For all i, the function f i positive linear combination of multimodular functions is multimodular. Remark B.4. Note that the functions f i a \characteristic set" of regular preserving in the same way sums of convex functions play an important role among Schur convex functions ([23]). Remark B.5. Functions of the form (21) have been used in [13] as a criterion to test regularity of a sequence. In that paper, the minimization problem of functions of this form is written as under quadratic assignment formulation. This makes computations possible for small n and T and can be used for optimization purposes. In order to prove that a balanced sequence minimizes a function f as in (21) ( f(s) is regular preserving and multimodular), one can choose two dierent methods. First, one can try to use the fact that f(s) is multimodular in N T . By referring to a general (complicated) theorem proved in [21, 3], f(s) will be minimized by a balanced sequence and then use the fact that if s is balanced, then s is a balanced sequence in P(T ; n). Another way to prove the same result, it by using the fact that f(s) is regular preserving directly. This second proof is easier and has a clear advantage on the previous approach since it provides a more precise result. It says that if s D s 0 , then f(s) > f(s 0 ). This fact will be illustrated in the following section, where we show that a given cost function is ri. The minimization will then derive directly from Lemma 2.8 as well as the comparison between several policies. Remark B.6. Finally, one may remark that regular preserving functions constitute an essential class among multimodular functions. Indeed, as shown in Section 3.2, if a given sequence in possible regular preserving functions, then this sequence in balanced. --R Admission control in stochastic event graphs. Balanced sequences and optimal routing. Multimodularity, convexity and optimization proper- ties Optimal open-loop control of vacations Optimal admission Towards a uni Complexity of trajectories in rectangular billiards. The book of L Mots, chapter Trac Optimal routing problems and multimodularity. Optimization of template-driven scheduling mechanisms Optimization of template-driven scheduling mechanisms Optimal allocation sequences of two processes sharing a resource. Optimal routing policy in two deterministic queues. Monotone optimal control of permutable gsmps. Covering the positive integers by disjoint sets of the form f Optimal control of two interacting service stations. Extremal splitting of point processes. Dinard Van Der Laan. Theory of Majorization and Its Applications Symbolic dynamics II- Sturmian trajectories Quasicrystals and geometry. --TR
admission control;regular ordering balanced sequences;schur convexity;plus algebra
593969
Fusion Via a Linear Combination of Scores.
We present a thorough analysis of the capabilities of the linear combination (LC) model for fusion of information retrieval systems. The LC model combines the results lists of multiple IR systems by scoring each document using a weighted sum of the scores from each of the component systems. We first present both empirical and analytical justification for the hypotheses that such a model should only be used when the systems involved have high performance, a large overlap of relevant documents, and a small overlap of nonrelevant documents. The empirical approach allows us to very accurately predict the performance of a combined system. We also derive a formula for a theoretically optimal weighting scheme for combining 2 systems. We introduce dthe difference between the average score on relevant documents and the average score on nonrelevant documentsas a performance measure which not only allows mathematical reasoning about system performance, but also allows the selection of weights which generalize well to new documents. We describe a number of experiments involving large numbers of different IR systems which support these findings.
Introduction In the past, neural network models which have been applied to the Information Retrieval problem have typ- Research was supported by UC Senate Bridge Grant #RW252G/B. ically used the very large feature vectors (document and query vectors) which are traditionally used in IR systems (see for example, [Crestani, 1994], [Wong et al., 1993], [Boughanem et al., 1993]). Unfortunately, the 2 Vogt and Cottrell resulting large networks generally require large numbers of training examples, a rare commodity in the IR setting. Although work using Latent Semantic Indexing [Deerwester et al., 1990] to reduce the number of features has met with some success [Vogt et al., 1997a], LSI is itself computationally expensive. Perhaps a wiser approach can be found in fusion, where the results from multiple IR systems are combined to generate a single (hopefully better) list of potentially relevant documents in response to presentation of a single query to a number of component systems. Fusion allows a significant reduction in the number of features, often to just one feature per system - typically the system's estimate of the document's probability of relevance. As such, a fusion neural net-work model can be much smaller than a traditional one based on document vectors. Fusion also allows leveraging of the component systems in several ways by exploiting a number of effects (from [Diamond, ffl The Skimming Effect happens when "retrieval approaches that represent their collection items differently retrieve different relevant items, so that a combination method that takes the top-ranked items from each of the retrieval approaches will push non-relevant items down in the ranking." ffl The Chorus Effect occurs "when several retrieval approaches suggest that an item is relevant to a query.this tends to be stronger evidence for relevance than a single approach doing so." ffl The Dark Horse Effect in which "a retrieval approach may produce unusually accurate (or inac- curate) estimates of relevance for at least some items, relative to the other retrieval approaches." It should be noted that when choosing how to combine the results from different IR systems, the Dark Horse Effect is at odds with the Chorus Effect. Likewise, a large Chorus Effect cuts into the possible gain from the Skimming Effect. These phenomena argue for a sophisticated fusion model which is able to predict when these effects will occur and take advantage of them. Furthermore, the expansion of the Word Wide Web has spawned a proliferation of search engines. Combining them would allow one to exploit the above effects. Such simple neural network fusion models may allow easier training, but their simplicity inflicts a penalty - they lack expressive power. Thus, it seems likely that the improvement in performance possible from these models would be limited. In this article, we examine in detail one such fusion model: the linear combination of scores (LC). The LC model has been used by many researchers with varying degrees of success [Bartell et al., 1994], [Kantor, 1995], [Knaus et al., 1995], [Sel- berg and Etzioni, 1996], [Shaw and Fox, 1995], and Fusion Via a Linear Combination of Scores 3 [Vogt et al., 1997b]. Our analysis of the model reveals what types of systems the model works best with and explores techniques for training the model. This article has the following format. First we describe the model and the specific problems we are ex- amining, along with the data we use and assumptions. Next, we both empirically and mathematically derive explanations for when it makes sense to use the LC model. This also leads to hard limits on the performance of an LC fusion system. Finally, we describe experiments in training the LC model, and describe a technique which generalizes to large numbers of systems as well as more complicated models. Through these analyses we gain insight into why and when the LC model works well. 2. Background 2.1. The LC Model The linear combination model calculates the real-valued relevance ae of a document x to a query q based on the weights given to each individual system and their estimates of relevance: ae(w; systems This value is then used to rank the documents. For only two IR systems, this simplifies to: However, all that really matters is the ranking given by the combined system. Thus, for the case of two systems, only the ratio of the two weights and the relationship of the signs on the weights are important, and equation (1) can be replaced by one using a single weight: ae(w; The LC model is more flexible than others which have been tried in the past. Most of these consist of a single choice of parameters, such as the sum of scores or the maximum score, or a fixed weighting based on individual system performance (see [Belkin et al., 1995], [Shaw and Fox, 1995], [Kantor, 1995]). The difference here is that we will use a search procedure to optimize the weights. The LC model is also equivalent to the simplest kind of neural network, a single layer net. The addition of a squashing function on the output unit does not add any power, as it would not change the ranking induced by the combined system. 2.2. The Data Our study of fusion necessitates the availability of a large number of IR systems. We have chosen to use the entries to the TREC5 adhoc track [Harman, 1997]. TREC (Text REtrieval Conference) is an annual conference sponsored by the National Institute of Standards and Technology, in which participants are given very large text corpora and submit the results of their 4 Vogt and Cottrell retrieval techniques in a sort of contest. Specifically, in the adhoc track, each participant submits the top 1000 documents returned by their system in response to 50 queries supplied by NIST, and each participant can submit up to 4 runs. 61 runs were submitted, for a total of 3050 lists of 1000 documents, each of which we treat as a separate IR "system." Although most of our work involves examining the routing problem, we choose the adhoc entries because there are far more of them than routing entries. One issue arises when combining such lists of top-ranked documents - what score should be given to documents returned by one system but not the other? We have assumed that for such documents, the system which did not return them gave them a score of zero. In so doing, we also had to eliminate any negative scores for documents, because otherwise the unreturned documents would get ranked above those with negative scores. We did this by adding the absolute value of the lowest score to all documents for any system which had negative scores. We believe that zero scores for unseen documents is a reasonable choice - the vast majority of documents are not relevant, and most systems give a zero score to nonrelevant documents. 2.3. Problems Addressed A common dichotomy used by fusion researchers is the difference between data fusion and collection fu- sion. The former takes place in a setting when all of the IR systems involved have access to the same text collection. The latter is used when the collections searched by all IR systems are disjoint. Because of our use of TREC entries, our work necessarily falls under the label of data fusion, but it is not data fusion in the purest sense, since we do not have scores for all of the documents in the corpus. Another distinction is in the type of IR problem solved. TREC distinguishes two main problems: ad-hoc and routing. Adhoc retrieval occurs when the text collection is relatively static, and new queries are constantly being submitted by users. In contrast, routing has a standing query, and new documents are arriving which need to be filtered according to that query. The difference between these two tasks is important in the context of the LC model. For the routing task, a new set of weights can be trained for each standing query. However, for the adhoc task, one set of weights must be applied for all queries. Clearly, the routing task should be easier to solve using this model, since the weights can be tuned on a per-query basis. 3. Limitations of the LC Model Using the LC model does not always result in an improvement in performance. Previous work using the TREC data has shown that even on the training set, significant improvement is achievable for less than half of the possible combinations [Vogt and Cottrell, 1998a]. Fusion Via a Linear Combination of Scores 5 Why is it that it works sometimes and not others? When does the LC model work? In this section we use two techniques, one empirical and one analytical, to determine the source of the LC model's power. In the process, we gain a deeper understanding of why it works, its limitations, and which systems can be successfully combined using the LC model. 3.1. Empirical Analysis We introduce here a technique for analyzing the behavior of fusion models and applied it to the LC model. The technique involves measuring various properties of the component IR systems, and using them in a linear regression to predict the performance of the combination [Vogt and Cottrell, 1998a]. By examining how the measures are weighted by the regression, we gain an intuitive feel for when using the model pays off. Our technique works surprisingly well - the resulting regression can predict the performance of unseen combinations very accurately (on the test set, r 3.1.1. Method Our data set is all 61 TREC5 entries on the first 20 queries - 1220 lists of up to 1000 docu- ments. We begin by making a number of measures of all 36,600 pairs of systems/queries. These measures are meant to indicate either how well each system performs or how similar the two systems are to each other. The performance measures include average precision (indicated by P) and a statistical measure of rank correlation between the system and the relevance judgments (J) [Bartell et al., 1994]. J is defined as: ae(w; indicates the user prefers document x to document x 0 on query q. Note that J has a maximum value of 1 when the numerator and denominator are the same (i.e., the IR system ranks documents exactly as the user would), and a minimum value of -1 when the opposite is true. These two measures are subscripted a for the better of the two systems and b for the worse. The pairwise similarity measures include: ffl Guttman's Point Alienation (GPA) - a measure of how similar two rankings are to each other and which is calculated as: ffl the number of documents in the intersection of the two lists of returned documents (") ffl the correlation coefficient from a linear regression of the scores of documents in the intersection of the two systems (C) ffl the number of relevant documents returned by one system but not the other divided by the total num- 6 Vogt and Cottrell ber of relevant documents returned by that system (U for uniqueness), ffl Lee's [Lee, 1997] overlap measures, O rel and O nonrel , which measure the proportion of relevant and nonrelevant documents in the intersection of the two lists. These two measures are calculated as: O O where R i is the number of relevant documents returned by system i, and N i is the number of nonrelevant. Also, because it seemed likely that measuring the similarity of the two systems on nonrelevant documents is less important than on relevant ones, the first three measures were also calculated using only relevant doc- uments, and are denoted: GPA rel rel . One last measure, GPA ni (for "not irrelevant") is the GPA using pairs of documents where at least one is relevant. After normalizing the scores for each system on each query by dividing by their respective means we found the optimal combination for each possible pair. For each query and each pair of systems, the single weight w was chosen by optimizing average precision using golden section search [Press et al., 1995], and the best w was used to generate a combined system (of 1000 documents) according to equation (2). Because systems were combined on a per-query basis, this experimental setup most accurately simulated the routing task. We then performed a multiple linear regression using the aforementioned measures as predictor variables and the average precision of the optimal combination as the target. 80% of the pairs (29,280 total - the "training were used in the regression. 3.1.2. Results Table 1 presents the results of the multiple regression. Measures are sorted by decreasing F value, indicating roughly how important each measure is in predicting the average precision of the optimally combined system. All measures above the horizontal line in the table contribute to some degree (as indicated by F values much larger than 1). The indicates that the fit of the model is very accurate. Furthermore, the model generalizes extremely well to new data - when the remaining 20% of the pairs (the "test set") were plugged into the model, Positive regression coefficients in Table 1 can be interpreted as meaning that the correponding measures should be maximized in order to maximize the performance of the combined system. Likewise, measures with negative coefficients should be minimized. This leads to the following conclusions: the better system should have high performance (P a and J a have large positive coefficients), whereas the performance of the Fusion Via a Linear Combination of Scores 7 Table 1. Results of Linear Regression Normalized Regression Measure Coefficient F Pa 0.8993 129141.5501 Ua -0.1202 405.5097 Ja 0.0431 346.1357 GPA rel -0.0359 220.1937 O rel -0.0519 55.8835 C rel 0.0125 35.8910 O nonrel -0.0427 20.9289 " rel 0.0088 17.5199 GPA ni -0.0099 8.9850 Results of a linear regression for predicting the combination's average precision (r 2 =0.94). Positive coefficients indicate the measure should be maximized, negative coefficients indicate it should be minimized. worse system may or may not be good (Pb and Jb have opposite signs). The positive coefficients on GPA and C rel indicate that the two systems should generally rank documents in their intersection similarly and the distribution of scores by both systems should be similar to each other. On the other hand, the negative coefficients on GPA rel and GPA ni indicate that each system should rank relevant documents differently than the other system. Finally, the negative coefficient on O nonrel means that the two systems should retrieve different sets of nonrelevant documents. The table leads to conflicting conclusions about the overlap of relevant documents. The negative coefficients for U a and U b indicate that both systems should not return unique relevant documents, whereas the negative coefficient on O rel indicates they should. As it turns out, the negative coefficient on O rel is inaccurate because O rel is directly related to U a and U b by2O rel . The regression simply accounted for the effect of uniqueness using U a and U b alone. In fact, by repeating the regression using only O rel and O nonrel , we can predict the combined system's precision with nearly the same accuracy as the original regression (r 2). This table shows that the following three conclusions about when it makes sense to use the LC model are strongest, namely when: ffl at least one exhibits good performance, ffl both return similar sets of relevant documents, and ffl both return dissimilar sets of nonrelevant documents 8 Vogt and Cottrell Table 2. Results of Linear Regression on a Subset of Predictors Normalized Regression Measure Coefficient F Pa 0.9366 191543.1029 O rel 0.1021 2249.4031 O nonrel -0.0581 975.4101 3.2. Mathematical Analysis The conclusions in the previous section were arrived at empirically, and give little insight into why they may be true. We now provide a mathematical justification for these hypotheses. Furthermore, we mathematically derive an equation for the optimal weighting and from this derivation arrive at an equation for the performance of an LC model. As in the empirical case, we concentrate on the routing problem only. 3.2.1. d as a Performance Measure Our analysis hinges on the use of d as a performance measure. d is equal to the difference between the mean score on all positive examples (relevant documents) - p and the mean score on all negative examples (nonrelevant doc- As such, d is most easily applied to evaluating the routing task, where documents need to be placed in one of two categories. Without loss of gen- erality, we assume the scores have been normalized to the interval [0; 1], to make values of d comparable across systems. d is equal to the numerator of d 0 from signal detection theory [Swets, 1996]. d 0 divides this difference by the standard deviation of the negative example score distribution (d oe N A variation of called the Swets measure has been examined before in the information retrieval setting ([van Rijsbergen, 1979], p.157): However, despite their statistically based theoretical attractiveness and the excellent argument put forth by Swets, neither d 0 nor S have ever caught on as a basis for performance evaluation in IR. We now justify our use of d as a performance measure by relating it to d 0 . First we note that IR systems typically distribute scores according to an exponential distribution. Figure 1 shows the empirical distribution of two typical IR systems from TREC5 after scores have been normalized to [0; 1]. Note that these distributions are summed over all 50 queries for each system. Over 60% of the 61 entries have distributions similar to these two. Figure shows two atypical distributions. The typical distribution not only has the appearance of an exponential, but as Figure 3 shows, it also has the property that its mean is approximately equal to its standard deviation - a property of all exponential distributions. Note that this property is stronger for the negative score distribu- Fusion Via a Linear Combination of Scores 920060010001400 Relevant Scores - Cor5A1se Nonrelevant Scores - Cor5A1se20060010001400 Relevant Scores - vtwnB1 Nonrelevant Scores - vtwnB1 Fig. 1. Smoothed Histogram of Typical Relevant and Nonrelevant Score Distributions (from entries Cor5A1se and vtwnB1) Relevant Scores - INQ301 Nonrelevant Scores - INQ30120060010001400 Relevant Scores - uwgcx0 Nonrelevant Scores - uwgcx0 Fig. 2. Smoothed Histogram of Atypical Relevant and Nonrelevant Score Distributions (from entries INQ301 and uwgcx0) Fusion Via a Linear Combination of Scores 112610-0.6 -0.4 -0.2 0 0.2 0.4 0.626101418 Fig. 3. Histograms of the difference between the mean and the standard deviation for positive example scores (-p \Gamma oe P , left) and negative example scores (-n \Gamma oe N , right) on the 61 TREC entries. Note that both are centered around 0 and the negative score histogram has smaller variance. tions (the negative score histogram has lower variance), indicating that it is closer to an exponential. Because of the aforementioned properties, it is reasonable to substitute the mean of the negative score distribution for the standard deviation in the equation for We note that maximizing d is the same as maximizing this approximation to a measure which has already been well defended for use in IR [van Rijsbergen, 1979], [Swets, 1996]. To see this, simply take the logarithm of both sides of the approximation after ignoring the subtraction of a log(-n). We further justify the use of d bynoting that on the 61 entries and 50 queries from TREC5, a linear regression of d to predict average precision yields an r Figure 4 shows this relationship. Furthermore, as we will show later, optimizing d has effects similar to optimizing average precision directly. 3.2.2. An Equation for d Our first step is to derive an equation for d when combining two systems. We use the following notation: the letters a and b are used to indicate the two systems being combined. The set of relevant documents returned only by system a are indicated P a (P for positive example), those by are P b and those returned by both are P ab , with their sizes shown as P , so that P a = jP a j, etc. The corresponding sets of nonrelevent documents and their sizes are indicated using N and N . We proceed by deriving an expression for - p in terms of - p a and - b , and then likewise for - n. As in our empirical study, we are trying to predict the performance of the combined system based on the performance (and other variables) of the component systems. ae min be the lowest score produced by the combined system before normalization, and r be the range of scores (ae These quantities are used to 12 Vogt and Cottrell -0.4 d Average Precision Fig. 4. Average Precision versus d for the 3050 TREC5 Entries normalize the scores of the combined system to [0; 1]. Then, by definition, ab sin(w) (ae a (x)\Gammaae min ) r x2Pa sin(w) (ae a (x)\Gammaae min ) r r Or, sin(w) sin(w) \Theta P ab ae a x2Pa ae a (x) cos(w) \Theta P Now, note that for system a ab ae a x2Pa ae a (x) Or, p a ab ae a x2Pa ae a (x) And likewise for system b. Substituting this into the equation for - p yields: By a similar series of steps, it can be shown that: sin(w)(-n a \Gamma ae min )(N ab +N a ) r(N ab +N a +N b ) r(N ab +N a +N b ) Fusion Via a Linear Combination of Scores 13 After introducing the following shorthand for the relative ratios of relevant and nonrelevant documents returned by both systems: N a +N ab ab +N a +N b ab +N a +N b We have, d =r 3.2.3. Maximizing d Because we have a formula for d (the performance of the combined system), we can find its maximum via calculus. Differentiating with respect to w and setting it equal to 0 yields: Or, For simplicity, define ffi a as the numerator of this expression and ffi b as the denominator so that tan(w opt . Then the equation for the optimal weight could equivalently be written: sin(w opt a Recall from trigonometry that: Using this fact, and the relationship between sin(w) and cos(w) noted above, we get the following expression for the performance of the optimally combined system: d opt =r4 At this point we make the simplifying assumption that both systems have positive weight, a realistic one if both systems perform reasonably well. Then the above equation simplifies to: 3.2.4. Support for the 3 Hypotheses The final equation from the previous section indicates that to maximize d opt , we need to maximize both ffi a and ffi b . Remember that, ab )(-n a \Gamma ae min ) ab +N a +N b Or, using shorthand, 14 Vogt and Cottrell It is easy to show that since 0 - ff ffi a is optimal when ff Similarly, when - a is maximal (i.e., when d a is maximized). Similar arguments hold for We are now in a position to address the three hy- potheses: Linear combination is warranted when, 1. at least one system exhibits good performance, 2. both systems return similar sets of relevant documents 3. both systems return dissimilar sets of nonrelevant documents. The first point follows directly from the fact that ffi a and are maximal when d a and d b are maximal. The second two points can be concretely supported as follows. Recall that ffi a is optimal when ff Pa+P b +P ab , it is equal to 1 when P Likewise, are optimal when both systems return no unique relevant documents (i.e., they maximize the overlap of relevant documents). Conversely, since ff Na+N b +N ab , it goes to 0 as N b ! 1. The same argument holds for are optimal when both systems are retrieving different sets of nonrelevant documents. 3.2.5. Summary We have provided mathematical support for our empirically derived conclusions about when to use the linear model. Namely, when at least one system exhibits good performance, both systems return similar sets of relevant documents, and both systems return dissimilar sets of nonrelevant documents. These conclusions basically state that the LC model primarily exploits only the Chorus Effect. We have also derived a formula for the optimal weighting as well as a formula for the performance of the combined system assuming the optimal weighting. 4. Training the LC Model We now turn to the practical side of things. In the previous section we derived an optimal weighting for two systems according to d. However, using this weighting is not a practical technique for several reasons. First, it is not general, in that it does not cover the case of more than two systems. Second, it would be more satisfying if we could choose a weight which optimizes a more traditional IR performance measure, such as average precision or exact precision. Finally, the optimal weights used in the previous section were not tested for generalization to new data. In this section, we describe a series of experiments in which we compare two different techniques for choosing the weights of an LC model. Fusion Via a Linear Combination of Scores 15 4.1. Advantages of Using d as an Optimization Cri- terion Recent work [Vogt and Cottrell, 1998b] has argued in favor of using d 0 as an optimization measure for ranking problems such as those seen in IR. The list of reasons to prefer d 0 also apply to d. Namely, ffl it is differentiable with respect to model parame- ters, thus it can be applied to more complex models (via gradient based techniques) than optimizing precision directly, which has no gradient, ffl it may be suitable for online learning, ffl it is roughly correlated with other well-known measures of performance like average precision, ffl it is generally cheaper to calculate than rank-order statistics, such as J [Bartell et al., 1994], ffl it has an intuitive meaning: as a measure of how well the scores on positive examples are separated from those on negative examples. Below we show that optimizing d to select a weighting scheme for combining two IR systems works nearly as well (as measured by average precision) as optimizing average precision itself. It is important that we distinguish between an optimization criterion and a performance measure. Often, the same function plays both roles. However, in the following experiments we may choose the weights by optimizing either d or average precision (P), but we will always evaluate the resulting combined system using P. 4.2. Training a Model for the Routing Problem As previously noted, the routing problem should be easier for the LC model (or any parameterized model) because a separate model can be trained for each query, assuming enough training data. For this reason, we begin by examining the routing problem. 4.2.1. Method We examined all 61 submissions to the TREC5 adhoc track. For each of the possible pairs, on each of the 50 TREC5 adhoc queries, and 70% of all possible documents, we chose w in one of two ways: ffl using a golden section search [Press et al., 1995] to optimize P directly, or ffl using a golden section search to optimize d directly Golden section search does not require a gradient, hence it is useful in both of these cases. Although we are able to calculate the optimal weight for d using equation (3), we use the golden section technique because in the future we would like to be able to apply the technique of optimizing d to models with more than one parameter, for which an equation will not be derivable. We did not use a hold out set to stop training since the LC model has only one parameter and thus is unlikely to overfit the training data. We tested each of the trained combinations on the remaining 30% of the documents, evaluating each combination using P. 4.2.2. Results and Discussion For each of the pair/query triples, we calculated the difference between the value of average precision on the test set as determined by the weight being trained on d and the weight being trained on P. That is, we are looking at the the quantity PP \Gamma P d . This measures how much better training on P is than training on d. Figure 5 displays the histogram of this difference. The average difference is -0.009. This means that on average, training using d gives better generalization when P is the measure of performance. Although an ANOVA shows that the PP and P d distributions are different (p ! 0:0001), to interpret whether or not this is really a meaningful difference, consider a query that has 75 relevant documents (the median for a TREC5 query). Then a difference of 0.009 in Pmeans that only one or fewer more relevant documents are returned. It is also interesting to note that there are a significant number of combinations where training on d is better than on P (differences less than zero). Table 3 gives some idea as to the amount of improvement gained by using the LC model. As noted previously, the LC model is not always capable of improving over the two component systems. Of the 91500 pair/query triples, 80324 (88%) saw improvement on the training set over both of the component systems when trained using P. Of those 80324, 32493 (36% of the total) also saw improvementon the test set. The table also shows that over the 80324 combinations which saw improvement on the training set, the average degradation on the test set over the better of the two component systems was 14%, a significant drop. The fact that the LC model tends to degrade performance for most pairs of systems should not be surprising - recall the results of the previous section, which indicate that the LC model should only be used in certain situations. Apparently, training using d allows better discrimination of when these situations arise. While training on d results in fewer systems which see improvement on the training set, it allows for much better generalization when those weights are used on the test data - in fact, it allows for an improvement in most cases! 4.3. Training a Model for the Adhoc Problem All of our analysis thus far has focused on the routing problem. One can certainly argue that the adhoc problem, where new queries are constantly being submitted to the IR system, is also an important problem, especially for the World Wide Web. The following experiments explore this aspect of IR. 4.3.1. Method For the routing case we were able to examine the full set of all the TREC5 queries. How- ever, for the adhoc problem, we reduce the space of Fusion Via a Linear Combination of Scores 1750001500025000-1 -0.8 -0.6 -0.4 -0.2 0 Fig. 5. Histogram of PP \Gamma P d on the Routing Test Set for 91500 TREC5 Combinations. Table 3. Training Results for Routing Training Test Combos Avg Degradation or Method Improved Improved Improvement on Test d 37661 (41%) 21028 (23%) +15% Number of pair/query triples (out of 91500) which achieved better performance (as measured by P) than both component systems on the routing problem. Average improvement on the test set is over all combinations counted in the first column. combinations by only examining three subsets of the total 61 systems. The first subset, labeled "chorus", consists of 10 systems (see Table 4). These systems were chosen because many of the 45 possible pairings have a high overlap of relevant documents and low overlap of nonrelevant documents, and thus are theoretically able to exploit the linear combination model. The second subset, labeled "diverse," was chosen to maximize the differences between the systems in the Table 4. Subsets Used in Adhoc Experiments Subset Systems chorus KUSG2 KUSG3 anu5man4 anu5man6 gmu96ma1 diverse CLTHES DCU961 anu5aut1 anu5man6 brkly17 random CLCLUS Cor5M2rf DCU961 ETHas1 ETHme1 subset. These differences were simply the average of all the pairwise measures used our empirical analysis above after they had all been normalized to z-scores. Recall that these pairwise measures included the correlation of scores, similarity of rank order as measured by Guttman's point alienation statistic, and the size of the intersection of the documents returned by both systems. The third subset of ten systems was chosen randomly from the 61 entries. For each subset and each of the possible 45 pairs of systems from that subset, w was chosen using golden section search to maximize either - d (P or d when averaged over 35 randomly chosen queries). We then tested each of the trained combinations on the 13 of the remaining queries, evaluating each combination using P. 4.3.2. Results and Discussion Once again, we use d on the test set as our comparison metric. Figure 6 displays histograms of this difference for the chorus and random subsets (the diverse histogram is similar to the one for chorus). The average difference for the chorus subset is 0.012, diverse is 0.011, and random is 0.006. This means that on average, training using - d gives worse generalization when - P is the measure of performance, just the opposite effect as observed for routing. Once again, although these differences are statistically significant, it is doubtful that they are practically different. Table 5 confirms that training the adhoc problem using - P is better than using - d, since more pairs of systems are able to successfully generalize and overall degradation is less, although the difference is slight. The table also confirms the hypothesis that adhoc is harder to train than routing, since on average, there is a degradation in performance. However, for a significant number of combinations (16% or 23%) there was an improvement. In future work, we would like to apply techniques similar to those we used in our empirical work for routing in order to predict when improvement is achievable for the adhoc task. 5. Conclusions Our analysis of the linear combination of scores fusion model has revealed a number of important points. The LC model has severe limitations, both in scope and power. Its effective use is limited to situations where the systems involved have high overlap of rel- Fusion Via a Linear Combination of Scores 1951525 Fig. 6. Histogram of - d on the Adhoc Test Set for the chorus (left) and random (right) subsets Table 5. Training Results for Adhoc Training Test Combos Avg Degradation Subset Method Improved Improved on Test chorus - diverse - random - Total/average chorus - d 19 (42%) 5 (11%) -15% diverse - d random - d Total/average 61 (45%) 21 (16%) -10% Number of system pairs (out of 45) which achieved better performance (as measured by - P) than both component systems on the adhoc problem. Average degradation on the test set is over all combinations counted in the first column. evant documents and low overlap of nonrelevant doc- uments. Luckily, we have also developed a technique which can very accurately (r predict when improvement is possible for the routing problem. Our analysis has also revealed several advantages and disadvantages to using d as a performance crite- 20 Vogt and Cottrell rion. Despite the facts that it fails to take into account recall, and its correlation with average precision is weak, its simple algebraic form has allowed us to mathematically support conclusions about the model which previously were only empirically based, and to derive theoretically optimal weightings for pairs of IR systems. Furthermore, when used as an optimization criterion for the routing problem, d selects weights which generalize much better than those chosen by optimizing average precision, resulting in an average improvement of 15% versus an average degradation of 14% when training average precision directly. Further- more, on the adhoc problem, training d works about as well as training average precision. Because d is also differentiable and cheap to calculate, it is also a likely candidate for use in gradient based optimization procedures, which will be necessary for combinations of more than two systems or for other parameterized models (e.g., neural nets). --R Automatic combination of multiple ranked retrieval sys- tems Combining evidence of multiple query representations for information retrieval. A neural network model for documentary base self-organising and querying Comparing neural and probabilistic relevance feedback in an interactive information retrieval system. Indexing by latent semantic analysis. Information retrieval using dynamic evidence combination. The Third Text REtrieval Conference (TREC-3) The Fifth Text REtrieval Conference (TREC-5) Decision level data fusion for routing of documents in the TREC3 context: A best case analysis of worst case results. SIGIR 93: Analyses of multiple evidence com- bination Numerical Recipes in C: The Art of Scientific Computing. Combination of multiple searches. Signal detection theory and ROC Analysis in Psychology and Diagnostics. Information Re- trieval Computation of term associations by a neural network. --TR --CTR Javed A. Aslam , Mark Montague, Bayes optimal metasearch: a probabilistic model for combining the results of multiple retrieval systems (poster session), Proceedings of the 23rd annual international ACM SIGIR conference on Research and development in information retrieval, p.379-381, July 24-28, 2000, Athens, Greece Elmrio Gomes Dutra, Jr. , Jose Valdeni de Lima, Supplement of partial ranks to the data fusion, Proceedings of the 12th Brazilian symposium on Multimedia and the web, November 19-22, 2006, Natal, Rio Grande do Norte, Brazil Shengli Wu , Fabio Crestani, Methods for ranking information retrieval systems without relevance judgments, Proceedings of the ACM symposium on Applied computing, March 09-12, 2003, Melbourne, Florida Anselm Spoerri, Using the structure of overlap between search results to rank retrieval systems without relevance judgments, Information Processing and Management: an International Journal, v.43 n.4, p.1059-1070, July, 2007 Holger Billhardt , Daniel Borrajo , Vctor Maojo, Using genetic algorithms to find suboptimal retrieval expert combinations, Proceedings of the 2002 ACM symposium on Applied computing, March 11-14, 2002, Madrid, Spain Eui-Hong Han , George Karypis , Doug Mewhort , Keith Hatchard, Intelligent metasearch engine for knowledge management, Proceedings of the twelfth international conference on Information and knowledge management, November 03-08, 2003, New Orleans, LA, USA M. Elena Renda , Umberto Straccia, Web metasearch: rank vs. score based rank aggregation methods, Proceedings of the ACM symposium on Applied computing, March 09-12, 2003, Melbourne, Florida David Lillis , Fergus Toolan , Rem Collier , John Dunnion, Probabilistic data fusion on a large document collection, Artificial Intelligence Review, v.26 n.1-2, p.23-34, October 2006 Shengli Wu , Fabio Crestani, Shadow document methods of resutls merging, Proceedings of the 2004 ACM symposium on Applied computing, March 14-17, 2004, Nicosia, Cyprus Paul Ogilvie , Jamie Callan, Combining document representations for known-item search, Proceedings of the 26th annual international ACM SIGIR conference on Research and development in informaion retrieval, July 28-August 01, 2003, Toronto, Canada Holger Billhardt , Daniel Borrajo , Victor Maojo, Learning retrieval expert combinations with genetic algorithms, International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems, v.11 n.1, p.87-113, February Kui Lam Kwok , Sora Choi , Norbert Dinstl, Rich results from poor resources: NTCIR-4 monolingual and cross-lingual retrieval of korean texts using chinese and english, ACM Transactions on Asian Language Information Processing (TALIP), v.4 n.2, p.136-162, June 2005 Dmitri Roussinov , Weiguo Fan, Discretization based learning approach to information retrieval, Proceedings of the conference on Human Language Technology and Empirical Methods in Natural Language Processing, p.153-160, October 06-08, 2005, Vancouver, British Columbia, Canada Sally McClean, Performance prediction of data fusion for information retrieval, Information Processing and Management: an International Journal, v.42 n.4, p.899-915, July 2006 David Lillis , Fergus Toolan , Rem Collier , John Dunnion, ProbFuse: a probabilistic approach to data fusion, Proceedings of the 29th annual international ACM SIGIR conference on Research and development in information retrieval, August 06-11, 2006, Seattle, Washington, USA Kui-Lam Kwok , Laszlo Grunfeld , Peter Deng, Employing web mining and data fusion to improve weak ad hoc retrieval, Information Processing and Management: an International Journal, v.43 n.2, p.406-419, March 2007 Mark Montague , Javed A. Aslam, Relevance score normalization for metasearch, Proceedings of the tenth international conference on Information and knowledge management, October 05-10, 2001, Atlanta, Georgia, USA B. Uygar Oztekin , George Karypis , Vipin Kumar, Expert agreement and content based reranking in a meta search environment using Mearf, Proceedings of the 11th international conference on World Wide Web, May 07-11, 2002, Honolulu, Hawaii, USA D. Lillis , F. Toolan , A. Mur , L. Peng , R. Collier , J. Dunnion, Probability-based fusion of information retrieval result sets, Artificial Intelligence Review, v.25 n.1-2, p.179-191, April 2006 D. Frank Hsu , Isak Taksa, Comparing Rank and Score Combination Methods for Data Fusion in Information Retrieval, Information Retrieval, v.8 n.3, p.449-480, May 2005 Yu-Ting Liu , Tie-Yan Liu , Tao Qin , Zhi-Ming Ma , Hang Li, Supervised rank aggregation, Proceedings of the 16th international conference on World Wide Web, May 08-12, 2007, Banff, Alberta, Canada Steven M. Beitzel , Eric C. Jensen , Abdur Chowdhury , David Grossman , Ophir Frieder , Nazli Goharian, Fusion of effective retrieval strategies in the same information retrieval system, Journal of the American Society for Information Science and Technology, v.55 n.10, p.859-868, August 2004 Weiguo Fan , Ming Luo , Li Wang , Wensi Xi , Edward A. Fox, Tuning before feedback: combining ranking discovery and blind feedback for robust retrieval, Proceedings of the 27th annual international ACM SIGIR conference on Research and development in information retrieval, July 25-29, 2004, Sheffield, United Kingdom Fabio Crestani , Shengli Wu, Testing the cluster hypothesis in distributed information retrieval, Information Processing and Management: an International Journal, v.42 n.5, p.1137-1150, September 2006 Weiguo Fan , Michael D. Gordon , Praveen Pathak, A generic ranking function discovery framework by genetic programming for information retrieval, Information Processing and Management: an International Journal, v.40 n.4, p.587-602, May 2004 Javed A. Aslam , Mark Montague, Models for metasearch, Proceedings of the 24th annual international ACM SIGIR conference on Research and development in information retrieval, p.276-284, September 2001, New Orleans, Louisiana, United States Sally McClean, Result merging methods in distributed information retrieval with overlapping databases, Information Retrieval, v.10 n.3, p.297-319, June 2007 Cai-Nicolas Ziegler , Sean M. McNee , Joseph A. Konstan , Georg Lausen, Improving recommendation lists through topic diversification, Proceedings of the 14th international conference on World Wide Web, May 10-14, 2005, Chiba, Japan Sally McClean, Improving high accuracy retrieval by eliminating the uneven correlation effect in data fusion, Journal of the American Society for Information Science and Technology, v.57 n.14, p.1962-1973, December 2006 Mark Montague , Javed A. Aslam, Condorcet fusion for improved retrieval, Proceedings of the eleventh international conference on Information and knowledge management, November 04-09, 2002, McLean, Virginia, USA Weiguo Fan , Michael Gordon , Praveen Pathak, On linear mixture of expert approaches to information retrieval, Decision Support Systems, v.42 n.2, p.975-987, November 2006 Wai-Kit Lo , Helen Meng , P. C. Ching, Cross-language spoken document retrieval using HMM-based retrieval model with multi-scale fusion, ACM Transactions on Asian Language Information Processing (TALIP), v.2 n.1, p.1-26, March Jacques Savoy, Comparative study of monolingual and multilingual search models for use with asian languages, ACM Transactions on Asian Language Information Processing (TALIP), v.4 n.2, p.163-189, June 2005 Weiyi Meng , Clement Yu , King-Lup Liu, Building efficient and effective metasearch engines, ACM Computing Surveys (CSUR), v.34 n.1, p.48-89, March 2002
routing;fusion;performance evaluation;neural networks;linear combination
593983
Simple Bayesian Model for Bitmap Compression.
Bitmaps are a useful, but storage voracious, component of many information retrieval systems. Earlier efforts to compress bitmaps were based on models of bit generation, particularly Markov models. While these permitted considerable reduction in storage, the short memory of Markov models may limit their compression efficiency. In this paper we accept the state orientation of Markov models, but introduce a Bayesian approach to assess the state&semi; the analysis is based on data accumulating in a growing window. The paper describes the details of the probabilistic assumptions governing the Bayesian analysis, as well as the protocol for controlling the window that receives the data. We find slight improvement over the best performing strictly Markov models.
Introduction This paper continues a series of papers (Bookstein et al. 1992, 1994, 1997) that apply advanced statistical models to compress bitmaps. The earliest models assumed that the 1-bits of a bitmap occur independently (Bookstein et al. 1992). In several environments, however, the independence assumption is not valid. For example, consider a concordance built to allow access to a large textual database. In the concordance, each term occurring in the text is associated with a bitmap, and each bit-site to a region of the text obtained by partitioning the text into logically coherent units. The bitmap representing a term has a 1-bit at the positions where the corresponding segment contains the word. Within the text, discussions of specific subjects are clustered, and this imposes a clustering tendency on the 1-bits within the bitmaps belonging to the terms associated with these subjects. Since bit encoding is determined by the probability of its having a given value (for example, using arithmetic coding (Witten et al. 1994)), we expect more accurate modeling to result in improved compression. In this paper we focus on the problem of modeling the clustering rather than on the details of generating codewords. Our first attempts to represent this clustering involved the use of a variety of Markov Models (Feller 1957). Although significant improvement resulted from using even simple Markov models with a small number of states (Bookstein et al. 1997), ultimately we were limited by the finite memory of such models. We now test an alternative approach, based on the principles of Bayesian statistical analysis (Press 1989), that may correct for this limitation. We continue with the assumption made in our earlier papers that bits are generated within two states, a cluster state, which tends to produce a high density of 1-bits, and a between-cluster state, which produces 1-bits more sparingly. However, we no longer model the state transitions. Instead, we use Bayesian reasoning to assess the likelihood of being in either state, given the data. The data is gathered within a growing window, which is closed when we decide a state transition was likely. The probability of the next bit taking Simple Bayesian model for bitmap compression 3 a value is determined by the data currently in the window, and the bit encoded accordingly. In the next section we describe, in general, the model on which the remainder of the paper is based. We illustrate these ideas by means of a simple version of the Bayesian model, which we then elaborate. In subsequent sections we describe the protocol used to control the window in which data is collected, as well as describe the experiments used to test the models. 2 . Bayesian Modeling We conceptualize a bitmap of length D bits as having been generated in a manner similar to that of the hidden Markov model described in Rabiner (1989). At any site in the bitmap, the generator is assumed to be in one of two states: a cluster state C or a between-cluster state B. These states determine the probability of generating a 1-bit. We diverge from the Markov assumption by no longer modeling state transitions. Rather, we collect data as we scan the bitmap, and use the data to guess the probability of being in either state, and hence the probability of a 1-bit being generated. Our overall strategy is to encode bits using a dynamic window to accumulate the data. We begin with a window that contains zero bits. As we scan, and encode, bits, the window grows. As we accumulate data in a window, we use Bayesian methods to improve our estimate of the probability of a 1-bit being generated. If at any point we sense a state change, we close the window and begin the process anew. Suppose then that we are just opening a window. Taking a Bayesian approach, we try to determine the probability distribution of the parameter governing data generation within the window. The parameter here is p, the probability of a 1- bit. We do this by starting with a prior distribution for p, based on not having 4 Bookstein, Klein and Raita any information, and then improve the distribution as the window grows and more information accrues. In this problem, the uncertainty in p before developing the window is primarily the result of our not knowing which state we are in, though also, possibly, each state may allow a range of probabilities. Most generally, we assume that in the cluster state C, the distribution of probabilities is given by f C (p), and that the corresponding distribution in the between-cluster state is denoted by f B (p). We further assume that ' is the a priori probability of being in a cluster state. When we begin a new window, we have no data on which to base our esti- mate, and use the prior, unconditional, distribution of p, P 0 (p). Given the above assumptions, we have: As we accumulate data in the current window, we want to integrate the information it contains to improve our estimate of P 0 (p). We denote this density function by P (pjW ), indicating that the density function incorporates information about the window W we are scanning. To derive the updated distribution, we use Bayes' formula to get, But the probability of the occurrence of a certain window, for a fixed p, is simply are, respectively, the number of 1's and the number of zeros in W , which now contains N 0 +N 1 values. We can now estimate the probability of a 1-bit, given the evidence W in the R integral is simply, - p, the expected value of p, conditional on the evidence in W : Simple Bayesian model for bitmap compression 5 To proceed, we have to specify the state based distributions. Below we consider two models of increasing complexity. 2.1. Sharp probability distribution In the simple first model, we assume that the probability is fully determined by the state; that is, we let p C be the probability of seeing a 1 while in a cluster, and p B the probability of seeing a 1 while between clusters. Note this allows us to see 1's even if not in the cluster state, and vice versa. Since the probability is determined by the state, we have and Here ffi(x) is the Dirac ffi-function, defined as 0 everywhere but whose integral satisfies, Z In other words, ffi(p \Gamma p C ) is a probability distribution that asserts that equation 1, the prior, unconditional, distribution of p, is given by, which is a mixture of two Dirac ffi-functions. As we develop the current window, we want to integrate the data it contains, as prescribed by equation 2. We now apply equation 4 to evaluate - p, the expected 6 Bookstein, Klein and Raita value of the posterior distribution P (p j W R 1[' R 1[' Special values As a check on the reasonableness of the model, it is useful to examine several limiting values. For example, if in such cases, the state is known, and the formula gives the appropriate probability of producing a 1-bit. We also find that if, as the window grows, the value of N 1 increases indefinitely while the value of N 0 remains fixed, then (assuming That is, as grows relative to N 0 , the evidence becomes increasingly strong that we are in the C state, and the appropriate probability is approached. Next we find that if While initially surprising, a little thought shows this result is correct. First note that p N 1 , and hence - p, is undefined we must be in a cluster state, the only state that permits a 1-bit. Similarly, if p p is defined only if N 0 ? 0; but then we must be in a between-cluster state, requiring - p to be as is indeed the case. Similar comments could be made for the cases but these are unrealistic given what these probabilities represent. Finally, if That is, there is in effect a single state, and the probability of a 1 being generated is the common probability. Simple Bayesian model for bitmap compression 7 2.2. Beta-Distributed Model If we relax the requirement that the state completely determines the probability parameter, we can create much more flexible models; doing this transcends in principle the Hidden Markov Model (HMM) used in our earlier papers. In the model we now examine, we still conceptualize the generator as being in one of two states. The state now influences, but doesn't fully determine, the probability of generating a 1-bit. We gain this additional flexibility by admitting as probability distributions, density functions f C (p) and f B (p) with some dispersion. The beta-distribution (Johnson et al. 1970) offers a class of models that allows a great deal of flexibility in controlling the shape of the distribution, while at the same time being relatively convenient analytically. We include this model as an example of how the Bayes approach allows us to reach well beyond the Markov framework: we now recognize the possibility that not all clusters are identical, and that in some clusters, the probability of a one-bit is greater than in others. A similar comment could be made for the non-cluster state, although a hybrid model, using a beta-model for the cluster state and a delta-function for the between-cluster state, is possible. In the beta-distributed model we assume, for parameters ff C ? 0 and fi C ? 0 where, a parallel distribution defines f B (p; ff (Note that here the factorial operator, for non-integer values of its argument, is defined in the usual way in terms of the Substituting the beta-functions for f C and f B in equation 4, we find after integration 8 Bookstein, Klein and Raita The result is analytically complex, and must be evaluated numerically. We must adjust the parameters, ff C , ff B , fi C , fi B and ' to optimize compression efficiency. However, it is easier to interpret the values of these parameters if we note that the expected value of the Beta-distribution is given by (suppressing the subscripts on the ff, fi parameters) E(p) j - fi), and the variance by V reminiscent of the binomial distribution, with ff successes, fi failures, but M j ff tries. It is in terms of - and M that the model is most understandingly parameterized, the former indicating the probability of a 1-bit, the latter, our confidence in the parameter. That is, for large M , our model approaches the sharp model described above. Thus, the parameter - a role parallel to that of p C for the sharp distribution, and similarly for - Experiments will indicate whether the benefit of flexible probabilities within and between clusters justify the extra effort and the need for an extra two parameters. 3 . Window Dynamics In the preceding section, we have assumed we are in a region where the state is fixed, but that we do not know what the state is. The parameters ' and the beta- distribution parameters allow us to make an information free guess of the likelihood of being in either state, and thereby the probability of the next bit taking the value 0 or 1. We then modify the probability estimate in accordance with the accretion of data as the window grows. In the preceding section, we indicated how to use the data within a window to estimate the probability of a 1-bit being generated. But as the bitmap is generated, we are shifting between states. To take this into account, Simple Bayesian model for bitmap compression 9 we must use a protocol that defines how the window grows and when it should be closed. We begin with a window of size 0, and with the initial probability estimate of eqn 1. Then, as bits are scanned, we upgrade the probability. In principle, we should let the window grow indefinitely. But doing so would put us at a severe disadvantage when the state changes. Thus we introduce a parameter, W max , and when the window-size reaches W max we stop growing the window. That is, at this point we begin shifting the window, dropping bits at the end while introducing bits at the front. In practice, a finite W max limits the size of the codeword when a bit inconsistent with the current state is scanned; we choose a value that optimizes performance during compression. But if at some site an unexpected bit value (or two) is scanned, it suggests that the state may have changed and that the estimate of probabilities should begin again - that is, that the old window may no longer be informative. At this point the old window is closed, and a new one is grown. It may be possible to estimate W max and the start-over point theoretically. The ideal way to assess the predictive power of the current window would be to actually look ahead and evaluate how well we are able to compress the next several bits (that is, the bits following those already in the window), given the information within the window. (The performance expected if the probability prediction mechanism is valid could be computed using information theoretic arguments as discussed in Cover et al. (1991) and Hamming (1980).) If the performance is less than expected, then it indicates we should close the window and begin again. However, the decoder would not be able to follow the same decision procedure, because it does not know the next, incoming bits. This situation is inherent to many adaptive processes (see, e.g. Vitter (1987) and Welch (1984)), and requires that we use a less efficient, delayed update strategy. The method we adopted is to simulate this strategy retrospectively: after each bit has been encoded, it is appended to the window W and we run a consistency check Next bit to be encoded Tail T History bits (do not have an effect on the representation of the unencoded bits)Full window Figure 1. Schema of window structure. of the bits recently encoded. We backtrack, in turn, one, two, the maximum number of bits we can backtrack, is a parameter of our model. Every time we backtrack, we ask whether the bits being reviewed are more consistent with the current window than they are with a new window. If the reviewed bits pass this test for each backtracking step, then we continue, encoding the following bit as an extension of the current window. Otherwise we start a new window. Though the reviewed bits for which the model retrospectively fails have already been encoded as if they were part of the preceding window (which is necessary for the decoder), they are considered as data for the new window. For example, if the first window is closed on the basis of its being inconsistent with the last three bits, those bits act as information for the new window when the next bit is encoded. We now describe our decision rule for opening a new window. For the purpose of this calculation, after backtracking, we consider the (retrospectively) "current" window, now called to contain the data up to the point the backtracked data begins; we treat the backtracked data, T , at the tail of the current window as "new" data, available to both the encoder and decoder. This is illustrated is Figure 1. We examine the data in W 1 , and consider whether, with hindsight, we should have continued extending W 1 , or begun a new window, W 2 . Suppose, looking for- ward, we see T consists of n 1-bits and m 0-bits. On the basis of W 1 we estimate a probability - while the corresponding probability for W 2 is - . For Simple Bayesian model for bitmap compression 11 simplicity, we consider these probabilities to be unchanged as we see successively the bits of T . We can now use the standard Bayesian argument to estimate the probabilities of W 1 or W 2 being the correct window, given that T follows, where one of the two windows is assumed to be correct. The probability that W 1 is correct is given by: the probability for W i . Thus the odds in favor of a new window, P (W 2 )=P (W 1 ) is given by Equation 7, which expresses the evidence infavor of a new window in terms of an odds ratio, can be used directly if the !'s can be properly assessed. This can be tricky. But notice that the !'s are fixed, and the evidence actually influenced by T is fully expressed by the other two factors. Thus we can empirically determine a threshold, and change the window provided the product of the two right-hand factors exceeds this threshold. It is instructive to rewrite equation 7 as: Note that the value n+m, which is just the number of bits observed while looking ahead, is fixed. Only the term in the brackets actually depends on the evidence, and this, in isolation, can serve as a measure of how strong is the claim to start a new window. The factor is the value of the odds favoring a new window, provided all the bits being tested are zero. Suppose this would result in one of the windows, say W being selected. Each 1-bit then modifies this by a factor of the odds ratio (-p 2 collectively this doesn't compensate for the initial weight, we use W i , else we use the other window. To estimate p 1 we used the last probability estimate made from the evidence in before the new evidence was evaluated. A few reasonable options are available for estimating p 2 . For example, we can use the a priori probability, to assess whether the evidence in the window is at this point any better than no evidence at all. An opposite bias can be obtained by estimating p 2 from the actual evidence itself, to see if this high-value estimate is sufficiently larger than that based on W 1 to justify opening a new window. In principle, the window update procedure could be applied recursively to the resulting tail T , perhaps allowing an even better (and shorter) basis for estimating the probabilities of the upcoming bits. However, since we restricted the length of T just to 6 bits at most, the recursive procedure was not applied in the experiments. We next illustrate the window update protocol by an example. Suppose we just encoded bit b 511 , and that W now consists of bits b 500 . Next we must test for the goodness of the window. We do this in stages: First stage: For our first consistency check, W 1 will consist of bits b 500 and T of the single bit, b 511 . On the basis of W 1 , we compute - of a 1-bit, given W 1 , and the size of the codeword for b 511 . Get the corresponding quantity for an empty window, and call it - Evidence for closing the window is given by - else it is given by This is equation 7, ignoring the !'s. We close the window if the ratio is larger than some threshold value fl, beginning a new window with b 511 . If the ratio is smaller than fl, we go to next stage. Second stage: First we backtrack, redefining W 1 as b 500 the probability of a 1-bit, given the reconstituted W again is the probability of a 1-bit given the empty window. We now use equation 7 to get the weight for a new window, based upon the evidence of T . If the evidence is large enough, we close the current window, and repeat the process beginning with Simple Bayesian model for bitmap compression 13 the window consisting of bits b 510 ; b 511 . Else we continue with the third stage (which we suppose, for this example, to be the last). Stage three: We now compute - p 1 on basis of the new W assuming the null window. We then compute the weight favoring a new window based on the evidence of the new T , consisting of the bit sequence b 509 as given by equation 7. If the weight supports opening a new window, we close the current window, and begin the cycle again starting with the window made up of bits b 509 ; b 510 ; b 511 . Else we accept the old window, make bits b 501 511 the current window, and continue. If additional stages are desired, we continue as above. Figure 2. Example of window development. An actual detailed example of window development, taken from the experiment described below, appears as Figure 2. The numbers on top of the figure denote bit positions of the specific bitmap we are studying and the actual bits are shown beneath. The maximal length of the window, W max , is set to 10 and the figure depicts the window dynamics during the encoding of bits At the start, the window contains the 1-bits at indices This run of 1-bits gives a high probability for the upcoming 1-bit at index 111. The bit is encoded and inserted into the window. In the update procedure, we notice that the window W 1 is always a better predictor for the corresponding tail T than an empty window, so only one 14 Bookstein, Klein and Raita change is made: we discard the oldest bit from the window to prevent an overflow. For the next two bits at indices 112 and 113, the same is repeated. When the 0-bit at index 114 has been encoded, we note that the threshold value used in the consistency check allows one spurious bit inside a run of 1-bits, so again, only the oldest bit is discarded. However, the next 0-bit gives already enough evidence about the end of the run of ones. This is noticed at stage two of the update process: the eight 1-bits in W 1 are not consistent with the pair of zeroes in the tail T , so we open a new window and initialize it with the contents of T . In this way, the update procedure captures the essential features of the bit sequence: if there are only a few 0-bits in a 1-bit cluster (or vice versa), we are satisfied with it, since the overall tendency is clear. Also, when we are scanning a truly heterogeneous bitmap region, there is no reason to change the window radically (see e.g. the situation when the encoding of the bit at index 144 commences). This implies that the change of the state is performed only when necessary, i.e. at the boundaries where the characteristics of the bit string change considerably. 4 . Experiments In order to compare the new technique with previous methods, we used the same test databases as in our earlier papers, namely: the King James Version of the Bible in English (its chapters acting as documents); and a subset of the Tr'esor de la Langue Fran-caise (TLF). The TLF is a database of 680 MB of French language texts (112 million words) of the 17 th -20 th centuries (see Bookstein et al. (1992) for more details of the collection), whose uncompressed concordance spans about 345 MB (excluding references to the 100 most frequent words, considered as stop-words). The subset used in this study consisted of the the 35070 terms belonging to the (lexicographic) range between elle and flaube. The models we are developing are intended for terms that show a considerable degree of clustering. But clustering strength is determined both by the nature of the Simple Bayesian model for bitmap compression 15 word and the choice of what we define as a document, which is represented as a single bit in our bitmaps. In our investigation of the TLF, we chose as the document the level immediately below the "book" level in the TLF hierarchy, resulting in bitmaps of length bits. The chosen unit is convenient for the construction of the concordance, but may obscure some of the underlying clustering because of its large size. Our first task is to estimate the parameters which give the best compression, that is, which minimize the function \Gamma log - We here use the fact that we can encode a bit with log - the encoding is based on - being the probability that the i-th bit of the bitmap takes the value it does; this probability is determined by the evidence in the window. Performing this minimization analytically is hard, so we reverted to search methods that may yield sub-optimal parameter values. The parameters divide into two sets: the model parameters (', p B , and p C for the sharp model); and three window parameters: the maximum window size W max , the maximal backtracking length b in the update process, and the threshold fl that controls when we should restart the window. For the Beta-model, we need . (Recall that the ff's and fi's of the Beta distribution can be expressed in terms of the - p's and M 's in a straightforward way.) To simplify our task of finding optimal values for our parameters, we note that, is the number of 1-bits in the map, the sharp model requires that: yielding as the value for For the Beta-model, there is no fixed probability for bit generation, but we can use the above formula, with (-p substituted for (p C as an initial estimate of '. It reasonable to expect that n 1 =D, the frequency of occurrence of 1-bits, satisfies the constraint - implies that 0 - 1. For each set of terms tested, we chose manually several combinations of parameter values obeying the restrictions above, and then searched for local optima. The large number of optimization runs performed showed, that the overall shape of the function is 'rippling' and with high probability, it does not have large minima or maxima. Due to this, we can safely fix most of the input parameters without a notable loss in compression efficiency. Some tuning may be appropriate, however, if the 1-bit densities or clumping properties change considerably from bitmap to bitmap. Table descriptive statistics for the tested files and summarizes the compression performance of the various models tested. For the Bible, only words appearing in at least 60 chapters, and thus occurring in between 60 and 929 chapters, were considered; for TLF, the terms (bitmaps) were partitioned into three classes according to the density of 1-bits. The threshold values were 78, 388 and 1162. Table I. Statistics and compression results Bible TLF terms 623 2032 619 381 occurrences 131874 352522 402890 1387698 independent 2.683 9.09 7.26 4.02 best 4-state 2.544 8.48 6.65 3.48 Bayes-Beta 2.523 8.38 6.55 3.42 Bayes-Sharp 2.556 8.44 6.65 3.52 Simple Bayesian model for bitmap compression 17 The upper part of Table 1 shows, for each class, the number of different terms, and their total number of occurrences. The lines of the lower part correspond to various bitmap compression methods. The independence model is the 1-state Markov model used in Bookstein et al. (1992); this is a very simple model that does not take clustering into account and therefore performs relatively poorly. In Bookstein et al. (1997) we studied the performance of different Markov models; the values cited here correspond to choosing, for each class, the model giving the best compression among all traditional 4-state Markov models. The line entitled HMM gives the compression figures of Bookstein et al. (1997) for the hidden Markov model. The omission of a HMM result is due to the enormous execution time needed for the run to produce this figure. The last lines correspond to the Bayesian models described herein. To understand the values in the table, recall that we are representing the top level of the concordance as follows: for each term, we list sequentially the documents in which the term occurs. This list can be conceptualized as a bitmap, D bits in length, with a 1-bit for each document in which the term occurs. As our measure of compression for the list corresponding to a term, we compute the number of bits needed to encode this list with our methods, and divide this value by the number of documents in which the term occurs. The table gives the average of this quantity for all the terms in a class. In other words, it is the average, per 1-bit within an uncompressed bitmap, of the number of bits needed to encode the 1-bits of all the bitmaps in this class. As in our earlier studies (Bookstein et al. (1997)), we have not included the cost of storing the necessary parameters, which is negligible in most of the cases. Since it seems that varying W max and the threshold value fl does not have a large effect on the results, they were fixed throughout the experiments, leaving only three or five parameters for the sharp and Beta models, respectively, as compared to four parameters for each of the HMM and 4-state models. Since only the high frequency terms are to be compressed, the average number, per term occurrence, of added bits caused by this overhead, is very small. The parameter values - define the generating model. For about 45% of the bitmaps we were able to set ' equal to zero, implying that a single beta-distribution suffices for the majority of terms. Note, however, that this does not invalidate an underlying multi-state generating process. Rather, the intrinsic variability of the beta-distribution seems adequate to represent the impact of the various states. For the other 55% of the terms, it was useful to incorporate the services of a second beta-distribution. However, for both the cluster and between cluster distri- butions, the probability of a 1-bit was low: the average value for - 0.144. The average standard deviations were 0.010 and 0.096, respectively. Thus, for these cases, the spread was comparable to the size of the values themselves. As can be seen, in all our tests, the Bayesian model outperformed the best 4-state models by 1-2%. Surprisingly, the new approach even improved upon the compression obtained by the highly CPU-intensive HMM model in one of the cases. We thus conclude that the Bayesian technique described in this paper gives a good time/space tradeoff, compressing better than the faster 4-state models, and using significantly less processing time than the HMM. 5 . Conclusion Bitmaps can be very convenient when representing concordances. But in realistic applications, they are large, and storing them is expensive. The simplest techniques for compressing bitmaps, based on models that represent bitmap generation as a sequence of independent events, in fact work quite well. But when dealing with large databases, even a small percent improvement in cost can yield large absolute benefits. Our investigations were directed to those applications for which the size of the database justifies the additional complexity. Simple Bayesian model for bitmap compression 19 As is typical in the design of optimal systems, we found that added complexity tends to improve overall performance, but with each increment in complexity yielding diminishing returns. For example, the best 4-state models in fact come quite close to the performance of the Hidden Markov Models, which are very flexible, and very expensive, models that we considered as a bound on our ability to improve performance using the route of simple probabilistic models. In this paper, we were interested in assessing whether alternatives to Markov modeling, that relieved the constraint of the Markov model's short memory, could match or improve performance. We used an underlying state model, but applied Bayesian reasoning, rather than Markov transition probabilities, to assess the state. Our results were comparable to that of the HMM, and at comparably little cost. The convergence towards the performance of the HMM, and the agreement of two rather different approaches on two very different databases, leads us to expect that striking compression improvement is unlikely without a radical increase in model complexity. However, we examined only one application - that of representing the concordance of a textual database. Other applications, in which clustering is more pronounced, may lead to quite different results, and for which the variety of models we developed could be more strongly differentiated. 6 . --R Model based concordance compression. In: Storer J. Markov models for clusters in concordance compression. A systematic approach to compressing a full text retrieval system. Elements of Information Theory. An Introduction to Probability Theory and Its Applications Coding and Information Theory. Distributions in Statistics: Continuous Univeriate Distributions-2 Exploiting Clustering in Inverted File Compres- sion Bayesian statistics A tutorial on hidden Markov models and selected applications in speech recognition. Design and analysis of dynamic Huffman codes. A technique for high performance data compression. Managing Gigabytes Address for Offprints: Timo Raita Dept. --TR --CTR Justin Zobel , Alistair Moffat, Inverted files for text search engines, ACM Computing Surveys (CSUR), v.38 n.2, p.6-es, 2006
markov modelling;bitmap compression;concordances;IR models
593994
Matching Index Expressions for Information Retrieval.
The INN system is a dynamic hypertext tool for searching and exploring the WWW. It uses a dynamically built ancillary layer to support easy interaction. This layer features the subexpressions of index expressions that are extracted from rendered documents. Currently, the INN system uses keyword based matching. The effectiveness of the INN system may be increased by using matching functions for index expressions. In the design of such functions, several constraints stemming from the INN must be taken into account. Important constraints are a limited response time and storage space, a focus on discriminating (different notions of) subexpressions for index expressions, and domain independency. With these contextual constraints in mind, several matching functions are designed and both theoretically and practically evaluated.
Introduction Searching large and dynamic information spaces is a di-cult task. Many authors have opted for an approach combining browsing and searching in a hypertext environment (see e.g. [12, 21, 5]). The motivated renements. In this overview, the hyperindex, navigation is used for query formulation. After formulation, the query (an index expression) is used to render relevant documents. The INN currently uses existing search engines for rendering, utilising only the terms (keywords) in the query. Thus, relations between terms, as modeled in the query, largely get lost. Exploiting similarity measures for index expressions in rendering, e.g. in re-ranking the documents, would alleviate this problem. The intended use of the similarity measures in the INN poses a number of constraints on their design. The documents that are rendered by the search engines are likely to contain (most of) the required keywords from the query. However, inter-term relations, as modeled in index expressions, do not have to be satised since matching is keyword-based. This may be repaired by re-ranking the rendered set of documents according to similarity measures that go beyond keywords. Thus, the similarity measures should focus on inter-term relations and, consequently, on the hierarchical composition of index expressions. The WWW contains documents on a very wide range of topics. Therefore, the INN has been designed as a domain independent exploration and search tool. This means that domain dependent approaches, such as conceptual IR ([13]) and coordinated concepts ([20]), which make heavy use of domain specic knowledge bases, are not adequately general for our purposes. The use of thorough (linguistic) analysis is not viable for the INN due to the need for quick responses and a limited amount of available memory. Since the INN runs on a web server, it potentially serves many users which may cause an explosion of the required memory. This especially holds if considerable (linguistic) knowledge is used, for instance in detailed parsing. Similarity values have to be computed real time, next to rendering documents form search engines, locating and extracting index expressions from these documents and parsing them. We thus focus on matching of basic syntactic renement structure of phrases, as captured in index expressions. In turn, this means that not all ambiguities encountered in content analysis can be resolved. The renements in the hyperindex directly stem from a particular view on the decomposition of index expressions, or, a notion of subexpressions. For instance, the order of subexpressions may be deemed relevant in computing the renements of the hyperindex. In addition, there are dierent views on embedding of index expressions. It is envisaged that the user may select his preferred notion of subexpression for constructing the hyperindex. Otherwise, the preferred notion may be derived by analysing the (types of) renements the user makes during navigational query formulation. In order to render documents consistent with the user's preferences, the similarity measure used for re-ranking should be in accordance with the preferred notion of subexpressions. The presented topic overview in the INN is based on a small subset (10-100 documents) of the complete collection. Statistics thus only apply to a very small part of the information space. As a consequence, (domain dependent) collection statistics cannot be globally computed. In addition, term weights for the query are not supplied by the INN system. In this article, therefore, we focus on the core function of our similarity, the structure of index expressions and provide abstract functions for comparing terms. The goal of this article is the design of (relatively simple and e-cient) similarity measures that focus on the structure of index expressions. In particular, the order of subexpressions and dierent notions of embedding are to be modeled in the similarity measures. The structure of this article is as follows. In section 2, our approach is put in the context of other work on linguistically motivated indexing and structural matching. Section 3 introduces two representations of index expressions. In section 4, elaborates several topics concerning the contents and structure of index expressions. These topics lead to the design of several similarity measures for index expressions in section 5. Section 6 provides an experimental evaluation of the proposed measures. Finally, section 7 provides concluding remarks and directions for further research. Related Work 2.1 Obtaining Index Expressions from Text Currently, the INN system directly extracts index expressions from the titles of the rendered doc- uments. This produces acceptable results since titles are often in a form which permits a ready transformation to index expressions. A similar point of departure is taken in a closely resembling sys- tem, the hyperindex browser ([10]). A simple parsing method based on a two-level priority scheme for connectors ([4]) is used. All allowed connectors are listed together with their priority. All other words encountered are interpreted as terms (keywords). A stoplist (list of stopwords) lters out unwanted words like, for instance, articles and determiners. During parsing, the priority of the connector at hand determines if the already parsed part of the index expression has to be broadened or deepened. Thus, no (explicit) linguistic grammar is used by the parser. The parser results in one parse tree only, thus possibly neglecting dierent interpretations. A test on the CACM document collection showed that this parsing method was accurate in approximately 90 percent of the cases. additional sources of linguistic knowledge are used in the INN. Although our approach is rather simplistic and although "syntax by itself cannot resolve the many ambiguities that complicate the content analysis task" ([16]), in our view, the advantages oered by the INN system, augmented with similarity measures for index expressions, produce a workable balance between keyword based approaches and sophisticated linguistic analysis. A more general approach to indexing documents with index expression could make use of a part- of-speech (POS) tagger and an extractor. For closely related descriptors, called phrase frames, this approach is described in ([1]). Phrase frames are head-modier structures in which connectors are (eventually) omitted. After preprocessing the documents, ltering non-textual parts and moulding the text into acceptable format, the POS tagger labels words with their part of speech. A well performing POS tagger is described in [3]. The extractor, which is applied next, uses the POS labels in identifying the linguistic constructs that should be parsed as index expressions. The extractor uses pattern matching with regular expressions. An extractor for index expressions would constitute a subset of the patterns used for phrase frames, albeit augmented with connectors. Finally, the same parser as above can be used, although more sophisticated parsers are available. This is, however, not in the scope of this article. The general architecture of an NLP-IR system advocated by [19] also includes a stemmer. Normalisation at term level (e.g. stemming) is largely orthogonal to our approach, so possible to include, but not considered in this article. However, we provide abstract similarity functions for terms and connectors that could be instantiated by measures that do take, for instance, stemming into account. Regularisation or syntactical normalisation at phrase level aims at recall enhancement by reducing syntactical variation ([19]). For phrase frames, syntactic normalisation is described in ([1]). This approach can, mutatis mutandis, also be applied for index expressions since phrase frames closely resemble index expressions. However, this requires the use of additional linguistic knowledge. We do not elaborate on this issue here. Instead of normalising descriptors, systematically generating (all) linguistically motivated alternative variants ([18]) has a similar aim. Semantic linguistic knowledge is exploited not to generate invalid variants. Although this restricts the number of generated variants, this number may explode, especially for complex composed descriptors. Matching is done via translation of the generated variants into a Boolean query. Due to the need for e-cient matching in the INN system, we do not generate all variants of index expressions. The Constituent Object Parser ([14]) also uses syntactic structure for matching descriptors. It produces binary dependency trees, which indicate (directed) dependency and scope of relationships between terms. Many dierent syntactic forms are represented by this single dependency relation. As a result, ambiguities are retained in the representation ([19]). A comparable approach is that of using tree structured analytics ([17]). Both approaches use external sources of linguistic knowledge and oer at best reasonable e-ciency. Index expressions make a distinction between the "types" of dependency, as given by connectors. Not taking the actual (contents of) connectors into account, e.g. by allowing wildcard connectors, in matching normalised index expressions eectuates a similar (directed) notion of dependency. An intermediate form of dependency, colliding connectors of the same type, is described in section 4.1.2. The CLARIT system ([7, 6]) integrates several NLP techniques with the vector space retrieval model. These techniques include morphological analysis, robust noun-phrase parsing, and automatic construction of rst order thesauri. Construction of the thesauri, used to support the selection of appropriate terms, requires a substantial (minimally 2 GB) sample of documents about a certain topic. Because of limited response times and available memory, we expect this approach too demanding for our application. 2.2 Comparing Structured Descriptors The similarity measures for index expressions described in section 5 focus on the structure of index expressions, i.e., on the way subexpressions are nested. Three similarity measures are provided: skeleton content, full-product, and a twig-based measure. These measures represent dierent views on the order of subexpressions and embedding. Headedness, or concept modication in head-modier structure, is considered as well. In [11], a language for querying structured text based on tree inclusion is described. Their approach, which exploits inclusion patterns to ensure preservation of binary properties between nodes, takes both structure and content into account. Example inclusion patterns are L for labels, A for ancestorship, and O for (left-to-right) ordered tree inclusion. Our skeleton-content approach resembles their fLAOg-embedding. That is, ancestorship and ordering are preserved and labels are taken into account. However, our approach does not require labels to be equal but supports approximate matching by oering abstract similarity functions for terms and connectors. In the introduction of the mentioned article the authors indicate that, indeed, such "standard IR techniques should be added to the language". Our similarity measures for index expressions capture both aspects by approximation: they are sensitive to both structure and, by using abstract term comparison functions, content. Our skeleton-content approach thus 'preserves' labels by taking into account their similarity. Essen- tially, it searches for the best fL 0 AOg-embedding and delivers the degree of embedding. A similar line of reasoning shows that our full product approach computes the degree of fLAg-embedding. Since in the full product measure the order of subexpressions is considered irrelevant, the corresponding inclusion pattern O is not satised. In [1], an approach to compare head-modier structures is described. Lexico-semantical relations are used in computing the similarity between their binary terms. Our set based similarity measures for twigs do not explicitly make use of a semantical network. However, this may be incorporated in our abstractly modeled comparison functions for terms and connectors. In addition, twigs are tertiary structures, also including information on the type of connection between head and modier. Furthermore, we focus on similarities between sets of twigs, rather than on their individual similarities. own advantages and disadvantages. We use the inductive and the structural representations. These representations describe the same language of index expressions but dier in denotational properties. adjectives, gerunds, and attribute values. Connectors denote relations between terms in the form of prepositions (showing place, position, time, or method) and some present participles (e.g. using, having, and being). 3.1 Inductive Representation The inductive representation, as described in denition 3.1, is used since it most basically describes the (de)composition of index expressions. The underlying idea is that index expressions can be augmented with subexpressions through connectors, as illustrated in gure 1. An advantage of such an elementary representation is that it allows several auxiliary functions to be readily designed. c I Figure 1: Basic setup of inductive representation. Denition 3.1 Let T be a nonempty set of terms and C be a set of connectors such that T Then, the language of non-empty index expressions is dened as: 1. if t 2 T , then t is a non-empty index expression, and 2. if I and J are non-empty index expressions and c 2 C is a connector, then add(I; c; J) is also a non-empty index expression, and 3. no other non-empty index expressions exist.Example 3.1 Single word queries or document representations are modeled by terms. Example terms are con- ference, biology, and Holland. Composed index expressions can be constructed through the add operator. This also exploits connectors, such as in, with, and on. For instance, the composed index expression add(conference; on; biology) represents information on a conference on biology. As a more complex example, consider add(add(conference; on; biology); in; Holland) denoting information on a conference on biology held in Holland. The semantics of index expressions depends on their structure. That is, dierences in nested subexpressions may cause dierences in se- mantics. As an example of this, compare the last index expression with the slightly dierent Holland)). The last one denotes a conference about biology as far as it is practiced in Holland. 2 Conns(add(I; c; Figure 2: Auxiliary functions on index expressions. Three auxiliary functions on index expressions are introduced in gure 2. These functions are used in dening similarity measures. The denitions are formulated in terms of the inductive denition of index expressions. The functions result in the terms of an index expression, its connectors, and its head, respectively. 3.2 Structural Representation The inductive representation provides a horizontal decomposition of index expressions: subexpressions are added on the right. Another representational formalism for index expressions, called the structural representation, serves better if all subexpressions at a certain depth have to be addressed at the same time. The structural representation provides a vertical decomposition of index expressions, allowing a direct and clear look on their structure. The structural representation is exploited in cases where the order of subexpressions is to be taken into account. In the structural representation, a composed index expression is denoted by where h is the head and the k subexpressions I i are connected with the head by connectors c i . The subexpressions are denoted by the structural representation as well. The composition operator provides a notational shorthand. I 1 I k Figure 3: Structural representation Figure 3 gives a schematic view on the structural representation. Example 3.2 The two last mentioned index expressions of example 3.1 are denoted in the structural representation as conference on (biology) in (Holland) and conference on (biology in (Holland)), respectively. Note that in the structural representation the dierences in semantics are modeled by brackets.4 Issues concerning Semantics The index expression can be seen as a description of a concept named h being further rened by relations called and concepts I As a consequence, comparison of both concepts names (contents) as well as the rening mechanism (structure) will be important issues when matching index expressions. This section presents several topics concerning contents and structure that rene the abovementioned view on the semantics of index expressions. The contents of index expressions is given by their terms and connectors. 4.1.1 Terms In order to match index expressions, their terms should be compared. Therefore, we assume a similarity function between terms, denoted sim [0::1]. The expression sim T (t; t 0 ) denotes the similarity between the concepts referred to by terms t and t 0 . The similarity between terms can be obtained in several ways. For instance, it can be computed by string comparison algorithms such as n-grams, the overlapping substrings of terms of length n. Furthermore, additional lexico-semantical knowledge can be used taking, for instance, hypernyms and synonyms into account. In IR, stemming is often performed to reduce morphological variance. We abstract from the particular techniques used in computing the similarity and concentrate on using this in computing similarity between index expressions. 4.1.2 Connectors Similar to terms, we also assume a similarity function sim connectors. This similarity expresses the strength of the relation between connectors. The similarity function for connectors can take several aspects into account. For instance, it can be based on the types of connectors as identied by Farradane (see [8] and [9]). In this approach, called relational indexing, connectors model the relationships between terms. Connectors that model the same relationship could be given a high similarity value. Furthermore, the priority of connectors within index expressions can be exploited. In addition, occurrence-frequencies of connectors could be used. Again, we abstract from the dierent approaches and focus on exploiting similarity between connectors for matching index expressions. 4.2 Structure The structure of index expressions partly determines their semantics and should, therefore, be taken into account in matching. Three issues concerning structure are considered in this article: the order of subexpressions, embedding, and headedness. 4.2.1 Order of Subexpressions An important question considering the semantics of index expressions is whether the order of subexpressions is relevant. Consider for example the following index expressions. hiking in mountains with friends hiking with friends in mountains One may argue that their meaning is equivalent. In other situations, however, there may be cases in which the order of subexpressions is relevant, i.e., causes a dierent meaning. For instance, if the sequential order of paragraphs in a text is represented in an index expression. The notion of order of subexpressions is formalised by the relation EqOrder as follows. Denition 4.1 We call two index expressions I = denoted by EqOrder(I; J), i 1. 2. there exists a projection of [1::k] to [1::l] such that for all 1 i k it holds that c and EqOrder(I i ; J (i) ).Two index expressions are equal modulo order i their heads are equal and (recursively) the subexpressions of the rst are equal modulo order to one of the subexpressions of the second. The second criterion holds, for instance, for index expressions of which the subexpressions (recursively) are a permutation of the other's. In addition, it holds for index expressions which contain several copies of the same subexpressions. This is, for instance, illustrated by the index expressions retrieval of information and retrieval of (information) of (information). Furthermore, equality modulo order is re exive. Later in this article, we describe a matching strategy that computes maximal similarity for index expressions that are equal modulo order. 4.2.2 Embedding Embedding or containment plays a prominent role in IR. For example, an often applied strategy to query-document matching is: if the query is contained or embedded in a document, then the document is deemed relevant to the query. Dierent notions of embedding exist. The subexpression relation for index expressions, for example, denes a connected variant of embedding: a subexpression is a connected part of its superexpressions. For instance, surng in Holland is a subexpression of surng in Holland in November but not a subexpression of surng in sunny Holland. The subexpression relation (see [23]) is denoted by 4. The expression I 4 J means that index expression I is a subexpression of J . In the abovementioned case, sunny modies the last term Holland of the path expression sunny Holland. In our view, surng in Holland is therefore embedded in surng in sunny Holland. To cater for these cases, we exploit a slightly more liberal version of embedment. Connectedness can, in the context of subexpressions, be described by direct ancestorship, or, parenthood. This means that for all terms in an index expression I, their parent must also be their parent in index expressions in which I is embedded. Instead of direct ancestorship we use the notion of (general) ancestorship. This means that index expression I is embedded in J i for all terms in I their ancestors in I are also ancestors in J . These considerations are re ected in our notion of embedding of index expressions which is formalised by an embedment relation as follows. Denition 4.2 Embedding of index expressions is captured in the binary relation , where I J means that I is embedded in J , which is dened as: I add(J; c; K) if I J or I K I K and J L (Stop) no other cases applyTerms are embedded in themselves, as described by case Same of denition 4.2. Case Sub states that an index expression is embedded in a composed one add(J; c; K), if it is embedded in either subexpression J or K. The third case Split shows that a composed index expression add(I; c; J) is embedded in another composed index expression add(K; d; L) if their connectors are equal, the leftmost subexpression I is embedded in the other leftmost subexpression K, and a similar argument holds for the rightmost subexpressions J and L. 4.2.3 Headedness The head of an index expression is considered to be its most important part. We will address this notion, implemented by head-modier pairs in ([19]), as headedness. The subexpressions modify the main concept stated in the head. Consequently, the lower terms occur, i.e., deeper with respect to nested expressions, the less important they are. This can be taken into account in matching by exploiting the depth of terms. Multiplying by a factor that is inversely proportional to the depth gives the desired result. We will indicate how each similarity function described below can be equipped with such a depth factor. 5 Similarity Measures for Index Expressions In this section, several similarity measures for index expressions are designed. They vary from simple measures that only consider the content of index expressions (section 5.1) to more comprehensive measures that consider both content and structure (section 5.2). Measuring the similarity between index expressions by sets of terms and connectors considers only provide similarity measures for (the contents of) index expressions. Example set measures are the Dice, Cosine, and Jaccard measures (see e.g. [15]). These measures do not use similarity functions for terms and connectors as described in sections 4.1.1 and 4.1.2, respectively. Instead, they use set primitives as union, disjunction, and cardinality which use equality of elements (terms and connectors). The Dice measure, which normalizes the intersection A \ B with the sum of constituents, the Cosine measure, relating the overlap of both sets to their geometric average, and the Jaccard measure, expressing the degree of overlap between two sets as the proportion of the overlap from the whole, are dened as in case the denominator does not equal zero. In case it does, the measures return similarity value zero. By considering the (sets of) terms and connectors of index expressions, the set based measures can be exploited for index expressions. As an example, consider the similarity measure for index expressions based on the Dice coe-cient as shown in gure 4. Dice (I; Figure 4: Dice's similarity measure for index expressions. The factor determines the relative in uence of terms and connectors on the similarity value. For only terms are considered and for only connectors are taken into account. Note that a depth factor cannot be taken into account directly since no information about structure is available in sets of terms. We say a measure is maximal for certain index expressions if maximal similarity is returned for the descriptors. The Dice measure is not maximal for embedded index expressions nor for subexpressions. Consider, for example, index expression surng in Holland and its subexpression Holland. Since the sets of terms and connectors are dierent, the Dice measure is not maximal for these index expressions. 5.2 Contents and Structure This section provides three similarity measures that take both content and structure of index expressions into account. The rst measure, coined full product, adheres to the idea that the order of subexpressions is irrelevant for the meaning of index expressions. On the contrary, the second mea- sure, called embedded content, deems the order relevant. In addition, embedded content is based on non contiguous embedding. Finally, the twigs measure is based on decomposing index expressions into elementary connections called twigs. 5.2.1 Full Product The full product similarity measure computes the degree to which an index expression is equal modulo order with another one. This means that the order of subexpressions is considered irrelevant. Since the structural representation appears most appropriate in this case, the full product measure, denoted by FP, is specied by it as depicted in gure 5. FP(h FP(h jTerms(h Figure 5: Full product algorithm The full product algorithm, as shown in gure 5, consists of four cases. They can be readily translated to a functional algorithm using pattern matching. The rst case Terms computes the similarity between single terms. In case Top, the embedment of a single term t in a composed index expression computed by comparing t with head h since both occur at the same depth. The third case Tall computes the similarity between two composed index expressions. Correspondingly to denition 4.1, this is the product of the similarity between the heads and the maximum similarity values of the subexpressions of I with some subexpression of J . The nal case Toll gives a penalty for the fact that a composed index expression cannot be equal modulo order with a term. The returned similarity value is smaller if the mismatch in size (number of terms) is larger. Note that the full product measure computes a similarity value layer by layer. That is, only terms and connectors that occur at the same depth are compared. The full product similarity measure is maximal for index expressions that are equal modulo order, as shown in the following theorem. Theorem 5.1 EqOrder(I; J) ) FP(I; Proof: The proof is done by structural induction to index expressions. We only investigate the (inter- esting) case of composed index expressions. Suppose I = are equal modulo the order of their subexpressions. This means that for each value 1. 2. for each 1 i k there exists a 1 j l such that d that sim C by the induction hypothesis, FP(I This means that, for the j of case (2), sim T (h; 1. The total similarity measure then comes down to 1 As a direct consequence of theorem 5.1, we conclude that the full product similarity measure is maximal for equal arguments, i.e. that FP(I; I) = 1. In addition, full product computes maximal similarity for index expressions of which the subexpressions (recursively) are a permutation of the other's or occur multiple times. The full product measure is not maximal for embedded index expressions. Consider, for example, index expressions surng in Holland and surng in sunny - Holland. Since the full product measure computes similarity values layer by layer and the keyword Holland resides at dierent depths, the given index expressions are not maximally similar. In addition, the full product measure is not maximal for subexpressions. Illustrating this, consider Holland and surng in Holland. Since the heads of both index expressions are unequal, the full product measure is not maximal. FP d sim d;T (1) The depth factor d can be incorporated in the full product algorithm (gure 5) by replacing case Tall by equation (1). Note that it is only shown how the depth factor is to be correctly computed in this case. Multiplying each case by a factor inversely proportional to this depth factor takes headedness into account. 5.2.2 Embedded Content The embedded content measure (see gure computes the best way in which an index expression can be embedded (as dened in denition 4.2) in another one. This means that the order of subexpressions is considered relevant. This eases the use of the inductive representation of index expressions and thus enables an elementary decomposition of index expressions. Figure Embedded content algorithm. The cases considered by the embedded content measure are the same as for the full product measure. The case Terms of gure 6 is exactly the same as for full product and calls the similarity function for terms sim T . The second case Top states that the degree to which a term t is embedded in a composed index expression add(I; c; J) is equal to the maximal similarity to one of the subexpressions I and J . Case Tall computes the strength of embedding of a composed index expression add(I; c; J) in another one add(K; d; L). It computes the maximum similarity of the following three cases: (1) add(I; c; J) is completely embedded in the leftmost subexpression K, (2) it is completely embedded in the rightmost subexpression L, and (3) subexpression I is embedded in K, subexpression J is embedded in L, and connectors c and d are similar. As before, case Toll gives a penalty for mismatch since composed index expressions can never be embedded in terms. The embedded content measure is maximal for embedded index expressions, as stated more formally in the next theorem. Theorem 5.2 I J Proof: The theorem is proven by induction on the structure of index expressions. Suppose I J . Basis. Suppose I is a term t. Only the rst two cases of denition 4.2 of embedded content, Same and Sub, are to be examined since the third case Split cannot apply (I cannot be composed). Same: Suppose this case of denition 4.2 applies. Then, consequently Sub: Suppose this case applies which means that J is composed, say add(K; c; L) and thus that t K or t L. Combining this with the induction hypothesis implies that 1. This means that, by case Top of the embedded content algorithm, EC(I; Induction step. Suppose I is a composed index expression, say add(K; c; L). Furthermore, without loss of generality, let only cases Sub and Split of denition 4.2 need to be examined. Both result in EC(I; Sub: Here, we have (1) I M implying EC(I; or (2) I N which means that case Tall of the algorithm is applied, which computes the maximum of these cases, both options cause EC(I; Split: In this case, we have that . By the induction hypothesis, this implies that Together with sim C causes case Tall of the embedded content algorithm to compute EC(I; We have shown that in all possible cases in which I is embedded in J the embedded content measure computes EC(I; The embedded content measure is also maximal for subexpressions. This stems from the observation that every index expression that is a subexpression of another is also embedded in it. This, in turn, follows from the observation that in the subexpression relation, direct ancestorship is preserved, which is stricter than the general ancestorship which is preserved by the embedment relation. In addition, the embedded content measure is maximal for equal arguments, since every index expression is embedded in itself, i.e. embedment is re exive. The embedded content measure is not maximal for index expressions that are equal modulo order. Consider, for example, the following pair of index expressions that are equal modulo the order of their subexpressions: conference on (biology) in (Holland) and conference in (Holland) on (biology). Since in the embedded content measure the order of subexpressions is relevant, index expressions I and J are not maximally similar. EC d (add(I; c; J); add(K; d; d) EC d+1 The depth factor d can be incorporated in the inductive representation by the scheme of equation (2). Note that the rightmost subexpressions J and L reside one layer deeper than the head. 5.2.3 Twigs The twigs (see [2]) of an index expression are its subexpressions that consist of exactly two terms and one connector. Twigs are the elementary connections in the concept graph which is formed by an index expression. Twigs have essentially the same structure as the head-modier pairs described in [19], except that they also contain a connector. Twigs enable us to form a global picture about similarity while focusing in on the elementary renements. To denote subexpressions of size two, we use the subexpression relation as described in section 4.2.2. Twigs can be dened constructively in terms of the inductive representation of index expressions. This shows that the twigs of an index expression can be produced by a straightforward syntactic process. Below, twigs are accompanied by their depth factor, modeled by a positive integer. Since the head of an index expression is assigned depth one, the expression twigs(I; 1), according to the denition below, computes the twigs of I with their correct depths. Single terms do not contain subexpressions of size two and thus have no twigs. Composed index expressions add(I; c; J) lead to at least one twig: Head(I) c Head(J ). Additional twigs may result from the subexpressions, as modeled by the recursive calls. As observed in [2], twigs conserve most of the structure of index expressions. In fact, if all terms are dierent the complete structure can be correctly reconstructed without any additional information. For twigs with depth factor, this can be relieved further to all terms at the same depth should be dierent. Twigs of index expressions seem to resemble tri-grams for strings. Considerations similar to tri-grams are therefore expected to hold for twigs. For instance, an advantage of the use of twigs over equality-matching is their robustness for 'spelling variations'. For example, if to index expressions are equal modulo order, then their twigs are the same, as stated in the following lemma. Lemma 5.1 EqOrder(I; J) Set-based approaches, described for terms in section 5.1, can also be equipped with twigs. Instead of calling the set based approaches with sets of terms or connectors as arguments, the (sets of) twigs are used. This yields similarity measures using twigs. As an example, the Dice measure for twigs is dened as The Dice measure for twigs is maximal for index expressions that are equal modulo order and for identical index expressions, since their twigs are equal. The Dice measure for twigs is not maximal for embedded index expressions nor for subexpressions. This is illustrated by, for instance, index expressions surng in Holland and Holland. 6 Experimental Evaluation Subexpressions are the basic notion on which the hyperindex is generated. The INN system thus heavily uses subexpressions. We therefore evaluate how the dierent similarity measures behave in the context of subexpressions. A number of experiments, described in this section, give insight in this matter. retrieval workshop Amsterdam holidays during on in of for by agents information thesis on disk in Java classification for November Similarity Subexpressions Dice-twigs Dice (terms) Figure 7: (a) Index expression used for experiments. (b) Overview of similarity measures. The experiments exploit a non-trivial index expression (see gure 7(a)) consisting of 11 terms nested in subexpressions of dierent forms. The similarity between this index expression and all its subexpressions was computed, according to the dierent similarity measures. Both contiguous and non-contiguous (e.g. workshop during November) subexpressions were generated from the original. In total, the experiment featured 439 such subexpressions. Similar tests may be performed with other notions of subexpressions. For instance, the order of subexpressions may be varied. Figure 7(b) gives an overview of the dierent similarity measures. The graphs provided in this gure serve as reference for the next sections. Not all set-based measures are provided; the Dice measure (for terms and twigs) is given as representative. For several reasons, the subexpressions on the x-axis are sorted according to their similarities such that non-increasing graphs are obtained. Sorting similarities non-increasingly and therefore grouping together subexpressions that are equally similar to the original index expression in intervals, clearly shows the results of the similarity measures. For instance, without the sorting, dierent measures in the same graph would be hard to distinguish. Furthermore, since subexpressions within the same interval have the same similarity value, intervals rather than the positions on the x-axis should serve as indication of the ordering. For instance, the number of intervals gives an indication of the granularity of the measure. We therefore compare similarity measures by examining dierences between their intervals. The order of subexpressions on the x-axis may be dierent for dierent similarity measures since we sort non-increasingly per measure. The dierent orders per measure are not explicitly visualised in the graphs. Therefore, our examination used detailed information, as kept in the test results, next to the graphs. The dierences between the measures are examined in the rest of this section. In our experiments, we examine dierences between similarity measures by imposing the order of one similarity measure (the so called dominant measure) onto other similarity measures (the subordinate measures). That is, the order of subexpressions on the x-axis is set to conform to the non increasing order of the dominant measure. According to this ordering, the subordinate measures are plotted. In this way, the dierences in ordering are visualised. plateau 10.60.2 Subordinate measure1.0 100 200 300 400 Subexpressions Similarity Dominant measure Figure 8: Generic sketch of results. Figure 8 gives a generic sketch of the graphs that resulted from the experiments. In this gure, the solid line depicts the similarity values of the dominant measure. In later gures, where actual similarity measures are used, the line of the dominant measure equals its similarity graph of gure 7(b): the subexpressions on the x-axis are sorted such that a non-increasing line for the dominant measure is obtained. According to this dominant ordering, the subordinate measures are plotted. Note that the example dominant measure has three horizontal plateaus, i.e. intervals of subexpressions that all receive the same similarity value, or, which are not discriminated by the dominant measure. For subexpressions out of the same plateau, the dierences in positions on the x-axis are thus irrelevant to the dominant measure. However, a subordinate measure may assign dierent similarity values to subexpressions out of the same dominant plateau. Therefore, we re-sort subexpressions per dominant plateau according to subordinate measures. In this way, we obtain non-increasing graphs for the subordinate measures per dominant plateau. Note that the number of subordinate plateaus may dier per dominant plateau. However, this number at most equals the number of plateaus in the original graph of the subordinate measure of gure 7(b). As a result of sorting non-increasingly, all peaks in subordinate graphs occur at the beginning of dominant plateaus. The form of actual graphs later conforms to the generic sketch of gure 8 but has much ner granularity because of more and smaller plateaus. We rst compare the measures that use both content and structure: the full product, embedded content, and set based measure for twigs. In sections 6.1 and 6.2, respectively, the full product and embedded content are used as primary measure. Set-based measures for twigs are compared in section 6.3. 6.1 Full Product The results of imposing the full product ordering onto the embedded content and Dice's measure for twigs are shown in Figure 9. The three measures dene dierent orderings, as can be concluded from the many peaks (embedded content) and drops (Dice-twigs) with respect to their original lines in gure 7(b). The ordering imposed by full product is related to the removal of leaves starting at maximal depth, or, a layer by layer (bottom up) defoliation. Thus, the full product measure scores high if the heads are (recursively) equal. In other words, a subexpression that equals a topmost (layer by layer) part of the original gets a high score. Full product thus demands the main concepts to be equal and is less sensitive for modication of the most specic (maximal depth) modiers. The gradually decreasing graph of the full product measure combined with its high granularity implies that this measure delicately distinguishes most subexpressions. The rightmost interval of full product is a large zero-plateau. In this large interval, the subordinate measures resemble their original graphs of gure 7(b). Inspection of the test results showed that the subexpressions in this plateau, compared to the original, either have a dierent head or contain a single subexpression that has a dierent head. For comparing index expressions that represent dierent main concepts, the full product measure is thus inadequate.0.20.61 Similarity Subexpressions Dice-twigs Figure 9: Full product order imposed. In general, each full product interval shows a peak for embedded content and a drop for the Dice measure for twigs. By examining which subexpressions cause these drops and peaks, the dierences between the measures are studied. Basically, two types of subexpressions cause the peaks for the embedded content measure. Of both types, a stereotypical example is investigated. Consider, for instance, the peak in gure 9 that occurs at position 43, having similarity value 0:5. The subexpression causing this peak consists of 10 terms, missing only the depth two leaf Amsterdam. However, full product favours subexpressions which are obtained by removing leaves at larger depth. The many possibilities to construct subexpressions in this way makes that full product has many higher scoring subexpressions. As concluded from analysing the test results, embedded content would have placed it in the interval at positions 2-5 of gure 7(b). The removal of the highest leaf (i.e. Amsterdam) is, to embedded content, similar to removing a leaf at largest depth, such as disk. Similar subexpressions cause the peaks at positions 17, 80, 101, and 171. Thus, when the in uence of concept modication is desired to gradually decrease with depth, full product is a better measure than embedded content. Another type of subexpression causes the peak at position 277, which is visually somewhat obscured by a simultaneous drop in the graph of the Dice-twig measure. This peak of similarity value 0:25 is surrounded by values of 0:09. It is caused by an index expression that does not have workshop as head, but has a large embedded subexpression consisting of 7 terms. Since the main concepts are dierent, full product scores it low. However, embedded content scores it higher and resorts it to the beginning of the value zero interval. Embedded content, by its own order, would have placed it in the interval at positions 2-5. Thus, full product, more strongly than embedded content, targets at equal main concepts. The drop in Dice's graph at position 68 to similarity value 0:375 is caused by a non-contiguous subexpression, ve of whose six twigs do occur in the original. Full product, by its layer by layer computation, does not favour non-contiguous (layer skipping) subexpressions. A similar argument applies to the drops at positions 133, 214 and 263. Twig based measures thus are more robust for slightly diering non-contiguous subexpressions than full product. Dice's drop to the x-axis at position 271 is caused by the single term workshop which consists of the head of the original index expression only. As stated before, single terms have no twigs and Dice's twig measure thus returns value zero. 6.2 Embedded Content The second experiment imposes the embedded content order onto the full product and Dice measure for twigs. The results are given in gure 10. Consider the line of the embedded content measure. A striking feature is the large rightmost plateau, stemming from the following. We computed the embedment of the original index expression in its subexpressions, not the other way around since this would result in maximal embedment for all subexpressions. Thus, the embedding of the original is computed in (the components of) the subexpressions. For many subexpressions, this eventually results in the embedding of the original in the single terms of the subexpression. This, in turn, results in a constant value which is inversely proportional to the size of the original. As a consequence, embedded content is in this context best used for discriminating between highly similar subexpressions.0.20.61 Similarity Subexpressions Dice-twigs Figure 10: Embedded content order imposed. In general, the order imposed by embedded content is relative to the removal of terms, irrespective of their depth. For instance, removing Amsterdam or disk results in equally similar subexpressions. Thus, embedded content makes no distinction to the specicity (depth) of the removed term. As a consequence, it is not so sensitive for the removal or addition of a complete modication (subexpres- sion). This makes the embedded content measure suitable for matching queries with rich document characterisations. After position 88, the graphs of full product and Dice-twigs seem to equal their original graphs of gure 7(b). This is caused by the large dominant plateau, starting at that position, which, by resorting the subordinate measures, ensures that all the subexpressions are in the same order as in their original graph. Dierences with their original graphs are caused by subexpressions that are placed in the last interval by the dominant measure but not by the original order of the subordinate measure, or vice versa. A number of stepwise drops occur in the line of full product in gure 10. For instance, full product drops to zero at positions 22, 23, and 84-87. This is caused by subexpressions with dierent heads. As was also observed for the rst experiment, embedded content is more robust to changes in the main concept. The downward steps towards value zero are caused by subexpressions that miss several terms, but preserve other parts of subexpressions. Dice's twig measure shows a clear drop at position 65, which is caused by a (rather small) non-contiguous subexpression which has only a few twigs in common with the original. However, embedded content, being more robust to non-contiguous subexpressions, ranks it relatively high. A similar argument holds for the drops at positions 22 and 87. Thus, highly non-contiguous subexpressions are still recognised as relevant by embedded content. 6.3 Set-based Twig Similarity In the nal experiment, we compared set-based approaches for terms with those for twigs. In particular, the Dice measure for terms is used as baseline for comparison with the Dice, Jaccard, and Cosine measures for twigs. Figure 11 provides an overview of the used set-based measures, each measure being non increasingly ordered as was done for the other measures in gure 7(b).0.20.61 Similarity Subexpressions Dice (terms, 0.5) Dice (terms, 1.0) Cosine-twigs Dice-twigs Jaccard-twigs Figure 11: Set-based approaches for terms and twigs. In our experiments, twig based measures have ner granularity than term based measures. This is concluded from the number of intervals imposed by these measures. The set based measures divide the subexpressions in classes (intervals) corresponding to the number of constituents in common with the original. As can be seen from gure 11, the Dice measure for terms has 10 intervals for connectors considered) and 20 for 0:5. The measures for twigs each have 26 intervals. This number directly depends on the number of dierent twigs that can be generated from the subexpressions. Since we included non-contiguous subexpressions in our experiment, new twigs are introduced that are not in the original. The original index expression has 10 twigs thus leaving room for newly introduced twigs. This suggests that twig based measures should be favoured over term based measures if delicate matching is required. The lines for the twig-based measures eventually drop to zero whereas the term-based measure does not. The reason for this is that single terms do not result in twigs although every (non-empty) subexpression has a minimal overlap of at least one term with the original. Clearly, twig based measures should thus not be used for matching single terms. The minimal similarity value according to Dice for terms (for 0:167. In order to further compare the set-based approaches, the Jaccard ordering for twigs was imposed on the others. Figure 12 shows the results. Dice and Jaccard for twigs impose the same ordering: they have exactly the same intervals. As can thus be concluded from the denitions of these measures (see section 5.1), in comparing two subexpressions, no distinction (in terms of the ordering) is made Similarity Subexpressions Dice (terms) Cosine-twigs Dice-twigs Jaccard-twigs Figure 12: Jaccard (twigs) order imposed. between the sum of twigs and the number of dierent twigs. Their similarity values, however, dier in size. This basically means that, practically, the dierences between both measures can be overcome by using rightly set thresholds. The Cosine measure for twigs has closely resembling intervals: of its 26 intervals, only six dier slightly (maximally nine subexpressions). There is a big dierence between set-based approaches for terms and those for twigs, since terms do not capture structure. The peaks in the line of Dice's term measure (gure 12) are caused by non-contiguous subexpressions, not having many twigs in common. Irrespective of their structure, however, they share terms with the original. Clearly, when the structure of descriptors is to be taken into account, twig based measures are more suitable than term based measures. The experiments illustrated characteristics of the dierent similarity measures. A number of conclusions can be drawn. First, approaches only using terms should be avoided if the structure of descriptors is relevant. This is what one would expect, as well as that set based measures using terms provide less granularity than those using twigs. Second, the dierent criteria about subexpressions resulted in similarity measures with distinct properties. The layer by layer computation of the full product measure, derived from the idea of headedness, favours descriptors that only dier in their most specic subexpressions. However inadequate for comparing index expressions that represent dierent main concepts, it naturally supports delicate discrimination of most subexpressions and a gradual notion of concept modication. The embedded content measure is not sensitive to the depth of the deleted subexpression. Furthermore, since it is based on embedding, it is more robust to non-contiguous subexpressions. Embedded content is, in the context of our experiments, best used for discriminating highly similar subexpressions. Twig based measures fall somewhere in between. They are somewhat sensitive to non-contiguous subexpressions: this introduces new twigs, but may preserve other twigs. There is little dierence between twig based measures themselves. It appears mainly a matter of thresholds, viz size of the similarity values, to distinguish between them. Conclusions In this article, we devised several similarity measures for index expressions aiming at use in a dynamic hypertext system, the INN system. This envisaged application implied our focus on dierent notions of subexpressions for index expressions and several constraints concerning high e-ciency, domain independency, and restricted memory. These constraints prohibit the use of sophisticated natural language processing techniques. However, several issues on the semantics of index expressions and corresponding notions of subexpressions were formulated in criteria. These issues were the order of subexpressions, non contiguous embedding, and headedness. The devised similarity measures were both theoretically and experimentally evaluated showing their suitability for the mentioned criteria. Although the experiments described in this article are rather small, they illustrate some of the essential characteristics of the similarity values. Future research can be directed towards augmenting index expressions with wildcards and the design of corresponding matching functions. This could aim at using the structure of index expressions for fact nding and the generation of partly specied overviews. In the augmented language, descriptors like ying from Tasmania to * can be used to nd all documents that deal with air transport from the island of Tasmania to any destination. Furthermore, it would be interesting to see if our approach is also viable for other retrieval related tasks such as clustering, routing, and ltering. This would enable an augmentation of the functionality of the INN system. For instance, the INN system could then present grouped results or it could work o-line by using stable information needs in the case of ltering. It is also interesting to investigate dierent instantiations of the abstract term and connector comparison functions. For instance, stemming could be included by computing the similarity of stems. In addition, connectors can be grouped by their function type. In matching, connectors with the same function could be given higher similarity. Including term weights potentially increases the eectiveness of matching. We did not focus on this issue here since it is not primarily concerned with the structure of index expressions. However, it seems a valuable extension worth further research. Acknowledgment The authors kindly acknowledge the valuable remarks of the referees. --R Navigational Query Construction in a Hypertext Environment. Some advances in rule-based part of speech tagging Relational Indexing Part I. Relational indexing part II. The open information locator project. Information Retrieval from Hypertext: An Approach using Plausible Inference. Retrieval Performance in FERRET. The Constituent Object Parser: Syntactic Structure for Information Retrieval. Information Retrieval. On the Application of Syntactic Methodologies in Automatic Text Indexing. Using Morpho-Syntactic Language Analysis in Phrase Matching Automatic Search Term Variant Generation. Natural Language Information Retrieval. Integrated Information Access via Structure. INdex Navigator for Searching and Exploring the WWW. Nesting and Defoliation of Index Expressions for Information Retrieval. --TR
matching;index expressions;similarity;information retrieval
593997
Skeleton Trees for the Efficient Decoding of Huffman Encoded Texts.
A new data structure is investigated, which allows fast decoding of texts encoded by canonical Huffman codes. The storage requirements are much lower than for conventional Huffman trees, O(log^2 n) for trees of depth O(log n), and decoding is faster, because a part of the bit-comparisons necessary for the decoding may be saved. Empirical results on large real-life distributions show a reduction of up to 50% and more in the number of bit operations. The basic idea is then generalized, yielding further savings.
Introduction The importance and usefulness of Data Compression for Information Retrieval (IR) Systems is today well-established, and many authors have commented on it [1, 17, 31, 27]. Large full-text IR Systems are indeed voracious consumers of storage space realtive to the size of the raw textual database, because not only the text has to be kept, but also various auxiliary files like dictionaries and concordances, which are usually adjoined to the system to make the retrieval process efficient. Moreover, certain data structures, such as decoding tables or trees, have to be resident in RAM, so that large systems require more and more powerful machines. It is therefore quite natural that efforts have been made to compress the text and other necessary files, thereby reducing the demand for storage or RAM, or equivalently, for a fixed machine with given resources, effectively increasing the size of the data base that can still be handled efficiently. Most of the popular compression methods are based on the works of Lempel and Ziv [29, 30], but these are adaptive methods which are not always suitable for IR applications. In the context of full-text retrieval, a large number of small passages is accessed simultaneously, e.g., when producing a KWIC (Key-Word-In-Context) index in response to a submitted query, and all these text fragments should be decodable, regardless of their exact location. When an adaptive coding method has been used, this would then force us to start the decoding at the beginning of the text or the logical block that contains the retrieved passage. So we would either decode much more than needed, which may imply increased processing time, or prepare a priori smaller blocks, which would cost us compression efficiency. In both cases, the advantage of using adaptive methods, which often yield better compression than static ones, may be lost. Huffman coding [14] is still one of the best known and most popular static data compression methods. While for certain applications, such as data transmission over a communication channel, both coding and decoding ought to be fast, for other applications, like the IR scenario we focus on in this paper, compression and decompression are not symmetrical tasks. Compression is done only once, while building the system, whereas decompression is needed during the processing of every query and directly affects response time. There is thus a special interest in fast decoding techniques (see e.g., [15]). The data structures needed for the decoding of a Huffman encoded file (a Huffman tree or lookup table) are generally considered negligible overhead relative to large texts. However, not all texts are large, and if Huffman coding is applied in connection with a Markov model [2], the required Huffman forest may become itself a storage problem. Moreover, the "alphabet" to be encoded is not necessarily small, and may, e.g., consist of all the different words in the text, so that Huffman trees with thousands and even millions of nodes are not uncommon [23]. We try, in this paper, to reduce the necessary internal memory space by devising efficient ways to encode these trees. In addition, the new suggested data structure also allows a speed-up of the decompression process, by reducing the number of necessary bit comparisons. The manipulation of individual bits is indeed the main cause for the slow decoding of Huffman encoded text. A method based on large tables constructed in a pre-processing stage is suggested in [5], with the help of which the entire decoding process can be performed using only byte oriented commands (see also [26]). However, the internal memory required for the storage of these tables may be very large. Another possibility to avoid accessing individual bits is by using 256-ary instead of the optimal binary Huffman codes. This obviously reduces the compression efficiency, but de Moura et al. [6] report that the degradation is not significant. In the next section, we recall the necessary definitions of canonical Huffman trees as they are used below. Section 3 presents the new suggested data structure and includes experimental results. In Section 4, the main idea is then extended, yielding yet smaller trees and even faster decoding. 2. Canonical Huffman codes Figure 1: Canonical Huffman code for For a given probability distribution, there might be quite a large number of different Huffman trees, since interchanging the left and right subtrees of any internal node will result in a different tree whenever the two subtrees are different in structure, but the weighted average path length is not affected by such an interchange. There are often also other optimal trees, which cannot be obtained via Huffman's algorithm. One may thus choose one of the trees that has some additional properties. The preferred choice for many applications is the canonical tree, defined by Schwartz and Kallick [25], and recommended by many others (see, e.g., [15, 27]). Denote by (p we assume that be the length in bits of the codeword assigned by Huffman's procedure to the element with probability is the depth of the leaf corresponding to p i in the Huffman tree. A tree is called canonical if, when scanning its leaves from left to right, they appear in non-decreasing order of their depth (or equivalently, in non-increasing order, as in [22]). The idea is that Huffman's algorithm is only used to generate the lengths f' i g of the codewords, rather than the codewords themselves; the latter are easily obtained as follows: the i-th codeword consists of the first ' i bits immediately to the right of the "binary point" in the infinite binary expansion of Many properties of canonical codes are mentioned in [15, 3]. The following will be used as a running example in this paper. Consider the probability distribution implied by Zipf's law, defined by the weights is the n-th harmonic number. This law is believed to govern the distribution of the most common words in a large natural language text [28]. A canonical code can be represented by the string hn 1 called a source, where k denotes, here and below, the length of the longest codeword (the depth of the tree), and n i is the number of codewords of length k. The source corresponding to Zipf's distribution for 74i. The code is depicted in Figure 1. We shall assume, for the ease of description, that the source has no "holes", i.e., there are no three integers This is true for many real-life distributions, and in particular for all the examples below. On the other hand, the distribution of one of the alphabets used for compressing a set of sparse bitmaps in [8] is All the techniques suggested herein can be easily adapted to the general case using a vector succ(i), giving for each codeword length i, the next larger codeword length j for which But to make the exposition clearer, we shall suppress reference to succ(i), since for all distributions without holes, One of the properties of canonical codes is that the set of codewords having the same length comprises the binary representations of consecutive integers. For example, in our case, the codewords of length 9 bits are the binary integers in the range from 110011100 to 111011010. This fact can be exploited to enable efficient decoding with relatively small overhead: once a codeword of ' bits is detected, one can get its relative index within the sequence of codewords of length ' by simple subtraction. The following information is thus needed: let be the length of the shortest codeword, and let base(i) be the integer value of the first codeword of length i. We then have denote the standard s-bit binary representation of the integer k (with leading zeros, if necessary). Then the j-th codeword of length i, for Let seq(i) be the sequential index of the first codeword of length i: Suppose now that we have detected a codeword w of length '. If I(w) is the integer value of the binary string w (i.e., is the relative index of w within the block of codewords of length '. Thus seq(') is the relative index of w within the full list of codewords. This can be rewritten as I(w) \Gamma diff('), for Thus all one needs is the list of integers diff('). Table 1 gives the values of n i , base(i), seq(i) and diff(i) for our example. 9 Table 1: Decode values for canonical Huffman code for Zipf-200 We suggest in the next section a new representation of canonical Huffman codes, which not only is space-efficient, but may also speed up the decoding process, by permitting, at times, the decoding of more than a single bit in one iteration. 3. Skeleton trees for fast decoding The following small example, using the data above, shows how such savings are possible. Suppose that while decoding, we detect that the next codeword starts with 1101. This information should be enough to decide that the following codeword ought to be of length 9 bits. We should thus be able, after having detected the first 4 bits of this codeword, to read the following 5 bits as a block, without having to check after each bit if the end of a codeword has been reached. Our goal is to construct an efficient data-structure, that permits similar decisions as soon as they are possible. The fourth bit was the earliest possible in the above example, since there are also codewords of length 8 starting with 110. 3.1 Decoding with sk-trees The suggested solution is a binary tree, called below an sk-tree (for skeleton-tree), the structure of which is induced by the underlying Huffman tree, but which has generally significantly fewer nodes. The tree will be traversed like a regular Huffman tree. That is, we start with a pointer to the root of the tree, and another pointer to the first bit of the encoded binary sequence. This sequence is scanned, and after having read a zero (resp., a 1), we proceed to the left right) child of the current node. In a regular Huffman tree, the leaves correspond to full codewords that have been scanned, so the decoding algorithm just outputs the corresponding item, resets the tree-pointer to the root and proceeds with scanning the binary string. In our case, however, we visit the tree only up to the depth necessary to identify the length of the current codeword. The leaves of the sk-tree then contain the lengths of the corresponding codewords. f tree pointer /\Gamma root start length of string f if string else tree pointer /\Gamma right (tree pointer) if value (tree pointer) ? 0 f codeword string [start output tree pointer /\Gamma root start else Figure 2: Decoding procedure using sk-tree The formal decoding process using an sk-tree is depicted in Figure 2. The variable start points to the index of the bit at the beginning of the current codeword in the encoded string, which is stored in the vector string [ ]. Each node of the sk-tree consists of three fields: a left and a right pointer, which are not null if the node is not a leaf, and a value-field, which is zero for internal nodes, but contains the length in bits of the current codeword, if the node is a leaf. In an actual implementation, we can use the fact that any internal node has either zero or two children, and store the value-field and the right-field in the same space, with serving as flag for the use of the right pointer. The procedure also uses two tables: table giving the j-th element (in non-increasing order of frequency) of the encoded alphabet; and diff [i] defined above, for i varying from m to k, that is from the length of the shortest to the length of the longest codeword. The procedure passes from one level in the tree to the one below according to the bits of the encoded string. Once a leaf is reached, the rest of the current codeword can be read in one operation. Note that not all the bits of the input vector are individually scanned, which yields possible time savings. 9 91010 Figure 3: sk-tree for Zipf-200 distribution Figure 3 shows the sk-tree corresponding to Zipf's distribution for 200. The tree is tilted by 45 ffi , so that left (right) children are indicated by arrows pointing down (to the right). The framed leaves correspond to the last codewords of the indicated length. The sk-tree of our example consists of only 49 nodes, as opposed to 399 nodes of the original Huffman tree. An idea similar to the sk-tree, but based on tables rather than on trees, has been suggested by Moffat and Turpin [22]. Instead of identifying roots of subtrees in which all codewords have the same depth, they essentially form a complete tree to a fixed depth no less than the depth of the code tree (by extending any shorter branches), and examine the code tree nodes at that depth to determine the minimum codeword length in each subsidiary subtree. To find the length of a codeword, a fixed-sized window of the compressed bitstream, considered as a binary value, is compared with left-justified base values in a sequence of hard-coded cascading if-statements. Each such comparison is equivalent to a transition to a left or right child of the sk-tree, and the replacement of bit comparisons by equivalent byte or word based comparisons is reminiscent of a mechanism suggested in [5]. 3.2 Construction of sk-trees While traversing a standard canonical Huffman tree to decode a given codeword, one may stop as soon as one gets to the root of any full subtree of depth h, for h - 1, i.e., a subtree of depth h that has 2 h leaves, since at this stage it is known that exactly h more bits are needed to complete the codeword. One way to look at sk-trees is therefore as standard Huffman trees from which all full subtrees of depth h - 1 have been pruned. A more direct and much more efficient construction is as follows. The one-to-one correspondence between the codewords and the paths from the root to the leaves in a Huffman tree can be extended to define, for any binary string path P (S) induced by it in a tree with given root r 0 . This path will consist of e e, where for i ? 0, r i is the left (resp. right) child of r For example, in Figure 3, P(111) consists of the four nodes represented as bullets in the top line. The skeleton of the sk-tree will consist of the paths corresponding to the last codeword of every length. Let these codewords be denoted by L are, for our example, etc. The idea is that P (L i ) serves as "demarcation line": any node to the left (resp. right) of P (L i ), i.e., a left (resp. right) child of one of the nodes in P (L i ), corresponds to a prefix of a codeword with length - As a first approximation, the construction procedure thus takes the tree obtained by (there is clearly no need to include the longest codeword L k , which is always a string of k 1's), and adjoins the missing children to turn it into a complete tree in which each internal node has both a left and a right child. The label on such a new leaf is set equal to the label of the closest leaf following it in an inorder traversal. In other words, when creating the path for L i , one first follows a few nodes in the already existing tree, then one branches off creating new nodes; as to the labeling, the missing right child of any node in the path will be labeled (basing ourselves on the assumption that there are no holes), but only the missing left children of any new node in the path will be labeled i. A closer look then implies the following refinement. Suppose a codeword L i has a zero in its rightmost position, i.e., L string ff of length i \Gamma 1. Then the first codeword of length It follows that only when getting to the i-th bit one can decide if the length of the current codeword is i or i + 1. But if L i terminates in a string of 1's, L with a ? 0 and jfij then the first codeword of length so the length of the codeword can be deduced already after having read the bit following fi. It follows that one does not always need the full string L i in the sk-tree, but only its prefix up to and not including the rightmost zero. Let L prefix. The revised version of the above procedure starts with the tree obtained by The nodes of this tree are depicted as bullets in Figure 3. For each path P (L there is a leaf in the tree, and the left child of this leaf is the new terminal node, represented in Figure 3 by a box containing the number i. The additional leaves are then filled in as explained above. 3.3 Space complexity To evaluate the size of the sk-tree, we count the number of nodes added by path P (L k. Since the codewords in a canonical code, when ordered by their corresponding frequencies, are also alphabetically sorted, it suffices to compare L i to L i\Gamma1 . Let empty string, and for i ? m, let fl(i) be the longest common prefix of L i and L is the string 10 in our example. Then the number of nodes in the sk-tree is given by: since the summation alone is the number of internal nodes (the bullets in Figure 3). The maximum function comes to prevent an extreme case in which the difference might be negative. For example, if L then the longest common prefix is but since we consider only the bits up to and not including the rightmost zero, we have L In this case, indeed, no new nodes are added for P (L An immediate bound on the number of nodes in the sk-tree is O(min(n; k 2 )), since on the one hand, there are up to 2, but on the other hand, it cannot exceed the number of nodes in the underlying Huffman tree, which is 2n \Gamma 1. To get a tighter bound, consider the nodes in the upper levels of the sk-tree belonging to the full binary tree F with leaves and having the same root as the sk-tree. The depth of F is 1)e, and all its leaves are at level d or d \Gamma 1. The tree F is the part of the sk-tree where some of the paths P (L must be overlapping, so we account for the nodes in F and for those below separately. There are at most 2k \Gamma 1 nodes in F ; there are at most k \Gamma 1 disjoint paths below it, with path P (L extending at most F , for log This yields as bound for the number of nodes in the sk-tree: There are no savings in the worst case, e.g., when there is only one codeword of each length (except for the longest, for which there are always at least two). More generally, if the depth of the Huffman tree is \Omega\Gamma n), the savings might not be significant. But such trees are optimal only for some very skewed distributions. In many applications, like for most distributions of characters or character pairs or words in most natural languages, the depth of the Huffman tree is O(log n), and for large n, even the constant c, if the depth is c log 2 n, must be quite small. For suppose the Huffman tree has a leaf on depth d. Then by [16, Theorem 1], the probability of the element corresponding to this leaf is F j is the j-th Fibonacci number, and we get from [18, Exercise 1.2.1-4], that 5)=2 is the golden ratio. Thus if d ? c log 2 n, we have To give a numeric example, in Section 4 below one of the Huffman trees corresponds to the different words in English, with leaves. The probability for a tree of this size to have a leaf at level 3 log 2 n is less than 4:4 \Theta 10 \Gamma12 , meaning that such a word occurs only once every 4400 billion words; the existence of such a rare word then puts a lower limit on the size of the text, which in our case must be large enough to fill about 35,000 CD-Roms! For all the distributions given in Table 2 in the experiments below, the ratio of the depth of the Huffman tree to log 2 n is between 1.31 and 2.61. But even if the original Huffman tree would be deeper, it is sometimes convenient to impose an upper limit of O(log n) on the depth, which often implies only a negligible loss in compression efficiency [10]. In any case, given a logarithmic bound on the depth, the size of the sk-tree is about log n (log 3.4 Time complexity When decoding is based on a standard Huffman tree, the average number of comparisons per codeword is the sum, taken over all the leaves i, of the depth of i in the tree times the probability to get to i. A similar sum holds for sk-trees, with the difference that a leaf does not correspond to a single element, but to several consecutive codewords of the same length. Let w be the prefix of a codeword corresponding to a leaf of the sk-tree labeled and denote codewords corresponding to this leaf of the sk-tree are correspond, using the notations of Section 2, to indices in the range from diff('). The average number of comparisons per codeword using the sk-tree can thus be evaluated as: i2fleaves in sk-treeg@ d i where w i is the binary string corresponding to the leaf i, d is the depth of i in the tree, ' i is a shortcut for label(i), and Prob(j) is the probability of the element with index j. As an approximation, we assume that the probability of an element on level i in the tree is 2 \Gammai . This corresponds to a dyadic probability distribution, where all the probabilities are integral powers of 1. There cannot be too great a difference between the actual probability distribution and this dyadic one, since they both yield the same Huffman tree (see [20] for bounds on the "distance" between such distributions). Given this model, eqn. (2) becomes i2fleaves in sk-treeg A similar sum, but taken over all the leaves of the original Huffman tree gives the average codeword length for a dyadic distribution. There are therefore large savings whenever the number of nodes in the sk-tree is much smaller than in the underlying full Huffman tree. 3.5 Experimental Results To test the effectiveness of the use of sk-trees, the following real-life distributions were used. The data for French was collected from the Tr'esor de la Langue Fran-caise, a database of 680 MB of French language texts (115 million words) of the 17 th -20 th centuries [4]; for English, the source are 500 MB (87 million words) of the Wall Street Journal [24]; and for Hebrew, a part of the Responsa Retrieval Project , 100 MB of Hebrew and Aramaic texts (15 million words) written over the past ten centuries [7]. The first set of alphabets consists of the bigrams in the three languages (the source for English for this distribution was [13]); for the next set, the elements to be encoded are the different words, which yields very large "alphabets"; and the final set contains the distribution of trigrams in French. For completeness, the Zipf-200 distribution used in the above examples was also added. total average number of average relative Source k number of codeword nodes in number of savings in # elements length sk-tree comparisons comparisons bigrams French 29 2192 7.784 285 4.620 40.6% Hebrew 24 743 8.037 127 4.183 48.0% English 26 289101 11.202 425 5.726 48.9% words French 27 439191 10.473 443 5.581 46.7% Hebrew trigrams French 28 25781 10.546 381 5.026 52.3% Table 2: Time and Space requirements for real-life distributions Table 2 displays the results. The first three columns give some statistics about the various distributions: the depth k of the Huffman tree, the size n of the encoded alphabet, and the weighted average length of a codeword, measured in bits, which equals the average number of comparisons if the standard Huffman tree is used. The next two columns bring the number of nodes in the sk-tree, as given in eqn. (1), and the average number of comparisons per codeword when decoding is based on the sk-tree, as given in eqn. (2). The final column shows the relative savings in the number of comparisons, measured in percent. We see that for large distributions, roughly half of the comparisons may be saved. Note that these savings are in spite of the fact that the high-probability symbols with short codewords have relatively few bits in common. The weighted average takes this into account: few bits are saved for the shorter codewords, and the savings are multiplied by higher probabilities; more bits are saved for the longer codewords, but even if their probabilities are very small, their large number have a cumulative effect. Note also that the cost of storing the sk-tree is only several percent of the cost for the full Huffman tree. 4. Reduced sk-trees We now wish to explore what might be gained by pruning the sk-tree at some internal node: one would thus get to leaves at which it is not yet possible to deduce the length of the current codeword, but at which some partial information is already available. For example, in Figure 3, if the bits already processed were 111 (corresponding to the internal node in the rightmost upper corner), we know already that the length of the current codeword is either 9 or 10. We therefore need only one more comparison to know the exact length: concatenate the following seven bits with the three already processed to get a 10-bit string w; if the binary value of w is smaller than base(10), the next codeword must be of length 9, otherwise it is of length 10. If we had used the original sk-tree as explained in the previous section, we would have had at least one more comparison, possibly even more, e.g., if the bits after 111 were 0110, we would have performed four more comparisons and still not know if the length is 9 or This reflection leads to the idea of a reduced sk-tree, which is obtained from the sk-tree by pruning some of its branches. On the one hand, this reduced tree is obviously smaller, on the other, as we saw, it may also decrease the number of comparisons. More formally, define for each node v of the sk-tree two values lower(v) and upper(v) by: if v is a leaf lower(v) if v is an internal node lower(v) = lower(left(v)) that is, for each node v, the codewords corresponding to leaves of the sub-tree rooted by v have their lengths in the interval [lower(v); upper(v)]. In terms of our earlier notation we have We define the reduced sk-tree as the smallest sub-tree of the sk-tree for which all the leaves w correspond to a range of at most two consecutive codeword lengths, i.e., 5-6 Figure 4: Reduced sk-tree for Zipf-200 distribution Figure 4 is the reduced sk-tree obtained from the sk-tree of Figure 3. Leaves are now also indicated as bullets, with the corresponding range underneath. Recall that the original Huffman tree had 399 nodes, and the sk-tree 49; in the reduced sk-tree we are left with only 13. Note that all the leaves of the original sk-tree are deleted, but also entire sub-trees. The nodes corresponding to the part of P (L which is not overlapping with P (L eqn. (3), but since we seek the minimal tree, for each such path, only the node highest up in the tree need be kept, so the rest of this branch and its offsprings are pruned. A generalized view of both regular and reduced sk-trees would be as follows: consider a full canonical Huffman tree and assign to each node the values lower and upper. Delete now some of the nodes, starting at any leaf and proceeding to the parent nodes, until you get to the smallest tree for which every leaf w satisfies lower(w) = upper(w), i.e., all the corresponding codewords have the same length; this is the sk-tree. If the process is continued and more nodes are deleted until the codewords corresponding to the new leaves have lengths i or for some i, we get the reduced sk-tree. We henceforth adopt the notation sk 1 -tree and sk 2 -tree for the original and the reduced sk-trees, respectively, the subscript referring to the maximal size of the set of codeword-lengths associated with each of the leaves of the tree. We cannot use equality in eqn. (3), which would impose a range of exactly two codeword lengths for each leaf of the sk 2 -tree. In the example of Figure 4 all the leaves do have equality Last elements of codeword blocks 5-6 6-7 12Corresponding sk 2 -tree Figure 5: Example of sk 2 -tree with special leaves in eqn. (3), but for other examples, leaves may exist, the parent nodes of which correspond already to ranges of 3 or more codeword lengths. In this case, the original leaf of the sk 1 - tree must be kept. Let us call such leaves in the sk 2 -tree special leaves. For the example distributions above, special leaves exist only for the distribution of the Hebrew bigrams, the first few elements of the source being h0; 0; 0; 0; In the left part of Figure 5, the last codewords L i for the codeword lengths up to 13 are listed, and the right part of Figure 5 is the corresponding section of the sk 2 -tree. The special leaves w are indicated as rectangles, containing the value lower(w) (which is equal to upper(w)), the other leaves are depicted as bullets as above. For example, we see that only codewords of length 8 can have the prefix 011, but the parent node of the corresponding leaf is associated with the prefix 01, which may be extended to codewords of lengths 6, 7 or 8. Similarly, a prefix 11110 implies the codeword length 11, but 1111 is the prefix of codewords of lengths 11 to 24. The decoding procedure for sk 2 -trees is similar to that of the sk 1 -trees given in Figure 2, and only the if-block has to be replaced by the one in Figure 6. We now use a f lag field for each leaf w, with f otherwise. The value field of w stores lower(w) if w is a leaf, and 0 if w is an internal node. if value (tree pointer) ? 0 f len /\Gamma value (tree pointer) codeword string [start if f lag (tree f codeword string [start len output tree pointer /\Gamma root start Figure Decoding using sk 2 -tree When a leaf w is reached, the current codeword is initialized as having length lower(w). This is the correct setting if w is a special leaf or if the next codeword has indeed length lower(w). When w is not a special leaf (f lag(w) = 1), we check if by appending a zero at the right end of the codeword, we get an integer value larger or equal to that of the first codeword of length lower(w) + 1. If so, we update the current codeword to include also the following bit. The construction of sk 2 -trees is similar to that of the underlying sk 1 -tree. We again consider the paths of nodes only those nodes that appear in at least two different paths; these are the internal nodes of the sk 2 -tree. The leaves are then added by filling in the missing left and right children, some of which may be special leaves. As to the space complexity of sk 2 -trees, note that in principle, several special leaves may emanate from a single branch P (L leaves the upper bound for the number of nodes at O(min(n; k 2 )) as for sk 1 -trees. But in practice, special leaves are rare, because they appear only in the very particular case when P (L contained in either P (L In the former case, the special leaves are right children of nodes in P (L in the latter they are left children. For example, referring to the tree in Figure 5, L is a prefix of L and generates a special leaf as a right child, whereas L contains L as a prefix and generates special leaves as left children. If for a given Huffman tree, there are no special leaves in its associated sk 2 -tree, as was the case in all our examples beside the one of Figure 5, then the number of nodes is clearly there is exactly one leaf for each range and the sk 2 -tree is a complete tree, i.e., each internal node has exactly two children. The sizes of the sk 2 -trees for our earlier example distributions are listed in Table 3. As can be seen, even for huge Huffman trees with hundreds of thousands of nodes, this size is reduced to several tens, and there is a 70-90% reduction even relative to the sizes of the sk 1 -trees. number of Savings average Savings Source nodes in rel. to number of rel. to sk2 -tree sk1 -tree comparisons sk1 -tree English 15 78% 3.444 18.0% bigrams French 47 84% 3.757 18.7% Hebrew English 41 90% 4.842 15.4% words French 41 91% 4.725 15.3% Hebrew 33 90% 4.715 17.2% trigrams French 43 89% 4.157 17.3% Table 3: Time and Space requirements for sk 2 -trees To evaluate the average number of comparisons, we take a sum similar to eqn. (2) over all the leaves of the sk 2 -tree. For the special leaves, the formula of eqn. (2) applies. For the others, let w be the prefix of the corresponding codeword, assume the leaf is labeled ', jwj. Then the codewords corresponding to this leaf of the sk 2 - tree are w0 . The first few of them are of length ' and the following ones of length '+1. The exact cutoff point is not important, as the codewords correspond to the consecutive indices in the range from I(w0 t 1). The probability of each of these codewords should be multiplied by the number of necessary comparisons to detect them, which is jwj + 1, since we need an additional comparison to decide if the length is ' or 1. This yields, using the same notations as for eqn. (2), the following formula: i2fleaves in sk-treeg@ d Table 3 gives the resulting averages for our examples. For the real-life examples, they give a reduction of 50-64% relative to the regular Huffman decoding algorithm, and of 15-19% relative to the algorithm using sk 1 -trees. 5. Final remarks If pruning the skeleton tree turned out to be profitable in terms of both time and space, shouldn't we climb up even higher and define sk d -trees accordingly, for d ? 2? We can associate a value range-size with each node v of the sk 1 -tree, giving the size of the set of the corresponding codeword lengths. The leaves of the sk 1 -tree have all range-size those of the sk 2 -tree have range-size - 2. Consider a path that starts at any leaf of the moves through parent pointers towards the root. The range-size values of the nodes in this path form a non-decreasing sequence, the first value being 1, followed possibly by several 2's, etc. Fixing, for all such paths, the last node with value 2 (if it exists) as a new leaf yields the sk 2 -tree. Similarly, proceeding even further up to the last node with value 3 would yield an sk 3 -tree, etc. However, the savings incurred by passing from the sk 1 -tree to the sk 2 -tree were caused by the fact that several consecutive nodes on these paths had the range-size value 2, so that the new leaves were several levels higher, and accordingly several comparisons could be saved. But if the parent node of a node with range-size value 3 has also range-size value 3, the other child of this parent node must have range-size value 1, which means that it is a special leaf. We argued earlier that such cases are rare. Therefore, whenever no special leaf is involved, passing from the lowest node with range-size value 2 to the lowest node with range-size value 3 would just let us climb one level and save one comparison. On the other hand, we need now an additional comparison within the range of 3 values, so that in all these cases, nothing is gained. Of course, for the price of two additional comparisons, we could process, using binary search, ranges of size 4 and not just 3. More generally, we need only r additional comparisons after reaching a leaf of an sk 2 r -tree. Pushing this idea to its extreme with would be no skeleton-tree at all, and we would find the correct length of a codeword using a sequence of binary search steps in the list of first (or last) codewords for every codeword length, as suggested in [22]. But with standard binary search, the search in a code with maximal codeword length k takes exactly dlog 2 ke comparisons, which would be 4 or 5 for our example distributions. Note that while the average number of comparisons with sk 1 -trees is above that threshold for all the examples (Table 2), all the corresponding values for the are below it (Table 3). It does therefore not seem necessarily worthwhile to pass to sk d -trees, for d ? 2. Moffat and Turpin [22] further suggest to use a biased binary search, since the probability distribution of the codeword lengths is itself very skewed. For the first few bits of any code- word, this approaches then a linear search. The skeleton-trees introduced in this paper are a convenient data structure to perform a similar search efficiently. --R Is Huffman coding dead? Ziff D. Perl Y. Fast searching on compressed text allowing errors All about the Responsa Retrieval Project you always wanted to know but were afraid to ask Rabinowitz J. Moore E. Information Retrieval A method for the construction of minimum redundancy codes Nemetz T. The Art of Computer Programming An application of informational divergence to Huffman codes On the implementation of minimum redundancy prefix codes Text compression for dynamic document databases Kallick B. Fast decoding of the Huffman codes The Psycho-Biology of Language A universal algorithm for sequential data compression Compression of individual sequences via variable-rate coding Adding compression to a full-text retrieval system --TR --CTR Dana Shapira , Ajay Daptardar, Adapting the Knuth-Morris-Pratt algorithm for pattern matching in Huffman encoded texts, Information Processing and Management: an International Journal, v.42 n.2, p.429-439, March 2006
canonical Huffman codes;skeleton trees;huffman trees;decoding
594016
Measuring Search Engine Quality.
The effectiveness of twenty public search engines is evaluated using TREC-inspired methods and a set of 54 queries taken from real Web search logs. The World Wide Web is taken as the test collection and a combination of crawler and text retrieval system is evaluated. The engines are compared on a range of measures derivable from binary relevance judgments of the first seven live results returned. Statistical testing reveals a significant difference between engines and high intercorrelations between measures. Surprisingly, given the dynamic nature of the Web and the time elapsed, there is also a high correlation between results of this study and a previous study by Gordon and Pathak. For nearly all engines, there is a gradual decline in precision at increasing cutoff after some initial fluctuation. Performance of the engines as a group is found to be inferior to the group of participants in the TREC-8 Large Web task, although the best engines approach the median of those systems. Shortcomings of current Web search evaluation methodology are identified and recommendations are made for future improvements. In particular, the present study and its predecessors deal with queries which are assumed to derive from a need to find a selection of documents relevant to a topic. By contrast, real Web search reflects a range of other information need types which require different judging and different measures.
Introduction Publication of accurate and meaningful evaluation of the quality of results returned by public Web search engines not only enables informed consumer choice but also assists and encourages search engine operators to improve their standard of service. Since many search engines claim to be using novel techniques, effectiveness comparisons between these engines and systems employing published methods are potentially of interest to the Information Retrieval (IR) research community. It is known that some search engine operators already perform extensive internal evaluations, but it is not clear that these evaluations have the properties (blindness, independence, reproducibil- ity) which would normally be expected of published studies in IR. On the other hand, it is possible that the model of the user retrieval task implicit in these evaluations more accurately represents everyday Web search than do the models assumed in much published IR work. Countless evaluations of text retrieval systems eg. [16, 18] have measured the effectiveness of relevance ranking schemes in the context of a well-defined static document collection for which sufficiently complete relevance information is available. Such evaluations are reproducible and can provide accurate measurements of recall (the proportion of all relevant documents which have been retrieved so far) as well as precision (the proportion of retrieved documents which are relevant). Early test collections, such as those described in [4] and [16] were small enough to permit relevance judgments for every document against every request. Exhaustive judging is infeasible with collections the size of TREC [18] and "almost complete" judgments are collected by assessing only the pool of documents retrieved by a large and diverse set of retrieval methods. All unjudged documents are assumed to be irrelevant. TREC is subject to continual evolution and now models some specific features of Web search. For example, in TREC-8, the Large Web Task involved blind effectiveness evaluations of retrieval systems operating over 18.5 million pages crawled from the Web [11]. Each participant was required to process 10,000 queries taken from Web search engine logs and return the top 20 ranked documents for each. Approximately 50 queries were selected post hoc for relevance assessment using binary measures. The large number of queries and the size of the data set served to rule out the use of unrealistically inefficient algorithms. Here, we report on the application of an extended TREC-8 Large Web task methodology to the effectiveness evaluation of 20 public search engines. When evaluating public search engines without the knowledge or cooperation of their operators, there is no fixed test collection and it is consequently necessary to evaluate crawling and document retrieval algorithms in combination. Finally, we present an analysis of the extent to which characteristics of Web search have been captured by this experiment and argue for further methodological developments both to improve the applicability of evaluation results and to reduce the cost of obtaining them. 2 Relationship to other studies Gordon and Pathak [8] distinguish between two types of search engine evaluation: testimonials, encompassing informal and impressionistic appraisals and feature-list comparisons; and shootouts, which correspond more closely to traditional IR effectiveness experiments. Here, only the latter are considered. Gordon and Pathak present a table 1 of twelve earlier shootout studies but identify only three (including their own) which make use of "appropriate experimental design and evaluation." Of these, that of Gordon and Pathak is the most comprehensive and most recent. Ding and Marchionini Table 1, p. 148. [5] evaluated three engines on only five topics in 1996 and found no statistically significant difference between effectiveness means. In 1997 2 Leighton and Srivastava [14] compared five engines using topics and found that three of the engines were superior to the other two, using a new measure based on precision at cutoff 20, but assigning different weights to the top three, next seven and next ten results. Leighton and Srivastava constructed a set of topics from a variety of sources which were intended to model the information needs of undergraduate students. They themselves generated queries from the topics. Seven queries were simple "bag of words" queries, seven were structured in some way and one attempted to locate information about a person. Leighton and Srivastava used automatic scripts to submit queries to the engines and fetch results. After suppressing information about which engine had retrieved each page, they judged relevance themselves using four categories of relevance. They also identified dead links and duplicate pages. They analysed their data several times using different relevance thresholds and with and without penalising duplicates. Gordon and Pathak obtained 33 real information needs from volunteers among the faculty members in a university business school. These were recorded in considerable detail and passed to skilled search intermediaries who were given the task of generating near-optimal queries for each of eight search engines by an interactive, iterative process. The top 20 live results generated by each of the engines in response to the final queries were then printed and returned to the originating faculty member for assessment on a four point relevance scale. They found that search effectiveness was generally low, that there were significant differences between engines and that the ranking of engines was to some extent dependent upon the strictness of the relevance criterion. The characteristics of the Leighton and Srivastava and Gordon and Pathak studies are compared with the present experiment in Table 1. We compare our results with those of Gordon and Pathak in Section 4.3. 2.1 Evaluation philosophy Gordon and Pathak 3 present a list of seven evaluation features which they claim should be present to maximise accuracy and informativeness of evaluation. Paraphrasing for brevity, these are: 1. Searches should be motivated by genuine user need. 2. If a search intermediary is employed, the primary searcher's information need should be as fully captured as possible and transmitted in full to the intermediary. 3. A large number of search topics must be used. 4. Most major search engines should be included. 5. The most effective combination of specific features of each search engine should be exploited. Ie. the queries submitted to the engines need not be the same. 6. Relevance judgments must be made by the individual who needs the information. 7. Experiments should be well designed and conducted. Features 3 and 7 are essential features of a useful scientific study and Feature 2 is certainly desirable (when applicable). However, the value of Feature 4 is dependent upon the goals of the evaluation. For general-purpose evaluation, we propose an additional feature: 2 the work was not formally published until 1999 3 p. 146 8. The search topics should represent the range of information needs both with respect to subject and to type of results wanted. (See Section 6 below.) The remainder of the features in the original Gordon and Pathak list merit detailed discussion. Restricting searches to genuine user needs (Feature 1) and requiring that the person with the original need should evaluate the results (Feature must inevitably limit the scope of evaluations. For example, the Gordon and Pathak study itself was limited to the information needs of faculty members in a business school. We acknowledge that judgments of relevance vary from person to person and from time to time. However, there is empirical evidence that retrieval system rankings remain stable across different sets of relevance judgments[21]. We would prefer to say that searches should be representative of genuine user need and that relevance judging across the aggregated search results should be consistent (one judge should evaluate all responses to a topic in as short a time as possible.) Feature 5 is quite contentious. All well-known public search engines are designed to produce a list of results when a set of query words (without operators or special syntax) is typed into the search box provided. It is therefore perfectly reasonable to compare the quality of results produced by search engines given identical input queries in this form. Furthermore, such queries are far more typical of Web search than are sophisticated queries exploiting advanced query language features. Analyses of query logs ([17] for example) show that typical search engine users very rarely use any form of query operator and, when they do, they frequently make errors. Measures obtained by studies which adopt the approach of trying to find the best query formulation for each search engine (Feature 5) are certainly interesting, particularly if they indicated that these "near-optimal" queries performed dramatically better than simple queries. However, the Gordon and Pathak study did not compare the performance of the highly tuned queries with simpler versions such as the initial Boolean queries supplied by the search originators or simple word lists. Unfortunately, in order to achieve the fifth desideratum within their own study, Gordon and introduced a set of confounding variables due to the introduction of skilled search intermediaries charged with converting user topic specifications into near-optimal queries for each engine. What they have evaluated is in fact a set of combinations of human intermediary and search engine, with all the human variability that that entails. The use of human intermediaries also seems somewhat inconsistent with the reasoning behind requiring relevance assessments to be conducted by the originator of the information need (Feature 6). In order to form queries from the original user's topic statement, and to judge which variant of a query performs best, the intermediary must make the same type of interpretations which would need to be made by relevance judges other than the topic originator. 3 The experiment We applied TREC-style methodology to the evaluation of Web search engines, as they operated on September 20, 1999. Because we used queries taken from real query logs, our study exhibits all of the eight features listed in the preceding section, except for those numbered 2 (not applicable), 5 and 6. As discussed in the preceding section, we feel that there are strong reasons not to adopt Feature 5 and reasons why Feature 6 may not necessary. Table 1: Comparison of the present experiment with the two most closely related previous studies. Notes: a - including stopwords. b - Subject to the result page changeover bug for some engines (see text). c - The operation of the RAT (Relevance Assessment Tool) is described in the text. d - plus duplicate and dead link categories. Detail Gordon and Leighton and The present Number of search engines 8 5 20 Number of topics 33 15 54 No. relevant found per topic 5-125, mean 42.2 not stated 2-255, mean 88.5 Date queries submitted 1998 early 1997 20 Sep. 1999 Information needs presumed originators faculty members library clients anonymous searchers - length of topic statement approx. 100 wds 4.9 wds 5.9 wds - range of subjects business broad broad - types of answer required selection of rel. pages selection of rel. pages a variety Queries How generated? human intermediary by experimenter from from Web logs verbal requests (verbatim) Average number of words not stated 4.9 5.9 a Used query operators yes some no - How many results judged top 20 live top 20 top 20 live b How were pages judged? - by whom inquirers experimenters research assistants - by topic originator yes no no - result lists merged? yes yes yes - blind judging yes yes yes - follow hyperlinks no no no images both raw HTML rendered text presentation printed paper EMACS text editor browser page truncation after ten pages no no - order of presentation random not stated increasing length - auto. search aids for judging none not stated RAT c Relevance categories 4 4 d 2 Measures reported - P@20 in graphs weighted yes b - P@15-20 yes no no average P@1-5 yes no yes Prec. based on strict rel. yes yes no - MRR1 no no yes - relative recall @ 20 yes no no TREC-style ave. prec. no no yes 3.1 What was evaluated? Most current Web search engines consist of two key components. One is a crawler [12] whose job it is to create a Web snapshot 4 " by identifying, selecting and fetching documents and the other is a text retrieval system (TRS) operating over the snapshot collection. The quality of search results clearly depends upon the performance of both search engine com- ponents. Relevant document documents may fail to be retrieved either because they were not fetched by the crawler or because the TRS failed to rank them appropriately. There is evidence [10] which suggests that precision at fixed cutoff increases with collection size. We evaluated both components in combination. This is a necessary extension of conventional evaluation methodology, in which there is a static, well-defined test collection and only the TRS is evaluated. 3.2 Which search engines? Lawrence and Giles[13] report measurements taken in February 1999, of the coverage and freshness of indexes maintained by Web search engines. These measurements relate principally to the performance of the crawler rather than the TRS. They found that none of the most popular search engines covered more than 16% of their estimated total of 800 million indexable web pages 5 . To add an extra dimension to the data presented by Lawrence and Giles, we measured the result-list precision of all eleven engines considered in their study: Northern Light, Snap, AltaVista, HotBot, Microsoft, Infoseek, Google, Yahoo, Excite, Lycos and Euroseek. We added: 1. Two metasearch engines: MetaCrawler (mentioned but not studied by Lawrence and Giles and accessing About.com, AltaVista, DirectHit, Excite, GoTo.com, Infoseek, Looksmart, Lycos, Thunderstone, WebCrawler and Yahoo) and Inquirus (operated by Lawrence and Giles on a limited-access basis and accessing AltaVista, DirectHit, Euroseek, Excite, Fast, Google, Hotbot, Infoseek, Lycos, NorthernLight, Open Directory, Snap, Thunderstone, Yahoo and Yahoo/Inktomi). Metasearchers do not index documents themselves but rather forward queries to a number of primary search engines and form a composite result list. 2. FAST (also known as "AllTheWeb", a new large search engine which aims to index the whole of the Web; 3. EuroFerret, another Europe-focused search service we wished to compare with EuroSeek; 4. DirectHit, a system which takes page popularity (in terms of access frequency) into account when ranking pages; 5. Three search engines providing search within the Australia and New Zealand area: ANZwers, ExciteAus, and Web Wombat; 6. A directory service (LookSmart Australia). This service presents Australian results first but then appends results from its Looksmart Worldwide counterpart. Note that several of these services are not fully automatic and generate responses at least partially on the basis of classifications made and stored by human editors. Prominent examples include the Yahoo, Lycos and LookSmart Australia directory services. Given that these services 4 This term captures the intention but not the long-drawn-out and incomplete nature of actual crawls. 5 Any item accessible via the Web is conventionally called a Web page. When printed, a Web page may cover many pages or just a few lines. accepted queries and presented results in a fashion which appeared to be ordered, we considered that it was reasonable to evaluate them alongside search services whose degree of manual intervention is less. Our list of twenty engines included all of those in the Gordon and Pathak [8] study except for OpenText and Magellan. The overlap between the two studies therefore comprises six engines: AltaVista, Excite, Infoseek, Hotbot, Lycos and Yahoo. Leighton and Srivastava [14] studied the first five of these, but did not report comparable measures. 3.3 Queries and result lists. We started with two sets of 100,000 single-sentence "natural language" queries from logs supplied to us by Alta Vista [1] and Electric Monk. [7] We used natural language queries because we believed that it would be easier for our judges to interpret what it was that the inquirer was actually seeking. We merged the two sets and filtered out queries which were either likely to be offensive to some people or which contained fewer than two non-stopwords 6 . We then randomly selected queries from the merged set and asked an experienced judge to decide whether she and her colleagues would be able to interpret what it was the inquirer was seeking and be able to judge the relevance of results returned. She rejected 719. who are the current supreme court justices? historic preservation where can i find gif files? where can i find film reviews? how do i create a web site? where can i find the best jokes? where can i find information on russia? how does a digital camera work? thalidomide and multiple sclerosis slobadan milosevic hindenburg disaster old japanese science fictions movies where can i find information on school violence? what are some psychological principles and attitudes for advertising why do feet smell? armstrong louis where can i find statistics for education in the united states? methodist sermons egyptian history wall street journal how to start business find information about american anarchists reasons for studying marketing where can i find the saints and the catholic church? show me a list of vegetarian restaurants in new york city. Figure 1: A sample of the queries used in the experiments. 6 A stopword is a functional word like "the", "and", "of ", and "but". A total of 59 queries were accepted but due to a pair of near-duplicate queries and problems of assessor availability, 57 were judged for the TREC Large Web track and a subset of 54 for the search engines. A sample of the judged queries is shown in Figure 1. We did not correct typographical errors in the queries (such as "slobadan" in the figure). They were submitted to the search engines exactly as shown. We used scripts to present each of the queries to each search engine. The first 20 "live" results from each engine for each query were retrieved and merged into a single pool of pages to be judged. To enable fair comparison with TREC systems, "dead" or inaccessible results were ignored and engines were not penalised for returning them. 3.4 Relevance assessment We employed a team of six judges, each of them an Australian with a University degree, but with no specific expertise in Computer Science or Information Retrieval, and no allegiance to any search engine or retrieval system. We asked them to judge whether documents were relevant to particular queries. The pages retrieved by all search engines were combined into a single pool for each query and presented to the judges without indication of which search engine(s) had retrieved them. The documents were rendered using a text-only browser (Lynx, [15]) and images or sounds were not presented. Hyperlinks were presented as numbered references, eg. [4], but the judges could not follow them. We asked judges to imagine that they themselves had submitted the queries and to evaluate the answers on that basis. However, we also instructed them to judge each document independently of the others and to score as relevant any page which included material which was "on-topic" and which contributed information not contained in the query. We asked them not to make any judgment about the correctness of information given and not to require that pages returned in response to a question were in the form of an answer. In line with past TREC methodology, judgments were binary and made on the basis of textual content alone. Either a document contained relevant content and was judged relevant or it was judged irrelevant. To ensure consistency of results, all the documents retrieved for a query were judged by the same person. Each document retrieved in response to a query was judged by only one judge, as earlier work ([9] and [21]) failed to demonstrate any particular benefit of multiple judgments. Our Relevance Assessment Tool We used special relevance judging software known as the RAT (Relevance Assessment Tool) which was developed by Jason Haines and Paul Thistlewaite in 1996 for use in the TREC Very Large Collection and Web tracks. Since then it has been maintained by Nick Craswell. Figure 2 shows a screen snapshot of the RAT in action. Before commencing judgments for a new query, each judge was required to enter concepts they would use in evaluating relevance and a criterion of relevance. For example, the concepts for the last query in Figure 1 were "vegetarian", "restaurants" and "new york" and the criterion of relevance was the conjunction of all concepts. We encouraged judges to accumulate lists, using RAT facilities, of words, phrases and part- words which in their view constituted evidence for the presence of each concept. For example, "ny", "yc" 7 , "manhattan" and "new york city" as well as "New York" constituted evidence for the presence of the "New York City" concept. Judges entered evidence either by typing into an add evidence box or by selecting text from presented pages. 7 We don't know why this was chosen. Figure 2: A screenshot of the Relevance Assessment Tool (RAT) used in the present experiments. Coloured rectangles representing the concepts are at the top. (In this case, Concept 0 was a dummy entry created by the assessor because she didn't like the colour assigned to it!) The list of documents seen so far is shown at top left, and the text of the current document (with highlighting of concept evidence) on the right. The bottom left panel includes the text of the current query as a heading but is otherwise unused. The description and narrative fields would be filled in if TREC ad hoc topics were being judged. A judgment for the current document is recorded by clicking on the appropriate icon at the far top right of the screen. Once the judgment has been made, the Next document button is enabled unless this is the last document for the topic, in which case the Next topic button is enabled. At any point within a topic, the judge may revisit previously seen documents (which will be displayed with all the concept evidence currently available) and, if desired, change their judgment. The RAT displays concept and criterion information throughout the judging process as an aid to maintaining consistency. However, the information was provided for assistance only and judges were free to make judgments regardless of whether criteria were actually satisfied or not 8 . As shown in Figure 2, occurrences of evidence were highlighted in each document by the RAT in the colour associated with the concept, enabling the judge to easily spot multi-coloured, potentially relevant sections in long documents. A hypothetical document containing the sentence, "Luigi's pasta restaurant at 2301 7th Avenue, Manhattan caters for both vegetarians and vegans." would thus show highlighting in all three colours. The judge could instantly home in on the sentence, and decide that the document is relevant. He or she might also decide to save the part-word "vegan" as additional evidence for the vegetarian concept. Over the 54 topics, the number of concepts defined per topic by the judges ranged from 1 to 8 (mean 3.7) and the average number of pieces of evidence defined (for all concepts combined) was 30.9. The order of presentation of documents for judging has been shown to affect the judgments made [6]. When we first used the RAT in TREC-5 (1996), we compared three different presentation orders: ascending length, descending length and random. We found that descending length was impractical because judges felt that they needed to read all of a long document in order to be sure that sections of it were not relevant. In the most extreme case, a judge spent more than six hours on a single document 9 . By contrast, judges found that working from the shortest documents first enabled them to accumulate sufficient evidence to be confident that when they reached a very long document, they could use scrolling to quickly locate the paragraphs which contained potentially relevant material with a low probability of missing relevant material. We concluded that presentation in ascending length order increased the probability that short relevant sections in long documents would be identified by maximising the highlightable evidence available. Accordingly, we have used this presentation order in all judging since 1996. In the Eisenberg and Barry study, the documents had already been assessed on a multi-level relevance scale. They found that when documents were presented in relevance order, either increasing or decreasing, subjects tended to "hedge" on the first documents seen by allocating scores closer to the mean than they would have otherwise. It seems very unlikely that this effect was present in our experiment as: 1. Hedging is not possible when there are only two possible relevance values. 2. Our judges were free to reassess previously judged documents if they felt that they had been too harsh or too lenient. 3. The association between document length and relevance is unlikely to be strong. The judge would have no reason to expect that the first document would or would not be relevant. 4. Order effects would only have confounded our comparative results if certain engines tended to disproportionately return very short or very long documents. 8 The RAT does have a mode in which relevance of documents is automatically determined using the concepts, criteria and evidence supplied by the judge. Any discrepancies between the automatic and the manual judgments result in an alert, which requires the judge to either accept the automatic judgment or to override it and record a reason. This mode was not used in the present experiments. 9 NIST judges working with documents from the same TREC collections judge several hundred documents per day on average. 3.6 Measures used Judgment of the first 20 live results allows calculation of a range of precision-oriented measures including precision at n 20 documents retrieved (P@n), mean reciprocal rank of first relevant document (MRR1) and TREC-style average precision (TSAP) 10 . Unfortunately, data analysis long after the data was collected revealed a bug in the scripts used to send queries and fetch result lists and documents. This meant that for some engines (nine altogether) where the queries contained a question-mark (29 out of 54), the script failed to move to the second page of ten results 11 meaning that some results lists were artificially truncated. For the engines affected by this bug, P@n (10 ! n 20) would certainly be underestimated. It is also likely that comparison of engines on precision at earlier cutoff (10, 9, .) would be unfair because dead links in the first ten results were replaced with live ones for the engines not subject to the bug. Available evidence suggests that most Web users generally look at very few results indeed. Silverstein et al [17, Table 7, p. 10] report that for a sample of over half a billion queries submitted to Alta Vista, 85.2% requested only a single result page. Accordingly, and to completely avoid the effects of the bug, our main analysis is based on the first seven live results. However, we acknowledge that precision at very early cutoff is likely to result in less stable system rankings (see Buckley and Voorhees [3]), and statistical significance is likely to be harder to achieve than for later cutoffs. For compatibility with Gordon and Pathak's study, we performed our principal analyses on P@1-5, the average of the precisions at cutoffs 1.5. We also present results for P@1, MRR1, and TSAP and plot P@n against n separately for the groups of engines affected and not affected by the bug. Unlike Gordon and Pathak, we did not measure recall, since the meaningfulness of the recall values depends heavily upon the accuracy of estimates of how many relevant documents there really are. Our main results are presented in Table 2 and in Figures 3 to 9. As shown in Figure 3 the range of values recorded for P@1-5 was quite large. A multiple analysis of variance (MANOVA 12 ) of the P@5 data, confirms that there is a significant difference in the performance of the search engines, F (19; In TREC, average precision for a topic is based only on precision values at each point in the ranking where a relevant document is retrieved. Average precision is obtained by dividing the sum of these values by the number of relevant documents in the collection. Here, it is unknown how many relevant documents there really are and rankings are cut off at 7 rather than 1000. Accordingly, here the sum is divided in each case by 7 and the result is called TREC-style average precision at cutoff 7. TSAP is an average of precisions and unlike the true TREC measure does not include a recall component. A system for which all 7 documents retrieved for a query were judged relevant would score 1.0 on this measure; One for which only one retrieved document was relevant would score between 1=7 and 1=49, depending upon the position of the relevant document in the ranking. Note that although for each individual topic TSAP scores are a constant multiple of what the TREC average precision would have been at the same cutoff, the multiplier varies from topic to topic. Consequently, across-topic averages of the two measures are not simply related. 11 For some engines we requested a generous number of results in a single page. For other engines we didn't know the necessary URL syntax to do this and requested results page by page. 12 In this MANOVA, the repeated factor was Queries, the independent group factor was Search Engines and the dependent measure was P@1-5. A multiple rather than a univariate model was selected as the latter is sensitive to violations in sphericity which are common in repeated measures designs. NorthernLight Inquirus Google InfoSeek AltaVista Microsoft LookSmart MetaCrawler HotBot EuroFerret Lycos Snap Fast Excite DirectHit Yahoo Exciteaus WebWombat ANZwers EuroSeek Search engines as at 20 Sep 19990.20.40.6 Srch. Eng. Types General Metasearcher European Australia/NZ Figure 3: The 20 public search engines compared on the basis of precision averaged across cutoffs 1-5 (P@1- 5). The type of search engine (general, metasearch, or regional) is color-coded according to the key at top right. Each horizontal black line gives the corresponding performance for a 1999 TREC system. See Section 4.2 and Table 4 for details. Table 2: Summary of results for the twenty search engines. These are based on the first seven live results. TSAP means TREC-style average precision, MRR1 is the mean reciprocal rank of first relevant document and P@n means precision at n documents retrieved. P@1-5 is the average of precision values at each ranking cutoff from 1 to 5. Run TSAP MRR1 P@1 P@5 P@1-5 ANZwers 0.1111 0.2673 0.1852 0.1667 0.1787 AltaVista 0.3714 0.582 0.5 0.4519 0.4691 EuroFerret 0.284 0.5553 0.4444 0.3704 0.4043 Excite Exciteaus Fast Google 0.3939 0.6133 0.463 0.4889 0.4848 HotBot 0.3064 0.5705 0.463 0.4111 0.4196 InfoSeek Inquirus 0.404 0.5833 0.463 0.4963 0.4965 LookSmart MetaCrawler 0.3075 0.5764 0.4259 0.4222 0.4329 WebWombat Northern Light is the top ranked engine on the basis of P@1-5 but differences between it and the next nine engines are not statistically significant. Differences between NorthernLight and the engines below EuroFerret (Lycos, SNAP, Fast, Excite, DirectHit, etc.) are all significant, All 190 pairwise correlations between the engines were positive, Pearson r values ranging from 0.066 to 0.885. Of these, 155 of the correlations achieved significance (p ! 0:05). Four pairs exhibited correlations in excess of 0.7. They were (Looksmart, AltaVista), (Microsoft, AltaVista), (SNAP, Hotbot) and (Microsoft, Looksmart). It is understood that at the time of the evaluation, that SNAP and Hotbot both used Inktomi search technology and that there were commercial relationships between (Microsoft, Looksmart), (Microsoft, AltaVista) and (AltaVista, Looksmart). Table 2 shows that the search engine rated by Lawrence and Giles as having the largest coverage (NorthernLight) also scored highest on all of the measures except TSAP. However, the FAST engine which claimed a coverage nearly 50% larger than NorthernLight, performed relatively poorly. Figure 4 shows a scatter plot of coverage against P@5 for the Lawrence and Giles engines. The correlation between P@5 and coverage is not statistically significant. The two metasearch engines performed creditably, but not as well as the best individual engines. A metasearcher works by broadcasting incoming queries to a large number (11 and 14 in these two examples) of primary search engines and merging their results lists. Metasearchers therefore achieve a very high effective coverage. Not surprisingly, given the general nature of the test queries, the unrestricted-domain search engines generally achieved better results than those restricted to European and Australasian domains Figure 5 shows the mean reciprocal rank of first relevant document (counted as zero if no relevant document appeared in the first 7 results). If the first relevant document appears in the second position, the reciprocal rank is 0.5. Figure 6 compares performance on the TSAP measure(explained in Section 3.6). Figure 7 compares performance on the basis of P@1, which is equivalent to the probability that the first result is relevant. A nonparametric test revealed a highly significant difference in the performance of the engines on this measure, Cochran There are many more tied scores on P@1 than on P@20, reflecting quantisation due to the fact that there are only 54 possible scores in the former case and 1080 in the latter. Note that Euroseek did not find a single relevant document at rank 1 and scored zero on this measure. Figure 8 documents the variation of P@n as n varies from 1 to 20 for the search engines not affected by the result changeover bug. Figure 9 shows the variation in P@n for the remaining enginges over the range of n which was not affected by the bug. Ignoring fluctuations at small exhibited a very gradual decline in P@n across the range studied. This general pattern was also found for all of the engines shown in the corresponding Gordon and Pathak figure Figure 1, p. 156), except OpenText (not studied here.) DirectHit and Yahoo in the present study were also exceptions. For them, P@n declined more sharply. 4.1 Inter-correlation of measures Table 3 shows the Pearson r intercorrelations between four different measures, based on the top 7 results. All are significantly inter-correlated (p ! 0:01). Similar, and in some cases even higher, intercorrelations were observed for the combined TREC and search engine groups. Considering only the eleven engines not affected by the result-page bug, there is a correlation of 0.965 (p ! 0:01) between P@5 and P@20 and of 0.868 between P@1 and P@20 (p ! 0:01) 4.2 How do search engines compare with 1999 TREC systems? The performance of groups participating in the TREC-8 Large Web Task is reported in detail in [11] and summarised in Table 4. The queries judged in the Large Web Task were a slightly larger superset of the queries used in the present study. TREC results reported here have been re-analysed to take into account only the data for the 54 queries judged for both the TREC systems and the search engines. These TREC results are also represented as horizontal black lines in Figure 3, 5 and 7. Instead of crawling their own data from the current web, these runs used the first 18.5 million pages (100 gigabytes) of a crawling run carried out in early 1997. Each TREC run thus simulates the operation of a search engine with small and out-of-date coverage (comparable to the size estimated by Lawrence and Giles for Lycos). Documents for the TREC systems were judged in a separate batch from the search engine documents but the judges assigned to each query were the same, the presentation of documents was identical and all other conditions were held constant. The TREC batch was judged first, with judging of the search engine batch following immediately afterward. Concepts and evidence accumulated with the Relevance Assessment Tool (RAT, see Section 3.5) during the TREC judging was carried over to the search engine batch. The number judged was 11,654 of which 25.8% (3008) 13 The Cochran Q test is a non-parametric test for determining if a series of related samples differ in circumstances where data is based on a nominal or dichotomised ordinal scale of measurement. The P@1 data was of the latter type as P@1 scores for individual queries must be either 1 or 0. A related samples test was appropriate because each query was sent to each of the 20 engines. coverage (% of Web) Figure 4: The relationship between coverage (percentage of estimated publicly indexable Web, as reported by Lawrence and Giles) and P@5, for the eleven Lawrence and Giles engines plus FAST, whose claimed size was known at the time of the experiments. The Pearson R coefficient of correlation was 0.370, which is not significant at the 0.05 level(two-tailed.) Table 3: Intercorrelation of various measures for the twenty search engines. TSAP means TREC-style average precision, MRR1 is the mean reciprocal rank of first relevant document and P@n means precision at n documents retrieved. Measure TSAP MRR1 P@1 P@5 NorthernLight Google Microsoft InfoSeek Inquirus AltaVista MetaCrawler HotBot LookSmart EuroFerret Lycos Snap Excite Fast Exciteaus DirectHit Yahoo WebWombat ANZwers EuroSeek mean recip. rank of first rel. page Search engines as at 20 Sep 19990.20.40.6Srch. Eng. Types General Metasearcher European Australia/NZ Figure 5: The 20 public search engines compared on the basis of mean reciprocal rank of the first relevant page found. Engines scored zero for a query if no relevant document was found in the first 7 results. Colors identify the type of search engine. Each horizontal black line gives the corresponding performance for a 1999 TREC system. Table 4: Summary of results for the participating groups in the TREC-8 Large Web task. The results presented are the averages of all the runs submitted by each group. TSAP means TREC-style average precision, MRR1 is the mean reciprocal rank of first relevant document and P@n means precision at n documents retrieved. Run TSAP MRR1 P@1 P@5 P@1-5 UMass 0.3927 0.5698 0.463 0.4741 0.4741 ACSys AT&T Research 0.507 0.7074 0.6111 0.6074 0.6178 Microsoft Research 0.4561 0.6528 0.5741 0.5333 0.5449 MSR/City Univ. 0.457 0.6582 0.5926 0.5259 0.5512 Inquirus Google NorthernLight AltaVista InfoSeek LookSmart Microsoft MetaCrawler HotBot Fast Snap Lycos EuroFerret Excite DirectHit Exciteaus Yahoo WebWombat ANZwers EuroSeek TREC-Style Average Precision Search engines as at 20 Sep 19990.20.4Srch. Eng. Types General Metasearcher European Australia/NZ Figure The 20 public search engines compared on the basis of TREC-style average precision, based on top 7 rankings. Colors identify the type of search engine. Each horizontal black line gives the corresponding performance for a 1999 TREC system. NorthernLight AltaVista Microsoft Inquirus HotBot LookSmart Google EuroFerret InfoSeek MetaCrawler DirectHit Fast Excite Lycos Exciteaus Yahoo Snap WebWombat ANZwers EuroSeek Probability that first result is relevant Search engines as at 20 Sep 19990.20.40.60.8 TREC systems General Metasearcher European Australia/NZ Figure 7: The 20 public search engines compared on the basis of P@1. Colors identify the type of search engine. Each horizontal black line gives the corresponding performance for a 1999 TREC P@n Inquirus Google EuroFerret Metacrawler DirectHit Excite Lycos ANZwers Figure 8: The variation of P@n with n for the engines not affected by the bug (ranked on P@1). were judged relevant. The number relevant per query ranged from 1 - 136, with a mean of 55.7. The maximum achievable mean P@5, based on total known relevant pages, was 0.9259. The group of TREC systems was compared with the group of search engines using separate Mann-Whitney U 14 tests for each of the following measures averaged across queries: P@1, P@5, P@1-5, MRR1 and TSAP. On each measure the TREC systems are significantly better (p !! 0:05 in all cases). 4.3 How do the present results relate to those of Gordon and Pathak? Figure compares P@1-5 scores (lenient cutoff) obtained by Gordon and Pathak with the corresponding scores for the same engines obtained in the present study. As shown the results are highly correlated. In Figure a line of best fit which passes through the origin has a slope of about 1.5. In other words, our P@1-5 scores tended to be 50% higher than those observed by Gordon and Pathak (based on their lenient encoding), despite the advantage presumably conferred by use of highly tuned queries and advanced query syntax. This may be explained either by a systematic difference in topic difficulty, or by a difference in judging standards. Our main findings are as follows: 1. There was a high correlation (0.89) between the P@1-5 scores from this study and those reported by Gordon and Pathak for their experiment conducted a year earlier than ours. This was surprising because we expected that updating of indexes and improvements to algorithms would have caused much more volatility. Furthermore, there were considerable differences in the way the two studies were conducted. 2. We found no significant correlation between index coverage and early precision (P@5). (Figure 4.) Provided that coverage is sufficient to include enough relevant documents for the chosen topics, the ranking algorithm seems to be the determining factor. Coverage would become more critical if the queries had very few relevant answers or if the task were different (for example, if it involved locating obscure known items, or locating ALL documents matching a specification). 3. In the present experiment, TREC Large Web systems as a group outperformed the search engines. However, the performance of the best search engines approached (and in one case surpassed) the median performance of the participants in the TREC-8 Large Web Task. Figures 5, 6 and 7.) This suggests that some search engines are indeed using reasonably good retrieval methods. However, it should be noted that the Large Web collection was only a small fraction of the size of the data indexed by NorthernLight and Google. Previous work [10] shows that precision at fixed cutoff tends to increase as collection size increases. Furthermore, TREC participants almost certainly sacrificed some of the effectiveness of their methods in order to achieve speed of processing and are therefore not fully representative of TREC state-of-the-art. There is still room for improvement. The VLC2 content was nearly three years out of date with respect to the Web at the time of our experiments. More importantly, we believe it was out-of-date with respect to the query 14 The Mann-Whitney is a non-parametric test of the difference between two independent groups. A non-parametric test was selected because of the small sample size of the TREC group. P@n Alta Vista HotBot Looksmart Infoseek Fast ExciteAus Yahoo! WebWombat EuroSeek Figure 9: The variation of P@n with n for the search engines which were affected by the bug (ranked on (present study) P@1-5 (Gordon and Pathak) Figure 10: The 6 public search engines included in both the Gordon and Pathak study and the present experiment. The correlation coefficient of the scores is 0.89 which is significant at the 0.05 level (two-tailed). set as well. It is not known precisely when the queries employed were originally submitted but they were obtained from the search engine companies in late 1998 and early 1999. However, none of the queries seem to relate to subjects which would have been well covered on the Web in 1997 but not in 1999. Any slight bias due to the age of the VLC2 crawl is likely to have operated in favour of the search engines rather than against them. 4. For the great majority of engines, P@n declined very slowly with increasing n. (Figures 8 and 9.) 5. A very high intercorrelation of measures was observed. (Section 4.1.) P@20 for the engines with complete result lists is very strongly predicted by P@5 and almost as strongly predicted by P@1. 6. Despite very large effective coverage, the metasearchers did not outperform the best individual engines. We have calculated that a hypothetical metasearcher which broadcast queries to all 20 engines in this study could achieve a P@5 score of 0.9741 on the test queries if it had used an ideal merging algorithm 15 . Its precision would be less than perfect because of the 3 queries for which the total number of relevant pages returned by all engines was less than 5. The large gap between actual and potential performance of metasearchers suggests a possibly fruitful avenue for the improvement of search engine quality. Given that some metasearchers (such as Inquirus) perform merging by downloading all the result documents and reranking on the basis of their content, it may be possible to find merging methods which are capable of substantially narrowing the gap. 5.1 Reliability of results How subject are the above search engine evaluations to the vagaries of human judgment? It is known that agreement between human relevance judges is less than perfect. Voorhees and Harman [20, p.8] reported a 71.7% rate of unanimous agreement between three assessors over 14,968 documents. However, as previously noted, Voorhees [21] found that, while substituting relevance judgments made by different human assessors changed the magnitude of scores, it had almost no effect on the rank order of systems. 5.2 Representativeness of queries Three potential biases affected our method for selecting queries (Section 3.3). The 200,000 queries constituting the initial pool were taken from the query logs of two different search engines. They may have been biased by user perceptions of the characteristics of the two particular engines, by the client demographics of those engines, and by the fact that the queries were in both cases submitted to a "natural language" interface. We introduced further biases by excluding adult-content queries and by allowing a single judge to select queries which she felt that she and her colleagues would be able to judge. To what extent are these biases compatible with desirable features of search engine evaluation, particularly our own modification and extension (Figure 11) of Gordon and Pathak's list? Against Feature 1, we can be reasonably confident that most of the judged queries were submitted as a result of some real information need. Apparently flippant queries such as, "who are you?", and empty or incomplete queries are occasionally found in the logs but no obvious examples were accepted for An ideal merging algorithm is one capable of producing a merged list in which all relevant documents in the primary results lists appear first. judging. Some queries may have been submitted for testing purposes (as in our own study) but the proportion is almost certainly negligible. Feature 4 suggests that queries should be representative of the full range of information needs over which it is desired to draw conclusions, both with respect to subject and to type of results wanted. As can be seen from Figure 1, the queries we employed covered a very broad range of subjects, including politics, history, psychology, music, medicine, religion, humour, technology, geography, science and finance. However, all our queries appear to be at the "general-knowledge" level. Our results may not, in general, accurately predict the performance of search engines when queries are posed and evaluated by subject experts, despite the fact that they are correlated with the Gordon and Pathak results for expert queries. Despite the selection biases identified above, we are fairly confident that our results would predict search engine performance reasonably well for general, non-adult-content queries, provided the same evaluation methodology were employed. In our view, the major limitation of our study (and most of its predecessors) lies in our failure to represent the range of result types wanted and to apply appropriate evaluation for the different types. Performance on a "find a range of relevant pages" task may not predict at all well the performance on a "find a homepage" task. This issue is discussed in detail below. 5.3 Effect of spelling errors As noted earlier, some of the test queries contained spelling errors. These errors were not corrected for use in TREC-8 Large Web. This was something of a departure from convention for TREC participants because TREC ad hoc topics are almost free of such errors. Judges looked for documents satisfying the apparent intent of the query, rather than literal matches. Accordingly, systems able to detect and correct spelling errors may have been at a small advantage relative to others. We have no information about which systems (if any) employed spelling correction. 6 Future Web search evaluation Unfortunately, in the absence of a standardised Web test collection with complete relevance judgments and a willingness of search engine companies to use it, there is little prospect of reproducible Web search engine evaluation. Nonetheless, search engine effectiveness evaluation is necessary. Future Web search evaluations should adhere to sound experimental methods, such as blind judging and statistical significance testing while improving on the methodology adopted by the present study and its predecessors. There are a number of ways in which future Web search evaluations could more accurately model Web search. 1. Future evaluations should recognise that retrieval corresponding to different types of information need require different evaluation techniques. Searcher information needs may be broadly classified on the basis of the type of answer expected: A. A very short (eg. single sentence) answer to a question. For example, "The melting point of lead is 210C." Evaluation techniques for this type of information need were considered in the TREC-8 Question and Answer Track. [19, various papers] B. A single document. For example, the home page for the Information Retrieval journal, or the current timetable for the Washington Metro Red Line, or a directory of accommodation in Athens, Greece. This type subsumes traditional known-item search. However, Web searchers often deduce that a particular document, such as a company homepage or a bus timetable, ought to exist without having previously seen it, a situation more accurately described as suspected-item search. Sometimes, no single document matching the information need exists and the searcher must synthesize the required information from a group of documents. On the other hand, there may be several matching documents. For example, there may unexpectedly be several homepages for the same individual. C. A selection of documents. These documents might be sought for research purposes (for example a range of documents relating to US policy on North Korea) or in order to access on-line services (for example on-line auction sites or sites which allow download of recorded music.) D. Every document matching a criterion. For example, every document which discusses orbital engines. Evaluation against this type of need is obviously related to coverage, but also to the sophistication of the engine in expanding the query so as to find documents which do not literally match the query but do match one which is semantically equivalent, for example "Sarich engine", "motor developed by Ralph Sarich" etc. Another complication arises from the expression of needs in terms of explicit or implicit document metadata, such as "all documents authored by Doug Engelbart". Blair and Maron [2] discuss retrieval effectiveness in a high-recall legal application. In each of the first two types, evaluation requires access to only very few documents. In Type B, relevance, even strong relevance, is usually not sufficient and evaluation must be based on asessment of whether a document is the one sought. In these cases, performance can be evaluated on the basis of the position of the desired document in the ranking and it makes little sense to talk of "relevance". To illustrate this point, the Australian National University intranet contains thousands of documents which are relevant to the topic of libraries and one of the most common queries submitted to its search engine is library. In the cases where the submitter of this query was looking for the library homepage or the library catalogue, the ability of the engine to retrieve pages which merely talk about libraries is unlikely to correlate well with searcher satisfaction. In Type C, evaluation may be limited to the first 10 or 20 results, as most users would expect to find the information they require in fewer than this number of documents. However, it is important that the documents presented in a result list should not repeat the same material and evaluation methodology for this type of information need should take this into account. Evaluation of recall-oriented 16 information needs (Type D) is difficult, given limited assessment resources. It must inevitably rely on sampling techniques. Classification of information needs is notoriously difficult. The above four-way classification illustrates the need for a range of evaluation paradigms, but is still a gross over-simplification. Information needs are highly dynamic and are conditioned by preliminary results. For ex- ample, someone searching for a complete timetable may find that no such single document can be found and be forced to seek multiple individual timetables which together provide the needed information. A Type B search for a single homepage may or may not transform into a Type D search when it becomes clear that multiple home pages exist. Note that high precision is also required. Within any of the above types there may be a need for a range of judging instructions and a need for a range of different measures. In the case of Type C(reseach) needs it is probably preferable to use multi-valued relevance judgments, as Gordon and Pathak did, as it may be possible to exploit the additional relevance information. We note, however, that Gordon and Pathak do not report any results relying on the distinction between irrelevant and highly irrelevant. They also convert four-value relevance judgments into binary in order to compute their measures. By contrast, in the case of Type C(services) needs, the judgment to be made is not whether the page is relevant or not but rather whether the page allows direct access to a service of the right type. For example, whether the page allows downloading of MP3 files, as opposed to discussing the topic of downloading MP3 files. A defect of the present study is that we treated all queries as though they corresponded to Type C(reseach) information needs. The problem is not that we may have misinterpreted the intention of the searchers who submitted these particular queries. What is important is that the ability of search engines to process queries of other types was not evaluated. In the Web context, it may be desirable to accord partial scores to retrieved documents which contain no useful content but which link to it. A number of issues in indirect relevance are canvassed in the discussion of the TREC-8 Small Web Task [11]. 2. Evaluations should be carried out at frequent intervals to maintain an up-to-date picture of relative performance and to smooth out performance "glitches." 3. The cost of evaluations based on Type C queries is largely attributable to assessor salaries. Consequently, it is desirable to maximize the value of assessments made. Potentially, the high correlation between different measures, might be used to limit how many pages per topic need be judged. There may be greater value in judging fewer documents per topic for a larger number of topics, particularly as measures derived from the very top of the ranking may be the best predictors of user satisfaction for certain information need types. 4. In comparing performance from one evaluation to the next, a normalising baseline is needed. The median performance of a large number of engines could be used in this role but would be unable to detect a general increase or decline in performance. Use of a standard set of queries in each evaluation does not solve this problem because of variations in a human's judgments over time and because of the possibility that search engine companies would tune their results for these queries. The only possibility we have been able to identify involves including a known, unchanging engine, operated by the evaluators in each successive evaluation and using its (rejudged) performance as a normalising baseline. 5. The present study did not penalise search engines for returning defunct pages in the results list to enable fair comparison with the TREC systems. Future evaluations which do not compare against TREC should treat a dead link as a worthless answer. However, it may be considered necessary to make several attempts to access the page, in case inaccessibility is only temporary. 6. Considerable care is needed in writing scripts for automatically posing queries to search engines and retrieving results. The format of result pages changes from time to time and may cause errors. Thorough sanity checks should be applied to all results obtained by such automatic means. 1. Search topics should be representative of genuine user need. 2. If a search intermediary or third-party relevance judge is employed, the primary searcher's information need should be as fully captured as possible and transmitted in full. 3. A large number of search topics should be used. 4. If the purpose of the study is to enable informed consumer choice, the study should aim to include most major search engines. 5. If the study seeks to compare maximal as opposed to typical effectiveness of search engines, it is appropriate to exploit the full range of features of each engine and to do so may require the use of skilled search intermediaries. 6. The search topics should represent the full range of information needs over which it is desired to draw conclusions, both with respect to subject and to type of results wanted. (See type categories in Section 6.) 7. Result judging should be appropriate to the type of result sought. (See Section 6.) 8. Judging should be consistent within a topic. All judgments for a topic should be made by the same person and within a short period of time. 9. Judging should be blind and be conducted by the person with the need for information or by independent judges. 10. Documents should be presented for judging in the same way that they would have been seen by a real Web searcher, ie. rendered by a browser. Images should be viewable and it should be possible to follow links. However, careful instruction may be necessary to ensure that judges remain on task and do not waste time on fruitless link following. 11. For general comparisons of search engines, dead links should count as useless answers. 12. Experiments should be well designed and conducted. Figure 11: Summary of desirable features of future Web search evaluations. Based on Gordon and Pathak's list, extended and modified. 7. In presenting documents for assessment, it is important to do so in a way which simulates normal Web search. See point 10 in Figure 11. Figure 11 summarises the features which we believe should be present in future Web search evaluations. 7 Conclusions and Future Work We have performed a Web search engine evaluation using appropriate scientific methodology on a larger scale than previously published studies. This evaluation shows that choice of search engine makes a difference, that early precision (P@5) is not significantly correlated with index coverage, and that the best search engines are now using algorithms approaching the effectiveness (on relevance based tasks) of those used by TREC-8 Large Web task participants. We found an unexpectedly high correlation between P@1-5 results obtained by us and those obtained by a study conducted by others a year earlier. There were very high inter-correlations between four different measures derivable from the top 7 live results. Further work is needed to refine the classification of information need types in Web search and to identify the appropriate evaluation paradigms for each type. User studies to determine the effectiveness measures which best correlate with searcher satisfaction ratings would be very valuable. User studies will also be required to determine the proportion of Web queries which are derived from each information need type, as inspection of query logs frequently fails to reveal the searchers' intentions. In order to present a complete picture of search engine performance, future studies should include queries derived from range of commonly expressed information need types and use appropriate judgments and measures for each type. To reduce the multi-dimensional results from such a study to a single index of effectiveness will require that each dimension be weighted by frequency of occurrence of the corresponding information need type. --R Alta Vista web page. An evaluation of retrieval effectiveness for a full-text document-retrieval system Evaluating evaluation measure stability. The Cranfield tests on index language devices. Comparative study of web search service performance. Order effects: A study of the possible influence of presentation order on user judgments of document relevance. Electric Knowledge LLC. Finding information on the world wide web: The retrieval effectiveness of search engines. Overview of TREC-6 Very Large Collection Track Scaling up the TREC Collection. Overview of TREC- 8 Web Track The web robots pages. Accessibility of information on the web. First 20 precision among world wide web search services (search engines). Computer evaluation of indexing and text processing. Analysis of a very large web search engine query log. Overview of the fifth Text Retrieval Conference (TREC- 5) Variations in relevance judgments and the measurement of retrieval effec- tiveness --TR --CTR Eric C. Jensen , Steven M. Beitzel , Ophir Frieder , Abdur Chowdhury, A framework for determining necessary query set sizes to evaluate web search effectiveness, Special interest tracks and posters of the 14th international conference on World Wide Web, May 10-14, 2005, Chiba, Japan Steven M. Beitzel , Eric C. Jensen , Abdur Chowdhury , David Grossman , Ophir Frieder, Using manually-built web directories for automatic evaluation of known-item retrieval, Proceedings of the 26th annual international ACM SIGIR conference on Research and development in informaion retrieval, July 28-August 01, 2003, Toronto, Canada Mathias Gry, Non-linear reading for a structured web indexation, Proceedings of the 25th annual international ACM SIGIR conference on Research and development in information retrieval, August 11-15, 2002, Tampere, Finland Fazli Can , Rabia Nuray , Ayisigi B. Sevdik, Automatic performance evaluation of web search engines, Information Processing and Management: an International Journal, v.40 n.3, p.495-514, January 2004 Judit Bar-Ilan, Comparing rankings of search results on the web, Information Processing and Management: an International Journal, v.41 n.6, p.1511-1519, December 2005 Liwen Vaughan , Mike Thelwall, Search engine coverage bias: evidence and possible causes, Information Processing and Management: an International Journal, v.40 n.4, p.693-707, May 2004 Yves Rasolofo , David Hawking , Jacques Savoy, Result merging strategies for a current news metasearcher, Information Processing and Management: an International Journal, v.39 n.4, p.581-609, July King-Lup Liu , Weiyi Meng , Jing Qiu , Clement Yu , Vijay Raghavan , Zonghuan Wu , Yiyao Lu , Hai He , Hongkun Zhao, AllInOneNews: development and evaluation of a large-scale news metasearch engine, Proceedings of the 2007 ACM SIGMOD international conference on Management of data, June 11-14, 2007, Beijing, China Paul Thomas , David Hawking, Evaluation by comparing result sets in context, Proceedings of the 15th ACM international conference on Information and knowledge management, November 06-11, 2006, Arlington, Virginia, USA Ayman Farahat , Geoff Nunberg , Francine Chen, AuGEAS: authoritativeness grading, estimation, and sorting, Proceedings of the eleventh international conference on Information and knowledge management, November 04-09, 2002, McLean, Virginia, USA Liwen Vaughan, New measurements for search engine evaluation proposed and tested, Information Processing and Management: an International Journal, v.40 n.4, p.677-691, May 2004 Peter Bailey , Nick Craswell , David Hawking, Engineering a multi-purpose test collection for web retrieval experiments, Information Processing and Management: an International Journal, v.39 n.6, p.853-871, November Thanh Tin Tang , Nick Craswell , David Hawking , Kathy Griffiths , Helen Christensen, Quality and relevance of domain-specific search: A case study in mental health, Information Retrieval, v.9 n.2, p.207-225, March 2006 Ziv Bar-Yossef , Maxim Gurevich, Random sampling from a search engine's index, Proceedings of the 15th international conference on World Wide Web, May 23-26, 2006, Edinburgh, Scotland Steven M. Beitzel , Eric C. Jensen , Abdur Chowdhury , David Grossman, Using titles and category names from editor-driven taxonomies for automatic evaluation, Proceedings of the twelfth international conference on Information and knowledge management, November 03-08, 2003, New Orleans, LA, USA Michael P. Evans , Richard Newman , Timothy Putnam , Diana J. M. Griffiths, Search Adaptations and the Challenges of the Web, IEEE Internet Computing, v.9 n.3, p.19-26, May 2005 Theodora Tsikrika , Mounia Lalmas, Combining evidence for Web retrieval using the inference network model: an experimental study, Information Processing and Management: an International Journal, v.40 n.5, p.751-772, September 2004 Judit Bar-Ilan , Mazlita Mat-Hassan , Mark Levene, Methods for comparing rankings of search engine results, Computer Networks: The International Journal of Computer and Telecommunications Networking, v.50 n.10, p.1448-1463, 14 July 2006 Caroline M. Eastman , Bernard J. Jansen, Coverage, relevance, and ranking: The impact of query operators on Web search engine results, ACM Transactions on Information Systems (TOIS), v.21 n.4, p.383-411, October Jaroslaw Baliski , Czeslaw Daniowicz, Re-ranking method based on inter-document distances, Information Processing and Management: an International Journal, v.41 n.4, p.759-775, July 2005 Bernard J. Jansen , Paulo R. Molina, The effectiveness of web search engines for retrieving relevant ecommerce links, Information Processing and Management: an International Journal, v.42 n.4, p.1075-1098, July 2006 Robert M. Losee , Lewis Church Jr., Information Retrieval with Distributed Databases: Analytic Models of Performance, IEEE Transactions on Parallel and Distributed Systems, v.15 n.1, p.18-27, January 2004 I. S. Nekrestyanov , N. V. Panteleeva, Text Retrieval Systems for the Web, Programming and Computing Software, v.28 n.4, p.207-225, July-August 2002
evaluation;search engines;web search
594035
Some Formal Analysis of Rocchio''s Similarity-Based Relevance Feedback Algorithm.
Rocchio's similarity-based Relevance feedback algorithm, one of the most important query reformation methods in information retrieval, is essentially an adaptive supervised learning algorithm from examples. In spite of its popularity in various applications there is little rigorous analysis of its learning complexity in literature. In this paper we show that in the binary vector space model, if the initial query vector is 0, then for any of the four typical similarities (inner product, dice coefficient, cosine coefficient, and Jaccard coefficient), Rocchio's similarity-based relevance feedback algorithm makes at least i>n mistakes when used to search for a collection of documents represented by a monotone disjunction of at most i>k relevant features (or terms) over the i>n-dimensional binary vector space &lcub;0, 1&rcub;i>n. When an arbitrary initial query vector in &lcub;0, 1&rcub;i>n is used, it makes at least (i>n mistakes to search for the same collection of documents. The linear lower bounds are independent of the choices of the threshold and coefficients that the algorithm may use in updating its query vector and making its classification.
Introduction Research on relevance feedback in information retrieval has a long history (Baeza-Yates and Ribeiro-Neto, 1999; Frakes and Baeza-Yates, 1992; Ide, 1971b; Ide, 1971a; Raghavan and Wong, 1986; J.J. Rocchio, Salton, 1989). It is regarded as the most popular query reformation strategy (Baeza-Yates and Ribeiro-Neto, 1999). The central idea of relevance feedback is to improve search performance for a particular query by modifying the query step by step, based on the user's judgments of the relevance or irrelevance of some of the documents retrieved. In the vector space model (Salton, 1989; Salton et al., 1975), The extended abstract of this paper was published in Proceedings of the Eleventh International Symposium on Algorithms and Computation (ISAAC'00), Lecture Notes in Computer Science 1969, pages 108-119, Springer-Verlag, December, 2000. c 2001 Kluwer Academic Publishers. Printed in the Netherlands. Z. Chen and B. Zhu both documents and queries are represented as vectors in a discretized vector space. In this case, relevance feedback is essentially an adaptive supervised learning algorithm: A query vector and a similarity are used to classify documents as relevant and irrelevant; the user's judgments of the relevance or irrelevance of some the classied documents are used as examples for updating the query vector as a linear combination of the initial query vector and the examples judged by the user. Especially, when the inner product similarity is used, relevance feedback is just a Perceptron-like learning algorithm (Lewis, 1991). It is known (J.J. Roc- chio, 1971) that there is an optimal way for updating the query vector if the sets of relevant and irrelevant documents are known. Practically it is impossible to derive the optimal query vector, because the full sets of the relevant and irrelevant documents are not available. There are many dierent variants of relevance feedback in information retrieval. However, in this paper we only study Rocchio's similarity-based relevance feedback algorithm (J.J. Rocchio, 1971; Ide, 1971a; Salton, 1989). In spite of its popularity in various applications, there is little rigorous analysis of its complexity as a learning algorithm in literature. This is the main motivation for us to investigate the learning complexity of Rocchio's similarity-based relevance feedback algorithm. Wong, Yao and Bollmann (Wong et al., 1988) studied the linear structure in information retrieval. They designed a very nice gradient descent procedure to compute the coe-cients of a linear function and analyzed its performance. In order to update the query vector adaptively, their gradient descent procedure must know the user preference which is in practice the unknown target to be learned by an information retrieval system. More discussions of their gradient descent procedure will be given in section 5.1. The main contribution of our work in this paper is that linear lower bounds on classication mistakes are proved for the algorithm when any of the four typical similarities (inner product, dice coe-cient, cosine coe-cient, and Jaccard coe-cient) listed in (Salton, 1989) is used. Technically, our work in this paper is enlightened by the work in (Kivinen et al., 1997) on lower bounds of the Perceptron algorithm. Precisely, we borrow the method developed in (Kivinen et al., 1997) for constructing an example sequence with pairwise constant inner prod- ucts. We extend the method to cope with other similarities besides inner product. We also design a new method for selecting trial sequences and prove in a uniform way our lower bounds for Rocchio's similarity-based relevance feedback algorithm. It should be pointed out that the lower bounds established in this paper for Rocchio's similarity-based relevance feedback algorithm is based on the following worst case considerations: The user acts as an Some Formal Analysis of Rocchio's Algorithm 3 adversary to the algorithm; the algorithm is required to precisely search for the collection of all documents relevant to the given search query; and the algorithm is allowed to receive one document example judged by the user as relevance or irrelevant at each step 1 . In practical applications , in contrast to the above worst case considerations, the user in general may not act as an adversary to the algorithm; the algorithm is usually required to search for a short list of top ranked documents relevant to the given search query; and at each step of the similarity-based relevance algorithm, the user may judge a few documents as relevance feedback to the algorithm. In other words, the appropriate situation in real-world information retrieval applications would be a kind of \sympathetic oracle" model, where the user is not an adversary to the information retrieval system but a \sympathetic judge" who provides the most useful possible information in order to help the system help him/her to accomplish his/her work. Hence, our lower bounds proved in this paper for Rocchio's similarity-based relevance feedback algorithm may not aect the algorithm's eective applicability to the real-world problems despite of their theoretical signicance. The formal analysis of the algorithm helps us understand the nature of the algorithm well so that we may nd new strategies to improve its eectiveness or design new algorithms for information retrieval. Recently, we have made some progress along this line in (Chen, 2001). We have designed two types of multiplicative adaptive algorithms for user preference retrieval with provable better performance: One has better performance than Rocchio's algorithm in learning a class of linear classiers over non-binary vector space. The other boosts the usefulness of an index term exponentially, while the gradient descent procedure in (Wong et al., 1988) boosts the usefulness of an index term linearly. We refer the readers to the work of (Salton and Buckley, 1990; Lewis, 1991) for discussions of the Perceptron-like learning nature of the similarity-based relevance feedback algorithm. It was stated in (Lewis, 1991) that the most important future direction for research in information retrieval is likely to be machine learning techniques which can combine empirical learning with the use of knowledge bases. This paper is organized as follows. In section 2, we give a formal presentation of Rocchio's similarity-based relevance feedback algorithm. In section 3, we prove several technical lemmas based on the techniques developed in (Kivinen et al., 1997). In section 4, we prove linear lower bounds for Rocchio's similarity-based relevance feedback algorithm with any of the four typical similarities listed in (Salton, 1989). This last restriction is not critical to the proof of the lower bounds, but it would make the analysis easier. 4 Z. Chen and B. Zhu In section 5 we give some discussions of our result and also show that Rocchio's algorithm can be used to learn many other target document classes. We conclude the paper and list several open problems in section 6. 2. Rocchio's Similarity-Based Relevance Feedback Algorithm Let R be the set of all real values, and let R + be the set of all non-negative real values. Let n be a positive integer. In the binary vector space model in information retrieval (Salton, 1989; Salton et al., 1975), a collection of n features or terms T are used to represent documents and queries. Each document d is represented as a vector such that for any i, 1 i n, the i-th component of v d is one if the i-th feature T i appears in d or zero otherwise. Each query q is represented by a vector v any i, 1 i n, the i-th component of v q 2 R is a real value used to determine the relevance (or weight) of the i-th feature T i . Because of the unique vector representations of documents and queries, for convenience we simply use d and q to stand for their vector representations v d and v q , respectively. A similarity in general is a function m from R n R n to R + . A similarity m is used to determine the relevance closeness of documents to the search query and to rank documents according to such close- ness. In the binary vector space model of information retrieval (Salton, 1989; Salton et al., 1975; Baeza-Yates and Ribeiro-Neto, 1999), to retrieve relevant documents for a given query vector q with respect to a similarity m, the system searches for all the documents d, classies those with similarity values m(q; d) higher than an explicit or implicit threshold as relevant, and returns to the user a short list of relevant documents with highest similarity values. This information retrieval process is in fact determined by a linear classier, as dened later in this section, which is composed of a query vector q, a similarity m, and a real-valued threshold . Unfortunately, in the real-world information retrieval applications, usually an ideal query vector cannot be generated due to many factors such as the limited knowledge of the users about the whole document collection. A typical example is the real-world problem of web search. In such a case, the user may use a few keywords to express what documents are wanted. However, it is nontrivial for both the user and a web search engine to precisely dene the collection of documents wanted as a query vector composed of a set of keywords. The alternative Some Formal Analysis of Rocchio's Algorithm 5 solution to the query formation problem is, as stated in (Salton, 1989), to conduct searches iteratively, rst operating with a tentative query formation (i.e., an initial query vector), and then improving formations for subsequent searches based on evaluations of the previously retrieved materials. This type of methods for automatically generating improved query formation is called relevance feedback, and one particular and well-known example is Rocchio's similarity-based relevance feedback (J.J. Rocchio, 1971; Ide, 1971a; Salton, 1989). Rocchio's similarity-based relevance feedback algorithm works in a step by step adaptive renement fashion as follows. Starting at an initial query vector q 1 , the algorithm searches for all the documents d such that d is very close to q 1 according to the similarity m, ranks them by m(q; d), and nally presents a short list of the top ranked documents to the user. The user examines the returned list of documents and judges some of the documents as relevant or irrelevant. At step t 1, assume that the list of documents the user judged is x the algorithm updates its query vector as q where the coe-cients t j the algorithm uses the updated query vector q t and the similarity m to search for relevant documents, ranks the documents according to m, and presents the top ranked documents to the user. In practice, a threshold is explicitly (or implicitly) used to select the highly ranked documents. Practically, the coe-cients t j may be xed as (Baeza-Yates and Ribeiro-Neto, 1999; Salton, 1989). The following four typical similarities were listed in (Salton, 1989): For any q; x 2 R n , dice coefficient To make the above denitions valid for arbitrary q and x, we dene that the similarity between two zero vectors is zero, i.e., It should be pointed out that the cosine similarity m 3 is nothing but the inner product similarity m 1 when the vectors are normalized. It is also easy to show that Jaccard similarity m 4 is a strictly monotonic 6 Z. Chen and B. Zhu transformation of the dice similarity m 2 . This implies that both the inner product similarity and the cosine similarity may achieve equivalent classication for any document document with respect to the given query vector. The subtle, but important, dierence between the two similarities is that dierent rank values may be obtained for a document with respect to the given query vector so that dierent \ranking gaps" are obtained between documents. For example, when the rank gap between documents x and y based on the inner product similarity m 1 is 0:8; the rank gap based on the similarity m 3 may be Analogously, the above analysis applies to the dice similarity and the Jaccard similarity. Dierent rank gaps between pairs of documents based on dierent similarities dene dierent user preference structures of the documents, hence dierent document groups or clusters may be obtained. Therefore, dierent information retrieval performances might be achieved for dierent similarities. In other words, if the performance of an information retrieval system is not concerned, then cosine and inner product similarities may be regarded as two equivalent similarities, and so may Jaccard and dice similarities. In (Wong et al., 1988) user preference was studied in terms of structures of weak order and, in particular, linear order. In designing an adaptive information retrieval system, the system performance is denitely not a negligible factor. To our best knowledge, we do not know any provable theoretical results about in uences of the dierent similarities on the performance of the similarity-based relevance feedback. Our linear lower bounds proved in this paper tell us that in the worst case, the four dierent similarities have the same aect on the performance of Rocchio's relevance feedback. But we do not know what the result will be in the average case. As stated in (Baeza-Yates and Ribeiro-Neto, 1999), the main advantage of the relevance feedback is its simplicity and good results. The simplicity is due to the fact that the modied term weights (query vector components) are computed directly from the set of retrieved documents. The good results are observed experimentally and are due to the fact that the modied query vector does re ect a portion of the intended query semantics. The similarity-based relevance feedback algorithm is essentially an adaptive supervised learning algorithm from examples (Salton and Buck- ley, 1990; Lewis, 1991). The goal of the algorithm is to learn some unknown classier to classify documents as relevant or irrelevant. The learning is performed by modifying (or updating) the query vector that serves as the hypothetical representation of the collection of all relevant documents. The method for updating the query vector is similar to the Some Formal Analysis of Rocchio's Algorithm 7 Perceptron algorithm. We given the necessary formal denitions in the following. DEFINITION 1. Let m from R n R n to R + be a similarity. A classier with respect to m over the n-dimensional binary vector space f0; 1g n is a triple (q; ; m), where q 2 R n is a query vector, and R is a threshold. The classier (q; ; m) classies any documents d) or irrelevant otherwise. The clas- sier m) is called a linear classier with respect to the similarity if m is a linear function from R n R n to R + . For simplicity, we may just call (q; ; m) a classier, or a linear classier when m is linear. The following are examples of classiers: In particular, (q is a linear classier but the other three are not, because m 1 is linear and m i are not linear for DEFINITION 2. An adaptive supervised learning algorithm A for learning a target classier (q; ; m) over the n-dimensional binary vector space f0; 1g n from examples is a game played between the algorithm A and the user in a step by step fashion, where the query vector q and the threshold are unknown to the algorithm A, but the similarity m is. At any step t 1, A gives a classier (q m) as a hypothesis to the target classier to the user, where q t 2 R n and t 2 R. If the hypothesis is equivalent to the target, then the user says \yes" to conclude the learning process. Otherwise, the user presents an example such that the target classier and the hypothesis classier dier at x t . In this case, we say that the algorithm A makes a mistake. At step t + 1, the algorithm A constructs a new hypothetical classier m) to the user based on the received examples x The learning complexity (or the mistake bound) of the algorithm A is in the worst case the maximum number of examples that it may receive from the user in order to learn some classier. If the readers are familiar with on-line learning from equivalence queries (Angluin, 1987; Littlestone, 1988), then an adaptive supervised 8 Z. Chen and B. Zhu learning algorithm as dened above is a proper on-line learning algorithm for learning the class of classiers from equivalence queries over the n-dimensional binary vector space. We now give the formal denition of Rocchio's similarity-based relevance feedback algorithm. DEFINITION 3. Rocchio's similarity-based relevance feedback algorithm is an adaptive supervised learning algorithm for learning any classier m) over the n-dimensional binary vector space f0; 1g n from examples. Let q 1 be the initial query vector. At any step t 1, the algorithm presents a classier (q m) as its hypothesis to the target classier to the user, where t 2 R is the threshold, and the query vector q t is modied as follows. Assume that at the beginning of step t the algorithm has received a sequence of examples x then the algorithm uses the following modied query vector q t for its next classication: are called additive updating factors. Please note that our denition above is a generalized version of Rocchio's original algorithm. In our denition, any function m from R n R n to R + can be used as a similarity; arbitrary real values can be used in computing the updated query vector; and nally our denition allows adaptive learning until the target is obtained. REMARK 4. We would like to give the following remarks about Roc- chio's similarity-based relevance feedback algorithm. (a) In the above denition, Rocchio's similarity-based relevance feed-back algorithm can use any real-valued threshold and any real-valued additive updating factors at each step. But in practice the additive updating factors t j may be xed as 1 (or 0:5) to promote the relevance of relevant examples and 1 (or 0:5) to demote the irrelevance of irrelevant examples (Baeza-Yates and Ribeiro-Neto, 1999; Salton, 1989). t 0 is usually set to 1. Also in practice the threshold is usually implicitly used for selecting a short list of top ranked documents. (b) For the purpose of the worst case analysis of the mistake bounds of the algorithm, in the denition we only allow the algorithm to receive one example at each step. We consider that the user acts as an adversary to the algorithm and that the algorithm is required Some Formal Analysis of Rocchio's Algorithm 9 to precisely learn (or search for) the target linear classier. In practical applications, in contrast to the above worst case consid- erations, at each step the algorithm may receive several examples from the user; the user in general may not act as an adversary to the algorithm; and the algorithm is usually required to search for a short list of top ranked relevant documents. Hence, our lower bounds proved in this paper may not aect the algorithm's eective applicability to the real-world problems. The formal analysis of the algorithm helps us to understand its nature well so that we may be able to nd new strategy to improve its eectiveness or to design new algorithms for information retrieval. (c) When the similarity m is the inner product of two vectors, then Rocchio's algorithm is similar to Rosenblatt's Perceptron algorithm (Rosenblatt, 1958). We will use the sets of documents represented by monotone disjunctions of relevant features to study the mistake bounds of Rocchio's algorithm. The e-cient learnability of monotone disjunctions of relevant features (or attributes) has been extensively studied in machine learning (for example, (Littlestone, 1988)). Although very simple in for- mat, monotone disjunctions are very common ways of expressing search queries, especially in the case of web search. All existing popular search engines support disjunctions of keywords as search query formations. For any k with 1 k n, classiers can be dened to precisely classify a monotone disjunction of at most k relevant features i.e., to precisely classify whether any given document satises the monotone disjunction of (2) or not. If we choose a vector u 2 R n such that all its components are zero except that those at positions i are one, then it is easy to verify that for any d 2 f0; 1g n , each of the following four expressions is a necessary and su-cient condition for deciding whether d satises (2): kn Z. Chen and B. Zhu This implies that (u; 1 are all respectively classiers for (2). 3. Technical Lemmas The technique used in (Kivinen et al., 1997) to prove linear lower bounds for the Perceptron algorithm (or, in general, linear additive on-line learning algorithms) is the construction of an example sequence l )) with pairwise constant inner products such that for any given initial query vector (or weight vector as used in (Kivinen et al., 1997)) and any linear classier with the inner product similarity, if the initial query vector and the linear classier dier on the sequence B, then the Perceptron algorithm makes one mistake at one of the two examples in every pair of B. In other words, each pair of the sequence B preserves the classication dierence of the linear classier and the initial query vector for the Perceptron algorithm when the examples in B are used to update the query vector. It was shown in (Kivinen et al., 1997) that row vectors of Hadamard matrices can be used to construct the required example sequence B. We will borrow the above technique from (Kivinen et al., 1997) to prove linear lower bounds for Rocchio's similarity-based relevance feedback algorithm. However, one must note that when a similarity (for example, the Jaccard coefcient similarity) other than the inner product similarity is used, the pairs of the sequence B as used in (Kivinen et al., 1997) may not still preserve the classication dierence of the target classier (which may not necessarily be linear) and the initial query vector. The rotation invariant concept (Kivinen et al., 1997) is in general not applicable to learning algorithms with a non-zero initial query vector, nor applicable to non-rotation variants of the linear additive learning algorithms. Therefore, we need to design new methods for constructing example sequences that are applicable to Rocchio's similarity-based relevance feedback algorithm with arbitrary initial query vector and any of the four similarities dened in section 2. In the following we extend Denition 7 given in (Kivinen et al., 1997) to deal with any similarity. DEFINITION 5. Let the sequence l )), where z 0 t and z 00 t are in f0; 1g n for all t. Let q 1 2 R n be a query vector, m a similarity, and (u; ; m) a classier. Dene q R. We say that the query vector q t and the classier (u; ; m) dier on the sequence B Some Formal Analysis of Rocchio's Algorithm 11 with respect to m if either We now prove a weaker version of Lemma 8 in (Kivinen et al., 1997) in which only the initial query vector (or weight vector in their term) is required to dier from the target linear classier on the example sequence, whereas we require that all the query vectors (the initial one and the updated ones) dier from the target classier which in general may not be linear. We do not use the pairwise constant inner product property, because we may not have such a property for other similarities. LEMMA 6. Let m from R n R n to R + be a similarity and let q 1 2 R n be the initial query vector. Let the sequence l )); where z 0 t and z 00 t are in f0; 1g n for all t. For any classier (u; ; m) over the domain f0; 1g n , if q t , which is as dened in Denition 3, and m) dier on B with respect to m for similarity-based relevance feedback algorithm makes at least l mistakes for learning the classier (u; ; m). Proof. Let A be Rocchio's similarity-based relevance feedback algorithm for learning (u; ; m). Consider the trial sequence in which x t 2 fz 0 t g, and y t is the classication value of x t determined by the unknown target classier (u; ; m) for In other words, y m) classies x t as relevant or y At any step t with 1 t l, the hypothesis of the algorithm A is m) with according to expression (1) of Denition 3. By the assumption, q t and m) dier on the sequence B with respect to m. That is, we have either m(q t ; z or In the rst case, if t m(q t ; z 0 t ), the adversary chooses x t . In this case, i.e., the target classier classies x t as irrelevant, but the hypothesis issued by A classies it as relevant, thus A makes a Z. Chen and B. Zhu mistake. If t > m(q t ; z 0 t ), the adversary chooses x t . In such a case, y i.e., the target classier classies x t as relevant, but the hypothesis issued by A classies it as irrelevant, thus again A makes a mistake. In the second case of m(q t ; z 0 t ), with the same manner we can show that the learning algorithm A makes a mistake at either z 0 t or z 00 t . Therefore, A makes l mistakes on the trial sequence S. This means that A makes at least l mistakes for learning the unknown target classier (u; ; m).We now follow the approach in (Kivinen et al., 1997) to construct example sequences from row vectors of Hadamard matrices. Such sequences are essentially applicable to Rocchio's similarity-based relevance feedback algorithm when the zero initial vector is used, but not when a non-zero initial vector is used. Let I n be the identity matrix of order n. DEFINITION 7. A Hadamard matrix H n of order n is an n n matrix with elements in f1; 1g, such that H n is normalized if the rst row and the rst column consist of ones only. The above (2) implies that any two distinct rows (or columns) of a H n are orthogonal. Normalized Hadamard matrices can be constructed as follows. Let (1). For any Two examples of Hadamard matrices are given as follows. Some Formal Analysis of Rocchio's Algorithm 13 The following property follows from (3) and (4). PROPOSITION 8. For be the t-th row of the normalized Hadamard matrix H n for DEFINITION 9. Let positive integers d and k. For be the t-th row of the normalized Hadamard . We dene BH to be the sequence ((z 0 where z 0 z 00 PROPOSITION 10. Let positive integers d and k and let be the sequence as dened in Denition 9. For any i and j with 1 Proof. We only prove (a), because (b) and (c) can be coped with similarly. By Proposition 8, h is Similarly, is 2 14 Z. Chen and B. Zhu h is We now introduce a new method for constructing example sequences that are applicable to Rocchio's similarity-based relevance feedback algorithm with any of the four similarities and an arbitrary initial query vector. We expand a Hadamard matrix by adding rows and columns with zeroes, and exchange rows and columns of the expanded matrix according to the initial query vector. The new method is given in the proof of Proposition 11. PROPOSITION 11. Given any positive integers d and k, for any query vector q 2 f0; 1g n , there is a sequence such that v 0 t and v 00 t are in f0; 1g n for and the sequence satises all the three properties given in Proposition 10 with each occurrence of 2 d replaced by 2 d 1 and each occurrence of z replaced by v, respectively. Furthermore, we have q has at least 2 d 1 zero components or m 1 (q; v Proof. Given any vector for positive integers d and k, we have This means that we can choose 2 d 1 components of q, denoted by , such that they are either all one or all zero. Dene the is the 2 d 1 2 d 1 Hadamard matrix. We move the rst of C n to the rows , respectively. This process can be achieved through a sequence of exchanges of two rows. In other words, there is an n n transformation matrix A such that AC n does the work and AA . We now move the rst 2 d 1 columns of AC n Some Formal Analysis of Rocchio's Algorithm 15 to the columns , respectively. Similarly, this process can be achieved through a sequence of exchanges of two columns. Moreover, does the work. Now, for any j with 1 row of AC n A T . Then, for 1 s 2 d 1 , the component of d i j denoted by x i s is in fact the s-th component of the j-th row of the Hadamard matrix H 2 d 1 . In other words, d i j has all zero components except these 2 d 1 components x i s forming a subvector that is the same as the j-th row of H 2 d 1 . We nally construct j from d i j by changing all its 1 components to zero and keeping all its other components. We also construct v 00 j from d i j by changing all its 1 components to one and all its one components to zero, and keeping all the other components. In other words, v 0 are constructed as follows: We rst construct z 0 j and z 00 j from the 2 d 1 2 d 1 Hadamard matrix as in Denition 10. We construct v 0 j by adding n 2 d 1 zero components to the end of z 0 j and then moving the rst 2 d 1 components of z 0 through exchanging columns. Similarly, we construct v 00 j by adding n 2 d 1 zero components to the end of z 00 j and then moving the rst 2 d 1 components of z 00 j to exchanging columns. Hence, Proposition 11 follows from Proposition 8 in a manner similar to Proposition 10. 2 In the following two lemmas we show that the sequence BH enables the query vector q t to preserve m 1 similarity for any pair of z 0 t and z 00 t in BH when the zero initial query vector is used, and the sequence enables the query vector q t to preserve m 1 similarity for any pair of v 0 t and v 00 t in D(q 1 ) when the arbitrary initial query vector q 1 is used. LEMMA 12. For positive integers k and d, let BH be the sequence dened in Denition 9. Let where the initial query vector q 2 R, and g. Then, Z. Chen and B. Zhu Proof. For any 1 j < t 2 d , for x (b) and (c) we have m 1 For for any 1 t 2 d , LEMMA 13. Let positive integers k and d. Given any initial query vector q 1 2 f0; 1g n , let D(q 1 ) be the sequence given in Proposition 11, and g. Then, q 1 has at least 2 d 1 zero components. Proof. For any 1 g, by Proposition 11, properties similar to (b) and (c) of Proposition 10 hold for the sequence D(q 1 ). Hence, we have m 1 any 2 t 2 d 1 , we have by Proposition 11 Some Formal Analysis of Rocchio's Algorithm 17 Thus, for any 2 t 2 d 1 has at least 2 d 1 zero components, then again by Proposition 11 we have The following lemma allows us to choose examples in some subdomain to force a learning algorithm to make mistakes. LEMMA 14. Given n > k 1 0, there is an adversary strategy that forces any adaptive supervised learning algorithm to make at least k 1 mistakes for learning the class of disjunctions of at most k 1 variables from fx i 1 g over the binary vector space f0; 1g n . Moreover, the adversary chooses examples in the vector space with nonzero values only for variables in fx i 1 g. Proof. For any given adaptive supervised learning algorithm, at any step t for 1 t k 1, the adversary uses the example x t to defeat the learning algorithm as follows, where x t has all zero components except that its i t -th components is one: If the learning algorithm classies x t as relevant, then the adversary classies it as irrelevant, otherwise the adversary classies it as relevant. 2 One may easily note that the strategy we used above to prove Lemma 14 can be generalized to prove the following fact: Any given adaptive supervised learning algorithm makes at least n mistakes to learn the class of disjunctions of at most n variables from fx over the binary vector space f0; 1g n , because the algorithm can be forced to make one mistake to determine whether or not each of the n variables is in the target disjunction of at most n variables. Unfor- tunately, one must note that this kind of strategy cannot be used to prove that an adaptive supervised learning algorithm (such as Rocchio's algorithm) makes at least n mistakes for learning a disjunction of at most k variable when k is less than n, especially, when k is a small constant. That is the reason why we must design sophisticated methods in this paper to prove linear lower bounds for Rocchio's algorithm for learning disjunctions of at most k variables, where k can be any value between 1 and n. For example, k can be 1, 3, log n, or n. Disjunctions of a small number of variables are the common ways for users to specify their information needs. For example, in the real-world of web search, the number of keywords used in a query session is usually very small. The problem of learning disjunctions of a small number of variables Z. Chen and B. Zhu has been studied by many researchers (see, for example, the work in (Littlestone, 1988)). Careful readers may have observed that we do not rule out the choice of in the decomposition of Proposition 8 to Lemma 13 in this section. When 14 can be simply used to prove linear lower bounds for any adaptive supervised learning algorithm (such as Rocchio's algorithm) to learn disjunctions of at most variables over the binary vector space. However, as we pointed out in the above paragraph, for any k such that must reply on the choice of d 6= 0 in the decomposition of to prove our linear lower bounds. For example, in the case of with the choice of d 6= 0 such that d we can prove that Rocchio's algorithm makes at least mistakes for learning disjunctions of one variable when the zero initial query vector is used. As we will show in the next section, the proof is accomplished through the construction of 2 d pairs of examples from row vectors of Hadamard Once again, the strategy of Lemma 14 or the like is not applicable at all to the case of learning disjunctions of one variable, nor to the case of learning disjunctions of a small number of variables. 4. Linear Lower Bounds We are now ready to prove linear lower bounds for Rocchio's similarity-based relevance feedback algorithm when any of the four typical similarities is used. Throughout this section, we let two positive integers d and k, and let u be the vector in f0; 1g n such that its rst component is one, its last k 1 components have all k 1 or fewer ones (however, these one-components are not specied at this point), and all other components are zero. Given any query vector be its 2 d 1 components such that they are either all zero or all one. Dene u(q 1 ) to be the vector in f0; 1g n such that its i 1 -th component is one, its i j -th components are all zero among the remaining n 2 d 1 components there are at most k 1 one components (again, setting which of these components to be one is not determined at this point). Note that both respectively a monotone disjunction of at most k relevant features. We use E(u) and E(u(q 1 )) to denote the monotone disjunctions represented by u and u(q 1 ), respectively. LEMMA 15. Let 2 d) be the example sequence dened in Denition 9. For any similarity m i , 1 i 4, there Some Formal Analysis of Rocchio's Algorithm 19 is a 2 R such that the query vector and the classier dier on BH with respect to m i for are arbitrary values in R, and x g. Proof. As noted in section 2, 1= are respectively classiers for the monotone disjunction E(u) of at most k relevant features. It follows from Denition 9 that the rst component of z 0 t is one, the rst of z 00 is zero, and the last k components of each of both z 0 t and z 00 t are all zero for 1 t 2 d . Hence, for any 1 t 2 d , we have by Propositionm 1 (u; z 0 kn By Lemma 12 and the above (8) and (12), for any 1 t 2 d , q t and the classier (u; 1=2; dier on the sequence BH with respect to m 1 . For the similarity m 2 , for any t 2 we have by Proposition 10 and Lemma 12 For Because z 00 according to the denition. Thus, for 1 t 2 d , we have hence by (9) and (12) the query vector q t and the classier (u; 2=(k dier on the sequence BH with respect to m 2 . Z. Chen and B. Zhu For the similarity m 3 , for any t 2 we have by Proposition 10 and Lemma 12 For as for the similarity m according to the denition. Thus, (12) the query vector q t and the classier (u; 1= dier on the sequence BH with respect to m 3 . Finally, for the similarity m 4 , for any t 2 we have by Proposition For Thus, for 1 t 2 d , we have m 4 and (12) the query vector q t and the classier (u; 1=(k dier on the sequence BH with respect to m 4 . 2 LEMMA 16. Given any initial query vector ) be the example sequence dened in Proposition 11. For any similarity m i , 1 i 4, there is a 2 R such that the query vector and the classier dier on D(q 1 ) with respect to m i for are arbitrary values in R, and x g. Moreover, if q 1 has at least 2 d 1 zero components, then q 1 and dier with respect to m i , too. Some Formal Analysis of Rocchio's Algorithm 21 Proof. The proof is the same as what we just did for Lemma 15, but we need to replace u by u(q 1 ), z by v, and 2 d by 2 d 1 , respectively. We also need to use Proposition 11 and Lemma 13 to complete our proof.We now prove the following main results in this paper. THEOREM 17. Let positive integers d and k. For any given similarity m i with similarity-based relevance feedback algorithm makes at least n mistakes for learning the class of monotone disjunctions of at most k relevant features over the binary vector space f0; 1g n , when the initial query vector and the similarity m i are used. Proof. Let A be the Rocchio's similarity-based relevance feedback algorithm with the similarity m i and the initial query vector q We analyze the number of mistakes that A must make in learning E(u), the disjunction of at most k relevant features represented by u. As we noted before, there is a 2 R such that the classier classies E(u), i.e., it is logically equivalent to E(u). By Lemma 15, the classier and the query vector q t with 1 t 2 d dier on the example sequence d) with respect to the similarity m i . Thus, by Lemma 6, the adversary can use examples from BH to force the algorithm A to make 2 d mistakes. Note that the last k 1 components of all examples in BH are zero and the last components of u are unspecied but may have at most k 1 one components. Hence, by Lemma 14, the adversary can further force A to make at least k 1 mistakes to learn the values of the last k 1 components of u. Putting all together, A makes at least 2 d mistakes for learning E(u). 2 THEOREM 18. Let positive integers d and k. For any given similarity m i with similarity-based relevance feedback algorithm makes at least (n+k 3)=2 mistakes for learning the class of monotone disjunctions of at most k relevant features over the binary vector space f0; 1g n , when an arbitrary initial query vector q 1 2 f0; 1g n and the similarity m i are used. Moreover, if the initial query vector q 1 has at least 2 d 1 zero components, then the algorithm makes at least (n mistakes. Proof. Let A be the Rocchio's similarity-based relevance feedback algorithm with the similarity m i and the arbitrary initial query vector We analyze the number of mistakes that A must make in learning E(u(q 1 )), the disjunction of at most k relevant features represented by u(q 1 ). As we noted before, there is a 2 R such that the 22 Z. Chen and B. Zhu classier classies E(u(q 1 )), i.e., it is logically equivalent to E(u(q 1 )). By Lemma 16, the classier (u(q 1 and the query vector q t with 2 t 2 d dier on the example sequence D(q 1 respect to the similarity m i . If q 1 has at least 2 d 1 zero components, then the classier and q 1 also dier with respect to m i . Thus, by Lemma 6, the adversary can use examples from D(q 1 ) to force the algorithm A to make 2 d 1 1 mistakes, or 2 d 1 mistakes if q 1 has at least 2 d 1 zero components. Note that there are positions such that the components of all examples in D(q 1 ) at those positions are zero, and the components of u at those components are unspecied but may have at most k 1 one components. Hence, by Lemma 14, the adversary can further force A to make at least k 1 mistakes to learn the values of the components of u at those positions. Putting all together, A makes at least 2 d 1 (n+k 3)=2 mistakes for learning E(u); and if q 1 has at least 2 d 1 zero components, A makes at least 2 d 1 mistakes. 2 The lower bounds obtained in Theorems 17 and are independent of the choices of the threshold and coe-cients that Rocchio's similarity-based relevance feedback algorithm may use in updating its query vector and in making its classication. 5. Discussions 5.1. The Gradient Descent Procedure As pointed out in (Wong et al., 1988), one primary concern in information retrieval is to ensure that those documents more relevant to the user information needs are ranked ahead of those less relevant ones. This means that a ranking is acceptable if it can guarantee that less preferred documents will not be listed in front of the more preferred ones (such a ranking is called an acceptable ranking in (Wong et al., 1988)). Let denote the user preference relation over the document vector space. Wong, Yao and Bollmann designed a very nice gradient descent procedure to compute the query vector q satisfying d d is called the dierence vector for documents d and d 0 . Gradient Descent Procedure. The procedure is outlined as follows (i) Choose an initial query vector q 0 and let Some Formal Analysis of Rocchio's Algorithm 23 (ii) Let q k be the query vector in the k-th step. Identify the set of dierence vectors is a solution vector), terminate the procedure. (iii) Let b: back to step (ii). The above gradient descent procedure is a very nice adaptive algorithm for computing the query vector. However, it must know the user preference relation ahead of the time and perform exhaustive search to identify the set (q k ) at step (ii). The exhaustive search is of exponential time complexity. In practice and in a machine learning setting, the user preference relation is the unknown target that must be learned by an information retrieval system. In this sense, the gradient descent procedure is not an adaptive learning process. On the other hand, Rocchio's similarity-based relevance feedback algorithm formally dened in section 2 is an adaptive learning algorithm without a priori knowledge of the user preference. The updating process for the query vector at each iteration is of linear time complexity. 5.2. Learning Other Document Classes In this paper, we study the learning of a document class represented by a monotone disjunction of index features (or terms) with Rocchio's algorithm. In fact, Rocchio's algorithm can be easily used to learn other target document classes such as the class represented by a conjunction of disjunctions of index features (or terms). Conjunctions of disjunctions are the most common forms for search queries as constructed by humans. We give several examples here. From those examples we know that the lower bounds proved in previous section on the learning performance of Rocchio's algorithms hold for those learning cases. Example 1 (Learning Arbitrary Disjunctions). An arbitrary disjunction is a general case of monotone disjunctions and may have negated index features in it. A negated index feature in a disjunction means that the feature should not occur in the desired documents. Let _ _ be an arbitrary disjunction. Let z document vector that makes g false, i.e., g(z must have z Z. Chen and B. Zhu 1. We can use z to transform g into a monotone disjunction Please also note that the same z can be used to transform f back into g in the following manner The above transformation methods tell us that g can be learned as follows. First, use Rocchio's algorithm with an initial query vector q to learn g. When a document vector judged by the user as irrelevant, then one can use this z to transform any obtained document vector continue the learning. With this kind of transformation, the algorithm actually learns the function f . But as we observed before, f can be easily transformed into g with the document vector z. The above process implies that learning an arbitrary disjunction with Rocchio's algorithm has the same performance as learning a monotone disjunction. Example 2 (Learning Monotone Conjunctions). Let be a monotone conjunction, i.e., all index features occurred in g are positive. The negation of g is a disjunction _ _ Note that the vector z can be used to transform g into a monotone disjunction and to transform f back into g Hence, to learn g we only need to learn its negation g, and by example 1 this can be done with Rocchio's algorithm. Example 3 (Learning Arbitrary Conjunctions). Let be an arbitrary conjunction. The negation g of g is _ _ Some Formal Analysis of Rocchio's Algorithm 25 which is an arbitrary disjunction. To learn g we only need to learn its negation g, and by Example 1 this can be done with Rocchio's algorithm. Example 4 (Learning Disjunctions of Conjunctions). Let be a disjunction of conjunctions, G i is a conjunction for 1 i s. Using the virtual variable technique developed in (Maass and Warmuth, 1998), one can learn G as follows: Introduce one new input variable (virtual variable) for each of the possible 3 n conjunctions that can be constructed from variables x . When the values of x are known, then the value of each virtual variable is easy to be determined. With the help of virtual variables, G is a monotone disjunction and thus can be learned with Rocchio's algorithm. Example 5 (Learning Conjunctions of Disjunctions). Let be a conjunction of disjunctions, F i is a disjunction for 1 i s. Because the negation of F is a disjunctions of conjunctions, as in Example 4 the virtual variables technique developed in (Maass and Warmuth, 1998) can be used to learn the negation of F and hence F itself. 5.3. A Remark on the Optimal Criterion of Rocchio's Algorithm According to (J.J. Rocchio, 1971; Salton, 1989), the construction of an ideal query vector would target to maximize the average query-document similarity for the relevant documents and at the same time minimize the average of the query-document similarity for the irrelevant documents. It is known in (J.J. Rocchio, 1971; Salton, 1989) that under appropriate assumptions such an ideal query vector has the form (R Rel Nonrel where R and N R are the assumed number of relevant and irrelevant documents, and the summations range over the sets of normalized relevant and irrelevant documents, respectively. However, the optimal query vector cannot be adopted in practice, because the sets of relevant and irrelevant documents with respect to the queries are not known 26 Z. Chen and B. Zhu before an exhaustive search. Recall that the gradient descent procedure in (Wong et al., 1988) has a similar problem. On the other hand, the optimal query vector can be adaptively approached in the following (see (Salton, 1989)): where R 0 is the set of documents judged by the user as relevant by the end of iteration i, and N 0 is the set of documents judged as irrelevant. In the above two approximation formulas, 1 , and are updating factors (or coe-cients). In our formalization of Rocchio's algorithm given in section 2, arbitrary updating factors are allowed. The lower bounds we have proved are in fact independent of the choices of the updating factors. That is, our lower bounds hold even if 1 and 1 are used as updating factors. 5.4. Counting Arguments For any k with 1 k n, given any disjunction of k variables it is easy to know that there are 2 n k documents in the binary vector space f0; 1g n that are irrelevant to Q, i.e., the vectors of those documents make Q false; and that there are 2 n 2 n k documents that are relevant to Q, i.e., the vectors of those documents make Q true. When k is small, say 3 is a huge value for a very large n. That is, there are a huge number of documents that are irrelevant to Q. One might ask whether or not this kind of observation helps us to nd some simple ways to prove linear lower bounds for Rocchio's algorithm. For example, an adversary could, in response to almost any formulated query, produce some document which matches the query to some degree but it not relevant. As a matter of fact, this sort of idea is in essence an example of the decision tree technique developed in (Littlestone, 1988) to prove lower bounds for general learning algorithms: In the decision tree, each inner node is labeled with a document vector in the binary vector space f0; 1g n . There are two edges leaving each inner node, labeled \relevant" and \irrelevant", respectively. Some Formal Analysis of Rocchio's Algorithm 27 Each leaf is labeled with a disjunction of at most k variables in such a way that the disjunction at the leaf is consistent with all the labels along the path leading from the root to the leaf ( the number of inner nodes along this path is the depth of the leaf). Assume that the depth of every leaf of the decision tree is at least t. Then, the mistake bound of any learning algorithm for learning disjunctions of at most k variables is at least t. Readers can refer (Littlestone, 1988) for the proof of this statement. The decision tree technique is useful because it reduces the problem of proving a lower bound for every learning algorithm to the problem of constructing a single decision tree with the required depth. The latter can usually be done through counting arguments to estimate the depth of the decision tree. To our best knowledge, (Maass and Turan, 1994) is the best article to prove lower bounds for learning a threshold gate, a general case of a disjunction of at most k variables. The article presents several powerful counting arguments to estimate the depth of decision tree for the class of threshold gates. Unfortunately, those counting arguments do not help us to obtain linear lower bounds for Rocchio's algorithm. The reason is that there are not many monotone disjunctions of at most k binary variables. One can easily nd out that the number of monotone disjunctions of at most k binary variables is The least depth of the decision tree for the class of monotone disjunctions of at most k binary variables is log 2 C(k). For example, log log log log 2 C(k) < log 2 C(n) < n; for any k < n: Therefore, counting arguments cannot yield linear lower bounds for Rocchio's algorithm when k is much less than n, especially when k is a small constant. For example, when the lower bound produced by the counting argument is just log 2 n. However, the non-counting technique that we use is this paper can yield linear lower bounds for Rocchio's' algorithm for any k with 1 k n, even recall that disjunctions of a small number of variables are the common 28 Z. Chen and B. Zhu ways for users to specify their information needs. For example, in the real-world of web search, the number of keywords used in a query session is usually very small. The problem of learning disjunctions of a small number of variables has been studied by many researchers (see, for example, the work in (Littlestone, 1988)). 6. Conclusions and Open Problems Rocchio's similarity-based relevance feedback algorithm is one of the most popular query reformation method in information retrieval and has been used in various applications. It is essentially an adaptive supervised learning algorithm from examples. However, there is little rigorous analysis of its learning complexity. In this paper we prove linear lower bounds for Rocchio's similarity-based relevance feedback algorithm when any of the four typical similarities listed in (Salton, 1989) is used. Because the linear lower bounds are proved with the worst case analysis, they may not aect the algorithm's eective applicability to the real-world problems. The lower bounds help us understand the nature of the algorithm well so that we may nd new strategies to improve the eectiveness of Rocchio's algorithm or design new algorithms for information retrieval. One possible way is to use the Winnow2 (Littlestone, 1988) algorithm as an alternative for the similarity-based relevance feedback algorithm in applications of information retrieval. For example, in our recent research on building real-time intelligent web search engines (Chen et al., 2000; Chen and Meng, 2000), we used a tailored version of Winnow2. We list the following open problems for future research. Problem 1. The lower bound in Theorem holds for an arbitrary initial query vector q 1 2 f0; 1g n . Choosing a zero-one initial query vector is a very common practice in applications. For example, in web search an initial zero-one query vector may be constructed with the query words submitted by the user. When the initial query vector q 1 is chosen from R n , we can prove the same lower bound with the similar but tedious approach for the similarities But we do not know whether the same lower bound still holds for m 4 . Problem 2. In this paper we have proved linear lower bounds for Rocchio's similarity-based relevance feedback algorithm in the binary vector space. We do not know whether our approach can be extended to study the learning complexity of Rocchio's algorithm in arbitrary discretized vector space. Problem 3. The linear lower bounds we have established for Roc- chio's algorithm in the binary vector is based on the worst-case anal- Some Formal Analysis of Rocchio's Algorithm 29 ysis. It would be very interesting to analyze the average-case learning complexity of Rocchio's algorithm. We feel that this problem is very challenging, because any nontrivial average case analysis will reply on realistic models of document distribution, index term distribution, and the user preference distributions as well. We feel that it is not easy to model those distributions nor to analyze the complexity under those distributions. The probabilistic corpus model proposed in (Papadimitrious et al., 2000) may shed some light on this problem. Problem 4. Although it follows from our linear lower bounds that the Winnow2 algorithm (Littlestone, 1988) has better worst case learning complexity, the authors do not know any provable theoretical analysis of the average learning complexity about the Winnow2 algorithm, nor about the similarity-base relevance feedback algorithm. More pre- cisely, we do not know whether the Winnow2 algorithm performs better in average than the Rocchio's similarity-based relevance feedback algorithm. Acknowledgements In early 1997, Dr. Stanley Sclaro asked the rst author whether the relevance feedback algorithm can be used to help the user search for the desired World Wide Web documents with about two dozens of examples judged by the user. At the time, his research group implemented ImageRover (Taycher et al., 1997; Sclaro et al., 1997), an image search engine from the user's relevance feedback, while the author and his colleagues started to build intelligent search tools (such as Yarrow (Chen and Meng, 2000), WebSail (Chen et al., 2000) and Features (Chen et al., 2001)) with the help of information retrieval and machine learning techniques. Dr. Sclaro's question together with the authors' own research on building intelligent web search tools inspired the work in this paper. The authors would also acknowledge that the example sequence selection method developed in (Kivinen et al., 1997) for proving linear lower bounds for the Perceptron algorithm is the key to the breakthrough of our proofs. Without knowing the method it would take longer for the authors to nish the work in this paper. We thank three anonymous referees and the journal editor, Professor Paul B. Kantor, for their critical and valuable questions and comments for helping us to revise this paper. We thank one referee for informing us the reference (Wong et al., 1988). Z. Chen and B. Zhu --R Queries and concept learning. Modern Information Retrieval. Multiplicative adaptive algorithms for user preference retrieval. Yarrow: A real-time client site meta search learner WebSail: From on-line learning to web search IEEE Press Information Retrieval: Data Structures and Algorithms. Relevance feedback in information retrieval. The Smart Retrieval System - Experiments in Automatic Document Processing The perceptron algorithm vs. Winnow: linear vs. logarithmic mistake bounds when few input variables are relevant. Learning in intelligent information retrieval. Learning quickly when irrelevant attributes abound: A new linear-threshold algorithm Information and Computation Latent semantic indexing: A probabilistic analysis. A critical analysis of the vector space model for information retrieval. The perceptron: A probabilistic model for information storage and organization in the brain. Automatic Text Processing: The Transformation A vector space model for automatic indexing. Cascia M and Sclaro Linear structures in information retrieval. --TR --CTR Ying Liu , Dengsheng Zhang , Guojun Lu , Wei-Ying Ma, A survey of content-based image retrieval with high-level semantics, Pattern Recognition, v.40 n.1, p.262-282, January, 2007
relevance feedback;similarity;supervised learning;lower bound;vector space
594068
Spectral Sequences on Combinatorial Simplicial Complexes.
The goal of this paper is twofold. First, we give an elementary introduction to the usage of spectral sequences in the combinatorial setting. Second we list a number of applications.In the first group of applications the simplicial complex is the nerve of a poset&semi; we consider general posets and lattices, as well as partition-type posets. Our last application is of a different nature: the {\cal S}_n-quotient of the complex of directed forests is a simplicial complex whose cell structure is defined combinatorially.
Introduction In this paper we use spectral sequences to compute homology groups of combinatorially given simplicial complexes, whether they come as nerves of posets or with an explicit combinatorial description of the cell structure. This idea is not new, in fact spectral sequences have been used for that purpose in a quite general setting, already in, e.g. [Ba75, Ba77, Ba80], [Q1]. Recently, these methods have started to take more concrete forms, for example Phil Hanlon used them in [H] to compute the homology groups of the so-called generalized Dowling lattices. In the joint paper [FK], Eva Maria Feichtner and the author used spectral sequences to attack an especially difficult case of subspace arrangements, namely the so-called D n;k -arrangements. In Section 2 we define some basic notions. Then, in Section 3, we give a thorough and elementary, from scratch, description of one possible way to use spectral sequences for poset homology computations. In Section 4 we derive several corollaries of the properties of the spectral se- quences, which can be applied to a wide class of posets. These results include both M-obius function computations and finding the Betti numbers of a poset. We take a look at the Whitney homology of a poset and the intriguing questions coming up in this context. In Theorem 4.1 we prove two inequalities for the Betti numbers of an arbitrary lattice. In Section 5 we apply these methods to different partition-type posets. In Sub-section 5.1 we consider the intersection lattices of orbit arrangements, \Pi Furthermore, we compute completely the homology groups of the particular lattice \Pi 3;2;2;1 . This example shows that the homology groups of orbit arrangements A - can have very irregular structure in general, which was not known before. We remind the reader that it was shown in [Ko1, Theorem 4.1] that when a partition - has no primitive partition identities then \Pi - is shellable, in particular it is Date: March 21, 2000 Research at IAS was supported by von Hoffmann, Arcana Foundation. homotopy equivalent to a wedge of spheres. In Subsection 5.2 we take a look at the subspace arrangements associated with certain partitions with restricted block sizes. In Section 6 we use spectral sequences to study homology groups of the Sn - quotient of the complex of directed forests \Delta(G n )=Sn . In [Ko2] it was shown that \Delta(G n ) is shellable. Here we derive a formula for the rational Betti numbers of \Delta(G n )=Sn and also detect torsion in its integer homology groups. 2. Basic notions and definitions In this section we introduce the basic notions which we use throughout the text. Definition 2.1. Let P be a finite poset. The nerve of P , \Delta(P ), (also known as the order complex of P ), is the abstract simplicial complex whose vertices are the elements of P and whose faces of dimension k are the chains x in P . See [Q2] for a more general definition. If P is explicitly given with adjoint elements then we consider the simplicial complex \Delta( - 1g. Where it causes no confusion we often write \Delta(P ) instead of \Delta( - We also use the convention that unless the bar ( - ) is explicitly written, the concerned poset always has adjoint elements For an arbitrary simplicial complex C, e will denote the kth reduced homology group of C (see [Sp] for a definition). For the sake of brevity we will often Furthermore we let -P (x; y) denote the value of the M-obius function on the interval (x; y) of the poset P . The definition and many properties of the M-obius function can be found for example in [St1]. We use the convention - Definition 2.2. A poset P is called Cohen-Macaulay (or just CM) if for every interval (x; y) of the poset P we have e 2. Recall that a chain complex C of vector spaces (resp. abelian groups) is a sequence \Gamma! Cn d \Gamma! of vector spaces (abelian groups) and maps between them, such that d A filtration on C, a collection of filtrations on each all i, j; here we denote F 3. Spectral sequences for the nerves of posets Spectral sequences constitute a convenient tool for computing the homology groups of a simplicial complex. Here we give a short description of one possible way to apply spectral sequences to compute homology groups of the nerve of a poset. A special case of this particular approach has been previously used by Phil Hanlon, in the work cited above. Of course, the filtrations considered here are very special, but we hope that this may be a good starting point for a combinatorialist. A few good sources for the material on spectral sequences are [McC, Sp, Mas]. SPECTRAL SEQUENCES ON COMBINATORIAL SIMPLICIAL COMPLEXES 3 3.1. The definition and some properties of spectral sequences. A spectral sequence associated with a chain complex C and a filtration F on C is a sequence of 2-dimensional tableaux (E r r=0 , where every component E r k;i is a vector space (for simplicity we start with considering field coefficients), E r unless k; i - 0, and a sequence of differential maps (d r ) 1 r=0 such that d r im(E r k+1;i+r d r Comments. It follows from (0) and (2) that E 1 1. In the general case k;i is defined using the notion of convergence of the spectral sequence. We will not explain this notion in general, since for the spectral sequence that we consider only a finite number of components in every tableau E r are different from zero, so there exists N 2 N, such that d one sets k;i . 2. For the case of integer coefficients, (3.2) becomes more involved: rather than just summing the entries of one needs to solve extension problems to get H (C). This difficulty will not arise in our applications, so we refer the interested reader to [McC] for the detailed explanation of this phenomena. When considering integer coefficients, E r ; are not vector spaces, but just abelian groups. 3. We would like to warn the reader that our indexing is different from the standard (but more convenient for our purposes). The standard indexing is more convenient for the spectral sequences associated to fibrations, an instance we do not discuss in this paper. For a finitely generated abelian group G, let rk G denote the dimension of the G. When specializing to a spectral sequence for the homology of the nerve of a finite bounded poset, we immediately observe that its M-obius function can be read off from the E r non-negative integer r. Proposition 3.1. Let P be a finite poset and (E r an associated spectral se- quence, then Proof. It is a well known fact that k;i , see for example [McC, Example 6, pp. 15-16]. Furthermore the theorem of Ph. Hall says that Formula (3.3) follows from (3.4) and (3.5). As we will see later, formula (3.3) specializes to several well-known formulae for M-obius function computations, once the spectral sequence is specified. Proposition 3.2. Let P be any poset and (E r spectral sequence for H (\Delta(P )). Then we have (for some r and, for any k, Proof. From (3.1) we have rk E r+1 and (3.6) follows. It also follows that We shall now prove (3.8). Let us denote d n+1;i+r , then we have the following diagram d2 n+1;i+r d3 From the definition of the spectral sequence we know that hence Comment. We use here the fact that if G is an abelian group and H is a sub-group of G then rk Summing over all i 2 Zin (3.9) we obtain hence using formula (3.2) we obtain SPECTRAL SEQUENCES ON COMBINATORIAL SIMPLICIAL COMPLEXES 5 3.2. A class of filtrations. In this subsection we consider all homology groups with coefficients in F, where F is either a field or the ring of integers. In fact, everything prior to (3.13) is valid for F being an arbitrary ring. Let us describe a special class of filtrations on the chain complex for \Delta(P ). First of all one chooses the following data: J a subposet of - P and a function N, such that f( - other words the preimage of each element in N forms an antichain in J . The most frequent choices of the function f are the rank function on J (when it exists) and an arbitrary linear extension of the partial order on J . The choice of J is more subtle and usually depends heavily on the structure of the poset P , see [FK, H]. Having chosen f and J , we will define an increasing filtration on the chain complex for \Delta(P ). be a chain (not necessarily maximal) in P . Define the pivot of \Gamma, piv(\Gamma), to be the element of \Gamma " J with the highest value of the function f . Since the preimages under f of each natural number form an antichain, we know that f takes different values on different elements in hence the notion of pivot is well defined. Furthermore, let the weight of \Gamma, !(\Gamma), be the value of f on the pivot, i.e., take - 0 as a pivot and say that the chain \Gamma has weight 0. This assignment of weights gives us the filtration of the chain complex C (P ff F with h \Delta i F denoting the linear span of the given chains with coefficients in F. Recall that by the definition of the nerve of a poset, Omitting an element other than the pivot does not alter the weight of the chain, omitting piv(\Gamma) turns another element into the pivot, on which f takes a lower value than on the former pivot, so the resulting chain has a strictly lower weight. Hence the differential operator @ respects the filtration. By construction, the filtration is bounded from below. By definition ff F The differential d k\Gamma1;i is induced by the simplicial boundary operator. be a generator of k;i , then since the weight of a chain is lowered by the omission of an element if and only if it is the pivot which is removed. Now we shall replace the chain complexes chain isomorphic complexes. The latter allow us to give an explicit description of 6 DMITRY N. KOZLOV the in terms of the homology groups of certain subposets of P . First we need some notations: for a 2 J , let S a f(a)g. There is an obvious isomorphism between the following chain complexes "divid- ing" each chain \Gamma in P with pivot a ? - 0 into two chains, namely its subchains below and above the pivot: with e denoting the augmented simplicial chain complex of the respective intervals. We need to use augmented complexes including the empty chain in order to get proper counterparts for chains which have the pivot as maximal element below - 1 or as minimal element above - 0. Let ~ id\Omega ~ with the usual sign conventions, namely ~ @c p\Omega ~ for c p 2 e One can see that the isomorphism commutes with the boundary operators d 0 and ~ @\Omega , respectively. Hence ' is actually a bijective chain map and we get For simply have In case F is a field, or F = Zand at least one of the subposets ( - 0; a) " S a and S a has free homology groups, we can apply the algebraic K-unneth theorem and deduce In this setting, Proposition 3.1 specializes to Special cases of formula (3.14) can be found in for example [St1]. Observe that when P is a lattice and P , and f is an arbitrary order preserving function on J [ f - 0g, then (3.14) gives Weisner's theorem: For the explicit derivation of the E 1 ; -tableau in this case see Theorem 4.1. For more information on Weisner's theorem itself the reader may want to consult [St1, Corollary 3.9.3]. SPECTRAL SEQUENCES ON COMBINATORIAL SIMPLICIAL COMPLEXES 7 When J is a lower ideal and f is an order-preserving function, i.e. if x ? y then specializes to a)\Omega e 4. Applications for general posets Let P be a pure poset. Form a spectral sequence by choosing P and according to (3.15) and (3.12), rk (a)=i e We can read off the so-called Whitney homology groups of P (first introduced and studied by Baclawski in [Ba75]) from the E 1 e now P be a CM poset, then 2. It means that under the first differential d 1 all of the groups W k (P ), except for the highest one, cancel in some intriguing way. It would be of a great interest to clarify the combinatorial nature of these cancellations. Theorem 4.1. Let P be a finite lattice, x an atom in P . Then the following inequalities hold: In particular, if fi k\Gamma2 ( - 0; Proof. Let P-x and let x be any linear extension of J . Consider the spectral sequence E which is associated to the ideal J , where we filtrate using the given linear extension of J . Observe first that - P-x is contractible. Also, for any a 2 J , we have (a; - P-x-a . This means that (a; - contractible (actually a cone with apex x- a) using formulae (3.12) and (3.15), we obtain 0; otherwise. The inequalities (4.2) and (4.3) follow from inequalities (3.7), resp. (3.8). Applications of Theorem 4.1 will be given in the next section. The following theorem may be occasionally useful. Theorem 4.2. Let P be a pure poset of rank r. Suppose that there exists a subposet J of P such that (2) for any a 2 J , both [ - 0; a] and [a; - 1] J are CM and rk [a; - e a)\Omega e Proof. Construct the spectral sequence r=0 with the subposet J as in the proof of the Theorem 4.1 and with f(x) =rk(x). Then it follows from the formulae and (3.12) that E 1 rk (a)=i;a2J a)\Omega e Using the fact that P nJ , [ - 0; a] and [a; - 1] J are CM and that rk (P nJ) =rk (P r\Gamma2. The spectral sequence collapses here, hence (4.4) and e 2, follow from (3.2). Let us recall a theorem proved in [BWal]. Theorem 4.3. (Complementation Theorem). If L is a bounded lattice, s 2 - L, and the complements of s form an antichain, then - wedge x?s Remark 4.4. In the special case, when the complements of an atom x 2 P form an antichain, the spectral sequence above allows us to derive the homology counter-part of the Complementation Theorem 4.3. Reason. If the complements of x form an antichain one can choose the function f so that it takes the same value v on all of the complements of x. Then there will be only one non-zero row in E 1 namely e All the differentials d r will be zero maps for r - 1, so we obtain e 5. Applications to partition-type posets 5.1. Orbit arrangements. A subspace arrangement A is a finite collection of affine subspaces in the Euclidean space R n . Let A be a central subspace arrangement (all the subspaces pass through the origin) and take all possible non-empty intersections t, ordered by reverse inclusion, that is This is a partially ordered set, which is actually a lattice. The SPECTRAL SEQUENCES ON COMBINATORIAL SIMPLICIAL COMPLEXES 9 bottom element is - and the top element is This lattice is called the intersection lattice and is often denoted by LA . We use the notation - for the partition of the number into blocks of sizes - and we always have these blocks ordered in non-increasing . By \Pi n we denote the partition lattice of the set [n]. It is the poset with elements all different partitions of [n] ordered under refinement. The following class of subspace arrangements was first introduced in [Bj94, sub-section 3.3]. If nontrivial partition of the set [n], let The type of - is the number partition of n given by the block sizes jB i j. Given a non-trivial number partition - ' n, let A - is called an orbit arrangement, expressing the fact that A - is the orbit of any single subspace K - under the natural action of Sn on R n . Let \Pi Note that \Pi Theorem 5.1. Consider a partition - notation means that we have m blocks of size 1). Let 2. Then \Pi - is Remark. For this bound to be useful, we should have much larger For example, for Proof. Take a coatom consider the spectral sequence associated with the ideal We have e Let d be the number of blocks in y, then [y; - 1] ' \Pi d (here we use that 2 occurs as a block size in -). We shall show that d - t. Let y have blocks of sizes . The set fs gives a number partition of n, means that - is a refinement of fs g. The condition x - means that there exists a block of y, without loss of generality we can assume it is s d , such that - is not a refinement of fs 1g. It means it is impossible to pack disjointly blocks of sizes - blocks of sizes s We will attempt to perform such a packing with a version of a greedy algorithm. Let us start with packing - 1 into some of the blocks s 1. If it is possible continue with - 2 and so on. At some point we will have to stop. Say we stopped at - i , i.e., it is impossible to pack - i into the rest (after packing - the blocks s 1. Then the rests of the blocks s have at most us an inequality or which implies hence d - t. The lattice \Pi d has nonzero homology group only in dimension d \Gamma 3, so using (3.6), we can conclude that \Pi - is Often spectral sequences can be used for a direct computation of the poset homology groups. We will give here an informative example. 1). We shall compute the homology groups of \Pi 3;2;2;1 . The poset \Pi 3;2;2;1 is pure and ranked by the function (the number of blocks in x). \Pi 3;2;2;1 , f(x) =rk(x), and construct the corresponding spectral sequence. As was described in Section 4 we obtain Whitney homology groups. It is easy to see that ( - 0; a) is CM for all a 2 - \Pi 3;2;2;1 except for the case when a has partition type (4; 4). These intervals are schematically shown in Figure 1. Figure 1. The Betti numbers of intervals ( - 0; a) are given in the following table. type of a number e e e 44 Observe that we can use formulae (3.15) and (3.12), since the intervals ( - 0; a) have . Hence, the f(x) =rk(x), can be easily computed. The only non-zero entries are It is easy to check that the first two rank levels of \Pi 3;2;2;1 form a connected poset, hence d 1 is exact in E 1 0;1 . It means that Already this shows that H 1 (\Pi 3;2;2;1 ) 6= 0. SPECTRAL SEQUENCES ON COMBINATORIAL SIMPLICIAL COMPLEXES 11 It is not difficult to show that d 1 is exact in E 1 1;2 too (this will be done later). Hence the associated spectral sequence collapses at its second term, and the non-zero entries of the tableau are: Hence, e In [Ko1, Theorem 4.1] it has been proved that \Pi - is shellable if - has no primitive partition identities. This of course does not apply to \Pi 3;2;2;1 , since has the identity 3+ 1. It is however not difficult to adapt the proof of [Ko1, Theorem 4.1] to show that is shellable. This adaptation is technical and requires to go into the details of the 4-pages proof of the mentioned theorem, so we shall omit this argument. Alternatively, one could show that P is shellable by a direct argument. Now, associate a spectral sequence ( e to P in the same way as above. The Whitney homology groups of P are subgroups of the Whitney homology groups of \Pi 3;2;2;1 . On the other hand, since P is shellable, d 1 must be exact in e 1;2 . Then, of course, d 1 is also exact in E 1 1;2 . 5.2. Partitions with restricted block sizes. Let \Pi n;1;:::;k denote the poset consisting of partitions with block sizes from the set n). These lattices were considered in [W] in connection with certain relative subspace arrangements. It is believed that \Pi n;1;:::;k is torsion-free. We can obtain some information on the homology groups of these lattices from the following proposition. Proposition 5.2. \Pi n;1;:::;k is Proof. \Pi n;1;:::;k is a lower ideal of the partition lattice \Pi n . \Pi n is a CM poset and \Pi n;1;:::;k contains the first k \Gamma 1 rank levels of \Pi n . Let J be a subposet of \Pi n;1;:::;k consisting of the complement of the first k \Gamma 1 rank levels of \Pi n , f(x) Then the formulae (3.13) specialize to e since (a; - Every interval ( - 0; a) is a CM poset of rank rk(a) - k, also P n J is CM of rank k, hence Using (3.2) we conclude that e 6. Sn-quotient of the complex of directed forests In this section we shall assume the following notions to be known: directed graph, a subgraph of a directed graph, directed tree, directed forest. If needed the reader may consult any textbook on graph theory for the definitions. We shall use V (G), resp. E(G), to denote the sets of vertices, resp. edges, of a directed graph G. We think of E(G) as a subset of (V (G) \Theta V (G)) n f(x; (G)g. Since all the graphs considered in this section are directed, we will often omit this word. Following a hint of R. Stanley, [St2], the following simplicial complexes were considered in [Ko2]. Definition 6.1. Let G be an arbitrary directed graph. Construct a simplicial complex \Delta(G) as follows: the vertices of \Delta(G) are given by the edges of G and k- simplices are all directed forests with are subgraphs of G. Let Gn be the complete directed graph on n vertices, i.e., a graph having exactly one edge in each direction between any pair of vertices, all together n(n \Gamma 1) edges. It was shown in [Ko2] that \Delta(G n ) is shellable, thus all its homology groups are 0 except for the top one, and one can show that fi n\Gamma2 (\Delta(G n Furthermore, there is an action of Sn on \Delta(G n ) induced by the permutation action of Sn on [n], thus one can form the topological quotient see Figure 2 for the case It was asked in [Ko2, Section 6, Question 2] what H was. The answer to that turned out to be more complex than we thought. In this section we show that the groups H are not, in general, free, and also give a formula for fi n\Gamma2 A combinatorial description for the cell structure of Xn . Clearly, the action of Sn on \Delta(G n ) is not free. What is worse, the elements of Sn may fix the simplices of \Delta(G n ) without fixing them pointwise: for example for permutation (23) "flips" the 1-simplex given by the directed tree 2 /\Gamma 1 \Gamma! 3. Therefore, one does not have a bijection between the orbits of simplices of \Delta(G n ) and simplices of Xn . To rectify the situation, consider the barycentric subdivision We have a simplicial Sn -action on Bn induced from the Sn - action on \Delta(G n ) and, clearly, Bn =Sn is homeomorphic to Xn . Furthermore, if an element of Sn fixes a simplex of Bn then it fixes it pointwise. In this situation, it is well-known, e.g. see [Br], that the quotient projection Bn ! a simplicial structure on Xn , in which simplices of Xn correspond to Sn -orbits of the simplices of Bn with appropriate boundary relation. Let us now give a combinatorial description of the Sn -orbits of the simplices of Bn . Let oe be a simplex of Bn , then oe is a chain of forests on labeled vertices, such that T i is a subgraph of T i+1 , for dim(oe). One can view this data in a slightly different way: it is a forest with dim(oe)+1 integer labels on edges (labels on different edges may coincide). Indeed, given a chain of forests above, take the forest T dim(oe)+1 and put label 1 an all edges of the forest T 1 , label 2 on all edges of T 2 , which are not labeled yet, etc. Vice versa, given a forest T with a labeling, let T 1 be the forest consisting of all edges of T with the smallest label, let T 2 be the forest consisting of all edges of T with one of the two smallest labels, etc. To make the described correspondence a bijection, one should identify all labeled forests on which labelings produce the same order on edges. Formally: the p-simplices of Bn are in bijection with the set of all pairs (T; OE T ), where T is a directed forest on n labeled vertices and OE Z, such that there exists an order-preserving injection / : Z! Z, such that The boundary operator is given by: for a p-simplex (T; OE T takes the same values as OE T except for the edges on which OE T takes ith and (i+1)st largest values (say a and b), on these edges OE T i takes value a. Furthermore, SPECTRAL SEQUENCES ON COMBINATORIAL SIMPLICIAL COMPLEXES 13 T p+1 is obtained from T be removing the edges with the highest value of OE T , OE Tp+1 is the restriction of OE T . Of course, this description of the boundary map is just a rephrasing of the deletion of the ith forest from the chain of forests in the original description. However, we will find it more convenient to work with the labeled forests rather than the chains of forests. The orbits of the action of Sn can be obtained by forgetting the numbering of the vertices. Thus, using the fact that simplices of Xn and Sn -orbits of simplices of Bn are the same thing, we get the following description. The p-simplices of Xn are in bijection with pairs (T; OE T ), where T is a directed forest on n unlabeled vertices and OE T is an edge labeling of T with modulo a certain equivalence relation. This equivalence relation and the boundary operator are exactly as in the description of simplices of Bn . Figure 2. On Figure 2 we show the case 3. On the left hand side we have \Delta(G 3 ), on the right hand side is X . The labeled forests next to the edges indicate the bijection described above, labeling on the forests corresponding to the vertices in X 3 is omitted. S 3 acts on \Delta(G 3 ) as follows: 3-cycles act as rotations around the line which goes through the middles of the triangles, each transposition acts as a central symmetry on one of the quadrangles, and as a "flip" on the edge which is parallel to that quadrangle. Filtration. There is a natural filtration on the chain complex associated to the simplicial structure on described above. Let F i be the union of all simplices has at most i edges. Clearly, The description of the E 1 tableau. Recall that E 1 we use the indexing from Section 3. In other words, the homology is computed with "truncated" boundary operator: the last term, where some edges are deleted from the forest, is omitted. Clearly, where the sum is over all forests with k edges and E T is a chain complex generated by the simplices (T; OE T ), for various labelings OE T , with the truncated boundary operator as above. Let us now describe a simplicial complex whose reduced homology groups, after a shift in the index by 1, are equal to the nonreduced homology groups of E T . The arrangement of k(k \Gamma 1)=2 hyperplanes x cuts the space S simplices, where H is the hyperplane given by the equation x 1 14 DMITRY N. KOZLOV Denote this simplicial complex A k . The permutation action of S k on [k] induces an S k -action on A k . It is easy to see that if an element of S k fixes a simplex of A k , then it fixes it pointwise. Hence, for any subgroup \Gamma ' S k , the \Gamma-orbits of the simplices of A k are in a natural bijection with the simplices of A k =\Gamma. Let T be an arbitrary forest with n vertices and k edges. Assume that vertices, resp. edges, are labeled with numbers acts on [n] by permutation, let Stab (T ) be stabilizer of T under this action, that is the maximal subgroup of Sn which fixes T . Then Stab (T ) acts on E(T ), i.e., we have a homomorphism does not depend on the choice of the labeling of vertices. However, relabeling the edges changes S(T ) to a conjugate subgroup. Therefore, for a forest T without labeling on vertices and edges, S(T ) can be defined, but only up to a conjugation. Proposition 6.2. The chain complex of A k =S(T ) and E T (with a shift by 1 in the indexing) are isomorphic. In particular, e Proof. Label the k edges of T with numbers As mentioned above, the p-simplices of E T are in bijection with labelings of the edges of T with numbers (using each number at least once). Taking in account the chosen labeling of the edges, this is the same as to divide the set [k] into an ordered tuple of non-empty sets, modulo the symmetries of [k] induced by the symmetries of T . Clearly, these symmetries of [k] are precisely the elements of S(T ). The (p \Gamma 1)-simplices of A k are in bijection with dividing [k] into an ordered tuple of p+1 non-empty sets: by the values of the coordinates. Therefore we conclude that the p-simplices of E T are in a natural bijection with the (p\Gamma1)-simplices of A k =S(T ). Here the unique 0-simplex of E T , (T; 1), (1 is the constant function taking value 1), corresponds in A k =S(T ) to the empty set, which is a (\Gamma1)-simplex. One verifies immediately that the boundary operators of E T and A k =S(T ) commute with the described bijection. Therefore E T and A k =S(T ) are isomorphic as chain complexes (after a shift in the indexing). In particular, e coefficients. Proposition 6.2 allows us to give a description of -entries in the case when the homology groups are computed with rational coefficients. Indeed, it is well known that, when a finite group \Gamma acts on a finite simplicial complex X, one has e is the maximal vector subspace of e acts trivially (more generally Q can be replaced with a field whose characteristic does not divide j\Gammaj). Since A k is homeomorphic to S k\Gamma2 we have e 2. It is easy to compute e In fact, for - Q), one has where sgn denotes the sign homomorphism sgn 1g. Therefore e 0; otherwise, where A k is the alternating group, A Combined with the Proposition 6.2 this gives H i cases. Therefore it follows from (6.1) that rk E 1 is equal to the number of forests T with k edges and n vertices, such that S(T ) ' A k . rk that fi i SPECTRAL SEQUENCES ON COMBINATORIAL SIMPLICIAL COMPLEXES 15 in [Ko2]), and fi i In particular, by computing the Euler characteristic of Xn in two different ways, we obtain Theorem 6.3. For n - 3, fi n\Gamma2 The first values of f k;n are given in the table below. Note that there are zeroes on and below the main diagonal and that the rows stabilize at the entry (k; (for k - 2). Zcoefficients. The case of integer coefficients is more complicated. In general, we do not even know the entries of the first tableau. However, we do know that it is different from the rational case, i.e., torsion may occur. For example, let T be the forest with 8 vertices and 6 edges depicted on Figure 3. Clearly, (12)(34)(56)g. It is easy to see that A 6 =S(T ) is a double suspension (by which we mean suspension of suspension) of RP 2 , thus the only nonzero homology group is e In particular, E 1 4;6 is not free. Figure 3. 2 6On the positive side, we can describe the values which d 1 takes on the "rational" generators of E 1 us call a forest admissible if S(T ) ' A jE(T )j . For every admissible forest T with k edges we fix some order on the edges, i.e., a bijection determines uniquely an integer generator e T of H by where we sum over all right cosets of S(T ), (we choose one representative for each coset). Observe that the sign of g, resp. the simplex (T; g ffi / T ), are the same for different representatives of the same right coset class, because S(T ) ' A k , resp. by the definition of S(T ). Proposition 6.4. For an admissible forest T , we have ff where the sum is over S(T )-orbits of E(T ), for which there exists a representative ff, such that T n ff is admissible, we choose one representative for each orbit; note that the admissibility of T n ff depends only on the S(T )-orbit of ff, not on the choice of the representative. Notation in the formula: T n ff denotes the forest obtained from T by removing the edge ff; ~ defined by ~ ~ consists of those permutations of edges of T n ff which can be extended to T by fixing the additional edge. Proof. For an admissible forest T with k edges and a bijection OE : E(T OE) denote a face simplex of (T; OE), where e T is obtained from T by removing the edge with the highest label, ~ OE is the restriction of OE to e T . In our notations However, for convenience, we use the notation "tilde" in the rest of the proof. According to the general theory for spectral sequences, d 1 @ denotes the usual boundary operator, and we view @(e T ) as embedded into the relative homology group H k\Gamma2 linear combination of simplices which are obtained from the simplices (T; g ffi / T ) by either merging two labels, or omitting the edge with the top label. e T 2 H means that the application of the "truncated" boundary operator to e T gives 0, therefore all the simplices obtained by merging two labels will cancel out. Furthermore, since 2, and the group H k\Gamma2 freely generated by e U , where U is an admissible forest with edges, we can conclude that also the contributions ( e OE), where e T is not admissible, will cancel out. Combining these arguments with (6.2) we obtain: sgn (g)( e where we have only those terms left in the sum, for which e T is admissible. After regrouping we get sgn (g)( e ff sgn (g)( e where in the second term the first sum is taken over all S(T )-orbits of [k], for which e T is admissible, while the second sum is taken over all right cosets S(T )g which have a representative g such that g ffi / T we take one representative per coset. To verify (6.5) we just need to observe that the S(T )-orbit of (g does not depend on the choice of the representative of S(T )g; this follows from the definition of S(T ). Finally, one can see that, for ff being an edge of T , such that T n ff is admissible, sgn (g)( e S(Tnff)h where the sum in the first term is again taken over all right cosets S(T )g which have a representative g such that g ffi / T and the sum in the second term is simply over all right cosets of S(T n ff). Indeed, on the left hand side we have a sum over all labelings of E(T ) with such that ff gets a label k, and we consider these labelings up to a symmetry of T ; each labeling comes in with a sign of the permutation g, which is obtained by reading off this labeling in the order prescribed by / T . On the right hand side the same sum is regrouped, using the observation that to label E(T ) with [k], so that ff gets a label k, is the same as to label E(T n ff) with [k \Gamma 1]. The only details which need attention are the multiplicity and the sign. Every S(T )-orbit of labelings of E(T ) with [k] so that ff gets a label k corresponds to [S(T S(T )] of S(T n ff)-orbits of labelings of E(T n ff) with [k \Gamma 1], since we identify labelings by the actions of different groups: S(T n ff) ' e S(T ). Each of this S(T n ff)-orbits comes with the same sign, because S(T n ff) ' A k\Gamma1 . The sign SPECTRAL SEQUENCES ON COMBINATORIAL SIMPLICIAL COMPLEXES 17 corresponds to the change of the order in which we read off the edges: instead of reading them off according to / T , we first read off along / Tnff and then read off the edge ff last. Formally: g ffi / hence sgn defined by ~ hj h, ~ Combining (6.4), (6.5) and (6.6) we obtain (6.3). Homology groups of just a point. As shown in Figure 2, X 3 ' S 1 , where ' denotes homotopy equivalence. With a bit of labour, one can manually verify that X 4 ' S 2 . Further- more, one can see that H 3 (X 5 We leave this to the reader, while confining ourselves to the case Figure 4 we have all forests on 6 vertices. We denote some of the forests by two digits. The numbers over the edges denote the order in which we read the labels, i.e., the bijection / T . It is easy to see that A k =S(T ) is homeomorphic to S k\Gamma2 for all admissible T , and is contractible otherwise. The only nontrivial cases are 41, 47, 48, 51, 55, and 59, all of which can be verified directly. Therefore, the only nontrivial entries of E 1 (Zcoefficients) will lie on the can be computed from the chain complex 0 k=1 k=2, not admissible31 33213 1 k=3, admissible k=3, not admissible k=4, not admissible 3 4k=5, admissible Figure 4. k=5, not admissible From Proposition 6.4 we have here the two-digit strings denote the corresponding forests on Figure 4. Thus e Therefore 6 is the smallest value of n, for which the homology groups H are not free. --R Whitney numbers of geometric lattices Galois connections and the Leray spectral sequence "Handbook of Combinatorics" "First European Congress of Mathematics, Paris 1992" Introduction to compact transformation groups A homotopy complementation formula for partially ordered sets The generalized Dowling lattices Fundamentals of the Theory of Groups Complexes of directed trees Exact couples in algebraic topology I User's Guide To Spectral Sequences Homotopy properties of the poset of nontrivial p-subgroups of a group Higher algebraic K-theory I group actions on subspace arrangements and combinatorics of discriminants --TR
shellability;homology groups;posets;spectral sequences;graphs
594153
Fast Subsumption Checks Using Anti-Links.
The concept of anti-link is defined (an anti-link consists of two occurrences of the same literal in a formula), and useful equivalence-preserving operations based on anti-links are introduced. These operations eliminate a potentially large number of subsumed paths in a negation normal form formula. Those anti-links that directly indicate the presence of subsumed paths are characterized. The operations have linear time complexity in the size of that part of the formula containing the anti-link.The problem of removing all subsumed paths in an NNF formula is shown to be NP-hard, even though such formulas may be small relative to the size of their path sets. The general problem of determining whether there exists a pair of subsumed paths associated with an arbitrary anti-link is shown to be NP-complete. Additional techniques that generalize the concept of pure literals are introduced and are also shown to eliminate redundant subsumption checks. The effectiveness of these techniques is examined with respect to some benchmark examples from the literature.
Introduction The logical consequences of a ground formula, expressed as minimal implied clau- ses, are useful in certain approaches to non-monotonic reasoning (Kean and Tsi- knis, 1992; Przymusinski, 1989; Reiter and de Kleer, 1987), where all consequences of a formula set (e.g., the support set for a proposed commonsense conclusion) are required. Minimal conjunctions that imply a formula are useful in situations where satisfying models are desired, as in error analysis during hardware verification. Such minimal implied clauses are the formula's prime implicates, and the minimal conjunctions that imply it are its prime implicants. Many algorithms have been proposed to compute the prime implicates of propositional boolean formulas. Most algorithms (de Kleer, 1992; Jackson and Pais, 1990; Jackson, 1992; Kean and Tsiknis, 1990; Slagle et al., 1970) assume that the input is either in conjunctive normal form (CNF) or in disjunctive normal form (DNF). The algorithm of (Ngair, 1993) requires the input to be a conjuncti- This research was supported in part by National Science Foundation Grant CCR-9101208 (Ramesh and Murray) and by Deutsche Forschungsgemeinschaft within the Schwerpunktpro- Deduktion (Beckert and H-ahnle). Some of the results in this paper appear in condensed form in the Proceedings of the 5 th International Conference on Logic Programming and Automated Reasoning, Kiev, Ukraine, July 16-21, 1994, and in the Proceedings of the National Conference on Artificial Intelligence, Seattle, WA, July 31-August 4, 1994. on of DNF formulas. In (Ramesh and Murray, 1993) we propose a set of techniques for finding the prime implicates of formulas in negation normal form (NNF). Our techniques are based on dissolution, an inference rule introduced in (Murray and Rosenthal, 1987b), and on an algorithm called Pi. We have discovered classes of formulas for which our techniques are polynomial but for which any CNF/DNF- based technique must be exponential in the size of the input. Ngair has also introduced similar examples; however, our method is more general than Ngair's which is based on order theory (Ngair, 1993). Coudert and Madre (1992) have also developed an algorithm for computing prime implicates and implicants of formulas in which binary decision diagrams (BDDs) are employed. Although we use prime implicate/implicant generation as an example to demonstrate the utility of anti-link operations, anti-link operations can be employed in any application that requires eliminating subsumed paths in an NNF formula. We have successfully used anti-link operations in a diagnosis system (Ramesh and Murray, 1995). In (Ramesh and Murray, 1993) we describe the Pi algorithm; there, Pi is used to enumerate all the prime implicates of a full dissolvent, an NNF formula that has no conjunctive links (defined later). Pi repeatedly does subsumption checks to keep intermediate results as small as possible. However these checks are expensive. Many result in failure, and they have to be done on sets which can be exponentially large. The time required for these operations can be reduced by using a more compact representation of the intermediate results (de Kleer, 1992), but avoiding as many such checks as possible is the focus of this paper. We show that the full dissolvent can be restructured before applying Pi such that many non-prime implicates are removed without doing subsumption checks at all. We define disjunctive and conjunctive anti-links 1 in NNF formulas, and we identify operations to remove such anti-links and their associated subsumed paths. This leaves fewer subsumption checks for the Pi algorithm. In the next section we describe our path semantics viewpoint and our graphical representation of formulas in classical logic. In Section 3 we introduce anti-links and develop useful equivalence-preserving operations based on them. In Section 4, complexity issues are discussed and some NP-completeness results are proven. Section 5 introduces further techniques based on strictly pure subformulas. The effectiveness of our techniques on certain benchmark formulas described by Ngair (1993) is explored. In Section 6 we introduce strictly pure full blocks and use them to develop a method that reduces the number of subsumption tests required. 1 Anti-links and some associated operators were first proposed by Beckert and H-ahnle - personal communication. The first motivation for studying anti-links arose in connection with regular clausal tableau calculi (Letz et al., 1992). The anti-link rule as it will be defined later can be viewed as an implementation of the regularity condition in (Letz et al., 1992) for the propositional non-clausal case (Letz et al. considered the first-order clausal case). There, refinements of general inference rules are considered, whereas the anti-link rule allows implementation as a preprocessing step. Fast Subsumption Checks Using Anti-Links 3 Foundations: Facts on Formulas in Negation Normal Form We assume the reader to be familiar with the notions of atom, literal, and formula from classical logic. We consider only formulas in negation normal form (NNF): The only connectives used are conjunction and disjunction, and all negations are at the atomic level. This restriction is reasonable, since formulas that contain negations and other operators at any level can be converted to NNF in polynomial time. In this section, we introduce a number of technical terms and definitions that are treated in detail in (Murray and Rosenthal, 1993). They are required for the development of the anti-link operations defined in Section 3, and they make the paper self-contained even for readers not familiar with dissolution. 2.1 Semantic Graphs Semantic graphs are a graphical representation of NNF formulas: Definition 1. A semantic graph consists either of 1. one of the constants true and false, 2. a literal A or A, 3. a c-arc, which is a conjunction of two semantic graphs, or 4. a d-arc, which is a disjunction of two semantic graphs. We use the notation (X; Y ) c for the c-arc from X to Y and similarly use (X; Y ) d for a d-arc; the subscript may be omitted when no confusion is possible. Each semantic graph used in the construction of a semantic graph G is called an explicit subgraph of G. If subgraph of G if X (Y ) is not a c-arc; otherwise the fundamental subgraphs of X (Y are fundamental subgraphs of G. Similarly if a fundamental subgraph of G if X (Y ) is not a d-arc, otherwise the fundamental subgraphs of X (Y ) are fundamental subgraphs of G. The set of nodes of a semantic graph G consists of all literal occurrences used in its construction; the same holds for the set of c-arcs of G and the set of d-arcs of G; i.e., these sets include the nodes, c-arcs, and d-carc, respectively, occurring in the explicit subgraphs of G. In the following, we identify a semantic graph G and the formula it represents 2 ; essentially, the only difference between the semantic graph and the formula is the point of view, and we will use either term depending upon the desired emphasis. For a more detailed exposition, see (Murray and Rosenthal, 1993). true, false, and positive literals represent themselves; a negative literal A represents :A; 4 Anavai Ramesh et al. In addition, we identify a semantic graph and the triple consisting of its set of nodes, its set of c-arcs, and its set of d-arcs. The only exception, were this latter identification is not possible, because it would be ambiguous, are the semantic graphs true and false (both correspond to (;; ;; ;)). Note, however, that when a semantic graph contains occurrences of true and false, the obvious truth-functional reductions apply. Unless otherwise stated, we will assume that semantic graphs are automatically so reduced. In pictorial representations, c-arcs and d-arcs are indicated by the usual symbols for conjunction and disjunction; the arguments of a c-arc are placed vertically above each other, the arguments of a d-arc horizontally besides each other. Example 1. Below, the formula E) - (:A - (B - C)) (1) is displayed as a semantic A Y (2) The boxes in (2) show the explicit subgraphs used in the construction of the semantic graph (since c-arcs and d-arcs are associative and commutative we do not show the explicit subgraphs in subsequent pictorial representations). Definition 2. If A and B are nodes in a graph, and if (X; Y ) ff is an arc or d) with A in X and B in Y , we say that is the arc connecting A and B, and that A and B are ff-connected. Example 2. In (2), C is c-connected to each of B, A, C, D, and E, and is d- connected to A. Definition 3. Let G be a semantic graph. A partial c-path through G is a set of nodes such that any two are c-connected, and a c-path through G is a partial c-path that is not properly contained in any partial c-path. (Partial) d-paths are defined accordingly using d-arcs instead of c-arcs. '(p) denotes the set of literals of a path p. Fast Subsumption Checks Using Anti-Links 5 Example 3. Below, the semantic graph (2) is shown with lines indicating its c- paths (on the left) and its d-paths (on the right): A A The c-paths are fC; A; Ag, fC; A; B; Cg, fD; Ag, fD; B; Cg, fE; Ag, fE; B; Cg; the d-paths are fC; D; Eg, fA; D; Eg, fA; Bg, fA; Cg. The following lemma is obvious. Lemma 1. Let G be a semantic graph. Then an interpretation I satisfies (falsi- fies) G iff I satisfies (falsifies) every literal on some c-path (d-path) through G. 2.1.1 Subgraphs We will frequently find it useful to consider subgraphs of a semantic graph that are not explicit. Definition 4. Given a semantic graph G and a non-empty subset N of the nodes of G, the subgraph of G that corresponds to N is that part of G that consists of nodes from N , where the logical structure of that part is preserved. denotes the subgraph of G corresponding to the set of nodes of G that are not in N . Two subgraphs H and H 0 of G meet each other if they have nodes in common. A non-empty subset N of nodes corresponds unambiguously to one subgraph of G. The empty set corresponds to both true and false; true and false are sub-graphs of all semantic graphs. For a more precise definition of subgraphs, see (Murray and Rosenthal, 1993). Example 4. Below the subgraph of (2) is shown that corresponds to the node set g. A - D A 6 Anavai Ramesh et al. 2.1.2 Blocks The most important subgraphs are the blocks: Definition 5. A c-block H is a subgraph of a semantic graph G with the property that any c-path p that includes at least one node from H passes through H , where p passes through H iff the subset of p consisting of nodes of H is a c-path through H . d-blocks are accordingly defined using d-paths. Example 5. In (2), the subgraph corresponding to the node set fA; D; E;A; Cg is a c-block. However, it is not a d-block since the d-path fA; Bg restricted to the subgraph is fAg, which is a proper sub-path of fA; Cg in the subgraph. Definition 6. A full block is a subgraph that is both a c-block and a d-block. One way to envision a full block is to consider conjunction and disjunction as n-ary connectives. Then a full block is a subset of the arguments of one connective, i.e., of one explicit subformula. Full blocks may be treated as essentially explicit subgraphs (up to the order of arguments), and the Isomorphism Theorem from (Murray and Rosenthal, 1987a) assures us that they are the only structures that may be so treated. Example 6. In (2), the subgraph corresponding to fC; A; Eg is a full block. It can be written as (fC; Ag; E) d ; i.e., we can regard the upper part of the graph as (fC; A; Eg; D) d . The fundamental subgraphs of the upper disjunction are and the literals D and E. Definition 7. Let H be a full block; H is a conjunction or a disjunction of fundamental subgraphs of some explicit subgraph M . If the final arc of M is a conjunction, then we define the c-extension of H to be M and the d-extension of H to be H itself. The situation is reversed if the final arc of M is a d-arc. We use the notation CE(H) and DE(H) for the c- and d-extensions, respecti- vely, of H . Example 7. In (2), In this paper, we compute c- and d-extensions of single nodes only. Single nodes are always full blocks and so testing for this property will be unnecessary. If we assume that formulas are represented as n-ary trees, computing these extensions can be done in constant time; we merely determine whether the given node's parent is a conjunction or a disjunction, and the appropriate extension is then either the node itself or the parent. Fast Subsumption Checks Using Anti-Links 7 2.2 Path Dissolution Path dissolution (Murray and Rosenthal, 1993) is an inferencing mechanism for classical logic that has several interesting properties. It is an efficient generalization of the method of analytic tableaux, is strongly complete in the propositional case, and can produce a list of satisfying interpretations of a formula. The latter feature is particularly valuable in this or in any setting in which one wishes to make use of satisfying interpretations rather than merely to determine whether any exist. Path dissolution works by selecting a link and restructuring the formula so that all paths through the link are eliminated. The nature of the restructuring is such that one cannot rely on CNF (conjunctive normal starts out in CNF, a single dissolution step produces an unnormalized formula. One consequence of eliminating all paths through a link is strong completeness: Any sequence of dissolution steps will eventually create a linkless formula. The paths that remain may be interpreted as models (satisfying interpretations) of the formula. Definition 8. A c-link is a complementary pair of c-connected nodes; d-connec- ted complementary nodes form a d-link. Unless stated otherwise, we use the term link to refer to a c-link. Path dissolution is in general applicable to collections of links; here we restrict attention to single links. Example 8. Consider the link fA; Ag in (2). Then the entire graph is the smallest full block containing the link. Definition 9. Let X be a semantic graph and H an arbitrary subgraph. 3 The c-path complement of H with respect to X , written CC(H;X), is the subgraph of X consisting of all literals in X that lie on c-paths that do not contain nodes from H . If no such literal exists, CC(H;X)= false. The c-path extension of H with respect to X , written CPE(H;X), is the sub-graph of X containing all literals that lie on c-paths that pass through H . If no such literal exists, In the development of anti-link operations, we will use operations that are the duals of CC and CPE. We use DC for the d-path complement and DPE for the d-path extension operators. Their definitions are straightforward by duality, Because we consider only single link dissolution, the first arguments of CC and CPE will be literals when these operators are used in the construction of dissolvents in the examples below. However, this is not the case in Section 3, and hence the above definitions of these operators are in full generality. usually is but does not have to be a subgraph of X. 4 Note, that CPE has two arguments whereas CE (Def. 7) has but one; intuitively, CE has an implicit second argument that is always the entire graph in which the explicit argument occurs. 8 Anavai Ramesh et al. Example 9. In (2), E) The above definitions of the operators CC and CPE are adequate for the definition of dissolution. However, (equivalent) more constructive definitions are given in Section 3, where they will be required in proving the correctness of the anti-link operations introduced there. The reader is referred to (Murray and Rosenthal, 1993) for the proofs of the lemmas below. Lemma 2. Let H be an arbitrary subgraph of G. The c-paths of CPE(H;G) are precisely the c-paths of G that pass through H . Corollary 1. CPE(H;G) is exactly the subgraph of G relative to the set of nodes that lie on c-paths that pass through H . Lemma 3. Let H be an arbitrary subgraph of G. The c-paths of CC(H;G) are precisely the c-paths of G that do not pass through H . Corollary 2. CC(H;G) is exactly the subgraph of G relative to the set of nodes that lie on c-paths that do not pass through H . Lemma 4. If H is a c-block, then CC(H;G)- CPE(H;G) and G have the same c-paths. The above lemmas and corollaries about CC and CPE all hold in dual form for DC and DPE. Suppose that we have literal occurrences A and A residing in conjoined sub-graphs X and Y , respectively. It is intuitively clear that the c-paths through (X - Y ) that do not contain the link fA; Ag are those through (CPE(A;X)-CC(A; Y plus those through (CC(A; X) Ag be a link, and let be the smallest full block containing H . DV (H; M ), the dissolvent of H in M , is defined as follows: If H is a single c-block, then DV (H; Otherwise (i.e., if H consists of two c-blocks), Fast Subsumption Checks Using Anti-Links 9 The only way that H can be a single c-block is if H is a full block (it is trivially a d-block). In that case, and A and A must be (up to commutations and reassociations) arguments of the same conjunction. The following proposition follows from the corollaries and Lemma 4: Proposition 1. Either of the two more compact graphs shown below has the same c-paths as DV (H; M ), and may thus be used instead: Y The semantic graphs from the above proposition are not identical to DV (H; M) as graphs, but they do have the identical c-paths: all those of the original full block except those of CPE(A;X)- CPE(A; Y ), i.e., except those through the link. Example 10. If we dissolve on the link fA; Ag in (2) (using the compact form (4) of dissolution from Proposition 1), the graph that results is: A Theorem 1. Let H be a link in a semantic graph G, and let M be the smallest full block containing H . Then M and DV (H; M) are logically equivalent. A proof of Theorem 1 (in a more general form) can be found in (Murray and Rosenthal, 1993). We may therefore select an arbitrary link H in G and replace the smallest full block containing H by its dissolvent, producing (in the ground case) an equivalent graph. We call the resulting graph the dissolvent of G with respect to H . Since the paths of the new graph are all that appeared in G except those that contained the link, this graph has strictly fewer c-paths than the old one. As a result, finitely many dissolutions (bounded above by the number of c-paths in the original graph) will yield a linkless equivalent graph. This proves: Theorem 2. At the ground level, path dissolution is a strongly complete rule of That means, that the result of applying dissolution repeatedly to an unsatisfiable semantic graph results in the graph false, independently of the choice of the link that is dissolved on at each step. 2.3 Prime Implicates/Implicants We briefly summarize basic definitions regarding implicates. The treatment for implicants is completely dual and is indicated by appropriate dual expressions in parentheses. Definition 11. A disjunction (conjunction) D subsumes another disjunction D 0 (conjunction true (false) is subsumed by all disjunctions (conjunctions). A disjunction is called true iff it is equivalent to true. A conjunction is called false iff it is equivalent to false. Lemma 5. If a disjunction (conjunction) D 0 that is not true (false), then D subsumes A true disjunction (false conjunction) subsumes another true disjunction (false conjunction) only. Definition 12. A disjunction (conjunction) P of literals is an implicate (impli- cant) of a formula G, iff G A disjunction (conjunction) D is a prime implicate (prime implicant) of a formula G iff 1. D is not true (false). 2. D is an implicate (implicant) of G. 3. For all literals A i in D, Gj= (D Note that the set of all prime implicates (implicants) of a formual G, when treated as a CNF (DNF) formula, is equivalent to G. Definition 13. Let D be the set of all prime implicates of a formula G. A prime implicate D of G is essential if D n fDg is not equivalent to G, otherwise D is inessential. 2.4 Fully Dissolved Formulas If we dissolve in a semantic graph G until it is linkless, we call the resulting graph the full dissolvent of G and denote it by FD(G). Observe that FD(G) is dependent on the order in which links are activated. However, the set of c-paths in FD(G) is unique: It is exactly the set of satisfiable c-paths in G. Because FD(G) is link-free, the consequences, i.e., implicates, of G are represented in the d-paths of FD(G). In a dual manner, we may define dissolution for disjunctive links; in that case, FD(G) has no disjunctive links, and the implicants of G are represented in the c-paths of FD(G). These relationships are made precise by Theorem 3 below. Fast Subsumption Checks Using Anti-Links 11 In the discussion that follows, we will often refer to subsumption of d- and c- paths rather than of disjuncts and conjuncts. Paths are defined as sets of literal occurrences, but with regard to subsumption, we consider the literal set '(p) of a path p. In this way, no change in the standard definitions is necessary. The theorem below was proved in (Ramesh and Murray, 1993). Theorem 3. In any non-empty formula in which no c-path (d-path) contains a link, every implicate (implicant) of the formula is subsumed by some d-path (c- path) in the formula. Corollary 3. Every prime implicate (implicant) of a reduced DNF (CNF) for- mula, i.e., one with no false conjuncts (true disjuncts), is subsumed by some d-path (c-path) in the formula. This follows directly from the theorem because such a DNF (CNF) formula has no c-paths (d-paths) with links. In (Ramesh and Murray, 1993), the prime implicates of G are computed by first obtaining FD(G); then, knowing that all implicates are present in the d-paths of FD(G), the Pi algorithm computes -(FD(G)), where is a d-path through F , '(p) is not true; for all d-paths q through When used in this way, Pi extracts all unsubsumed non-tautological d-paths from an NNF formula without c-links. In general, Pi computes -(F ) for an arbitrary 3 Subsumed Paths and Anti-Links Much of the material in this section is a detailed description of the results sketched in (Ramesh et al., 1994). Our goal is to first identify as many subsumed paths as possible in an efficient manner and then eliminate them. The presence of anti- links (both disjunctive and conjunctive) in a graph may indicate that subsumed d-paths are present in the graph. We now define anti-links and then discuss ways to identify and remove subsumed paths due to anti-links. Definition 14. If is a d-arc in a semantic graph G and if AX and A Y are nodes (occurrences of literal in X and in Y respectively, then we call a disjunctive anti-link. is a c-arc in G, then we call fAX ; A Y g a conjunctive anti-link. Note, that M is the smallest full block containing the anti-link. The following theorem relates subsumed paths to anti-links. The theorem is immediate for CNF formulas; there is an obvious dual theorem regarding subsumed c-paths that is immediate for DNF formulas. 12 Anavai Ramesh et al. Theorem 4. Let G be a semantic graph in which a d-path p is subsumed by a distinct non-tautological d-path p 0 in G. Then G contains either a disjunctive anti-link or a conjunctive anti-link. Proof. There are two distinct possibilities, either Suppose there must be a literal having two different occurrences. These two occurrences must be either d- or c-connected and thereby constitute either a disjunctive or a conjunctive anti-link. Suppose '(p) oe '(p 0 ). The proof is by induction on the structure of G. Basis: G is a literal. The result is vacuously true since there cannot be two distinct d-paths through G. Induction step: (a) Suppose are both from the same explicit subgraph (X or Y ) or from different explicit subgraphs. If they lie in the same explicit subgraph then the result follows directly from the induction hypothesis. If they are from different subgraphs then every literal in '(p 0 ) occurs at least once in X and at least once in Y . Any two such occurrences of some literal in '(p 0 ) constitute a conjunctive anti-link. (b) Suppose Y be the restriction of p to X and to Y respectively. Let p 0 Y be the restriction of p 0 to X and to Y respectively. Since p and p 0 are distinct, either p X and p 0 X must be distinct, or p Y and p 0 Y must be distinct (or both), so assume without loss of generality that p X and p 0 are distinct. If either p X subsumes p 0 X or vice versa, then by the induction hypothesis, must have an anti-link and so does G. On the other hand if p X and X do not subsume each other, then there must be some literal (say L) in '(p 0 which is not in p X . But since p 0 subsumes p, there must be an occurrence of L in p Y . The two occurrences of L, one in p 0 X and the other in p Y , constitute a disjunctive anti-link. Unfortunately, the presence of anti-links does not imply the presence of subsumed paths, and hence the converse of the above theorem is not true. 3.1 Redundant Anti-links We now identify those disjunctive anti-links which do imply the presence of subsumed paths. Definition 15. A disjunctive anti-link fAX ; A Y g with respect to the graph G is redundant if either CE(AX ) 6= A or Fast Subsumption Checks Using Anti-Links 13 Definition 16. Let fAX ; A Y g be a disjunctive anti-link in graph G, where is the smallest full block containing the anti-link. is the set of all d-paths of M which pass through both and A Y or through both and AX . Example 11. Consider the following graph Y A Y The two occurrences of A form a disjunctive anti-link. Because contains the d-path (indicated by a line). But since CE(AX there are no paths through is the only member of DP (fAX ; A Y g; G). The anti- link is redundant, and p is subsumed by p (with literal set fA; Cg). Notice that had G been embedded in a larger graph G 0 , every d-path q containing p in G 0 would be subsumed by a corresponding d-path q 0 that differs from q only in that q 0 contains p 0 instead of p. In general, one or both of the literals in a redundant anti-link fAX ; A Y g is an argument of a conjunction, and DP (fAX ; A Y In the above example, the two occurrences of C are both arguments of disjunctions, and thus comprise a non-redundant anti-link for which DP (fCX ; C Y Although only redundant disjunctive anti-links contribute directly to subsumed d-paths, non-redundant anti-links do not prohibit the existence of subsumed paths. However, such non-redundant anti-links do not themselves provide any evidence that such paths are in fact present. Theorem 5. Let fAX ; A Y g be a redundant disjunctive anti-link in a semantic graph G. Then each d-path in DP (fAX ; A Y g; G) is properly subsumed by a d- path in G that contains the anti-link. Proof. Recall that a d-path (c-path) in a graph G is said to pass through a sub-graph X of G if the path when restricted to the set of nodes in X forms a d-path (c-path) in X . Let p 2 DP (fAX ; A Y g; G), and assume without loss of generality that p passes through both CE(AX and A Y . Note that CE(AX is non-empty and that is the largest full block containing the anti-link. We may write CE(AX ) as are p restricted to X and to Y , respectively, and p o is p restricted to nodes outside of both X and Y . By construction, AX 62 p X and thus p X passes through some C restricted to C i , and hence . The d-path p 0 14 Anavai Ramesh et al. 3.2 An Anti-Link Operator The identification of redundant disjunctive anti-links can be done easily by checking to see if either CE(AX ) 6= AX or After identifying a redundant anti-link, it is possible to remove it using the disjunctive anti-link dissolvent (DADV) operator defined below; in the process, all d-paths in DP (fAX ; A Y are eliminated, and the two occurrences of the anti-link literal are collapsed into one. Definition 17. Let fAX ; A Y g be a disjunctive anti-link and let the smallest full block containing the anti-link. Then Example 12. Consider again the semantic graph (5) from Example 11. We have C), so the upper conjunct in DADV is E). For the middle conjunct, this conjunct is (B - A). Finally in the lower conjunct, DPE(AX and false, so this reduces to C). The result is: We point out that although DADV produces a CNF formula in the above simple example, in general it does not. In particular, the above graph can be simplified as the consequence of easily recognizable conditions, and the resulting graph is not in CNF. For the details, see Case 1 of Section 3.5. 3.3 Extension and Path Complement Operators A number of more primitive operators are used in the definition of DADV ; they are described in (Murray and Rosenthal, 1993) and have been defined in Section 2. We present equivalent constructive descriptions here in order to prove Lemma 6 below, and, in the next subsection, to verify that DADV has the desired properties. Fast Subsumption Checks Using Anti-Links 15 Proposition 2. Let G be a semantic graph and H an arbitrary subgraph. Then false if H does not meet G the final arc of G is a d-arc if the final arc of G is a c-arc true if H does not meet G the final arc of G is a c-arc if the final arc of G is a d-arc G if H does not meet G false if the final arc of G is a d-arc if the final arc of G is a c-arc G if H does not meet G true the final arc of G is a c-arc if the final arc of G is a d-arc are the fundamental subgraphs of G that meet H , and F i are those that do not. Lemma 6. If G is a graph and A is a literal occurrence in G, then CC(A;G) is logically equivalent to Proof. We prove the lemma by showing that the formula on the left and the formula on the right possess exactly the same set of d-paths; the result then follows from Lemma 1. The proof is done via induction on the syntactic structure of G (the lemma trivially holds if true or 1. If G is a literal, then and both the set of d-paths of CC(A;G) and the set of d-paths of are empty. Note, that 2. If loss of generality assume A belongs to X . Hence By the induction hypothesis, the d- paths of CC(A;X) are just those of (DPE(A;X) \Gamma fAg)-DC(CE(A);X). So CC(A;G) has the same d-paths as (DPE(A;X) \Gamma fAg)-DC(CE(A);X)-Y . Now consider the right hand side of the equation. Since A is in X , Therefore, be disjoint from Y , and thus Therefore we can write the right hand side of the equation as (DPE(A;X) \Gamma factoring out the subgraph Y we get an equivalent subgraph having the same d-paths. But this is just the semantic graph that has been shown to have the same d-paths as the left hand side. 3. Finally suppose again assume that A is in X . Now there are two subcases to consider. a) If false and thus has no d-paths, then A in X is not d- connected to any other subgraph in X . Hence X is of the form A-C 1 - and A. As a result, both DC(CE(A);G) and d-paths. (b) If CC(A;G) 6= false, then and thus CC(A;X) 6= false. Therefore, by the induction hypothesis, CC(A;G) has the same d-paths as Focusing now on the right hand side of the equation, DPE(A;X) by definition. The c-extension of A can only include nodes from X (otherwise, contrary to the subcase (b) condi- tion). Therefore, . Therefore the right hand side of the equation has the same d-paths as (DPE(A;X) \Gamma fAg)- (DC(CE(A);X)-Y ). This is just the result obtained for the left hand side in this subcase. 3.4 Correctness of DADV In Theorem 6 below we show that DADV (fAX ; A Y g; G) is logically equivalent to G and does not contain the d-paths of DP (fAX ; A Y g; G). Theorem 6. Let be the smallest full block containing fAX ; A Y g, a disjunctive anti-link in semantic graph G. Then DADV (fAX ; A Y g; M) is equivalent to M and differs in d-paths from M as follows: d-paths in DP (fAX ; A Y are not present, and any d-path of M containing the anti-link is replaced by a path with the same literal set having only one occurrence of the anti-link literal. Fast Subsumption Checks Using Anti-Links 17 Proof. Note that AX and A Y are literal occurrences (and hence d-blocks) in X and in Y respectively. By the dual of Lemmas 4, X is equivalent to DC(AX ; X)- X), and from the distributive law, M is equivalent to Similarly, Y is equivalent to we expand the upper occurrence of Y and distribute. Thus, M is equivalent to By the duals of Lemmas 2 and 3, not only have we rewritten M equivalently, but the d-paths of M have been preserved. We will continue to rewrite M ; our goal is to eventually put it in an equivalent form in which the d-paths of DP (fAX ; A Y have been omitted. Consider the d-paths of DC(AX ; X) - the d-paths in X that miss AX . They either miss CE(AX ), the c-extension of AX , or pass through CE(AX g. Hence DC(AX ; X) has the same d-paths as By replacing the lower occurrence of DC(AX ; X) in the previous graph, we get the following graph M 0 which is equivalent to M and has the same d-paths as Every d-path in the subgraph DPE((CE(AX Theorem 5, all these paths are subsumed by other d- paths. Therefore, we can remove the subgraph DPE((CE(AX preserving equivalence to get the graph M 00 shown below. Again by using arguments dual to the one given earlier for X , we have that Y and have identical d-paths. Replacing Y in M 00 , we find that every d-path in the subgraph DPE(AX ; X)- Again by Theorem 5, these paths are also subsumed by other d-paths. Therefore we can remove the subgraph preserve equivalence; M 000 results. The d-paths in M 000 are those of M excluding the d-paths in DP (fAX ; A Y Consider now the d-paths of DPE(AX . They are exactly those of M (and of M 000 ) that contain the anti-link: They each contain two occurrences of the literal A. Hence we can remove the node A Y from Fast Subsumption Checks Using Anti-Links 19 and apply Lemma 6 to get M 0000 . Applying Lemma 6 to M 0000 we get DADV (fAX ; A Y In constructing DADV (fAX ; A Y g; M)we have removed only subsumed d-paths and altered only d-paths that contain the anti-link by collapsing the double occurrence of the anti-link literal. Hence DADV (fAX ; A Y g; M) is equivalent to M , does not contain the anti-link, and does not contain any d-path of DP (fAX ; A Y Theorem 6 gives us a method to remove disjunctive anti-links and some subsumed d-paths: Simply identify a redundant anti-link and the smallest full block M containing it, and then replace M by DADV (H; M ). The cost of this operation is proportional to the size of the graph replacing M , and this is linear in M . Also, c-connected literals in M do not become d-connected in truly new disjunctive anti-links are not introduced. Howe- ver, parts of the graph may be duplicated, and this may give rise to additional copies of anti-links not yet removed. Nevertheless, persistent removal of redundant disjunctive anti-links (in which case DP (fAX ; A Y g; M) 6= ;) is a terminating process, because the number of d-paths is strictly reduced at each step. This proves: Theorem 7. Finitely many applications of the DADV operation on redundant anti-links will result in a graph without redundant disjunctive anti-links, and termination of this process is independent of the choice of anti-link at each step. Although we can remove all the redundant disjunctive anti-links in the graph, this process can introduce new conjunctive anti-links. Such anti-links may indicate the presence of subsumed d-paths, but the situation is not as favorable as with disjunctive anti-links - see Section 3.7. Simplifications Obviously, DADV (fAX ; A Y g; M) can be syntactically larger than Under certain conditions we may use simplified alternative definitions for DADV . These definitions result in formulas which are syntactically smaller than those that result from the general definition. The following is a list of possible simplifications. 20 Anavai Ramesh et al. 1. If then DC(CE(AX Therefore by (possibly non atomic) factoring on DC(AX ; X) and observing that has the same d-paths as Y , DADV (fAX ; A Y g; M) becomes It turns out that this rule applies to (2) in Example 1. Since CE(AX the simplified rule for this case results in the following graph. A 2. If then DC(CE(AX Hence 3. If both Case 1 and Case 2 apply, then CE(AX and the above formula simplifies to Note that in all the above versions of DADV , the roles of X and Y can be interchanged. Fast Subsumption Checks Using Anti-Links 21 3.6 Disjunctive Anti-Links and Factoring It is interesting to note that the DADV operation contains factoring (i.e., the ordinary application of the distributive law to a pair of conjunctions containing a common argument) as a special case. This is just the condition for Case 2 above except that both CE(AX A This is the graph obtained by disjunctive factoring (Murray and Rosenthal, 1993). The DADV operator also captures the absorption law (or merging). If AX and A Y are both arguments of the same disjunction, then Note, however, that technically the anti-link is not redundant in this case. 3.7 Conjunctive Anti-Links There are conjunctive anti-links that always indicate the presence of d-paths that are subsumed by others, and they are easy to detect. However, the conditions to be met are much more restrictive than those for redundant disjunctive anti- links. Consider a conjunctive anti-link fAX ; A Y g, where the smallest full block M containing the anti-link is (AX Every d-path in Y which passes through A Y will be subsumed by the d-path consisting of the single literal AX . Hence we can replace Y by This is a kind of dual to Case 3 of the simplified versions of DADV discussed earlier. There, the anti-link fAX ; A Y g is disjunctive and . The simplified DADV operation just replaces Y by Note that the conjunctive anti-link operation above removes subsumed d-paths, whereas the Case 3 disjunctive anti-link operation can either remove paths or merely remove the second occurrence of the anti-link literal on paths that contain the anti-link. Both operations involve d-paths, and both have strictly dual operations that would affect c-paths instead. 4 Complexity Considerations The problem of eliminating all subsumed paths in a graph in an efficient manner does not seem feasible. The following definition makes precise the notion of minimality with respect to subsumed d-paths. Then we show that it is NP-hard to achieve this property. Definition 18. Let G be a semantic graph; we say that a graph G 0 is a d-minimal equivalent of G if it satisfies the following conditions. 22 Anavai Ramesh et al. 1. G is logically equivalent to G 0 . 2. If p 0 and q 0 are two distinct d-paths in G 0 , then p 0 does not subsume q 0 and vice versa. 3. If p 0 is a d-path in G 0 , then there is a d-path p in G such that, 4. If p is a minimal d-path in G, then there is a d-path p 0 in G 0 such that The c-minimal equivalent of a graph is defined in the obvious dual way. Note that Property 1 above is implied by Properties 3 and 4, and that G 0 needs not be unique. However, the d-paths of G 0 will always include all essential (and possibly some inessential) prime implicates of G. Computing d-minimal equivalent graphs efficiently would be helpful for finding prime implicates. In a d-minimal equivalent graph of a full dissolvent, subsumption checks can be completely eliminated by Property 2 above. Hence to find the prime implicates of G, we can find a d-minimal equivalent G 0 of the full dissolvent FD(G), and then simply enumerate the d-paths of G 0 . A d-minimal equivalent of a given graph G can be trivially obtained by first enumerating all the d-paths of the given graph G and then eliminating all the subsumed d-paths. The above algorithm is exponential in the size of G, because G 0 is being constructed in CNF. However an NNF d-minimal equivalent G 0 of G may be small compared to a CNF d-minimal equivalent. Even so, the problem is NP-hard (proof follows) and hence is not likely to have an efficient algorithm. Theorem 8. The following problem (elimination of subsumed paths) is NP-hard. Given a graph G, find a d-minimal equivalent graph G 0 . Proof. To show NP-hardness we reduce from satisfiability of CNF formulas. Let C be an instance of the CNF satisfiability problem and fX be the set of variables in C. Let A; X 0 n be distinct variables not occurring in g. Let D be the semantic graph obtained by replacing X i by X 0 (by which we denote the NNF of the negation of C). We construct the following semantic graph G. A - D Fast Subsumption Checks Using Anti-Links 23 The size of the graph G is no more than a constant factor of the size of C and can therefore be constructed in linear time. It is easy to see that any d-path which includes the literal A must pass through D and vice versa. Let G 0 be any graph which is a d-minimal equivalent to G. We will show that C is satisfiable iff the literal A occurs in G 0 . Suppose C is satisfiable; then :C is falsifiable, and there are d-paths (in fact, clauses, since :C is in DNF) in :C that do not contain any disjunctive link g. All such d-paths through D do not contain any fX at least one such, say p, is not subsumed by another d-path through D. The d-path pA cannot be subsumed by any other d-path in G and hence there will be a path in G 0 which has the same literal set as pA. Hence the literal A must occur in G 0 . If C is not satisfiable, then :C is valid. Therefore for every d-path p in D (and hence every d-path through A), there is some i, 1 - i - n, such that the pair of literals fX '(p). But every such pair of literals forms a d-path in G and hence every d-path containing A will be subsumed by another d-path in G. Furthermore the subsuming path will not contain the literal A. By definition of d-minimal equivalent and by construction of G 0 , no d-path in G 0 can contain the literal A, and thus the literal A cannot occur in G 0 . If we can solve the elimination of subsumed paths problem in polynomial time then we have the following algorithm which can solve the satisfiability of CNF in polynomial time: Given any instance C of the CNF satisfiability problem, we can construct in polynomial time the graph G as shown earlier. We then find the graph G 0 using the algorithm for elimination of subsumed paths. The size of G 0 will be polynomial in the size of G (since computing it required only polynomial time). Now C is satisfiable iff the literal A occurs in G 0 and this check can be done in polynomial time. By a completely dual construction we obtain the following corollary. Corollary 4. Given a graph G, finding a c-minimal equivalent of G is NP-hard. We have seen that the general problem of computing d- or c-minimal graphs is NP-hard. Nevertheless, redundant disjunctive anti-links are easily recognized, and eliminating their corresponding subsumed d-paths can be done without direct subsumption checks. On the other hand, recognizable subsumed d-paths due to conjunctive anti-links are not likely to be as plentiful due to the strong restriction defining such useful anti-links. It is also difficult to find out if an arbitrary conjunctive anti-link results in subsumed d-paths. In fact, this problem is NP-complete. Theorem 9. The following problem is NP-complete. Given a conjunctive anti- link in a graph G, determine whether there are there two d-paths p X and p Y in G, such that p X passes through AX and p Y passes through A Y and either p X subsumes p Y or vice versa. Proof. It is easy see that this problem is in NP. To show NP-hardness we reduce from satisfiability of CNF formulas. Let C be an instance of the CNF satisfiability problem and fX be the set of variables in C. Let A; n be distinct variables not occurring in fX g. Let D be the semantic graph obtained by replacing X i by X 0 We construct the following semantic graph G. The subgraph D is a DNF formula, and A 1 and A 2 are two different occurrences of the literal A. These two literal occurrences form a conjunctive anti-link in G. Every d-path through A 2 contains the literal B, and only d-paths containing A 1 can possibly subsume d-paths through A 2 . The size of the graph G is no more than a constant factor of the size of C and can therefore be constructed in linear time. We must show that C is satisfiable iff there is a d-path through A 1 that subsumes another d-path through A 2 . Suppose first that C is satisfiable. Then :C is falsifiable, and there is at least one d-path in :C that does not contain any disjunctive link g. Therefore some d-path p through D does not contain any of the literal pairs fX Recall that '(p) is defined to be the literal set of path p. It is easy to see that since '(p) does not contain such a literal pair (corresponding to a d-link in :C), a d-path p 0 can be chosen that passes through the n rightmost disjuncts in the lower part of G, such that '(p 0 subsumed by A 1 p. To show the if-part, suppose there is some d-path (say p) through A 1 that subsumes another d-path (say p 0 ) through A 2 . Then p cannot contain any literal because no such pair occurs on any d-path in the lower part of G. Therefore p when restricted to D will not have such a literal pair, :C has a d-path without a d-link, and hence C is satisfiable. Corollary 5. The problem dual to the one described in Theorem 9 involving disjunctive anti-links and c-paths is NP-complete. 5 Some Benchmark Examples Ngair (1993) has investigated examples that prove difficult for many proposed prime implicate/implicant algorithms. In this section, we show that Pi links is effective for some of these examples. For other examples from (Ngair, 1993), applying anti-link techniques appears not to produce as significant an impro- vement. We develop an additional technique based on strictly pure full blocks that results in a dramatic improvement for these latter examples. Fast Subsumption Checks Using Anti-Links In (Ngair, 1993) a class of formulas is proposed for which reliance on an intermediate CNF form can result in an exponential increase in size and hence would be intractable for CNF-based algorithms. Dissolution does poorly for these examples: Although the full dissolvent can be computed quickly, a large number of subsumption checks must be performed by Pi. It turns out, however, that in this case the subsumed implicates correspond to easily recognizable anti-links of both the disjunctive and conjunctive kind. We show that if these anti-links are removed after dissolution is performed, dissolution + Pi can find all the implicates in polynomial time. Ngair's formulas are abbreviated with F n (n - 1) and are defined as: A A 2i In the left part of Figure 1 we show the graph of F n for a fixed n. A 3 A A 4 A 2n A A 2n A 3 A A 4 A 2n A A 2n A 3 A 4 Fig. 1. Semantic graph of Ngair's formulas before and after dissolution. F n has 4n literals, and 2n \Gamma 2 c-links; dissolution can remove these links by performing 2n dissolution steps. The full dissolvent that results is depicted in the right part of Figure 1. The structure of the full dissolvent depends on the order in which links are selected for application of dissolution; the above dissolvent is the one obtained by the current version of our propositional dissolution prover Dissolver. (The 26 Anavai Ramesh et al. compact version (3) from Proposition 1 of the dissolvent is used; X is chosen to be the smallest of the two c-blocks). We can now factor on all the occurrences of both A 1 and A 2 in the upper right hand part of the graph and on the two occurrences of A 1 in the lower left corner. The resulting graph is shown in Figure 2. (Since Dissolver is NNF-based, such factoring is not only feasible but is in fact implemented and routinely employed to produce the final output.) A 3 A A 4 A 2n A A 2n A 3 A 4 Fig. 2. Ngair's formulas after dissolution and factoring. The two occurrences of A 2 at the bottom left part of the graph form a redundant disjunctive anti-link; they can be removed using the special case covered by Rule 1 for disjunctive anti-links. The two occurrences of A 1 on the left hand side of the graph form a conjunctive anti-link and can be removed using the conjunctive anti-link rule. This produces: A A 2n A 3 A 4 By factoring on A 1 and removing the conjunctive anti-link comprised of the two occurrences of A 2 (or by just factoring on A 1 - A 2 ), the above graph reduces Fast Subsumption Checks Using Anti-Links 27 to the prime implicates are just fA 1 g and fA 2 g. To get this graph, factoring and 3 anti-link operations were required - obviously polynomial time. Hence dissolution + removal of anti-links + Pi can handle the above class of problems in polynomial time. Perhaps the most important point is that no subsumption checks whatsoever are required. 6 A Generalized Purity Principle 6.1 Strictly Pure Full Blocks Recall that a full block is essentially an explicit subgraph; it is a subset of the arguments of a conjunction or disjunction, and, via commutations and reassociations, can in fact be made explicit. Definition 19. A subgraph M in a graph G is pure iff all c-links or d-links that meet M at all are totally within M . 6 If, in addition, all conjunctive or disjunctive anti-links that meet M at all are totally within M , we say that M is strictly pure. 7 If M is a full block in G we speak of a (strictly) pure full block . When factored, some of the examples from (Ngair, 1993) contain surprisingly many strictly pure full blocks. Note that both factoring and recognizing strictly pure full blocks are polynomial operations. Intuitively, such full blocks can be replaced by single new variables, and the implicates of the resulting graph bear a strong relationship to those of the original. Of course, the full block in question must be satisfiable (since the new variable certainly is). At first, this may appear to be a heavy penalty. It is not, however, because the prime implicates of the full block itself must be computed anyway. In doing so, its satisfiability is determined as a byproduct. The following theorems characterize the properties of strictly pure full blocks with respect to prime implicates. In them we employ the following notation: let M be an explicit subgraph of a graph G and let X be a variable not occurring in G. By G X M we denote the graph obtained by the substitution of X for M in G. Similarly, if D is a disjunction of literal occurrences from G we denote by DM the disjunction of literals that occur in M and by DG\GammaM the disjunction of literals that do not. Obviously, holds. Finally, we set ae DG\GammaM otherwise Theorem 10. Let M be a satisfiable strictly pure full block in a satisfiable semantic graph G and let D be a non-tautological disjunction of literals from G. If DM 6= false, then the following statements are equivalent: 6 This is just the obvious generalization of the concept of a pure literal as it is used in the literature on CNF-based automated deduction. 7 Simply put, M shares no variables with the rest of G. 28 Anavai Ramesh et al. 1. D is a prime implicate of G. 2. D X M is a prime implicate of G X M , and DM is a prime implicate of M . This theorem turns out to be a special case of Theorem 11 to be proved in the following subsection. If a graph G contains several strictly pure full blocks M then the repeated application of Theorem 10 provides a potentially significant speedup in computing the prime implicates of G: Replace each strictly pure full block M i by a new variable X n) and compute the prime implicates of the resulting graph G X M . Then, all substitutions of prime implicates of M i for X i in the prime implicates of G X M result in prime implicates of G. The speedup is potentially dramatic: Each subsumption test performed within some M i would otherwise be performed once for every d-path in G X M that can be extended through M i to form a d-path in G. Observe that prime implicates of G X containing none of the variables are simply prime implicates of G that do not contain literals from any of the blocks M i . Several distinct strictly pure full blocks can be handled by repeated application of Theorem 10 as explained above. However, multiple occurrences M of a single full block M in G require an extended analysis. The problem is that the multiple occurrences themselves preclude any single M i from being strictly pure, even if M shares no variables with the rest of G. Intuitively, we would expect that by replacing each of the occurrences M i in G by the single new variable X , the prime implicates of the resulting graph G X would also bear a strong relationship to those of G. This is made precise in the next section. 6.2 Multi-Pure Full Blocks Definition 20. Suppose that M are occurrences of full blocks in G and that all of them are syntactically identical (up to associativity and commutativity of disjuncts and conjuncts). The subgraph M formed by taking all the nodes of the blocks M i is not necessarily a full block; 8 but let M be strictly pure. Then we call the M i multi-pure full blocks. In addition, suppose that there are occurrences M n+m of full blocks syntactically identical (up to associativity and commutativity of disjuncts and conjuncts) to M , the NNF of the complement of any of the M i (1 - i - n). We call complementary multi-pure full blocks. Note that each (complementary) multi-pure full block M i is not strictly pure, since it has anti-links (and possibly links) to its equivalent (complementary) full blocks in G. Observe that complementary non-atomic formulas must be recognized. For example, if could be A - B or B - A. In fact, M could 8 As the blocks M i could be single literals occuring in arbitrary positions, this is hardly surprising. Fast Subsumption Checks Using Anti-Links 29 have been input as or as :(:A oe B). If NNF formulas are stored in an appropriate canonical way, complementarity is easily (that is, in polynomial time) detectable; the situation is also straightforward when complementary formulas have the form M and :M prior to conversion to NNF. In any case, a detailed treatment of this issue is beyond the scope of this paper. We do note that in the absence of complements, multi-pure full blocks are recognizable in polynomial time via a canonical NNF representation. A modification of any algorithm for finding common subtrees (see (Grossi, 1993) for one such algorithm) can be used for recognizing multi-pure full blocks. It turns out that the results of Theorem 10 can be extended to the case in which a formula contains complementary multi-pure full blocks. We use the notation of Theorem 10 with the understanding that M denotes any occurrence of an and the M are treated as negated occurrences of M (thus G X M replaces the complementary occurrences M j as well). M is defined as the subgraph of G relative to the M i and are defined as before, but relative to M . Additionally, we define D X M in the obvious way (Intuitively, we use D X M when the literals of DM correspond to unnegated occurrences of M , and we use D X M when the literals in DM correspond to negated occurrences of M ). Theorem 11. Let M n+m be complementary multi-pure full blocks in a satisfiable semantic graph G, where all of the blocks M i and M j are satisfiable (we allow D be a non-tautological disjunction of literals from G. Then the following statements are equivalent: 1. D is a prime implicate of G. 2. false and D is a prime implicate of GX or DM 6= false and D X M is a prime implicate of G X M , and DM is a prime implicate of M or DM 6= false and D X M is a prime implicate of G X M , and DM is a prime implicate of M . Proof. Let G 0 X M be a graph without c-links that is equivalent to G X (for instance, M could be the full dissolvent of G X Similarly, let M 0 be a c-linkless equivalent of M , and M 0 be a c-linkless equivalent of M . Let G 0 be the graph obtained from G by replacing X by M 0 and X by M 0 . It is easy to see that G 0 is equivalent to G but has no c-links. By Theorem 3, every prime implicate of G is present as a d-path in G 0 and every prime implicate of G X M is present as a d-path in G 0 X M . To prove the only-if-part, let D be a prime implicate of G. Then there must be an unsubsumed d-path p in G 0 such that are complementary, p can never meet (and thus pass through) both M i and M j for any i and j. Suppose first that p does not pass through any of the occurrences of M or M . In this case, p must also be a d-path in G 0 X (technically, p is isomorphic to a d- path in G 0 X To prove that p is not subsumed by another d-path in G 0 X M , assume otherwise, namely, that there is a path p 0 in G 0 X M that subsumes p. But p 0 cannot contain X or X, and hence would also be a d-path in G 0 that subsumes p, which is a contradiction. Thus D is a prime implicate of G 0 X M and hence of G X M . Now suppose p passes through the full blocks be the restriction of p to M i j 0 . Notice that for 1 - j, k - q, '(p wise, the d-path obtained by replacing p j by p k in p would subsume p. Similarly, cannot be subsumed by another d-path in M i j 0 . Since p j is an unsubsumed d- path in M i j 0 which is a linkless equivalent of M , of M . Now let p X be the d-path obtained by replacing each of the p j by X in p; X is a d-path in G X M . Furthermore, p X cannot be subsumed by another d-path in G X M (again, such a subsuming path would induce a path in G 0 that subsumes p). Therefore D X prime implicate of G X M . Finally, in the case that p passes through the full blocks argument is similar as in the previous paragraph. To prove the if-part, first suppose DM = false and D is a prime implicate of G X M . Then there is an unsubsumed d-path (say p) in G 0 X M such that contains neither X nor X. Thus p is an unsubsumed d-path of G 0 , and D is a prime implicate of G. Now suppose that D X M is a prime implicate of G X M and that DM is a prime implicate of M . Then there are unsubsumed d-paths p X in G 0 X M and p M in M 0 , respectively, such that '(p X M and '(p M respectively. In the present subcase, by definition, D X contains X (and cannot contain X). Let p 0 be the result of replacing all occurrences of X in p X by p M ; p 0 is a d-path in G 0 . Since both p M and p X are unsubsumed in G X respectively, and since M 0 does not share any variables with the rest of G 0 , p 0 will also be unsubsumed in G 0 . Hence, by Theorem 3, prime implicate of G. Similarly, if D X M is a prime implicate of G X M and DM is a prime implicate of M , then D is a prime implicate of G. It is straightforward to see that in the case when Theorem 11 collapses into Theorem 10. On the one hand we expect to substitute new variables for multi-pure full blocks and achieve a savings in the computation of prime implicates comparable to that provided by Theorem 10. But note that some complementary occurrences of M may be c-connected; this means that such occurrences play a role in whatever inference process is employed prior to computation of the implicates themselves. In Fast Subsumption Checks Using Anti-Links 31 particular, we may treat them as literals and dissolve (indeed, the set of individual links between such full blocks would satisfy the requirements of a multiple link dissolution chain as it is defined in (Murray and Rosenthal, 1993)). That dissolving on two complementary full blocks accomplishes exactly what dissolving on all the corresponding single-links would is clear: All c-paths through both full blocks are eliminated from the graph. But the former operation is much more efficient than the latter. Therefore, recognizing such complementary full blocks and performing inference directly on them, rather than on their constituent literals, is desirable. Note also that for the inference phase of a prime implicate computation, complementary full blocks do not have to be multi-pure full blocks. This condition is necessary only for the extraction of implicates using Theorem 11 once all implicates are known to be present. Finally, the remarks above apply also to identical full blocks if they form appropriate non-atomic anti-links as discussed in Section 3. 6.3 More Examples Kean & Tsiknis (1990) provide a class of examples referred in the following to as Knm . They have mn+1 input CNF clauses and (m+1) n +mn prime implicates. This set of clauses can be factored to obtain a more compact representation in NNF as shown in Figure 3. Snm Fig. 3. Semantic graph of Kean & Tsiknis's formulas. Since the number of prime implicates is exponential, so is the number of subsumption checks required. The number of subsumption checks for the Ipia (de Kleer, 1992) and Gen-Pi (Ngair, 1993) algorithms are shown in Table I. Examples Ipia Gen-Pi Pi K33 5166 972 164 K44 506472 11600 887 K54 1730120 29074 887 I Number of subsumption checks needed for Kmn by Ipia, Gen-Pi, and Pi For each i, the literals S a full block M i , and all literals in it are strictly pure. Let K 0 mn be the graph obtained by replacing each full block M i by a new variable X i . By the corollary of Theorem 10, we can get the prime implicates of Kmn from the prime implicates of K 0 mn . Since each of the subgraphs M i has no c-links, the prime implicates of M i are present as d-paths by Theorem 3. Since they also have no anti-links, by the contrapositive of Theorem 4, neither are subsumption checks required to find these prime implicates. Thus the number of subsumption checks to be done is exactly that required for computing the prime implicates of K 0 mn , and this is significantly less than that needed for Kmn . Note that the number of prime implicates of K 0 mn is only n. For the problems in Table I, we applied the above technique in combination with anti-link operations. For mn , the full dissolvent depends only on n and can be defined recursively. The full dissolvent of K 0 m2 (basis) and of the general case of K 0 are shown in Figure 4. The number of subsumption checks required here is also shown in Table I. Cle- arly, our techniques produce a significant reduction in the number of subsumption checks required. Note that for the problem Kmn , the number of subsumption checks depends only on n and not on m, and is not reduced by applying the anti-link operations to the full dissolvent. Our techniques are not limited to NNF formulas. They can sometimes be used by other algorithms like Ipia and Gen-Pi which are not based on NNF formulas. For example K 0 mn turns out to be in CNF and hence both Ipia and Gen-Pi can handle these formulas, thereby reducing the number of subsumption checks nee- ded. However normal forms like CNF provide very little scope for applying these techniques directly. For example the literals S in the unfactored form of Kmn do not form a full block. Hence one cannot apply Theorem 10. They do form a full block after factoring. This provides stronger evidence that by avoiding less general normal forms like CNF/DNF, one can improve the performance of prime Fast Subsumption Checks Using Anti-Links 33 A n A Fig. 4. The full dissolvent of Km2 (left) and of Kmn , n ? 2 (right). implicate algorithms. We also note that we have implemented some of the anti-link operations presented above in a diagnosis system introduced in (Ramesh and Murray, 1995). As a result, we greatly increased the size of problems solvable on our system. (The anti-link operations used were the special cases that do not increase the size of the 7 Comparison with BDDs BDDs (Bryant, 1986) are commonly used in verification of boolean circuits. Coudert and Madre (Coudert and Madre, 1992) describe an algorithm which produces the prime implicates/implicants of a propositional formula represented as a BDD. Any algorithm including theirs which uses BDDs must perform large amounts of subsumption testing. Given any formula in NNF, a BDD based method would first construct the BDD and then extract the prime implicates/implicants from it. In contrast, our system would first compute the full dissolvent. But for either of these approaches, the next stage - extracting the prime implicates/implicants - requires extensive testing for subsumption. The size of the BDD depends critically on the ordering (of variables) chosen. There are classes of NNF formulas (Breitbart et al., 1990) for which any BDD will be exponentially large in the formula size. These formulas do not have any c-links, so the dissolution phase of our method does not change the formula. Hence the input to Pi would be a small formula, whereas the BDD based method would have to handle an exponentially larger intermediate representation. However these formulas have many prime implicates and prime implicants, and the subsumption checking is the bottleneck for both methods. 34 Anavai Ramesh et al. In all likelyhood, the relative performance of these two methods can be determined through experimental evaluation only; formulas will exist for which one method is superior, and vice versa. 8 Conclusions and Future Work We have introduced anti-links and defined useful equivalence-preserving operations on them. These operations can be employed so as to strictly reduce the number of d-paths in an NNF formula. Unlike path dissolution, which removes unsatisfiable (or tautological, in the dual case) paths, anti-link operations remove subsumed paths without any direct checks for subsumption. This is significant for prime implicate computations, since such computations tend to be dominated by subsumption checks. Although prime implicate/implicant problems are intractable in general, our techniques perform exponentially better than others on certain examples. In addi- tion, we are able to improve performance greatly on the inherently exponential examples of (Ngair, 1993). Some experimental results on a dissolution- and Pi-based system for computing prime implicates are reported in (Ramesh and Murray, 1993). That system is currently being extended; some anti-link operations are already implemented and have shown to improve performance. Operations based on strictly pure full blocks are under development, and their effectiveness in practice will be tested. --R Hunt III Implicit and incremental computation of primes and essential implicant primes of boolean functions. An improved incremental algorithm for computing prime implicants. On finding common subtrees. Computing prime implicants. Computing prime implicants incrementally. An incremental method for generating prime implicants/im- plicates Assumption based reasoning and clause management systems. Inference with path resolution and semantic graphs. Path dissolution: A strongly complete rule of inference. Dissolution: Making paths vanish. A new algorithm for incremental prime implicate generation. An algorithm to compute circumscription. An application of non clausal deduction in diagnosis. On anti-links Foundations of assumption-based truth maintenance systems: preliminary report A new algorithm for generating prime implicants. --TR --CTR A. K. Shiny , Arun K. Pujari, An Efficient Algorithm to Generate Prime Implicants, Journal of Automated Reasoning, v.22 n.2, p.149-170, February 1999
anti-links;subsumption
594161
The Use of Lemmas in the Model Elimination Procedure.
When the model elimination (ME) procedure was first proposed, the notion of lemma was put forth as a promising augmentation to the basic complete proof procedure. Here the lemmas that are used are also discovered by the procedure in the same proof run. Several implementations of ME now exist, but only a 1970s implementation explicitly examined this lemma mechanism, with indifferent results. We report on the successful use of lemmas using the METEOR implementation of ME. Not only does the lemma device permit METEOR to obtain proofs not otherwise obtainable by METEOR, or any other ME prover not using lemmas, but some well-known challenge problems are solved. We discuss several of these more difficult problems, including two challenge problems for uniform general-purpose provers, where METEOR was first in obtaining the proof. The problems are not selected simply to show off the lemma device, but rather to understand it better. Thus, we choose problems with widely different characteristics, including one where very few lemmas are created automatically, the opposite of normal behavior. This selection points out the potential of, and the problems with, lemma use. The biggest problem normally is the selection of appropriate lemmas to retain from the large number generated.
Introduction The Model Elimination (ME) procedure was defined in the 1960's [17, 18] nearly coincident with the Resolution procedure [22]. There was little early This work is supported in part by NSF grants CCR-8900383 and CCR-9116203. experimentation with ME, so the procedure received relatively little attention. (However, it is the basis of the SL-Resolution procedure which played an historical role in the beginning of Logic Programming). A major implementation of ME was completed in the early 1970s by Fleisig et al. [14] that did provide a direct comparison with the Unit Preference - Set-of-Support refinement of Resolution (Wos et al. [29]), as both procedures were implemented with the same underlying inference tools. Individual wins were found for both methods but no evidence to prefer ME over the known refinements of Resolution was forthcoming. The notion of lemma, introduced in the first papers of ME, was implemented by Fleisig et al. with a resulting indifference to their usefulness. The paper included this statement:"The use of lemmas usually was detrimental, although some short refutations were obtained using lemmas. The poor performance is due to the lack of selection rules for lemmas and is aggravated by the depth-first nature of search." ([14], p. 135). This paper will strongly refute the conclusion of that paper regarding the value of lemmas within ME. We report on results that to date are only obtainable within the ME framework by the use of lemmas, and the theorems proven include some not yet provable by Resolution techniques. We now understand that the notion of only direct competition with Resolution is oversimplistic; this procedure tends to do relatively well on non-Horn problems where Resolution methods now employed are not focused. Unlike chess, theorems are a very diverse lot and different proof methods may excel in different areas. As is well-known to many in this research area, the idea of ME was kept alive by the work of Mark Stickel. Most notably, Stickel exploited the fact that ME is an extension of SLD-Resolution (Prolog) to develop the Prolog Technology Theorem Prover (PTTP) [24, 25]. In the late 1980's almost simultaneously three groups took this one step further, building on the architecture (the Warren Abstract Machine (WAM)) developed for Prolog that the logic programming community had extended to parallel machines. These projects were PARTHENON (CMU) [9], PARTHEO (Munich) [23], and METEOR (Duke) [4, 3]. (The Munich [16] and Duke efforts included sequential provers also.) The work reported here has been implemented on one of the METEOR family of ME provers. What has changed in 20 years that makes the lessons of the Fleisig et al. paper invalid in part? A fair number of things: maybe most important is the WAM architecture ideas, but also important are use of iterative deepening (introduced to ME by Stickel), a sophisticated implementation exploiting the WAM ideas, vastly more powerful yet cheaper computers, and careful use of techniques to restrict lemma creation and use. We look in depth at three examples, theorems proven with use of lemmas that would otherwise not be proved using ME. Two theorems are from the theory of continuous functions and one is a collection of three challenge problem from the 1960's not all solved automatically until now. (Two of the three latter challenge problems were solved in the weak fully automated mode; that is, some initial restructuring of the problem was done to force generation of lemmas. This was done by splitting the problems into two and six cases respectively, without knowledge of the problem solution. The details are presented later. We mention here that the problem has subsequently been solved in (strong) fully automated mode by the Semantic Hyper-Linking prover of [11]. To our knowledge no other prover has succeeded on this problem in non-interactive mode.) Because the purpose of this paper is to look in-depth at the nature of lemma use, and because other papers [5, 3], contain results of METEOR on standard problems of the Automated Theorem proving (ATP) community, we omit such listings here. For example, the Astrachan and Stickel paper [5] discussing caching in includes two tables of results and a brief discussion on lemma use. (Caching is not applicable for the examples we treat here.) The Model Elimination procedure is a linear input procedure; that is, one parent clause is the preceding clause and the other clause is an input clause. This is the key property of SLD-resolution that allows compilation of Prolog programs, elegantly implemented in the WAM architecture. ME can use the architecture structure and also enjoy a high inference rate. However, like Prolog, ME does suffer high search redundancy due to the depth-first search mode of the WAM architecture. (ME does use iterative deepening rather than pure depth-first search, however.) Lemmas shorten proofs, thus reducing this redundancy. We show by example that the compression of proof achieved with lemma use can be striking, and the gain occurs on "real" (vs. toy) problems. We note that there is a version (restriction) of resolution that is in 1-1 correspondence with ME. The Resolution version is a linear input procedure except for a restricted "ancestor resolution" operation where resolution between two deduced clauses is necessary. (See the TOSS procedure in [19] 1 .) Thus, the procedure can be regarded as an encoding of the TOSS Resolution restriction using 2-sorted logic (framed and unframed literals), but it should be noted that ME is not, strictly speaking, a Resolution procedure. This is highlighted by the fact that Linear Input Resolution is complete only for a small extension of Horn clause logic whereas ME is complete for all of first-order logic. This paper is organized as follows: In Sections 2 and 3 we briefly describe the ME proof procedure and the realization of this procedure in METEOR. In Section 4, Section 5, and Section 6 we describe the successful application of lemmas to three theorems none of which can be proved without lemmas using We conclude with a short summary in Section 7. 2 The Model Elimination Procedure Like Resolution, Model Elimination is a refutation procedure, with the formula to be refuted presented in conjunctive normal form. Skolem functions are 1 The TOSS procedure corresponds to a slightly different ME procedure. 2 None of these problems yields a proof in less than 24 CPU hours using "vanilla" ME without lemmas. used to eliminate existential quantifiers. (See [10] or [19] for details on formula preparation; [19] contains a full description of the ME procedure.) We view the clauses as sets of literals and the formula as a set of clauses. The corresponding entity to a derived clause in Resolution is the (derived) chain, an ordered list of literals where each literal is of one of two classes, an A-literal or B-literal. ME, as currently defined, has two basic operations 3 , extension and reduction. The basic operation of extension is like the Resolution operation with retention of one of the literals resolved upon. The retained literal is promoted to a A- literal. (Intuitively, the A-literal is a type of ancestor literal.) We give a succinct presentation of the ME operations and lemma mechanism here, followed by an example. The appendix gives an expanded presentation of the procedure. The first chain in a deduction is an ordered input clause with all literals classified as B-literals. Below and in general we oversimplify the reference to occurrences of literals. In particular, we refer to a literal in successive chains as the same literal when in fact the later literal may be an instantiation of its "parent" literal. This simplification should cause no confusion. The extension operation glues a (shortened) input clause, ordered by user choice, to the left of the current chain if the leftmost B-literal of the chain unifies with the complement of some literal of the input clause. The new chain is the instantiation of the current chain by the unifier with the unifying literal of the input clause dropped, and the other unifying literal (the leftmost literal of the current chain) promoted to A-literal. Newly added literals are B-literals and other literals retain their classification from the current chain 4 . All leftmost A-literals, if any, are removed back to the leftmost B-literal. The reduction operation removes the leftmost B-literal of the current chain if it can be unified with the complement of an A-literal of the chain. The new chain is the instantiation of the current chain by the unifier with the leftmost B-literal missing. Again, all leftmost A-literals, if any, are removed back to the leftmost B-literal. The creation of lemmas occurs when the leftmost A-literals are removed, at the end of the extension and reduction operations. Here we use only unit lemmas, although a more general notion of lemma, allowing multiliteral lemmas, is defined in [19]. The lemma chains are created as the complements of the removed A-literals, but only some A-literals can produce a lemma. To create only unit lemmas the eligibility mechanism is simple. In a reduction operation, the A-literal that complements the leftmost B-literal is the reduction A-literal. During a reduction step, all A-literals strictly to the left of the reduction A- literal are marked. Every A-literal in a newly created clause inherits the mark, if any, of its parent A-literal. Any A-literal unmarked at the time of removal 3 Early papers [17, 18] had three operations. Traditionally, ME chains grew "to the right" but Prolog conventions of replacing the leftmost literal have influenced recent implementations. We choose to follow the convention used in the METEOR implementation. creates a lemma. Not all generated lemmas are retained; retention depends on a number of criteria discussed later, some under control of the user. If the lemma is retained it acts exactly as if it were a unit input clause regarding extension. (Lemmas are subject to restrictions and modifications not shared by input clauses, however.) The ME refutation in Figure 1, adapted from [19, 2], illustrates the ME mechanisms. 1. p(X) -q(Y) input clause 2. -p(X) r(Y) input clause 3. -p(a) -r(Y) input clause 4. q(X) p(a) input clause (and goal chain) begin proof 5. p(Y) [q(X)] p(a) extension with 1 variable renaming, clause 1 6. r(Z) [p(Y)] [q(X)] p(a) extension with 2 variable renaming, clause 2 7. -p(a) [r(Z)] [p(Y)] [q(X)] p(a) extension with 3 8. [r(Z)] [p(a)] [q(X)] p(a) prior to removal of A-literals p(a) reduction unit lemmas formed -p(a) 9. 2 extension with lemma -p(a) Proof completion if lemma mechanism is not used: 9. -r(Y) [p(a)] extension with 3 10. -p(X) [-r(Y)] [p(a)] extension with 2 reduction Figure 1: ME refutation In Figure 1 the unit lemmas are created during the final stage of the reduction that yields chain p(a) at step 8. The reduction creates A-literal p(a) from parent p(Y) and the subsequent removal of A-literals p(a) and q(X) create the lemmas. Removal of r(Z) cannot create a unit lemma. In Figure 2 we present proof trees associated with the example, one proof tree for the proof without a lemma use and one that incorporates the lemma use. In a proof tree, each clause used in an extension is represented, here by its clause number. The goal clause is listed as root of the proof tree. The literals of a clause are represented by the child nodes of the clause, where the clause (actually, the appropriate chain of the clause) that is used in the extension on that literal is named. Reduction is shown by a backwards arrow from a node labeled Rn, for the nth reduction. The arc to an ancestor node of the reduction is labeled by the same Rn to indicate which is the A-literal of the reduction. The proof tree gives sufficient information to reconstruct the proof without backtracking; the instantiations of the variables are determined during the tracing of the proof using the proof tree. The proof tree on the right side of Figure 2 is a method of displaying lemma use and also representing a proof tree without lemma use. The leftmost box encloses the subdeduction that defines the lemma and the other occurrences of that box in the deduction indicate positions where the lemma is used. As such, the box would be replaced by a single node. The subdeduction in the box provides a subtree that could occur if the lemma device is not used. (Note that the tree to the left shows that a slightly different deduction actually occurred when the lemma device was not used.) This reporting device is used later in the paper for deductions of considerably larger size, and does give a graphic feeling of the saving in proof size realized by lemma use. The reduction arrow in the boxes is somewhat confusing in this instance because the reduction is to the A-literal that creates or calls the lemma. For the right side box the reduction arrow refers back to the literal that calls the lemma, for that is the valid A-literal of reduction if the lemma is replaced by its deduction.3 without lemma with lemma Figure 2: proof trees 3 The METEOR Architecture We provide here only a brief description of the METEOR architecture. A full description can be found in [2]. METEOR is written in C and runs on workstations and in parallel and distributed computing environments. The same search engine is used in both the sequential and parallel setting. A discussion of the architecture of METEOR with emphasis on the parallel and distributed modes, appears in [3]. Iterative deepening is used to ensure completeness of the search strategy in all cases. METEOR is designed to conform to the principles of the Warren Abstract Machine (WAM) [1], the de facto standard for Prolog implementations. How- ever, clauses are not compiled into WAM code, but into a data structure that is then interpreted at runtime by either a sequential or parallel search engine. Several different depth measures can be used in bounding a potential proof during an iterative deepening search. We report on four measures here, these are described in Table 1. measure description D inf inference depth - bound total number of inferences in proof tree D Alit A-literal depth - bound depth of proof tree D roll rollback-reduction - combination of D inf and D Alit (preference given to reductions) Dweight weight depth - weight clauses (per branch of proof tree) Table 1: depth measures employed in METEOR The measures D inf and D Alit are commonly used in ME provers and are the default measures in PTTP and SETHEO respectively. As a bound on the entire proof tree, the use of D inf ensures that a minimal-length proof (in terms of extensions and reductions) is found. D Alit was also used in the first implementation of ME [14]. At any point in a proof search one can tell the current depth with respect to the D Alit measure by counting the number of A-literals in the current chain, whereas for the D inf measure it is the total number of extensions and reductions needed to derive the current chain. (Of course, one does not include subsearches removed by backtracking.) Note that the depth of a proof tree corresponds to the number of A-literals in the longest chain of the proof; this depth is the minimal D Alit depth for which this proof tree can be found. For the example deduction at the end of the last section, when no lemma is used the D inf is 7 and D Alit is 3. By this we mean that the proof can be found in the depth stated, and that a depth bound less than that stated would not permit the proof shown to be found. This can be seen from the given derivation or by consulting the appropriate proof tree. Note that for the proof tree the D inf measure is determined by counting the nodes in the proof tree (excluding the root), and the D Alit is determined by counting the depth of the proof tree with the root having depth 0. For the deduction using the lemma the D inf is 5 and the D Alit is 3. For the latter depth we assume that the left branch is searched first; if the right branch were searched first then the D Alit is 2. When D roll is employed, depth is calculated as for D inf regarding extension, but a successful reduction operation results in "rolling back" the depth to that depth in force when the A-literal of the reduction was introduced to the chain as a B-literal. The measure was discovered by accident, and has no clear intuition for its definition; we simply noted it worked well in many cases. One observation regarding its success is that it charges for extension by unit clauses, e.g. lemmas, (whereas D Alit has no charge) yet gives added resources to explore the consequences of a reduction as does the D Alit measure, indeed often more resources than D Alit gives. For the example deductions of the last section the D roll is 4 for both lemma use and no lemma use. None of these measures is uniformly superior to the others nor does it appear that simple syntactic criteria can be used to determine which measure is more appropriate (see [2]). Dweight is used in conjunction with any of these depth measures rather than as a replacement for them. Input clauses are annotated with an integral weight that contributes to the weight of a chain when the clause is used. The weight of a chain is the sum of the weight of all clauses whose literals "appear" in the chain. When Dweight is used it is incremented in an iterative manner just as other depth bounds are incremented in an iterative deepening search. Minimum Value Theorem We first consider a problem where the lemmas are very effective. The theorem we consider is the Minimum Value Theorem (MinVT). This is also known in the ATP literature as AM8 [27]. Theorem: (MinVT) A continuous function in a real closed interval [a; b] attains its minimum in the interval. In Table 2 we give a readable rendition of the axioms for MinVT. The numbers on the left refer to the clause numbers for these axioms as they are given in [27] and in the order submitted to METEOR. In Figure 3 we give the axioms as submitted to METEOR. The clause numbering relates the input clause to its counterpart in Table 2. The "nocontra" suppresses the contrapositive of the dichotomy axiom (axiom 1) from being generated as it is clearly not needed and is a very costly axiom to include, as is the dichotomy axiom itself. Line 16 is for naming Skolem functions, for which we do not permit self nesting. The semantics of the problem made it likely that no function needed to be so nested. In Figure 4 and Figure 5 we present in proof tree format the proof found by METEOR. The proof tree is given in a coded form with the lemmas labeled but unwound so that the full structure of the tree is apparent. Figure 4 (on the left) shows the proof actually discovered by METEOR. It is a proof involving seven input clauses and four lemmas. Using depth measure D inf , this would be a proof of depth 10. Actually, it took D inf depth 12 (it can be done in D inf depth 11) to discover the proof because some lemmas used in the proof required that depth to be discovered. Figures 4 (on the right) and 5 present the full proof, with each inference labeled by a lemma label or by a clause number. The dotted arrows at the bottom of the proof tree to the right in Figure 4 link to the appropriate subtrees in Figure 5. The node labels beginning with "L" refer to lemmas, with the numbering determined by the proof recovery procedure (not the order of lemma generation); other node labels refer to the input clauses by number. the axioms for total order 1. x - y - y - x 2. x - y - y - z Roughly, because of the symmetry of the antecedent 3. x - y - y l is a minimum point in [a,l] 4. a - l 5. l - b Any point greater than l has a smaller point with smaller f-value. 7-9. a - x For all x, h(x) is the smallest point in [a,b] s.t. 10-13. a - b - a - y For all x, there exists a point k(x) in [a,b] such that This is the negation of the theorem conclusion 14,15,17. a - x Table 2: Axioms for Minimum Value Theorem (MinVT) We made the point previously that lemmas are often discovered in earlier, "failed" parts of the search tree. That is very evident here from the proof trees. For example, some top lemmas are used only once in the proof and so, to be invoked as lemmas, they must have been established before the proof was found. (In particular, note lemmas L2, L3, and L4, labeled on the trees in Figure 4 with derivations continuing in Figure 5.) Lemmas used in the proof of these lemmas were clearly discovered by the time the more complex lemmas were obtained. Note that this does not mean that they were discovered earlier in the same proof tree iteration because lemmas are retained through successive iterations. Thus information from "later" in the tree can be used on early branches of the succeeding iterate. In general, iterative deepening gives a breadth-first form of search within which the search is depth-first. The effectiveness of lemmas is dramatically made by noting that the depth of the proof given here, when D inf (the depth measure used here) is employed with no use of lemmas, is over 600! Such search depth is well beyond the capability of any implementation of ME, or any other prover we know. It should be pointed out that resolution does exactly this kind of compres- sion, and that resolvents are lemmas also, in the sense of retaining intermediate results. Indeed, we noted that unit lemmas are resolvents from tree portions where reduction is not used. The difference, as noted before, is that our lemmas nocontra 1. p(X,Y) - p(Y,X) 2. -p(X,Y)-p(Y,Z) - p(X,Z) 3. -p(X,Y) -p(Y,X) - p(f(X),f(Y)) 4. p(a,l) 5. p(l,b) 7. -p(a,X) -p(X,b) - p(a,q(X)) - p(X,l) 8. -p(a,X) -p(X,b) -p(f(X),f(q(X))) - p(X,l) 9. -p(a,X) -p(X,b) - p(q(X),X) - p(X,l) 11. -p(a,X) -p(X,b) - p(h(X),b) 12. -p(a,X) -p(X,b) - p(f(h(X)),f(X)) 13. -p(a,X) -p(X,b) -p(a,Y) -p(Y,b) -p(f(Y),f(X)) - p(h(X),Y) 14. -p(a,X) -p(X,b) - p(a,k(X)) 15. -p(a,X) -p(X,b) - p(k(X),b) 17. -p(f(X),f(k(X)))-p(a,X)-p(X,b) Figure 3: Input clauses for Minimum Value Theorem (MinVT) are optional regarding completeness. That does not mean they are optional regarding realizing an actual proof, as this case demonstrates. However, not being a priori concerned about having to retain almost every nonredundant intermediate result we do achieve much more trimming. In the first proof we obtained of MinVT, taking 171 secs., we created and retained just 593 lemmas while making about 393,000 successful inferences. The only constraint we had (besides the unit lemma constraint, which is universal for us but still a con- straint) was a term-length stipulation of 10 on the lemmas and a disallowance of iterative nesting of any function. For speed considerations the term-length check is done only on the term replacing a variable at a substitution occurrence; thus the literal may contain many more constant and function letters than the bound indicates. This check can be done on literals in new chains as well as on lemmas, but here was applied only to lemmas. Note, by the way, that the nesting restriction applied here is almost totally safe for a first run as all functions but f are Skolem functions, for which one should never gain by nesting, and nesting f makes no heuristic sense given the problem nature. The sensitivity to term length is quite modest; if term-length is not limited at all, the same number of lemmas are retained. For this problem, the limit on nesting of the same function symbol is necessary in order to constrain the number of lemmas generated. (That is not always the case, however.) Thus we illustrate the point that we can trade off more search with input clauses for less retention goal goal Figure 4: top-level proof of AM8 and partial unrolling of proof of intermediate results than many strategies of resolution employ. 5 Wang Challenge Problem It is the usual case that we generate too many lemmas and seek ways of trimming lemmas without removing those helpful to the proof. Therefore, it is an interesting situation to encounter a problem where too few lemmas are produced. This situation arose when we pursued a challenge problem set that had not been fully proven by an automated theorem prover, to our knowledge. In 1965, Hao Wang put forth several challenge problems for automated theorem provers, problems beyond the capability of theorem provers at that time [26]. Moreover, they proved difficult for automated provers for some time. The problem set we consider were labeled ExQ1, ExQ2, and ExQ3 by Wang, and were treated in the section "Examples from Quantification Theory". Although Wang suggests that for a more intuitive understanding of these examples, we should think of F as the membership relation ffl, this suggestion did not achieve its purpose for us. We state the three problems, which are formulas to be refuted in Table 3, followed, in Figure 6, by one set of axioms used by METEOR. In Figure 6, replaces constant n of the axioms (the latter for purely historical reasons). Problem ExQ1 has been solved by a number of systems, we suspect, because it is also known as "Wos 31", in [28]. As such it has been solved by OTTER [20], but because this is one of numerous problems that circulate under labels as meaningful as "Wos 31", it is not easily determined what provers have solved Figure 5: proof of AM8, lemmas unrolled ExQ1: The conjunction of the following formulas is unsatisfiable. 1. m 6= n 2. n 6= k 3. 4. 5. ExQ2: Replace (2) and (3) above by 1. 2. ExQ3: Remove (1) from ExQ2. Table 3: Axioms for ExQ problems this first problem. OTTER has not solved ExQ2 or ExQ3 although ITP, in some sense an ancestor of OTTER, did solve ExQ2. We have solved it with some preconditioning on the manner of presentation of the problem; we broke it into cases so as to force generation of lemmas. As previously mentioned, the semantic hyper-linking theorem prover of Chu and Plaisted subsequently has provided a fully automated proof, the first to our knowledge [11]. (This problem is particularly well suited for their prover because it is highly non-Horn, like a logic puzzle. That is why it turns out to be poorly structured for us; the high number of reductions blocks the production of unit lemmas.) A proof of Wos31 (ExQ1) was obtained using D Alit without breaking the theorem into cases; using Dweight reduces the time to one-third of that when no weighting is employed. Note that our ExQ3 solution contains the ExQ1 case, in terms of the inequalities for the three basic constants. As for ExQ1 we again use D Alit for case 6 of ExQ3; for all other cases we use D roll . It is interesting 8 -p(Y,V),-p(V,Y),r(Y,m),p(Y,m),r(V,m),r(V,Y) 21 -r(X,Y),r(f(X),f(Y)) 9 -r(g(Y),b),r(Y,b),p(Y,b) 22 -r(X,Y),r(g(X),g(Y)) Figure Input clauses for ExQ2, case 1 that all other cases are solved using D roll and not easily solved (in some cases) using D Alit . When we first ran ExQ2, very few lemmas were generated and no proof was obtained. Indeed, an overnight run (14 hours) yielded only 6 lemmas. In con- trast, for ExQ1 87 lemmas were created in one successful run. Our experience with the success of lemmas made us consider immediately how more lemmas could be generated. Because unit lemma generation is inhibited by reductions we decided to introduce unit clauses to allow more frequent proof branch termination by unit extension rather than reduction. Although we did not understand the proof of the theorem (and still don't) it did seem apparent that ExQ1 and ExQ2 were in effect dealing with cases of equality and inequality of three con- stants; k,m,n. We broke up the problem into the cases listed in Table 4 and were able to prove each theorem. Note that the case definition for ExQ2 follows immediately from the axioms added to define ExQ2. Statistics for these cases are provided in Table 4. ExQ2 cases case case ExQ3 cases case case case case case case problem depth time # inferences # lemmas measure (secs) (case 1) D roll Dweight 4,562 14,462,880 47 ExQ2 (case 2) D roll 4.6 12,683 26 (case 1) D roll 108 170,185 67 ExQ3 (case 2) D roll 14 19,916 104 (case (case ExQ3 (case 5) D roll 9 12,741 26 (case Table 4: cases and statistics for ExQ1, ExQ2, and ExQ3 As reported in Table 4 for ExQ2(case 1), we used Dweight , although the same result was obtained by D roll alone in twice the time (9784 sec.) and double the inference total (29,886,276 inferences). The weights are the entries seen at the end of each input clause. It is interesting to view a case of use of weights; these weights were entered once prior to any weighted run, using an "algorithm" that accounts for clause length primarily, but with high penalty for known problem axioms such as the equality axioms. There is a small adjustment in general for the number of free variables in a clause. The major gain in using the weights here is the control of the equality axioms. We comment on other aspects of the METEOR listing of the axiom set in Figure 6. The lex predicate is used to order constant symbols for use in demodulation, and the skolem predicate lists the Skolem functions as mentioned in the previous section. Neither restriction was used in the runs reported in this section; demodulation was not used at all in these reported runs. (We obtained so few lemmas that use of demodulation was not a useful option, but the problem was initially specified so that demodulation could be used.) The input axioms are seen to differ in small ways from the natural translation of the Wang problem specification. Besides the replacement of b for n, there is the appearance everywhere of example. We had done our experiments with the axiom set obtained from [27] before tracing the origin of the problem in preparation for this paper. We have chosen to accept the variance from the original problem statement since [27] has become a standard source for problems, including this problem set. We now feel that we have discovered a new technique for ME, that of breaking problems into cases to generate useful lemmas. It may even pay to do this when many lemmas are generated, by severely limiting the natural lemmas being produced if they do not yield a proof, and trying to generate new lemmas by the case method. It is clear that this is not (yet) an automatable technique, as other tries at this have not been useful. That is, we tried to see if the discovery of a proof for ExQ1 could be speeded up by introduction of cases. However, all runs of ExQ1 using added cases resulted in increased time to proof. But we have seen that the technique of forcing lemma creation is fruitful. How widely applicable the idea is awaits further experience. 6 Bledsoe Challenge Problem The third problem we consider is actually a collection of variants on a problem investigated by Bledsoe starting in the early 1970's. Using a natural deduction style theorem prover with some special heuristics for limit theorems, Bledsoe and Boyer [7] proved that the sum of two continuous functions at a limit point is the limit of the sums at that point. They assigned the label LIM+ to this problem. The theorem has since been proven on the STR - +VE prover of Hines and Bledsoe [8, 15], a general theorem prover with some special devices for handling inequalities on the real line. Bledsoe recognized the difficulty of the theorem for the uniform provers and issued a challenge set of problems based on the LIM+ problem [6]. There are five different first-order axiom sets designated as the challenge set; two make explicit use of the equality predicate with one formulation requiring use of paramodulation or another rule that builds-in equality. Using METEOR we have been able to prove the first three formulations of the challenge problem. The lack of a full built-in equality mechanism (no mechanism such as paramodulation) precludes an attempt on the fifth formulation of the problem. The fourth formulation, using equality without any strong equality restriction or simplification mechanism stronger than the unadorned demodulation mechanism for lemmas that we do have, is too much for METEOR at this point. The first three formulations use the single predicate - (also used in the MinVT theorem discussed in Section 4). Several uniform provers have now obtained the third formulation of the challenge problem but METEOR was the first to achieve success at this level to our knowledge. (See Digricoli [13, 12] regarding proofs of the fourth and fifth formulations done in a partially interactive mode.) Our purpose here is to report on the experience of METEOR with lemma use on this problem set. The third formulation is the hardest problem where METEOR has succeeded, where by "hardest" we mean that the strongest set of restrictions we have ever used were needed to obtain a proof. One proof requires over 20 hours on a SUN SparcStation 2; shorter proofs required some insight as to the proof so do not count as legitimate datapoints regarding the capability of METEOR as a fully automated theorem prover. Proof attempts using proof knowledge are used to test various constraints in shorter time periods. We will make clear which experiments use proof insight when such results are presented. In Figure 7 we give the clauses used in the three formulations, in roughly the order that METEOR took them as input (read from left-to-right and top-to- bottom). The clause numbers are those of the original presentation by Bledsoe in [6]. METEOR normally re-orders clauses by number of literals (with preference for unit clauses), with attention to degree of generality also considered. The item "n at the end of each clause assigns a clause weight n, as discussed later. We do not give the original non-clausal formulation of the clause set because the meaning of most clauses is self-evident, once the constants are interpreted. The predicate lt denotes -, the function ab denotes absolute value, d1 and and d2 are Skolem functions for regions (see below), ha is one-half, pl is +, ng is negative, xs is a Skolem function for exception points in the delta regions, and min is, of course, minimum. Clauses 3 and 4 give the continuity condition for function f and g respectively. For example, clause 3 is a direct translation of Equation 1 %clause 5 % clauses 1,2 clause 12 % clause 8 clause 8.1 clauses 3,4 clause 6 clause 9.1 % clause 10.1 clause 9.2 % clause 10.2 clause 9.11 % clause 10.11 clause 10.3 % clause 11.3 lt(X,0),lt(Y,0),-lt(min(X,Y),0)"6 -lt(X,ha(Z)), -lt(Y,ha(Z)), lt(pl(X,Y),Z)"6 clause 14 % clause 15 clause 7.1 clause 7.2 Figure 7: clauses for Bledsoe challenge problems The goal clauses, 7.1 and 7.2, give the negation of the theorem we assert (in conjunction with clause 6). For example, clause 7.1 states Clause 7.1 is a more difficult goal to achieve because of the need to use the triangle inequality. However, that change and the related need for clause 8.1, are not significant changes. Formulation 2 is not much harder than formulation 1. The introduction of the dichotomy axiom (axiom 14) and especially the transitivity axiom (axiom 15) along with axiom changes for min makes formulation 3 a considerably harder problem. We list the three formulations by clause set, but will here only study the third formulation. The first two problem variations METEOR could handle relatively straightforwardly. The clauses for all three formulations are: formulation 1 clauses 1, 2, 3, 4, 5, 6, 7.2, 9.11, 10.11, 10.3, 11.3, 12. The goal is clause 7.2. formulation 2 clauses 1, 2, 3, 4, 5, 6, 7.1, 8.1, 9.11, 10.11, 10.3, 11.3, 12. The goal is clause 7.1. formulation 3 clauses 1, 2, 3, 4, 5, 6, 7.1, 8, 9.1, 9.2, 10.1, 10.2, 11.3, 12, 14, 15. The goal is clause 7.1. We include some data on runs for the first two formulations with little dis- cussion; see Table 5. In the lemmaizing run, only ground lemmas are stored and only ground subgoals are extended by lemmas. For reasons not important here these combined restrictions are labeled "ground generalizations only". thm search parameters time (secs) and number of inferences D Alit Dweight y lemmaizing z one 1222.7 2,854,399 3.94 14,682 26.37 36,232 two 3859.8 12,294,133 9.76 43,980 10242.9 11,682,410 yusing D Alit and Dweight zground generalizations only Table 5: time and inferences for formulations one and two We also give the proof tree for the second and third formulations in Figure 8 with the proof of the second formulation on the left. The dotted boxes surrounding subtrees denote lemma occurrences as discussed below. 6.1 The Third Formulation We now list the proof devices and restrictions used to tackle the third formula- tion. Many have been used here only, so our experience with them is limited. Others are more common, and we can comment on our more general experience with them. 1. We use a combination of depth measures, D Alit and Dweight . This has been tried often, usually obtaining considerably improved performance. This is one of the few times that the proof has not been obtained without this device. ( This is an oversimplification in that a proof has been obtained using a combination of D roll and Dweight but only under optimal limits on other parameters. It was considerably more time-consuming than with D Alit and Dweight , so more general attempts were not tried.) 2. Only ground lemmas are stored, and these are only used when no instantiation of the calling goal occurs. This trims the branching factor to one at this node. 6 10.1 Figure 8: proof trees for second and third formulations 3. Lemma literals are limited in number of symbols allowed. 4. The number of occurrences of any specific input clause allowed in a proof is controlled. METEOR is very sensitive to this parameter, so we could not set it by uninformed guess and realize a proof. For our no-proof- knowledge proofs METEOR increments the upper bound on the clause occurrence count as a function of current depth according to a ratio set by the user. The ratio needed is moderately insensitive to user setting in terms of obtaining a proof at all, but the time needed to realize the proof is strongly sensitive to that ratio. 5. Literals containing terms of certain forms are not retained. The exclusions range from obvious redundancies to possibly strong exclusions. We have results for several combinations of exclusions. 6. Reduction is restricted to B-literals whose source is the transitivity axiom. This restriction is more natural than it sounds, for reasons stated below. 7. The goal clause is reordered from the usual rule of ordering. There are only two literals so there is a binary choice, but normally we would expand the more constrained literal first. This is a significant deviation, because we rarely had to have to rotate through the literals of the goal clause, yet in this case no proof has been obtained using the usual literal as the literal first expanded. We now expand on some of the above. The depth measure Dweight utilizes the clause weight that appears appended at the end of each input clause. The depth measure functions as does D Alit except the clause weight is used to increment the count instead of the unit increase used by D Alit . This depth measure may be used in conjunction with each of the other measures; the proof development terminates appropriately when any depth count is exceeded, so one gets the conjunction of the depth criteria. The weights used in the clauses were guessed once and not changed, following a prescribed pattern. Some adjustment from a strict formula occurred. For example, the min axioms constitute one-third of the axioms, but are not likely to constitute one-third of the clauses used in a proof, so are weighted higher. The initial "guess" did not use proof knowledge, so is considered a non-sensitive constraint. Controlling the number of occurrences of each input clause was necessary for the any proof to succeed. In actuality it is control of dichotomy and transitivity that is crucial; experiments show that we can make quite safe guesses about upper bounds for all other axioms if the transitivity and dichotomy axioms are limited to the optimal count. We have solved the problem with iterative incrementing on all clauses, but this is effectively equivalent to binding only dichotomy and transitivity (clauses 14 and 15) as shown in Line 6 of Table 7 which indicates limits on clause use of optimal plus twenty except for dichotomy and transitivity. When iterative deepening on clause limits is employed the user still must pick a ratio of clause-limit increment to depth bound increment, but this is less sensitive than guessing the clause count. Data on Line 7 of Table 7 increases the count by one for every two levels of increase in the D Alit count. Literals can be excluded from retention when certain terms are created upon unification. This is an action at the chain level, not just at the lemma level. Thus this can effect completeness "in the small". This avenue is taken because of the large number of proof branches explored that seem to expand from along very questionable routes, dealing with terms that are either unlikely to occur in the proof or are redundant with other terms. The terms excluded range from clearly safe for exclusion to perhaps quite risky. We stress that we do obtain proofs of the theorem excluding only the very safe exclusions, but performance is definitely enhanced with more risky exclusions. In Table 6 we list the exclusions used. 1. d1(d2(X)) 2. d2(d1(X)) 3. d1(d1(X)) 4. d2(d2(X)) 5. d1(pl(X,Y)) 6. d2(pl(X,Y)) 7. min(min(X,Y),Z) 8. min(X,min(Y,Z)) 9. 13. min(ab(X),Y) 14. min(pl(X,Y),Z) 15. min(X,pl(Y,Z)) 16. min(X,X) 17. Table terms excluded in Bledsoe challenge problems We used four classes of risk in the exclusion heuristic: very safe, safe, some risk, more risk. In the very safe category are the nested Skolem functions, terms such as d1(d2(X)). Also included is the term min(X; X). Still considered safe is the term ha(ab(X)), because ab(ha(X)) is not excluded. Likewise for included. That is slightly riskier perhaps because the symbol count has increased and there is a limit on the total number of symbols allowed per literal. (The limit used was either 11 or 15, about equally used, but 20 symbols still allowed a proof.) Others in Table 6 are riskier; we discuss this further in the next paragraph. Line 3 of Table 7 gives data for a run using very safe exclusions only. (The phrase "ex. 1-4, 16" means that terms 1-4 and 16 are excluded terms for that run.) Lines 4 and 5 provide data using other subsets of the terms in Table 6. The column labeled att.infs gives attempted inferences, which are potential inferences identified prior to the unification attempt. run statistics description time infs. att.infs. # of lem.infs. of run (sec) 1. optimal (termsize=15) 27858 23,146 66,485 115 590 2. optimal (termsize=20) 23095 23,052 66,346 153 601 3.y very safe (ex. 1-4,16) 66609 65,871 196,968 186 1,136 4.y ex. 1-4,7-10,12,14,16 49206 47,861 143,409 133 1,037 5.y ex. 1-4,9,10,16 64368 61,499 186,207 180 1,124 limits 7. iter. deep.z 152077 129,419 355,250 181 2,443 ytermsize limit of 15 symbols zSparc ELC, increase all limits by 1 every other stage Table 7: altering limits on extensions for third Bledsoe problem We now briefly consider some of the riskier exclusions we considered. Excluding similar terms may seem very risky, but we noticed a proliferation of such terms and it led to asking if they were needed. Note that terms of this type could still arise through instantiation of variables. Here exclusion means that unification cannot instantiate a variable with an excluded term. To illustrate this, the goal \Gammalt(min(e0; e0); min(X; e0)) can be refuted with the exclusion of all terms in Table 6. In obtaining the proof the term min(min(e0; e0); e0) is generated when X in the goal is instantiated to e0). Thus, an excluded term's appearance is not forbidden, just the use of a term that embeds the excluded term as a substitution instance. The reader can see this by first extending with clause 9.2 followed by extension with clause 9.1. Even though completeness is endangered this is a reasonable action to take and this type of step may be needed to get very difficult theorems. We are adverse to this not because it endangers completeness if the exclusions are risky, but because we have been developing the thesis that basic ME should be hidden and the user should play with the lemmas. We give in Table 8 data on runs with exclusions, and note again that the exclusions are not needed for the proof. run statistics description time infs. att.infs. # of lem.infs. % succ. of run (sec) 1. all limits 23324 23,144 66,994 2. D roll , all limits 38663 37,928 116,069 71 1,081 33 3. ex. 1-17 38144 38,450 116,421 71 987 20 4. ex. 1-10,12-16,18 23533 23,709 68,196 5. ex. 1-8,18 43228 38,601 110,147 75 664 22 6. ex. 1-4,7,8,16-18 42101 37,633 107,665 75 659 22 7. ex. min terms 26529 31,290 89,898 8.y ex. 1-4,7-10,12,14,16 49206 47,861 143,409 133 1,037 21 9.y ex. 1-4,9,10,16 64368 61,499 186,207 180 1,124 12 10.y ex. 1-4,16 66609 65,871 196,968 186 1,136 12 14.z trm.size=15 27858 23,146 66,485 115 590 48 yterm-size zotherwise as in line 1 Table 8: performance characteristics for the third formulation That reductions are restricted to B-literals from the transitivity axiom actually came from noting that the simpler proofs for formulations 1 and 2 did not use reductions. That puts this in the category of "cheating" by our standards. Even though it was not based on knowing the proof of formulation three, it is not likely in the real world that simpler versions of the problem are around to break in our heuristics. Even though that is how we arrived at this guess, there is reason to guess that way again in similar situations. That is, it is a guess we would make in other similar situations. The reason is that we have seen many proofs having some non-Horn clauses where nevertheless reduction is not needed. So, after our initial tries that tells us the problem is hard, we would first try without reduction. Then we would try as here, allowing it on transi- tivity, because they often do use it. This is the first time that a problem has strongly benefited from this restriction; we have simply made the observation about when reduction occurs in other problems. More experience is needed to tell whether the pattern we just have suggested really holds. The interchange of literal order within the goal is the hardest to justify on uniform grounds. It works this way and not the other way. Usually, starting with the more instantiated literal first is the winning strategy. The normal ordering works in the first two formulations; the proof can be obtained without lemma use starting with the more instantiated literal. The proof is impossible to find without lemmas when the literal lt(D; 0) is expanded first. The situation reverses even for these easier formulations when lemmas are tried. The reason is clear from the proof tree. In Figure 8 we see that if one follows the second literal (the right branch) then the key lemma, in the dotted line box, is reached in D Alit depth 8 while pursuit down the other branch requires depth 13. (Recall that D Alit is proof tree depth, so the reader can check this directly by counting branch lengths in the proof trees of Figure 8). We note that the variable D in the goal is instantiated by the key lemma as follows: \Gammalt(min(d1(ha(e0)); d2(ha(E0))); 0). This instantiation gives the found in the typical proof found in analysis texts: for a given epsilon (ffl 0 ) the corresponding delta is the minimum of the deltas (ffi 1 and corresponding to half of the given epsilon. The min axioms are necessary for discovering this instantiation. However (and somewhat surprisingly), if the min axioms are removed from the input clause set a proof is obtained in less than 600 seconds with no limits on clause use. However, in this case "lt(d1(ha(e0)),0) or lt(d2(ha(e0)),0)" is deduced, giving less information than is provided when the min axioms are included, but in significantly less time. Summary Plaisted [21] has shown that ME with unit lemmas and caching is one of the few goal-sensitive procedures that has a polynomial search time within the propositional Horn clause setting in a suitable sense. The result assumes that ME is used with iterative deepening search, the normal mode for METEOR. Elsewhere (see Astrachan and Stickel [5]) data is given on use of unit lemmas and caching in the Horn clause domain. Here we have presented examples in the non-Horn setting (where caching is not implemented nor easily designed) but where we have demonstrated that unit lemmas are still very effective. We have considered examples illustrating very different situations: the nearly ideal, the insufficient lemma supply situation (where case analysis provided the needed lemmas), and a barely adequate lemma supply (where only one deep lemma is used and not even created until the correct proof is under development). This wide scope of situations we believe demonstrates the usefulness of unit lemma addition to the Model Elimination procedure. We hope to achieve enough knowledge about lemma use that henceforth we will view the basic ME procedure as a black box and the collection of lemmas as the data the user views and manipulates. Appendix We give a more formal definition of the Model Elimination(ME) procedure as used by METEOR. A fuller treatment of the ME procedure(s) is given in [19]. A chain is a sequence of literals, each literal an A-literal or a B-literal. (That is, a chain is an ordered clause with two types of literals.) An input chain is a chain of B-literals, obtained by placing an ordering on a clause from a conjunctive normal form (CNF) formula to be tested for unsatisfiability. Clauses from the CNF formula are called input clauses. A input clause of n literals from the CNF formula must be represented by at least chains, where each literal of the input clause is leftmost in at least one input chain. The empty chain is denoted by 2. If l is the leftmost literal in chain C we write (C \Gamma flg) to represent the chain C with the leftmost occurrence of l removed. There will be no occasion to remove a literal other than the leftmost literal, or a sequence of literals leftmost. In the latter case we regard the literals as being removed one at a time as the literal becomes leftmost in an intermediate chain. A chain is preadmissible iff 1. complementary literals are separated by an A-literal; 2. no B-literal is to the left of an identical A-literal; 3. no A-literal is identical or complementary to another A-literal. A chain is admissible iff it is preadmissible and the leftmost literal is a B- literal. In practice we will sometimes not enforce condition (2); that can be a user option. Let C 1 be an admissible chain and let C 2 be a variable-disjoint input chain. If there exists a most-general-unifier(mgu) oe of the leftmost literal l 1 of C 1 and the complement of the leftmost literal l 2 of C 2 , then the extension operation extends C 1 by C 2 to form chain C 3 by promoting the leftmost literal of C 1 oe to an A-literal and placing (C g)oe to the left, the literals of (C arranged in any order the user chooses. All literals in C 3 oe inherit the type of their parent literal in C 1 and C 2 with the noted exception of the promoted A-literal. If the chain that results is not preadmissible then the extension is not defined. If the chain that results is a preadmissible chain with an A-literal leftmost (i.e., nonadmissible) then the leftmost A-literals are removed back to the leftmost B-literal, which yields an admissible chain. Let C be an admissible chain. If there exists a mgu oe of the leftmost literal l of C and an A-literal of C, then the reduction operation yields chain (C \Gamma flg)oe. If the chain that results is not preadmissible then the reduction operation is not defined. If the chain that results is a nonadmissible preadmissible chain then the A-literals to the left of the leftmost B-literal are removed. An ME derivation is a sequence of chains composed of an input chain as first (goal) chain, followed by derived chains, each either an extension of the preceding chain by some input chain or a reduction of the preceding chain. An ME refutation is an ME deduction of the empty chain. Lemmas, an optional property of ME, are created when A-literals are removed in the transformation from preadmissible to admissible chain. A lemma is an informal label for either a lemma clause that results from the process of lemma creation, or the chains that are created from the clause. Here we consider only unit lemmas so the distinction is pedantic. From the lemma clause only one chain is created, consisting of the single B-literal and regarded as an ordered clause. A lemma chain is treated as an input chain after generation (if it passes filters that may eliminate it; these are generally user-controlled options). The lemma clause is the complement of the A-literal removed when the A- literal is eligible to create a lemma. The eligibility criterion is exactly as stated in the body of the paper but repeated here for completeness. This definition is only correct for unit lemmas. In a reduction operation, the A-literal that complements the leftmost B-literal is the reduction A-literal. During a reduction step, all A-literals strictly to the left of the reduction A-literal are marked. Every A-literal in a newly created clause inherits the mark, if any, of its parent A-literal. Any A-literal unmarked at the time of removal creates a lemma clause. --R Investigations in Model Elimination Based Theorem Prov- ing METEOR: Exploring model elimination theorem proving. METEORs: High performance theorem provers using model elimination. Caching and lemmaizing in model elimination theorem provers. Challenge problems in elementary calculus. Computer proofs of limit theorems. Variable elimination and chaining in a resolution-based prover for inequalities A parallel theorem prover for non-Horn clauses Symbolic Logic and Mechanical Theorem Proving. Semantically guided first-order theorem proving using Hyper-linking The Rue theorem-proving system: the complete set of LIM+challenge problems An implementation of the model elimination proof procedure. The central variable strategy of str - +ve Mechanical theorem proving by model elimination. A simplified format for the model elimination procedure. Automated Theorem Proving: A Logical Basis. Otter 2.0. The search efficiency of theorem proving strategies. A machine-oriented logic based on the resolution principle PARTHEO: A high performance parallel theorem prover. A Prolog technology theorem prover: Implementation by an extended Prolog compiler. A Prolog technology theorem prover: Implementation by an extended Prolog compiler. Hierarchical deduction. Efficiency and completeness of the set of support strategy in theorem proving. --TR --CTR Allen Van Gelder , Fumiaki Okushi, Lemma and cut strategies for propositional model elimination, Annals of Mathematics and Artificial Intelligence, v.26 n.1-4, p.113-132, 1999 Fumiaki Okushi , Allen Van Gelder, Persistent and Quasi-Persistent Lemmas in Propositional Model Elimination, Annals of Mathematics and Artificial Intelligence, v.40 n.3-4, p.373-401, March 2004 Allen Van Gelder, Autarky Pruning in Propositional Model Elimination Reduces Failure Redundancy, Journal of Automated Reasoning, v.23 n.2, p.137-193, August 1999 Allen Van Gelder , Fumiaki Okushi, A propositional theorem prover to solve planning and other problems, Annals of Mathematics and Artificial Intelligence, v.26 n.1-4, p.87-112, 1999
lemma;METEOR;model elimination
594176
Testing Positiveness of Polynomials.
Many problems in mathematics, logic, computer science, and engineering can be reduced to the problem of testing positiveness of polynomials (over real numbers). Although the problem is decidable (shown by Tarski in 1930), the general decision methods are not always practically applicable because of their high computational time requirements. Thus several partial methods were proposed in the field of term rewriting systems. In this paper, we exactly determine how partial these methods are, and we propose simpler and/or more efficient methods with the same power.
Introduction Many problems in mathematics, logic, computer science, and engineering can be reduced to the problem of testing positiveness of polynomials. In 1930, Tarski [33, 34] showed that the problem is decidable. In fact, he gave a decision method for a more general problem than just testing positiveness. Since many improvements and new methods were proposed 19, 18, 25, 36, 8, 21, 22, 5, 12, 23]. However, these methods are computationally expensive due to their generality. Naturally one is interested in devising more efficient methods for the sub-problem: testing posi- tiveness. But then, this sub-problem turns out to be still difficult. Thus, several authors (mainly from the field of rewriting systems, for proving termination) proposed some partial methods, partial in the sense that these method might not be able to decide the positiveness of some polynomials [24, 4, 10, 10, 31, 32, 11]. In this paper, we determine how much partial these methods are. In other words, we find proper subsets of the set of all positive polynomials that are decided by those meth- ods. Equivalently, we determine restrictions on the positiveness that are decided by them. Further, we propose simpler and/or more efficient methods for the same sets of polynomials The structure of the paper is as follows. In Section 2, we list three different restrictions on the property of positiveness (eventual, absolute, eventual absolute). In the subsequent sections, we show that each of the existing methods is a decision procedure for one of the above restricted positiveness properties. We also give several new methods. Several Types of Positiveness let P be a non-zero polynomial in R[x], and let - 2 R. Definition 2.1 (Types of Positiveness) We say that P is ffl positive from - iff 8x eventually positive 1 iff there exists a - such that P is positive from -. absolutely positive from - iff P is positive from - and every partial derivative P is non-negative from -, that eventually absolutely positive iff exists a - such that P is absolutely positive from -. 2 Example 2.1 Let polynomial. The polynomial P is positive from 0. In fact one sees easily that it is positive from all - ? \Gamma1. Therefore it is also eventually positive. However, it is absolutely positive only from - 0 because its second derivative (6x) is non-negative only for non-negative values of x. As it is absolutely positive from some - it is also eventually absolutely positive. 2 Example 2.2 Let 1. It is easy to see that the polynomial P is positive from every - and therefore it is eventually positive. On the contrary, it is not absolutely positive from any -, because the derivative always negative. Consequently it is not eventually absolutely positive. 2 1 The notion of eventual positiveness was first introduced by Lankford [24]. Naturally, we are interested in the relations among the properties. Let P - , EP, AP - , EAP respectively stand for positiveness from -, eventual positiveness, absolute positiveness from - and eventual absolute positiveness. The following proposition shows the relations among them. Proposition 2.1 The following diagram describes the implication relation among the different kinds of positiveness prop- erty. An arrow goes from A to B if A implies B. @ @ @ @R \Gamma\Psi @ @ @ @R \Gamma\Psi Proof. The proofs of all indicated implications follow easily from the definition. It is equally easy to see that the indicated implications are the only implications valid for the given properties. We can use polynomials from Examples 2.1 and 2.2 as counter examples to other implications. 2 3 Methods for Absolute Positive- ness We start with presenting a necessary and sufficient condition for absolute positiveness of a polynomial. The theorem will be used in several following subsections and will form a base for the new method we present later. From now on, let oe - stand for "shifting" by -. Precisely, oe - (P ) is the polynomial obtained from P by replacing each x i with polynomial P can be rewritten as e c e x e where the exponent vectors (e) are distinct and (distributed representation). We will call those c e 's coefficients of P . Theorem 3.1 Let P be a non-zero polynomial. Then P is absolutely positive from - iff every coefficient of the polynomial oe positive and the constant term is non-zero. 2 Proof. First note that since P is a non-zero polynomial, oe - (P ) is also a non-zero polynomial. Assume that every coefficient of oe - (P ) is positive and the constant term is non-zero. Then obviously oe - (P ) is positive from 0. Note that the graph of P is obtained by shifting that of oe - (P ) by - to the positive direction. Thus the polynomial P is positive from -. Let @ be an arbitrary but fixed partial differentiation operator as @ e 1 +\Delta\Delta\Delta+e n . Since the polynomial oe - (P ) has all co-efficients positive, the polynomial @(oe - (P )) also has all coefficients positive. Therefore @(oe - (P )) is non-negative from 0. Observing that differentiation and shifting commutes, that is, we see that oe - (@(P )) is non-negative from 0. Note again that the graph of @(P ) is obtained by shifting that of oe - (@(P )) by - to the positive direction. Thus @(P ) is non-negative from -. Therefore P is absolutely positive from -. We can use the same arguments as in the first part of the proof. The assumption of the polynomial P being absolutely positive from - is then equivalent to the fact of polynomial oe being absolutely positive from 0. Therefore we want to prove that if the polynomial oe - (P ) is absolutely positive from then it has all coefficients positive and non-zero constant term. The proof will be done by contradiction. Assume the constant term of oe - (P ) is zero. Then the value of oe - (P ) at 0 is 0, contradicting the fact that oe - (P ) is absolutely positive from 0. Assume there exists a term with negative coefficient in oe - (P ). Then there exists a derivative 2 of oe - (P ) with a negative constant term (it is the derivative eliminating all variables of the term with the negative coef- ficient). The derivative is negative at 0, contradicting the absolute positiveness of oe - (P ). 2 3.1 Partial Derivative Method Giesl [11] gave a partial method for testing the positiveness, by extending the idea of Lankford [24]. The method is based on the following implications (called differentiation rules): It might be the 0-th order derivative, that is, the polynomial oe - (P ) itself. The method, starting from the sentence repeatedly applies the implications (backward) until it results in a conjunction of inequalities among numbers. If all those numeric inequalities are satisfied, then the method outputs 'yes' (P is positive from -). If not, it outputs `don't know'. Since it is a partial method, a question naturally arises: How much partial is the method? In other words, we want to know the proper subset of the set of all positive polynomials that are decided by the method. As the reader might have guessed (from the fact we put the method into the section about absolute positiveness), the method is a decision procedure for absolute positiveness (Theorem 3.2). The rest of the subsection is devoted to the proving the theorem. For this we start with a lemma which will simplify the proof of the theorem. Lemma 3.1 The partial derivative method outputs 'yes' on an input P and - iff it outputs 'yes' on the input oe - (P ) and 0.Proof. Suppose that by applying Rule-1 on an inequality Q we generate two new inequalities Q is easy to see, from the commutativity of shifting and differ- entiation/evaluation, that by applying the same rule on the inequality oe - (Q) ? 0, we will generate two new inequalities The same holds for applying Rule-2 on Q - 0. The computation of the partial derivative method on input ends in a set A of numerical inequalities. Following the observation above, we easily see that applying the same computation path on oe - (P ) and 0 results in the set A 0 of numerical inequalities, which are obtained by shifting the inequalities in A. But then shifting a numerical inequality results in the same inequality. Thus we conclude that and so their truths are the same. 2 Now we can prove the exact power of the partial derivative method. Theorem 3.2 (Power of Partial Derivative Method) The partial derivative method outputs 'yes' for an input P and - iff the polynomial P is absolutely positive from -. Proof. Assume the partial derivative method outputs 'yes' on P and -, it means P is positive from -. From Lemma 3.1 we know it also outputs 'yes' on oe - (P ) and 0, and oe - (P ) is positive from 0. Now we will show, by contradiction, that oe - (P ) has all co-efficients positive and non-zero constant term. Assume that the polynomial oe - (P ) has either a negative monomial or constant term equal to zero. We start with the case when the constant term is zero. Then we know oe - (P ) is not positive at 0 and it contradicts the positiveness of oe - (P ) from 0. Next we assume that a monomial in oe - (P ) has a negative coefficient. We would like to show that the partial derivative method does not prove the positiveness of oe - (P ) from 0. For this, let us first observe the following. Let Q be a polynomial in an inequality and let M be a monomial in Q with negative coefficient. We claim that after applying a rule (Rule-1 or Rule-2) one of the resulting polynomials still contains a monomial with negative coefficient. To see this, let x i be the variable on which the rule is applied. Let Q 1 be the polynomial resulting from substituting 0 into x i in Q and let Q 2 be the polynomial resulting from differentiating Q in x i . Assume that x i occurs in M . Then the monomial @M occurs in since it cannot be canceled by other monomials. Further, the monomial obviously has a negative coefficient. Assume that x i does not occur in M . Then M occurs in Q 1 because there cannot be any cancelation among different monomials after substituting 0 into x i . Repeated applications of the rules results in a set of numerical inequalities. From the observation above we know that at least one of the inequalities must be of the form a - 0 or a ? 0 where a ! 0. Thus the partial derivation method does not output 'yes' for oe - (P ) and 0. Contradiction. Thus, we showed that oe - (P ) has all coefficients positive and the constant term is non-zero. From Theorem 3.1 we conclude that P is absolutely positive from -. Assume P is absolutely positive from -. From Theorem 3.1 we know that oe - (P ) has all coefficients positive and non-zero constant term. We will show that the method proves the positiveness of oe - (P ) from 0. In the final set of numerical inequalities there is only one strict inequality (resulting from the application of the first part of Rule-1). Note that the left hand side of the strict numerical inequality is the constant term of oe - (P ). Because the constant term is positive, the strict inequality is true. All other numerical inequalities are non-strict. Note that the applications of the rules cannot create a negative term. Thus all the numbers in the non-strict numerical inequalities are non-negative. Hence, all non-strict inequalities are true, We showed that the initial inequality is transformed into a valid set of numerical inequalities. Thus the partial derivative method outputs 'yes' for the input oe - (P ) and 0. From Lemma 3.1 we conclude that the method outputs 'yes' also on the input P and -. 2 3.2 Merging Method The method was first described by Ben Cherifa and Lescanne [4]. Steinbach [31, 32] increased the power of the method and removed some restriction imposed by the original version. The main idea of this method is deciding the positiveness of polynomial P by finding a sequence of polynomials such that The positiveness of polynomial Pn is checked with the "all coefficients are positive" rule. Of course this rule decides the positiveness of a polynomial only from positive -. The transformation of P i into P i+1 is done in two main steps. Firstly, a pair of monomials, one having negative and one positive coefficient are chosen. Secondly, the two monomials are merged into a new monomial. The procedures CHOOSE and CHANGE are the essential part of the method and for a detailed discussion on them we refer to Section 5.3 of [32]. We are again interested in the relation between the method and the positiveness properties we have defined. Giesl in [11] proved the following facts regarding the comparison of his method with the merging methods of Steinbach [32] and Ben Cherifa and Lescanne [4]. Theorem 3.3 (Power of Merging Method (Giesl 95)) ffl If the merging method can prove a polynomial P to be positive from -, then the partial derivative method can do so as well. ffl If partial derivative method can prove P to be positive from -, then there exists a - 0 - such that the merging method can prove P to be positive from - 0 . 2 Giesl also gives an example for the second observation, where the merging method cannot prove positiveness of P from - though the partial derivative method does. Therefore the partial derivative method is stronger than the merging method. We see that the merging method besides being a partial method for deciding positiveness of polynomial also only partially decides absolute positiveness of polynomial. Naturally one can be again interested in an exact characterization of the power of the method. We do not know a simple characterization for the polynomials positiveness of which is decided by the merging method and leave it as an open problem. 3.3 Shifting Method (New) We present a new and simple method for checking absolute positiveness of polynomial. The method is based on Theorem 3.1 which gives the necessary and sufficient condition for absolute positiveness. From the theorem the method is obvious. 1. Shift P by -, that is, compute the polynomial P 0 such that 2. If all the coefficients of P 0 are positive and the constant term is non-zero then output 'yes', else output `no'. Example 3.1 Check if the following polynomial 3 P is absolutely positive from 2: We apply the shifting method. For this, we first shift P by 2, Next, we find that all the coefficients are positive and the constant term is non-zero (8), thus we conclude that P is absolutely positive from 2. 2. 3.4 Comparing the Complexity of the Method Since all the above methods are of equal testing power, we are interested in complexity differences. In the following theorem, we recall the analysis of the merging method by Steinbach [31] and the analysis of the partial derivate method by Giesl [11]. Then we give the analysis of the shifting method. Below we assume that every arithmetic operation takes a constant time. Theorem 3.4 (Computing Times) The computing times for worst cases are bounded from above as follows: ffl Partial Derivative Method O(d n ') ffl Merging Method O(' ' ) Shifting Method O(d n ') where n the number of variables, ' the number of monomials, d the maximum of the degrees in each variable. 2 Proof. Partial Derivative Method: In [11], Giesl gave the following derivation: Let T (n; d; ') be the computing time for an input characterized by n, d and '. First we compute store the results for the later use. This 3 This polynomial arises from a certain polynomial interpretation of the term rewrite system for associativity and endomorphism [11]. takes O(d). Let S(n; d; ') be the computing time for the re- maing part (successive elimination of variables). The elimination of one variable involves the following: d partial differen- tiations, where each differentiation takes O('), d evaluations, where each evaluation takes O('), using the stored values of the powers of -. Further this process produces (d polynomials in variables to work on. Thus, we get the following recurrence formula: By solving the recurrence formula, we see that S(n; d; O(d n '): Thus, T (n; d; Merging Method: Steinbach [31] proved the following bound where s is the number of the positive monomials and t is the number of the negative monomials. He then, using the formula of Stirling(s! aproximates to ( s 2-s), simplified it to t, it can be rewritten in terms of ' as Shifting Method: First, we need to shift each monomial. 4 For this, we need to shift the power of each variable, that takes O(d 2 ) for each variable. Then we need to multiply out the shifted powers, that takes O(d n ). Thus, shifting each monomial takes O(d shifting all the monomials takes O(d n '). Next, we need to sum all the shifted monomials, which takes O(d n '). Thus, all together, the shifting method takes O(d These are asymptotic results (with the big O's), and thus one should be careful not to draw any too definitive conclusion about their performances on small/moderate size inputs, which often arise in application domains. 4 Methods for Eventual Absolute Positiveness Now we shift our focus to the (known and new) methods for testing eventual absolute positiveness. 4.1 Partial Derivative Method (eventual positiveness version) We have already mentioned the Lankford's method [24] which was suggested as a partial decision method for eventual pos- itiveness. Let P be a non-constant polynomial. Then the 4 Actually, there are practically better ways, such as Honer evaluation or synthetic division, etc. But they do not seem to give a better bound, thus we will stick to the brute-force way since it easier to analyze. method is based on a simple fact that if every non-vanishing first order partial derivative of P is eventually positive then P is eventually positive. We can apply the same reasoning on the partial derivatives repeatedly until all polynomials are numbers. If all of them are positive then the input polynomial is eventually positive. Otherwise, we don't know. Since it is a partial method, a question again naturally arises: How much partial is the method? We will answer this question by following the similar approach that we already used in studying the power of the partial methods for testing positiveness (Section 3). Lemma 4.1 The partial derivative method (Lankford) outputs 'yes' on an input P iff it outputs `yes' on the input Proof. Suppose that by applying differentiation on an inequality another inequality Q ? 0 . It is easy to see, from the commutativity of shifting and differentiation, that by differentiation in the same variable on the inequality oe 0, we will generate a new inequality oe - (Q The computation of the partial derivative method (Lank- ford) on input P ends in a set A of numerical inequalities. Following the observation above, we easily see that applying the same computation path on oe - (P ) results in the set A 0 of numerical inequalities, which are obtained by shifting the inequalities in A. But then shifting a numerical inequality results in the same inequality. Thus we conclude that and so their truths are the same. 2 Now we can prove the exact power of the partial derivative method (Lankford). Theorem 4.1 (Power of Partial Derivative Method) The partial derivative method (Lankford) outputs 'yes' on input P iff the polynomial P is eventually absolutely positive. Proof. Assume that the partial derivative method outputs 'yes' on an input P. We need to show that P is eventually absolutely positive. From the assumption, we know that P is eventually positive. Therefore there exists a - 0 such that P is positive from - 0 . Further from the details of the execution of the partial derivative method, we observe that every partial derivative of P is either a zero polynomial or eventually posi- tive. Thus for every non-zero partial derivative P of P , there exists a - such that P is positive (thus non-negative) from - . Now let - be the maximum of - 0 and all - 's. Then, P is absolutely positive from -. Thus, P is eventually absolutely positive. Assume that the polynomialP is eventually absolutely pos- itive. We need to show that the partial derivative method outputs 'yes' on this input. Since P is eventually absolutely positive, there exists - such that P is absolutely positive from -. Then, oe - (P ) is absolutely positive from 0: By Theorem 3.1, we see that all coefficients of oe - (P ) are positive and the constant term is non-zero. Note that the partial derivative method does not introduce negative terms during the execution on the input oe - (P ). Thus, the method will output 'yes' on the input oe - (P ). By Lemma 4.1, we conclude that the method will also output 'yes' on the input P . 2 4.2 Dominating Monomials Method (New) Now we propose a new and more efficient method for testing eventual absolute positiveness of polynomial. First note that the eventual (absolute) positiveness of a univariate polynomial can be trivially decided by checking whether its leading coefficient is positive. We generalize the observation to multivariate case. For this, the notion of leading monomial must be generalized. The new method is essentially based on one such generalization (which we call dominating monomial). Definition 4.1 (Dominating monomial) We say that a monomial s dominates a monomial m. iff m divides s and m is different from s. We say that s is a dominating monomial of P iff there is no monomial in P dominating s. 2 Example 4.1 Let us consider polynomials There are three dominating monomials in P there are two in Q univariate polynomials there is only one dominating monomial - the leading monomial. 2 Now we are ready to describe the new method: 1. Collect all the dominating monomials of P . 2. If their coefficients are all positive, then output 'yes', else output 'no'. The proof of the correctness of the method will be given later (Theorem 4.2). Let us first get familiar with the method by working on several small examples. Example 4.2 Check if the following polynomial is eventually absolutely positive: One sees immediately that the only dominating monomial is xyz. Since its coefficient is positive, the polynomial is eventually absolutely positive. Let us try another: One sees that the dominating monomials are x 4 , 2xy 3 and y. Since their coefficients are positive, the polynomial is eventually absolutely positive. Let us try still another (obtained by changing signs in the previous example), Again, the dominating monomials are x 4 , \Gamma2xy 3 and 3x 2 y. Since one coefficient is negative, the polynomial is not eventually absolutely positive. 2 Remark: Note that one can use the dominating monomial method as a quick pre-pruning method while testing absolute positiveness: Suppose that we would like to check whether a polynomialP is absolutely positive from -. We first apply the (very cheap) dominating monomial method. If it turns out that P is not eventually absolutely positive, then we do not need to carry out the more expensive absolute positiveness check, since we already know that the answer is 'no'. 2 The remaining of this section is devoted to proving the correctness of the dominating monomial method (Theorem 4.2). We begin with a lemma which "justifies" the notion of dominating monomial. Lemma 4.2 The set of dominating monomials of P and that of oe - (P ) are the same. 2 Proof. Let m be a dominating monomial of P . We need to show that m is a dominating monomial of oe - (P ). We prove it by contradiction. Assume that m is not a dominating monomial of oe - (P ). There are two cases: ffl m is not a monomial of oe - (P ) ffl m is a monomial of oe - (P ), but not dominating. Case 1: m is not a monomial of oe - (P ). Note that m is a monomial of oe - (m). Since m is not a monomial in the oe - (P ), there must be a monomial q (different from m) in P such that there is a monomial q 0 of oe - (q), that has the same degree vector as m. Every monomial of oe - (q) divides q, in particular, q 0 divides q. Since m and q 0 have the same degree vector, m divides q. Recall that m is different from q, thus q dominates m. Therefore m is not a dominating monomial of P . Contradiction. Case 2: m is a monomial of oe - (P ), but not dominating. Thus there is a monomial q 0 in oe - (P ) that dominates m. Hence there is a monomial q of P such that there is a monomial q of oe - (q), that has the same degree vector as q 0 . Every monomial of oe - (q) divides q, in particular, q divides q. Since q 0 and q have the same degree vector, q 0 divides q. Recall that q 0 dominates m. Thus m divides q 0 and they are differ- ent. Since q 0 divides q, we see that m divides q and they are different. Thus q dominates m, and m is not a dominating monomial of P . Contradiction. Thus m must be a dominating monomial of oe - (P ). Let m be a dominating monomial of oe - (P ). We need to show that m is a dominating monomial of P . This can be done in the exactly same way as before, since The following theorem relates the notions of eventual absolute positiveness and dominating monomials, proving the correctness of the dominating monomial method. Theorem 4.2 A polynomial P is eventually absolutely positive iff the coefficients of all dominating monomials of P are positive. 2 Proof. Assume that P is eventually absolutely positive. We show that the coefficients of all dominating monomials are positive. From eventual absolute positiveness of P we know that there exists a - such that P is absolutely positive from -. From Theorem 3.1 we know that oe - (P ) must have all coefficients positive and non-zero constant term. Therefore also all coefficients of the dominating monomials of oe - (P ) are posi- tive. By Lemma 4.2, the set of all dominating monomials of P is the same as that of oe - (P ). Thus all dominating monomials of P have positive coefficients. Assume that the coefficients of all dominating monomials of P are positive. We need to prove that P is eventually absolutely positive. We begin by taking care of a trivial case where all other (non-dominating) monomials of the polynomial P also have positive coefficients. It is easily seen, that such a polynomial is absolutely positive from any - ? 0 and therefore eventually absolutely positive. Therefore assume that there exists at least one monomial with negative coefficient (negative monomial) in P . We partition the polynomial P into a sum of several sub-polynomials such that one of the sub-polynomials consists of all positive non-dominating monomials of P and that each of the remaining sub-polynomials consists of exactly one dominatingmono- mial of P (which we will call the leading monomial) and some (possible none) monomials dominated by the leading monomial (thus, these monomials are all negative.) Obviously, if each sub-polynomial is eventually absolutely positive, so is P . Thus, from now on, we will show that each sub-polynomial is eventually absolutely positive. Let Q be one of sub-polynomials. If Q consists of only positive monomials, it is absolutely positive from any - ? 0 and therefore eventually absolutely positive. Thus, from now on assume that Q has at least one negative monomial. First we will prove that Q is eventually positive. Let We claim that Q is positive from - 5 . To see this, first recall that x m dominates x e i for every i. Hence, we see immediately that for every i and for every x -, a ab a Thus, we see that for every x -, Q Q is eventually positive. It remains to show that for every partial derivative Q of there exists a - such that Q is non-negative from - . It is obvious that every vanishing derivative is non-negative from any -. Thus from now on let Q be a non-vanishing partial derivative of the polynomial Q. Note that Q again consists of a positive leading monomial (a derivative of the leading monomial of Q) and some (possibly none) negative monomials dominated by the leading monomial. Thus, by applying the idea used for showing the eventual positiveness of Q, we see that there exists a - such that Q is positive from - and therefore it is also non-negative from - . As there are only finitely many derivatives, we can take -m the maximum of the bound for positiveness of the polynomial Q and the bounds for non-negativeness of its derivatives. Then the polynomial Q is absolutely positive from -m . Thus Q is eventually absolutely positive. We showed that every sub- polynomial of P is eventually absolutely positive. Therefore P is also eventually absolutely positive. 2 4.3 Comparing the Complexity of the Method We know that the partial derivative method (Lankford) and the dominating monomial method have the same power. The following theorem compares their asymptotic complexity bounds. Theorem 4.3 (Computing Times) The computing times for worst cases are bounded from above as follows: ffl Partial Derivative Method (Lankford) O(d n ') Dominating Monomial Method O(n' log ') where n the number of variables, ' the number of monomials, d the maximum of the degrees in each variable. 2 Proof. 5 Actually, one can find a much smaller value for - than the one given here. But here we are only interested in proving the existence of a bound, and thus any bound suffices. The bound for the partial derivative method is immediate from the fact that the method performs up to d n differentiations of polynomials, where each differentiation takes O('). The bound for the dominating monomial method is immediate from the number of comparisons needed for sorting the monomials according to the partial order defined by the domination property, where each comparison takes O(n). 2 Conclusions Several results were presented in the paper. We determined how much partial the several known methods are for testing positiveness of polynomials. After that we proposed simpler and/or more efficient methods with equal power. There are several directions for the future work. Regarding the shifting method, one can use more sophisticated methods for the shifting. For example, the number of monomials necessary to be shifted can be significantly reduced by exploiting the structure of the polynomial (not all monomials have to be shifted, the shifting can be stopped after no negative monomials can be generated, summing new monomials need not be done for only positive monomials etc). Also, it will be interesting/important to find out experimentally how the new methods (shifting, dominating monomial) behave on small/moderate size inputs. --R Algorithms for the Geometry of Semi-Algebraic Sets Cylindrical algebraic decomposition II: An adjacency algorithm for the plane. An adjacency algorithm for cylindrical algebraic decompositions of three-dimensional space Termination of rewriting systems by polynomial interpretations and its implementation. Simplification of Truth Invariant CAD's and Solution Formula Construction. Improved algorithms for sign and existential quantifier elimination. Quantifier elimination for the elementary theory of real closed fields by cylindrical algebraic decomposition. Quantifier elimination by cylindrical algebraic decomposition - 20 years of progress Partial cylindrical algebraic decomposition for quantifier elimination. Termination of rewriting. Generating polynomial orderings for termination proofs. A combinatorial algorithm solving some quantiifier elimination problems. The complexity of deciding Tarski algebra. On the complexity of semialgebraic sets. An improvement of the projection operator in cylindrical algebraic decomposition. Improvements in CAD-based Quantifier Elim- ination Simple solution formula construction in cylindrical algebraic decomposition based quantifier elimina- tion Heuristic search and pruning in polynomial constraint satisfaction. Parallelization of quantifier elimination on a workstation network. Quantifier elimination for formulas constrained by quadratic equations via slope resultants. Approximate quantifier elimination. Generic quantifier elimination. Approximate Quantifier Elimination. A finite termination algorithm. Applying linear quantifier elimination. An Improved Projection Operator for Cylindrical Algebraic Decomposition. An improved projection operator for cylindrical algebraic decomposition. Solving polynomial strict inequalities using cylindrical algebraic decomposition. On the computational complexity and geometry of the first-order theory of the reals A parallel implementation of the cylindrical algebraic decomposition algorithm. Proving Polynomials Positive. Termination of rewriting. The completness of elementary algebra and geometry. A Decision Method for Elementary Algebra and Geometry. The complexity of linear problems in fields. Quantifier elimination for real algebra - the cubic case --TR --CTR Salvador Lucas, Practical use of polynomials over the reals in proofs of termination, Proceedings of the 9th ACM SIGPLAN international symposium on Principles and practice of declarative programming, July 14-16, 2007, Wroclaw, Poland Jrgen Giesl , Aart Middeldorp, Transformation techniques for context-sensitive rewrite systems, Journal of Functional Programming, v.14 n.4, p.379-427, July 2004 Jrgen Giesl , Ren Thiemann , Peter Schneider-Kamp , Stephan Falke, Mechanizing and Improving Dependency Pairs, Journal of Automated Reasoning, v.37 n.3, p.155-203, October 2006 Nao Hirokawa , Aart Middeldorp, Tyrolean termination tool: Techniques and features, Information and Computation, v.205 n.4, p.474-511, April, 2007
termination proofs;positiveness of polynomials;term rewrite systems
594182
The Design of the CADE-13 ATP System Competition.
Running a competition for automated theorem proving (ATP) systems is a difficult and arguable venture. However, the potential benefits of such an event by far outweigh the controversial aspects. The motivations for running the CADE-13 ATP System Competition were to contribute to the evaluation of ATP systems, to stimulate ATP research and system development, and to expose ATP systems to researchers both within and outside the ATP community. This article identifies and discusses the issues that determine the nature of such a competition. Choices and motivated decisions for the CADE-13 competition, with respect to the issues, are given.
Introduction Running a competition for Automated Theorem Proving (ATP) systems 1 is a difficult and arguable venture. The reasons for this are that existing ATP systems are based on different logics, are designed for different types of reasoning, require different amounts of user interaction, have different input formats, and may run on specialized hardware. Moreover, there is no clear work profile with respect to which ATP systems should be evaluated. However, a competition will make a significant contribution to the following important needs: ffl to evaluate the relative capabilities of ATP systems, ffl to stimulate ATP research in general, ffl to stimulate ATP research towards autonomous systems, ffl to provide motivation for implementing and fixing systems, ffl to provide an inspiring environment for personal interaction between ATP researchers, ffl to expose ATP systems to researchers both within and outside the ATP community. In other disciplines competitions are regularly used to stimulate research and development. The annual ACM computer chess championship [New94] is very well known and attracts much interest. Participants in the chess championship are clearly motivated to improve their pro- grams. Programming competitions are also popular [CRRP90, AKK93, Dem96], and encourage programmers to improve their skills. Competitions in machine learning [MMPS94], algorithm implementation [CIJ + 94], and other disciplines have all stimulated effort in their respective fields. In ATP there have already been competitions at the propositional [BKB92] and 1st order [Ove93] levels, which have encouraged researchers to improve their ATP systems. As an example, Otter's autonomous mode [McC94] resulted in part from the competition described in [Ove93]. As well as being of interest in its own right, an ATP system competition, or at least an examination of the issues involved in a competition, will provide insight into the more general notion of ATP system evaluation. For a long time the testing and evaluation of ATP systems has been very ad hoc. Results being published seldom provide an accurate reflection of the intrinsic power of the ATP system being considered. Inadequate evaluation has the potential to damage a field of research; bad ideas may appear to be good and are thus adopted and perpetuated 2 , while good ideas may appear bad and are hence discarded. A first step towards improving ATP system evaluation has been the release of the TPTP problem library [SSY94, SS96]. A common library of problems is a necessary, but not sufficient, tool for meaningful system evaluation. There are also other issues that need careful consideration. This examination of the issues surrounding an ATP system competition goes a fair way towards capturing the requirements for meaningful ATP system evaluation. It should be noted that using benchmarks and test beds (such as the TPTP problem library) in system evaluation requires special care (see also [HPC93]). In order to benefit fully from a competition, a thoroughly organized event, with unambiguous and motivated rules, is necessary. One approach to ATP system competitions, as was taken in We use the term "ATP system" to refer to the functional unity represented by the implementation of a reasoning calculus and its control, i.e., the implementation of a proof procedure. Geoff thinks that paramodulation is an example of this in ATP. the competition at CADE-11 [Ove93], is to run very specialized competitions with specifically selected problems. While this allows a detailed analysis and comparison of the proofs obtained by ATP systems, it does not provide a realistic comparison of the ATP systems in terms of their general usefulness. In [Ove93] Overbeek notes that the ATP community had "never [been] able to formulate an acceptable mechanism for comparing different systems". In order for a comparison of different ATP systems to make sense, it is necessary that all the systems should be attempting to capture a common notion of "truth", as is described in the Realist viewpoint in [Pel91]. Given this commonality across all systems, we believe that it is possible to set out rules for a competition that can determine a winner, relative to some clearly specified constraints. For some issues relevant to an ATP competition, inevitable constraints emerge. For other issues there are several choices, and a decision has to be made for the competition. The arising issues, choices, and decisions are described in this paper. Although this paper focuses on the issues relevant to experimental comparison of ATP systems, it is important to acknowledge the contribution of analytic approaches, e.g., [Let93, Dun94, Pla94]. These approaches investigate metrics such as search space duplication, search tree size, Herbrand multiplicity, number of variables to bind, and connectivity. At the propositional level analysis can be reasonably accurate, and is thus important for that limited situation. At the 1st order level the analysis is significantly more difficult, and results obtained so far are insufficient. Complete analysis of search guidance at the 1st order level is of course impossible (for otherwise 1st order logic would be decidable!). It is for these reasons that we have focused on experimental comparison of ATP systems; in this way meaningful judgements can be made. Competition Issues There are four types of issues (and resultant choices and decisions) relevant to the design of an ATP system competition: 1. Issues based on the fundamental nature of ATP. 2. Issues associated with common practice in ATP research. 3. Issues arising from the application of ATP systems. 4. Issues directly related to the competition. An overview of the issues that are discussed in the following sections is given in Table 1. The table also shows the two types a particular issue most belongs to, according to the four types presented above. Type Issue ATP System Properties 2,3 Degree of Automatization : How much user interaction is allowed? 2,3 Soundness and Completeness: How do soundness and completeness affect evaluation? 2,3 Proofs and Satisfiability Checking: Should the systems be expected to find proofs or models? 2,3 Proof Objects : Are explicit proof objects required? 1,2 Monolithic and Compositional Systems: Does the system structure matter? Problem Presentation 1,2 Scope of Logic : Which logics (e.g., Classical, Modal) should be allowed? 1,2 Built-in Theories : Are built-in theories (e.g., equality) allowed? 2,3 Input Language : What input language (e.g., CNF, FOF) should be used? 3 Clause Type Information : Is usage of such information acceptible? Clause and Literal Orderings : How can bias be avoided? Problem Selection 2,4 Problem Source : Where should the problems come from? 4 Problem Difficulty : How difficult should the problems be? 2,4 Problem Axiomatization : What about biases in problem axiomatizations? 4 Number of Problems : How many problems should be used? Resource Limits - Should a runtime limit apply per problem or for all problems altogether? What runtime limit should be imposed? 4 Hardware and Software resources : What hardware and software should be available for the competition? System Evaluation Should a distinguished "winner" be determined? 2,3 Performance Metrics : What metrics should be used for evaluation? 3,4 Ranking Scheme : How should performance metrics be combined for ranking? Table 1: Relevant issues for an ATP system competition design. 3 ATP System Properties 3.1 Degree of Automation From a user's viewpoint, ATP systems are of two types : fully automatic systems 3 and interactive systems. Currently, and presumably in the future, the most important application of fully automatic ATP systems is as embedded systems in more complex reasoning environments, where they serve as core inference engines. This setting is shown in Figure 1. The environment may call on the fully automatic ATP system to solve subtasks within the operation of the overall system. The subtasks are thus generated by the environment. Examples of fully au- 3 Astrachan and Loveland [AL94] define three types of fully automated deduction (F.A.D.) systems . "pure F.A.D. systems, where only the problem can be entered ., the strong F.A.D. systems, where parameters can be set only once, and the weak F.A.D. system, where many experiments with parameters are permitted, but each adjustment requires a restart from the beginning of the problem. For the purposes of this discussion, automatic ATP systems are the pure and the strong F.A.D. systems; weak F.A.D. systems are considered to be interactive. tomatic systems are Otter [McC94], SETHEO [LSBB92], SATCHMO [SGar], the CLIN series [LP92, CP94, Ale95], and SPASS [WGRar]. Problem Solution Language Transformers Interface User Management Data System Control (fully automatic) ATP System atomic proof requests proofs/ models Figure 1: The embedding of an ATP system in some reasoning environment. Interactive systems find application in hardware and software verification research [BM90, HRS90, ORS92, WG92], the formalization of informal mathematical proofs [FGT90, Pau90], the teaching of logic [Por94], and as tools of mathematical and meta-mathematical research 86]. Examples of interactive systems NQTHM [BM90], and NUPRL [CAB systems typically embed some form of a fully automatic ATP system as a subsystem 4 , and are an example of the reasoning environments mentioned above. At any point the user of an interactive system may ask that some (sub)problem be submitted to the fully automatic component. Both fully automatic and interactive theorem proving systems are important. It is clear, however, that the two types of system are very different. In terms of a competition, it would not make sense to compare fully automatic systems with interactive systems; the qualities that are required of the two types are largely distinct. In particular, an interactive system must be judged largely on its ability to assist the user in the organization and tracking of complex proof objects [PS94]. While for fully automatic systems the system runtime is the dominating issue. It may be the case that there are ATP system evaluation criteria that are independent of whether the systems are automatic or interactive. However, at this stage no such criteria are clear. Thus the ATP competition can compare either fully automatic systems or interactive systems, but the two types cannot be compared with each other. Many of the assessment criteria for a competition between interactive theorem proving systems would be very subjective, as individual user abilities and tastes would necessarily factor into the comparison. There are some criteria, e.g., "user-friendliness", that do not appear to have any objective measure at all. In contrast, the criteria for comparing fully automatic ATP systems are mostly objective, as described in this paper. Also, recalling that interactive systems typically embed a fully automatic ATP system as a subsystem, a comparison of the fully automatic subsystems of interactive systems should be part of a comparison of interactive systems 5 . Decision : The ATP competition will be between fully automatic ATP systems. No 4 At the current stage, most interactive systems contain a quite rudimentary and simple theorem proving part, while (justifiably so) most effort is spent on the development of adequate logics, proof tactics, and interfaces. 5 Today, a direct comparison of the fully automatic subsystems would be difficult, because due to their human interaction will be allowed. In the rest of this paper attention is limited to fully automatic ATP systems, and they will be referred to simply as ATP systems. Note that the decision to focus on fully automatic systems does not exclude the possibility of a competition between semi-automated or interactive systems. Indeed, such a competition would be of interest, if the difficulties involved can be adequately resolved. A consequence of the limitation to fully automatic systems is that the competition will not appreciate system properties which relate to human computer interaction, and therefore will not necessarily determine the best "human assistant". Rather, the competition will assess the pure reasoning capabilities, and nothing else 6 . The intention is to foster automation. In particular, the competition aims to identify powerful general purpose deduction procedures, and to encourage automatic problem analysis and system configuration. The development of deduction procedures has been a focus of ATP research since its inception, and is certainly a mature enough area to evaluate. On the other hand, automatic problem analysis and system configuration are underdeveloped. In various research groups there is much knowledge about how to analyze problems, and how to configure that group's ATP system. This knowledge is exploited when the ATP system is evaluated, by having the knowledgeable person manually configure the system for each problem. Due to this common practice of manual configuration, the effort to capture and implement problem analysis and configuration techniques is rarely taken. Implemented systems often feature strong reasoning components but have very poor automatic control. This makes it difficult for other users of such systems to get the best possible performance from the systems, and makes the systems unsuitable for use as embedded systems. Hopefully this ATP competition will encourage researchers to formalize and capture their analysis and configuration skills in their ATP systems. 3.2 System Soundness and Completeness Traditionally, ATP systems are designed to be deductively sound and complete: every answer returned is deductively correct and if a solution exists it is eventually deduced. However, ATP systems are usually too complex to formally verify either property. Thus systems are tested experimentally to detect violations of these properties. While soundness can be assumed for extensively tested systems (at least if a large number of the produced solutions have been verified), completeness is more difficult to assess. The reason for this is that the inability to find a solution is not necessarily a sign of incompleteness - the problem could simply be too hard for the system. In fact, due to the finite amount of resources allocated to any particular ATP system run, every search for a solution has limited completeness, independent of the system's theoretical completeness. A special case of incompleteness is a bugged ATP system, which crashes on some problems. The question then arises, how should unsound and/or incomplete and/or bugged systems be treated in the ATP competition? From a users point of view, there is little difference between not finding a solution due to system incompleteness, not finding a solution due to a resource limit being reached, and not finding a solution due to a system crash. In all cases the user learns nothing about the solution to the problem. There are also positive aspects to each form of failure. If a resource limit is reached then the user can optimistically increase the resources available (although this is futile simplicity no modular system design is used and an isolated test is not easily possible. 6 But note: the success of ATP systems as automated reasoning assistants will depend on their usability by non-experts. For high usability the dependence on user interaction, in terms of low level guidance, needs to be minimized. This argues that even in interactive systems less (low level) interaction is better. if the resource limit was reached due to incompleteness, and further, the supply of resources typically cannot deal with the exponential growth in demand for them). If a system terminates quickly, due to incompleteness or a bug, then time is saved and there is hope of having the system repaired. In practice a bugged or incomplete system may solve more problems within some time limit than a bug free complete system, and therefore may be more useful. Decision: In the ATP competition, systems must be sound but may be incomplete and may be bugged. The soundness of competing systems will be assessed before the competition with specific test problems. Systems found to be unsound will be excluded from the competition. A property associated with soundness and completeness is that of stability. If an ATP system finds a solution to a problem, then the system must be able to find that solution again if required. ATP systems that simply 'get lucky' are of little use. Repeatability of testing is also a standard requirement of the scientific method. Decision: For every solution found in the ATP competition, the solution process must be reproducible by running the system again. If randomness plays a role, reproducibility must still be ensured, e.g., by using pseudo-random numbers and recording the seeds. 3.3 Proofs and Satisfiability Checking There are two distinct classes of problems that may be presented to ATP systems. Firstly there are problems that require the proof of a theorem, and secondly there are problems that require satisfiability to be established (via the generation of a model). Both types of problem are of interest. However, mixing proof performance and satisfiability assessment would blur the interpretation of competition results. One reason for this blurring is that most ATP systems are designed for one or the other purpose, but not both (exceptions are, for example, the RAMCEC algorithm [BCP94] and hyper-linking [LP92]). As a result, a comparison of specialized systems of the different types is arguable. In the future more dual purpose systems may be developed, in which case it may not be necessary to divide the work profile for a competition, as is now the case. Thus the work profile for a competition should not require both theorem proving and model generation. Historically, more emphasis has been placed on the ability to find proofs, and therefore this is considered to be the more important issue (albeit this is not the case in some applications). : The ATP competition will focus on theorem proving performance. In order to implement this decision, only theorems will be used (see Section 5.1). 3.4 Proof Objects Depending on its generality and purpose, there are various responses that an ATP system may make when given a problem. A proof or a model may be returned, or only an assurance that a proof or model exists may be given. There are some ATP systems, e.g., completion based systems, that conserve resources by not building a proof object during their search. Such systems merely retain enough information to build the proof object later, if required. Other ATP systems, e.g. model generation systems, are not able to build a proof object, and can only give an assurance that a proof exists. There is no evidence that usage of an ATP system as an embedded system typically either requires or does not require delivery of proof objects or models, and it is desirable to make the ATP competition accessible to as many ATP systems as possible. Decision : The ATP competition will not require ATP systems to return proof objects. The soundness testing (see Section 3.2) will ensure that competing systems do not make invalid claims to have found a proof. This decision evidently gives some bias towards those systems that do not build a proof object. This bias is considered to be slight, and justified in terms of extending the scope of the competition. However, the added functionality of ATP systems that do produce proof objects should be acknowledged. Decision : In the presentation of the ATP competition results, it will be noted which of the systems do produce complete proof objects. 3.5 Monolithic and Compositional Systems Today, a large number of different calculi, inference rules, and proof procedures exist. Ideally, an ATP system would have all known techniques available, and would, during a particular solution search, switch back and forth between the most suitable techniques. First steps in this direction are ATP systems which are formed from a collection of different proof procedures, and which analyze the given problem in order to decide which procedure or inference rules to use. A well known example of this type of system is Otter in its autonomous mode [McC94], which enables or disables inference rules according to a check whether the problem is propositional, whether it is Horn, and whether it contains equality and/or equality axioms. An ATP system in which no components are chosen as alternatives to others, based on the given problem's characteristics, is called a monolithic ATP system. An ATP system that runs as one of several possible distinct monolithic systems, as determined by the given problem's characteristics, is called a compositional system. Combining several monolithic systems in a parallel manner also results in a compositional system. The definition does not limit monolithic systems to a single calculus or rule of inference, but requires that any switching back and forth between different calculi and/or rules is done in an integrated way, and occurs continuously during the proof process 7 . Compositional systems provide a valuable approach to building more versatile systems, and compositional systems can be expected to outperform monolithic systems. Conversely, monolithic systems can be improved by integrating more specialized components. Altogether, both the construction of compositional systems and of monolithic systems are valuable. Research into compositional systems will improve knowledge of which approach is best for what problems, while research into monolithic systems will push the performance of uniform approaches (which form the core of compositional systems) to the limit 8 . The results may one day lead to a type of monolithic system which integrates all available techniques. Altogether, the competition should not put either of these approaches at a disadvantage. Decision : In the ATP competition there will be two categories: Open, which includes both types of systems, and Monolithic, which includes only monolithic systems. Any controversy about which category a system belongs to will be resolved by the competition panel (see Section 7.1). 4 Problem Presentation 4.1 Scope of Logic Many "hot logics" in ATP research, such as higher order logics, modal logics, and temporal logics, are very interesting. Sufficient research has been done to support a comparison of ATP 7 Rigid definitions of monolithic and compositional systems seem hardly possible. For any formal definition made it seems that an ATP system that violates the intuitive notion can be contrived. 8 This issue can also be viewed from the evaluation viewpoint: the goal of evaluating a system is not simply to see how well it performs, but also to learn why it achieves the observed performance. This is difficult for compositional systems, and therefore it is desirable to obtain evaluations of their individual components as well. \Phi \Phi \Phi \Phi \Phi \Phi 1st order \Phi \Phi \Phi \Phi \Phi \Phi propositional non-propositional no-equality pure-equality mixed unit non-unit Figure 2: Classification of 1st order logic problems. systems within many of these logics, as well as common classical logics. However, a comparison of ATP systems across logics would be unlikely to make sense, for several reasons: the types of output and expected performance vary from logic to logic, different logics may have different semantics, and specific logics may be particularly designed for specific applications (making a fair problem selection difficult). Thus the problems for the competition should include problems from any one (but only one) logic. The wide-spread use of 1st order logic in ATP suggests it as a starting point; specialized competitions for other logics can still occur in the future. Furthermore, transformation techniques (e.g., [Ker91, Ohl91]) allow the use of 1st order logic ATP systems for solving problems in some other logics. Decision : The ATP competition will be restricted to theorems expressed in classical 1st order logic. Classical 1st order logic can be subdivided as shown in Figure 2. The leaves in Figure 2 correspond to syntactically identifiable problem classes, and it is trivial to determine which class a given problem belongs to. It would be possible to use problems from all classes together, or to have different competition categories for the classes. For some of the problem classes specialist ATP systems, that employ the appropriate tech- niques, exist. Specialist systems can be expected to perform better on problems in their class than systems not specialized for that class. A best overall ATP system can be built easily by choosing the best system for each problem class, and using a trivial switch to invoke the appropriate system. Trivial switch systems are of high interest to users, because they provide better overall performance than any of their components. For the competition, however, there is little interest in evaluating the actual switches within such systems, due to their simplicity. Thus an overall performance evaluation of such systems is not desirable. Instead it is better to evaluate the component systems separately, and from that conclude the overall performance achievable by their combination. Decision : Specific problem classes of 1st order logic (according to the classification above), for which specialist ATP systems exist, should be treated in separate categories in the ATP competition. Ideally the ATP competition would have separate categories for all of the problem classes. However, merging or omitting some classes may still be appropriate. The development of specialized ATP systems for propositional logic is largely separated from ATP for non-propositional logic, and there are separate specialized competitions for evaluating propositional systems 9 . There is no need to duplicate those competitions. Decision : There will not be a category for propositional problems in the ATP competition. In the original competition announcement the no-equality and mixed problem classes were merged, for the following reasons. Firstly, systems that include specialized equality treatment 9 The next such event will be at the "International Competition and Symposium on Satisfiability Testing", to be held in Beijing, China. are typically competitive on no-equality problems, and therefore can be evaluated on such problems. Secondly, a natural formulation of many interesting problems involves equality, and therefore most state-of-the-art ATP systems include specialized equality treatment. These motivations indicate that it is acceptable to keep these two classes merged, and it is desirable for the competition to maintain its announced format. Decision : The no-equality and mixed problem categories will be combined in the ATP competition. Most pure equality problems known to the ATP community are written in Clause Normal Form (see Section 4.3) as unit equality problems. There are specialized unit equality ATP systems which cannot handle any other format of problem, so it is necessary to have a separate category for these systems. Decision : There will be a separate category for unit equality problems in the ATP competition. Non-unit pure equality systems are able to handle no-equality and mixed problems by translating such problems to pure equality, and are able to compete with no-equality and mixed systems (even if at a disadvantage, the extent of which is currently unknown). The small number of commonly known non-unit pure equality problems seems insufficient for a separate category. Decision : The no-equality, mixed, and non-unit pure equality problem categories will be combined in the ATP competition 10 . 4.2 Built-in Theories A theory can be built into an ATP system, rather than expressing it by a set of axioms (e.g., equality theory). This approach is chosen in order to provide specialized, and hopefully more efficient, treatment of that theory. Systems capable of such specialized action should not be put to a disadvantage by including axioms unnecessary for them. Decision : In the ATP competition, theory axioms that have been built into an ATP system can be removed from the input to that ATP system, using an automatic tool. 4.3 Input Language The problems submitted to an ATP system may be expressed in full First Order Form or possibly in some constrained subset of 1st order logic, such as clause normal form (CNF). Therefore the ATP system must, in general, be able to deal with FOF. An examination of current ATP systems reveals three possible approaches (see Figure 3): ffl A non-specialized clausifier is used to convert the submitted problems to CNF, and the ATP system uses the CNF versions of the problems (CNF systems). ffl A specialized clausifier is prepended to a CNF ATP system, and this combined system deals with the submitted problems directly (Clausifying systems). ffl The ATP system does it's processing in FOF, and thus deals with the submitted problems directly systems). A comparison of CNF systems and FOF systems is not sensible, as CNF systems are disadvantaged if FOF problems are used, and FOF systems are disadvantaged if CNF problems are used. A comparison of FOF systems and clausifying systems is possible using FOF problems, and a comparison of CNF systems and clausifying systems is possible using CNF problems. A comparison of CNF systems and clausifying systems would constitute a comparison of the clausifiers as well as a comparison of the ATP systems' basic theorem proving capabilities. In future ATP competitions these categories should be separate. Non-specialized Clausifier atomic proof requests in FOF System CNF proofs/ models Clauses atomic proof requests in FOF System CNF Specialized Clausifier proofs/ models proofs/ models atomic proof requests in FOF FOF System Figure 3: Alternative system designs regarding the input language. This would blur the interpretation of the results. Also, current ATP research is dominated by CNF systems. There are applications that generate CNF problems directly (e.g., test pattern generation for electronic circuits). These factors indicate that a CNF competition is particularly viable. Decision : The ATP competition will use theorems expressed in Clause Normal Form. This constraint does not exclude the other two types of ATP systems from the competition. However, such systems will not be able to take advantage of their abilities to deal with FOF problems 11 . CNF problems can be either Horn or non-Horn, and there are particular techniques designed for solving each type. In the same way that the different classes of problems shown in Figure 2 are treated separately, ideally Horn and non-Horn problems should be treated separately. It is not clear what techniques are best for which type of problem. ATP systems for non-Horn problems can perform well on Horn problems, and some non-Horn problems are handled well by systems designed for Horn problems. There is no corpus of ATP systems that are specialized for either type of problem. For switch systems, classifying problems as being Horn or non-Horn is easy, but selecting superior specialists for either type is not. Decision : Horn and non-Horn problems will be used together in the ATP competition. 4.4 Clause Type Information. A clause of a CNF problem may be classified, depending on whether it is part of the conjecture, part of assumptions made, or part of the underlying theory. This clause type information may be useful to an ATP system. In the application of ATP systems such information will typically be available, and therefore should be used. Decision : Systems may use clause type information in the ATP competition. 11 A competition using FOF problems is planned for the future. Having established the quality of their CNF ATP systems in this competition, developers will be able to spend time building specialized clausifying programs ready for a subsequent FOF competition. 4.5 Clause and Literal Orderings. A possible source of bias in the competition is the particular ordering of the clauses within each problem, and the literals within clauses. For system evaluation it is necessary to ensure that pre-tuning (possibly accidental) to a particular clause or literal order is of no benefit. Decision : In the ATP competition, random clause and literal orderings will be used. The same clause and literal ordering will be used for all systems. 5 Problem Selection 5.1 Problem Source There are several possible sources for the unsatisfiable sets of clauses to be used in the ATP competition. The competition could use problems supplied by the entrants, could use problems selected from those commonly used by the ATP community, or could use new problems designed especially for the competition (there are other options too). If entrants were to supply the problems, they would almost inevitably supply (pathological) cases for which their ATP system is particularly effective and that are difficult for other com- petitors. The problems would not reflect any real work profile of ATP systems, and the ability of the entrant to make such selections would influence the results. Designing new problems for the competition would be acceptable, but requires additional effort and could still (unintention- ally) be biased towards a particular system. Using problems from the ATP community will to some extent reflect common usage of ATP systems. The TPTP Problem Library [SSY94, SS96] contains a broad selection of such CNF theorems, as required. Decision : The unsatisfiable sets of clauses to be used in the ATP competition will be selected from the TPTP problem library 12 . The tptp2X utility, distributed with the TPTP problem library, will be used to remove unnecessary equality axioms (see Section 4.2) and to reorder the problem clauses and literals (see Section 4.5). The tptp2X utility can be used to produce a pure equality representation of problems for non-unit pure equality systems (see Section 4.1). TPTP problems provide clause type information (see Section 4.4). Use of the clause type information is possible either directly, by reading the TPTP problem, or indirectly, by using a tptp2X formatting option that takes clause type information into account (see [SS96] for details). In principle, knowing the problem source allows problem-specific information to be precomputed and accessed during the competition (in the extreme, proof look-up). Such activity is contrary to the purpose of the competition, and is disallowed. Global optimization of an ATP system for the TPTP is acceptable, because such an optimization seems likely to work well as a general technique, due to the size and scope of the TPTP. 5.2 Problem Difficulty The TPTP problem library contains a broad selection of problems in CNF, most of which are unsatisfiable sets of clauses, as required. Suitably difficult problems must be selected for the 12 In a sense, entrants influence the problem selection by their submission of problems to the TPTP. However, due to the selection process this constitutes a very weak influence. competition. The TPTP rating system categorizes problems as follows: solvable by all state-of-the-art ATP systems difficult solvable by some state-of-the-art ATP systems unsolved solvable by no state-of-the-art ATP systems 4 open theorem-hood unknown Here "state-of-the-art ATP systems" means specialist ATP systems for the class of the problem, according to an extended version of the classification given in Section 4.1. Using easy problems would not differentiate sufficiently between systems performances. On the other hand, each theorem must be provable by at least some of the systems. Decision : The ATP competition will use TPTP problems with difficulty rating 2. 5.3 Problem Axiomatization Many of the problems in the TPTP problem library have been taken from publications and existing problem libraries. Some of the theorems contain only those axioms necessary for a particular proof, or contain lemmas that assist a proof. Also, the axioms are often selected to suit a particular ATP system, and are therefore biased towards that system. For each problem using such a non-standard axiomatization, the TPTP provides a version of the problem using a standard axiomatization containing all the axioms and no lemmas. The use of a standard axiomatization is desirable because it reflects the typical initial situation when formulating a new problem, and also avoids any bias towards a particular system. Decision : Only standard versions of TPTP problems will be used in the ATP competition. 5.4 Number of Problems Ideally, each ATP system would be evaluated using all eligible TPTP problems. However, the limited time and hardware resources available determine a maximal number of problems 13 . At the same time, a minimal number of problems is necessary for meaningful system evaluation. The number of problems to be used needs to be determined. The minimal number of problems will be determined to ensure sufficient confidence (say 85%) that the competition results are the same as would be obtained using all eligible problems. For an evaluation based on the number of problems solved, as is essentially the case for the ranking schemes to be used in the ATP competition (see Section 7.3), and assuming a worst case proportion of problems solved (50%), this is computed by first computing n 0 , the minimal number assuming an infinite number of eligible problems: and then adjusting for the actual number of eligible problems as follows [Wad90]. Decision : The minimal number of problems to be used in the ATP competition will be: 13 A complete evaluation pursued in a larger time frame is an interesting project by itself. Available TPTP result data for various systems have been collected in [SS95]. These results, however, are difficult to compare, because they have been obtained under differing constraints and on differing hardware platforms. min number of problems = 23:04 \Theta number of eligible problems number of eligible problems The rating of TPTP problems is currently being completed. From that the number of eligible problems, and hence the minimal number of problems, will be determined. The maximal number of problems is determined by the resources available for the competition (the number of workstations and the total time), the number of ATP systems competing, and the minimal time limit that can be imposed on each proof attempt (see Section 6.1). The maximal number of problems to be used in the ATP competition is: number of problems = number of workstations \Theta time for competition number of ATP systems \Theta minimal time limit 6 Resource Limits 6.1 Time Resource Problems may be presented to an ATP system one-at-a-time or in batches. In both cases the ATP system will have to deliver results within some time limit. Also, the time available for the competition is limited. It is thus necessary to impose a time limit in the competition, either on individual proof attempts or on the total time for all proof attempts. If the problems are presented as a batch and an overall time limit is used, that will allow the ATP systems to examine all the problems and perform once-only tuning before starting the proof attempts. In terms of the competition, an overall time limit would simplify the winner assessment; the system that proves the most theorems within the time limit is the winner, and if more than one system proves the same number of theorems then the total time taken to prove those theorems is compared (this ranking scheme has been adapted in Ranking Scheme A in Section 7.3). In applications of ATP systems, problems are likely to be presented one-at-a-time, as and when the problems are generated. Even if an ATP system is presented with a batch of problems, they are likely to be treated individually, again requiring individual time limits to decide when an unsuccessful attempt to find a solution should be abandoned 14 . Also, an evaluation based on individual time limits still provides relevant performance information for the batch case, but not vice versa. Decision : In the ATP competition, a time limit will be imposed on individual proof attempts. A minimal time limit is necessary so that the ATP systems spend most (say 95%) of their time on proof search. This ensures that their deductive capabilities are compared, rather than their initialization and termination efficiencies. The non-proof time of each ATP system will be estimated by measuring the time that it takes to prove a trivial theorem, e.g., :p. The geometric average of the system's non-proof times will be taken to determine the non proof time. Then, to ensure that the systems spend 95% of their time searching for a proof, the minimal time limit = non proof time \Theta 20. The time limit to be imposed on each proof attempt depends on the resources available for the competition, the number of ATP systems, and the number of problems (which in turn is determined by the minimal time limit; see Section 5.4). Decision : The time limit to be imposed on each proof attempt in the ATP competition will be: 14 Processing all problems on a single processor via multiprogramming would not require imdividual time limits, but realistically allows only as many proving processes as can fit in memory. number of workstations \Theta time for competition number of ATP systems \Theta number of problems The way that the number of problems is determined in Section 5.4 ensures that time limit - minimal time limit. In Section 7.2 it is explained that the granularity of timing will be one second. Thus the minimal non proof time is one second, so time limit - seconds. It is important to note that the imposition of a time limit means that the competition results will have to be viewed as modulo that time limit. A different time limit may produce a different ranking of the systems. Results could be computed for a range of time limits, up to the time limit used, and this may give further insight into the relative abilities of the systems. If the ranking remains stable for a reasonable range of time limits, this provides empirical evidence that the ranking reflects a time limit independent ranking of the ATP systems. Decision : The ATP competition results will be plotted as a function of the time limit. 6.2 Hardware and Software Resources ATP systems have been developed using a wide range of hardware and software resources. The most commonly used computers in ATP research are UNIX based workstations. Many current ATP systems do, or can be made to, run on such workstations. Other hardware used includes PCs, LISP machines, parallel computers, and distributed networks of computers. The ATP competition could limit itself to the common workstation environment, or could allow the use of other hardware. Running all ATP systems on the same hardware allows a direct comparison of the systems, and is thus the preferred environment for the ATP competition. The host institution of the competition is able to provide a homogeneous set of UNIX workstations, thus making this option available. However, some ATP systems have been specially designed for a particular computer, and cannot easily be ported. The host institution will be unable to supply the range of specialized hardware that the ATP systems could use. Although a direct comparison of ATP systems running on widely disparate hardware is not meaningful, performance data from such systems does allow a comparison of the overall performance from a user's perspective. Since it would be of interest to the ATP community to see the performance of all systems, ATP systems that run on specialized hardware need to be catered for in the ATP competition. Decision : Each of the ATP competition categories will be divided into two subcategories : General Hardware and Special Hardware. ATP systems in the General Hardware subcategories will all be executed on the same type of UNIX workstation, as supplied by the host institution. ATP systems in the Special Hardware subcategories will be allowed to use any hardware brought to the competition or accessible via the InterNet. Common languages used to implement ATP systems are C, Prolog, and LISP. Other languages have also been used. For the ATP competition, the major difference between implementation languages is whether the code is interpreted, as is often the case for Prolog and LISP, or compiled to machine code, as for C. ATP systems that run in an interpreted environment are disadvantaged by the time taken to start up the interpreter, and by the relative slowness of interpreters compared to compiled code. It would be possible to provide separate categories for the two types of implementation. There are some implementation environments that do not fall clearly into one or other of the two types. Many vendors of Prolog and LISP interpreters also provide compilers, thus alleviating the problems to at least some extent. Further, it is also important to acknowledge the effort taken to code an ATP system in a language such as C, and the benefit derived is competition will not be concerned with the languages used to implement the ATP systems. 7 System Evaluation 7.1 Winner Assessment It is arguable if an overall "winner" should be assessed. The reason for this is that potential entrants may be frightened off by the fact that some other system may be believed to perform better than theirs. The ATP competition could avoid this problem by simply reporting the performance data of the ATP systems and allowing observers to draw their own conclusions. Not determining a winner would leave out much of the spice of the competition, and would remove much of the motivation for improving the capabilities of current systems 15 . It must be remembered that the assessment done in the competition will be with respect to the decisions made in this paper, i.e., the winner may not be the best ATP system in a general sense. Rather, the ranking will simply provide feedback to potential users about which system is currently the most adequate for their particular use. An useful byproduct of the competition will be the charting of the performance data; that may reveal interesting properties of systems that do not win. Decision : The ATP competition will both determine a specific ranking of the ATP systems and present a listing of the performance data. Two possible ways of determining a ranking are to have a quantitative ranking scheme or to have a judging panel. A quantitative ranking scheme can be implemented mechanically (as a computer program, such as the RoboJudge described in [AKK93]) and checked for bias, but has little flexibility. A judging panel is more flexible, can take into account more aspects, and can impose intuitive judgment on the results. The benefits of both approaches are desirable. Decision : The ranking of ATP systems in the ATP competition will be done by quantitatively evaluating system performance and having that evaluation vetted by a panel of ATP researchers. 7.2 Performance Metrics There are many criteria by which ATP systems can be evaluated. Issues such as completeness, soundness, correctness, and proof performance are all of direct interest to ATP researchers. Soundness (correctness) should be required in-so-far as falsely stating theoremhood of a non- theorem renders the system untrustworthy, and therefore useless. Completeness plays a role in-so-far as systems capable of proving more theorems should be ranked higher than systems proving less. Incomplete systems, systems that fail to find a proof due to resource constraints, and those that crash because they are bugged must be compared (see Section 3.2). Proof performance can be measured in many ways, including runtime, number of inference steps (proof size), and proof tree depth (if proofs are represented as trees). In the broader context of computing science, issues such as software quality (portability, modularity), user friendliness, execution environment, etc., may also be of interest. It is necessary to decide what criteria should be used to evaluate the ATP systems in the ATP competition. Would you train hard for the Olympics if you could not stand on a pedestal when you won? The panel will consist of Peter Andrews, Alan Bundy, and Jeff Pelletier. The broader issues of computer science are of lesser interest to the ATP community, and factoring them into the evaluation would blur the focus on ATP. Also, no generally accepted metrics exist for the broad issues, and evaluation would become a matter of taste. Decision : The ATP competition will evaluate the ATP systems in terms of issues directly relevant to the ATP community. The quantitative ranking scheme needs to observe performance metrics of the ATP systems, and combine those values to produce a ranking of the systems. Quantitative performance metrics available for evaluation are: ffl Calculus-relative measures number of inferences performed (ideally split according to the inference rules used) number of unifications performed (ideally split into failed and succeeded unifications) proof size (e.g., length, depth) ffl Absolute measures number of problems solved runtime (possibly split into startup time and search time) memory usage For the purposes of the ATP competition the measures used must be independent of the ATP systems, so that values can be meaningfully compared. Measures such as the number of inference steps and proof length are not suitable because the units of measure can vary from calculus to calculus, and from implementation to implementation. System independent measures that are readily obtained are the number of problems solved, runtime, and memory usage. The number of problems solved and the runtime are direct indicators of the quality of an ATP system. Memory usage is important in so far as it can affect runtime. If an ATP system requires less than the available memory, then the effect on runtime is negligible. If an ATP system requires more than the available memory, then either the system cannot handle the problem, or swapping increases the wall-clock time of the computation. Therefore the effect of memory usage can be subsumed in a proper definition of runtime. Decision : The number of problems solved and the runtime will be used for winner assessment in the ATP competition. The memory usage will also be recorded and presented. The are two reasonable ways of measuring runtime: CPU time taken, and wall-clock time taken. The advantage of CPU time is that it is easy to measure and it is independent of system influences such as external load, daemon processes, memory size, disc performance. However, CPU time seems inappropriate if swapping occurs because it does not reflect the user's perception of the runtime. Wall-clock time takes swapping into account, but is dependent on system influences, and therefore can be difficult to measure in a reproducible manner. From the developer's viewpoint, CPU time is more interesting. From the user's viewpoint, wall-clock time is more relevant. Decision : For runtime, both CPU time and wall-clock time will be measured and reported in the ATP competition. In the General Hardware category, the choice of which time measurement is used for winner assessment will depend on the computing environment. If no swapping occurs then: Decision : In the General Hardware category CPU time will be used for winner assessment. If swapping does occur, and the wall-clock time measurements are stable and representative of the time required for the computation (essentially CPU time plus time required for swapping), then: Decision : time will be used for winner assessment. Otherwise: Decision : CPU time plus an estimate of the swapping time will be used for winner assessment. In the Special Hardware category CPU timings are typically incomparable. In contrast, wall-clock times can be compared, in the context of the hardware used. Decision : In the Special Hardware category wall-clock time will be used for winner assessment. The precision of time measurement on a computer system is limited. In order to reduce the effect of this, and to emphasize significant rather than marginal differences, the competition timing will be discrete. Decision : In the ATP competition timing will be done in units of one second. In particular, the minimal time a system can take to find a proof is one second. 7.3 Ranking Schemes The ATP competition must have a ranking scheme that combines the performance metrics to produce a ranking of the ATP systems. The ranking scheme must have certain properties: Monotonicity requirements: ffl The ranking improves as the number of problems solved increases. ffl The ranking improves as the time taken to solve the problems decreases. Required invariants: ffl For systems that solve the same number of problems, the one that takes the least time to solve them obtains a better ranking. ffl For systems that take the same amount of time, the one that solves the most problems obtains a better ranking. ffl Systems which solve the same number of problems in the same total amount of time obtain the same ranking. The issue of how to combine the performance metrics to obtain a ranking scheme is a contentious one. Two different quantitative schemes for determining a system ranking, representing different emphases between number of problems solved and time taken, have been developed. For both schemes higher scores are better. Both schemes use the total time takenas a parameter. This value is the sum of the time taken over all problems, including those for which no proof is found, in which case the time limit is used. Ranking Scheme A. This ranking scheme focusses on the ability to find as many solutions as possible. The idea is to rank the systems according to the number of problems solved, and to further differentiate by considering the number of erroneous claims that no proof exists, the number of system errors, and the runtime. Since for erroneous claims and system errors the full time limit is used as system runtime, considering only the runtime accounts for all the latter issues. Thus, each system is given a score: number of problems solved \Gamma total time taken \Gamma best total time taken where best total time taken = least total time taken by a system that solves the same number of problems, and number of problems \Theta time limit. Achievable scores range from 0 to the number of theorems. Ranking Scheme B. This scheme measures solutions per unit time. Each system is given a score: number of problems solved total time taken Achievable scores range from 0 to 1. The schemes place different emphasis on the two performance metrics: scheme A puts emphasis on the number of problems solved while scheme B balances the emphasis. This difference can lead to different rankings of ATP systems. Under scheme A a system which solves the most problems will win, while under scheme B this is not necessarily so. Each ranking scheme suits specific user requirements, and thus neither can be ignored. Decision : In the ATP competition, in each competition category a winner according to Scheme A and a winner according to Scheme B will be determined. As well as the scheme specific rankings, an overall assessment is desirable. However, any attempt to combine the scores from Schemes A and requires a trade-off between the two metrics: number of problems solved and time taken. Any particular trade-off is unlikely to be acceptable to everyone. However, if a particular ATP system is the winner in both schemes, then it is superior in terms of both performance metrics. In the ATP competition, iff a single system is the winner of a category according to both Schemes A and B, it will be the overall winner of the category. Note: If only one ATP system is registered for a particular category, no winner will be announced for that category, but the results for that system will still be presented. Example. The example in Table 2 gives some system runtime values ("-" denotes an unsolved problem). The example is designed to illustrate the differences between the schemes, to the extent that each scheme produces a different ranking of the systems. It also shows the scores and rankings that are produced by the different schemes, assuming that time limit = 20. System Runtime for problem Scores by scheme Ranking by scheme 2. 1. 2. Table 2: Example showing the runtimes of three different systems on five different problems and the unscaled scores and rankings produced by the ranking schemes. 8 Conclusion From the numerous issues that need to be resolved for organizing a useful competition, and the impossibility of making an indisputable decision in several cases, it becomes clear that alternative competition designs are possible. However, we believe that this rationally planned competition will provide most of the benefits hoped for in Section 1. We see a clear potential for improved future competitions, by extending their scope; additional coverage of FOF systems (instead of just CNF systems) and model generation (instead of just theorem proving) are the most important issues. However, it seems preferable to start with a core design for a competition, and to add such extensions when more experience has been gained. After more than 30 years of research, automated theorem proving abounds with techniques developed and ideas proposed. The future requires that these techniques and ideas be evaluated to determine which are viable, and to integrate them into systems which are far more flexible than the simple monolithic and compositional systems available today. The most important challenge ahead in theorem proving research will be to provide adequate control, a subject still in its infancy, since it is difficult to approach theoretically. For all these goals, system evaluation is a crucial research tool, and competitions such as this one will provide stimulus and insight that can lay the basis for the development of future ATP systems. Acknowledgments . We are indebted to many researchers who discussed and commented on our ideas; especially Alan Bundy, Bertram Fronh-ofer, Reinhold Letz, Bill McCune, David Plaisted, and Christoph Weidenbach. --R The Internet Programming Contest: A Report and Philosophy. Measuring the Performance of Automated Theorem Provers. Proving First-Order Equality Theorems with Hyper-Linking A Method for Building Models Automati- cally A Theorem Prover for a Computational Logic. Implementing Mathematics with the Nuprl Proof Development System. The 4th DIMACS Interantional Algorithm Implementation Challenge. Semantically Guided First-order Theorem Proving using Hyper-linking The ACM Scholastic Programming Contest - <Year>1977</Year> to <Year>1990</Year> The 2nd Annual Prolog Programming Contest. Search Space and Proof Complexity of Theorem Proving Strategies IMPS: An Interactive Mathematical Proof System. Introduction to HOL Controlled Experimentation and the Design of Agent Architectures. Tactical Theorem Proving in Program Verification. How to Prove Higher Order Theorems in First Order Logic. On the Polynomial Transparency of Resolution. Eliminating Duplication with the Hyper-Linking Strat- egy OTTER 3.0 Reference Manual and Guide. 3 Inductive Learning Com- petitions The 24th ACM International Computer Chess Championship. Based Translation Methods for Modal Logics. PVS: A Prototype Verification System. The CADE-11 Competitions: A Personal View The Next 700 Theorem Provers. The Philosophy of Automated Theorem Proving. The Search Efficiency of Theorem Proving Strategies. Automated Advice in Fitch-style Proof Construction Theorem Proving in Interactive Verification Systems Efficient Model Generation through Compilation. ATP System Results for the TPTP Problem Library (upto TPTP v1. The TPTP Problem Library (TPTP v1. The TPTP Problem Library. Handbook of Statistical Methods for Engineers and Scientists. RVF: An Automated Formal Verification System. Automated Reasoning Contributes to Mathematics and Logic. --TR --CTR Geoff Sutcliffe , Christian Suttner, The Procedures of the CADE-13 ATP System Competition, Journal of Automated Reasoning, v.18 n.2, p.163-169, April 1997 Geoff Sutcliffe , Christian Suttner, The Results - of the CADE-13 ATP System Competition, Journal of Automated Reasoning, v.18 n.2, p.271-286, April 1997 Francis Jeffry Pelletier , Geoff Sutcliffe , Christian Suttner, Conclusions about the CADE-13 ATP System Competition, Journal of Automated Reasoning, v.18 n.2, p.287-296, April 1997 G. Sutcliffe , C. B. Suttner, The CADE-15 ATP System Competition, Journal of Automated Reasoning, v.23 n.1, p.1-23, July 1999 Francis Jeffry Pelletier , Geoff Sutcliffe , Christian Suttner, The development of CASC, AI Communications, v.15 n.2, p.79-90, September 2002 Francis Jeffry Pelletier , Geoff Sutcliffe , Christian Suttner, The development of CASC, AI Communications, v.15 n.2,3, p.79-90, August 2002 G. Sutcliffe, The CADE-17 ATP System Competition, Journal of Automated Reasoning, v.27 n.3, p.227-250, October 2001 Geoff Sutcliffe, The CADE-16 ATP System Competition, Journal of Automated Reasoning, v.24 n.3, p.371-396, April 2000 Christian Suttner , Geoff Sutcliffe, The CADE-14 ATP System Competition, Journal of Automated Reasoning, v.21 n.1, p.99-134, August 1998
automated theorem proving;competition;design
594200
The TPTP Problem Library.
This paper provides a detailed description of the CNF part of the TPTP Problem Library for automated theorem-proving systems. The library is available via the Internet and forms a common basis for development and experimentation with automated theorem provers. This paper explains the motivations and reasoning behind the development of the TPTP (thus implicitly explaining the design decisions made) and describes the TPTP contents and organization. It also provides guidelines for obtaining and using the library, summary statistics about release v1.2.1, and an overview of the tptp2X utility program. References for all the sources of TPTP problems are provided.
Introduction This paper describes the TPTP (Thousands of Problems for Theorem Provers) Problem Library. The TPTP is a library of problems for automated theorem proving (ATP) systems, for 1st order logic. The TPTP is comprehensive, and thus provides an overview, and a simple, unambiguous reference mechanism for problems. All problems in the TPTP are presented in an unambiguous format, and automatic conversion to other known ATP formats is provided. The principal motivation for this project is to move the testing and evaluation of ATP systems from the present ad hoc situation onto a firm footing. This is necessary, as results currently being published seldom provide an accurate reflection of the intrinsic power of the ATP system being considered. A common library of problems is necessary for meaningful system evaluations, meaningful system comparisons, repeatability of testing, and the production of statistically significant results. Guidelines for using TPTP problems and presenting results are given in Section 4.2. 1.1 State of the Art A large number of interesting problems have accumulated over the years in the ATP community. Besides publishing particularly interesting individual problems, from early on researchers have collected problems in order to obtain a basis for experimentation. The first major publication 3 in this regard was [MOW76], 3 The first circulation of problems for testing theorem provers to our knowledge is due to L. Wos in the late sixties. which provides an explicit listing of clauses for 63 problems, many of which are still relevant today. In the same year Wilson and Minker [WM76] tested six resolution strategies on a collection of 86 named problems. The problem clauses are not supplied in [WM76], however. A second major thrust was provided by [Pel86], which lists 75 problems. Other more recent collections are [BLM [Qua92a], [MW92], and [McC93], to name a few. Also, the Journal of Automated Reasoning's Problem Corner regularly provides interesting challenge problems. However, problems published in hardcopy form are often not suitable for testing ATP systems, because they have to be transcribed to electronic form. This is a cumbersome, error-prone process, and is feasible for only very small numbers of problems. A problem library in electronic form was made publicly available by Argonne National Laboratories (Otter format, [McC90]) in 1988 [ANL]. This library has been a major source of problems for ATP researchers. Other electronic collections of problems are available, but have not been announced officially (e.g., that distributed with the SPRFN ATP system [SPR]). Although some of these collections provide significant support to researchers, and formed the early core of the TPTP library, none (with the possible exception of the ANL library) was specifically designed to serve as a common basis for ATP research. Rather, these collections typically were built in the course of research into a particular ATP system. As a result, there are several factors that limit their usefulness as a common basis for research. In particular, existing problem collections are often: - hard to discover and obtain, - limited in scope and size, outdated, - formatted and tuned for a particular ATP system, - inconsistent in their presentation of equally named problems, - provided without a regular update service for new problems, - provided without a reliable error correction service. As a consequence, several problems arise. Firstly, system development and system evaluations typically rely on a limited range of problems, depending on the collections of problems available to the researcher. Secondly, the presentation format used in a collection may not be appropriate for the desired purpose, and a comparatively large effort is required just to make the problems locally usable (which in practice often means that such a collection of problems is simply ignored). Thirdly, using a particular collection may lead to biases in results, because the problems have been designed and tuned specifically for a particular ATP system. Fourthly, the significance and difficulty of a problem, with respect to the current state-of-the-art in ATP systems, is hard to assess by newcomers to the field. Existing test problems are often not adequate anymore (e.g., Schubert's Steamroller [Sti86]), while others may be solvable only with specialized techniques (e.g., LIM+ [Ble90]) and therefore are much too hard to start with. Finally, many copies and variants of the same "original" problem may exist in different libraries. Coupled with a lack of documentation, this means that unambiguous identification of problems, and therefore a clear interpretation of performance figures for given problems, has become difficult. The problem of meaningfully interpreting results is even worse than indi- cated. Commonly a few problems are selected and hand-tuned (clauses and literals are arranged in a special order, irrelevant clauses are omitted, lemmas are added in, etc) specifically for the ATP system being tested. However, the presentation of a problem can significantly affect the nature of the problem, and changing the clauses clearly makes a different problem altogether. Nevertheless the problem may be referenced under the same name as it was presented else- where. As a consequence the experimental results reveal little. Some researchers avoid this ambiguity by listing the clause sets explicitly. But obviously this usually cannot be done for a large number of problems or for large individual prob- lems. The only satisfactory solution to these issues is a common and stable library of problems, which the TPTP provides. 1.2 What is Required? The goal for building the TPTP has been to overcome the current drawbacks, and to centralize the burden of problem collection and maintenance to one place. The TPTP tries to address all relevant issues. In particular, the TPTP - is available by anonymous ftp. The TPTP is thus easily available to the research community. Awareness of the TPTP is assured by extensive formal and informal announcements. - spans a diversity of subject matters. This reduces biases in the development and testing of theorem provers, which arise from the use of a limited range of problems. It also provides an overview of the domains that ATP is used in. The significance of each domain may be measured by the number of problems available. - is large enough for statistically significant testing. In contrast to common practise, an ATP system should be evaluated over a large number of problems, rather than a small set of judiciously selected examples. The large size of the TPTP makes this possible. - is comprehensive and up-to-date. The TPTP contains all problems known to the community.There is no longer a need to look elsewhere. - is presented in a well structured and documented format. All problems in the TPTP are presented in an unambiguous format. Automatic conversion to other known formats is provided, thus eliminating the necessity for transcription. The design and arrangement of the problems is independent of any particular ATP system. - provides a common, independent, source for problems. This provides unambiguous problem reference, and makes the comparison of results meaningful. contains problems varying in difficulty, from very simple problems through to open problems. This allows all interested researchers, from newcomers to experts, to rely on the same problem library. provides statistics for each problem and for the library as a whole. This gives information about the syntactic nature of the problems. - will provide a rating for the difficulty of each problem. This is important for several reasons: (1) It simplifies problem selection according to the user's intention. (2) It allows the quality of an ATP system to be judged. (3) Over the years, changes in the problem ratings will provide an indicator of the advancement in ATP. The problem ratings are currently being worked on, and will be part of a future TPTP release. documents each problem. This contributes to the unambiguous identification of each problem. - provides standard axiomatizations that can be used in new problems. This simplifies the construction of new problems. - is a channel for making new problems available to the community, in a simple and effective way. - removes the necessity for the duplicated effort of maintaining many libraries. specifies guidelines for its use for evaluating ATP systems. As the TPTP is a standard library of problems, it is possible to set ATP evaluation guidelines which make reported results meaningful. This will in turn simplify and improve system comparisons, and allow ATP researchers to accurately gauge their progress. The TPTP problem library is an ongoing project, with the aim to provide all of the desired properties. Current Limitations of the TPTP. The current version of the TPTP library is limited to problems expressed in 1st order logic, presented in clause normal form. In particular, there are no problems for induction and nonclassical theorem proving. However, see Section 5 for upcoming and planned extensions. 2 Inside the TPTP Scope. Release v1.0.0 of the TPTP contains 1577 abstract problems, which result in 2295 ATP problems, due to alternative presentations (see Section 2.2). Tables 1, 2, and 3 provide some statistics about release v1.0.0 of the TPTP. The problems have been collected from various sources. The two principal sources have been existing problem collections and the ATP literature. Other sources include logic programming, mathematics, puzzles, and correspondence with ATP researchers. Many people and organizations have contributed towards the TPTP. In particular, the foundations of the TPTP were laid with David Plaisted's SPRFN collection [SPR]; many problems have been taken from Argonne National Laboratory's ATP problem library [ANL] (special thanks to Bill McCune here); Art Quaife has provided several hundred problems in set theory and algebra [Qua92b]; the Journal of Automated Reasoning, CADE Proceedings, Table 1. Statistics on the TPTP. Number of problem domains 23 Number of abstract problems 1577 Number of problems 2295 Number of non-Horn problems 1449 (63%) Number of problems with equality 1805 (79%) Number of propositional problems 111 (5%) being non-Horn 56 (50%) Total number of clauses 322156 Total number of literals 970456 Table 2. Statistics for non-propositional TPTP problems. Measure Minimum Maximum Average Median Number of clauses 2 6404 124 68 Percentage of non-Horn clauses 0% 99% 6% 4% problems 0% 99% 9% 4% Percentage of unit clauses 0% 100% 32% 22% Number of literals 2 18966 276 162 Percentage of equality literals 0% 100% 47% 47% problems 4% 100% 57% 47% Number of predicate symbols 1 4042 12 4 Number of function symbols 0 94 Percentage of constants 0% 100% 46% 33% Number of variables 0 32000 311 271 Percentage of singletons 0% 100% 6% 7% Maximal clause size 1 25 4 5 Maximal term depth Table 3. Statistics for propositional TPTP problems. Measure Minimum Maximum Average Median Number of clauses 2 8192 444 36 Percentage of non-Horn clauses 0% 99% 20% 1% problems 1% 99% 40% 45% Percentage of unit clauses 0% 100% 16% 1% Number of literals 2 106496 3303 132 Number of predicate symbols 1 840 69 13 Maximal clause size 1 21 5 3 and Association for Automated Reasoning Newsletters have provided a wealth of material; smaller numbers of problems have been provided by a number of further contributors (see also the Acknowledgements at the end of Section 5). The problems in the TPTP are syntactically diverse, as is indicated by the ranges of the values in Tables 2 and 3. The problems in the TPTP are also semantically diverse, as is indicated by the range of domains that are covered. The problems are grouped into 23 domains, covering topics in the fields of logic, mathematics, computer science, and more. The domains are presented and discussed in Section 2.1. Releases. The TPTP is managed in the manner of a software product, in the sense that fixed releases are made. Each release of the TPTP is identified by a release number, in the form v!Version?.!Edition?.!Patchlevel?. The Version number enumerates major new releases of the TPTP, in which important new features have been added. The Edition number is incremented each time new problems are added to the current version. The Patch level is incremented each time errors, found in the current edition, are corrected. All changes are recorded in a history file, as well as in the file for an affected problem. 2.1 The TPTP Domain Structure An attempt has been made to classify the totality of the TPTP problems in a systematic and natural way. The resulting domain scheme reflects the natural hierarchy of scientific domains, as presented in standard subject classification literature. The current classification is based mainly on the Dewey Decimal Classification [Dew89] and the Mathematics Subject Classification used for the Mathematical Reviews by the American Mathematical Society [MSC92]. We define four main fields: logic, mathematics, computer science, and other. Each field contains further subdivisions, called domains. These TPTP domains constitute the basic units of our classification. The full classification scheme is shown in Figure 1. 2.2 Problem Versions and Standard Axiomatizations. There are often many ways to formulate a problem for presentation to an ATP system. Thus, in the TPTP, there are often alternative presentations of a prob- lem. The alternative presentations are called versions of the underlying abstract problem. As the problem versions are the objects that ATP systems must deal with, they are referred to simply as problems, and the abstract problems are referred to explicitly. Each problem is stored in a separate physical file. The primary reason for different versions of an abstract problem is the use of different axiomatizations. This issue is discussed below. A secondary reason is the use of different formulations of the theorem to be proven. TPTP Logic Mathematics Computer science Other Combinatory logic Logic calculi Henkin models Theory Graph Theory Algebra Boolean Algebra Robbins Algebra distributive Lattices Groups Rings General Algebra Number Theory Topology Analysis Geometry Category theory Computing Theory Planning Program Verification Syntactic Puzzles Miscellaneous COL HEN GRA BOO ROB LDA GRP ALG NUM ANA GEO COM PLA PUZ MSC # of Problems abs. / vers. 284 / 321 Fig. 1. The domain structure of the TPTP. Different axiomatizations. Commonly, different axiomatizations of a theory exist, and, in general, most are equally acceptable. In the TPTP an axiomatization is acceptable if it is complete (in the sense that it captures some closed theory) and it has not had any lemmas added. Such axiomatizations are called standard axiomatizations. Within the ATP community some problems have been created with non-standard axiomatizations. An axiomatization is non-standard if it has been reduced (i.e., axioms have been removed) and the result is an incomplete axiomatization, or if it has been augmented (i.e., lemmas have been added) and the result is a redundant axiomatization. Non-standard axiomatizations are typically used to find a proof of a theorem (based on the axiomatization) using a particular ATP system. In any 'real' application of an ATP system, a standard axiomatization of the application domain would typically have to be used, at least initially. Thus the use of standard axiomatizations is desirable, because it reflects such 'real' usage. In the TPTP, for each collected problem that uses a non-standard axiomatization, a new version of the problem is created with a standard axiomatization. The standard axiomatization is created by reducing or augmenting the non-standard axiomatization appropriately. The standard axiomatizations used in the TPTP are kept in separate axiom files, and are included in problems as appropriate. If an axiomatization uses equality, the required axioms of substitution are kept separate from the theory specific axioms. The equality axioms of reflexivity, symmetry, and transitivity, which are also required when equality is present, are also kept separately, as they are often used independently of the substitution axioms. 2.3 Problem and Axiomatization Naming Providing unambiguous names for all problems is necessary in a problem library. A naming scheme has been developed for the TPTP, to provide unique, stable names for abstract problems, problem versions, and axiomatizations. File names are in the form DDDNNN-V[MMM].p for problem files, and DDDNNN-E.TT for axiom files. DDD is the domain name abbreviation, NNN is the index within the domain, V is the version number, MMM is an optional generation parameter, E is the axiomatization extension number, .p indicates a problem file, and TT is either .ax for theory specific axioms or .eq for equality substitution axioms. The complete file names are unique within the TPTP. Abstract problems and axiomatizations have also been allocated semantic names. The semantic names can be used to augment file names, so as to give an indication of the contents. While these names are provided for users who like to work with mnemonic names, only the standard syntactic names are guaranteed to provide unambiguous reference. The semantic names are formed from a set of specified abbreviations. The semantic names can be added to the syntactic file names using a script that is provided. 2.4 Problem Presentation The physical presentation of the TPTP problem library is such that ATP researchers can easily use the problems. The TPTP file format, for both problem files and axiom files, has three main sections. The first section is a header section that contains information for the user. The second section contains include instructions for axiom files. The last section contains the clauses that are specific to the problem. The syntax of the problem and axiom files is that of Prolog. This conformance makes it trivial to manipulate the files using Prolog. All information in the files that is not for use by ATP systems is formatted as Prolog comments, with a leading %. All the information for ATP systems is formatted as Prolog facts. A utility is provided for converting TPTP files to other known ATP system formats (see Section 3). A description of the information contained in TPTP files is given below. Full details are given in [SSY93]. The Header Section The header contains information about the problem, for the user. It is divided into four parts. The first part identifies and describes the problem. The second part provides information about occurrences of the problem. The third part gives the problem's ATP status and a table of syntactic measurements made on the problem. The last part contains general information about the problem. An example of a TPTP header, extracted from the problem file GRP039-7.p, is shown in Figure 2. Released v0.0.0, Updated v0.11.5. Theory (Subgroups) Subgroups of index 2 are normal O is a subgroup of G and there are exactly 2 cosets in % G/O, then O is normal [that is, for all x in G and y in O, % x*y*inverse(x) is back in O]. Problems and Experiments for and with Automated Theorem Proving Programs IEEE Transactions on Computers C-25(8), 773-782. % Number of literals : % Number of predicate % Number of function % Number of variables : 42 ( 0 singletons) Maximal clause size : 4 Maximal Comments : element-in-O2(A,B) is A-1.B. The axioms with force index 2. fixed. Fig. 2. Example of a problem file header (GRP039-7.p). The header fields contain the following information: - The % File field contains the problem's syntactic and semantic names, the TPTP release in which the problem first appeared, and the TPTP release in which the problem was last updated. - The % Domain field identifies the domain from which the problem is drawn. - The % Problem field provides a one line, high-level description of the abstract problem. - The % Version field gives information that differentiates this version of the problem from other versions of the problem. - The % English field provides a full description of the problem if the one line description is too terse. - The % Refs field provides a list of references to items in which the problem has been presented. - The % Source field acknowledges, with a citation, where the problem was (physically) obtained from. - The % Names field lists existing known names for the problem (as it has been named in other problem collections or the literature). - The % Status field gives the problem's ATP status 4 , one of satisfiable, unsatisfiable, open, or broken. - The % Syntax field lists various syntactic measures of the problem's clauses. - The % Comments field contains free format comments about the problem. - The % Bugfixes field documents any changes which have been made to the clauses of the problem. The Include Section The include section contains include instructions for TPTP axiom files. An example of an include section, extracted from the problem file GRP039-7.p, is shown in Figure 3. %-Include the axioms of equality include('Axioms/EQU001-0.ax'). %-Include the equality axioms for group theory in equality form include('Axioms/GRP004-0.ax'). %-Include the subgroup axioms in equality formulation include('Axioms/GRP004-1.ax'). Fig. 3. Example of a problem file include section (GRP039-7.p). Each of the include instructions indicates that the clauses in the named axiom file should be included at that point. Axiom files are presented in the same format as problem files, and include instructions may also appear in axiom files. Full versions of TPTP problems (without include instructions) can be created by using the tptp2X utility (see Section 3). 4 This field has not yet been filled in all problems of the TPTP. The Clauses Section TPTP problems are presented in clausal normal form. The literals that make up a clause are presented as a Prolog list of terms. Each literal is a term whose functor is either ++ or -, indicating a positive or negative literal respectively. The single argument of the sign, i.e., the atom of the literal, is a Prolog term. The signs ++ and - are assumed to be defined as prefix operators in Prolog. Each clause has a name, in the form of a Prolog atom. Each clause also has a status, one of axiom, hypothesis, or theorem. The hypothesis and theorem clauses are those that are derived from the negation of the theorem to be proved. The status hypothesis is used only if the clauses can clearly be determined as status is theorem. The name, status, and literal list of each clause are bundled as the three arguments of a Prolog fact, whose predicate symbol is input clause. These facts are in the clauses section of the problem file. Two examples of clauses, extracted from the problem file GRP039-1.p, are shown in Figure 4. %-Definition of subgroup of index 2 input-clause(an-element-in-O2,axiom, ++subgroup-member(A)]). input-clause(prove-d-is-in-subgroup,theorem, [-subgroup-member(d)]). Fig. 4. Examples of problem file clauses (GRP039-1.p). 2.5 Physical Organization The TPTP is physically organized into five subdirectories: - The Problems directory contains a subdirectory for each domain, as shown in Figure 1. Each subdirectory contains the problem files for that domain. - The Axioms directory contains the axiom files. - The TPTP2X directory contains the tptp2X utility, described in Section 3. - The Scripts directory contains some useful C shell scripts. - The Documents directory contains comprehensive online information about the TPTP, in specific files. This provides quick access to relevant overview data. In particular, it provides a simple means for selecting problems with specific properties, by using standard system tools (e.g., grep, awk). 3 The tptp2X Utility The tptp2X utility converts TPTP problems from the TPTP format to formats used by existing ATP systems. Currently, the tptp2X utility supports the following output formats: - the MGTP format [FHKF92]; - the Otter .in format [McC90] (the set of support and inference rules to be used are also specified, and are included in the output file); - the PTTP format [Sti84]; - the SETHEO .lop format [STvdK90]; - the SPRFN format [Pla88]; - the TPTP format, substituting include instructions with the actual clauses. It is simple to add new formatting capabilities to the tptp2X utility. The tptp2X utility can also be used to rearrange the clauses and literals in a problem. This facilitates testing the sensitivity of an ATP system to the order in which the clauses and literals are presented. Full details of the tptp2X utiltity are given in [SSY93]. 4 You and the TPTP 4.1 How to FTP the TPTP The TPTP can be obtained by anonymous ftp from the Department of Computer Science, James Cook University, or the Institut f?r Informatik, Technische Uni- versit-at M-unchen. The ftp hosts are coral.cs.jcu.edu.au (137.219.17.4) and flop.informatik.tu-muenchen.de (131.159.8.35), respectively. At both sites the TPTP is in the directory pub/tptp-library. There are three files. Information about the TPTP is in the ReadMe file (9.3 KByte), a technical report [SSY93] is in TR-v1.0.0.ps.Z (357 KByte), and the library itself is packaged in TPTP-v1.0.0.tar.Z (2.4 MByte). Please read the ReadMe file, as it contains up-to-date information about the TPTP. 4.2 Important: Using the TPTP By providing this library of ATP problems, and a specification of how these problems should be presented to ATP systems, it is our intention to place the testing, evaluation, and comparison of ATP systems on a firm footing. For this reason, you should abide by the following conditions when using TPTP problems and presenting your results: - The specific version, edition, and patch level of the TPTP, used as the problem source, must be stated (see the introduction of Section 2). Each problem must be referenced by its unambiguous syntactic name. may be added/removed without explicit notice. (This holds also for removing equality axioms when built-in equality is provided by the prover.) - The clauses/literals may not be rearranged without explicit notice. If clause or literal reversing is done by the tptp2X utility, the reversals must be explicitly noted. - The header information in each problem may not be used by the ATP system without explicit notice. Any information that is given to the ATP system, other than that in the input clauses, must be explicitly noted (including any system switches or default settings). Abiding by these rules will allow unambigous identification of the problem, the arrangement of clauses, and further input to the ATP system which has been used. 5 Present and Future Activity 5.1 The Present Users' experiences with the TPTP will tell if we have achieved the design goals described in Section 1, except for the problem ratings. The problem ratings are currently being worked on. Note that these ratings should decrease as advances in automated theorem proving are made. Therefore the long term maintenance of the individual problem ratings will provide an objective measure of progress in the field. An upcoming addition to the TPTP is a set of Prolog programs for generating arbitrary sizes of generic problems (e.g., the "pigeon holes" problem). This will implicitly provide any size of such problems, and also reduce the physical size of the TPTP. These programs are being worked on. In order to ensure the usability of the programs, they will be tested on a public domain Prolog. The tptp2X utility will soon be modified to be capabile of applying arbitrary sequences of transformations to problems. The transformations will include those currently in tptp2X, and others that become available (e.g., Mark Stickel has provided a transformation for his upside-down meta-interpretation approach [Sti94]). Automatic removal of equality axioms will also be provided, to ease the life of researchers whose ATP systems have built in mechanisms for equality reasoning. 5.2 The Future We have several short and long term plans for further development of the TPTP. The main ideas are listed here. Performance tables for popular ATP systems. Extensive experimental results, for publically available ATP systems, are currently being collected (based on all the TPTP problems). These will be published in some form. - ATP system evaluation guidelines. General guidelines outlining the requirements for ATP system evaluation will be produced. - The BSTP Benchmark Suite. A benchmark suite (the BSTP) will be selected from the TPTP. The BSTP will be a small collection of problems, and will provide a minimal set of problems on which an ATP system evaluation can be based. The BSTP will be accompanied by specific guidelines for computing a performance index for an ATP system. Full 1st order logic. The TPTP will be extended to include problems in full 1st order logic no- tation. ATP systems with automatic conversion to clausal form then can derive additional information regarding the problem, such as which functors are Skolem functors. - Various translators. Translators between various logical forms will be provided. ffl from full 1st order logic to clausal normal form ffl from non-Horn to Horn form ffl from 1st order to propositional form - Non-classical and higher order logics. In the longer term, the TPTP may be extended to include problems expressed in non-classical and higher order logics. 5.3 Acknowledgements We are indebted to the following people and organizations who have contributed problems and helped with the construction of the TPTP: Geoff Alexander, the ANL group (especially Bill McCune), the Automated Reasoning group in Munich, Dan Benanav, Woody Bledsoe, Maria Poala Heng Chu, Tom Jech, Reinhold Letz, Thomas Ludwig, Xumin Nie, Jeff Pelletier, David Plaisted, Joachim Posegga, Art Quaife, Alberto Segre, John Slaney, Mark Stickel, Bob Veroff, and TC Wang. --R Available by anonymous ftp from info. Challenge Problems in Elementary Calculus. Set Theory in First-Order Logic: Clauses for G-odel's Axioms Dewey Decimal Classification and Relative Index. Model Generation Theorem Provers on a Parallel Inference Machine. Otter 2.0 Users Guide. Single Axioms for Groups and Abelian Groups with Various Operations. Problems and Experiments for and with Automated Theorem-Proving Programs American Mathematical Society Experiments in Automated Deduction with Condensed Detachment. Automated Deduction in von Neumann-Bernays-Godel Set The- ory Problems based on NBG Set Theory. The TPTP Problem Library (TPTP v1. A Prolog Technology Theorem Prover. Schubert's Steamroller Problem: Formulations and Solutions. SETHEO/PARTHEO Users Manual. --TR --CTR Robert Nieuwenhuis, The impact of CASC in the development of automated deduction systems, AI Communications, v.15 n.2, p.77-78, September 2002 Robert Nieuwenhuis, The impact of CASC in the development of automated deduction systems: Guest-editorial, AI Communications, v.15 n.2,3, p.77-78, August 2002 Monty Newborn , Zongyan Wang, Octopus: Combining Learning and Parallel Search, Journal of Automated Reasoning, v.33 n.2, p.171-218, September 2004 John Slaney, More Proofs of an Axiom of ukasiewicz, Journal of Automated Reasoning, v.29 n.1, p.59-66, 2002 G. Sutcliffe , C. B. Suttner, The CADE-18 ATP System Competition, Journal of Automated Reasoning, v.31 n.1, p.23-32, Bernd Lchner, Things to Know when Implementing KBO, Journal of Automated Reasoning, v.36 n.4, p.289-310, April 2006 Steven Trac , Yury Puzis , Geoff Sutcliffe, An Interactive Derivation Viewer, Electronic Notes in Theoretical Computer Science (ENTCS), v.174 n.2, p.109-123, May, 2007 Urban , Grzegorz Bancerek, Presenting and Explaining Mizar, Electronic Notes in Theoretical Computer Science (ENTCS), v.174 n.2, p.63-74, May, 2007 Geoff Sutcliffe , Christian Suttner, The CADE-19 ATP System Competition, AI Communications, v.17 n.3, p.103-110, August 2004 Geoff Sutcliffe , Christian Suttner, The CADE-19 ATP system competition, AI Communications, v.17 n.3, p.103-110, July 2004 Justin Ward , Garrin Kimmell , Perry Alexander, Prufrock: a framework for constructing polytypic theorem provers, Proceedings of the 20th IEEE/ACM international Conference on Automated software engineering, November 07-11, 2005, Long Beach, CA, USA Geoff Sutcliffe, The IJCAR-2004 Automated Theorem Proving Competition, AI Communications, v.18 n.1, p.33-40, January 2005 Geoff Sutcliffe, The IJCAR-2004 automated theorem proving competition, AI Communications, v.18 n.1, p.33-40, January 2005 Francis Jeffry Pelletier , Geoff Sutcliffe , Christian Suttner, The development of CASC, AI Communications, v.15 n.2, p.79-90, September 2002 Francis Jeffry Pelletier , Geoff Sutcliffe , Christian Suttner, The development of CASC, AI Communications, v.15 n.2,3, p.79-90, August 2002 G. Sutcliffe , C. B. Suttner , F. J. Pelletier, The IJCAR ATP System Competition, Journal of Automated Reasoning, v.28 n.3, p.307-320, April 2002 G. Sutcliffe , C. B. Suttner, The CADE-15 ATP System Competition, Journal of Automated Reasoning, v.23 n.1, p.1-23, July 1999 Geoff Sutcliffe, The CADE-16 ATP System Competition, Journal of Automated Reasoning, v.24 n.3, p.371-396, April 2000 Geoff Sutcliffe, The CADE-20 Automated Theorem Proving Competition, AI Communications, v.19 n.2, p.173-181, April 2006 Geoff Sutcliffe, The CADE-20 automated theorem proving competition, AI Communications, v.19 n.2, p.173-181, January 2006 Dominique Pastre, Strong and weak points of the MUSCADET theorem prover - examples from CASC-JC, AI Communications, v.15 n.2,3, p.147-160, August 2002 G. Sutcliffe, The CADE-17 ATP System Competition, Journal of Automated Reasoning, v.27 n.3, p.227-250, October 2001 J. Avenhaus , Th. Hillenbrand , B. Lchner, On using ground joinable equations in equational theorem proving, Journal of Symbolic Computation, v.36 n.1-2, p.217-233, July Josef Urban, MPTP -- Motivation, Implementation, First Experiments, Journal of Automated Reasoning, v.33 n.3-4, p.319-339, October 2004 Jia Meng , Claire Quigley , Lawrence C. Paulson, Automation for interactive proof: first prototype, Information and Computation, v.204 n.10, p.1575-1596, October 2006 Andreas Meier , Volker Sorge, Applying SAT Solving in Classification of Finite Algebras, Journal of Automated Reasoning, v.35 n.1-3, p.201-235, October 2005 Evaluating general purpose automated theorem proving systems, Artificial Intelligence, v.131 n.1-2, p.39-54, September 2001 Stephan Schulz, E - a brainiac theorem prover, AI Communications, v.15 n.2,3, p.111-126, August 2002 Stephan Schulz, E - a brainiac theorem prover, AI Communications, v.15 n.2, p.111-126, September 2002 Geoff Sutcliffe , Christian Suttner, The state of CASC, AI Communications, v.19 n.1, p.35-48, January 2006 Geoff Sutcliffe , Christian Suttner, The state of CASC, AI Communications, v.19 n.1, p.35-48, January 2006 Jens Otten , Wolfgang Bibel, leanCoP: lean connection-based theorem proving, Journal of Symbolic Computation, v.36 n.1-2, p.139-161, July Simon Colton , Stephen Muggleton, Mathematical applications of inductive logic programming, Machine Learning, v.64 n.1-3, p.25-64, September 2006 Jose Hernandez-Orallo, Beyond the Turing Test, Journal of Logic, Language and Information, v.9 n.4, p.447-466, October 2000 Christian Suttner , Geoff Sutcliffe, The CADE-14 ATP System Competition, Journal of Automated Reasoning, v.21 n.1, p.99-134, August 1998 Christoph Weidenbach, Combining superposition, sorts and splitting, Handbook of automated reasoning, Elsevier Science Publishers B. V., Amsterdam, The Netherlands, 2001 R. Sekar , I. V. Ramakrishnan , Andrei Voronkov, Term indexing, Handbook of automated reasoning, Elsevier Science Publishers B. V., Amsterdam, The Netherlands, 2001
problem library;TPTP;experimental evaluation
594215
What You Always Wanted to Know about Rigid E-Unification.
This paper solves an open problem posed by a number of researchers: the construction of a complete calculus for matrix-based methods with rigid E-unification. The use of rigid E-unification and simultaneous rigid E-unification for such methods was proposed by Gallier et al., in 1987. After our proof of the undecidability of simultaneous rigid E-unification in 1995. (Degtyarev and Voronkov, 1996d), it became clear that one should look for more refined techniques to deal with equality in matrix-based methods. In this article, we define a complete proof procedure for first-order logic with equality based on an incomplete but terminating procedure for rigid E-unification. Our approach is applicable to the connection method and the tableau method and is illustrated on the tableau method.
Introduction Algorithms for theorem-proving based on matings or tableaux in first-order logic without equality comprise two kinds of rules. Rules of the first kind construct matrices or tableaux from a given formula using a suitable amplification. Rules of the second kind try to close paths or branches using substitutions making the paths or branches inconsistent. These substitutions are unifiers of some atoms laying on a path or branch. Until recently, most approaches to introducing equality in such matrix-based methods tried to generalize such algorithms by a suitable modification of the notion of a unifier. Such a modification using simultaneous rigid E-unification was introduced in (Gallier et al., 1987) for the method of matings (Andrews, 1981) or the connection method (Bibel, 1981). It can be easily represented in the tableau formalism. The method of matings interleaves two steps: amplification by quantifier duplication and search for mating for a given amplification. For formulas in disjunctive normal form this method was formulated earlier in (Prawitz, 1983). In this case amplification is represented by a matrix and mating is represented by a set of simultaneously satisfiable substitution conditions (mated pairs). Prawitz proposed a procedure for constructing substitution conditions one by one, closing the corresponding paths in the matrix through search with backtracking. Procedures of this kind were used in later formalizations and implementations of the tableau method, the method of matings or the connection method for formulas without equality. For example, in (Fitting, 1990) the search for mating consists of repeated applications of MGU atomic closure rule. Supported by grants from INTAS, TFR and the Swedish Royal Academy of Sciences y Partially supported by a TFR grant (Gallier et al., 1987) tried to describe a similar procedure for logic with equality. For example, (Gallier et al., 1992) describe such a procedure in which a substitution condition is formalized via rigid E-unification, and the set of substitution conditions via simultaneous rigid E-unification. Simultaneous rigid E-unification can be formulated as follows. Given equations finite sets of equations find a substitution oe such that means provability in first-order logic with equality). The corresponding instance of the simultaneous rigid E-unification problem is denoted by the system of rigid equations EXAMPLE 1.1. In this and further examples we shall often omit parentheses in terms with unary function symbols, for example we write ffb instead of f(f(b)). Assume that we want to prove the formula several applications of tableau expansion rules to the negation normal form of :' (ff-, fi- and fl-rules in the terminology of (Smullyan, 1968; Fitting, 1990) we obtain the following tableau (we only consider the part of the tableau containing literals, omitting non-literal a ' b c ' d Collecting formulas lying on the two branches in this tableau we obtain the following two rigid equations expressing inconsistency of this tableau: This system of rigid equations has one solution ffa=x; fb=y; fc=u; fd=vg. This substitution can be found by applying the functional reflexivity rule and MGU replacement rule of (Fitting, 1990) (in fact, the reformulation of the paramodulation rule for tableaux). Obviously, this tableau cannot be closed without the use of the functional reflexivity. 1 1 The system of clauses corresponding to this example improves a result proved in (Plaisted, 1995) by using a more complicated example. In this system of clauses a a there is no refutation even by unrestricted rigid paramodulation (i.e. using non-ordered rigid paramodulation and paramodulation into variables), while (Plaisted, 1995) gives an example showing incompleteness of ordered rigid paramodulation only. What You Always Wanted to Know About Rigid E-Unification 3 Since the invention of simultaneous rigid E-unification by (Gallier et al., 1987), there have been a number of publications on simultaneous rigid E-unification itself and its use in theorem proving, for example (Gallier et al., 1988; Gallier et al., 1989; Gallier et al., 1990; Gallier et al., 1992; Baumgartner, 1992; Beckert and H-ahnle, 1992; Becher and Petermann, 1994; Beckert, 1994; Goubault, 1994; Petermann, 1994). Some of these articles were based on the conjecture that simultaneous rigid E-unification is decidable. There were several faulty proofs of the decidability of this problem (Gallier et al., 1988; Gallier et al., 1990; Goubault, 1994). The refutation procedure for first-order logic with equality using simultaneous rigid E-unification (Gallier et al., 1992) was based on a faulty assumption that solutions to simultaneous rigid E-unification can be found by consecutive combination of finite complete sets of solutions for (non-simultaneous) rigid E-unification (Gallier et al., 1990; Gallier et al., 1988). Later, (Degtyarev and Voronkov, 1996d; Degtyarev and Voronkov, 1996e) proved that simultaneous rigid E-unification is undecidable, which implied that Gallier et.al.'s procedure cannot, in general, find solutions to simultaneous rigid E-unification. However, it is not clear whether this implies incompleteness of their procedure for first-order logic with equality: there are examples when the procedure cannot find a solution for a given amplification in spite that such a solution exists, but can find a solution for a bigger amplification. Completeness of Gallier et.al.'s procedure or existence of a procedure complete for first-order logic with equality based on some set of solutions to rigid E-unification was an open problem (Petermann, 1994; Beckert, 1995). Our paper gives a positive solution to this problem. An advantage of Gallier et.al.'s procedure is that it allows one to extend the proof-search technology developed for tableaux without equality to the case with equality, using solutions to rigid E-unification instead of most general unifiers. In particular, for a given amplification Gallier et.al.'s procedure always terminates. A procedure of this kind is used in the theorem prover 3 T A P (H-ahnle et al., 1994, R.H-ahnle, private communication). In this paper we define a procedure extending the tableau method to logic with equality based on an incomplete procedure for rigid E-unification. Nevertheless, our procedure is complete for first-order logic with equality. Hence, we rehabilitate Gallier et.al.'s program for adding equality to semantic tableaux. Moreover, our procedure solves rigid equations laying on different tableau branches independent- ly. This strongly improves Gallier et.al.'s procedure which uses solutions of some rigid equations to solve rigid equations on other branches. A similar approach has already been defined in (Kanger, 1983) based on a more straightforward way of variable instantiation. As a method for finding a closing substitution, Kanger proposed an algorithm which can now be characterized as an incomplete (but terminating) algorithm for simultaneous rigid E-unifiability. Variables in a matrix (or a tableau) could be consecutively substituted by ground terms already occurring in the matrix. This procedure does not solve simultaneous rigid E-unifiability, but it gives a procedure complete for first-order logic with total.tex - Date: June 3, 1997 4 Anatoli Degtyarev and Andrei Voronkov equality. In the terminology of (Fitting, 1990) it means that a closing substitution can be found after a sufficiently high (but not necessarily minimal) number of applications of the fl-rule. The approach to substitutions based on this idea has been characterized as minus-normalization in (Matulis, 1963; Maslov, 1967). However, for a language with function symbols minus-normalization is interesting mostly theoretically. Even in simplest cases, minus-normalization requires a huge number of instantiations. For example, in the tableau of Example 1.1, we have to consider 8 4 possible instantiations of variables x; by terms in the set fa; b; c; d; fa; f b; f c; fdg. Moreover, it has been proved that the use of minus- normalization can lead to considerable growth of derivations. Some results on minus-normalization are proved by (Norgela, 1974). In this paper we describe a logical calculus BSE for rigid E-unification based on the rigid basic superposition rule that is an adaptation of basic superposition of (Bachmair et al., 1992; Nieuwenhuis and Rubio, 1992), for "rigid" variables. For a given rigid E-unification problem (called rigid equation in this paper), there is only a finite number of BSE-derivations for this problem. Thus, BSE gives us an algorithm returning a finite set of solutions to this rigid equation. We use these solutions to close a tableau branch in the same way as most general unifiers are used to close a branch in the MGU atomic closure rule of (Fitting, 1990). Preliminaries We present here a brief overview of notions and preliminary definitions necessary for understanding the paper. We assume basic knowledge of substitutions and unification. Let \Sigma be a signature, and X be a set of variables. T (\Sigma; X) denotes the set of all terms in the signature \Sigma with variables from X. The set of all ground terms in the signature \Sigma is denoted by T (\Sigma). A literal is either an atomic formula or a negation of an atomic formula. An equation is a literal s ' t, where s; t 2 T (\Sigma; X). We do not distinguish equations Literals of the form :(s ' t) are denoted by s 6' t and called disequations. For simplicity, we assume that ' is the only predicate symbol of our first-order language. As usual, arbitrary first-order languages can be represented in such language by introducing a sort bool and replacing any non-equational atom A by A ' true (for details see e.g. Bachmair et al., 1995). By a ground expression (i.e. term or literal) we mean an expression containing no variables. For any expression E, var(E) denotes the set of all variables occurring in E. For a sequence of variables - x, we shall sometimes denote - x also the corresponding set of variables. We write A[s] to indicate that an expression A contains s as a subexpression and denote by A[t] the result of replacing this occurrence of s in A by t. By Aoe we denote the result of applying the substitution oe to A. If A is a formula, we can as usual rename bound variables in A before applying oe. We total.tex - Date: June 3, 1997 What You Always Wanted to Know About Rigid E-Unification 5 shall denote '(x) a formula ' with zero or more free occurrences of a variable x and write '(t) to denote the formula 'ft=xg. A substitution ' whose domain is a subset of fx is denoted A substitution oe is called grounding for a set of variables V if for every variable the term voe is ground. be two substitutions with disjoint domains, The union of ' 1 and is the substitution oe defined as follows. For every variable v we have Note that we use the union notation only for substitutions with disjoint domains. The inference systems used in this paper are defined with respect to a reduction ordering, denoted by - which is total on ground terms. Our results are valid for any such ordering. A formula is in the Skolem negation normal form if it is constructed from literals using the connectives - and the quantifier 8. There is a satisfiability- preserving structure-preserving translation of formulas without equivalences into formulas in Skolem negation normal form consisting of the standard skolemization and a translation into negation normal form used e.g. in (Andrews, 1981). In order to prove an arbitrary formula ', we translate :' in Skolem negation normal form obtaining a formula / and try to establish unsatisfiability of /. For this reason, theorems in this paper are formulated in terms of unsatisfiability. For an equation s ' t and a multiset of equations E we write E ` s ' t to denote that the formula ( V is provable in first-order logic with equality. For such formulas provability can be tested by the congruence closure algorithm (Shostak, 1978). For the inclusion of multisets we shall use notation 3 Rigid basic superposition The term "rigid paramodulation" has already been used in (Becher and Peter- mann, 1994; Plaisted, 1995) for systems of inference rules in which all variables are treated as "rigid". For example, rigid clause paramodulation of (Plaisted, 1995) is essentially paramodulation and resolution over a set of clauses, where all substitutions are applied to the whole set of clauses. A similar system for resolution has been proposed earlier in (Chang, 1972) as V-resolution and in (Lee and Chang, for resolution with paramodulation as V-resolution and V-paramodulation. 6 Anatoli Degtyarev and Andrei Voronkov We shall use the term "rigid basic superposition" to denote a "rigid" version of basic superposition. We formalize rigid basic superposition using constraints that is close to the presentation of (Nieuwenhuis and Rubio, 1995). DEFINITION 3.1 (Constraints). By an (ordering) constraint we mean a set of expressions which can be of two kinds: an equality constraint s ' t, or an inequality constraint s - t, where s; t are terms. A substitution ' is a solution to a constraint (respectively, a constraint s - t) if ' is grounding for coincides with t' (respectively, s' - t'). A substitution ' is a solution to a constraint C if ' is a solution to every equality or inequality constraint in C. A constraint C is satisfiable if it has a solution. Constraints C 1 and C 2 are called equivalent if they have the same sets of solutions. We assume that there is an effective procedure for checking constraint satisfiability. For example, there are efficient methods for solving ordering constraints for lexicographic path orderings given by (Nieuwenhuis, 1993; Nieuwenhuis and Rubio, 1995). DEFINITION 3.2. A rigid equation is an expression of the form E ' 8 s ' t, where E is a finite multiset of equations and s; t are terms. Its solution is any substitution ' such that E' ' s' ' t' 2 . Below we shall introduce a system BSE for solving rigid equations. The derivable objects of BSE are constraint rigid equations: DEFINITION 3.3 (Constraint rigid equation). A constraint rigid equation is a pair consisting of a rigid equation R and a constraint C. Such a constraint rigid equation will be denoted R \Delta C. DEFINITION 3.4 (Calculus BSE). The calculus BSE of constraint rigid equations consists of the following inference rules: Left rigid basic superposition: (lrbs) Right rigid basic superposition: 2 The term "rigid equation" could be more adequately expressed as "instance of a (non- simultaneous) rigid E-unification problem", but this would be too lengthy. What You Always Wanted to Know About Rigid E-Unification 7 Equality resolution: Application of the rules is restricted to the following conditions: 1. The constraint at the conclusion of the rule is satisfiable; 2. The right-hand side of the rigid equation at the premise of the rule does not have the form q ' q. 3. In the basic superposition rules, the term p is not a variable. 4. In the left basic superposition rule, s[r] 6= t. The basic restriction in BSE is formalized by representing most general unifiers through equality constraints. Condition 1 has two purposes. The satisfiability of equations in constraints is needed to preserve correctness of the method. The satisfiability of inequality constraints is needed to ensure termination (Theorem 3.9 below). Conditions 3-4 are not essential, they are added as standard optimizations used in paramodulation-based methods. Condition 2 prohibits to apply any rules to rigid equations of the form E ' 8 q ' q. We denote by R \Delta C the fact that R 0 \Delta C 0 is obtained from R \Delta C by an application of one of the inference rules of BSE. The symbol ; denotes the reflexive and transitive closure of ;. 3.5. Consider the rigid equation ha ' a; hx ' a; hb ' fy ` 8 y ' gfy. The ordering - is the Knuth-Bendix ordering (see Martin, 1987) in which all weights of symbols are equal to 1 and which uses the precedence relation f ? h ? a. Under this ordering we have ht - a and f t - hb for every ground term t. The following is a BSE-derivation for this rigid equation: fy ` 8 y ' gfy fy ` 8 y ' ghb \Delta ffy - hb; gfy - fy ' fyg fy ` 8 y ' ga \Deltaff y - hb; gfy - fy ' fy; hx - a; ghb - fy ' fy; hx - a; ghb - By using constraint simplification, i.e. replacement of constraints by equivalent "more simple" constraints we can rewrite this derivation as fy ` 8 y ' gfy fy ` 8 y ' ghb fy ` 8 y ' ga \Delta fghb - 8 Anatoli Degtyarev and Andrei Voronkov THEOREM 3.6 (Soundness of BSE). Let R substitution satisfying C is a solution to R. In particular, R is solvable. Proof. For any constraint C, denote by C ' the constraint obtained from C be removing all inequality constraints. First we note that for every application of an inference rule of BSE of the By induction on the number of inference steps and using the fact C i ' C i+1 , we prove the same statement for multi-step derivations Let R have the form E Applying the obtained statement to the multi-step derivation R t. Hence, any solution to C. We have Any constraint in C ' ` has the form u ' u. Hence, solution to This theorem leads to the following definition: DEFINITION 3.7 (Answer constraint). A constraint C is called an answer constraint for a rigid equation R if for some rigid equation We note that BSE is an incomplete calculus for solving rigid equations. It means that there are solvable rigid equations R that have no answer constraint. For instance, consider the rigid equation 3 x ' a ` 8 gx ' x. It has one solution fga=xg. However, no rule of BSE is applicable to x ' a ` 8 gx ' x \Delta ;. This means that BSE can yield less solutions to a rigid equation than any other known procedure, for example that of (Gallier et al., 1992) because all these procedures are existentially complete. At the same time, BSE can yield more solutions than the procedure of (Gallier et al., 1992) as the following example shows. For the rigid equation a ' fa ` 8 x ' fa the procedure of (Gallier et al., 1992) will find one solution fa=xg, but there are two answer constraints whose solutions are the substitutions fa=xg and ffa=xg respectively. In order to show that there is only finite number of derivations in BSE from a given constraint rigid equation, we prove an auxiliary statement. be an infinite sequence of terms in a finite signature all whose variables belong to a finite set. Then there are numbers and the constraint t i - t j is unsatisfiable. Proof. Following (Kruskal, 1960) we introduce a partial ordering - on terms as the smallest reflexive and transitive relation satisfying 1. 2. if s - t then r[s] - r[t]. 3 Suggested by G.Becher (private communication). What You Always Wanted to Know About Rigid E-Unification 9 By Kruskal's Tree Theorem (Kruskal, 1960) there exist It is easy to see that the constraint t i - t j is unsatisfiable. Similar statements have been proven in (Dershowitz, 1982; Plaisted, 1986). THEOREM 3.9 (Termination of BSE). For any constraint rigid equation R \Delta C, there exists a finite number of derivations from R \Delta C. Proof. Suppose that there exists an infinite number of derivations from R \Delta C. Then, by K-onig's lemma there exists an infinite derivation R \Delta consisting of superpositions. Consider any application of superposition it have the form (lrbs) (the case of right rigid basic superposition is similar). We prove that for every the constraint C n is equivalent to C n [fs[p] - s[r]g. Indeed, the constraint (and hence the constraint C n ) contains pg. By the definition of reduction orderings, if p - r, then s[p] - s[r]. This implies that C n is equivalent to s[r]g. Since every application of rigid basic superposition replaces a literal s[p] ' t (or s[p] 6' t) in R i by a literal s[r] ' t (respectively, s[r] 6' t), there is an infinite sequence of terms t and an increasing sequence of natural numbers with the following property. For every positive natural number k the constraint C n k is equivalent to C n k . Since all terms t k are in the same finite signature and have variables in the same finite set, by Lemma 3.8, there are and the constraint t for all k - j, the constraint C n j is equivalent to C n j , for all k - j. Hence, the constraint is equivalent to C n j . Thus, C is satisfiable. But satisfiability of C implies satisfiability of t i - t j . Contradiction. Note. In (Degtyarev and Voronkov, 1996f) the left rigid basic superposition has been formulated incorrectly in the following way: (lrbs) With this formulation, termination is not guaranteed as the following example shows. Consider the rigid equation fgx ' gx; gx ' a ' 8 a ' b and arbitrary reduction ordering - total on ground terms. The following is an infinite sequence of applications of (lrbs): ag (lrbs) \Deltafgx ' gx; fgx - gx; gx - a; gx - fag (lrbs) ag It is easy to see that the constraint ag is satisfied by the substitution ff a=xg. Inequality constraints are not needed for soundness or completeness of our method. The pragmatics behind inequality constraints is to ensure that the search for solutions of a rigid equation is finite. In addition, the use of ordering constraints prunes the search space. To illustrate this theorem, we consider Example 3.5. The rigid equation of this example has an infinite number of solutions including fb=x; gh n a=yg, for every natural number n. However, all possible BSE-derivations starting with ha ' a; hx ' fy ` 8 y ' gfy \Delta ; give only two answer constraints, one is ffy - hb; gfy - fy ' fy; hx - a; ghb - shown in Example 3.5, another is ffy - hb; gfy - fy ' fy; y ' ghbg obtained from the following derivation: fy ` 8 y ' gfy fy ` 8 y ' ghb \Delta ffy - hb; gfy - fy ' fyg fy ' fy; y ' ghbg This answer constraint can be simplified to fy ' ghbg. Theorem 3.9 yields THEOREM 3.10. Any rigid equation has a finite number of answer constraints. There is an algorithm giving for any rigid equation R the set of all answer constraints for R. What You Always Wanted to Know About Rigid E-Unification 11 In the rules (fl) the variable y does not occur in the premise. Fig. 1. Tableau Expansion Rules 4 Answer constraints and the tableau method In this section we consider how to use the system BSE for theorem proving by the tableau method. Since we only consider skolemized formulas, we have no ffi-rules in tableau calculi. DEFINITION 4.1 (Branch and tableau). A branch is a finite multiset of formulas. A tableau is a finite multiset of branches. A tableau with branches be denoted by \Gamma . The tableau with is called the empty tableau and denoted by #. Often, tableaux are represented in tree form. Representation of tableaux as multisets of branches is more convenient for us for several reasons. For this representation we introduce a logical system allowing to expand tableaux: DEFINITION 4.2 (Tableau expansion rules). The rules (ff), (fi) and (fl) of Figure are called tableau expansion rules. DEFINITION 4.3. Let \Gamma be a branch of a tableau. The set of rigid equations on \Gamma is defined in the following way. A rigid equation t is on \Gamma if E is the multiset of all equations in \Gamma and (s 6' t) 2 \Gamma. We extend the notion of answer constraints to tableau branches: DEFINITION 4.4. A constraint C is an answer constraint for a tableau branch \Gamma if C is an answer constraint for some rigid equation on \Gamma. By Theorem 3.10, we obtain THEOREM 4.5. Any tableau branch has a finite number of answer constraints. There is an algorithm which for any tableau branch \Gamma computes the set of all answer constraints for \Gamma. The following theorem states soundness and completeness of the tableau method with answer constraints: THEOREM 4.6 (Soundness and completeness). Let - be a sentence in Skolem negation normal form. Then - is unsatisfiable if and only if there is a obtained from - by tableau expansion rules and answer constraints C respectively, such that Proof. Soundness follows from soundness of BSE . The proof of completeness is quite lengthy and is given in Appendix A. It is based on the completeness of the equality elimination method (Degtyarev and Voronkov, 1994; Degtyarev and Voronkov, 1995a; Degtyarev and Voronkov, 1996b). To illustrate this theorem, consider the formula of Example 1.1. Assume that we want to prove the formula d oe g(u; x; y) ' g(v; fa; fb))). The negation normal form of :- is 8xyzu((a ' The ordering - is the lexicographic path ordering (see e.g. Nieuwenhuis and Rubio, 1995) based on the precedence g ? f ? a d. For purely illustrative purpose, we shall display tableaux in the tree form. After one quantifier duplication (application of a fl-rule) and some other tableau expansion rules we obtain the following tableau: a ' b c ' d There is one rigid equation on each branch of the tableau: Rigid basic superposition is applicable to none of this rigid equations. Rigid equation (1) has one answer constraint fg(x; u; v) ' g(y; f c; fd)g obtained by an application of the equality resolution rule: Similarly, rigid equation (2) has one answer constraint fg(u; x; y) ' g(v; fa; fb)g. The union of these constraints fg(x; u; v) ' g(y; f c; fd); g(u; x; y) ' g(v; fa; fb)g is total.tex - Date: June 3, 1997 What You Always Wanted to Know About Rigid E-Unification 13 unsatisfiable. Thus, our method does not find solution after one quantifier dupli- cation. After three quantifier duplications and some other tableau expansion steps we obtain the following tableau: a ' b c ' d a ' b c ' d a ' b It has four branches: Consider the following rigid equations R 1 -R 4 on the branches \Gamma 1 -\Gamma 4 , respectively We can apply the following BSE-derivations to R 1 a ' b; a ' b ' 8 a a 14 Anatoli Degtyarev and Andrei Voronkov a a ag The union of the answer constraints of these derivations is a - b; a ' a; This constraint is satisfiable. To check this, we can consider the following substitution 5 Tableau basic superposition As a simple consequence of our results, we prove a completeness result for a paramodulation rule working on tableaux. A paramodulation rule working directly on tableaux was proposed in (Loveland, 1978) in the context of model elimination and later in (Fitting, 1990). However, their formulations have all disadvantages of the early paramodulation rule of (Robinson and Wos, 1969): 1. Functional reflexivity rule is used; 2. Paramodulation into variables is allowed; 3. Increasing applications of paramodulation are allowed (for example, x can be rewritten to f(x). As a consequence, for a given tableau expansion there may be an infinite sequence of paramodulations, in particular due to the use of functional reflexivity or increasing applications of paramodulation. Since the publication of the book (Loveland, 1978), no improvements of the paramodulation-based tableau calculi have been described except for (Plaisted, 1995) who has shown how to transform derivations with resolution and paramodulation to tableaux by introducing a tableau factoring rule. Here we show that paramodulation is complete under the following restrictions: What You Always Wanted to Know About Rigid E-Unification 15 1. No functional reflexivity is needed; 2. Paramodulation into variables is not allowed; 3. Orderings are used so that there are no increasing applications of paramodulation 4. Basic restriction on paramodulation prohibits paramodulation into terms introduced by unification. All these refinements are a consequence of our main result (Theorem 4.6). In order to formalize the basic strategy, we keep the substitution condition as a set of constraints, as before. Thus, we work with constraint tableaux: DEFINITION 5.1 (Constraint tableau). A constraint tableau is a pair consisting of a tableau T and a constraint C, denoted T \Delta C. Now we adapt the tableau rules of (Fitting, 1990) to the case of constraint tableaux. For simplicity, we only consider signatures whose only predicate symbol is '. When we prove a formula ', we construct the Skolem negation normal form / of :' and, starting with the constraint tableau / try to derive the empty tableau # with some satisfiable constraint. DEFINITION 5.2 (Calculus T BSE). The free-variable tableau calculus T BSE is shown in Figure 2. DEFINITION 5.3 (Constraint tableau expansion rules). The rules (ff), (fi) and (fl) of T BSE are called constraint tableau expansion rules. The calculus T BSE has the required completeness property: THEOREM 5.4 (Soundness and completeness). Let ' be a formula in the Skolem negation normal form. It is unsatisfiable if and only if there is a derivation from the constraint tableau ' \Delta ; of a constraint tableau # \Delta C. Proof. Straightforward from Theorem 4.6 by noting that the rules of BSE can be simulated by the corresponding tableau rules. This logical system has one more pleasant property: THEOREM 5.5 (Termination). For any constraint tableau T \Delta C there is only a finite number of derivations from T \Delta C not using constraint tableau expansion rules. (lrbs) In the rules (fl) the variable y does not occur in the premise. The conditions on the rules (lrbs) and (rrbs) are the same as for the corresponding rules of BSE. Fig. 2. Calculus T BSE Proof. Similar to that of Theorem 3.9. This means, that for a given amplification, we cannot have infinite search. Infinite search without any expansion steps is possible in Fitting's system. To illustrate the connection between the tableau rigid basic superposition rule and rules of BSE, we reconsider the example of Section 4. On the branch containing the literal g(x fd) and the equation c ' d, we can apply rigid basic superposition that adds g(x fd) to the branch. Similarly, we can apply rigid basic superposition to the branch containing This results in the tableau given in Figure 3 (the picture does not include the constraint, it is discussed below). After four application of the (er) rules all branches of this tableau become closed. The resulting constraint of this derivation is the same as the union of the answer constraints shown at the end of Section 4. 6 Related work The problem of extending tableaux with equality rules is crucial for enriching the deductive capabilities of the tableau method. Despite the fact that this problem has been attacked by a growing number of researchers during the last years, known solutions are not yet convincing. At the same time tableau methods of automated deduction play an important role in various areas of artificial intelligence and computer science - see e.g. special issues of the Journal of Automated Reasoning, v. 13, no. 2,3, 1994. These issues contain a survey (Schumann, 1994) of implemen- What You Always Wanted to Know About Rigid E-Unification 17 a ' b c ' d a ' b c ' d a ' b Fig. 3. tations of tableau-based provers. Among 24 systems mentioned in the survey only two are able to handle equality. The systems PROTEIN (Baumgartner and Furbach, 1994) and KoMeT (Bibel et al., 1995) implement the modification method of (Brand, 1975). This method transforms a set of clauses with equality into a set of clauses without equality. This transformation usually yields a considerably larger set of clauses. In partic- ular, the symmetry and the transitivity axioms must be explicitly applied to all positive occurrences of the equality predicate. Recently, (Degtyarev and Voronkov, 1996c) proposed a new translation method based on the so-called basic folding demonstrated for Horn clauses. According to (Schumann, 1994), the system 3 T A P uses the method of (Beckert and H-ahnle, 1992). Paper (Beckert and H-ahnle, 1992) claims the completeness of the method, but this claim is not true. The method expands the tableau using the standard tableau rules, including fl-rules. For finding a closing substitution, an analogue of linear paramodulation without function reflexivity has been pro- posed. As it is well known, linear paramodulation is incomplete without function reflexivity. The same is true for the method of (Beckert and H-ahnle, 1992), as the following example shows. Suppose that we prove the formula 9x(a ' b-g(fa; fb) ' In order to prove it using paramodulation, we need to paramodulate a ' b into g(fa; fb) ' h(fa; fb). The method of (Beckert and H-ahnle, 1992) only allows for paramodulation into copies of g(x; x) ' h(x; x) obtained by applications of the fl-rule. Thus, this (provable) formula cannot be proved using the method of (Beckert and H-ahnle, 1992). Consider now approaches based on simultaneous rigid E-unifiability by (Gallier et al., 1987; Gallier et al., 1992) and related methods. We do not consider numerous total.tex - Date: June 3, 1997 works dedicated to (non-simultaneous) rigid E-unifiability. This problem is NP-complete and there exist a number of complete algorithms for its solution (Gallier et al., 1988; Gallier et al., 1990; Goubault, 1993; Becher and Petermann, 1994; De Kogel, 1995; Plaisted, 1995). Since simultaneous rigid E-unification is undecidable (Degtyarev and Voronkov, 1996d; Degtyarev and Voronkov, 1996e), these completeness results are useless from the viewpoint of general purpose theorem proving as proposed in (Gallier et al., 1987; Gallier et al., 1992). Our system BSE can easily be extended to a calculus complete for rigid E-unifiability, but such completeness is not our aim. We tried to restrict the number of possible BSE- derivations preserving completeness of the general-purpose method of Section 4. It is not known whether the procedure described in (Gallier et al., 1992) is complete for theorem proving 4 . Even if it is complete, our procedure based on BSE has some advantages over Gallier et.al.'s procedure. For example, for every tableau branch with p equations and q disequations, we consider q rigid equations, while Gallier et.al.'s procedure checks q \Delta 2 p rigid equations. (Gallier et al., 1988; Gallier et al., 1990) introduced the notion of a complete set of solutions for rigid E-unification, proved finiteness of such sets and gave an algorithm computing finite complete set of solutions. According to this result, (Goubault, 1994) proposed to solve simultaneous rigid E-unifiability by using finite complete sets of solutions to the components of the simultaneous problem. Paper (Goubault, 1994) contained faulty results. The undecidability of simultaneous rigid E-unification shows that finite complete sets of solutions do not give a solution to the simultaneous problem. The reason for this is that substitutions belonging to complete sets of solutions for different rigid equations are minimal modulo different congruences. (Petermann, 1994) introduces a "complete connection calculus with rigid E- unification". Here the completeness is achieved by changing the notion of a complete set of unifiers so that solutions to all subproblems are compared modulo the same congruence (generated by the empty equality theory). In this case, a non-simultaneous problem can have an infinite number of solutions and no finite complete set of solutions. For example, for the rigid E-unification problem f(a) ' a ` 8 x ' a the complete set of solutions in the sense of (Gallier et al., 1992) consists of one substitution fa=xg (and there is only one answer constraint ag obtained by our method), but the complete set of solutions in the sense of (Petermann, 1994) is infinite and consists of substitutions ff n (a)=xg, for all :g. This implies that proof-search by the method of (Petermann, 1994) can be non-terminating even for a limited number of applications of the fl-rule (i.e. for a particular tableau), unlike algorithms based on the finite complete sets of unifiers in the sense of (Gallier et al., 1992) or based on minus-normalization (Kanger, 1983). The implementation of the method of (Petermann, 1994) uses a completion-based procedure by (Beckert, 1994) of generation of complete sets of 4 For example, the completeness of Gallier et.al.'s procedure does not follow from our method because, as noted above, our calculus BSE can give more solutions to some rigid equations. What You Always Wanted to Know About Rigid E-Unification 19 rigid E-unifiers. This procedure is developed with the aim of solving a more general problem - so-called mixed E-unification and has been implemented as part of the tableau-based theorem prover 3 T A P . Complete sets of unifiers both in the sense of (Gallier et al., 1992) and in the sense of (Petermann, 1994) can be computed by this procedure in the case when all variables are treated as rigid. However, the termination is not guaranteed even for complete sets of rigid E-unifiers in the sense of (Gallier et al., 1992). (Plaisted, 1995) gives "techniques for incorporating equality into theorem prov- these techniques have a rigid flavor". His method of path paramodulation guarantees termination for a given amplification and, in the case of success "solves the simultaneous rigid E-unification problem, in a sense". However, this does not solve the problem attacked by a number of researchers: extend the method of matings to languages with equality by rigid E-unification. First, unlike (Gallier et al., 1992) the search for solutions for a given amplification is not incremental (the method does not allow "branch-wise" computation of solutions to rigid E- unification for separate branches). Second, within a given amplification Plaisted uses factoring rules which involves two branches (paths). As a consequence, even when the original formula contains no equality, his method results in the standard tableau calculus plus the factoring rule. In fact, path paramodulation of (Plaisted, 1995) simulates resolution-paramodu- lation inference in a connection-like calculus. Although it is not noted in (Plaisted, 1995), this technique has been demonstrated for resolution in many papers, for example in (Bibel, 1981; Eder, 1988; Eder, 1991; Mints, 1990; Baumgartner and Furbach, 1993; Avron, 1993). The generalization of this simulation to paramodulation is straightforward. However, this simulation technique is insufficient for proving the results of our paper since, in particular, it gives no insight on how to avoid factoring in tableaux with equality. The use of factoring prevents not only from the independent search for solutions for tableau branches, but even from the incremental solving of rigid equations on tableau branches as proposed by Gallier et.al. Our equality elimination method (Degtyarev and Voronkov, 1996b; Degtyarev and Voronkov, 1995b; Degtyarev and Voronkov, 1995a) is based on extending a tableau prover by a bottom-up equation solver using basic superposition. Solutions to equations are generated by this solver and used to close branches of a tableau. Thus, the method combines (non-local) tableau proof search with the (local) equation solving. Only completely solved equations are used in the tableau part of the proof, thus reducing non-determinism created by applications of the MGU replacement rule of (Fitting, 1990). The equation solution is even more restricted by the use of orderings, basic simplification and subsumption. A similar idea: combination of proof-search in tableaux with a bottom-up equality saturation of the original formula is used in (Moser et al., 1995) for constructing a goal-directed version of model elimination and paramodulation. One of advantages of the tableau method is its applicability to non-classical logics. However, handling equality in non-classical logics seems to be a much more difficult problem than that in classical logic. For example, it is shown in (Voronkov, 1996) that procedures for intuitionistic logic with equality must handle simultaneous rigid E-unification. This implies that our method based on BSE does not give a complete procedure for intuitionistic logic with equality. Other results on relations between simultaneous rigid E-unification and intuitionistic logic are considered in (Degtyarev and Voronkov, 1996a; Degtyarev et al., 1996a; Degtyarev et al., 1996b; Veanes, 1997). An extensive discussion of equality reasoning in sequent-based systems may be found in (Degtyarev and Voronkov, 1997). Acknowledgments We thank Gerard Becher and Uwe Petermann for helpful discussions, anonymous referees for their valuable comments and the guest editors Jose Julio Alves Alferes and Luis Moniz Pereira for their patience. --R Theorem proving via general matings. Basic paramodulation and superposition. Basic paramodulation. Consolution as a framework for comparing calculi. PROTEIN: A PROver with a Theory Extension INterface. An ordered theory resolution calculus. Rigid unification by completion and rigid paramodulation. An improved method for adding equality to free variable semantic tableaux. A completion-based method for mixed universal and rigid E-unification Issues in theorem proving based on the connection method. On matrices with connections. Proving theorems with the modification method. Theorem proving with variable-constrained resolution Rigid E-unification simplified Equality elimination for semantic tableaux. Equality elimination for the inverse method and extension proce- dures General connections via equality elimination. Decidability problems for the prenex fragment of intuitionistic logic. Equality elimination for the tableau method. Handling equality in logic programs via basic folding. Simultaneous rigid E-unification is undecidable The undecidability of simultaneous rigid E-unification What you always wanted to know about rigid E-unification Equality reasoning and sequent-based methods Simultaneous rigid E-unification and related algorithmic problems "Logic in Computer Science" Orderings for term rewriting systems. Consolution and its relation with resolution. First Order Logic and Automated Theorem Proving. Theorem proving using rigid E-unification: Equational matings IEEE Computer Society Press Rigid E-unification is NP-complete Rigid E-unification and its applications to equational matings Rigid E-unification: NP-completeness and applications to equational matings Theorem proving using equational matings and rigid E-unification The many-valued tableau-based theorem prover 3 T A P Technical Report 30/94 A simplified proof method for elementary logic. Well quasi ordering Symbolic Logic and Mechanical Theorem Proving. Automated Theorem Proving: a Logical Basis. How to choose weights in the Knuth-Bendix ordering Invertible sequential variant of constructive predicate calculus (in Russian). On variants of classical predicate calculus with the unique deduction tree (in Russian). Model elimination with basic ordered paramodula- tion Basic superposition is complete. Theorem proving with ordering and equality constrained clauses. Journal of Symbolic Computations On the size of derivations under minus-normalization in Russian A complete connection calculus with rigid E-unification A simple non-termination test for the Knuth-Bendix method Special cases and substitutes for rigid E-unification An improved proof procedure. Paramodulation and theorem-proving in first order theories with equality An algorithm for reasoning about equality. On Simultaneous Rigid E-Unification Proof search in intuitionistic logic with equality --TR
equality reasoning;tableau method;rigid E-unification
594229
Integrating Computer Algebra into Proof Planning.
Mechanized reasoning systems and computer algebra systems have different objectives. Their integration is highly desirable, since formal proofs often involve both of the two different tasks proving and calculating. Even more important, proof and computation are often interwoven and not easily separable.In this article we advocate an integration of computer algebra into mechanized reasoning systems at the proof plan level. This approach allows us to view the computer algebra algorithms as methods, that is, declarative representations of the problem-solving knowledge specific to a certain mathematical domain. Automation can be achieved in many cases by searching for a hierarchic proof plan at the method level by using suitable domain-specific control knowledge about the mathematical algorithms. In other words, the uniform framework of proof planning allows us to solve a large class of problems that are not automatically solvable by separate systems.Our approach also gives an answer to the correctness problems inherent in such an integration. We advocate an approach where the computer algebra system produces high-level protocol information that can be processed by an interface to derive proof plans. Such a proof plan in turn can be expanded to proofs at different levels of abstraction, so the approach is well suited for producing a high-level verbalized explication as well as for a low-level, machine-checkable, calculus-level proof.We present an implementation of our ideas and exemplify them using an automatically solved example.Changes in the criterion of rigor of the proof engender major revolutions in mathematics. H. Poincar, 1905
Introduction The computer and the development of high-level programming languages made possible the mechanisation of logic as well as the realisa- M. KERBER, M. KOHLHASE, V. SORGE. tion of mechanical symbolic calculations, we could witness in the last forty years. This has lead to two rather disjoint academic fields: mechanised reasoning and computer algebra, which each have their own meth- ods, interests and traditions, even though they share common roots: none of the two fields is imaginable without the underlying foundation of mathematical logic or the mathematical study of symbolic calculations (leading to such algorithms and methods as the determination of the GCD or the Gau-ian elimination). Only in the last decade we have seen a move towards an integration of the fields driven by the insight that real-world formal problems often involve a mixture of both computation and reasoning, hence an integration of mechanised reasoning systems and computer algebra systems is highly desirable (cf. [Buc85]). This is the case in particular, since deduction systems are very weak, when it comes to computation with mathematical objects, and computer algebra systems manipulate highly optimised representations of these objects, but do not yield any formally checkable proof information (if they give any explanation at all). In the remainder of this introduction we briefly summarise key points of mechanised reasoning systems as well as of computer algebra systems and then give a short preview on the integration approach advocated in this paper. By its nature, such a short description has to abstract from many details and to simplify considerably. 1.1. Mechanised Reasoning Systems Mechanised reasoning systems (for short MRS in the following) are built with various purposes in mind. One goal is the construction of an autonomous theorem prover, whose strength achieves or even surpasses the ability of human mathematicians. Another is to build a system where the user derives the proof, with the system guaranteeing its cor- rectness. A third purpose consists in modelling human problem-solving behaviour on a machine, that is, cognitive aspects are the focus. Advanced theorem proving systems often try to combine the different goals, since they can complement each other in an ideal way. Let us roughly divide existing theorem-proving systems into three categories: machine-oriented theorem provers, proof checkers, and human-oriented (plan-based) theorem provers. Normally all these systems do not exist in a pure form anymore, and in some systems like our own\Omega mega system [BCF + 97] it is explicitly tried to combine the reasoning power of automated theorem provers as logic engines, the specialised problem solving knowledge of the proof planning mechanism, and the interactive support of tactic-based proof development environments. We think that the combination of these KeKoSo.tex; 24/06/1998; 14:49; no v. INTEGRATING COMPUTER ALGEBRA INTO PROOF PLANNING 3 complementary approaches inherits more advantages than drawbacks, because for most tasks domain-specific as well as domain-independent problem-solving know-how is required and for difficult tasks more often than not an explicit user-interaction should be provided. While such an approach seems to be general enough to cope with any kinds of logic-level proofs, it neglects the fact that for many mathematical fields, the everyday work of mathematicians only partially consists in proving or verifying theorems. Calculation plays an equally important r-ole. In some cases the tasks of proving theorems and calculating simplifications of certain terms can be separated from each other, but very often the tasks are interwoven and inseparable. In such cases an interactive theorem proving environment will only provide rather poor support to a user. Although theoretically any computation can be reduced to theorem proving, this is not practical for non-trivial cases, since the search spaces are intractable. For many of these tasks, however, no search is necessary at all, since there are numerical or algebraic algorithms that can be used. If we think of Kowalski's equation purpose procedures do not (and cannot) provide the control for doing a concrete computation. 1.2. Computer Algebra Systems Early computer algebra systems (CAS for short) developed from collections of algorithms and data structures for the manipulation of algebraic expressions like the multiplication of polynomials, or the derivation and integration of functions [Hea95]. Abstractly spoken, the main objective of a CAS can be viewed in the simplification of an algebraic expression or the determination of a normal form. Today there is a broad range of such systems, ranging from very generally applicable systems to a multitude of systems designed for specific applications. Unlike MRS, CAS are used by many mathematicians as a tool in their everyday work, they are even widely applied in sciences, engineering and economics. Their high academic and practical standard reflects the fact that the study of symbolic calculation has long been an established and fruitful subfield of mathematics that has developed the mathematical theory and tools. Most modern systems [Wol96, CGG have in common that the algebraic algorithms are integrated in a very comfortable graphical user interface that includes formula editing, visualisation of mathematical objects and even an interface to programming languages. As in the case of MRS the representation languages of CAS differ from system to system, which complicates the integration of such systems as well as the cooperation between them. This deficiency has been attacked in 4 M. KERBER, M. KOHLHASE, V. SORGE. the OpenMath initiative [AvLS96], which strives for a standard CAS communication protocol. Currently the main emphasis is laid on standardising the syntax and the computational behaviour of the mathematical objects, while their properties or semantics are not considered. That means there is no explicit representation format for theorems, lemmata and proofs. Some specific systems allow to specify mathematical domains and theories. For instance in systems like MuPad [Fuc96] or Axiom [JS92], computational behaviour can be specified by attaching types and axiomatisations to mathematical objects; but this also falls short of a comprehensive representation of all relevant mathemat- ics. Furthermore, almost all CAS fail to give an explanation or proof of their solution to the problem at hand, even though some mathematical theories like that of Gr-obner bases can be successfully applied to theorem proving in elementary geometry [Cho88, Kap88, CGZ94, Wu94]. 1.3. Contributions of this Paper Not only the fact that a mutual simulation of the tasks of an MRS and a CAS can be quite inefficient, but more that the daily work of mathematicians is about proving and calculating points to the integration of such systems, since mathematicians want to have support in both of their main activities. Indeed two independent systems can hardly cover their needs, since in many cases the tasks of proving and calculating are hardly separable. As pointed out by Buchberger [Buc96a] the integration problem is still unsolved, but it can be expected that a successful combination of these systems will lead to "a drastic improvement of the intelligence level" of such support systems. Our paper addresses two immediate questions occurring in the integration of automated reasoning and computation systems. How can the algorithms be integrated, so that the underlying mathematical knowledge is mutually respected and a synergy effect is achieved? How can the correctness problem inherent in any such combination be addressed? In particular, how can results from the CAS be integrated into a proof without having to completely trust the We advocate an integration of computer algebra into mechanised reasoning systems using the proof planning paradigm. This approach allows to encapsulate the computer algebra algorithms into meth- ods, that is, declarative representations of the problem solving knowledge specific to a certain mathematical domain. The proof planning paradigm enables a user to guide a proof or to fully hand over the con- INTEGRATING COMPUTER ALGEBRA INTO PROOF PLANNING 5 trol to a planner, which in turn can use computer algebra systems, if the specifications for the corresponding algorithms are met. The use of hierarchic proof plans at the method-level gives a suitable granularity of integration, since it allows to directly use existing (human) control knowledge about the interplay of computation and reasoning. A proper integration into the proof planning approach answers the question about the correctness automatically, since the corresponding questions are solved for proof planning. In this area a proof plan can either be rejected (the tactics are not executable, hence the plan cannot be used to build a proof) or executed. The later results either in a further planning phase to fill in possible gaps or in an accepted machine-checkable proof. Hence a proper integration requires that the computer algebra system produces high-level protocol information that can be processed by an interface to derive proof plans which themselves can be seamlessly integrated into the overall proof plan generated in the problem solving attempt. Since this can be expanded into an explicit, checkable proof in order to obtain a correctness guarantee for the combined solution, we have also given a principled answer to the correctness problem. The feasibility of the approach advocated in the sequel has been verified by integrating a simple CAS into the\Omega mega proof planning system. Therefore, we organise the paper around this experiment and describe the relevant features with a system perspective. Our approach requires a mode of the CAS that generates information from which it is possible to generate a proof plan. For that reason the integration of a standard CAS makes major adaptations unavoidable (in particular it is necessary to change the source code of these systems). Our approach is not committed to the particular systems involved, in particular, the work reported here should be understood rather as a proof of principle than as the development of a state-of-the-art integrated system. Moreover, we will make the details of the approach more concrete by explaining them by means of an example that cannot easily be solved by either a mechanised reasoning system or a computer algebra system alone, but that needs the combined efforts of systems of each kind. 2. Related Work We give a short description of some of the experiments to combine MRS and CAS and roughly categorise them into three classes with respect to the treatment of proofs that is adopted, that is, with respect to the correctness issue. In doing so we only describe in detail the approaches of integrating CAS into MRS, that is, essentially the MRS is the master KeKoSo.tex; 24/06/1998; 14:49; no v. 6 M. KERBER, M. KOHLHASE, V. SORGE. and the CAS the slave, since our approach is also of this kind. With the same right, one can of course follow the converse direction, namely to approach the integration from the point of the CAS and indeed such approaches are also successfully undertaken (see e.g. [CZ92, Buc96]). The question about the granularity of integration is treated uniformly by all these experiments. The application of the CAS is treated as another (derived) rule of inference at the level of the (tactic) calcu- lus, so the granularity of integration depends on the granularity of the calculus or the tactics involved. In the first category of attempts (see e.g. [HT93b, BHC95]) one essentially trusts that the CAS properly work, hence their results are directly incorporated into the proof. All these experiments are at least partly motivated by achieving a broader applicability range of formal methods and this objective is definitively achieved, since the range of mathematical theorems that can be formally proved by the system combinations is much greater than that provable by MRS alone. However, CAS are very complex programs and therefore only trustworthy to a limited extent, so that the correctness of proofs in such a hybrid system can be questioned. This is not only a minor technical problem, but will remain unsolved for the foreseeable future since the complexity (not only the code complexity, but also the mathematical complexity) of a CAS does not permit a verification of the program itself with currently available program verification methods. Conceptually, the main contribution of such an integration is the solution of the software-engineering problem how to pass the control between the programs and translating results forth and back. While this is an important subproblem, it does not seem to cover the full complexity of the interaction of reasoning and computation found in mathematical theorem proving. In an alternative approach that formally respects correctness, but essentially trusts CAS, an additional assumption standing for the CAS is introduced, so that essentially formulae are derived that are proved modulo the correctness of the computer algebra system at hand (see e.g. [HT93b]). The second category (for which [HT93a] is paradigmatic) is more conscious about the r-ole of proofs, and only uses the CAS as an oracle, receiving a result, whose correctness can then be checked deductively. While this certainly solves the correctness problem, this approach only has a limited coverage, since even checking the correctness of a calculation may be out of scope of most MRS, when they don't have additional information. Indeed from the point of applicability, the results of the CAS help only in cases, where the verification of a result is simpler than its discovery, such as prime factorisations, solving equations, or symbolic integration. For other calculations, such as symbolic addition or multiplication of polynomials and differentiation, the verification is just as complex as the calculation itself, so that employing the CAS KeKoSo.tex; 24/06/1998; 14:49; no v. INTEGRATING COMPUTER ALGEBRA INTO PROOF PLANNING 7 does not speed up the proof construction process. Typically in longer calculations, both types of sub-calculations are contained. A third approach of integrating computer algebra systems into a particular kind of mechanised reasoning system, consists in the metatheoretic extension of the reasoning system as proposed for instance in [BM81, How88] and been realised in Nuprl [Con86]. In this approach a constructive mechanised reasoning system is basically used as its own meta-system, the constructive features are exploited to synthesise a correct computer algebra system and due to bridge rules between ground and meta-system it is possible to integrate the so-built CAS that it can be directly used as a component. The theoretical properties of the meta-theoretic extension guarantee that if the original system was correct then the extended system is correct too. This method is the most appealing one from the viewpoint of correctness, although the assumption that the original (also rather complex) system must be correct can hardly be expected to be self-evident for any non-trivial system. A disadvantage compared to the other two approaches is that it is not possible to employ an existing CAS, but that it is necessary to (re)implement one in the strictly formal system given by the basic MRS. Of course this is subject to the limitations posed by the (mathematical and software engineering) complexities mentioned above. The main problem of integrating CAS into MRS without violating correctness requirements is that CAS are generally highly optimised towards maximal speed of computation but not towards generating explanations of the computations involved. In most cases, this is dealt with by meta-theoretic considerations why the algorithms are adequate. This lack of explanation makes it not only impossible for the average user to understand or convince himself of the correctness of the com- putation, but leaves any MRS essentially without any information why two terms should be equal. This is problematic, since computational errors have been reported even for well-tested and well-established CAS. From the reported categories of approaches only the last one seriously addresses this problem. 3.\Omega mega as an Open System for Integrating Computation \Omega mega is a proof development system, based on the proof planning paradigm. In this section we describe its architecture and components and show how this supports the integration of computer algebra sys- tems. Since the goal of this paper is not to present a system description of\Omega mega, but to document the integration of computer algebra into it, we try to be as concise as possible and introduce the relevant parts KeKoSo.tex; 24/06/1998; 14:49; no v. 8 M. KERBER, M. KOHLHASE, V. SORGE. only, the general architecture, the proof planner, and the integration possibilities for external reasoners. 3.1. The Proof Development Environment\Omega\Gammamen The entire process of theorem proving in\Omega mega can be viewed as an interleaving process of proof planning, execution and verification centred around a hierarchical proof plan data structure. Several integrated tools support the user in interacting with the system. Some of them are also available to the proof planner. Theory Database Since methods and control knowledge used in proof planning are most- ly domain-specific,\Omega mega organises the mathematical knowledge in a hierarchy of theories. Theories represent signature extensions, axioms, definitions, and methods that make up typical established mathematical domains. Each theorem has its home theory and therefore has access to the theory's signature extensions, axioms, definitions, and lemmata without explicitly introducing them. A simple inheritance mechanism allows to incrementally build larger theories from smaller parts. We give an overview of the part of\Omega mega's theory database that is necessary for solving our extended example in Figure 1. Proof Explanation Proof presentation is one important feature of a mathematical assistant that has been neglected by traditional deduction systems.\Omega mega employs an extension of the Proverb system [HF96] developed by our group that allows for presenting proofs and proof plans in natural lan- guage. In order to produce coherent texts that resemble those found in mathematical textbooks, Proverb employs state-of-the-art techniques of natural language processing. Due to the possibly hierarchical nature of\Omega mega proofs, these can be verbalised at more than one level of abstraction, which can be selected by the user. To summarise our view of proofs, for every theorem an explicit proof has to be constructed so that on the one hand it can be checked by a proof checker, on the other hand the system provides support to represent this proof in a high-level form that is easily readable by humans [HF96]. Neither the process of generating proofs nor that of checking them is fully replaced by the machine but only supported. If a human mathematician wants to see a proof, he/she can do so at an appropriate level of abstraction. INTEGRATING COMPUTER ALGEBRA INTO PROOF PLANNING 9 3.2. Proof Planning The central data structure for the overall process is the Proof plan Data Structure (PDS). This is a hierarchical data structure that represents a (partial) proof at different levels of abstraction (called proof plans). It is represented as a directed acyclic graph, where the nodes are justified by (LCF-style) tactic applications. Conceptually, each such justification represents a proof plan (the expansion of the justification) at a lower level of abstraction that is computed when the tactic is executed 1 . In \Omega mega, we explicitly keep the original proof plan in an expansion hier- archy. Thus the PDS makes the hierarchical structure of proof plans explicit and retains it for further applications such as proof explanation or analogical transfer of plans. Once a proof plan is completed, its justifications can successively be expanded to verify the well-formedness of the corresponding PDS. This verification phase is necessary, since the correctness of the different components (in particular, that of external ones like automated theorem provers or computer algebra systems) cannot be guaranteed. When the expansion process is carried out down to the underlying ND-calculus (natural deduction), the soundness of the overall system relies solely on the correctness of the verifier and of ND. This also provides a basis for the controlled integration of external reasoning components if each reasoner's results can (on demand) be transformed into a sub-PDS. The level to which the proofs have to be expanded depends on the sophistication of the proof checker. As pointed out by Barendregt [Bar96], a more complex proof-checker that accepts proofs in a more expressive formalism may drastically reduce the length of the communicated proofs. If the high-level justifications are not expanded but accepted as they are, our approach reduces to one in which the computer algebra system is fully trusted. In short, the hierarchical nature of the PDS supports the full spectrum of user preferences, from total trust in the CAS, over partial trust in certain levels to full expansion of the proofs in a detailed calculus level description that is machine checkable. A PDS can be constructed by automated or mixed-initiative plan- ning, or pure user interaction that can make use of the integrated tools. In particular, new pieces of PDS can be added by directly calling tac- tics, by inserting facts from a database, or by calling some external reasoner (cf. Section 3.3) such as an automated theorem prover or a computer algebra system. Automated proof planning is only adequate for 1 This proof plan can be recursively expanded, until we have reached a proof plan that is in fact a fully explicit proof, since all nodes are justified by the inference rules of a higher-order variant of Gentzen's calculus of natural deduction (ND). M. KERBER, M. KOHLHASE, V. SORGE. problem classes for which methods and control knowledge have already been established. The goal of proof planning is to fill gaps in a given PDS by forward and backward reasoning [HKKR94] (proof plans were first introduced by Bundy, see [Bun88, BSvH 93]). Thus from an abstract point of view the planning process is the process of exploring the search space of planning states that is generated by the plan operators in order to find a complete plan from a given initial state to a terminal state. \Omega mega's proof planner is an extension of the well-known STRIPS algorithm that can be evoked to construct a proof plan for a node (the goal node) from a set I of supporting nodes (the initial state) using a set Ops of proof planning operators, here called methods. A method is a (partial) specification of a tactic in a meta-level language. In\Omega mega planning is combined with hierarchical expansion of methods and precondition abstraction. The plans found by this procedure are directly incorporated into the PDS as a separate level of abstraction. In this model, the actual reasoning competence of the planner and the user builds upon the availability of appropriate methods together with meta-level control knowledge that guides the planning. At the moment,\Omega mega provides user-defined method ratings as a means of control and can use analogy as a control strategy of the planner. Two examples of methods are displayed in the section on the extended exam- ple, Section 3.4. 3.3. Integration of Computer Algebra Systems as External Reasoners According to the different modes of\Omega mega there are different levels on which an external reasoning system, RSys, can be integrated: Interactive calls, RSys is represented as a command call-RSys that invokes the reasoner on a particular subproblem and returns the result, Proof planning, RSys is represented as a method whose specification contains knowledge about the problem solving behaviour and option settings for RSys. Justifications, RSys can serve as a justification of a declaratively given subgoal that is left to be proved by RSys. In any case, the proof found by RSys must eventually be transformed into a PDS, since this is the proof-theoretic basis of\Omega mega. For automated theorem provers like Otter [McC94], we described the integration in [HKK + 94] and the necessary proof transformation to PDS in [HF96], so we will not pursue this matter here. The integration of KeKoSo.tex; 24/06/1998; 14:49; no v. INTEGRATING COMPUTER ALGEBRA INTO PROOF PLANNING 11 CAS follows the same paradigm and is the main topic of this paper, so we will develop the paradigm for the case of external computations in\Omega mega. We will see examples for the three different levels of integrations of a CAS into\Omega mega in the example in the next section, so we will not go into that here. This leaves us with the question of the transformation of the CAS results into PDS. If we take the idea of generating explicit PDS seriously also for computations we can neither just take existing systems nor follow the approach of meta-theoretic extensions, since\Omega mega is a classical proof system and does not use constructive logic. On the other hand we cannot using them even in cases, where the verification of a calculation is much easier than the calculation itself (e.g., integration of functions); the computation needed for verifying alone is in many cases still much too complicated to be automatically checked without any guidance. For instance even the proof for the binomial formula problem for any computer algebra system) needs more then 70 single steps in the natural deduction calculus 2 . Thus using theorem provers or rewriting systems to find such proofs can produce unnecessarily large search spaces and thus absorb valuable resources. On the other hand such proofs show a remarkable resemblance to algebraic calculations themselves and suggest the use of the CAS not only to instantly compute the result of the given problem, but also to guide a proof in the way of exploiting the implicit knowledge of the algorithms. We propose to do this extraction of information not by trying to reconstruct the computation in the MRS after the result is generated - as we have seen, even in case of a trivial example for a CAS this may turn out to be a very hard task for an MRS - but rather to extend the CAS algorithm itself so that it produces some logically usable output alongside the actual computation. Surely in most cases a user would not like to see proofs at a level where the binomial formula is explained (although a novice might want to). This means that a hierarchical approach to proof generation is appropriate, in which the abstraction level of the proof presentation can be chosen by the user. Our approach is to use the mathematical knowledge implicit in the CAS to extract proof plans that correspond to the mathematical computation in the CAS. So essentially the output of a CAS should be transferable into a sequence of tactics, which presents a high-level description for the proof of correctness of the computation the CAS has performed. Note that this does not prove general correctness of the algorithms involved, instead it only gives a proof for a particular instance of computation. The high-level description can then be used to 2 Proofs of this length are among the hardest ever found by totally automatic theorem provers without domain-specific knowledge. M. KERBER, M. KOHLHASE, V. SORGE. produce a readable explanation or further expanded to a level that can be automatically checked by proof checkers. The level of abstraction on which the checking can take place, depends on the level of sophistication of the proof checker. For a naive proof checker, the proof must be expanded to an explicit calculus level. The decision to extract proof plans rather than concrete proofs from the CAS is essential to the goal of being verbose without transmitting too much detail. For our purpose, we need different modes, in which we can use the CAS. Normally, during a proof search, we are only interested in the result of a computation, since the assumption that the computation is correct is normally justified for established CAS. When we want to understand the computation - in particular, in a successful proof - we need a mode of the CAS that gives enough information to generate a high-level description of the computation in terms of the mathematics involved. This is is described in the next section in detail. Before doing so we describe how the integrated system automatically solves an extended example from an economics examination. 3.4. Extended Example The concrete task at hand is to minimise the costs for running a machine while producing a certain product. Problem: The output of a machine can range over a certain interval, the interval I = [1; 7]. The cost of the product prod is determined by the costs of water and electricity for producing prod , which are given by the functions prod ffl r prod and the prices for water and electricity kWh Determine the output d in I of the machine such that the total costs are minimal. This example serves our purposes for several reasons. Firstly, it allows us to show the interaction of proof planning with symbolic computation and the extraction of proof plans from calculations. Secondly, the mathematics involved is simple enough to be fully explained (only simple polynomial manipulations are necessary). Thirdly, it is not an example we created, but the problem is a slightly varied version of a KeKoSo.tex; 24/06/1998; 14:49; no v. INTEGRATING COMPUTER ALGEBRA INTO PROOF PLANNING 13 minimisation problem from a masters examination in economics at the Universit?t des Saarlandes, Saarbr-ucken [WiW89]. In order to solve problems like this, we have integrated a simple CAS into\Omega mega, called -Cas 3 . The -Cas-system is very simple and can at the moment only perform basic polynomial manipulations and differentiation, but it suffices for automatically solving the example at hand. Clearly, for a practical system for mathematical reasoning, a much more developed system like Maple [CGG + 92], Reduce [Hea95], Axiom [JS92], or Mathematica [Wol96] has to be integrated. The technicalities of the integration will be described in Section 4. For the formalisation of the example, we use the theory mecha- nism of\Omega mega to create a theory economy (see Figure 1) that contains the domain-specific knowledge (both the factual and the method knowledge) needed for the problem solution. Obviously, we need a background theory of costs in economics (that handles both numerical parts and denomination of cost functions) and one of minimisation of real functions, therefore, our theory inherits material from the theories costs and calculus. The calculus theory is provided by\Omega mega and contains relevant parts of the knowledge of an elementary calculus textbook: For instance, the real numbers are introduced as a complete, dense archimedian field (based on elementary algebraic notions such as groups and rings defined in the respective theories). The set of real numbers (showing the existence of such a complete, dense archimedian field) are constructed as the quotient field of the ring of sequences of rational numbers over the ideal of null-sequences. The rational numbers in turn are constructed as signed fractions of natural numbers that are defined from the Peano axioms in theory natural. All of these mathematical theories are based on the theories function, set, and relation, that specify naive (simply typed) set theory and the properties of functions and relations on such sets. Finally, the whole hierarchy builds on the theory base, which declares the underlying logic by providing the logical connectives and quantifiers and the basic ND inference rules. The theory economy provides a type AE of units that covers the different units of denominations - in our example (for work), prod (for product) and DM (for the price). We then formalise prices as triples consisting of one real number and two units and cost functions as a real function together with two units (read as input/output units). Note, that just as in the real world, addition (\Phi) 3 The -Cas system is part of the standard distribution of\Omega mega which can be obtained from http://www.ags.uni-sb.de/software/deduktion/omega. The example is accessible as WiWi-Exam in the theory economy. 14 M. KERBER, M. KOHLHASE, V. SORGE. base set relation struct ordered topology ordered-ring ordered-field metric-space natural rational sequences integer semigroup group ring field monoid real calculus economy costs Figure 1. Theory hierarchy in\Omega mega's knowledge base multiplication (\Omega ) and comparison of costs and cost functions is defined as that of their real parts with respect to the denominations. For these calculations we have the axioms CF1 and CF2. If two denominations differ, we can relate them by their prices, for this purpose we use axiom Pr. INTEGRATING COMPUTER ALGEBRA INTO PROOF PLANNING 15 Optimisation in economy is formalised by a predicate Opt on a cost function cf(f; DM ; prod ) and an interval I that is true, whenever f has a total minimum 4 on I. O Opt(cf(f; DM ; prod ); I) , 9x TotMin(x; f; I) Thus we can state the problem as the following formula 5 where H is a set of hypotheses that are needed for the complete proof, for instance the price axioms The planner solves the problem by generating a high-level proof plan consisting of methods from its domain specific method base on economics exam questions 6 . We are going to outline this process by describing its major steps. In particular, we will demonstrate how the proof planner of\Omega mega and the -Cas-system interact, and make explicit, on which entries of a mathematical database this interaction depends. The planner finds the following simple proof plan: 4 The predicate TotMin and the problem solving knowledge related to it is inherited from the theory calculus. 5 Actually the formalisation of the problem is not fully correct, since the examiner is not only interested in the proof that there exists such an x, but he/she wants to know the value of x as well as a proof that this value fits the requirements. Obviously, such an answer cannot be obtained from the formula here, but only from a proof that is constructive for the variable x, where we can extract a witness term. This is no problem for a CAS nor for an MRS based on constructive logic, but for a traditional MRS based on classical logic, the proof construction process has to be refined to guarantee constructivity for x. Note that the arguments, why the witness for x meets the requirements can still be classical and non-constructive. this means that the proof planner may only use methods in our proof plan that are constructive to get the wanted answer as presented here and not a non-constructive abstract argument. Finally note that this phenomenon is another argument in favour manipulating explicit proofs. Without this, one may find oneself in the position, that one is convinced (by meta-theoretic arguments) of the existence of a (constructive) proof, but in fact without one from which to extract a term witness to answer the exam question. 6 Questions for certain standard exams are a good example for a very restricted mathematical domain, since the proofs and calculations involved are highly stan- dardised. Therefore finding the proof plan in this example is not a big problem for \Omega mega. M. KERBER, M. KOHLHASE, V. SORGE. Mult-by-Price 3 Add-by-Denom 5 TotMin-Rolle where the first three methods compute the actual cost function by adjusting the denominations and adding. Method 4 uses Axiom O for optimisation. As the example only contains polynomials of degree two, the planner selects a method TotMin-Rolle (cf. Figure 3) for finding total minima that makes implicit use of Rolle's theorem from the calculus theory: Let f be a polynomial of degree two, then f has a total minimum at has a minimum at x and f(a) - f(x) - f(b). Formally we get the following equivalence: TotMin TotMin(x; f; [a; b]) , x 2 [a; Note that Rolle's theorem is accessible in the current theory and, to ensure correctness, the database has to contain its formal proof. Now let us take a closer look at some of the methods in order to get a feeling of how this initial proof plan can be expanded. In Figures 2 and 3 we have given slightly simplified presentations of the Mult-by-Price and TotMin-Rolle method 7 . The declaration slot of the method simply defines the meta-variables used in the body of the method. The premises, conclusions, and the constraint describe the applicability of the method. In the example of Mult-by-Price, for instance, line L 4 has to be present and to be an open subgoal, while L 1 and L 3 are lines that can be used in order to has to be given already, whereas L 3 is generated by the application of the method (indicated by the \Phi). Since the method is intended to prove L 4 , after the application of the method, this line can be deleted from the current planning state (we indicate this by the \Psi). In the constraint slot further applicability criteria are described, which cannot be formulated in terms of proof line schemata. Declarations, premises, constraints, and conclusions form the specification part of the method. In order to be able to mechanically adapt methods the tactic part is further subdivided into the declarative content and the procedural content. (However, this particular feature is not important for the purpose of this paper.) In our examples the procedural content 7 We have especially adjusted the syntax of the constraint in a way that is more comprehensive for the reader. INTEGRATING COMPUTER ALGEBRA INTO PROOF PLANNING 17 Declarations just f; Premises Constraint Conclusions \PsiL 4 Declarative Content (=subst L 3 Procedural Content schema \Gamma interpreter Figure 2. The Mult-by-Price method from theory cost. consists of a schema-interpreter, which essentially inserts the declarative content (using the bindings made in the planning phase) at the correct place in the current partial proof tree. In the concrete example the lines L 1 through L 4 are inserted (Note that we adopted a linearised version of ND proofs as introduced in [And80]). In order to understand to which piece of actual proof these methods evaluate, we have to examine the declarative content and the bindings performed in particular in the constraint. The constraint of the Mult-by-Price-method states a rather simple computation: if there is a cost function in the given open line which has a denomination other than DM, it is multiplied with the appropriate price. The multiplication of the real parts is carried out by the CAS and the corresponding cost function is constructed. As this point is crucial for understanding the working scheme of a method we will view the bindings in the constraint step by step: When applied to the current plan the method is matched with the open goals of the planning state. The first pass of the planner yields that L 4 can be matched with our theorem THM. Thus its formula Opt([cf(-d 0:5d is bound to the meta-variable OE. It is then examined to find an occurrence of a cost function. If such a subterm exists its arguments are bound to g; v; w and by matching line L 1 we receive the numerical part of price in f (if the appropriate price is not provided the application of the method would fail here). Afterwards the new cost function is com- M. KERBER, M. KOHLHASE, V. SORGE. puted (according to axiom Pr) in / 0 and finally OE 0 contains the result of replacing the old cost function in OE by / 0 . Hence in the first plan step the optimisation formula stored in OE 0 contains the cost function cf(-d prod ) as a subterm. With all these meta-variables instantiated the subproof contributed by the Mult-by-Price-method consists of lines L 2 and L 3 in the declarative content. Here we observe that L 2 results from applying the price-axiom Pr (which is fetched from the database) to line L 1 . Furthermore note that in L 3 we have a call to the CAS as a justifying method for the line. This means that at this point in the proof planning procedure, the CAS is called in order to compute the product of price and original cost function. The line resulting from this calculation is then used as the new open subgoal in the planning state. Summarising the effects of the method Mult-by-Price can be observed in two steps. First the goal line THM is justified with the method yielding the following subproof: Then the method in the justification of line THM (which has been abbreviated due to a lack of space) could be expanded thereby inserting the intermediate steps as described above by instantiating the macro steps of the method. Note that the following expanded subproof is at a more detailed level of abstraction in the PDS. In particular, the justification of THM itself is different at this level. cf(-d In the proof of THM, the method Mult-by-Price is applied twice in order to normalise both summands. To preserve space we will not present the next two methods of our proof plan as extensively as the Mult-by-Price-method. Add-by-Denom is very similar to Mult-by-Price and applies axiom CF1 inside the optimisation function Opt to compute the final cost function. In its course the CAS is called once to perform a polynomial addition. Then the Optimise-method simply introduces the definition for the Opt function of axiom O. INTEGRATING COMPUTER ALGEBRA INTO PROOF PLANNING 19 Declarations just Premises Constraint degree(OE) y / compute with CAS(minimum; OE) Conclusions \PsiL 12 Declarative Content (a Procedural Content schema \Gamma interpreter Figure 3. The TotMin-Rolle method from theory calculus. Far more interesting than these two methods is the TotMin-Rolle- method as it contains a different example for the use of a CAS in \Omega mega. Again the presentation of the method in Figure 3 is simplified. The TotMin-Rolle method is applied at a stage of the proof where the actual minimum of the cost function has to be introduced. This task is fulfilled within the constraint of the method. The compute with CAS statement actually calls the CAS in quiet mode to compute the minimum of the function OE and store it in the meta-variable y. At this stage, the CAS is used as an oracle here, just as in [HT93a]. In our example the minimum of the cost function is at and the ND-line of the form will be transformed by eliminating the existentially quantified variable: M. KERBER, M. KOHLHASE, V. SORGE. The rest of the proof plan is devoted to proving that the result is actually a total minimum. This is done by using the definition for TotMin from the database and furthermore by using the definitions for minimum and interval which correspond to line L 1 and L 2 in the method TotMin-Rolle. These definitions are introduced in lines L 9 through L 11 by applying them to the correct assertions given in lines L 3 through L 8 . This is expressed by the justifications in the corresponding lines; for instance, the justification of line L 10 states that we can infer y 2 [ff; fi] from the lines L 5 and L 6 with the definition of interval in line L 2 . A closer look at the justifications of lines L 3 through L 8 reveals that these contain methods themselves. Lines L 3 and L 4 again depend on calculations of the CAS which computes the first and second derivative of our cost function. The justifications Simplify correspond to a method performing basic arithmetic simplifications and comparisons. Consisting of only 5 methods the above proof plan gives the impression of a small proof and on an abstract level it is indeed; an experienced mathematician might not want to see more. But expanding the plan into a partially grounded ND proof gives it a length of 90 lines, containing lines justified by the CAS. The proof on this level may roughly correspond to a proof that a novice would like to see and that would form a reasonable solution of the exam problem once it is presented in natural language by the Proverb system. By rerunning the CAS in a proof plan generating mode on the CAS-justifications and extracting proof plans, the proof can be expanded to a more detailed proof plan containing an account of the mathematics behind the calculations. This proof plan already contains 135 plan steps and - if the user does not feel comfortable with the level of detail yet - can then be expanded to a calculus-level ND proof of length 354. Note that even this proof is not a stand-alone proof of the minimisation theorem, but depends on the proofs of a number of lemmata from a database. Furthermore, in these proofs the simplification of ground arithmetic expressions is not expanded, for instance, into a representation involving zero and the successor function either, which would be necessary to obtain a detailed logic-level proof. 4. Integrating Computations into Explicit Proofs In this section we describe Sapper (System for Algorithmic Proof Plan Extraction and Reasoning), which generates proof plans from CAS output. As mentioned in Section 3.3, for the intended integration KeKoSo.tex; 24/06/1998; 14:49; no v. INTEGRATING COMPUTER ALGEBRA INTO PROOF PLANNING 21 it is necessary to augment the CAS with mathematical information for a proof plan generating mode in order to achieve the proposed integration at the level of proofs. For the -Cas system, which we have developed to demonstrate the feasibility of the approach, this was rather simple, as we will demonstrate below. Enriching a state of the art CAS with such a mode for producing the necessary additional protocol information, would of course require a considerable amount of work. 4.1. Architecture The Sapper system can be seen as a generic interface for connecting \Omega mega (or another proof plan-based mechanised reasoning system) with one or several computer algebra systems (see Figure 4). An incorporated CAS is treated as a slave to\Omega mega which means that only the latter can call the first one and not vice versa. From the software engineering point of view,\Omega mega and the CAS are two independent processes while the interface is a process providing a bridge for com- munication. Its r-ole is to automate the broadcasting of messages by transforming output of one system into data that can be processed by the other 8 . Unlike other approaches (see [HC95, GPT96] for example) we do not want to change the logic inside our MRS. In the same line, we do not want to change the computational behaviour of the computer algebra algorithms. In order to achieve this goal the trace output of the algorithm is kept as short as possible. In fact most of the computations for constructing a proof plan is left to the interface. The proof plans can directly be imported into\Omega mega. This makes the integration independent of the particular systems, and indeed all the results below are independent of the CAS employed and make only some general assumptions about the MRS (such as being proof plan-based). Moreover, the interface approach helps us to keep the CAS free of any logical computation, for which such a system is not intended anyway. Finally, the interface minimises the required changes to an existing CAS, while maintaining the possibility of using the CAS stand-alone. The only requirement we make for integrating a particular CAS is that it has to produce enough protocol information so that a proof plan can be generated from this information. The proof plan in turn can be expanded by the MRS into a proof verifying the concrete computation. The interface itself can be roughly divided into two parts; the translation part and the plan generator. The first performs syntax translation between\Omega mega and a CAS in both directions while the latter 8 This is an adaptation of the general approach on combining systems in [CMP91]. 22 M. KERBER, M. KOHLHASE, V. SORGE. Hypotheses Theorems Methods Tactics Tactics Methods Theorems Hypotheses Interface Structured Database Theory 1 Theory 1.1 Theory 1.2 Theory 2 result access access plan tactics knowledge call result Plan Generator Translator Function Mappings call args. Abstract -CAS Figure 4. Interface between\Omega mega and computer algebra systems only transforms verbose output of the CAS to\Omega mega proof plans. Clearly only the translation part depends on the particular CAS that is invoked. For the translations a collection of data structures - called abstract CAS 9 - is provided each one referring to a particular connected CAS (or just parts of one). The main purpose of these structures is to specify function mappings, relating a particular function of\Omega mega to a corresponding CAS-function and the type of its arguments. Furthermore it provides functionality to convert the given arguments of the mapped \Omega mega function to CAS input. In the same fashion it transforms results of algebraic computations back into data that can be further processed by\Omega mega. The functionality in this part of our interface offers us the possibility of connecting any CAS as a black box system, as in the first approach we have described in Section 2. For instance, we may want to use a very efficient system without a mode for generating proof plans in 9 In a reimplementation of Sapper we would probably use the OpenMath protocol [AvLS96] as a lingua franca on the CAS side. INTEGRATING COMPUTER ALGEBRA INTO PROOF PLANNING 23 proof search as black box system, and then another less efficient system with such a mode for the actual proof construction, once it is clear what the proof should look like. This corresponds to recent techniques used in knowledge based systems, where the explanation component is not just a trace of the rules applied during the search, but the explanation is reconstructed by an independent component. The plan generator solely provides the machinery for our main goal, the proof plan extraction. Equipped with supplementary information on the proof by\Omega mega it records the output produced by the particular algebraic algorithm and converts it into a proof plan. Here the requirements of keeping the CAS side free of logical considerations and on the other hand of keeping the interface generic seem conflicting at the first glance. However, this conflict can be solved by giving both sides of the interface access to a database of mathematical facts formalising the mathematics behind the particular CAS algorithms. Conceptually, this database together with the mappings governing the access, provides the semantics of the integration of\Omega mega with a particular CAS. Thus expanding the plan generator is simply done by expanding the theory database by adding new tactics. While\Omega mega itself can access the complete database, Sapper's plan generator in the interface is only able to use tactics and lookup hypotheses of a theory (cf. Figure 4). The CAS does not interact with the database at all: it only has to know about it and references the logical objects (methods, tactics, theorems, or definitions) in the proof plan generating mode. Thus knowledge about the database is compiled a priori into the algebraic algorithms in order to document their calculations 4.2. Proof Plan Extraction Let us now take a closer look at the implementation of the proof plan generation in -Cas and at the expansion process of its output. This should demonstrate how proofs can be extracted from computer algebra calculation and provide an intuition on the requirements that our approach poses on the CAS side. As an example we will consider a polynomial addition from the example above. Normally, an experienced mathematician would not like to see any proof at all for that, while a high-school student would like to. As we have seen in our example, the main purpose of the Add-by-Demon-method is to compute the final cost function prod ). This is done in -Cas by adding the two polynomials -d d 3. In the remainder of this subsection we will expand this addition in several steps and thereby obtain a calculus level proof for the computation. M. KERBER, M. KOHLHASE, V. SORGE. Before examining this example in detail, let us consider the general scheme of the proof plan generation inside the polynomial addition algorithm of -Cas. We first take a look at the different representations of a polynomial p in the variables x r . The logical language of\Omega mega is a variant of the simply typed -calculus (indeed we use a stronger type system, but here we want to keep things as simple as possible), so the polynomials are represented as polynomial functions, that is, as -expression where the formal parameters are -abstracted (mathematically, p is a function of r argu- For the notation, we use a prefix notation; the symbols +, and " denote binary functions for addition, multiplication and exponentiation on the reals. In this representation, we can use fi-reduction for the evaluation of polynomials. In -Cas, we use a variable dense, expanded representation as an internal data structure for polynomials (as described in [Zip93], for instance). Thus every monomial is represented as a list containing its coefficient together with the exponents of each variable. Hence we get the following representation for p: Let us now turn to the actual -Cas algorithm for polynomial addition. This simple algorithm adds polynomials p and q by a case analysis on the exponents 10 with recursive calls to itself. So let r and r . We have presented the algorithm in the jth component of p and the kth component of q in a Lisp-like pseudo-code in Figure 5. Intuitively, the algorithm proceeds by ordering the monomials, advancing the leading monomial either of the first or the second arguments; in the case of equal exponents, the coefficients of the monomials are added. Obviously, the only expansions of the original algorithm needed for the proof plan generation are the additional (tactic.) statements 11 . We assume a lexicographic monomial order and employ it for ordering the exponents. Thus we make use of the operators ?, !, and = in an intuitive sense. Furthermore we can define the rank of a monomial as the vector given by its exponents and the rank of a polynomial as the maximum rank of its monomials with respect to the lexicographic monomial order. 11 Observe that in this case, the called tactics do not need any additional argu- ments, since our plan generator in the interface keeps track of the position in the KeKoSo.tex; 24/06/1998; 14:49; no v. INTEGRATING COMPUTER ALGEBRA INTO PROOF PLANNING 25 (poly-add (p q) (tactic "mono-add") (cons-poly er j r (poly-add r r (tactic "pop-first") (cons-poly ff j x er j r (poly-add r r (tactic "pop-second") (cons-poly r (poly-add r r ))) Figure 5. Polynomial addition in -Cas. They just produce the additional output by returning keywords of tactic names to the plan generator and do not have any side effects. In particular, the computational behaviour of the algorithm does not have to be changed at all. If we now apply this algorithm to the two polynomials we obtain the following proof plan: (mono-add, pop-second, mono-add) First the two quadratic monomials from p and q are added, then the linear term of q (the second argument) is raised, since it only appears in one argument, and finally the remaining monomials are added up. In the case of the polynomial addition, each of the methods (proof plan operators) directly corresponds to a tactic with the same name, that is, the list of the three methods above directly represents a concrete proof plan for polynomial addition of the concrete polynomials p and proof and thus knows on which monomials the algorithm works when returning a tactic. This way we need not to be concerned which form a monomial actually has during the course of the algorithm. 26 M. KERBER, M. KOHLHASE, V. SORGE. q. (In the following representation we omitted the context in which the polynomials are embedded in the actual proofs.) These four lines correspond to a step-by-step version of the basic High School algorithm. So far the expansion of the call-cas-method has been exclusively done by -Cas proof plan generation mode. But at this stage -Cas cannot provide us with any more details about the computation and the subsequent expansion of the next hierarchic level can be achieved without further use of a CAS. Let us for instance take a look at the pop-second tactic to understand its logical content. The tactic itself describes a reordering in a sum that looks in the general case as follows: For the current example we can view a and c as arbitrary polynomials and b as a monomial of rank greater than that of the polynomial a. It is now obvious that the behaviour of pop-second is determined by the pattern of the sum it is applied to. If in equation (1) the polynomial c does not exist, pop-second is equivalent to a single application of the law of commutativity. Otherwise, like in our example, the tactic performs a series of commutativity and associativity steps. The pop-second step above can thus be expanded in a plan which reflects the single step applications of the laws of commutativity and associativity Assuming we have expanded the two mono-add tactics as well, we have constructed a representation of the proof at a level where it only needs the axioms in the polynomial ring. To finally expand this to a fully explicit calculus level proof, we further expand all three justifications of the above lines. This leads to a sequence of eliminations of universally quantified variables in the corresponding hypothesis, the axioms of commutativity and associativity. In our example the commutativity axiom would be transformed in the following fashion: INTEGRATING COMPUTER ALGEBRA INTO PROOF PLANNING 27 Here, the justification (THM) in the first proof line indicates that the commutativity of was imported from the theory real in\Omega megas mathematical database, where it was established as a theorem. The remaining lines are natural deduction inferences: universal eliminations that instantiate a with the number 6 and b with the term \Gamma12x. Altogether this single application of the pop-second-tactic is equivalent to a calculus-level proof of 11 inference steps. The length of the subproof for this trivial polynomial addition is 43 single steps. This example shows how it is possible to mechanically construct a proof verifying the correctness of any particular CAS computation without verifying the CAS algorithm (or their implementation) in the general case. However, the calculus level proofs for the computations are very long and rather boring and therefore hardly any human user might actually want to see much less read them. Therefore the Proverb proof explanation system in\Omega mega provides a more realistic alternative, since it gives the user access to representations of the parts of the proof on various levels of abstractions making use of the hierarchical structure of the underlying PDS. For instance, it is then possible to present the computations with some intermediate steps, as it is customary in textbooks. For example, we could include the three steps of the High School algorithm mentioned above, to illustrate the polynomial addi- tion. (The decision which steps should be included and which omitted, depends of course on the expertise of readers for which a particular proof presentation is intended.) Despite all these abstractions in both developing and presenting the proof, we can still use any proof checker for ND-calculus to verify all steps including computations. Furthermore, if we assume we have a more sophisticated proof checker, for example one that works modulo the axioms of polynomial rings, it is also possible to check the proof on an abstract level. As already mentioned, the more sophisticated the proof checker is, the more concise the communicated proofs can be. We have tested proof plan extraction from simple recursive and iterative CAS algorithms, where it works quite well, since these algorithms closely correspond to the mathematical definitions of the corresponding concepts. However, more complicated schemes like divide-and-conquer algorithms (for instance, the polynomial multiplication of Karatsuba and Ofman [KO63]) cannot be adapted to our approach so easily without extending the mathematical knowledge base by corresponding lemmata 28 M. KERBER, M. KOHLHASE, V. SORGE. The example of the polynomial addition is surely a trivial one, we have chosen it solely for presentation reasons. In particular it is very likely to be correct in any real-world implementation, since it is well tested and does not depend on sophisticated mathematical theorems for which fuzzy boundary cases must be considered. For the sake of argument, let us assume an error in the implementation, for instance, in the second case of the polynomial addition algorithm in Figure 5 the cons-poly statement was forgotten, so that the algorithm has the following (incorrect) form (tactic "pop-first") (poly-add r In the computation of ((x 2 that we have discussed above, the second case is never used, and the computation would be correct although the program is not. If we now change the order of addition of our polynomials p and q to we get the following incorrect result from the changed algorithm: Inserting the proof plan generated by the faulty algorithm then yields In checking, the proof checker would see that the pop-first step is not justified, since the expansion corresponds to the application of the law of associativity. This would yield and thus would not be applicable during the expansion. Thus the proof plan and consequently the calculation would be rejected by\Omega mega. Note that in a large system with literally millions of possible cases, the correctness of a calculation like depends only on a tiny subset of the whole program. It is a strength of our approach, that only the calculations that are necessary for a given proof would be checked. This has the advantage that errors on different levels can be detected (in particular, on the levels of algorithms, of compilers, and of processors). Of course, for very long computations checking can be pretty expensive. Moreover, highly elaborated and efficient algorithms in state of the art CAS might be hard to augment with proof plan KeKoSo.tex; 24/06/1998; 14:49; no v. INTEGRATING COMPUTER ALGEBRA INTO PROOF PLANNING 29 generation modes. As we have seen in the example above, the mathematical knowledge in the database has to reflect the mathematical knowledge in the algorithm in order to easily decorate the algorithms by a proof plan generation mode. However, to extend and prove corresponding lemmata is not a trivial task for sophisticated algorithms. In particular such an approach would go very much in the direction of program verification. Even if it proves practically impossible to extract the information that is valuable at the conceptual, mathematical level, it is always possible to reserve these elaborated techniques for the quiet mode used in proof discovery, and use more basic algorithms, for which the mathematics is easier and that are more easily decorated by a proof plan generation mode, for the proof extraction phase. Systems like Axiom [JS92] or MuPAD [Fuc96] seem to come closest among standard CAS to the needs for a proof plan generation, since one can already attach axiomatisations to algorithms. 5. Conclusion In this work we have reported on an experiment of integrating a computer algebra system into the interactive proof development environ- ment\Omega mega, not only at the system level, but also at the level of proofs. The motivation for such an integration is the need for support of a human user when his/her proofs contain non-trivial computations. We have shown that the proof planning paradigm in general and the \Omega mega system in particular provide an open environment for such an extended integration that supports different integration levels. In our approach it is not possible to use a standard CAS for the integration as it is, since such a system provides answers, but no directly usable justifications from which proof plans can be extracted. This, however, turned out to be essential in an environment that is built to construct communicable and checkable proofs. In order to achieve a solution that is compatible with such a strong requirement, we have adopted a generic approach, where the only requirement for the CAS is that it has a proof plan generation mode for the generation of communicable and checkable proofs. Since we want to achieve the two goals simultaneously, namely to have high-level descriptions of the calculations of the CAS for communicating them to human users as well as low-level ones for mechanical checking, we represent the protocol information in form of high-level hierarchical proof plans, which can be expanded to the desired detail. Fully expanded proof plans correspond to natural deduction proofs which can be mechani- M. KERBER, M. KOHLHASE, V. SORGE. cally checked by a simple proof checker. In the case that the CAS has made a mistake the proof checker will detect it. The general idea and the fundamentals of the integration of a CAS into an MRS are independent from the concrete proof development environment\Omega mega and the concrete computer algebra system -Cas. It can be realised in any plan-based theorem prover. Proof extraction can even be realised on any tactic-based system and with any CAS that can protocol its calculations in form of tactics. Axiom [JS92] and MuPAD [Fuc96] seem to be best suited for a corresponding extension since one can already attach axiomatisations to algorithms. If in addition the algorithms could be enriched in a way that they produce protocol information in every computation step, that is, state which of the attached axioms are used and what the particular instantiations are, the system would probably fit in with our approach pretty well. A useful extension of our approach would consist in the usage of various algorithms for the same computation, for instance, one as a fast and efficient algorithm that is not suitable for knowledge extraction while searching for a proof. Afterwards, when actually documenting the whole proof a less efficient algorithm, which is optimised to find short proofs, can provide a complete proof plan. Although the correctness issue can be achieved by a tactic-based approach as well and does not need the specifications that are used in proof planning, the full strength of an integration where considerable automated support is provided cannot be achieved on this level, since it is not possible to perform mechanical reasoning about the tactics. Such an automation can, however, be achieved by the proof planning approach, where the proof planner can automatically call a CAS pro- cedure, when the conditions in the corresponding method are met. The usefulness of an integration on this level can already be seen in the case of our simple -Cas: After the integration we are able to prove optimisation problems which were out of reach without such a support. On the other hand, the system is able to give explanations of the involved computations at various levels of abstraction. A feature that is missing from todays CAS. From our experiments we expect that the successful integration of any powerful computer algebra systems would considerably enhance the reasoning power of any mechanised reasoning system. Acknowledgements The work presented in this paper was supported by the "Deutsche Forschungsgemeinschaft" in SFB 378, project OMEGA. It benefited a lot from discussions in the Calculemus interest group. INTEGRATING COMPUTER ALGEBRA INTO PROOF PLANNING 31 The authors would like to thank Lassaad Cheikhrouhou for his help with\Omega mega's proof planner and with coding the methods for our examples. Furthermore, we would like to thank Deepak Kapur and the anonymous referees for carefully reading earlier versions of the paper and for their detailed comments that helped us to improve the presentation considerably. --R Transforming matings into natural deduction proofs. Objectives of OpenMath. Technical Report 12 Computations and formal proofs in type theory. Theorems and algorithms: An interface between Isabelle and Maple. Rippling: A heuristic for guiding inductive proofs. Symbolic Computation (An Editorial). Using Mathematica for Doing Simple Mathematical Proofs. "Intelligenzniveaus" The use of explicit plans to guide inductive proofs. Machine Proofs in Geome- try: Automated Production of Readable Proofs for Geometry Theorems Mechanical geometry theorem proving. Integrated software compo- nents: A paradigm for control integration Implementing Mathematics with the Nuprl Proof Development System. Reasoning theories - towards an architecture for open mechanized reasoning systems An open environment for doing mathe- matics Reduce user's manual: Version 3.6. Computational metatheory in Nuprl. Adapting methods to novel tasks in proof planning. Presenting machine-found proofs Extending the HOL theorem prover with a computer algebra system to reason about the reals. Reasoning about the reals: The marriage of HOL and Maple. AXIOM: The Scientific Computation System. A refutational approach to theorem proving in geometry. Otter 3.0 reference manual and guide. The Mathematica Book: Version 3.0. Mechanical Theorem Proving in Geometries: Basic Principles. Texts and monographs in symbolic computation. Effective Polynomial Computation. Address for correspondence: Manfred Kerber School of Computer Science The University of Birmingham Birmingham B15 2TT --TR --CTR Andreas Meier , Martin Pollet , Volker Sorge, Comparing approaches to the exploration of the domain of residue classes, Journal of Symbolic Computation, v.34 n.4, p.287-306, October 2002 Paul Cairns , Jeremy Gow, Integrating Searching and Authoring in Mizar, Journal of Automated Reasoning, v.39 n.2, p.141-160, August 2007 Micheal Kohlhase , Andreas Franke, MBase: representing knowledge and context for the intergration of mathematical software systems, Journal of Symbolic Computation, v.32 n.4, p.365-402, October 2001
computer algebra;proof checking;hierarchical proof planning;mechanized reasoning
594248
Ordered Semantic Hyper-Linking.
The ordered semantic hyper-linking strategy is complete for first-order logic and accepts a user-specified natural semantics that guides the search for a proof. Any semantics in which the meanings of the function and predicate symbols are computable on ground terms may be used. This instance-based strategy is efficient on near-propositional problems, is goal sensitive, and has an extension to equality and term rewriting. However, it sometimes has difficulty generating large terms. We compare this strategy with some others that use semantic information, and present a proof of soundness and completeness. We also give some theoretical results about the search efficiency of the strategy. Some examples illustrate the performance of the strategy.
Introduction There are at least two basic approaches to the study of automated deduction. One approach concentrates on solving hard problems, especially those of interest to human mathematicians. There have been some notable successes in this area, and even some proofs of hitherto unproven conjectures. This has served to give the field of automated deduction some respectability among mathematicians and the general public. Such proofs may be done with or without human interaction. For this approach, it is of secondary interest whether the prover that solves these hard problems is very weak for other, easy problems. Another approach to automated deduction concentrates on building provers that perform well on a broad range of problems, with a minimum of human interaction. In this approach, it does not make sense to try hard problems if one's prover still has difficulty with easy problems. The philosophy is to develop as far as possible theorem provers that have general problem solving ability. The advantage of this approach is that one is likely to obtain provers that are more well-rounded, and in the long run possibly more powerful; a disadvantage is that the results may appear less spectacular, especially in the early stages. We have concentrated on the latter approach. In this endeavor, we have developed a number of provers over the recent years, including the modified problem reduction format of [Pla88] and its extensions, the clause linking method of [LP92], and clause linking with semantics [CP94a], among others. We have emphasized first-order logic without equality. These provers have become increasingly more powerful, each able to solve a considerable range of problems out of reach of its pre- decessor, with little or no human guidance. In addition, these provers can compete respectably with well-known powerful theorem provers on certain types of problems. For this, we have emphasized Prolog implementation, as a way to rapidly implement and test a wide variety of ideas with little manpower. This means that our provers have a disadvantage with respect to provers carefully implemented in C or LISP, since the underlying language is less efficient; despite this, the performance has been impressive. We now would like to carry this investigation a step further, and incorporate some kind of ordering methods. hyper-linking [CP94a] was developed to retain the propositional advantages of hyper-linking [LP92] while adding natural semantics and goal-sensitivity [Pla94a]. We discuss the issues of propositional efficiency and goal-sensitivity in [Pla94a, Pla94b], thereby highlighting what we feel are some inefficiencies in many common theorem proving strategies. Even hyper-linking without semantics performs much better than resolution and similar strategies on some hard problems, particularly non-Horn problems. We show in [CP94a] that hyper-linking with semantics is sometimes much better still. However, there are still some problem with semantic hyper-linking that we would like to solve. The cooperation between semantic hyper-linking and rough resolution does not seem as clean as it could be. Rough resolution [CP93b] is a version of ordered resolution developed to eliminate large literals from proofs; this is helpful because semantic hyper-linking has difficulties generating such large literals. Therefore the cooperation of these two methods seems attractive, and indeed, we show in [CP93b] that this cooperation improves the effectiveness of semantic hyper-linking on a number of problems. However, the definition of rough resolution seems arbitrary; the logical way to eliminate large literals is to use ordered resolution, as described in [BG90, HR91]. Also, the manner in which the semantic tree is constructed and searched seems to have an arbitrary element to it; this also makes this part of the method harder to describe formally. In addition to semantic hyper-linking and rough resolution, UR (unit resulting) resolution is a component of the prover described in [CP94a]. The motivation for this is that rough resolution eliminates large literals from proofs, UR resolution eliminates the Horn parts of proofs, and what remains is a non-Horn proof with small literals; clause linking performs well on such problems. However, the Horn property is really irrelevant here, since clause linking performs well on problems with small literals, whether they are Horn problems or not. Therefore it seems logical to eliminate UR resolution. Also, the choice of which rough resolutions and UR-resolutions are performed seems to be arbitrary, to some extent; we prefer small clauses and small proofs, essentially. Our motivation in the development of ordered semantic hyperlinking is to simplify semantic hyper-linking as much as possible, and in this way hopefully to extend its peaks of strength to a wider class of problems. This should also allow for a small and easily understood implementation. In addition, we are interested in removing some of the propositional inefficiencies from term-rewriting based theorem proving strategies. Such strategies essentially reduce to A-ordering [Sla67] on first-order logic without equality. However, in [Pla94b] we show some simple sets of clauses on which A-ordering produces an exponential search space, regardless of the ordering chosen. On the other hand, term-rewriting methods are often very efficient, and extend naturally to other specialized theories [BG94]. Furthermore, certain sets of clauses are easily decided by strategies based on ordering [FLTZ93]. That is, these ordering based strategies are a decision procedure, always terminating and indicating whether the set of input clauses is satisfiable or not. However, these sublanguages of first-order logic are not decidable in this way by clause linking. Therefore, we would like to present semantic clause linking in a format that facilitates the transition to term-rewriting based theorem proving strategies, thereby also providing a way to remove some of their notable inefficiencies, while preserving some of their advantages. The idea of semantic hyper-linking is to show that a set S of clauses is unsatisfiable by the failure of a systematic search for a model of S. Ordered semantic hyper-linking is similar, but it organizes the search a little differently. The basic principle behind ordered semantic hyper-linking is the following: Suppose we have a set of independent choices to make in a process of examining a set of possibilities. Thus there are potentially 2 n combinations of choices altogether. Then we assume that there is an ordering on these choices, and we make the simplest choice first. Suppose p 1 is the simplest of the choices; then we first consider the alternatives p 1 and not(p 1 ). For each such alternative, we recursively attempt to solve the problem. The reasoning is that we may solve the problem before the more complex choices are even seen, thereby saving effort. This seems to be a natural strategy from the standpoint of human problem solving. In the application to theorem proving, the set of choices is infinite, and the order in which they are made has other implications, but the idea is still the same. Furthermore, a problem with semantic hyper-linking is that sometimes the enumeration of ground terms is necessary. We would like to have a method that is based on unification instead of on the enumeration of ground terms. The proposed method incorporates unification in a natural way, and for certain kinds of semantics we show that the enumeration phase can be done in polynomial time. There is an additional reason to believe that this new version will have better performance. The work required by semantic hyper-linking is strongly influenced by the number of "eligible literals" that are generated. The proposed method should reduce this number, thereby making the method more efficient and permitting proofs that require more rounds of search. We also consider some complexity issues, as discussed in part by [Gou94]. His approach is non-clausal, but the same analysis applies to a clausal framework. He shows that the fundamental problem associated with the mating [And81] or connection [Bib87] approach is \Sigma p 2 complete. These approaches first choose a number of copies of each of the input clauses, and then seek to find a single substitution that makes the given set of copies of the clauses propositionally unsatisfiable. (For a general set of clauses, an arbitrary number of copies may be needed.) Goubault's result has the consequence that the effort to prove a theorem is at worst exponential in the size of a minimal proof, using a suitable implementation of these approaches, where proof size is measured by the number of instances of the input clauses that appear in the proof. More precisely, if S is the set of input clauses and each clause appears at most n times in the proof, then the effort is exponential in n times the size of S, written as a character string. This is actually not a bad bound; many other methods are considerably worse, at least relative to this measure of proof size. For example, clause linking can be double or even triple exponential in this measure of proof complexity. The reason is that a proof containing n copies of the input clauses may involve a number of unifications proportional to n; each such unification can increase the size of the terms by a constant factor. Thus terms that are exponentially large can be generated, the number of terms within this exponential size bound is double exponential, and the time to handle a double exponential set of ground clauses can be triple exponential. By altering the measure of term size, so that repeated subterms are counted only once, this can be reduced to double exponential. We consider the behavior of ordered clause linking and argue that although the worst-case bound is double exponential, there is reason to believe that in many cases this performance will be single exponential or even better. Of course, there may be proofs in which many copies of the clauses are needed but the term sizes are all small; for such clauses, clause linking would probably be faster than the mating approach. It is interesting that our proposed theorem prover incorporates a number of well-known AI techniques, including case analysis, explanation-based generaliza- tion, procedural semantics (to describe the semantics of the set of clauses), back- tracking, ordering criteria, and of course unification and first-order logic. Therefore this prover may have some independent interest from the standpoint of artificial intelligence. 2 Orderings on Interpretations and Clauses The idea of ordered clause linking with semantics is to place a total ordering on the set of atoms, and then based on this to define a lexicographic total ordering on the set of interpretations of these atoms. Now, semantic hyper-linking can be seen as a systematic search for a model of a set S of clauses; if this search fails, then we know that S is unsatisfiable. During this search, a semantic tree is essentially constructed. Ordered semantic hyper-linking works in a similar way. In fact, both strategies are somewhat similar to model elimination in this respect [Lov69]; this similarity may be more apparent for ordered semantic hyper-linking than for semantic hyper-linking. However, instead of a semantic tree, we have a transfinite semantic tree, as in [HR91]. Also, in ordered semantic hyper-linking we specify more precisely how this tree is constructed. That is, the interpretations are examined in a sequence I consistent with the total ordering; the first interpretation I 0 to be considered is the one that is least in this ordering. For this interpretation we find a clause C 0 not satisfied by I 0 ; this clause is a minimal such contradicting clause in a specified ordering on clauses. The next interpretation I 1 considered is the smallest interpretation that is not "obviously" contradicted by the clauses found so far (in this case, C 0 ). The search continues in this manner, so we have I only is the sequence in which the interpretations are examined completely determined in this way, but the choice of which clauses C i are found is also largely determined (up to minimality). If S is unsatisfiable, eventually the set fC will be unsatisfiable; this will be detected by the prover, and the search will stop. We now define these orderings on atoms and interpretations more precisely. We assume that some first-order language is specified in terms of a finite set F of function and constant symbols, a finite set P of predicate symbols, and a list X of variables. Then we are interested in the satisfiability problem of sets S of first-order clauses over this language. Let T be the set of terms formed from the function symbols F and the variables X , and let T (F) be the set of ground terms (terms without variables) formed from T and F . Let A be the set of atoms, that is, expressions of the form P For orderings !, we define x ? y as equivalent to y ! x. A partial ordering is said to be well-founded if there are no infinite sequences x We assume that there is a total well-founded ordering ! on A. If the ordering is order-isomorphic to !, then A can be enumerated as A 1 this will not be possible if the ordering corresponds to a higher ordinal. The ordering on A may be based on the size (number of symbol occurrences) in the atoms A i , or it may be one of the orderings used to show the termination of term-rewriting systems [DJ90, Pla93]. Of course, there are also other possibilities. We now specify the ordering on interpretations. A literal is an atom or an atom preceded by a negation sign :. A literal without a negation sign is called positive and one with it is called negative. The literals L and :L are called complementary. We write L for the complement of L; thus if L is positive then L is :L, and :L is L. If A 2 A, then we call A and :A literals over A. An interpretation I is (for our purposes) a subset of A, or, equivalently, a mapping from A to ftrue; falseg. If I maps A to true then we write I j= A and say that I satisfies A. Otherwise, I 6j= A and we say that A contradicts I. We say that I j= :L iff I 6j= L. We say that two interpretations I and J agree on a literal L if (I assign L true or they both assign L false. If I and J are two distinct interpretations of A, let d(I; J) be the least atom A (with respect to !) such that I and J do not agree on A. Such a least atom must exist, because the ordering on atoms is well- founded. Let I 0 be a special interpretation called the initial interpretation; this is typically supplied by the user. We order the interpretations with respect to I 0 as follows: Let I and J be two different interpretations of A. Let A be d(I; J). Then if I agrees with I 0 on A, we say I I. Thus the smallest literals have the greatest influence on the ordering, and the interpretations that agree with I 0 are smaller in the ordering than interpretations that disagree with I 0 , other things being equal. Note that I 0 is the minimal interpretation in this ordering. This ordering on interpretations is not well-founded, if A is infinite. Still, it turns out that certain sets of interpretations of interest to us have minimal elements. We will mainly be interested in interpretations that differ from I 0 in finitely many places; we now develop some of their properties. Definition 2.1 Given an interpretation I of A and literals L i over A, let I[L be defined as follows: (a) Lng or if (b) L 62 fL Lng and I Thus like I except for the finite list [L ceptions;" for an earlier use of this notation see [CP94a]. Theorem 2.2 Suppose I is an interpretation over A and Proof. If I ! J , then I and J are unequal; thus there must be a literal d(I; J). If agrees with I 0 on the atom d(I; J), so J ! I. 2 This result classifies the interpretations that are smaller than I into n distinct groups, depending on which L i is equal to d(I; J). We now specify the ordering on literals and clauses. We order literals so that if A and B are atoms and A ! B then A ! B (as literals), A ! :B, :A ! B, and :A ! :B. However, the literals A and :A are not ordered with respect to one another. A clause is a finite set of literals, regarded as their disjunction. A clause is a tautology if it contains a literal and its negation. A clause is a ground clause if it contains no variables (and similarly for literals, atoms, and terms). We define a partial ordering ! on non-tautologous clauses by the multiset extension of the ordering on literals: If C is the empty clause (which contains no literals), then C ! D for every non-empty clause D. Also, if C and D are clauses, then let and M be their maximum respective literals, which exist if C and D are non- tautologous. Then if fMg. If L and M are complementary, then C and D are not ordered. We note that this multiset ordering on clauses is well-founded, because the underlying ordering on literals is. We will not further use this ordering on clauses, and present it here mainly to make this point clear. An interpretation I satisfies a ground clause C if I satisfies some (that is, at least one) literal in C, and in this case we write I Otherwise, we say that C contradicts I and write I 6j= C. A substitution is a mapping from variables to terms. If \Theta is a substitution and C is a clause then C \Theta represents the clause obtained by replacing variables of C by terms, as specified by \Theta. Such a clause C \Theta is called an instance of C. A similar terminology applies to terms, atoms, and literals, so we can apply substitutions to them, for example. For our purposes, we say that an interpretation satisfies a non-ground clause if it satisfies all of its ground instances. An interpretation I satisfies a set S of clauses if it satisfies every clause C in S. An interpretation I is a model of a ground clause C if I I is a model of a set S of ground clauses if I j= C for all C 2 S. Least models of single ground clauses and finite sets of ground clauses exist, as we now show. Theorem 2.3 Suppose Lng is a ground clause, and suppose the least model of C is I 0 [Ln ]. Proof. We note that I 0 [L n ] is a model of C. We show that no smaller interpretation J is a model of C. Suppose J ! I 0 [L n ]. Then by theorem 2.2, d(I; This implies that J agrees with I 0 on Ln , and so C contradicts J . 2 Theorem 2.4 Let G be a finite set of ground clauses. If G is satisfiable, then G has a least model. Proof. Let A An be the atoms that appear (positively or negatively) in clauses in G. There are only finitely many interpretations of these atoms; at least one of them, say I, is a model, since G is satisfiable. Let I min be the least such model, in our ordering on interpretations. This must exist, since there are only finitely many such finite interpretations. Extend I min to an interpretation I of A where each L i is either A i or :A i and I min i. Then J is a model of G, since I min is, and it is a least model, because I min is as small as possible among the interpretations of the A i and the interpretations of the other literals have been chosen as small as possible. Any smaller interpretation would have to differ from J on some literal L i , by theorem 2.2, which is not possible by the way I min was chosen. 2 Definition 2.5 If G is a set of ground clauses over A, let lm(G) be its least model, that is, the smallest interpretation I in the ordering on interpretations such that I satisfies all elements of G. In ordered clause linking with semantics, we accumulate such a set G and in the process keep track of its least model lm(G). However, we do not store G in its original form, but apply certain simplifications to it so that only the features relevant for the current minimal model are retained. For this, we not only make use of least models but also least contradicting clauses of interpretations. However, this necessitates the introduction of another ordering on clauses. We would have liked to define ordered semantic hyper-linking entirely in terms of the ordering ! on clauses, but it turns out that this is incomplete if the ordering ! has order type greater than !. Therefore we introduce another ordering ! cl on clauses and use it to choose contradicting instances. This means that for purposes of completeness, ordered semantic hyper-linking may need to use two different orderings, which seems to be somewhat remarkable. For this purpose, we assume that ! cl is a partial ordering such that for all ground clauses C, all but finitely many ground clauses D satisfy cl D. An example of such an ordering is to order the clauses by size, that is, the total number of occurrences of symbols, with two clauses unordered if they have the same number of symbols. Since there are only finitely many non-variable symbols that can appear in a clause, this ordering satisfies the finiteness property given above. We note that ! cl is well-founded. We say a clause C is minimal in a set of clauses if there is no clause D in the set such that D ! cl C; note that a set of clauses can have more than one, but at most finitely many, minimal elements. We now show that minimal contradicting clauses always exist; that is, if I is not a model of a set S of (non-ground) clauses, then there is a minimal ground instance D of a clause C of S such that D contradicts I. Theorem 2.6 Suppose S is a set of (possibly) non-ground clauses over A and I is an interpretation of A that is not a model of S. Then there is a ground instance D of some clause C of S such that for all other ground instances D 0 of clauses in S, if I 6j= D 0 then either D ! cl D 0 or D and D 0 are unrelated by ! cl . Proof. The set of such ground clauses D 0 such that I 6j= D 0 is non-empty, since I 6j= S. Since the ordering ! cl on ground clauses is well-founded, this set of ground clauses has a minimal element D, as claimed. Note that there may be more than one such minimal D, but at most only finitely many, by the way ! cl is defined. This result holds even if S is infinite, by the way. 2 Definition 2.7 Let mi(S; I) be some such clause D, that is, a ground instance D of some clause in S such that I 6j= D and such that there is no other ground instance C of a clause in S such that C ! cl D and such that I 6j= C. If there is more than one such instance D, we assume that mi(S; I) is one of them, chosen in an arbitrary manner. We call such a clause D a minimal contradicting instance for I. Later we will discuss algorithmic aspects of computing mi(S; I). 3 The Search Procedure The task now is to devise a procedure that will search through the set of all interpretations in a manner consistent with the ordering, finding clauses that contradict each interpretation. For this, we maintain a finite list C of "relevant" ground clauses that record the progress made in the search so far; this list contains instances of clauses from S as well as instances of clauses derived from S by A-ordering resolu- tion. The goal of the search is to continually increase the least model lm(C) of C, that is, to make lm(C) larger and larger in the ordering on interpretations. This least model is called the current interpretation. At the beginning C is empty and lm(C) is I 0 . As elements are added to C, and sometimes removed, this least model becomes larger and larger in our ordering on interpretations. If S is unsatisfiable, then eventually C will contradict all interpretations, and there will not be a least model any more. At this point, the empty clause will be in C and the search will stop. The invariant that C possesses is captured by the following definitions. Definition 3.1 If C is a (non-tautologous) clause, let max(C) be the maximal literal in C in the ordering !. This exists because clauses are finite. Definition 3.2 A list C of clauses is ascending if it is of the form C contradicts the least model of C 1 , and so on. The literals are called eligible literals, in harmony with the use of this term in [CP94a]. Theorem 3.3 If then so are all its prefixes. Also, Furthermore, this latter interpretation disagrees with I 0 on the literals max(C j ). Proof. The part about prefixes is immediate. Let I be I 0 [max(C 1 max(Cn )]. For the rest, we use induction on i. We show that I 0 6j= i. By induction, lm(fC Since the literal max(C i ) is larger than any literal [max(C 1 agrees with I 0 on max(C i ). By the definition of ascending, lm(fC To show that I and I 0 disagree on the literals max(C i ), we have just shown that I 0 6j= max(C i ) for all i, but I by the way it is constructed. To show that I is lm(C), we show that I is a model of C but no smaller interpretation J is a model of C. Now, I is a model of C since it satisfies all the literals I then J must differ from I on some eligible literal, by theorem 2.2. Suppose d(I; J) is max(C I and J differ on max(C i )). It remains to show that J does not satisfy the other literals of C i . We know that lm(fC by the definition of ascending. By induction, we know that lm(fC does not satisfy C i . Therefore J does not satisfy C i , since J agrees with I 0 [max(C 1 on all literals smaller than max(C i ), and thus the literals of C i other than are not satisfied by J either. 2 Now, when the search procedure begins, the set C is empty, and this set is then successively modified by adding to it some clause instance D contradicting its least model. Whenever this is done, it is necessary to do some processing on C to preserve the ascending property. This processing involves performing certain resolutions, as well as deleting certain elements from C, in a manner that will be described. Letting simp(C; D) denote the result of this processing, we have the following overall algorithm for ordered semantic hyper-linking: while fg 62 C do if lm(C) else D / mi(S; lm(C)); return "unsatisfiable" 3.1 Processing the list C of relevant clauses There are two kinds of operations that take place during the processing of C involved in the call to "simp." The first kind is to perform ordered resolutions between D and the last clause of C, when possible. For this, we define a res(C,D) as follows: Definition 3.4 Suppose C and D are ground clauses and suppose there is a literal L such that Now, a res(C,D) will be an (A-ordering) resolution involving the maximal literals in C and D. We note that A-ordering resolution is in itself a complete theorem proving method for propositional logic, and has natural extensions to first-order logic. The second kind of operation that takes place during "simp" is to eliminate elements from C that are made irrelevant by these A-ordering resolutions, that is, elements of C that can be eliminated without affecting lm(C). We have the following procedure for "simp": procedure if max(Cn ) and max(D) are complementary then return else if max(Cn ) ? max(D) then return else return [C To show correctness, we define an interpretation I (C; D) where C is an ascending list of clauses and D is a clause contradicting lm(C). This interpretation will have the property that lm(C) ! I (C; D) and if I I (C; D) then I 6j= C [ D. Therefore, if C [ D is satisfiable, then I (C; D) ! lm(C [ D). Thus this definition gives a lower bound on the least model (if it exists). We show that each processing step in "simp" does not decrease I (C; D). And, at the beginning, I (C; D) is larger than lm(C). Therefore, at the end we will obtain an ascending set of clauses whose least model is larger than it was at the beginning. Or, if I (C; D) is the maximal in- terpretation, that is, the interpretation that disagrees with I 0 everywhere, then this property is preserved, and we will eventually derive the empty clause fg. I (C; D) can be the maximal interpretation only if C [D is unsatisfiable. Definition 3.5 Suppose C is an ascending list of clauses. Let Cn be the last clause in the list C. Suppose D is a ground clause contradicting lm(C). Define I (C; D) as follows: I (C; D) 1. L ? max(D) and I 0 6j= L or 2. L max(D) and lm(C) We note that since I (C; D) imitates lm(C) for literals less than or equal to max(D), and max(D) contradicts lm(C), therefore I (C; D) does not satisfy D. Also, for literals larger than max(D), I (C; D) is chosen to be as large as possible in the ordering on interpretations, and for smaller literals, I (C; D) agrees with lm(C). Therefore, I (C; D) ? lm(C). (We cannot have equality because I (C; D) differs from I 0 on infinitely many literals, assuming that A is infinite.) Also, if J is an interpretation and J ! I (C; D) then J does not satisfy C [ D. If J ! lm(C) this is immediate. If contradicts J . If J ? lm(C) then d(J; lm(C)) ? max(D) since I (C; D) and lm(C) agree on literals not larger than max(D). Therefore D contradicts J in this case, too. Thus lm(C [ D) ? I (C; D). And if there is no least model for C [D, then we can at least say that all interpretations J less than or equal to I (C; D) fail to satisfy C [ D. To show correctness, we need to show that every processing step in "simp" preserves the property that D contradicts lm(C) and does not decrease I (C; D). Let Cn be the last element of C. If max(D) ? max(Cn ) then the list C with D added to the end, is ascending, because D contradicts lm(C). If max(D) then Cn does not enter into the definition of I (C; D) at all, so deleting Cn does not affect I (C; D). Also, D still contradicts lm(fC since this cannot be distinguished from lm(C) on literals smaller than max(Cn ). If max(D) and max(Cn ) are complementary (the only remaining case), then the A-ordering resolution replaces D by a resolvent containing all the literals of D and Cn except the two maximal ones. This resolvent will still contradict the least model of since Cn did (by the definition of ascending) and all the literals of D except max(D) also contradicted this least model (since D contradicted lm(C)). Also, I (fC because the maximal literal of a res(Cn ; D) is smaller than the maximal literal of D. We now consider the case in which C [ D is unsatisfiable. If at some stage in this processing, I (C; D) becomes the maximal interpretation, then no ascending list can be produced, since all ascending lists of clauses are satisfiable. This means that the empty clause fg is derived, and the search terminates. 3.2 An example Suppose that the set A of atoms is and that S contains the following clauses: Suppose I 0 interprets P i as true for all i, that is, I 0 the following sequence of current interpretations I and the corresponding clauses mi(S; I i I 0 (now the clause :P 5 ; :P 8 is chosen) I (resolvent , from the above two clauses) I I I is generated from (resolvent P 6 is generated from P 10 and P 6 (resolvent :P 5 is generated from P 6 and :P 5 I (resolvent fg is generated from :P 5 and P 5 We show the five ascending lists of clauses that are generated, too: (after "simp" is called) 3.3 Completeness We can argue the completeness of this method in a manner similar to the completeness proof in [PACL92]. However, the fact that our partial orderings may have order type greater than ! complicates the argument a little. Also, the fact that clauses D that are chosen at one point may be discarded later complicates the proof. The general idea is to show that if S is unsatisfiable then there are only a finite number of ground instances that can ever be chosen as minimal contradicting instances. Therefore there are only a finite number of atoms that ever appear in C. This essentially reduces the problem to one involving finite interpretations. Since there are only finitely many finite interpretations, and by the ordering on interpretations, the same one cannot be seen more than once, eventually the method must stop; if S is unsatisfiable, then the only way for the method to stop is to generate the empty clause. We now show that only a finite number of clauses can be chosen. Theorem 3.6 Suppose S is an unsatisfiable set of clauses. Then the set fmi(S; I) I is an interpretation over Ag is finite. Proof. Let T be a finite unsatisfiable set of ground instances of clauses in S; such a set T exists, by the so-called Herbrand's theorem. Then for every interpretation I there is a clause D in T such that I 6j= T . Note that mi(S; I) is minimal in ! cl among clauses contradicting I; thus we cannot have D ! mi(S; I). However, for all but finitely many ground clauses C, which is finite. 2 Definition 3.7 Suppose we arbitrarily choose some finite unsatisfiable set T of ground instances of S. Then we say a ground instance C of S is small if it is in the set [D2T fD it is large otherwise. We say that an atom is small if it appears (positively or negatively) in a small clause, and it is large otherwise. Also, a literal :A is small if A is small, otherwise :A is large. Theorem 3.8 Ordered semantic hyper-linking is complete, that is, if a set S of clauses is unsatisfiable, then eventually the empty clause fg will be derived. Proof. As above, we observe that the clauses mi(S; I) will be small for all current interpretation I, and there are only finitely many such clauses. It follows that all the current interpretations I constructed will be of the form I 0 [L], where L is a subset of the small literals. This is a finite set of interpretations. Each current interpretation constructed is larger than its predecessor in the ordering ? on interpretations; thus the same interpretation cannot be seen twice, and the search must eventually stop. The only way that this can happen is for the empty clause to be generated, if S is unsatisfiable. 2 3.4 Efficiency We briefly note one advantage of this approach over semantic hyper-linking as described in [CP94a]; that is that the growth in the number of eligible literals is better controlled. The number of eligible literals has a strong effect on the work required to find a minimal clause D contradicting the current interpretation. The procedure "simp" will automatically perform A-ordering resolutions when the maximal literal of the contradicting clause D is the complement of an existing eligible literal; each such resolution has the effect of removing an eligible literal from C. In semantic hyper-linking without an ordering, it is rare for eligible literals to be removed. Typically the set of eligible literals grows rapidly, making the search procedure time-consuming after a few rounds of search and making it difficult to find proofs that require more than a few rounds of search. However, the other parts of semantic hyper-linking are powerful enough so that many proofs can be found within three or four rounds of search. In addition, the way that the literals are ordered in semantic hyper-linking makes the propositional satisfiability test fast. Still, we think it would be an advantage to be able to handle many rounds of search efficiently. Finding Minimal Contradicting Instances The preceding discussion has not dealt with the practical aspects of how the minimal contradicting instance mi(S; I) is found. We have dealt with this to some extent in [CP94a]. The problem is, given an interpretation of the form I 0 [L a set S of (possibly) non-ground clauses, to test whether I 0 [L if not, to find a minimal ground instance D of some clause C in S such that D contradicts I (if such a clause D exists). For this purpose, as we shall see, it is helpful to choose an I 0 that is decidable, that is, given a non-ground clause C we can decide whether I 0 There are a number of kinds of interpretations I 0 that are decidable; among them are the syntactic interpretations, interpretations with a finite domain, and interpretations over the reals in which all the functions and predicates can be expressed in terms of linear arithmetic with inequality. We say that an interpretation I is syntactic if for any two atoms A and B with the same predicate symbol, I depending only on the signs and predicate symbols of literals, are fairly limited in expressiveness, but are sometimes useful anyway, as we show in [CP94a]. As an example of an interpretation in terms of linear arithmetic with inequality, we might interpret P (x; y) as 2x ? y and we might interpret f(x; y) as if (x ? y) then x else y. Now, if I 0 6j= C, then we will need to find a ground instance C \Theta of C such that I 0 6j= C \Theta. A problem is that even if I 0 6j= C, such a ground instance may not exist; the reason is that some of the elements of the domain of I 0 may not be values of any finite ground terms. The question, given clause C, does there exist a substitution \Theta such that C \Theta is a ground clause and I 0 6j= C \Theta, seems to be harder than deciding if I 0 We say an interpretation I 0 is Herbrand decidable if this question about \Theta is decidable. We note that syntactic interpretations and interpretations with a finite domain are Herbrand decidable. We don't know whether interpretations over the reals in which the functions and predicates can be expressed using linear arithmetic with inequality, are Herbrand decidable. If I 0 is not Herbrand decid- able, then we may have a current interpretation I that satisfies S without being able to detect this, and we may then spend an infinite amount of time fruitlessly searching for a ground instance C \Theta that contradicts I. However, this will not affect the completeness of ordered semantic hyper-linking, because if such a C \Theta exists, it will eventually be found. If I 0 is Herbrand decidable, then the question whether I decidable, as we will now show. We first specify in more detail how such an instance C \Theta can be found, if it exists, and moreover an instance that is minimal, assuming that I 0 is decidable. Our approach is to construct a set Z of literals with the following property: The set of ground instances of literals in Z is exactly the set of ground literals L such that I 0 6j= L and neither L nor its complement L appear in the list [L of eligible literals. Then one can show that D contradicts I 0 [L literal of D is either the complement of an eligible literal or an instance of a literal in Z. For, literals of D that are complements of some L i will be false in I 0 [L since I 0 [L Also, literals that are instances of elements of Z will be false in I 0 [L agrees with I 0 on such literals. To obtain such instances D, then, we can take a clause C of S and apply a substitution \Theta such that all literals L of C \Theta are instances of some literal in Z or complements of some eligible literal. That is, we find most general \Theta such that for all L in C \Theta, there exists a literal M in Z or in the set of complements of eligible literals such that L and M are identical. We note that finding such a \Theta is reminiscent of the hyper-linking method of [LP92]. If all the literals in Z are ground literals, this can be done by a matching procedure, in which the literals of C are matched one by one; if some literals in Z are non-ground, we may need to do successive unifications. Then if the resulting instance D is non-ground, it is necessary to instantiate the variables with ground terms in some manner. Since we are only interested in minimal ground instances, we can replace each variable by a ground term that is minimal in the clause ordering, assuming that the clause ordering ! cl is monotone. For our purposes, we say that the clause ordering is monotone if it can be extended to a partial ordering on terms such that for terms s and t, s ! cl t implies C[s] ! cl C[t] for a clause C[s] containing an occurrence of s. Furthermore, we require that for any term t there are at most finitely many terms s for which :(t ! cl s). Note that this implies that there are at most finitely many minimal terms, by the condition just given. However, in order to find a minimal instance, it may be necessary to replace each variable by all minimal terms in all possible ways, which can be expensive. For this purpose, we write s#t to mean :(s ? t)":(t ? s) and we write C#D similarly for clauses C and D. Then we can choose an ordering ! cl so that s#t implies C[s]#C[t]; for example, ordering terms by their size satisfies this condition. For such an ordering, we can obtain a minimal instance of C by replacing all variables by arbitrarily chosen minimal terms. Now, the problem is to generate Z. Without knowing more about I 0 , not much can be said. If the test I 0 j= L is decidable for ground L, then one can enumerate all ground literals L, discard eligible literals and their complements, and test whether I 0 L, and in this way generate, or at least enumerate, Z. The fact that Z is infinite need not be a problem, because one only needs to enumerate Z in ascending order in the ordering ! cl in order to find a minimal contradicting instance D. As soon as one instance D has been found, literals L such that D ! cl L need not be examined, and this eliminates all but finitely many literals. For this we need to assume that if L 2 C and L 6= C then L ! cl C. In [CP94b] we indicate how specialized decision procedures can be used to aid in the generation of such sets Z of literals, in some cases. Of course, such an enumerative method cannot detect if the current interpretation satisfies C (or S). Here we choose a different approach that permits a relatively small set Z 0 to be generated independent of I 0 , and delays the consideration of I 0 to a later stage. This approach also permits us to detect if the current interpretation satisfies S, if I 0 is Herbrand decidable. We make some comments about the complexity measures used. These are defined in terms of the sizes of various structures, considered as character strings. Alternatively, the size of a clause, set of clauses, etc. is the number of occurrences of symbols in it. So if we say that something can be computed in time polynomial in a set S of clauses, we mean that the running time is bounded by a polynomial in the length of S, written out as a character string; this is the usual complexity measure. 4.1 Disunification Let Z 0 be a set of literals such that a ground literal L is an instance of a literal in exactly when neither L nor its complement appear in the set of eligible literals. Thus we have a kind of disunification problem. It turns out that a finite set Z 0 always exists, since the eligible literals are ground and there are finitely many of them. We are using the fact that there are only finitely many function and predicate symbols in all. Also, Z 0 can be computed in time polynomial in the list of eligible literals. To see this, let Z 0 (E) be a desired Z 0 of literals as desired where E is the set of literals that are either eligible literals or their complements. We first note that the positive and negative literals can be handled separately. Let L pos be the set of all positive literals and let L neg be the set of all negative literals. Let pos be the positive literals in E and let E neg be the negative literals in E. Then We can then extend this further, to consider the sets of positive and negative literals having specified predicate symbols, and solving for Z 0 for each such subclass. If there are n predicate symbols in all, we obtain a total of 2n subproblems to solve, whose solutions can then be combined to obtain the desired Z 0 . If a predicate symbol with a specified sign does not appear in E, the subproblem has a simple solution; then P is in Z 0 (E) for suitable P , since none of its instances will be in E. Otherwise, we can consider the sets of literals having various function or constant symbols in some chosen position. Each such further division reduces the problem into a number of subproblems equal to the number of function and constant symbols, and partitions the set E further into disjoint subsets; eventually we obtain trivial problems, whose solutions can be combined. 4.2 Generating eligible instances Suppose the set Z 0 of literals has been generated, as specified above. We generate instances of the input clauses as follows: For each clause C in S, we partition C into two disjoint sets of literals C ff and C fi . This is done in all possible ways. For each such partition, we unify the literals in C ff with complements of eligible literals, and those in C fi with literals from Z 0 . This also must be done in all possible ways. In this way, we obtain an instance D of C which can be expressed as D ff [D fi where the literals in D ff are complements of eligible literals and the literals in D fi are instances of literals of Z 0 , that is, they do not unify with eligible literals or their complements. We call such an instance an eligible instance. Note that there are only finitely many eligible instances, and they can be found by a simple enumeration procedure. In order to ensure that D contradicts I 0 [L it is necessary to find a substitution fl such that I 0 6j= (D fi )fl. We also want this instance to be minimal in the clause ordering ! cl . If fl is as specified, then I 0 [L reasoning as above: The literals in D ff are not satisfied by the current interpretation because they are complements of eligible literals. The literals in (D fi )fl are not satisfied by the current interpretation I 0 [L because they are not satisfied by I 0 . If I 0 is Herbrand decidable, then the test whether such a fl exists is decidable, and in this way we can test if the current interpretation satisfies S. If such a fl exists, we want to find one such that Dfl is minimal in the ordering ! cl . The search for depends of course on I 0 and ! cl . If the clause ordering is based on the directed acyclic graph size of the terms, that is, the size is the number of distinct subterms that appear, independent of how often they appear, then this computation seems to be rather complicated, despite the favorable theoretical properties of this size measure. However, for certain types of interpretations and literal orderings, the search for such a fl producing a small clause can be done quickly, as we will show in the next section. This is also dealt with to some extent in the paper [CP94b]. It is likely also that this search can always be done fast if I 0 is an interpretation with a (small) finite domain and if the clause ordering ! cl is monotone. However, we can say something more in case I 0 is syntactic. If I 0 is syntactic, then I 0 6j= (D fi )fl iff I 0 6j= D fi , which latter condition can be checked just by examining the signs and predicate symbols of the literals. This permits some clauses to be rapidly eliminated from further processing. 4.3 Complexity analysis Now, the generation of all the eligible instances requires an amount of work that depends on the number and sizes of the eligible literals and the number of literals in the clauses. However, suppose that there is a constant bound on the number of literals in each clause of S. Then the generation of the eligible instances can be done in time polynomial in S and the set of eligible literals, since one has to look at all combinations obtained by unifying each literal in C with each complement of an eligible literals or with each element of Z 0 . If the set of eligible literals unioned with Z 0 has n elements and a clause C has k literals, then there are n k possibilities, which is still polynomial, for a fixed k. For each possibility, the work is polynomial (or even linear), using efficient unification algorithms. Note that n is polynomial because the number of elements of Z 0 is polynomial, by the disunification arguments given above. The assumption that k is bounded is reasonable, because it is possible to convert clauses with many literals into clauses with three literals in a satisfiability- preserving manner by introducing new predicate symbols. We have just shown that if the number of literals in clauses of S is bounded, then the eligible instances can be generated in polynomial time, regardless of I 0 . However, in order to generate a minimal contradicting instance, it is necessary to instantiate the eligible instances with a substitution fl as specified above. If I 0 is syntactic and the clause ordering ! cl is monotone and has properties concerning # as specified in section 4, then such a fl can be obtained simply by checking the signs and predicate symbols and replacing all variables by a minimal term; the entire process of generating a minimal contradicting instance can then be done in time polynomial in S and the set of eligible literals, that is, polynomial in their lengths written out as character strings. If I 0 has a finite domain, then we can find for each domain element d a minimal ground term t d whose interpretation under I 0 is d; such terms can be found by a simple iteration procedure. Then to find a minimal instance Dfl such that I 0 6j= (D fi )fl, it is only necessary to consider fl replacing variables by terms of the form t d for various d in the domain. This is a finite number of possibilities, exponential in the number of variables in D fi . If we use an ordering based on directed acyclic graphs, that is, counting the number of distinct subterms that appear, but not counting how often they appear, then the generation of a minimal contradicting instance can be more complicated. This is because such directed acyclic graph-based orderings are not monotone, since the contribution of a subterm to the clause depends on subterms that appear elsewhere in the clause. However, because of the favorable complexity properties of the directed acyclic graph-based orderings, it may be worthwhile to use them anyway. In general, it makes sense to look at the eligible instances in order of their size; then it is possible that a small contradicting instance may be found early and the instantiation of the larger eligible instances may be avoided. 5 Additional Ordered Resolutions Since the basic search procedure performs A-ordering resolution, it seems reasonable to do additional resolutions; these might have the effect of doing part of the work in advance, and thereby speed up the search. So we might add to the input clause set S a set S 0 of clauses generated by A-ordering resolution from S, and consider these clauses S 0 as additional input clauses. One would expect that these additional clauses might lead to the finding of certain contradictions earlier. They also have the effect of eliminating large literals from proofs, as mentioned in [CP94a]. Since the basic search procedure has difficulty generating large literals, this combination is reasonable, and we have found that in practice such a combination (with rough resolution instead of A-ordering resolution to eliminate large literals) often works well. We can balance the work between ordered semantic hyper-linking and A- ordering resolution in some way so that both are done in parallel; one method to use is to divide the total time spent equally between them. The question remains what ordering to use for these A-ordering resolutions, and which A-ordering resolutions to do. As for the ordering, we can choose some literal ordering ! lit compatible with the ordering ! on ground literals. That is, we can say lit M (for non-ground literals L and M ) if there is a ground substitution \Theta such that L\Theta ! M \Theta. Note that we really need to use the ordering ! here, not ! cl , to compare L\Theta and M \Theta. Then we can restrict literals resolved on, to literals that are maximal in this ordering ! lit . As for which A-ordering resolutions to perform, we observed in section 4.3 the strong dependence of the efficiency of ordered semantic hyper-linking on the number of literals in clauses of S, so it is reasonable to keep the number of literals small. However, we want to perform the A-ordering search in a complete manner, so that a reasonably large portion of the search space is explored. For this, it is necessary to consider not just the number of literals in a clause, but the size (number of symbol occurrences) of the clause. Preferring clauses of small size can be done in a completenesss-preserving manner, and will also tend to keep the number of literals small. The question remains exactly how this can be done. One way to do this is to keep a list of all pairs of clauses that have not yet been resolved together, and always resolve the pair of clauses whose sum of sizes is as small as possible. This produces clauses of small size (number of symbol occurrences), which will tend to have few literals, but requires a quadratic amount of space to store these possible resolutions. Another method is the Otter [McC89] approach, in which a small clause is repeatedly chosen and resolved against all other clauses. This avoids the expensive bookkeeping, but has the problem that this small clause will also resolve against large clauses, possibly producing large clauses very early. We propose a compromise, in which a list of clauses is constructed as follows: The clauses are entered into this list L smallest first. Whenever a clause is entered into the list, it resolves (using A-ordering resolution) against all clauses that are already on the list. In this way, all resolutions will eventually be done, but we have some guarantee that when two clauses are resolved, both of them are small. Also, the amount of bookkeeping is kept to a minimum. When we are done resolving C i against C j for all j i, then we find the smallest clause D such that D is in S or has been produced by an earlier resolution, and such that D is not already in the list L. We then enter this smallest clause D in the list as C i+1 and resolve D against C j for 1 j i, continuing the process. Now, for certain problems, it is important to produce large A-ordering resolvents early, if they are directly derived from the negation of the theorem; this is the case for example in Bledsoe's five limit problems [Ble90]. Therefore, we would like to make this A-ordering resolution procedure sensitive to support criteria as well as size. For this purpose, we can say that a clause C is semantically supported if it contradicts the user-given interpretation I 0 . Both input clauses and clauses generated by resolution can be semantically supported. We then can modify the above A-ordering search strategy so that on alternate choices of D, a semantically supported clause is chosen. Thus, we alternate between choosing D as the smallest clause not yet in L, and the smallest semantically supported clause not yet in L; this D is then added to the list and resolved against all clauses already in L. This will tend to favor resolutions involving supported clauses, even if the clauses are large. This can only be done systematically if I 0 is decidable (or better, Herbrand decidable). 5.1 Explanation-based generalization The idea of explanation-based generalization (EBG) is to extract some general principles from a specific argument, enabling the argument to be applied to a wider range of situations. This principle can be applied in the ordered resolution phase of ordered semantic hyper-linking. We note that in the procedure "simp" of section 3.1, A-ordering resolutions are performed between ground clauses Cn and D; both Cn and D are instances of clauses in S or clauses produced from S by prior A-ordering resolutions. Suppose C 0 n and D 0 are the more general (possibly non-ground) clauses of which Cn and D, respectively, are instances. Then by general properties of res- olution, it follows that there is an A-ordering resolvent C of Cn and D such that a res(C n ; D) is an instance of C. Therefore, it seems reasonable to store these more general clauses C 0 along with their respective instances Cn , D, and a res(C n ; D). In this way, we produce lemmas that can be added to the set of input clauses; such lemmas are likely to be relevant to the proof and may be generated again during the search. By adding them to S, we may avoid sections of the search in which the same resolutions are performed over and over again. It makes sense to generate these lemmas, in addition to performing the A-ordering resolutions of the previous section, since these lemmas may not be generated by the list-based search method described there. The fact that a uniform ordering is used in the search may make the lemma mechanism more effective; the search method used in semantic hyper-linking without ordering can vary the literal ordering for different interpretations, making it less likely that a lemma found earlier will be useful later on. 6 Replacement Rules We found that the use of replacement rules considerably enhanced the performance of semantic hyper-linking, and so it is reasonable to include them in ordered semantic hyper-linking, too. However, it is necessary to adapt them to the current context. Suppose is the current set C of relevant clauses. From this we construct a set EL of explicit literals; these are the literals satisfied by lm(C) that actually appear in the clauses C i . Thus, the set EL is defined as the set of literals L such that either L or L is a member of [ i C i and such that lm(C) Now, if we can show that EL[S is unsatisfiable, then we have already contradicted the current interpretation lm(C) and need not search for a minimal contradicting instance. For, if we can show that EL [ S is unsatisfiable, then we know that the clause is a logical consequence of S, and can be used as if it were the minimal contradicting instance. In fact, we typically find a small subset of the clause that is a consequence of S, which may be more useful for the search. To see whether EL[S is unsatisfiable, we use some incomplete but often effective methods. In the general AI context, these may be viewed as "obvious inferences" or "associations" that are readily made. One method is to use natural replacement rules; these are implications of the form L such that all variables in the literal L appear elsewhere in the implication, and such that some clause is a member of S. These replacement rules may be used by unifying the L i with elements of EL; then the corresponding instance of L will be a ground literal (since all variables of L also appear elsewhere). This ground literal is a logical consequence of EL and can be (temporarily) added to EL. This operation may be repeated a number of times, and if complementary literals appear in EL, then we know that EL [ S is unsatisfiable. From this derivation, a clause that is a subset of fL : L 2 ELg may be extracted, which is a logical consequence of S, and can be used in place of the minimal contradicting instance. Another kind of replacement rules are definitional replace rules; this is a slight simplification of the minimal replace rules of [CP93a]. The idea of definitional replace rules is to capture clauses that represent definitions, and the effect of applying them is to expand the definitions; this is particularly useful in set theory and modal logic. A definitional replace rule is of the form L such that all variables in L 1 appear in L and such that some clause fL; L is a member of S. If one has a definition of the form L j A where L is a literal and A is a formula involving no quantifiers, one can verify that such clauses will be generated when this definition is converted to clause form. Sometimes such formulae are generated even if A contains quantifiers. Such replacement rules are used in an inverse way to natural replacement rules; if L has an instance L\Theta in EL then the implication L\Theta can be added to EL. Also, if L has an instance L\Theta appearing on the right-hand side of some such implication that has previously been added to EL, then this instance can be added to EL. In this way, EL is augmented by a set of ground instances of clauses in S. These instances are the instances that one would consider when expanding definitions. We can then test if EL together with these ground instances is unsatisfiable, using something like Davis and Putnam's method [DP60]. If so, then a clause that is a subset of fL : L 2 ELg may be extracted, which is a logical consequence of S; thus we know that EL [ S is unsatisfiable. In order to control the application of replace rules, it seems most reasonable to use a time bound based on the time used to search for a minimal contradicting instance. It seems reasonable to first search for a contradiction using natural replace rules. If this fails, then definitional replace rules can be applied, again controlled by the time bound. If this fails, then we have generated a set G of ground clauses that can be processed a little more; this set G contains EL (considered as unit clauses) together with the ground instances generated by definitional replacement. Since G consists of ground clauses, it has only finitely many models. We can examine all these models J of G one by one, and for each such model J , we can again apply natural replacement to J to see if it can be contradicted. We note that since G is finite, the models J are essentially finite, too. Each such model J can be considered as a set U of unit clauses, that is, the set of literals L such that J j= L and such that L or its complement appears in G. Then we can perform natural replacement on this set U ; this may find some contradictions that were nearly, but not quite, found by definitional replacement. Such contradictions (demonstrations that S [U is unsatisfiable) need to be found for all models J of G in order to demonstrate that 7 Complexity Analysis We now consider the complexity required by ordered clause linking, in terms of the complexity of the shortest proof from S. We consider both worst-case bounds and give plausibility arguments for better performance. In the introduction, we noticed that clause linking has a triple exponential bound in the complexity of the proof, which can be reduced to double exponential if a suitable ordering on literals is used. The same arguments apply to ordered clause linking. However, we have reason to believe that the performance will be better than this. We know that if S is unsatisfiable then there is an unsatisfiable set T of ground instances of S. Suppose we measure the complexity of the proof by the complexity c(T ) of T , that is, its length when written out as a character string. Or, equivalently for our purposes, we can measure the complexity of the proof by the complexity of the largest clause in T . Note that this complexity measure does not economize on repeated occurrences of the same subterm. Now, suppose our clause ordering is based simply on the length of the clause, written out as a character string, that is, the sum of the sizes of the literals in the clause. Then each clause appearing in the proof has complexity at most c(T ), so we will find the proof when all clauses of complexity c(T ) or less have been generated (or earlier). The number of such clauses is exponential in c(T ), and so the time required to test their satisfiability may be double exponential in c(T ). We note that this measure is independent of how many clauses appear in T , if the "largest clause" complexity measure is used. This will give our method a better comparison with the methods of [Gou94], which need to count the number of elements in T . Another favorable factor for our method is that the satisfiability of a set of ground clauses can be tested in expected polynomial time, for many probability distributions. And in practice, methods similar to Davis and Putnam's method often decide satisfiability of sets of propositional clauses very fast. We have also found this to be the case in the clause linking theorem prover. Since our search procedure is similar to Davis and Putnam's procedure in its systematic search for a model, we would expect a similar time bound to apply to it, too; thus we may expect in practice that the time required by our method is single exponential. We note further that this is based on the assumption that all clauses of complexity c(T ) or less are generated. Our method is very selective about which clauses are generated, so that it is reasonable to assume that only a small subset of the clauses of complexity c(T ) or less will be generated. For example, we can show that our method will only generate logically minimal ground instances, that is, ground instances that are not logical consequences of smaller (with respect to ! cl ) ground instances. Equivalently, a ground instance C is logically minimal if there is some interpretation I such that C is a minimal instance of S contradicting I. The question whether a clause is a logical consequence of simpler clauses is also relevant for the methods of [BG90, BG94], it turns out. One would expect that the number of logically minimal ground instances of a given size is much smaller than the total number. In fact, we do not even generate all the logically minimal clauses. The fact that A-ordering resolutions are done in the "simp" procedure means that many interpretations are not even examined. That is, we only generate clauses that are logically minimal when such A-ordering resolvents are also considered, so this can eliminate some clauses from being logically minimal. But, as a worst case bound, our method is exponential in the number of logically minimal ground instances; this is always finite when S is unsatisfiable. One might ask whether it would be just as efficient to generate all ground instances of clauses in S, and then consider them in order of size. Thus we would have a list C of all ground instances of S, with the smallest ones occurring earlier in the list. We could then test if C i is a logical consequence of C i , for i j, and if so, delete C i from the list. This produces a smaller sublist C 0 of ground instances, of which finite prefixes can be tested for satisfiability. Our method is more efficient in that the non-logically minimal instances are never even generated. Furthermore, even some of the logically minimal instances are avoided, as explained above. 7.1 Estimating the size of the tree We can give additional evidence that the work required by ordered semantic hyperlinking is often small. We note that it is only the small atoms that influence the search, as defined in definition 3.7. Let us consider the smallest n atoms in A and the probability that a clause C over these atoms will contradict an arbitrary interpretation I. Suppose that there are m 3-literal ground clauses C in all. Then the chance that a random interpretation I will not satisfy a specific clause C is 1=8 (since each of 3 literals must be mapped to false). Therefore the chance that the interpretation will satisfy the clause C is 7=8;, and the chance that I will satisfy m 3-literal clauses is (7=8) m , if the clauses are chosen independently. The expected number of models of m independently chosen clauses (considering only the first n literals) is then 2 n (7=8) m since there are 2 n interpretations altogether for n atoms. We note that and the expected number of models is less than one. This means that the probability is very small that we will have to search past the first n atoms to find a contradicting instance. This is evidence that if there are many clauses then the size of the tree is small, on the average. What we are given is non-ground clauses, in general, instead of ground instances, so the determining quantity for this analysis is the number of their ground instances of various sizes, and how they depend on one another. --R Theorem proving via general mating. On restrictions of ordered paramodulation with simplification. Ordered chaining for total or- derings Automated Theorem Proving. Challenge problems in elementary calculus. Model finding in semantically guided instance-based theorem proving Rough resolution: a refinement of resolution to remove large literals. Semantically guided first-order theorem proving using hyper-linking The use of presburger formulas in semantically guided theorem proving. Rewrite systems. A computing procedure for quantification theory. Resolution Methods for the Decision Problem. The complexity of resource-bounded first-order classical logic Proving refutational completeness of theorem-proving strategies: the transfinite semantic tree method A simplified format for the model elimination procedure. Eliminating duplication with the hyper-linking strategy Otter 1.0 Users' Guide. Conditional term rewriting and first-order theorem proving Equational reasoning and term rewriting systems. The search efficiency of theorem proving strategies. The search efficiency of theorem proving strategies: an analytical comparison. Automatic theorem proving with renameable and semantic resolution. --TR --CTR Kahlil Hodgson , John Slaney, TPTP, CASC and the development of a semantically guided theorem prover, AI Communications, v.15 n.2,3, p.135-146, August 2002 Kahlil Hodgson , John Slaney, TPTP, CASC and the development of a semantically guided theorem prover, AI Communications, v.15 n.2, p.135-146, September 2002 R. Janvier , Y. Lakhnech , M. Prin, Certifying cryptographic protocols by abstract model-checking and proof concretization, ACM SIGBED Review, v.3 n.4, p.37-57, October 2006 David A. Plaisted , Adnan Yahya, A relevance restriction strategy for automated deduction, Artificial Intelligence, v.144 n.1-2, p.59-93, March Adnan Yahya , David A. Plaisted, Ordered Semantic Hyper Tableaux, Journal of Automated Reasoning, v.29 n.1, p.17-57, 2002 Carsten Sinz, Visualizing SAT Instances and Runs of the DPLL Algorithm, Journal of Automated Reasoning, v.39 n.2, p.219-243, August 2007 Peter Baumgartner, First-order logic Davis-Putnam-Logemann-Loveland procedure, Exploring artificial intelligence in the new millennium, Morgan Kaufmann Publishers Inc., San Francisco, CA,
automated theorem proving;hyper-linking;semantics
594258
Semi-Thue Systems with an Inhibitor.
A i>semi-Thue system with an inhibitor is one having a special symbol, called an i>inhibitor, that appears on the right side of every rule but does not appear on the left side of any rule. The main result of this paper is that the uniform halting problem is decidable for the class of such systems. The concept of i>inhibitor is related to the concept of i>well-behaved derivation in systems without an inhibitor. The latter concept has received some attention from those interested in the open question as to whether the uniform termination problem for one-rule semi-Thue systems is decidable.
April, 1998 (revised) 1. Introduction, notation and terminology. Semi-Thue systems constitute a universal model of computation, in the sense that any decision problem about computations is reducible to a problem about semi-Thue systems. Every recursively enumerable language is the language of some semi-Thue system. (E.g., see Chapter 7 of [3].) And yet the concept is simple: A semi-Thue system is an ordered pair (\Sigma; \Pi), where \Sigma is a finite alphabet of characters and \Pi a finite set of rules (u; v), where in each case u and v are words over \Sigma. These systems are called "semi-Thue systems" because the rules are not necessarily reversible, u being referred to as "the left side" and v "the right side" of the rule (u; v). Semi-Thue systems are distinguished from Thue systems (named for their originator Axel Thue [15]) whose rules operate in both directions. The important notion in the study of semi-Thue system is that of a derivation from one string to another. We write to mean that w 2 is derived in one step from w 1 , i.e., that there exist words x and y such that w (u; v) is a rule of the system. We say that z is derivable from w or that there is a derivation from w to z, if there exist strings w Each w i is a line of the derivation, which has lines, and has p steps; p is the length of the derivation. An infinite derivation is an infinite sequence w nonnegative integers i. Although the class of semi-Thue systems is quite powerful, there have been some applications in which not all this power is appropriate. An example is the context-free Supported by Grant No. CCR-9500182 from the National Science Foundation. grammar, which had as its origin the phrase-structure grammar put forth tentatively and critically by Chomsky [2] for natural languages, but which by now has its chief application in the domain of formal languages (see [11]). A context-free grammar is a semi-Thue system that (1) distinguishes between "terminal" and "nonterminal" characters of the alphabet, (2) restricts rules to those having single nonterminals as their left sides, and (3) restricts derivations to those whose first lines all consist of a single occurrence of a particular nonterminal designated as the start symbol (usually S in the literature) and whose last lines have only terminal characters. Another application has arisen in the area of computer science concerned with theorem proving by machine, where semi-Thue systems are used and thought about extensively. Typically, these semi-Thue systems, often called "rewrite systems," are used to reduce words to simpler equivalent words (see, e.g., [7] or [1]). Rewrite systems and context-free grammars are quite different both in their purpose and their operation. Nevertheless, these enterprises do have in common their use of computationally weak semi-Thue systems. The research program into which the present paper fits is not concerned with particular applications of semi-Thue systems such as grammars and rewrite systems. Rather it considers various subclasses of semi-Thue systems that appear to be usefully weaker than the entire class (without regard to any particular application), and then attempts to assess how useful and weak they are. One method of assessing a subclass is to determine whether certain decision problems are decidable for it. If so then that is evidence that the subclass is weak, since most significant decision problems are undecidable for the full class of all semi-Thue systems. And, if the decision problem is a computationally important one, it is evidence also that the subclass may be useful. There are several decision problems that can be used as criteria in this way. Let us focus on three of them: (1) The halting problem. Given a semi-Thue system (\Sigma; \Pi) of the subclass, and given x 2 \Sigma , is every derivation whose first line is x finite? (2) The uniform halting problem. Is every derivation in a given semi-Thue system of the subclass finite? (3) The derivability problem. Given a semi-Thue system of the subclass, and given x; y 2 \Sigma , does there exist a derivation of y from x? For example, for both the class of context-free grammars and the class of rewrite systems used by many theoreticians of machine theorem proving, all three of these problems are decidable. Moreover, any subclass of the semi-Thue systems for which any of these three problems is solvable is weaker and possibly more useful than the class of all semi-Thue systems. Weaker because all three problems are undecidable for the class of all semi-Thue systems. And more useful, albeit in a restricted set of possible applications, because the ability to tell whether or not derivations will terminate gives the user an advantage. The importance of termination in the practical use of rewrite systems is described in detail in [6]. The uniform halting problem seems more difficult than the other two, although there seems to be no general proof that if that problem is solvable for any subclass of semi-Thue systems then the other two problems are solvable also. At any rate, several research workers have decided to focus on the uniform halting problem; in particular, the question of whether or not that problem is decidable for the subclass of one-rule semi-Thue systems has received some attention, so far without an answer, in [4], [5], [8], [9], [10], [12], [13], [14] and [16]. This subclass will be discussed in Section 4. The subclass that is the main focus of this paper is the subclass of semi-Thue systems with an inhibitor, i.e., an alphabetic character that occurs at least once on the right side of every rule, but does not occur on the left side of any rule. The main result is that the uniform halting problem for this subclass is decidable, proved in Section 3, where it is also proved that the halting problem and the derivability problem are decidable. Derivations in semi-Thue systems with an inhibitor will be analyzed in detail in Section 2. The subclass of semi-Thue systems with an inhibitor was noticed in the course of studying the subclass of one-rule systems [13]. The results of the present paper on semi-Thue systems with an inhibitor lead to a distinction that promises to be fruitful in analyzing derivations in one-rule systems without an inhibitor. That distinction, between well behaved derivations and ill behaved derivations, is explained in Section 4. In this paper lower-case Greek iota (') will be used as an inhibitor. So, in a semi-Thue system with an inhibitor, ' 2 \Sigma, every right side has at least one occurrence of ', and no left side has any occurrence of '. Our convention will be that a semi-Thue system has k rules: We shall write x ! n y to mean that there is a derivation of n steps (n - 0) from x to to mean that there is a derivation from x to to mean that there is such a derivation of at least one step. For mean that there exist z 1 ; z 2 2 \Sigma such that x ! z 1 yz 2 . We shall write x . + y to mean that there exist z 1 ; z 2 2 \Sigma such that The well known fan theorem states that an infinite rooted tree in which each node has only finitely many children nodes has an infinite path. The following generalization will be useful in Section 2; the proof is left to the reader: Theorem 1.1. If a forest of finitely many rooted trees has infinitely many nodes, each of which has only finitely many children nodes, then the forest has an infinite path. Definition (the y). If x; y and z are words and xyz is a line in a derivation then it may be that y occurs several times as a factor (i.e., substring) of that line. For example, then there are three occurrences of y in xyz, the second of which we shall refer to as "the apparent occurrence of y in xyz," or "the y in xyz," or (frequently, when the designation xyz is understood) "the y." The article "the" will indicate that we are referring to a particular occurrence of y and not to the word y itself. When we wish to talk about the word apart from any occurrence in a line we shall simply say "y." Definition (the same occurrence). If x 1 bx 2 u h x 3 th line of a derivation where b is a letter and x 1 bx 2 v h x 3 is the (j st line of that derivation, then we talk of the apparent occurrence of b in the j th line as being the same as the apparent occurrence of b in the (j st line. On the other hand, no occurrence of a character in the apparent occurrences of u h or v h exists in the other line: any character occurrence in the apparent u h is destroyed, and any character occurrence in the apparent v h is created, in going from the j th line to the (j st line. 2. Analysis of derivations. Section 3 will present the decision procedure to determine whether a given semi-Thue system with an inhibitor is uniformly terminating. In preparation, this section will develop a method of analyzing derivations in such systems. To begin, it is stipulated that there be no occurrence of the inhibitor ' in the first line of a derivation. This stipulation will help us establish important structural concepts. It is justified by the fact that (since ' does not occur on the left side of any rule) if there is an infinite derivation from x 1 'x then there is an infinite derivation from one of the x i 's. All semi-Thue systems discussed in this section and Section 3 will be systems with the inhibitor '. Definition (vital). For b 2 \Sigma, if xby is a line other than the first in a derivation, then the apparent occurrence of b is vital if it is not ' and has been created either in that line or in some preceding line of the derivation (in other words, if that occurrence of b does not exist in the first line). Definition (S-occurrence, the set S). (1) If w 6= - and xwy is a line of a derivation then the w is an S-occurrence in that line if all its character occurrences are vital, the x does not end in a vital character occurrence, and the y does not begin with a vital character occurrence. (That is to say, the x must either be null or end in a nonvital character occurrence. And the y must either be null or begin with a nonvital character occurrence.) In short, a nonnull S-occurrence is a maximal factor of vital character occurrences in a line of a derivation. (2) If a line of a derivation begins with (ends in) ' then the occurrence of - at the left end (right end) of the line is an S-occurrence. And if bc is a factor of a line where both the b and the c are nonvital character occurrences, at least one of the two being ', then the - between the b and the c is an S-occurrence. (3) S is the set of all words having an S-occurrence in some line of some derivation. Note that if x's'y ! z and s has no ' then and y ! y 0 . This follows from the segregating power of the inhibitor ', and shows the importance of S-occurrences in lines all of whose non-iota character occurrences are vital. Even more revealing of the importance of S-occurrences is the next theorem. Definition (S-sequence). The sequence s is an S-sequence if, for each i, Theorem 2.1. In a semi-Thue system with an inhibitor, there exists an infinite derivation only if there is an infinite S-sequence. Proof: If there is an infinite S-sequence s then an infinite derivation can be defined as follows: For each i, define z i and z 0 i so that s (guaranteed by the definition of s i . s is an infinite derivation. Now assume there is an infinite derivation w where the character occurrences of w 1 that are never destroyed in the derivation are precisely those that occur in the apparent factors y . So, for every i, for some z . Consequently, for some j, 1 - j - m, the sequence z is an infinite derivation possibly with repetitions; that is to say, there is an infinite sequence For convenience we alter the notation and set z is an infinite derivation in which every character occurrence of z 1 is eventually destroyed. Since z 1 is only finitely long, there exists a p such that, for all i - p, the line z i has no character occurrences in common with z 1 . Thus for all i - p, where each s i;h 2 S, and q is the number of "s in z i . For each line i - p, there is exactly one h such that the s i;h is rewritten by a rule, becoming (for some in the (i st line; for all is copied to become the very same S-occurrence in the st line (but with a different subscript for to mean that the s i;h 0 is copied to become the s i+1;h 00 in the st line. Then define C to be the transitive, symmetric and reflexive closure of the relation C 0 . Thus C(! and the s i;h is copied, recopied, etc., until it becomes the s i 0 ;h 0 in the (i 0 ) th line, or is copied, recopied, etc., to become the s i;h . Clearly, C is an equivalence relation; furthermore if C(! then the word s . Note also that, for any i and h - q i , if there is no h 0 - q i+1 such that for finitely many values of h 0 , s i;h . s i+1;h 0 Consider now the graph H whose nodes are the equivalence classes of the C relation: The roots of H are the classes having the respective pairs as members; there is an edge from a node N 1 to a node N 2 of H if and only if there are and the s i+1;h 0 in the (i st line is part of what is obtained by rewriting from the s i;h in the i th line. Clearly, H is a forest of rooted trees. Define Since the z-derivation is infinite, H has infinitely many nodes, each of which has only finitely many children nodes. Hence, by Theorem 1.1, there is an infinite path through the Putting s Definition (v-section, L(v h ); R(v h )). The word w is a v-section if ' does not occur in w and there is a v h such that, for some x and y, either v or x'w. Thus the null word is a v-section if and only if some v h either begins in ', ends in ', or has '' as a factor. If v In particular, L(v h ) (R(v h only if v h begins with (ends in) '. Note that every v-section has an S-occurrence in some derivation. Whenever a rule applied in a derivation where v the apparent occurrences of the internal v-sections always become S-occurrences. The apparent occurrences of L(v h ) and R(v h ) become either S-occurrences or parts of S-occurrences. E.g., for the system in the third line of the derivation a a'bc'd a'be'f'g'd the occurrence of g is an S-occurrence, but the occurrence of of the S-occurrence be. An S-occurrence of - must, in its first line, be the result of an application of a rule having lambda as a v-section. However, if L(v h ) or R(v h the application of the rule need not result in an S-occurrence of -. E.g., in the system In the derivation a a'bc'd a'b'f"d the application of the second rule results in the third line in which R(v 2 ) becomes an S-occurrence of - but L(v 2 ) does not, being absorbed into the S-occurrence b. Generally, we can think of a nonnull S-occurrence in a line as a maximal substring occurrence consisting of consecutive vital character occurrences; it is brought about by one or more rule applications each of which is responsible for creating some of its vital character occurrences. The earliest S-occurrences are v-sections, which come about as a result of rule applications. The more complex S-occurrences are brought about by the modification of simpler S-occurrences. The theorems that follow will describe this process in detail. Theorem 2.2. If the s is an S-occurrence in a line xsy of a derivation, then either x ends in ' or y begins with ' or both. Proof: Let s be any word having an S-occurrence in the derivation. Let xsy be the first line that has that occurrence of s as an S-occurrence. Then some ff; i; fi, the preceding line being ffu i fi. Let us make three observations. (1) The v i must have either the rightmost character occurrence of the x, a character occurrence of the s or the leftmost character occurrence of the y. (Otherwise s would have an S-occurrence in the preceding line.) (2) The v i must have a character occurrence in common with either the x or the y. (Since s has no ' but v i does, s cannot account for all of v i .) (3) If the v i has any two character occurrences of the line xsy then it has every character occurrence between. From (1), (2) and (3) we infer that the v i contains either the rightmost character occurrence of the x or the leftmost character occurrence of the y. Case I: is a single character, and the v i contains the b. Then since the s is an S-occurrence, the b is not vital. And since it is part of the v i , it is not a character occurrence of the first line. Hence Case II: is a single character, and the v i contains the c. Then, by similar reasoning, We have proved that our theorem is true of the first line having that s as an S-occurrence. But since "s are indestructible in a derivation, our theorem is true of all the lines having that s as an S-occurrence.2 We consider two examples. The first is the system with the two rules (ab; aac') and (c; b'), and the derivation ab aac' whose S-occurrences are underlined. These are the first four lines of an infinite derivation, which is obtained by applying the two rules alternately. The lines of this derivation have increasingly long S-occurrences at their left ends, followed by several S-occurrences of -. Our second example is the system with the two rules (ab; b'ac) and (c; b'a) and the derivation ab b'ac b'ab'a b'b'ac'a b'b'ab'a'a Each line of the resulting infinite derivation will have several S-occurrences of b, followed by an S-occurrence of either ab or ac, and then several S-occurrences of a. Definition (the sets S L and SR ). S fxjsome derivation has a line x'y in which the x has no nonvital character occurrencesg. fyjsome derivation has a line x'y in which the y has no nonvital character occurrencesg. It is not difficult to see that is the number of rules.) It is clear from the definitions that S L [ SR ' S. Theorem 2.3. If xbsy (xsby) is a line of a derivation in which the s is an S-occurrence and the apparent letter b is an occurrence from the first line, then Proof for xbsy: By Theorem 2.2, . The result of deleting all lines of the derivation after the line xbsy, and then deleting the noted b and all characters to its left from all lines, is a derivation possibly with repeated lines, having the line sy = s'y 0 in which the apparent s is an S-occurrence, and hence is vital. Thus The proof that s 2 SR when the line is xsby is symmetric.2 Theorem 2.4. (a) S the smallest set C such that, for all j, (1) L(v j the smallest C 0 such that, for all j, Proof that C ' L from the two-line derivation y has no ' and there is a derivation of xu j y'z, for some z. This derivation can be extended to become a derivation of xv j y'z, which shows that Proof that S L ' C: In any derivation D, put th line does not begin with a character occurrence of the first lineg. If and D has a j 0th line then also. For each j 2 J(D), let the j th line be x j 'y j , where x j is iota-free. Clearly, if J(D) 6= ; and j 0 is the smallest member of J(D) then x for some h. Also, j(D)g. We have proved that L(D) where the union is taken over all derivations D in the system, we have S L ' C. Thus S C. The proof that We shall generally focus on S L knowing that, whatever we prove about S L , an appropriate similar assertion can be proved about SR . Definition (The set S Li ). For each The proofs of the following two theorems are straightforward: Theorem 2.5. S Theorem 2.6. If xu j y 2 S Lj 0 then xS Lj ' S Lj 0 Continuing the examples given above, in the system whose rules are (ab; aac') and (c; b') we have S f-g. In the system whose rules are (ab; b'ac) and (c; b'a), we have S ag. Theorem 2.7. the smallest set S 0 such (1) every v-section is in S Proof: To prove that S 0 ' S we verify that (1)-(5) are all true of S. Items (1) and (2) are clear. The verification for item (3) is as follows: Assume u some z 0 there is a derivation whose last line is yz'z 0 where the yz is vital. The result of appending x to the left end of every line is a derivation whose last line is . We can extend this derivation by adding the line v h z'z 0 , in which the word R(v h )z will have an S-occurrence, showing that R(v h )z 2 S. The verification for item (4) is similar. To verify that (5) is true of S, assume xu h y 2 S. Then, for some z and z 0 , there is a derivation whose last line is zxu h yz 0 in which the xu h y is an S-occurrence. If we append to this derivation the line zxv h yz 0 the result is a derivation whose last line has both xL(v h ) and R(v h )y as S-occurrences, showing that both these words are in S. (Examples illustrating (3) and (5) are provided after this proof.) To prove that S ' S 0 we assume an arbitrary derivation D and prove that every S-occurrence in D is in S 0 . We do so by proving the following proposition by mathematical induction on i: If D has at least i lines then every S-occurrence in the i th line is in S 0 . This proposition is true for since the first line has no S-occurrences. Assume it is true for the i th line and let the (i st line be xsy in which the s is an S-occurrence. Assume also that s does not have an S-occurrence in the i th line, and that is the rule by means of which the i th line is rewritten as the (i st line. either (a) at the left end of the line, which begins in ', (b) at the right end of the line, which ends in ', (c) flanked by two consecutive ''s, (d) flanked by nonvital b 6= ' on the left and ' on the right, or (e) flanked by ' on the left and nonvital c 6= ' on the right. Since - is not an S-occurrence in the preceding line in the derivation, the s must be a factor of the v h in the (i st line. In each of the cases (a), (b), (d) and (e), the ' flanking the s must be part of the v h and - is a v-section of v h . In case (c) either both "s are part of the v h and hence - is an interior v-section of v h , or only one of the two "s is part of the v h and - is an end v-section of v h . Henceforth we assume s 6= -. Let the noted occurrence of s in the (i st line consist of the B(s) th through the E(s) th characters of that line, in left-to-right order. And let the noted occurrence of v h in that line consist of the B(v h ) th through the E(v h ) th characters. Note that B(v h E(s). To carry through the proof, we divide into cases based on the relative positions of s and v h in the st line. Case I: B(v h ) - B(s) and E(s) - E(v h ). Then s is a factor of v h , and since s is an S-occurrence of the (i st line, s must be a v-section of v h . By (1), s 2 S 0 . In the remaining cases either E(v h v h has an occurrence of ' but s does not, the v h cannot be wholly inside the s. The possibility E(v h give rise to Cases II and III, while the possibility give rise to Cases IV and V, which are right-left symmetric to Cases II and III, respectively. Case II: E(v h E(s). Then there would be no change in the vicinity of the s in going from the i th line to the (i st line, and the s being an S-occurrence in the st line would also be an S-occurrence in the i th line, contrary to our assumption. Furthermore, the rightmost character of v h must be ', otherwise that character occurrence would be vital and the s would not be an S-occurrence in the (i st line. Thus R(v h since s does not have an S-occurrence in the i th line, the s must be a proper suffix of an S-occurrence rs in that line. Case IIa: r is a proper suffix of u h , i.e., u rs is an S-occurrence in line i preceded by a character occurrence of the first line. (The rightmost character occurrence of r 0 cannot be ' since it is part of u h , and cannot be vital since it is not part of the S-occurrence rs.) It follows by Theorem 2.3 that rs 2 S L . Taking apply (3) getting R(v h )s 2 S 0 . Since Case IIb: r is not a proper suffix of u h . Then . Here we apply (5) to get completing Case II. Case III: E(v h cannot be a substring of s, we get E(s). Consider the character occurrence b immediately to the left of the leftmost character occurrence of s. It must be an ', for otherwise it would be vital and would have to be part of the S-occurrence s. Furthermore, it must be the rightmost ' in the v h , for otherwise, there would be an ' inside of the S-occurrence s. Therefore, we can assume that the (i st line is x y, where v th line is x 0 u h z 3 y. Case IIIa: all the character occurrences of the u h are vital in the i th line. Then, for some x 00 ; x 000 where x the x 000 u h z 3 is an S-occurrence in the i th line. By the inductive hypothesis, x 000 u h z 3 2 S 0 , and by (5), Case IIIb: not all the character occurrences in the u h in the i th line x 0 u h z 3 y are vital. Then u all the character occurrences in the u 00 are vital but the rightmost character occurrence in the u 0 is not vital. This character occurrence is not an ', so it must be a character occurrence of the first line of the derivation. It follows that u 00 z 3 is an S-occurrence, since all its character occurrences are vital but it is not adjacent to a vital character occurrence in the i th line, either on the left or on the right. Moreover, by Theorem 2.3, u 00 z 3 2 S L . Thus Case IV: This case is left-right symmetric to Case II. (It is obtained from Case II by simultaneously interchanging ! and ? and interchanging B and E.) Case V: This case is left-right symmetric to Case III.2 We illustrate (3) of Theorem 2.7 by the Thue system Taking Hence by (3) we get derivation with an S-occurrence of ga is: bc baa'e h'ga'e We illustrate (5) by the Thue system Assume for the moment that abcd 2 S 0 , and take by (5) we get derivation with S-occurrences of abcd, ah and jd is h'abf Definition (ABC property). A word w has the ABC property if there exist words A; B; C such that Theorem 2.8. Every s 2 S has the ABC property. Proof: Theorem 2.7) we can complete our proof by showing that every s 2 S 0 has the ABC property. To this end we note that (1) if s is a v-section then we can take For (2), if s 2 S L (S R ) then we can take For (3), if then we can take y. For (4), if then we can take For (5) we should demonstrate that if xu h y has the ABC property then both it. Accordingly, we assume that xu h We shall prove that s 1 has the property, leaving the similar proof for s 2 to the reader. A division into cases is required according to whether or not x is a prefix of A 0 and, if not, whether or not it is a prefix of A 0 B 0 . Case I: A which satisfies the requirement, since Ax Case II: which satisfies the requirement, since A is the same as A 0 , B is no longer than B 0 and x 2 Case III: y, we have C y. We now take 2 . Clearly the requirement, since they are the same as A 0 , x 0 It remains to prove x 2 C 2 S L . Note first that So, by Theorem 2.4, x 0 our proof is complete.2 Theorem 2.9. If S is infinite then either S L or SR is infinite. Proof: If S is infinite then fjsjjs 2 Sg is an unbounded set of lengths. By Theorem 2.8, each accord with the ABC property. Since fjB s jjs 2 Sg is bounded, either fjA s jg or fjC s jg is unbounded. But each A s (C s ) is a factor of a member of SR (S L ). It follows that either SR or S L is infinite.2 3. The algorithm. This section presents the algorithm to determine whether a given semi-Thue system with an inhibitor is uniformly terminating, 1 using the analysis of the preceding section. At the conclusion of this section, it is proved that the halting problem and the derivability problem are also decidable. The algorithm will begin by constructing finite automata for the languages S L and SR . Definitions related to finite automata. A nondeterministic finite automaton (automaton for short) is a finite directed graph each of whose arcs has either a letter or the symbol - as a label; one node is designated as the initial node and any number of nodes are designated as accepting nodes. A walk through the graph is a sequence where the N i 's are nodes and the A i 's are arcs, such that for each i, A i goes from N i\Gamma1 to N i . This walk goes from N 0 to N p . The word spelled out by this walk is the result of deleting all -'s from the word a 1 a 2 \Delta \Delta \Delta a p where, for each i, a i is the label of N i . In particular, the word spelled out is the null word itself if a i. The language of an automaton is the set of all words spelled out by walks from the initial node to an accepting node. We begin by constructing, for the finite language fL(v h )j1 - h - kg, a loopless finite automaton G 0 (1) exactly one initial node N I with no arc entering it; (2) exactly one accepting node N T with no arc leaving it; (3) exactly one path from N I to N and exactly one path from N to N T , for each node N other than N I and N T ; (4) nodes N Lk such that, for each i, the path from N Li to N T spells out L(v i ); in particular, if L(v i there is simply a lambda arc, i.e., arc labeled -, from N Li to N T ; and (5) for each i, a lambda arc from N I to N Li . The construction is done so that there are no nodes in G 0 other than those required by (1)-(4), and there are no lambda arcs except those explicitly mentioned in (4) and (5). From G 0 we construct a finite automaton G L for the language S L by repeating the following step as often as possible: for a pair of nodes N; N 0 and an integer I am grateful to Friedrich Otto for pointing out a defect in a previous version of this algorithm in [13]. Figure 1: G 0 Figure 2: G L if there is a walk from N to N 0 spelling out u i , insert a lambda arc from N to N Li , provided that there is not one already there. Note that G L and G 0 , having the same set of nodes, differ only in that G L has certain lambda arcs that G 0 does not have. It follows that the construction step is repeated only finitely many times and the graph G L so constructed is a nondeterministic finite automaton. As an example, G 0 and G L for the semi-Thue system are shown in Figures 1 and 2, respectively. Thus S The automaton GR for the language LR is like G L except that N I and N T are interchanged and all arrows are reversed. More explicitly, we first construct G 0R for the same as (1),(2),(3); Rk such that, for each i, the path from N I to N Ri spells out R(v i this is simply a lambda arc; and each i, a lambda arc from N Ri to N T . The construction is done so that are no nodes in G 0R other than those required by (1 0 )-(4 0 ), and there are no lambda arcs except those explicitly mentioned in (4 0 ) and (5 0 ). From G 0R , GR is constructed by repeatedly finding N , N 0 and i such that there is a walk from N to N 0 spelling out u i ; and then inserting a lambda arc from N Ri to N 0 . GR is similar enough to G L so that we can carry through detailed reasoning about G L knowing that corresponding things about GR will also follow. Theorem 3.1. The language of G L (GR ) is S L (S R ). Proof for G L : With Theorem 2.4, the proof that S L is a subset of the language of G L is straightforward and is left to the reader. For the converse let A q be the lambda arcs in G L other than those of in the order in which they are added in the construction. For each h, 1 - h - q, let G h be the graph that results from G 0 by adding the lambda arcs A . Thus for Where A h+1 goes from N to N Li , there is a walk in G h from N to some node N 0 spelling out u i . Let P (h; n), for q - h - 1 and n - 0, be the following assertion: For all i, spelled out by a walk in G h from N Li to N T in which the lambda arc A h occurs at most n times then w 2 S Li . Our objective will be to prove that P (q; n) is true for all n. First we note that is true, since for each i there is only one walk in G 1 without the arc A 1 from N Li to N T , which is a walk in G 0 , and that walk spells out the word L(v i Next we prove that, for each h - q and n, n) and let w be spelled out by a walk W in G h from N Li to N T in which A h occurs times. Let occurs n times in W 1 but does not occur in respectively. Assume A h goes from node N to N Lg . By the construction of G h from G h\Gamma1 there is a walk W 3 in G h\Gamma1 from N to N T spelling out a word u g y for some y. The walk W 1 W 3 spelling out w 1 u g y has only n occurrences of the arc A h . Thus P (h; n) implies w 1 u . The walk W 2 from N Lg to N T has no occurrences of the arc A h . Hence w 2 2 S Lg , which, by Theorem 2.6, implies From this it follows by mathematical induction that, for all n and h - q, P (h; implies P (h; n). But, for all h ! q, P (h; n) for all n is equivalent to P Putting all this together we get the proposition For all n; P (q; n) which, by Theorem 2.5, clearly implies that the language of G is included in S L , concluding the proof that the language of G L equals S L . The proof that the language of GR is SR is similar.2 Next it is proved that certain derivations in the semi-Thue system can be obtained from certain walks in G L and GR . In particular, loops in the automaton graphs will yield derivation loops in the semi-Thue system. We confine our attention to G L , knowing that corresponding results about GR will also be valid. We refine our consideration of the algorithm in obtaining the graph G L from G 0 , considering the sequence as defined in the proof of Theorem 3.1. Definition (w 0 (N)). For any node N other than N I , w 0 (N) is the word spelled out by the unique walk from N to N T in G 0 . Theorem 3.2. For each i, 0 - i - q, and for any two nodes N and N 0 in G i other than N I , if there is a walk from N to N 0 spelling out a word x, then z. If the walk has at least one lambda arc then some z. These derivations are obtainable effectively. Proof: We begin by proving the first sentence by mathematical induction on i. That sentence is clearly true for 0: in this case w 0 since the only relevant walks are segments of the walks from the N Li 's to N T , which are disjoint from one another. We now assume it is true for inductive hypothesis) and prove it is true for i + 1. This proof is itself by mathematical induction on the length of the walk from N to N 0 . The proposition is clearly true when this length is 0. We assume it for walks of length e (the e inductive hypothesis) and we prove it for walks of length e + 1. Thus let be this walk in G i+1 , and let the word x a be the word spelled out by it, x e being the word spelled out by the walk N of length e. By the e inductive Case I: a 6= -. Then the arc from N e to N e+1 labeled a is on the walk in G 0 from N e to N T . Consequently, w 0 (N e Case II: h, and (whether the lambda arc from N e to N e+1 is the new lambda arc of G i+1 or one already in G i ) there is a walk from N e to N T in G i spelling out a word u h z 0 , for some z 0 . By the i inductive hypothesis, . Thus we have, for some z 000 : since x This concludes the proof of the first sentence in the statement of our theorem. The second sentence follows from the fact that a lambda arc in the walk causes Case II to apply, insuring that the derivation has at least one step. Clearly, all these derivations are obtained effectively.2 Definition (loop derivation). If w . y and y . + y then we say the system has a loop derivation on y from w. Note that a loop derivation provides us with one kind of infinite derivation: from w we get the infinite derivation From the proof of Theorem 3.2 we also get Theorem 3.3. If node N is on a loop in G L then there is a loop derivation on in the semi-Thue system. (It should be mentioned that a loop does not imply that S L is an infinite set. It is possible for the loop to consist entirely of lambda arcs and for S L to be finite. For example, this happens with the system with the two rules (a; b'c), (b; a'c), for which bg. Following the constructions we get a . + a, which we could have inferred in this simple example from the observation that a ! 2 a'c'c.) Theorem 3.4. If s 2 S and jsj ? (2k 1)m then there is a loop derivation from s. Proof: That s has the ABC property (by Theorem 2.8) and that jsj ? (2k together imply that either jAj ? km or jCj ? km. Case I: jCj ? km. Since x 2 C 2 S L , there is a walk in G L spelling out C ending at N T . But, by the construction of G L , all words spelled out by loop-free paths in G L have length - km. Thus there is a loop in the walk spelling out C, and so by Theorem 3.3 there is a loop derivation on some suffix of C, and hence a loop derivation from s. Case II: jAj ? km. The proof is similar, using the graph GR for SR :2 Theorem 3.5. There is an algorithm that determines whether both G L and GR are without loops and, if so, enumerates the finite set S. Proof: By Theorem 3.1, S It is easy to tell whether both graphs are without loops. Assume now that they are without loops. Then the sets S L and SR are finite and, by Theorem 2.9, so is S. S L and SR are enumerable from G L and GR . Let S 0 be the smallest set satisfying (1)-(4) of Theorem 2.7. From the finite enumeration of S L and SR we can enumerate S 0 . Noting that S is the smallest class that contains S 0 and is closed under (5) of Theorem 2.7, let us recursively define the sets S i+1 , for all i - 0: Clearly, S i+1 is computable from S i ; and . But since S is finite by Theorem 2.9, in computing the successive S i 's, eventually we shall reach an i such that which implies that, for this i, Theorem 3.6. If a semi-Thue system with an inhibitor in which S is finite has an infinite derivation then there exists an s 2 S such that s . Proof: Since there is an infinite derivation, there is an infinite S sequence Theorem 2.1. Because S is finite there must exist p and q, q ? p, such that s Theorem 3.7 (Main Theorem). There is an algorithm that produces either a loop derivation in a given semi-Thue system with an inhibitor, or the information that the system is uniformly terminating. Proof: The algorithm begins by constructing the automaton G L for S L . If G L has a loop then from that loop a loop derivation is effectively determined, by Theorem 3.3. If G L has no loop, the analogous automaton GR for SR is constructed. If GR has a loop, analogously a loop derivation is effectively determined. If neither G L nor GR has a loop then, by Theorem 3.5, S is finite and can be enumerated. The . relation on S is computed and, from this, the . + relation on S. If there is an s 2 S such that s . + s then we have a loop derivation, effectively. Otherwise, by Theorem 3.6, the system has no infinite derivation.2 Corollary 1. The uniform termination problem for semi-Thue systems with an inhibitor is decidable. Corollary 2. If a semi-Thue system with an inhibitor has an infinite derivation then it has a loop derivation. In studying the complexity of the algorithm of Theorem 3.7, we assume that the expression T naming the semi-Thue system is simply the list of its rules, The following assertions should be clear to the reader: The automaton G 0 is constructible in polynomial time. Each G i+1 is constructible from G i in polynomial time. Since all the automata G have the same set of nodes and each G i+1 is obtained from G i by adding an arc, q is bounded by a polynomial in the number of these nodes. Thus the construction of the nondeterministic finite automaton G L is accomplished in polynomial time, and similarly for GR . It is possible to determine in polynomial time whether G L (GR ) has a loop, and if so to produce the loop derivation in polynomial time. However, I cannot prove that the enumeration of S L (S R ), if it is finite, can be done in polynomial time, since jS L j (jS R j) may be exponential in the size of G L (GR ). Thus if G L or GR has a loop, the main algorithm produces a loop derivation in polynomial time and terminates. But if neither G L nor GR has a loop then the main algorithm has to enumerate the finite set S. Since there is no polynomial bound on jSj for those S that are finite, the algorithm as written in the proof of Theorem 3.7 is not a polynomial-time algorithm. However, this does not imply that the following has a negative answer: Open question 1. Is there a polynomial-time algorithm for the problem of whether a given semi-Thue system with an inhibitor is uniformly terminating? This section closes by settling the two remaining problems of Section 1, the halting problem and the derivability problem for semi-Thue systems with an inhibitor. Theorem 3.8. The halting problem for semi-Thue systems with an inhibitor is decidable. If such a system has an infinite derivation from a word w then it has a loop derivation from w. Proof: For a given w and be the set of derivations from w of length i, and let\Omega i be the set of all S-sequences that can be taken from those derivations (as in the proof of Theorem 2.1). Note that \Delta 1 and each \Delta i+1 and\Omega are readily computable from \Delta i and\Omega i . Consider three possibilities: (1) For some i ? 1, ;. Then there is no infinite derivation from w. (2) There is an i and an S-sequence in\Omega i with a repeated S-expression. Then there is an loop derivation from w. (3) There is an i and an S-sequence in\Omega i with an S-expression s such that m. Then by Theorem 3.4 there is a loop derivation from s, and hence a loop derivation from w. If any of these possibilities occurs, then we have the answer to the question after a finite amount of time. It remains to prove that one of them must occur. Theorem 1.1 can be used to prove that if there is no infinite derivation from w then the set of lengths of the derivations from w has an upper bound and possibility (1) will occur. If there is an infinite derivation from w then by Theorem 2.1 there is an infinite S-sequence \Delta. The proof of that theorem makes it clear that w . s 1 , and hence w . s i , for all i. If that S-sequence has a repeated element then possibility (2) will occur. If not there will be no bound on the length of the elements occurring in that S-sequence and possibility (3) will occur.2 Theorem 3.9. The derivability problem for semi-Thue systems with an inhibitor is decidable. Proof: We define I(w) to be the number of iotas in the word w. Where is a derivation and be the weight of the st step, which equals I(v), (u; v) being the rule used. Let the weighted length of a derivation be the sum of the weights of all the steps of the derivation. Given x and y, any derivation of y from x must have a weighted length of I(y) \Gamma I(x). The algorithm that decides whether y is derivable from x simply enumerates all derivations from x whose weighted length equals I(y) \Gamma I(x). Because all weights are positive, no line z such that I(z) ? I(y) can be part of a such a derivation. Consequently, the list of such derivations can be enumerated readily. Finally, y is derivable from x if and only if the last line of one of these derivations is y:2 Open question 2. Does there exist an algorithm for the following problem: Given a semi-Thue system with an inhibitor and words x and y, does x . y hold? 4. Well behaved derivations. 2 We now turn our attention to semi-Thue systems without an inhibitor, with an emphasis on those having only one rule. Some derivations in these systems turn out to be like those in systems with an inhibitor. Definition (inhibited rule, inhibition system). If u a rule of a semi-Thue system T without ' then is an inhibited rule of T . (v 0 or v 00 can be the null string.) The inhibition system of T is the semi-Thue system whose rules are all the inhibited rules of T . An immediate consequence of this definition is Theorem 4.1. If x is a finite or infinite derivation in the inhibition system of the semi-Thue system T then, where each x 0 i is x i with all "s erased, is a derivation in T . Definition (well behaved, ill behaved). A derivation D in a semi-Thue system T without ' is well behaved if there is a derivation in the inhibition system of T from which D is the result of deleting all "s. Otherwise D is ill behaved. From Theorems 4.1 and 3.7 we get Theorem 4.2. There is an algorithm that produces, given a semi-Thue system without ', either a well behaved loop derivation in the system or the information that the system has no well behaved infinite derivation. Example 1. The inhibition system T 0 of the system T whose one rule is (cb; bbcc) has five rules: This section is based on material from [13]. This system T 0 has an infinite derivation. In fact, all we need for this infinite derivation is the one rule (cb; bb'cc). The infinite derivation is based on the following loop of length 2: ccb cbb'cc bb'ccb'cc (In this and the examples to follow, the part of the line with an underscore is the occurrence of u that is rewritten as the v in the next line, which has an overscore.) The th line in the infinite derivation is (bb') st line is Accordingly, the original system also has an infinite well behaved derivation based on the loop ccb cbbcc bbccbcc The (2n) th line of this infinite derivation is (bb) st line is Example 2. The following is an ill behaved derivation in the system with the one rule (ccb; bbccc): ccccbb ccbbcccb bbcccbcccb bbcbbccccccb bbcbbccccbbccc bbcbbccbbcccbccc To prove that this derivation is ill behaved we note that the inhibition system has six rules, whose right sides are, respectively, 'bbccc, b'bccc, bb'ccc, bbc'cc, bbcc'c and bbccc'. Thus the second line in the corresponding derivation in the inhibition system has six possibilities. It is left to the reader to verify that in each of these six cases the ' will inhibit the replacement of the occurrence of bbc in one of the lines below. (For example, if the ccb in the first line is rewritten as bb'ccc in the second line then the ccb in the fifth line cannot be rewritten.) However, the first five lines of the above derivation form a well behaved derivation, which can be verified by considering the following derivation in the inhibition system: ccccbb ccbbc'ccb 'bbcccbc'ccb 'bbc'bbcccc'ccb If all the "s from this derivation are deleted, what remains are the first five lines of the above derivation in T . The system with the single rule (ccb; bbccc) has an infinite derivation. This is clear from the first five lines of our original derivation, which shows that ccccbb . 4 ccccbb The sixth line of this infinite derivation is the sixth line of the above derivation, which shows that the infinite derivation is ill behaved. This system has no infinite well behaved derivation. To verify this fact we can refer to Theorem 4.1 and prove that its inhibition system has no infinite derivation. Using the algorithm of Section 3 for this is tedious as it involves enumerating S. Rather than do this we work with a superset of S: 3g. Then it is rather simple to verify using Theorem 2.4 that Thus in particular, for each rule (u; v), we have L(v); R(v) 2 T . Next, we use Theorem 2.7 to verify that by proving the following: (1) every v-section is in TT ; (3) (since all rules have the same left side ccb) if then R(v)z 2 TT (for all right sides v); Parts (1) and (2) are clear. For (3), yz 2 S L implies . The reasoning for (4) is similar to the reasoning for (3). For (5), implies that . So, a fortiori, From L(v); R(v) 2 T , we then get xL(v); R(v)y 2 TT . Having proved that S ' TT , we complete the proof that the inhibition system has no infinite derivation by proving there is no infinite S sequence (invoking Theorem 2.1). The proof is by contradiction. Assume there is an infinite S sequence s Then each s h with and 3. The following must be true: nothing could be derived from u h .) (b) Either (Part (b) can be verified by first noting that there is only one occurrence of ccb in . Thus if s h ! w, then w is a word resulting from by placing a single ' anywhere between the two slashes and deleting the slashes; s h+1 is either the word to the left of the ' or the word to its right. If to the left then if to the right, k Taking , we have by (b) either 1. By (a) again we get k Where s we get by (b) either j By (a) we then see that s 4 cannot exist, which completes our proof that the system with the one rule (ccb; bbccc) has no infinite well behaved derivation, ending our discussion of Example 2. In [13] there is a much more expeditious algorithm for the problem of whether a given one-rule semi-Thue system has an infinite well behaved derivation. That algorithm does not require consideration of the inhibition system of the given system, but involves a structural analysis of one-rule systems that is well outside the purview of this paper. That structural analysis having been established, the proof in [13] that the one-rule system (ccb; bbccc) has no infinite well behaved derivation takes one quarter the space used in the proof given above as part of Example 2. This example generalizes. Zantema and Geser [16] prove that a system with one rule in which (1) p ? n, either p is a multiple of n or q is a multiple of m, has an infinite derivation. In [13] it is proved that such a system in which either 2m has no infinite well behaved derivation. From these two results it follows that any such system in which either 2n in for i an integer - 2 or else 2m ? in has an infinite ill behaved derivation but no infinite well behaved derivation. Interestingly, it is proved in [16] that there is no infinite derivation at all in any of the following cases: (1) q is not a multiple of m, is not a multiple of n. (Senizergues [14] has extended these results of Zantema and Geser.) In [13] it is proved that if p - 2n and q - 2m then there is an infinite well behaved derivation. The two examples discussed in this section illustrate the distinction between well behaved infinite derivations and ill behaved infinite derivations. They are intended to suggest the importance of this distinction to the question of whether the uniform halting problem for one-rule semi-Thue systems is decidable. It is generally conjectured that this problem is decidable, and some progress has been made in proving partial results along that line. However, the question of whether the uniform halting problem for one-rule semi-Thue systems is decidable is very much open. It seems to me that further progress on this question will come only if research workers achieve a structural understanding of ill behaved derivations. There are partial results towards this end in [13]. The halting problem is open for one-rule semi-Thue systems. On the other hand, we have Theorem 4.3. The derivability problem is decidable for one-rule semi-Thue systems. Proof: Given x, y and a semi-Thue system whose one rule is (u; v) our algorithm to determine whether y is derivable from x divides into three cases according to the relative lengths of u and v: Case I, jvj ! juj; Case II, In Cases I and II, the finite set of words derivable from x can be enumerated, and the presence or absence of y in the set easily determined. In Case III, the finite set of words derivable from x whose length does not exceed that of y can be enumerated, again yielding an answer to the question.2 For a class of semi-Thue systems an interesting question is, does every semi-Thue system in the class with an infinite derivation have a loop derivation? The result in Section 3 shows that this question has an affirmative answer for the class of semi-Thue systems with an inhibitor. When restricted to well behaved infinite derivations, it has an affirmative answer for all semi-Thue systems. However, it is an open question for one-rule semi-Thue systems; which means that the question restricted to infinite ill behaved derivations is open for one-rule semi-Thue systems. --R "Three models for the description of language," "Termination of rewriting," "Topics in termination," "On proving uniform termination and restricted termination of rewriting systems," Monographs on Theoretical Computer Science Termination und Konfluenz von Semi-Thue-systemen mit nur einer Regel "Explanations to the text," "One-rule semi-Thue systems with loops of length one, two or three," "The development of formal language theory since 1956," "The uniform halting problem for one-rule semi-Thue systems: progress report," "Well behaved derivations in one-rule semi-Thue systems," ''On the termination problem for one-rule semi-Thue systems" "Probleme - uber Ver-anderungen von Zeichenreihen nach gegeben Regeln," "A complete characterization of termination of 0 --TR --CTR Matiyasevich , Graud Snizergues, Decision problems for semi-Thue systems with a few rules, Theoretical Computer Science, v.330 n.1, p.145-169, 31 January 2005
string rewriting;uniform termination;well behaved derivation;semi-Thue system;termination
594264
Efficient Algorithms to Detect and Restore Minimality, an Extension of the Regular Restriction of Resolution.
A given binary resolution proof, represented as a binary tree, is said to be i>minimal if the resolutions cannot be reordered to generate an irregular proof. Minimality extends Tseitin"s regularity restriction and still retains completeness. A linear-time algorithm is introduced to decide whether a given proof is minimal. This algorithm can be used by a deduction system that avoids redundancy by retaining only minimal proofs and thus lessens its reliance on subsumption, a more general but more expensive technique.Any irregular binary resolution tree is made strictly smaller by an operation called i>Surgery, which runs in time linear in the size of the tree. After surgery the result proved by the new tree is nonstrictly more general than the original result and has fewer violations of the regular restriction. Furthermore, any nonminimal tree can be made irregular in linear time by an operation called i>Splay. Thus a combination of splaying and surgery efficiently reduces a nonminimal tree to a minimal one.Finally, a close correspondence between clause trees, recently introduced by the authors, and binary resolution trees is established. In that sense this work provides the first linear-time algorithms that detect minimality and perform surgery on clause trees.
Introduction The regular restriction of binary resolution [14] states that a resolution step resolving on a given literal should not be used to deduce a clause containing that literal. In other words, that resolution step should not be an ancestor of such a clause in the binary derivation tree. We extend this restriction so that it applies also when a reordering of the resolutions brings such a clause below that step. We use rotations of edges in the binary tree to reorder the resolution steps, and require that the rotations neither weaken what is proved nor increase the size of the tree. If a binary resolution proof cannot be made irregular by such rotations, we call it minimal. This extension of regularity depends on sequences of rotations, and thus appears to be expensive to compute. However we characterize it with a condition that can be checked efficiently by examining the static tree. The condition is stated in terms of history paths in the binary resolution tree. Each history path tells the story of a given literal. The tail of the history path is a leaf of the tree and it tells where the literal was introduced in an input clause. The history path is said to close at the node where its literal is resolved away. When one history path closes at a node that occurs in another history path, and the two paths are disjoint, we say the first history path directly precedes the other. The precedes relation on history paths is the reflexive transitive closure of directly precedes. History paths and the precedes relation are basic in our understanding of how binary resolution trees behave when rotations are performed. We provide simple conditions on history paths that characterize when one node can be rotated below another (we say it is visible from the other) and when one node cannot be rotated from below another (we then say it supports the other.) By examining the history paths in a static tree, we can decide much about what can and cannot be accomplished by sequences of rotations. In particular, we can say whether rotations can convert a regular tree to an irregular one. A theorem prover can use this ability to screen the proofs it builds, and retain only the minimal ones. Since every clause with a non-minimal proof is subsumed by some clause with a minimal proof, this theorem prover uses minimality to lessen its reliance on sub- sumption. The runtime of full subsumption grows with the number of retained clauses, which may become very large, while detecting minimality depends linearly on the size of each proof tree, which is typically much smaller. Instead of using minimality simply as a filter, a theorem prover can convert a non-minimal proof to a minimal one, using operations defined in this paper. Consider the branch of an irregular derivation tree that makes it irregular. This branch contains the node where some given literal is resolved and, further down, contains a clause in which that literal occurs. Why resolve the literal away, only to have it reappear later? We can remove this resolution but otherwise reconstruct the branch as closely as possible to the original, with this literal appearing additionally in some of the clauses. The constructed tree is smaller and the result it proves is at least as general as the original one. We call this operation surgery, and define a second operation, splay, which rearranges a non-minimal tree so that it is irregular. Both operations run in time linear in the size of the tree. A combination of both operations will eventually reduce any non-minimal tree to some minimal one. The first section below presents the regular restriction on binary resolution trees. In that section we introduce the surgery operation for irregular binary resolution trees. Then minimal binary resolution trees are introduced. This section also discusses our rotations and the set of rotation equivalent trees, which is all of the trees that can be generated from a given tree by a sequence of these rotations. In the following sec- tion, we discuss the "precedes" relation of history paths, and use that to define the "holds" relation on nodes. We then relate holds and visibility by showing that if the nearest common descendant of two given nodes holds one of them, then that one is not seen by the other, i.e. it cannot be rotated below the other. Based on this condition, we give an efficient algorithm for deciding visibility. We show how a theorem prover can be restricted to retain only minimal proofs and disregard non-minimal ones, decreasing the number of proofs that must be considered. Then we show the splay operation for efficiently converting a non-minimal proof into an irregular one, so that surgery can further convert it to a minimal one. Thus another theorem prover can efficiently convert its non-minimal results to minimal ones if desired. In the next section we use history paths to characterize support: the condition where one node in a binary resolution tree must be a descendant of another after any sequence of rotations. Both minimality and surgery were first developed for clause trees [5]. The clause tree is a tool for developing ideas in automated reasoning, and the binary resolution tree is an efficient, compact data structure to implement clause trees. In the second last section we show the close relation between binary resolution trees and clause trees. Thus, this paper provides the first efficient algorithm for surgery in clause trees. We close with some remarks on the relation between clause trees and binary resolution trees, and related work. This paper is an extension of [13]. 2. Binary Resolution Trees We use standard definitions [2] for atom, literal, substitution, unifier and most general unifier. In the following a clause is an unordered disjunction of literals. We do not use set notation because we do not want multiple occurrences of a literal to collapse to a single literal automat- ically. Thus our clauses can be viewed as multisets. An atom a occurs in a clause C if either a or :a is one of the disjuncts of the clause. The clause C subsumes the clause D if there exists a substitution ' such that C' ' D (as sets, not as multisets). A variable renaming substitution is one in which every replacement of a variable maps to another variable, and no two variables map to the same variable. Two clauses C and D are equal up to variable renaming if there exists a variable renaming substitution ' such that Two clauses are standardized apart if no variable occurs in both. Given two parent clauses which are standardized apart (a variable renaming substitution may be required) their resolvent is the clause (C 1 - C 2 )' where ' is the most general unifier of g. The atom resolved upon is a 1 ', and the set of resolved literals is g. It is convenient to define a mapping ae of literals for the resolution operation. This is used later to define history paths. (Resolution mapping). For each resolution operation we define the resolution mapping ae from each occurrence of a literal c in each parent clause to either the atom resolved upon if c is a resolved literal, or otherwise to the occurrence of c' in the resolvent. The reader may be missing the usual factoring operation on a clause, which consists of applying a substitution that unifies two of its literals with the same sign and then removing one of these literals. This operation is not needed in binary resolution trees, however, since if a clause contains two identical or unifiable literals, both can be resolved upon whenever the clause is used in a resolution. By allowing several literals to be resolved on, instead of merging them before the resolu- tion, we have just one type of internal node in our binary resolution tree, instead of two. (De Nivelle uses resolution nodes and factorization nodes [3].) Moreover, an implementation is free to merge or factor literals if desired. Factoring may be seen as an optimization if the factored clause can be used in several resolution steps, since the factoring is done only once. A binary resolution derivation is commonly represented by a binary tree, drawn with its root at the bottom. Each edge joins a parent node, drawn above the edge, to a child node, drawn below it. The ancestors (descendants) of a node are defined by the reflexive, transitive closure of the parent (child) relation. The proper ancestors (proper descendants) of a node are those ancestors (descendants) not equal to the node itself. If T and H are nodes in a tree then path(T ; H) is the unique path from T to H . Here, T is called the tail and H the head. Definition 2. A binary resolution tree on a set S of input clauses is a labeled binary tree. Each node N in the tree is labeled by a clause label, denoted cl(N ). Each node either has two parents and then its clause label is the result of a resolution operation on the clause labels of the parents, or has no parents and is labeled by an instance of an input clause from S. In the case of a resolution, the atom resolved upon is used as another label of the node: the atom label, denoted al(N ). Any substitution generated by resolution is applied to all labels of the tree. The clause label of the root of the binary resolution tree is called the result of the tree, result(T ). A binary resolution tree is closed if its result is the empty clause, 2. a-d a:b-d-e d:b-c-e e:b-c-f-g c:a- b-b-f-g a: b-b-f-g b:f-g -a-b-e c-d e-f-g a-b-c h:-a -b -g -a-h -h Figure 1. An irregular binary resolution tree. For the binary resolution tree in Figure 1, :gg. The labels of a node N are displayed beside the name of the node and separated by a colon, e.g. the node N 4 has atom label c, and clause label a - b - b - f - g. The order between the parents of a node is not defined. If instead of labelling the internal nodes by atoms, one labels each edge with the complement of the literal resolved upon, a binary resolution tree would become a semantic tree, and the leaves of the binary resolution tree would become failure nodes of the semantic tree. Using the resolution mapping ae for each resolution operation in the tree, we can trace what happens to a literal from its occurrence in the clause label of some leaf, down through the tree until it is resolved away. Clearly if all literals are eventually mapped to the atom label of some internal node, the clause label of the root is empty. In this case by soundness of resolution, the clause labels of the leaves is an b: a-d d:a-c c:a-a-b a:b c-d a-b-c -b -a-h -h Figure 2. Surgery: Operation 5 on Figure 1. unsatisfiable set. Thus we are primarily concerned about tracing the "history" of a literal starting from its appearance in a leaf. Definition 3 (History Path). Let the nodes (N occur in a binary resolution tree T such that N 0 is a leaf whose clause label contains a literal a, and for each is a parent of N i . Let ae i be the resolution mapping from the parents of N i to N i , and let occur in cl(N i ). Suppose N n either is the root of T , or has a child N such that ae a is mapped by the resolution at N to the atom resolved upon at N . Then (N is a history path for a. The history path is said to close at N if N exists, and then N is written as close(P ). The node N n is the head, the leaf N 0 is the tail and a is the literal of P , written head(P ); tail(P ) and literal(P ), respectively. For example in Figure 1, (M for c which closes at N 4 . The two history paths for b in Figure 1, corresponding to the two occurrences of b, are (M 3 Both of these close at N 6 . The only which does not close is the one for f , which is Definition 4 (Regular). A binary resolution tree T is regular if there does not exist a node N of T and a descendant M of N , such that al(N) occurs in cl(M ). The tree in Figure 1 is irregular because al(N 1 ) is a and a occurs in cl(N 4 ). Irregular trees are never necessary. Why resolve away the a twice? One could choose to leave out the resolution at N 1 , leaving the a in the clause, do the other resolutions as necessary (not all will be necessary) and later resolve a away, as was done at N 5 . Operation 5 makes this idea more formal. A new binary resolution tree T 0 is constructed from T in which M 0 and all of its ancestors, and possibly other M i are removed. However all leaves of T 0 are also leaves of T . Thus every history path in T 0 has a corresponding history path in T . The converse cannot be true, since the history paths of T through M 0 do not exist in T 0 . A new sequence (N 0 n ) is defined in which either N 0 M i is removed) or N 0 i corresponds to N i in T . Operation 5 (Surgery on irregular trees). Let T be an irregular binary resolution tree with being the path of nodes from N 1 to the root N n , and with being the first node in this path whose clause label contains a. Let N i\Gamma1 and M i\Gamma1 be the parents of N i for be the parents of so that a occurs in the clause label of N 0 with the same sign as in cl(N Procedure all ancestors of N 0 be in T 0 . (M 0 and all of its ancestors are removed.) Let N k be the node at which some history path P a for a in N j closes, if k exists. For do If there is a history path in T 0 containing N 0 for which the corresponding history path in T closes at N i then Put M i\Gamma1 and all its ancestors into T 0 . i as a new node in T 0 which is the child of N 0 Let the resolution at N 0 i be defined so that the history paths of i that close at N 0 i correspond to history paths of T that close at Let the history path corresponding to P a also close at N 0 k . (That is, the same literals as far as possible are resolved at N 0 i as at N i with one possible addition.) else There is no history path containing N 0 which corresponds To a path that closes at N i . . (M i\Gamma1 and all of its ancestors are removed.) endif endfor Note that if the occurrence of a in N j is never resolved away, i.e. its history path continues to the root, then the two literals corresponding to a may occur in the root of T 0 . However, they are both at least as general as a. Figure 2 shows the effect of surgery on Figure 1. Surgery is performed at using N 4 as N j . M are not needed in T 0 . By insisting that both occurrences of a are closed at the same node, we ensure that does not have a in its result when T does not. Thus N 0 5 closes both history paths for a from N 0 Theorem 6. Let T be an irregular binary resolution tree on a set S of clauses and T 0 is constructed by Operation 5. Then T 0 is also a binary resolution tree on S, T 0 is smaller than T and the result of T 0 subsumes the result of T . Proof. We use the following lemma: If clauses C 1 and C resolve to give clause C 2 subsumes C 1 then either the resolution of C 2 and C is not possible and C 2 subsumes R 1 , or it is possible and its result, This assumes that all literals from C resolved in the first resolution are also resolved in the second, and furthermore that literals from C 2 are resolved in the second resolution if they correspond (in the subsumption) to literals from C 1 resolved in the first resolution. Each leaf in T 0 has the same label as a leaf in T and therefore T 0 is defined on S. Also, each internal node is defined by a resolution of its parents, so T 0 is a binary resolution tree. Note that cl(N 0 subsumes cl(N 1 ) - a because cl(N 1 ) contains all the literals in cl(N 0 ) except possibly a. Using repeated applications of the lemma, it follows that cl(N 0 cl(N j ) since a occurs in cl(N j ). Then cl(N 0 the result of T 0 subsumes that of T . Since M 0 is not in T 0 and since all other nodes in T 0 are taken at most once from T , it follows that T 0 has fewer nodes than T . 2 Theorem 7 (Completeness [14]). If S is unsatisfiable there exists a closed regular binary resolution tree on S. Furthermore the smallest closed binary resolution tree is regular. Proof. If S is unsatisfiable, there exists a closed binary resolution tree [11]. If it is irregular, apply Operation 5 repeatedly until it is regular. This process must terminate since the tree is smaller at each step. If the smallest closed binary resolution tree is not regular, surgery can be applied to it, making a smaller closed tree. 2 Figure 3. A binary tree rotation 3. Minimal Binary Resolution Trees A rotation of an edge in a binary tree is a common operation, for example with AVL trees [1]. Before we apply it to binary resolution trees, we review the operation on binary trees. Given the binary tree fragment on the left of Figure 3, a rotation is the reassignment of edges so that the tree on the right of Figure 3 is produced. The parent C of becomes the child of E and the parent B of C becomes the parent of E. If E has a child in T , then C takes that child in T 0 . In other words, the edges (B; C), (C; E) and (E; F ) if it exists, are replaced by the edges (B; E), (E; C) and (C; F ) if necessary. Operation 8 (Edge Rotation). Let T be a binary resolution tree with an edge (C; E) between internal nodes such that C is the parent of E and C has two parents A and B. Further, suppose that no history path through A closes at E. Then the result of a rotation on this edge is the binary resolution tree T 0 defined by resolving cl(B) and cl(D) on al(E) giving cl(E) in T 0 and then resolving cl(E) with cl(A) on al(C) giving cl(C) in T 0 . Any history path closed at C in T is closed at C in T similarly any history path closed at E in T is closed at E in T 0 . Also, the child of E in T , if it exists, is the child of C in T 0 . A rotation may introduce tautologies to clause labels of internal nodes. For instance, if al(C) occurs in cl(D) then cl(E) in T 0 may be tautological. However the clause label of the root is not changed (Corollary 11). We prove a slightly more general result first, which is also used later. Definition 9. Let T 1 and T 2 be two binary resolution trees defined on the same set of input clauses. Then T 1 and T 2 close history paths similarly if there is a one-to-one and onto mapping - from nodes in T 1 to those in T 2 , such that: 1. If N is a leaf then -(N) is a leaf and both are labeled with instances of the same input clause. Thus there is a natural one to one cor- respondence, from literals in cl(N) to those in cl(-(N )). Moreover this mapping of literals provides a mapping from history paths in T 1 to those in T 2 , defined so that they start from the same literal in the input clause, up to variable renaming. We represent these other two mappings also with -. We require for all history paths P in T 1 that tail(-(P up to variable renaming. 2. For every history path P of T 1 , P closes at a node N if and only if closes at -(N ). Thus two binary resolution trees close history paths similarly if they resolve the same literals against each other, albeit in a possibly different order. Lemma 10. If two binary resolution trees T 1 and T 2 close history paths similarly, the result of T 1 and the result of T 2 are the same, up to variable renaming. Proof. Note that result(T 1 ) and result(T 2 ) are composed entirely of literals from history paths that do not close, and since the same history paths are closed in each, the same literals are not resolved away. Also the composition of mgus in T 1 and that in T 2 are unique up to variable renaming since, given a node N , the same literals are unified at N and -(N ), up to variable renaming. 2 Corollary 11. Given a binary resolution tree T with an internal node C and its child E, Operation 8 generates a new binary resolution tree and cl(E) up to variable renaming. Proof. Observe that Operation 8 produces a tree which closes history paths similarly. 2 A rotation changes the order of two resolutions in the tree. Rotations are invertible; after a rotation, no history path through D closes at C, so another rotation at (E; C) can be done, which generates the original tree again. We say that two binary resolution trees are rotation equivalent if one can be generated from the other by a sequence of rotations. For instance, the first binary resolution tree in Figure 4 is produced by rotating the edge (N 4 ; N 5 ) in Figure 1. The second tree Figure 4. From Figure 1 rotate in Figure 4 is then produced by rotating the edge (M 4 ; N 5 ). Thus both trees are rotation equivalent to Figure 1. Rotation equivalent is an equivalence relation. It is not surprising that rotation equivalent binary resolution trees must close history paths similarly, but the converse is true as well. Theorem 12. Two binary resolution trees T 1 and T 2 are rotation equivalent if and only if they close history paths similarly. Proof. Since one rotation of T 1 creates a binary resolution tree that closes history paths similarly to it, so too does the sequence of rotations creating T 2 . The converse is proved by induction on the number of internal nodes. Suppose T 1 and T 2 close history paths similarly. Then they must have the same number n of internal nodes since they have the same number of leaves. If no rotation is possible and the theorem holds. Let N be a node in T 1 with parents L 1 and L 2 that are leaves. Then in T 2 , -(N) has proper ancestors -(L 1 ) and -(L 2 ) which also are leaves, and -(N) closes only history paths with tails -(L 1 ) and -(L 2 ). We create T 0 2 by rotating edges so that -(L 1 ) and -(L 2 ) are parents of -(N ), if this is not already the case. Let C be either parent of -(N) and let A and B be the parents of C. If -(L 1 ) and -(L 2 ) are both ancestors of C then neither is an ancestor of the other parent of -(N ). But -(N) must close a history path from that other parent, contradiction. Thus the edge (C; -(N)) can be rotated, since not both A and B contain a history path closing at -(N ). This rotation reduces the total number of non-leaf ancestors of -(N ). After a finite number of such rotations, both parents of -(N) are leaves. Call this tree T 0 . Let T 1 be T 1 with leaves L 1 and L 2 deleted, and let T 2 be T 0 2 with leaves close history paths similarly since T 1 and T 0 close history paths similarly. By induction T and T are rotation equivalent. The sequence of rotations to convert 1 to T will also convert T 1 to T 0 2 which is rotation equivalent to T 2 .We focus on the set of rotation equivalent trees that do not contain an irregular binary resolution tree. Any tree in this set is said to be minimal, since surgery cannot be applied to make it smaller. Definition 13. A binary resolution tree T is minimal if no sequence of rotations of edges generates a tree T 0 that is irregular. Theorem 14. If a binary resolution tree T on S is non-minimal, there exists a minimal binary resolution tree T 0 on S which is smaller than T and the result of T 0 subsumes the result of T . Proof. If T is not minimal, apply Operation 8 and Operation 5 so that a regular tree is produced. If this tree is minimal then let T 0 be this tree. Otherwise repeat from the beginning until T 0 is defined. This process must terminate because the tree is getting smaller at each application of Operation 5. Also the old result is subsumed by the new result at each step. 2 Thus a smallest binary resolution tree is minimal. Goerdt has shown [4] that a smallest regular binary resolution directed acyclic graph (DAG) may be exponentially larger than an irregular binary resolution DAG. Thus by considering only regular or minimal binary resolution DAGs, a theorem prover may not find the smallest proof, and hence may require more inferences to prove a given theorem. However in most cases we believe that the space of minimal binary resolution DAGs is much smaller than the space of all binary resolution DAGs and that considerable time can be saved by restricting to minimal ones. (Similar arguments can be made for and against most restrictions of resolution, including set-of-support and hyperresolution.) 4. Checking Minimality Determining whether a given binary resolution tree is minimal seems to be labourious, since the straightforward application of the definition, as is done in the proof of Theorem 14, checks every possible sequence of rotations, and there can be exponentially many. In this section we give an efficient algorithm for determining visibility - which nodes can be rotated below which - so deciding minimality is efficient. Definition 15 (Visible). In a given binary resolution tree with internal nodes N and M , we say that M is visible from N , and that N can see M , if there exists a sequence of rotations such that M is a descendant of N . Otherwise M is invisible from N . Thus a node can see the nodes that can be rotated below it. Although this is a property defined in terms of rotations, it is possible to inspect a static tree, without doing any rotations, to determine visibility. Because of this, visibility can be computed in linear time. That static property, holds, is defined after one more concept, precedes. Definition directly precedes a history path Q if P and Q have no nodes in common, and P closes at some node in Q. We write P OE Q. Moreover we say P precedes Q, and there is a sequence of history paths (P directly precedes P i+1 for A history path P precedes a node N if N closes some history path Q and P OE Q. The precedes relation is the reflexive and transitive closure of directly precedes. In particular a history path precedes itself, even though it does not directly precede itself. Also note that precedes defines a partial order on the set of history paths. In most cases a rotation does not change the precedes relation on history paths. Lemma 17. Let the history path P precede the history path Q in the binary resolution tree T and let P be the images of and T respectively after a rotation of the edge (C; E) as in Definition 8. Further suppose that the head of Q is not C. Then P 0 precedes Q 0 in Proof. Let and TD be the subtrees of T rooted at A; B and C respectively. If close(P ) and head(Q) are in the same one of these subtrees, the rotation has no effect on whether P ! Q. We know that head(Q) is a proper descendant of close(P ). Also head(Q) is not C so head(Q) must be a descendant of E. Now if tail(P ) and tail(Q) are in the same subtree, then the rotation does not affect P OE Q because the node of Q which P precedes is also in that subtree. Let P 1 be the path closing at that node. Thus P OE P 1 OE Q. We have these cases, which are illustrated in Figure 5. 1. tail(Q) is in . If tail(P ) is in closes at C. After the rotation, 1 , which is one node longer, closes at C in T 0 . P 0 OE Otherwise tail(P ) is in TD . Let P 0 denote the path through B closing at C. After the rotation P 0 OE P 0 2. tail(Q) is in TB . If tail(P ) is in after the rotation P 0 OE Otherwise if tail(P ) is in TD then after the rotation, P 0 OE P 0 3. tail(Q) is in TD . Since no path through A closes at E, we know tail(P ) is not in . Thus tail(P ) is in TB . After the rotation, directly hold an internal node M of a binary resolution tree if M is the first node that occurs on both P 1 and Q 1 , (i.e. the parent of M does not occur on both.) We say that P and Q hold M if there exist history paths P 1 and directly hold M . A node N holds M if P and Q hold M and both P and Q close at N . Also, if P and Q hold M and are in a set of history paths, we say that the set holds M . The following theorem relates invisible, a property that depends on all rotation equivalent trees, to held, a property that can be checked by examining just the one tree of interest. Figure 5. Cases of Lemma 17 Theorem 19. The nearest common descendant of M and N holds M if and only if M is invisible from N . Proof. ()) We show that if the nearest common descendant of M and holds M , then after a rotation, the (possibly new) nearest common descendant of M and N holds M . Thus M can never be a descendant of N for if it were then the nearest common descendant would be M , and a node cannot hold itself. Cases in the proof of the forward direction are in Figure 6. Let F be the nearest common descendant of N and M , and let the rotated edge be (C; E), and let nodes A; B and D adjacent to it be as defined in Operation 8. Let P and Q hold M and close at F , while P 1 and directly hold M , and P 1 OE P and Q 1 OE Q. Consider the case where F 6= E, so that after the rotation F is still the nearest common descendant of M and N . By Lemma 17, P 0 Suppose M 6= E. Then after the rotation, M is still the first common node on P 0 1 , so F still holds M . Now suppose that Without loss of generality assume that P 1 contains C and Q 1 contains D. (Case 1a) If P 1 contains B then after the rotation, P 0 hold E, so F holds M . (Case 1b) If P 1 contains A then consider the path and closing at C. After the rotation, P 0 that P 0 hold E, so again F holds M . Now suppose that Consider the case (Case 2) where M is an ancestor of C so that N is either E or an ancestor of D. Since no history path can contain A and C and close at E, M 6= C. For the same reason, P and Q contain B and close at E. (Case 2a) If M is an ancestor of A then the paths that directly precede P and Q close at C and hold M . Thus C holds M . After the rotation the nearest common descendant of M and N is C, and C still holds M . (Case 2b) Otherwise if M is an ancestor of B then after the rotation the nearest common descendant of M and N is E and E still holds M . Finally consider the case (Case is an ancestor of D. (Case 3a) If N is either E or an ancestor of B then after the rotation, holds M and is still the nearest common descendant of N and M . (Case 3b) If N is C or is an ancestor of A, then consider path R with head at B which closes at C. After the rotation, the nearest common descendant of M and N is C, while R 0 directly precedes both after the rotation. Conversely, suppose that M is not held by the nearest common descendant F of M and N . If then M is a descendant of N and therefore visible from M . If F is the child of M then the edge can be rotated, making M a descendant of F and therefore of N . Assume that there exists a path (N A A A A Case 3a Case R' R Q' Q' A A Q' Case A A A Case 1a Case R R' A A Case 2a Figure 6. Cases of Theorem 19 so that N i\Gamma1 and K i\Gamma1 are the parents of N i , for use induction on n. Without loss of generality N 0 is chosen so that al(N 2 ) occurs in cl(N 0 ); thus a history path through N 0 closes at N 2 . If there exists i in ng such that no history path through N i\Gamma2 closes at N i then the edge (N can be rotated, shown as Rotation 1 in Figure 7. We say that N i has been rotated to the side of the path(M; F ), and so M is now closer to F and the theorem holds by induction. Thus for each i in path through N i\Gamma2 closes at N i . If there is no j in ng such that N j is held by F then in particular N n\Gamma1 is not held so the edge (N can be rotated. But then the distance from M to the nearest common descendant of M and N is and the theorem holds by induction. Choose the smallest j in ng so that N j is held by F . Note that j 6= 1 by assumption. If there is a history path through K j \Gamma2 closing at N j then this path and the path through N j \Gamma2 closing at N j directly hold Thus held by F , contradicting the definition of j. Therefore no such path through K j \Gamma2 exists and the edge can be rotated as illustrated in Rotation 2 of Figure 7. If is closer to F after the rotation; we say that N j has been rotated off the top of the path(M; F ), and the theorem holds by induction. Otherwise consider the history paths through N j \Gamma2 closing at N j . If one of these includes another includes K j \Gamma3 then N j \Gamma2 is held by F after the rotation. This means that N j \Gamma2 was held by F before the rotation, contradicting the definition of j. Thus the edge N j \Gamma2 N j can be rotated. The result of this rotation is that N j will be rotated either to the side of the path(M; F ), or closer to M . By a second induction on the distance from M to N j , N i will eventually be rotated off the top or to the side. Thus the distance from M to F will decrease by the constructed sequence of rotations. 2 The sequence of rotations, constructed in the proof of Theorem 19 rotates the visible node M to below the node N that can see it. Thus if the tree were regular but non-minimal, the rotations would make it irregular, exposing the non-minimal parts. Using surgery, we can remove all non-minimal sections of a tree, and reconstruct a minimal tree from the pieces left behind. Unfortunately the number of rotations required to expose a non- minimality may be quadratic in the length of the path in the tree. Since the trees are not balanced, this is quadratic in the size of the tree in the worst case. Later we give a linear time operation, called splaying, to bring a visible node to a descendant. In the next section, however, we detect and avoid non-minimal trees. Rotation 1: to the side F F Rotation 2: off the top F F Figure 7. Rotations in the proof of Theorem 19 5. Minimality as a restriction Now we turn our attention to a theorem prover that keeps only minimal binary resolution trees. Since every non-minimal tree is subsumed by some minimal tree, this strategy uses the minimal criteria to reduce redundancy, lessening reliance on subsumption. Definition 20. Let T be a binary resolution tree. Then atoms(T fal(N)jN is a node of Tgis called the set of atoms of T . A subbrt of T is a binary resolution tree whose node set consists of a node of T called the root of the subbrt, and all of its ancestors. For a subbrt T 0 of T , is a node of T 0 and visible from the root of Tg is called the set of visible atoms of T 0 . Theorem 21. Let binary resolution tree T consist of a root node R and two subbrt's T 1 and T 2 . T is minimal if and only if 1. T 1 and T 2 are minimal, 2. no atom in cl(R) is in atoms(T ), 3. 4. Proof. Assume that T is minimal. If T 1 or T 2 were not minimal, then there would be a sequence of edge-rotations which would make the subbrt irregular. The same sequence performed on T would make T irregular as well. Hence the first condition is true. If the second condition were false, then T would be irregular immediately. Assume that the third condition is false. Then there are two nodes, whose atom labels are the same, and M is visible from R. Hence M can be rotated below R, without rotating any edges in T 2 . Now M is the descendant of N , making T irregular. The fourth condition is symmetric to the third. Conversely, assume that T is not minimal. Then there is a sequence of rotations that create an irregular tree T 0 . Node N has a descendant M 1 in T 0 such that al(N) occurs in cl(M 1 ). Since the rotations do not change cl(R), if al(N) occurs in the result of T 0 , it occurs in cl(R) in T and then T violates the second condition. Thus al(N) does not occur in cl(R) so there is a descendant M of N in T such that If M and N are in the same T i then T i violates the first condition. Assume M and N are in different T i . Since M has been rotated below is visible from N in T , and by Theorem 19 M is not held by the nearest common descendant R of M and N . Thus M is visible from R. Therefore al(M) is in vis(T i ) while al(N) is in atoms(T 3\Gammai ). 2 Any theorem prover based on binary resolution that keeps only minimal trees has already satisfied the first condition in Theorem 21 for any newly constructed tree, since only minimal trees are used in the construction. It is easy to check that the new result does not contain an atom in atoms(T ). What is left is to find is an easy way to calculate those atoms in each subbrt T i which are visible from the root of T . The idea in Procedure 22 is that a node is visible in a subbrt T i if and only if it is not held by the root of T , by Theorem 19. So for each node N we need to calculate the history paths going through it that precede the root. If some of these paths go through one parent of N , and some go through the other, then N is held by the root; otherwise N is visible from the root. In each procedure call visible(N; PN ), the first argument N is a node in the tree and we want to know whether or not it is held by the root. Initially it is one of the parents of the root, but as we traverse upward it becomes instantiated with each ancestor of this parent. The invariant we maintain is that the second argument, PN , is the set of paths that include N and precede the root. The paths through the parent of the root that precede the root are simply those whose heads are that parent, so the invariant is easy to establish in the first place. Suppose N has parents A and B. To calculate the paths PB through B that precede the root, we start with PN , and remove those paths that go through A if there are any. If there are none then PB is PN . If there is some path in PN through A, we need to add to PB any path with head at B, since these paths precede that path through A, and thus precede the root. Since we know that the paths in PN must go through at least one of the parents of N , we assume that B is that parent. Procedure 22 (Visible). Given a node N in a binary resolution tree and a set PN of history paths that precede the root of the tree, vis(N; PN ) returns the atoms labels of nodes at and above N visible from the root. Procedure vis(N; PN ) If (N is a leaf) return OE Let A and B be the parents of N and partition PN into PA and PB , which are the sets of paths which go through A and through B, respectively. Assume without loss of generality that B is chosen so that PB is nonempty. Let CA and CB be the history paths with head at A and B respectively. if PA is nonempty then //N is held return else and N is not held, so it is visible return endif The third and fourth conditions in Theorem 21 require us calculate which is done by calling vis(R i is the root of the binary resolution tree, R 1 and R 2 are the parents of R, and R i is the set of paths with heads at R i . Procedure 22 runs in a number of set union calculations which is proportional to the number of nodes in the tree. With hashing, these operations can in principle be performed in time proportional to the size of the clauses. Hence vis is a linear time algorithm, which is as fast as one could expect. We have implemented a prototype theorem prover for propositional logic. It resembles OTTER[9], but it retains only minimal binary resolution trees (so that the recursive calls in the first condition of Theorem 21 are not needed), whereas the proofs built by OTTER correspond to non-minimal trees in some cases. Our implementation actually builds binary resolution DAGs, instead of trees, to save space. This is important because bottom-up theorem provers are limited by space as well as by time. Note that Procedure 22 traverses the entire implicit tree, so it may visit a single stored node more than once. Thus its runtime is not guaranteed to be linear in the size of the DAG. The prototype includes another restriction, discussed in [6], which ensures that from each set of rotation equivalent binary resolution trees, exactly one is found. 6. Restoring Minimality There are two basic ways to restore minimality. If a binary resolution tree is irregular then the surgery operation will remove the non- minimality. For a binary resolution tree that is regular but non-minimal, some rotations need to be done to make it irregular - we call this exposing the non-minimality. This irregularity involves two internal nodes M and N that resolve upon the same atom a and neither is a descendant of the other but (at least) one is visible from the other. Suppose M is visible from N . When M is rotated below N , one parent of M becomes a descendant of N . Since a occurs in that parent's clause label, surgery from N can be done. In the proof of the converse of Theorem 19 we gave a quadratic length sequence of operations for moving visible nodes down to become descendants. Operation 24, called splaying, performs this with a linear number of rotations. It is related to the splay operation [12] for binary trees in that it brings a node closer to the root. In our case, to splay a binary resolution tree at M , we divide the descendants of M into two sets: the observers O for those that can see M , and the non-observers O for those that cannot. Then we rearrange the tree so that all the nodes in O are ancestors of M , leaving just those nodes in O as descendants of M . Thus M is brought as close to the root as possible. This is done in a way that guarantees N is an ancestor of M , so that surgery can be accomplished. The first step in the splay is to determine which descendants of M can see M , as computed by Procedure 23. Procedure 23 (Obs). Given a node M in a binary resolution tree T such that P 1 is the complete set of history paths containing both M and one parent of M and P 2 is the paths containing M and the other returns the set of proper descendants of M that can see M , called the observers of M . Procedure obs(M; if M is the root of T then return OE else Let D be the child of M and K be the other parent of D. Let C 1 be the paths from P 1 that close at D and C 2 be the paths from P 2 that close at D. Let K be the paths that contain both K and D. holds M return else if C return fDg [ obs(D; K[ else endif endif The invariant of Procedure 23 is every pair of paths one from P 1 and the other from P 2 holds M . Then if we ever find a descendant D that closes a member of each, then D holds M and so cannot see M ; otherwise D can see M . To maintain the invariant as we go from M to D, if D closes paths in P i we need to add to P i all of the paths K that come from K through D, because each one of these paths precedes a path that contains M . If D does not close any path through P i , then we have nothing to add to P i . In either case we remove from P i any path that closes at D since these do not contain D. Operation 24 uses the same code as is in Procedure 23 (but not the procedure itself) to decide how the resolutions below M should be re-ordered in the new tree. The parents S 1 and S 2 of M are considered stubs to be built upon. The observers of M are put into one of two queues Q 1 and Q 2 , and the non-observers are put into Q 3 . Along with each node put in the queue, we also put the subtree rooted at the other parent of the node, so that the resolution can be reconstructed later. Those put into Q 1 become descendants of S 1 , while those in Q 2 become descendants of S 2 . After this is done, the two subtrees are joined by a resolution corresponding to M . Finally all of the nodes in Q 3 are made descendants of M . Each of the resolutions is done so that all of the history paths closed at a node in the given tree are closed at the same node in the constructed tree. Thus the resulting tree closes history paths similarly with the given tree. Operation 24 (Splay). A splay at an internal node M in a binary resolution tree T produces a new binary resolution tree T 0 such that all descendants of M cannot see M in T 0 . Operation Initially are three empty queues. Initially P 1 is the set of history paths through M and one parent S 1 of M , while P 2 is the set of history paths through M and the other parent S 2 of M . call splay1(M; return processQ( Procedure splay1(M; if M is the root of T then return else Let D be the child of M and let K be the other parent of D. Let C 1 be the paths from P 1 that close at D and C 2 be the paths from P 2 that close at D. Let K be the paths that contain both K and D. holds M enqueue D and the subtree rooted at K into Q 3 else if C enqueue D and the subtree rooted at K into else enqueue D and the subtree rooted at K into endif endif The resolutions must be done so that the resulting binary resolution tree closes history paths similary with T . Procedure processQ(T ; Q) If Q is empty then return T else dequeue T 1 and N from Q leaving Q 1 construct T 2 by resolving T and T 1 making them parents of N return endif For example, the binary resolution tree in Figure 8 shows the result of performing a splay from the node M = N 5 on the second tree in Figure 4. The subtrees S 1 and S 2 are the leaves labelled a- b -:c, and h, respectively. The descendants of N 5 are M 4 which N 3 and N 4 are put into Q 1 , while M 4 is put into Q 2 . These end up above N 5 in the resulting tree. The only non-observer, N 6 , is put into Q 3 and ends up below N 5 . Lemma 25. If T is a binary resolution tree and M is an internal node of T , then the result T 0 of splay(T ; M) is a binary resolution tree. T 0 is defined on the same set of clauses as T ; is the same size as T and has the same result as T . Proof. T 0 closes history paths similarly with T . 2 Operation 26 (Splay Surgery). Let T be a binary resolution tree T containing a pair M;N of nodes such that al(M) = al(N) and N can see M Procedure if M is a descendant of N in T then return surgery(N) else return endif To see that a splay actually brings M below N , so that surgery is possible, suppose that the nearest common descendant of M and N is not M . Thus either N is that nearest common descendant or N and M are on different branches. N is either D or an ancestor of K at some point in splay1. We also know N can see M so it must be put into one of the queues Q 1 or Q 2 , and cannot be part of Q 3 . Thus when the tree a-d a:b-d-e d:b-c-e e:b-c-f-g -a-b-e c-d e-f-g a-b-c a: b-b-f -b g: a- b-b-f -g c:a- b-b-f-g h:-a -a-h Figure 8. Result of Operation 24 at N5 on Figure 4 is rebuilt, N is put in while we are processing either Q 1 or Q 2 . M is put in after Q 1 and Q 2 are empty, so it is below N . Returning to our example, after the splay, the node N 5 in Figure 8 is below N 1 and surgery from N 1 can be done. The result of this surgery is the same as that shown in Figure 2. Notice that in this example we also have the option to splay at N 1 and do surgery from N 5 , since N 1 is also visible from N 5 . The resulting tree would be different: it would not require the leaf :a - h nor :h, and its result would have been f . Thus surgery does not always produce a more general result, but always produces a smaller tree. Splay surgery does a linear number of set operations since it does at most n resolution steps where n is the length of a branch, and then performs surgery, which itself does a number of resolutions limited by the length of a branch. Thus it is effectively a linear time algorithm. 7. Support Often a node that is visible from another can be rotated so that it is not a descendant, but sometimes no sequence of rotations can bring a node from below another. In this case we say the node supports the other, since it is always beneath it. Support, like visibity, is a property that depends on the entire set of rotation equivalent trees. In this section we characterize support in terms of history paths, so it can be checked by examining a static tree. Definition 27 (Support). A node N in a binary resolution tree T supports a node M if after every sequence of rotations, M is an ancestor of N . Definition 28 (Tightly holds). Two history paths tightly hold a node N if there exist two sequences of history paths such that for directly holds M , and for no pair does the head of P i equal the head of Q j . A node N tightly holds M if both P and Q close at N . Tightly holds is the special case of holds, where no two paths in the sequences have the same head except the bottom two. This turns out to be the exact condition for one node being below another after every sequence of rotations. Theorem 29. A node N tightly holds M if and only if N supports M . Proof. ()) We show that if N tightly holds M before a rotation, it also does afterwards, and thus must be a descendant of M . Consider a rotation of edge (C; E) where nodes A; B; C; D and E are defined as in Definition 8. We will indicate the image of any history path P after the rotation by P 0 . At first assume that fC; Eg is disjoint from fM;Ng. By . Before the rotation all heads of paths in the sequences were distinct so they must be distinct after the rotation, unless there is a new path in either of the sequences. There is one case where such a new path occurs. (In Lemma 17, the new path introduced when tail(Q) is in .) In that case the head of new history path was not on path(M; N ). Thus the heads of paths of the new sequences (P 0 are distinct except head(P 0 Also M is still directly held by 1 so N tightly holds M . Now suppose fC; Eg is not disjoint from fM;Ng. (Case 1) If then without loss of generality let P 1 contain A and Q 1 contain B. We know that P 1 does not close at E since the rotation is possible. Suppose first that Q 1 does not close at E, as in Figure 9.1a. Then after the rotation C is directly held by P 0 1 and so is tightly held by N . Now suppose Q 1 closes at E, as in Figure 9.1b. Note that there must exist Q 2 since otherwise E holds N and and so Q 1 and P 1 both close at and then the rotation is not possible. After the A C=N A Qm Pn Q'n Case 2b A A Case 3a A A Case A A Case 1a A A Case 1b A C=N A Pm Qn Q'n Case 2a Figure 9. Cases showing tightly hold is invariant under rotations Case 1 Case 2 A A A Figure 10. Cases showing support implies tightly holds rotation, C is directly held by P 0 2 , and so is still tightly held by N . (Case close at N , and either they both include A, as in Figure 9.2a, or they both include B as in Figure 9.2b. In either case, after the rotation, N tightly holds M by the sequences (P 0 (Case then assume without loss of generality that C is in P 1 and D is in Q 1 . If A is also in P 1 as in Figure 9.3a, then there is a path with head at B and P 0 OE P 1 . After the rotation, the sequences ensure that N tightly holds M . If B is in P 1 , as in Figure 9.3b, then after the rotation the paths (P 0 and (Q 0 tightly holds M . (Case Finally if then either Pm and Q n both contain C or both contain D. If they both contain C then note that they both contain the same parent B of C, else the rotation is not possible. Then after the rotation, B is a parent of E and although P 0 are one node shorter than before, N tightly holds M by the paths (P 0 both contain D, then after the rotation there is no change to the history paths; thus N tightly holds M . We induct on the length of path(M; is the parent of N i+1 for 1. If is a parent of N and then a rotation of the edge (M; N) is possible unless there exist history paths that tightly hold M and close at N . Suppose k ? 2. First assume (case 1) that N k does not support . Then by the first half of this theorem N k does not tightly hold and so the edge (N can be rotated. Since M must remain an ancestor of N; M must be an ancestor of B in the rotation, as shown in Figure 10.1. Thus the path from M to N is shorter, and by induction tightly holds M in the new binary resolution tree. By rotating the edge back again, one sees that N tightly holds M in T , because tightly holds is invariant under rotation. As this case is covered, from now on we assume that N k supports N k\Gamma1 . Next assume (case 2) that N k does not support N k\Gamma2 . Then N does not support N k\Gamma2 , for otherwise N k supports N k\Gamma1 supports N k\Gamma2 and, by transitivity of support, there is a contradiction. By the first half of this theorem, we may rotate the edge (N k\Gamma2 ; N k\Gamma1 ). Then we may rotate the edge (N k\Gamma2 ; N k ), since N k does not tightly hold N k\Gamma2 . In this second rotation, N k must remain a descendant of N k\Gamma1 so it must be as shown in the third binary resolution tree in Figure 10.2. Thus path(M; N) is shorter, so by induction N tightly holds M . Because tightly holds is invariant under rotation, N tightly holds M in T also. From now on we assume that N k supports N k\Gamma2 . By similar arguments we assume that N k supports each of N (Case If the edge (M; N 2 ) can be rotated in T then path(M; N) is shorter in the resulting tree T 0 and by induction N tightly holds M in T 0 . Since tightly holds is invariant under rotation, N tightly holds in T . Finally if the edge (M; N 2 ) cannot be rotated then there are paths from each parent of M closing at N 2 in T . We know by induction that tightly holds N 2 and assume that (P are the paths that make this so. Assume without loss of generality that includes M . Then Q 1 must include the other parent of N 2 since N 2 is the first node that P 1 and Q 1 have in common. No matter which parent of M is in P 1 , there is a path closing at N 2 containing M and 's other parent. Call this path Q 0 . Since the head of Q 0 is M , it is distinct from the heads of the other P i and Q j in the sequences. Thus tightly holds M by the paths (Q In this final theorem, we relate the notions of visibility and support. Theorem 30. In a binary resolution tree, a node M is invisible from a node N iff there is a support S of M on path(M; N ). Proof. ()) Let D be the nearest common descendant of M and N . Consider the path sequences by which D holds M . Consider the least i and j such that head(P i and let S be the node where P i and Q j both close. These i and j must exist since the head(Pm the heads of paths before are distinct, S tightly holds M and thus S supports M . be the sequences of paths by which S tightly holds M . Since S is on path(M; N ), and S is a descendant of M , S must be an ancestor of the nearest common descendant D of M and N . Thus there are paths R such that S is on R 1 and R k closes at D. Thus D holds M via the sequences 8. Relation to Clause Trees The results in this paper were developed after we understood clause trees [5], and then primarily as a means to implement clause trees. We eventually found that most of our ideas from clause trees could be expressed in binary resolution trees. Binary resolution trees are simpler in some ways since they are easier to implement, but more often clause trees are easier to use when exploring new ideas. In particular, visibility and support can be read almost directly from a clause tree, whereas the properties held and tightly held are somewhat harder to see in binary resolution trees. We use whichever is appropriate. In this section we relate the two. Refering once again to the example, the rotation equivalent binary resolution trees in Figures 1, 4 and 8 all correspond to the first clause tree in Figure 11. The binary resolution tree in Figure 2 corresponds the second clause tree in Figure 11. A reader familiar with clause trees will note that the second clause tree in Figure 11 can be constructed directly from the first by adding a merge path between the a atom nodes and then doing surgery to remove the subtree beyond the tail of the new merge path. Some of the correspondences between binary resolution trees and clause trees are obvious. The leaves of a binary resolution tree are the clause nodes of the clause tree; both are labeled by an instance of an input clause. The internal nodes of a binary resolution tree are the atom nodes of a clause tree. A history path in a binary resolution tree, which we associate with a literal, corresponds to a labeled edge in a clause tree, also associated with a literal. If the history path is not closed, it corresponds to an edge incident with an atom node leaf; a closed history path corresponds to an edge incident with a closed atom node. _ a c d a e f _ _ a c d a _ _ _ Figure 11. Clause trees corresponding to Figures 1, 4, and 8 and to Figure 2 Minimal clause trees correspond to minimal binary resolution trees whose result does not contain two identical atoms. The minimal condition on clause trees does not allow any legal unchosen merge path or legal tautology path, including leaf to leaf paths. A minimal binary resolution tree may correspond to a clause tree with a legal unchosen leaf to leaf merge path, or legal leaf to leaf tautology path. This is because the regularity condition in binary resolution trees requires one of the nodes to be internal. Finally, visibility (resp. support) between internal nodes in a binary resolution tree correspond to visibility (resp. support) between closed atom nodes in a clause tree. Table I shows a number of the corresponding notions. Table I. Corresponding notions Clause Trees Binary resolution trees clause node leaf node internal atom node internal node edge history path open leaf atom node literal in result cl(T internal-to-internal surgery splay surgery internal-to-leaf surgery surgery merge path two history paths closing together equivalence classes of reversal equivalent equivalence classes of rotation equivalent minimal (up to leaf-to-leaf) minimal visible internal atom nodes visible support on internal atom nodes support path reversal no structural change no structural change rotation (change in derivation) merge set all history paths closing at a given node 9. Related, Past and Future Work Regularity is one of the most important restrictions and some form of it is used in many theorem proving methods related to resolution, including tableau [10], and all variants of model elimination[8]. Permuting inference steps has been investigated by Kleene [7], in the context of Gentzen's sequence calculus, both classical and intuitionistic. Kleene's permutations sometimes increase the size of the proof. It is interesting to note he defines the ancestor relation between instances of formulas in each inference, allowing him to state which instances of formulas in the deduction belong to a given instance in the end sequent. This is analagous to our notion of history paths. In the context of binary resolution derivations, de Nivelle [3], has two types of nodes: resolution nodes have two parents and factoring nodes have one. He defines four types of edge rotations, depending on the type of nodes incident with the edge. We disallow the rotation where a factorization node is parent to a resolution node, because in this case the size of the derivation must be increased. His application is to construct resolution games which are then used to show various completeness results for restrictions of resolution based on ordering literals. Both de Nivelle's paper and ours show that basic properties of resolution may be exposed by considering the set of trees equivalent modulo permutations. The main contribution of this paper is to present the minimal restriction of resolution, originally developed in terms of clause trees, using the well known proof format of binary resolution derivations. The original motivation for doing so was to implement bottom up algorithms for constructing minimal clause trees. A direct implementation, based on the structural definition of clause trees was done, but each new tree needed its own storage space and the visibility algorithm was cumber- some. To remedy both of these problems we used the notion of binary resolution derivations and implemented each new clause tree as a single storage cell with two parent clause trees. (For propositional logic this solves the space problem, since each tree can be used as part of other trees.) The surprising result is that visibility can be expressed easi- ly, and requires a linear algorithm using this data structure. We then determined to explain as much of our clause tree work as we could using binary resolution trees, to make them more accessible to readers familiar with resolution. This task turned out to be difficult until we discovered that edge rotations, history paths and the "precedes" relation between history paths are the fundamental concepts needed. Then we related visibility and support to history paths using the "holds" and "tightly holds" relations. In retrospect the authors still believe that it is easier to work conceptually with clause trees. We have anecdotal evi- dence, from a graduate course in automated reasoning given two times by the second author, once with clause trees only and once with binary resolution trees followed by clause trees, that the intuitions for support and visibility are quite understandable using either data structure. We are presently developing the full theorem provers described here, one to use minimality as a restriction and another to use splay and surgery to improve proofs after they are constructed by resolution. Redundancy elimination by subsumption is always an important consideration for theorem provers. Unfortunately the minimal restriction is not complete with full subsumption. For instance one can not refute with a minimal binary resolution tree if one resolves p against :p-q, generating q, and then uses back subsumption to remove before the resolution between :p - q and :p - :q is done. The latter resolution is part of the only minimal binary resolution refutation of these clauses. The former resolution step leads to a binary resolution tree in which p is resolved at two different nodes on one branch. However, we have discovered [6] that one can retain completeness by giving up some of the power of minimality, without giving up any of the power of subsumption. The space of minimal binary resolution trees is interesting for the following reasons: (1) it is refutationally complete, (2) it extends the well known regularity restriction of resolution (3) it contains the smallest binary resolution tree, (4) non-minimal (sub)trees can be identified in time linear in the size of the tree, and (5) non-minimal trees can be reduced to minimal ones efficiently. --R An algorithm for the organizaton of information. Symbolic Logic and Mechanical Theorem Proving. Resolution games and non-liftable resolution orderings Regular resolution versus unrestricted resolution. Clause trees: a tool for understanding and implementing resolution in automated reasoning. Bottom up procedures to construct each minimal clause tree once. On the permutability of inferences. Mechanical theorem proving by model elimination. Otter 3.0 users guide. A machine-oriented logic based on the resolution principle Extending the regular restriction of resolution to non-linear subdeductions On the complexity of derivation in propositional calculus. --TR
clause tree;resolution;binary tree;regularity;theorem proving
594272
A Full Formalization of SLD-Resolution in the Calculus of Inductive Constructions.
This paper presents a full formalization of the semantics of definite programs, in the calculus of inductive constructions. First, we describe a formalization of the proof of first-order terms unification: this proof is obtained from a similar proof dealing with quasi-terms, thus showing how to relate an inductive set with a subset defined by a predicate. Then, SLD-resolution is explicitely defined: the renaming process used in SLD-derivations is made explicit, thus introducing complications, usually overlooked, during the proofs of classical results. Last, switching and lifting lemmas and soundness and completeness theorems are formalized. For this, we present two lemmas, usually omitted, which are needed. This development also contains a formalization of basic results on operators and their fixpoints in a general setting. All the proofs of the results, presented here, have been checked with the proof assistant Coq.
Introduction This paper is concerned with programming languages semantics. Its aim is to formalize SLD-resolution in the calculus of inductive constructions (Coq). First, terms and substitutions are dened, and the unication theorem, constantly used in the denition of the semantics of denite programs, is proved from a similar theorem for quasi-terms [26]. Then, we present the formal denitions of syntactic objects used in logic programming (atoms, queries, clauses, etc) and we introduce their interpretation schemes. Last, declarative and operational semantics of denite programs are introduced, and the proofs of fundamental results of the theory of logic programming are formalized (lifting and switching lemmas, soundness and completeness theorems). The reader is assumed to be familiar with the terminology and the basic results on the semantics of logic programs, as introduced in [2, 19]. Of course, there exists an extensive literature on logic programming [3, 2, 17, 19, 28] where all these theorems are proved, and one can wonder about the usefulness of such a formalization. For this, recall that formalizing a proof rst consists in giving formal denitions of the objects used, then explicitely stating a priori the hypotheses needed, and last inferring the conclusion. Therefore, since our proofs have been encoded using the Coq proof assistant, all objects and properties are explicitely dened as well as all hypotheses needed for these proofs. Each step of a proof also explicitely results from denitions or applications of induction schemes, lemmas, or the- orems. This makes a dierence from the initial proofs, which can be found in many books on logic programming. Hence, formalizing proofs brings us at a level of detail that is often left to the reader, and we will see that if we take into account some details, often considered as minor in more informal presentations, proofs are getting a bit complicated. Furthermore, we will see that some assumptions, usually overlooked, are required in some of the theorems proved. The main diculty of this development is concerned with the denition of the variable renaming process used in an SLD-derivation, which is made explicit, thus introducing some complications, usually omit- ted. For instance, we will see that the proof of the completeness theorem is more delicate than it seems and requires, from a formal point of view, several technical lemmas. Few papers are devoted to these details; however, in [27], J. C. Shepherdson points out inaccuracy in some of the basic results in the theory of logic programming and presents a discussion about the role of standardizing apart in SLD-resolution. Let us give the most famous ex- ample, coming from this paper, and showing how a classical result can be incorrectly stated: in a famous book [19], J. W. Lloyd gives the fundamental theorem of completeness of SLD-resolution in the following form: Let P be a denite program and G a denite goal. For every correct answer for P [fGg, there exists a computed answer for P [ fGg and a substitution such that . In 1994, J. C. Shepherdson wrote about this theorem: This is not true [27] and gave a very simple counterexample. Of course, it is true that we can nd two substitutions and such that G; however, although and act only on the variables occurring in G, this does not imply that can be restricted to yield . Another typical example is the lifting lemma generally stated as follows: be a program, N a goal and a substitu- tion. Suppose that there exists an SLD-refutation of P [ fNg with the sequence of mgu's . Then, there exists an SLD- refutation of P [ fNg with the sequence of mgu's 0 n such that 0 0 is more general than No assumptions on the substitution are supposed here. However, in order to formalize the corresponding proof, which builds the derivation starting from N , we have to assume some hypotheses. Let x be a variable occurring in the domain of and in N that does not occur in N . Whereas x can be used in the renamed clauses during the derivation starting from N , it cannot be used during the derivation starting from N . Hence, in order to exactly use the same sequences of variant clauses (which is always suggested) in the two derivations, we have to suppose that no variable used for the renaming of the input clauses in the derivation starting from N occurs in N . Furthermore, for similar reasons, in order to prove 0 we have to suppose that no variable used for the renaming of the input clauses in the derivation starting from N occurs in the domain of . A similar problem happens during the proof of the switching lemma asserting that if during an SLD- derivation, two atoms L 1 and L 2 are successively selected, then they can also be selected in the reverse order and the derived states are the same up to renaming of variables. | {z } | {z } Here again, if we want to exactly use the same variants of C 1 and C 2 in the two derivations, some assumptions are required. By hypotheses (coming from the initial derivation), we have However, if we want to select L 2 before L 1 with C 2 , C 2 must satisfy But nothing in the switching lemma ensures such a property, and, depending on the standardization apart used for SLD-derivations, it is possible that a variable occurring in R 0 but not occurring in R 1 , occurs in C 2 . To avoid this problem, C 1 and C 2 must satisfy holds if we use the standardizing apart coming from [2]). Hence, formalizing a proof brings us to a greatest level of detail. Most of problems are due to variables occurring in the input clauses (variants of clauses used in the resolution steps) coinciding with other variables in the derivation. Surprisingly, another problem comes from the necessity of an explicitation a priori of all the assumptions needed to obtain a proof. Curiously, even if this requirement seems natural, it is a common practice in logic programming, to suppose, during the proofs (on paper) of most of basic results, some assumptions on the variables occurring in the clauses used, without any justication. Of course, there exists an independence result asserting that the existence of a derivation does not depend on the choice of variables in the input clauses (even if this choice is very important during the proofs), and these results are still correct. However, these properties on variables have to be formally proved from this independence result, and, furthermore, they can only express properties on the objects occurring in the theorem (for instance, we cannot suppose during a proof that a clause does not share variables with a goal built in the proof if this goal does not appear in the theorem). Standard semantics ( la Herbrand) of denite programs deals with well-known objects: substitutions and renaming. However, several non-equivalent denitions for these objects can be found in the literature, thus introducing some confusion (for example, the meaning of renaming is often assumed simpler than its formal denition implies). In [15], H. P. Ko and M. E. Nadel point out that dierent authors give subtley dierent denitions of substitu- tions. They cite six non-equivalent denitions and point out the implication for the is an instance relation, thus giving dierent versions of the lifting lemma. Therefore, substitutions are not as simple as they might rst ap- pear, and this motivates some authors, as R. S. Kemp and G. A. Ringwood in [14], to take the instance relation as primitive rather than taking substi- tutions. There is also a confusion in the literature concerning the denition of most general uniers (mgu in short). In [18], J. L. Lassez, M. J. Maher, and K. Marriot compare four non-equivalent denitions of mgu. Of course, logic programming is concerned with these problems and is also subject to such confusions. For an SLD-derivation the following denitions for the renaming process required in an SLD-deriv- ation can be found: and the confusion arises because all these conditions in fact dene dierent hypotheses. There are at least two reasons for these conditions. First, the standardization apart is needed in order to be able to unify the selected atom with the head of the variant of the clause used. The last reason is to obtain the most general possible answer (substitution). Therefore, this formalization settles down the problem of variable renaming, which has long been a cause of errors in the eld of logic programming. Description of Coq We use here version 6.2 of the proof assistant Coq [4], which allows the interactive development of formal proofs. Coq is based on a logical framework known as the calculus of inductive constructions [5], which is an extension of a typed -calculus supporting dependent types and polymorphic types. The basic idea underlying this logical framework, based on the Curry-Howard isomorphism [9, 17], is that a proof of a proposition can be seen as a functional object. For instance, a proof of a proposition of the form A ) B is a function mapping every proof of A to a proof of B. The type of this function is isomorphic to the proved proposition, so types and proposition are identied, as are objects and proofs. Furthermore, this framework allows the denition of inductive and co-inductive types (which are specied by their constructors). A classical example of an inductive type is the denition of natural numbers, which is dened in the system Coq by Inductive Three induction principles are (automatically) generated by this denition: nat_id :(P:nat->s)(P O)->((n:nat)(P n)->(P (S n)))->(n:nat)(P n) (id; s) is either (ind,Prop) or (rec,Set) or (rect,Type). A classical example of a co-inductive type is the type of innite sequences, formed with elements of type A, or streams for shorter. In Coq, it can be introduced by CoInductive Set Stream [A:Set]:=cons:A->(Stream A)->(Stream A). Note that there is no induction principle for co-inductive types. Coq also allows the denition of mutually recursive types. We shall use here two distinct sorts: Prop is the type of logical propositions; Set is the type of data type specications. The type of Prop and Set is Type. Constructing a proof within Coq is an interactive process: given a goal, the user species which deduction rule should be applied, and Coq does all the computations. The theorem prover solves successive subgoals with tactics (i.e., functions that build a proof of a given goal from proofs of more elementary subgoals). Goal apply a tactic Subgoal n 3 Terms and Unication The rst step of this development consists naturally in formalizing uni- cation of terms. The unication algorithm deals with two sets: terms and substitutions. Let us rst present their formal denitions. Given a functional signature (i.e., a countable set with an arity function ar) and a countable set X of variable symbols, terms (T [X]) can be dened by induction as follows: if f is a functional symbol in such that l is a list of terms of length n, then tf(f; l) is a term. At the same time, we introduce the dependent type L n; [X] of lists of terms of length n (T [X] and L n; [X] are mutually recursive). Substitutions (S T ) are dened as applications from X to T [X]. Recall now that given two terms t 1 and t 2 , if for a substitution s, we have called a unier of t 1 and t 2 (unif_t). s is called a minimal unier (min_unif_t), if for every unier s 0 of t 1 and t 2 , we have s s 0 (i.e., there exists a substitution s 0 such that s 0 which is denoted by the predicate less_subst_t). A unier s of t 1 and t 2 is called a most general unier if it is minimal, idempotent (i.e. 8x 2 X which is denoted by the predicate idempotent_t) and relevant to t 1 and t 2 (i.e., every variable occurring in the domain of s also occurs in the list formed by t 1 and and every variable occurring in the range of s also occurs in the list formed by t 1 and t 2 (under_lt)). These denitions are extended to lists of terms. Instead of formalizing unication on terms by following the standard approach (based on the unication algorithm), and since a similar (formal) proof, dealing with quasi-terms, has been obtained by J. Rouyer [26], we have transposed the unication property from quasi-terms to terms. Originally, quasi-terms were introduced by Z. Manna and R. J. Waldinger in order to illustrate their technique for program synthesis: in [21], an unication algorithm has been derived by proving that its specication can be satised. This proof has been veried within the Cambridge LCF theorem prover [10], by L. C. Paulson [23]. A similar proof has been obtained by J. Rouyer [26] for quasi-terms dened as follows: are used to obtain variable quasi-terms, constant quasi- terms and functional quasi-terms, ConsArg is used to build non-empty lists of quasi-terms (terms and lists of terms are merged in a single data type). For instance, the term f(x; a; g(x)) is obtained as follows More precisely, J. Rouyer has proved that if there exists an unier for q 1 and q 2 , then if these two quasi-terms are well formed, this unier, which is a quasi-substitution, is indeed a substitution. Clearly, T [X] is a proper subset of Q [X] (ConsArg(ConsArg(x; x); x) is neither a term nor a list of terms), and, in order to transpose the unication property, we have to dene a mapping f from well-formed quasi-terms (i.e., quasi-terms representing terms) to terms. However, Coq does not incorporate the notion of partiality: any function is total. The current solution we have taken in this development consists in considering simultaneously a quasi-term and the proof that it is well formed. Hence, a predicate P on Q [X] characterizing well-formed quasi-terms is dened. Depending on the type of P , two approaches are possible. The rst one corresponds to a programming language the type of P is dened recursively on the structure of the quasi-term q considered. When q is not well formed, we use the applied to Q [X] and to a proof of the proposition False obtained from q and the proof that it is well-formed. The second approach corresponds to a proofs language approach: the type of dened recursively on the structure of the proof that the quasi-term considered is well formed. In addition to the interest of the unication theorem, we are motivated by developing an approach based on the reusability of formal proofs. Therefore, this development shows how to dene partial functions. More precisely, we outline how from two inductively dened sets, a predicate P on E 1 , such that there exists a bijection between E 2 and we can dene the following functions: The two characteristics of this problem are as follows: 1 is dened by a predicate, that does not allow any (direct) denition of functions or predicates on E P 1 . Therefore, every expression of the stands for 8e 2. We just know the bijection between such that Furthermore, we would like to have The function f 21 is easy to obtain, seeing that every e 2 in E 2 is in relation with an element in E 1 . Furthermore, for every e 2 in The denition of f 12 is rather critical because of the denition, by a pred- icate, of the domain of f 12 . The technique used here consists in dening inductively for every e 1 in E 1 , the collection P[e 1 of the proofs of P logically equivalent, only P[e 1 has a computational content). Of course, for every e 1 in E P 1 , there must exist a proof of P In this way, we are in position to dene, recursively on the structure of such a proof, the function Y such that 8e 1 Hence, according to these def- initions, proofs can be viewed as objects (data structures) of a mathematical collection (inhabiting the Set type) with an explicit algorithmic content, thus allowing to proceed by induction over these proofs or to dene recursive functions on the structure of these proofs. Since E 2 is isomorphic to a subset of our aim is to map result about E 1 to a result about E 2 by proving only few conservativity lemmas. Let U 1 and U 2 be two formal properties, respectively on E 1 and E 2 (quasi-unication and unication), such that and assume that the property U 1 holds on E 1 . Given the two functions f 12 and f 21 , we can prove the property U 2 on E 2 , by the following reasoning. Let e 2 be an element in by hypothesis, we have U 1 (f 21 (e 2 )), and there exists a proof p for f 21 And the claim now follows, since . In this manner, we have proved the unication theorem without really working on unication, but by proving only few conservativity lemmas (for a full presentation of this proof, see [13]). Theorem 3.1 (Unication) Given two terms t 1 and t 2 , either t 1 and t 2 are not uniable, or they are and there exists an idempotent minimal unier relevant to these terms. 4 Objects Used in Logic Programming We present in this section the formal denitions of the objects used in logic programming, and we introduce, for each of them, an interpretation schema. First, let be a relational signature. The collection At; [X] of atoms is inductively dened as follows: if p is a predicate symbol in such that l is a list of terms of length n, then pl(p; l) is an atom. Then, queries, denite clauses, denite programs, and Horn clauses are dened from atoms by (arity ar):fun -> nat Mutual inductive ar f)) -> Term with list_term : nat -> Set := cons: (n:nat)Term -> (list_term n)->(list_term (S n)). subst := var -> Term. Inductive Unification_t[t1,t2:Term]:Set:= (cons (S O) t1 (cons O t2 nil))) (cons (S O) t1 (cons O t2 nil))) ->(Unification_t t1 t2). Inductive quasiterm : Set := V:var->quasiterm | C:fun->quasiterm | Root:fun->quasiterm->quasiterm | ConsArg: quasiterm->quasiterm->quasiterm. Table 1: Terms, substitutions, and unication. In order to be able to give an interpretation of these objects, it is necessary to attach some meaning to each of the symbols occurring in them. Therefore, let us dene the interpretations of the signature [. For this, let ' n [A] be the dependent polymorphic type of lists of elements of type A of length n, and bool the set of booleans (with its usual operators, order to dierentiate them from the operators :, ^, _ on Prop). We dene two collections of mappings: F n [A] is formed by mappings from ' n [A] to A, and P n [A] is formed by mappings from ' n [A] to bool. According to these denitions, interpretations, over a domain D, are dened, in a classical way, by I Y F ar(f) [D] I Y Hence, given an interpretation I, over D, of (resp. ), a n-ary function f predicate p) is interpreted by a mapping f I bool). We are now in position to give the following interpretation At; [X] Inductive atom : Set := pl:(p:predic)(list_term (arity_p p))->atom. Inductive request : Set := true_req:request| cons_req:atom->request->request. C; [X] Definition clause:Set:=(atom*request). Inductive program : Set := nil_pgm:program| cons_pgm:clause->program->program. Inductive horn : Set := hp:program->horn| hr:request->horn->horn. Table 2: Syntactic objects. schemes. First, we dene A X as the collection of valuations over A (mappings from X to A). Then, terms (and lists of terms) are interpreted as follows: let I be an interpretation of , over a domain D, a term t (resp. a list of terms l of length n) is interpreted by an application dened by f I (l I (v)) if l I where nil_' and cons_' are the constructors of ' n [D] and cons is a constructor of L n; [X]. Henceforth, we will nd it more convenient to write I to denote an interpretation of [ I [D] I [D]). An object clause, .) is interpreted by an application dened by I (l I (v)) r I false if (: b a I true if a I I (v)) if ri I (v) = a I (v) _ b r I (v) true if c I (v) I (v) if r I (v) Inductive [D:Set][n:nat](LIST D n)->D. I [D] Definition f_i : Set -> Set := [D:Set](f:fun)(f_n D (arity ar f)). A X Definition valuation:Set->Set:=[A:Set](var->A). Fixpoint Interp_t [D:Set;I:(f_i D);t:Term] :(valuation D)->D:=[v:(valuation D)] Case t of [f:fun][l:(list_term (arity ar f))] ((I f) (apply_l D I (arity ar f) l v)) end with apply_l [D:Set;I:(f_i D);n:nat;l:(list_term n)] (valuation D) -> (LIST D n) := [v:(valuation D)]<[n:nat](LIST D n)>Case l of (NIL D) [n0:nat][t0:Term][l0:(list_term n0)] (CONS D n0 (Interp_t D I t0 v) (apply_l D I n0 l0 v)) end. bool Inductive bool:Set:= true:bool | false:bool. [D:Set][n:nat](LIST D n) -> bool. I [D] Definition p_i :Set -> Set := [D:Set](p:predic)(p_n D (arity_p p)). Table 3: Interpretations of [ . An interpretation I, over a domain D, is said to be a model of an object (atom, query, .), if for every valuation v, over D, we have In this case, we write (depending on the interpretation schema used for a query r, we write either I r or j= I ~ r). Note that, whereas given a valuation v 2 D X , the proposition true is decidable, since D can be an innite set, the proposition I o is not necessarily decidable. This will introduce many complications in the following. To terminate, we dene the relation of logical consequence (j=) between a Horn clause and a query as follows: General interpretation schema of Definition Interp_O :(D:Set)(f_i D)->(p_i D)->O-> (valuation D)->bool:=[D:Set][If:(f_i D)][Ip:(p_i D)][o [v:(valuation D)] Case o of I end. [p0:predic][l0:(list_term (arity_p p0))] request false (orb (neg (Interp_p D If Ip l0 v)) (Interp_r D If Ip r0 v)) request true (andb (Interp_p D If Ip l0 v)) (Interp_rn D If Ip r0 v)) clause [a:atom][r:request] (orb (Interp_r D If Ip r v) (Interp_p D If Ip a v)) true (andb (Interp_c D If Ip c0 v) (Interp_P D If Ip p0 v)) [p:program](Interp_P D If Ip p v) (andb (Interp_r D If Ip r0 v) (Interp_h D If Ip h0 v)) I Definition O_valid:(D:Set)(f_i D)->(p_i D)->O->Prop:= ((v:(valuation D))(Interp_O D If Ip l v)=true). request -> Prop := (h_valid D If Ip h)->(reqn_valid D If Ip r) ). Table 4: Interpretations and models. 5 SLD-Resolution The logical approach of denite programs can be split in two parts: the declarative semantics determines what can be computed; the operational semantics (given by SLD-resolution) describes how it is computed. Execution of denite programs is based on a combination of two mechanisms: replacement and unication. This form of computing is a specic form of theorem proving, called SLD-resolution (for Selection Linear Denite). 5.1 Resolution and Transitions The resolution rule is an inference rule that denes a deductive relation, C be the denite clause C + C in the program P , is an atom, called the head of C, and C a query, called the body of C, R a query and r a renaming substitution such that no variable occurring in r(C) also occurs in R. We can write the resolution rule in the following abbreviated manner: is a mgu of R =n and r(C where R =n stands for the (n+ 1)-th atom of R (at_n_req), and where R[n r(C )] is the query obtained by replacing in R, the (n 1)-th atom by the query r(C ) ((change_n_r n (rsubst_req r (body_c C)) R)). In each resolution step, two choices are made: the choice of the selected atom and the choice of the input clause whose head unies with the selected atom. Therefore, the resolution rule can be viewed as a rule that moves from a state to another, dened as follows. First, we introduce SR as the collection of mappings from X to X (the classical properties on substitutions are extended to SR ) and a predicate P R on SR , characterizing the renaming substitutions r such that occurring in the domain of r also occurs in the range of r (r is idempotent). A resolution state is a pair :R, where is a substitution and R a query. We give now the formal denitions of the renaming conditions that have to be satised during a transition i :R i . First, the initial resolution state i :R i must satisfy i R Then, the renaming substitution r has to rename all the variables occurring in C and only these variables; this is denoted by PRC (r; C). Last, we dene two predicates satised when no variable occurring in the range of r also occurs in the domain of i (P RS in the query R i (P RR We are now in position to dene the predicate P , on the collection of transitions, by where MGU(; a 1 ; a 2 ) (at_mgu) means that is a most general unier of a 1 and a 2 . 5.2 SLD-Resolution and Derivations A nite valid SLD-derivation is a nite sequence of composable transitions satisfying the predicate also Hence, derivations can be dened inductively, from , as lists. Most of proofs of properties of SLD-resolution are obtained by induction over a derivation by using a left induction schema. However, during the soundness proof, we shall need the right one. Hence, we dene two collections of derivations as follows (D l and D r are isomorphic): D l ::= d l c (D l ; | {z } left induction | {z } right induction Two functions are dened, in order to relate these two sets. As we did for the transitions, we now dene a predicate, on D l , characterizing valid derivations (sequence of composable transitions satisfying the predicate P and an additional renaming condition). For this, we introduce the following deni- tions. Two transitions e 1 f and e 2 f are said to be composable . A function #, on D l , that computes the list of variables occurring in the input clauses of a deriva- tion, is dened. The predicate P D , on D l , can now be recursively dened by t (t) and P (t), then P D (d) Resolution states Definition state:Set:=(subst*request). (Fst and Snd are projections on a product type.) SR Definition rename:=var->var. Definition good_rename:rename->Prop:=[sr:rename] (((y:var)(x=y)->(rdom sr x)->(rdom sr y)-> (sr x)=(sr y))^((rdom sr x)->(rrange sr x))). Definition rename_c:rename->clause->Prop:= PRR Definition rename_out_req:rename->request->Prop:= PRS Definition rename_out_dom:rename->subst->Prop:= ((x:var)(rdom r x)->(s (r x))=(tv (r x))). Inductive trans:Set:=trans_cons:state->program->nat-> clause->rename->subst->atom->state->trans. Inductive rsl [ei:state;p:program;n:nat; c:clause;r:rename;s:subst;a:atom;ef:state]:Prop:=rsl_init: (le (S n) (Length_r (Snd ei))) (rsubst_at r (head_c c))) (*H4*) ->(good_rename r) (([x:var](Subst_t s ((Fst ei) x))), (subst_req s ->(rsl ei p n c r s a ef). Definition RSL:trans->Prop:=[t:trans] Case t of [a:atom][ef:state](rsl ei p n c r s a ef) end. Table 5: Transitions. D l Inductive deriv:Set := deriv_cons : deriv -> trans -> deriv. Inductive deriv_f_cons : trans -> deriv_f -> deriv_f. Inductive couple_trans_ok [t1:trans;t2:trans]:Prop:= couple_trans_ok_init:(p_trans t1)=(p_trans t2)-> (state_end_t t1)=(state_init_t t2)->(RSL t1)->(RSL t2) ->(couple_trans_ok t1 t2). Fixpoint list_var_c_d [d:deriv]:listv:= Case d of (var_cl (rsubst_cl (sr_trans t0) (c_trans t0))) [d1:deriv][t1:trans](Appv (list_var_c_d d1) (var_cl (rsubst_cl (sr_trans t1) (c_trans t1)))) end. Fixpoint Deriv_ok [d:deriv]:Prop:= Case d of ((couple_trans_ok (end_d d2) t2)^(Deriv_ok d2)^ ->(rrange (sr_trans t2) x))) end. Table Derivations. where t d stands for the last transition of d 0 , R stands for the initial query of d, and r stands for the renaming substitution used during the transition t 0 . These denitions allow a full formalization of classical properties of SLD- resolution, presented in the next section. 6 Basic Results SLD-resolution denes a relation satisfying several important properties. The logical ones are soundness and completeness; two others are particular to this form of computation: lifting and switching lemmas. 6.1 Two Classical Lemmas 6.1.1 Switching Lemma At each resolution step of a derivation, an atom and a clause must be selected. The following well-known lemma ensures that the non-determinism in the choice of atom does not matter: this is called don't care non-determinism. Lemma 6.1 (Switching) If during a valid derivation, two atoms A and B are successively selected, then they can also be selected in the reverse order, and the derived states are the same up to renaming of variables. This lemma is proved by decomposing the two transitions: s id :R 0 | {z } d s id stands for the identity substitution tv). Furthermore, note that we suppose the rst selected atom is before the second in the query R, that's why we have the assumption This proof (presented in details in [12]) is classical and requires constantly the denition of the variable renaming process used during a valid derivation. Note that we explicitely build, during this proof, the substitution relating the two derived states, thus allowing us to assert they are the same up to renaming of variables. 6.1.2 Lifting Lemma The lifting lemma is another classical lemma and is used during the completeness proof. As we said in the introduction, we have to assume some additional conditions on the renaming used, in order to prove it. Lemma 6.2 (Lifting) Let d : s id :R be a valid derivation. If for a nite list of variables l, we have dom() l, then if there exists a valid derivation s id :R , such that and for a query R f , we have R The proof is obtained by induction over the derivation (using a left induction schema). Here again, the two derived states are explicitely related (for a full presentation, see [12]). 6.2 Two Fundamental Properties Answers and solutions dene the operational and declarative semantics of denite programs. Given a denite program P , the solutions to a query (Deriv_ok d) -> (Fst (state_init_d d))=tv -> (Snd (state_init_d d))=(subst_req eta r) -> ->(rrange (sr_trans t) x))) (p_trans t1)=(p_trans (Fst (Ex [rf:request]((Snd (state_end_d d0))= (subst_req (Fst (state_end_d d0)) rf) ^(Snd (state_end_d d))=(subst_req (Fst (state_end_d d)) (subst_req eta rf)))) ([x:var](Subst_t (Fst (state_end_d d)) (eta x)))))). (Deriv_ok (deriv_cons (deriv_cons d t1) t2)) -> (le (plus (n_trans t1) (Length_r (body_c (c_trans t1)))) (n_trans t2))-> (Fst (state_init_d d))=tv -> ((Deriv_ok (deriv_cons (deriv_cons d (n_trans t3)=(S (minus (n_trans t2) (Length_r (body_c (c_trans (n_trans t4)=(n_trans ((Snd (state_end_d (deriv_cons (deriv_cons d t1) t2)))= (subst_req r (Snd (state_end_d (deriv_cons (deriv_cons d (Snd (state_end_d (deriv_cons (deriv_cons d t3) t4))))) -> (Ex [v:var] (r x)=(tv v)))))))). Table 7: Lifting and switching lemmas. (Fst (state_init_d d))=tv->(Snd (state_end_d d))=true_req (subst_req (answer d) (Snd (state_init_t (head_d d))))). Table 8: Soundness of SLD-resolution. R are the substitutions such that h p (P ) a declarative description of the desired output from denite program and query. Conversely, we have the operational counterpart of solutions: answers. If s id :R is a valid derivation, then is an answer to R. 6.2.1 Soundness of SLD-Resolution The soundness theorem expresses that every answer to a query is also a solution to this query. Theorem 6.1 (Soundness) If, from a denite program P and a query R, we get the valid derivation s id :R The corresponding proof requires the notion of free transitions that are transitions satisfying the resolution rule, except that the uniers used are not necessarily most general uniers. Hence, the condition on the initial step (condition (*H5*) of the predicate rsl) does not necessarily hold. This proof is obtained by using a right induction schema. 6.2.2 Completeness of SLD-Resolution Let us now turn to the formalization of the completeness theorem of SLD- resolution: if there exists a solution to a query, then there exists an answer to this query (in this form, the completeness theorem is weaker than the usual completeness theorem and the soundness theorem proved; we don't relate here the solution with the answer). As usual, we need some new denitions and lemmas rst: let us summarize them. First, we have to dene Herbrand interpretations with respect to our denitions. For this, we dene inductively the collection, T [;], of ground terms (Ground_Term) that are terms not containing variables, obtained using only function symbols occurring in (this set is also called the Herbrand universe of ). At the same time, the collection, L n; [;], of lists of ground terms is dened. Similarly, we dene the collection, At; [;], of ground atoms (Ground_Atom), which can be formed by using relation symbols from with ground terms from the Herbrand universe of as arguments (this set is also called the Herbrand base of [ ). We also dene S ; , as the collection of ground substitutions (Ground_subst), which are mappings from X to T [;] (which exactly corresponds to T [;] X ). Furthermore, to avoid some uninteresting complications, we assume, from now, on that the signature used, contains a constant (i.e., a 0-ary function symbol); hence the above-dened sets are nonempty (in [8], J. H. Gallier discusses in detail the need for at least one constant symbol in in the context of the Herbrand theorem). We can now introduce the denition of the Herbrand interpretation of , written H. For every f 2 , f H is a mapping from ' ar(f) [T [;]] to T [;] dened by: where tfg is the unique constructor of ground terms and where l 0 has type L ar(f) [;] and is obtained from l. Using this denition, we dene a Herbrand interpretation of [ as any interpretation over ground terms, based on H. Since, for Herbrand interpretations, the assignment to constants and functions is xed, it seems possible to identify a Herbrand interpretation I of [ with a subset of the Herbrand base, dened by a predicate (I), characterizing the ground atoms a, such that I a However, note that specications of I and (I) correspond to two dierent kinds of specications: whereas for every Herbrand interpretation I and every ground atom a, the proposition I a is decidable, for an arbitrary predicate p, on At; [;], the proposition p(a) is not. Therefore, later, we'll use in theorems, which need to get from a predicate on the Herbrand base, a Herbrand interpretation, this proposition as an assumption (we will use this assumption only as a mapping). Fixpoint characterization of the least Herbrand model and completeness of SLD-resolution will be obtained only at the cost of this strong assumption, implicitly suggested by the identication: Type. For every denite program P , we dene, by a predicate on At; [;], the least Herbrand model of P as the intersection of all Herbrand models of According to this denition, and from [28], we prove that M )g. We wish now to obtain a deeper characterization of this set using xpoint concepts. To study Herbrand models of denite programs, R. Kowalski and M. van Emden associate with a denite program P , an operator T P , called immediate consequence operator, which provides the link between the declarative and operational semantics of de- nite programs. Using the identication of a Herbrand interpretation with a predicate on the Herbrand base, this operator is usually dened as follows. Table 9: Declarative semantics. Let P be a denite program, p a predicate on At; [;], and a a ground atom. If for a clause c occurring in P and for a ground substitution such that then T P (p)(a). Since it is possible that some variables occur in the atoms a 0 , we cannot establish the decidability of T P (p), even if we know the decidability of p: given an (not necessarily ground) atom and a Herbrand interpretation, we cannot prove the decidability of the proposition this interpretation is a model of this atom, needed to evaluate atoms not necessarily ground in the body of the clause aected by T P . Since Herbrand models of a program P are exactly the xpoints of T P , we formalize some results on operators and their xpoints in a general setting. In fact, T P is easily seen to be upward continuous, and it follows that its least xpoint is T "! P and is equal to M P . us turn to the denition of the operational counterpart of the least Herbrand model: the success set of a denite program, dened by a predicate S P on the Herbrand base, characterizing the ground atoms a such that there exists an SLD-refutation starting from a. From the two technical lemmas presented in the end of this section, we prove that M P is equal to S P and, under the assumption that every predicate on the Herbrand base is decidable, we are now in position to prove the completeness theorem in the following Theorem 6.2 (Completeness) Let P be a denite program, R a non-empty query, and a solution for R. Then, there exists the following valid derivation s id :R operators A:Set Definition PA:=A->Prop. Definition OP:=PA->PA. Inductive ORD:Set:= Fixpoint ORD_FIRST [n:nat]:ORD:= Case n of So:ORD->ORD | Oo Lo:(nat->ORD)->ORD. [p:nat](So (ORD_FIRST p)) end. ordinal powers Fixpoint [a:A]False [l:(nat -> ORD)]([a:A](Ex [n:nat]((POF f (l n)) a))) end. Theorem Table 10: Operators, ordinals, and xpoints. One delicate step in the proof of the completeness theorem of SLD-resolution is concerned with the combination of derivations. This proof is made by induction over the length of the initial query, and the following argument is always used: has a refutation Because each B i is ground, these refutations can be combined into a refutation of P [ However, the combination of derivations is rather delicate, and the way it is made is never given. Therefore, during the formalization of this proof, we need a (technical?) lemma, whose proof is not so immediate and shows (in a constructive way) how the nal refutation can be obtained from the initial ones. Of course, in order to ensure the good properties on the variables occurring in the nal refutation, we have to be able to rename the initial refutations. First, we have the following variant lemma, which is seldom stated and whose proof is not immediate. It is proved, in a dierent form, in [2] and stated in [20]. Lemma 6.3 (Variant lemma) Let l be a list of variables and d 1 the valid derivation s id :R There exists a valid derivation d 2 :s id :R Inductive Tp Inductive Success [p:program;a:Ground_Atom]:Prop := (p_deriv d)=p->(Deriv_ok d)-> (Fst (state_init_d d))=tv->(Snd (state_init_d d))= (cons_req (Ground_to_atom a) true_req)-> (Snd (state_end_d d))=true_req-> (Success p a). (r=true_req)-> (semantic_csq (hp p) (subst_req s r))-> (Fst (state_init_d (Snd (state_end_d Table 11: Completeness of SLD-resolution. , such that no variable occurring in #(d 2 ) also occurs in l. Furthermore, for a substitution s r such that dom(s r ) var(R 2 ), we have R In other words, the existence of a derivation does not depend on the choice of variables in the input clauses: it suces that the good properties hold. Formalizing this lemma consists in building from a derivation d 1 starting from R and from a nite set of variables Z, a derivation d 2 starting from R such that no variable occurring in the input clauses of d 2 also occurs in Z. R | {z } Z | {z } d2 As a (not so immediate) consequence, we obtain the following lemma. Lemma 6.4 (Combination of derivations) Let P be a denite program, R be a nonempty query, and be a ground substitution. If every atom occurring in R is in the success set of P , then, there exists a valid derivation s id :R Proof. We proceed by induction over R. For the claim is obvious. For induction hypothesis, if for every atom a 0 occurring in r, we have S P (a 0 ), then, there exists a valid derivation s id :r Moreover, by hypothesis, we have By the induction hypothesis, we can build the valid derivation I :r ; . Furthermore, by (1), we have S P (a) and, by denition, there exists a valid derivation d is a ground substitution, we can prove, by induction over d 1 , that d 3 : s id :c r (a; r) I :a is also a valid derivation. Moreover, from d 2 and by Lemma 6.3, there exists a valid derivation d 4 : I :a now, by induction over d 4 , we prove that d 5 : I :a valid derivation. Hence, by combining d 3 and d 5 , which is suitable for the renaming conditions, one obtains the valid derivation s id :c r (a; r) which settles the claim. J As we said, the formal completeness proved is weaker than the soundness theorem proved: we don't relate the solution and the answer. This is due to the following usual step in the proof: be the variables of N . Enrich the language of P by adding new constants a be the substitution fx 1 =a there exists an SLD- refutation of P [ f Ng (. By textually replacing in this refutation a i by x i , for This operation cannot be formalized by using our denitions (which do not allow to add new constants in the functional signature or to textually replace constants by variables and variables by constants). A possible solution could consists in following the S-semantics approach developed in [7] which allows variables in the Herbrand base. 7 Related Work As we said, unication on quasi-terms has been formalized by L. C. Paulson [23] using LCF and by J. Rouyer [26] using Coq. Concerning the deni- tion of partial functions, C. Dubois and V. Vigui Donzeau-Gouge [6] have recently proposed a method to generate automatically the predicate P characterizing the domain of a partial function f from equations dening f in a ML-style. Formal completeness proofs have been envisaged by several au- thors. In [24], H. Persson presents a formalization, using the proof assistant ALF, of a constructive completeness proof for intuitionistic predicate logic w.r.t. models based on formal topology. In [11], J. Harrison discusses about a formalization in HOL of basic rst-order model theory including Compact- ness, Lowenheim-Skolem, and Uniformity theorems. In [16], J. L. Krivine describes a formalization in second order-logic of an intuitionistic proof of the completeness theorem of classical logic. Finally, note that in the area of formal methods applied in logic programming, and starting from an operational semantics for Prolog, the soundness and completeness proofs, for each renement step toward the Warren Abstract Machine (WAM), have been elaborated by C. Pusch [25] with the theorem prover Isabelle. 8 Conclusion In order to prove formal properties about programs, it seems natural to give a formal specication of their operational semantics. In this paper, SLD-resolution has been formalized, and fundamental properties of this form of computation have been proved (more than 600 technical lemmas have been formalized during this development). The proofs we have machine-checked are based on those in J. W. Lloyd's book [19] and then follow the traditional we have chosen to take as the referential notion of truth a semantical notion (an alternative could consist in taking another referential notion like natural deduction or sequent calculus). First, note that, instead of proving the unication theorem by following the unication algorithm, as J. Rouyer did in [26], we have used his result on quasi-terms to obtain our proof. The technique proposed is based on dening a bijection between terms and quasi-terms satisfying a given predicate and proving the preservation of the unication property. In this way, we have proved the unication theorem without really dealing with unication theory, thus avoiding a heavy proof requiring to handle sophisticated theories. This construction is interesting because it represents an alternative to the classical approach and shows how to take advantage of a previous formal proof, even if the desired proof does not exactly deal with the same objects. Whereas the unication theorem proved in [26] requires about 140 lemmas, only 75 lemmas have been proved in our development. This shows the interest that a proofs library be provided. In more informal presentations, variable renaming problems are often ignored; however, these problems cannot be ignored here. As we said, the results presented here have been proved in the calculus of inductive construc- tions, thus making a dierence from the proofs one can nd in many books on logic programming. The main dierence is concerned with the renaming process used in an SLD-derivation: it is made explicit. A typical example is the proof of the completeness theorem that requires a variant lemma (assert- ing that the existence of an SLD-derivation does not depend on the choice of variables in the input clauses used) and a lemma on combination of derivations (which explicitely allows this operation), whose proofs are rather long. Therefore, formalizing SLD-resolution introduces many complications. Even if combining some SLD-derivations seems to be easy, this operation requires several renamings in order to satisfy the standardization apart assumed in the denition of a derivation. Usually, these lemmas are implicitly used without any proof or any justication. Not surprisingly, a full formalization of these proofs shows how minor details could be crucial. An interesting discussion about these ne points, at the foundations of logic programming, can be found in [15, 27]. Of course, the main problems come from the explicitation of all the substitutions used during the unication and the variable renaming processes: substitutions are a quite hard matter to deal with. In [1], substitutions are said to be the minence grise of the -calculus, this can also be said for SLD-resolution. Like SLD-resolution for clauses; -reduction denes a relation on -terms, and it may be interesting to compare properties of these relations. Whereas -reduction is conuent, SLD-resolution just satises the switching property, which is weaker than conuence. Fur- thermore, the variables renaming process used in an SLD-derivation can be viewed as an implicit -conversion: the renaming is made explicit and allows a full formalization of SLD-derivations, along the lines of the calculus of explicit substitutions [1] for -reduction (the -calculus is a renement of the -calculus, where substitutions are manipulated explicitely, and provides a link between the classical -calculus and concrete implementations). The renaming process required in an SLD derivation plays an essential role in logic programming: it constitutes an important step in the denition of operational semantics of denite programs from the declarative reading of Horn clauses. Acknowledgments Many thanks to Ren Lalement for enlightening discussions about this work, as well as the anonymous referees for some very useful comments. --R Explicit substitutions. Logic programming. Contributions to the theory of logic programming. Project Coq. The calculus of constructions. A step towards the mechanization of partial functions: domains as inductive predicates. Declarative modeling of the operational behavior of logic languages. Logic for computer science Proofs and Types Edinburgh LCF Formalizing basic Formalization of SLD-resolution in the calculus of inductives constructions Reynolds and Heyting models of logic programs. In Koichi Furukawa Une preuve formelle et intuitionniste du thorme de compltude de la logique classique. Computation as Logic. Foundations of Logic Programming. Partial evaluation in logic program- ming Deductive synthesis of the uni Logic Programming and Prolog. Verifying the uni Constructive completeness of Intuitionistic Predicate Logic: A formalization in Type theory. Dveloppement de l'algorithme d'uni The role of standardising apart in logic programming. The semantics of predicate logic as a programming language. --TR
calculus of inductive constructions;SLD-resolution;standardization apart;formal proofs;semantics of logic programs
594273
Some Lambda Calculus and Type Theory Formalized.
We survey a substantial body of knowledge about lambda calculus and Pure Type Systems, formally developed in a constructive type theory using the LEGO proof system. On lambda calculus, we work up to an abstract, simplified proof of standardization for beta reduction that does not mention redex positions or residuals. Then we outline the meta theory of Pure Type Systems, leading to the strengthening lemma. One novelty is our use of named variables for the formalization. Along the way we point out what we feel has been learned about general issues of formalizing mathematics, emphasizing the search for formal definitions that are convenient for formal proof and convincingly represent the intended informal concepts.
Introduction "This paper is about our hobby." That is the first sentence of [MP93], the first report on our formal development of lambda calculus and type theory, written in autumn 1992. We have continued to pursue this hobby on and off ever since, and have developed a substantial body of formal knowledge, including Church-Rosser and standardization theorems for beta reduction, and the basic theory of Pure Type Systems ( PTS ) leading to the strengthening theorem and type checking algorithms for PTS . Some of this work is reported in [MP93, vBJMP94, Pol94b, Pol95]. In the present paper we survey this work, including some new proofs, and point out what we feel has been learned about the general issues of formalizing mathematics. On the technical side, we describe an abstract, and simplified, proof of standardization for beta reduction, not previously published, that does not mention redex positions or residuals. On the general issues, we emphasize the search for formal definitions that are convenient for formal proof and convincingly represent the intended informal concepts. The LEGO Proof Development System [LP92] was used to check the work in an implementation of the Extended Calculus of Constructions with inductive types [Luo94]. LEGO is a refinement style proof checker, publicly available by ftp and WWW, with a User's Manual [LP92] and a large collection of examples. Section 1.3 contains Submitted to Journal of Automated Reasoning y A version of this paper appears as technical report ECS-LFCS-97-359, University of Edinburgh. z Laboratory for Foundations of Computer Science, University of Edinburgh x Basic Research in Computer Science, University of Aarhus. Centre of the Danish National Research Foundation. information on accessing the formal development described in this pa- per. Other interesting examples formalized in LEGO include program specification and data refinement [Luo91], strong normalization of System F [Alt93], synthetic domain theory [Reu95, Reu96], and operational semantics for imperative programs [Sch97]. 1.1 Why? PTS have a beautiful meta-theory, developed informally in [Bar92, Ber90, GN91, vBJ93, Geu93]. These papers are unusually clear and mathematical, and there is little doubt about the correctness of their re- sults, so why write a machine-checked development? The informal presentations leave many decisions unspecified and many facts unproved. They are far from the detail of representation needed to write a computer program for typechecking PTS , and the lemmas needed to prove correctness of such a program. At the start, our long-term goal was to fill these gaps in order to increase confidence in proofchecking programs (such as LEGO) based on type theory. That goal is largely met in [Pol95]. Also, while the basic informal theory of PTS is well under- stood, the difficulties of formalization suggested reformulations which clarify the presentation. Another goal of the project is to develop a realistic example of formal mathematics. In mathematics and computer science we do not prove one big theorem and then throw away all the work leading up to that theorem; we want to build a body of formal knowledge that can continually be extended. This suggests some design criteria for formalization. Representations and definitions must be suitable for the whole development, not specialized for a single theorem. The theory should be structured, like computer programs, by abstraction, providing "isolation of components" so that several parts of the theory can be worked on simultaneously, perhaps by several workers, and so that the inevitable wrong decisions in underlying representations can later be fixed without affecting too seriously a large theory that depends on them. The body of knowledge we want to formalize is itself still grow- ing, e.g. [vBJMP94] reports advances on typechecking for PTS done later than our original formalization, that became part of our formal development. The work on typechecking benefited from the basic formalization of PTS , since proofs about several related systems could be easily adapted from proofs already done for PTS . Further, new subjects were included; e.g. the standardization theorem, not used in the type theory, was formalized by the first author. On the other hand, we do not claim that type theory is a realistic example for all formal math- ematics: it is especially suitable for formalization because the objects are inductively constructed, their properties are proved by induction over structure, and there is little equality reasoning. Perhaps the most compelling reason for our continuing fascination with this work is the lure of completely concrete, yet simple, proofs of results whose conventional presentation seems to require some notions that are "messy" to formalize, e.g. the standardization theorem discussed in section 3.4. We see such proofs as beautiful, both by their simplicity and their concreteness. There is a tendency in formalization to throw simplicity to the winds in frustration to get the proof to work at all; but once it is checked, it can be beautified relatively easily, as improved definitions and arguments are mechanically checked, easily pointing out new glitches and suggesting how to fix them. Also, a formal development is easy to come back to a year later, as all the details you would not otherwise have written down are explicit, and don't have to be rediscovered. 1.2 Related Work There are many formalizations of the Church-Rosser theorem [Sha85, Hue94, Nip96, Pfe92]; the only formalization of a standardization theorem we know of is [Coq96a], for lazy combinator expressions. Formalizations of type theory include [DB93, Bar96]; both of these address limited aspects of very special type theories (essentially the Calculus of Constructions), although [Bar96] is very interesting work in which the program extraction mechanism of Coq is used to extract an executable typechecker from a proof of decidability of typechecking. In contrast to all the cited work except [Bar96], our development hasn't been terminated by reaching one specified theorem, but continues to grow in various directions guided by our interests, and by other work we come across that we feel needs checking. For example, both authors have checked parts of type theory papers we were asked to referee. A novelty in our presentation is the use of named variables. Most of the formalizations of type theory or lambda calculus that we know of use de Bruijn indices ("nameless variables") [Sha85, Alt93, Hue94, Nip96, Bar96] or higher order abstract syntax [Pfe92] to avoid formalizing the renaming of variables to prevent unintended capture during substitution. While de Bruijn notation is concrete and suitable for formalization, there are reasons to formalize the theory with named variables. For one thing, implementations must use names at some level, whether internally or only for parsing and printing; in either case this use of names must be formally explained. More interesting is the insight to be gained into the meaning of binding. Many researchers agree that de Bruijn representation "really is" what we informally mean by lambda terms, in the sense that there is no need to quotient terms by alpha-conversion, i.e. intensional equality on de Bruijn terms corresponds with what is informally meant by identity of terms. Nonetheless, de Bruijn representation is a coding of the informal notion of binding, and doesn't address at all the relationship between free and bound variables, namely how free variables become bound. In our formalization, syntactic terms using named variables are themselves concrete: the names of bound variables actually occur (parametrically) in meta-formulas containing them, just as the names of free variables do. This is done using a formulation suggested by Coquand [Coq91], based on syntactically distinguishing free from bound variables 1 . Other work on formalization of binding and substitution using names includes [Coq96b, GM96, Owe95, Sat83, Sto88], but these do not work out any large examples using their binding notions. It would be interesting to compare our development with some similar example using the terms up-to alpha conversion of [GM96]. A presentation of type theory based on treating terms with named variables concretely is Martin-L-of's calculus of explicit substitutions [Tas93], but this presentation is not closed under alpha-conversion, as our presentation is (section 5.5.3), and we view this as a failure of concreteness of Martin-L-of's system. 1.3 This paper and the formal development The source files for the development described in this paper, along with a README file explaining how to check them, is available on the LEGO WWW homepage http://www.dcs.ed.ac.uk/home/lego/. LEGO uses a module system (described in [JP93]) based on Cardelli's mock modules [Car91]. Each source file is a module, and each module has a header saying which modules it depends on. Thus the directory of modules associated with this paper contains parallel, and even incompatible, developments. If you type Load strengthening, the file strengthening.l (which contains the proof of strengthening for PTS ) will be loaded, preceeded by every module it depends on 2 . This distinction is already present in Gentzen [Gen69, pages 71-2, 116-7, 141, 216-7] and Prawitz [Pra65] 2 The dependencies are determined from the module headers, not by examining There are over 70 proof source files with extension .l containing over 1500 definitions and lemmas. This is a large amount of formal knowledge, which we can only survey here. This paper uses informal mathematical notation, but almost every definition and lemma that we mention is given with its formal name in typewriter font (often in parentheses). You can then use grep to find the file in which it is defined and the files in which it is used. This is not particularly elegant, but it's how we do it too. Keeping track of a large amount of formal knowledge is a serious problem that we have not addressed very well. 1.3.1 About notation As mentioned, this paper uses informal notation, which is arrived at by manually translating from the formal LEGO notation into L A T E X. Further, the translation is not purely syntactical; we chose to surpress some technical details to have a readable presentation. Errors are quite likely, arising from both our translation and your interpretation. This paper may be an informative outline of the formal work, but if you want to believe one of our results you must read its formal statement, and all the formal definitions used in its statement; see [Pol97] for discussion of believing a large formal development. In [MP93] we used formal notation, verbatim text manually extracted from LEGO source files; no translation errors occur, but there is no reason to believe the verbatim text in the paper actually appears in the files. Indeed, the the document and the files drifted apart over time. In [Pol94b] we again used formal notation, mechanically extracting marked sections of the source files, following the idea of Knuth's WEB. We could rerun the extraction to update the document to the formal source, but many readers complained the document was as unreadable as the formal source. Presenting a formal development is a serious problem. Perhaps mechanical extraction with mechanical translation to informal notation is the right direction to pursue. For better or worse, we have sanitised this presentation so that very little purely formal detail shows through. For example, we mostly sur- press the distinction between boolean values and propositional values. However, we don't want to hide the fact that formalization requires many details that don't appear in informal presentations. the actual dependencies in the files. 3 The .l files are the ones we wrote; LEGO generates "compiled" files with a .o extension. These are the fully annotated -terms generated by the LEGO tactics called in the .l file. A few basic notations The development uses LEGO's built-in library of impredicative definitions for the usual logical connectives and their properties; we use standard notation for these connectives. Quantifiers are typed in ECC, but we reserve symbols to range over certain types, and drop the type labels almost everywhere; e.g. p will be reserved to range over parameters, PP , so we Well known computer science notations are used, e.g. if as if-then-else, list( ) for the type of lists over , @ (or sometimes just concatenation) for list append. All funtions of ECC are total (as opposed to functions in the object theory of lambda terms and PTS ), so some operations take extra arguments for a "failure value", e.g. (assoc a b l) returns b if a is not the first element of a pair occurring in l . Pure Languages In this section we discuss a formalization of the language of PTS , including terms, occurrences and substitution. We derive a strong induction principle for well-formed terms. A Pure Language ( PL ) is a triple (PP; VV; SS) where ffl PP is an infinite set of parameters, ranged over by p , q , r ; these are the global, or free, variables. ffl VV is an infinite set of variables, ranged over by v , x , y ; these are the local, or bound, variables. ffl SS is a set of sorts, ranged over by s , t u ; these are the constants PP , VV and SS have decidable equality. That PP and VV are infinite is captured by the assumption that for every list of parameters (variables) there exists a parameter (variable) not occurring in the list; e.g. We are not assuming mathematical principles, but working parametrically in types PP , VV and SS having the stated properties. These can be instantiated with particular types that provably do have these properties, e.g. the natural numbers, or lists over some finite enumeration type. By working parametrically we are preserving abstractness: only the stated properties are used in our proofs. 4 In this formula, member(p; l) is decidable because PP has decidable equality. 2.1 Terms The terms of a PL , Trm , ranged over by M , N , A , . , E , a , b , are given by the grammar atoms: variable, parameter, sort binders: lambda, pi application To be precise, Trm is inductively generated by six constructors: every term can be thought of as a well-founded tree whose leaves are variables, parameters and sorts, and whose interior nodes are lambda and pi (having three branches each) and application (having two branches). We often define functions on Trm by structural (primitive) recursion over this inductive definition. As usual, we intend [v:A]B and fv:AgB to bind v in B but not in A . However the intended binding structure is not determined by the definition of Trm , but is made explicit by the definitions of substitution and occurrence below. Equality on terms is defined by recursion over Trm ; it inherits decidability from PP , VV and SS . Remark 2.1 (Notation) Often when doing case analysis by term structure, we want to say that the binders, lambda and pi, behave the same way. We introduce a notation hv:Aia to allow combining these cases. The actual formalization does not have such a notation, but this would have saved much cutting and pasting in developing the proofs. The length of a term is used as a measure for well-founded induction Two properties of this measure are used applications: if A is a proper subterm of B then (used in induction on the length of terms), and every term has positive length (used in reasoning about PTS by induction on the sum of the lengths of the terms in a context). 2.2 Occurrences of Parameters and Sorts The list of parameters occurring in a term is computed by primitive recursion over term structure, and the boolean judgement whether or not a given parameter occurs in a given term is decided by the member function on this list of parameters. params(p) , [p] params(hv:Aia) , params(A) @ params(a) Similarly are defined. 2.3 Substitution For the machinery on terms, we need two kinds of substitution, for parameters and for variables, both defined by primitive recursion over term structure. Write [a=p]M (formally psub) for substitution of a for a parameter, p , in M . This is entirely textual, not preventing capture. Since parameters have no binding instances in terms, there is no hiding of a parameter name by a binder. [a=p]q , if(p=q; a; q) [a=p]ff , ff ff 2 VV; SS [a=p]hv:Bib , hv:[a=p]Bi[a=p]b Substitution of a for a variable, v , in M , written [a=v]M (formally vsub), does respect hiding of bound instances from substitution, but does not prevent capture. [a=v]x , if(v=x; a; x) [a=v]ff , ff ff 2 PP; SS [a=v]hx:Bib , hx:[a=v]Biif(v=x; b; [a=v]b) will be used only in safe ways in the type theory and the theory of reduction and conversion, so as to prevent unintended capture of variables. Note that these operations are total functions, and do not rename variables. Also, occurrences of a in [a= ] are shared, regardless of whether they occur within different binding scopes, in contrast to the situation with de Bruijn indices. Some important lemmas can now be proved: Vcl-atom Vclosed(ff) ff 2 PP [ SS Vcl-bind Table 1: Inductive definition of the relation Vclosed . and we have a ready supply of terms of the shape [p=v]M , with Many other properties of these operations are proved in the formal development. 2.4 No Free Occurrences of Variables Intuitively parameters are the free names in terms; variables are intended to be the bound names, and we do not consider terms with free variables to be well formed. We define inductively a predicate Vclosed (variable-closed) over terms (table 1). This is analagous to the way a typing relation specifies another kind of well-formedness. (It will turn out that every PTS -typable term is Vclosed ). Thus Vclosed is used as an induction principle over well formed terms. As this relation is a simple case of ideas that recur many times in what follows, we will discuss it at some length. Of course all terms of form s and p are Vclosed (rule Vcl-atom), and no terms of shape v are Vclosed (there is no rule to introduce Vclosed(v) ), but how do we define Vclosed for binders? The approach to "going under binders" is a central idea of our formal handling of names: for hv:Aia to be Vclosed , we require Vclosed(A) and Vclosed( [p=v]a) for some parameter p . That is, to go under a binder, first fill the hole with parameter, p . But p doesn't appear in the conclusion of rule Vcl-bind; which parameter are we to use? In the definition of Vclosed we say that any parameter will do, but there is another possible choice: that Vclosed( [p=v]B ) be derivable for all p . This is not a formal question; it is one of the tasks of a reader of formal mathematics to decide if the formalisation correctly captures her informal understanding. But a formaliser can help readers by pointing out alternatives, and formally proving some relationship between them. This is especially interesting when alternative definitions lead to easier proofs in some cases. We will see this below for Vclosed . Remark 2.2 Vclosed is equivalent to having no free variables (Vclosed vclosed, vclosed Vclosed). This observation may be of informal interest ("the definition of Vclosed is reasonable"), but we do not use it formally because Vclosed allows us to avoid all talk of Vclosed Generation Lemmas Suppose you have a proof of examining it you know it must be constructed by Vcl-bind from proofs of Vclosed(A) and no other rule for Vclosed has a conclusion of shape Vclosed( hv:AiB ) . The very fact that a relation is inductively defined means that its judgements can only be derived by using its rules. This is often called case analysis, and more generally, the lemmas that express such properties are called generation lemmas [Bar92], or inversion principles [DFH 93]. Note that inversion principles are determined by the shape of a definition, not by its extension. LEGO has new and very useful tactics to automate the use of inversion [McB96], but most of what we describe in this paper was done before the tactics were available. We will frequently use inversion on inductive definitions in the rest of this paper without further comment. The generation lemmas from the definition of Vclosed are Notice how the existential quantifier in the case for binders expresses the failure of the subformula property in Vclosed . 2.4.1 A better induction principle for Vclosed . Here are three "obvious" facts about Vclosed (alpha Vclosed lem, Vclosed alpha). They are all directly provable, but appear to need length induction (which appeals to well-founded induction and then subsidiary case analysis; e.g. the proof of claim aVclosed alpha below), for the usual reason that statements about change of names are proved by length induction rather than structural induction: e.g. [q=v]M is not generally a subterm of ( M N ) , but it is shorter than ( M N ) . We will derive a new induction principle which packages up such arguments once and for all. Consider an alternative definition, called aVclosed , differing only in the rule for binders, in which the right premise requires aVcl-bind We will show that Vclosed and aVclosed derive the same judgements. Induction over aVclosed is the principle which Melham and Gordon rediscovered [GM96, Section 3.2]. It is worth saying that Vclosed is a type of finitely branching well-founded trees; i.e. Vcl-atom are the leaves, and Vcl-bind and Vcl-app are binary branching nodes. On the other hand, aVclosed contains infinitely branching well-founded trees, where aVcl-bind creates a branch for each parameter p . Notice also that for any term, A , there is at most one derivation of aVclosed(A) , while there may be many derivations of Vclosed(A) , differing in the parameters used in the left premises of instances of aVcl-bind. Equivalence of Vclosed and aVclosed (aVclosed Vclosed, Vclosed aVclosed) Both directions follow easily by structural inductions once we have the following claim (aVclosed alpha): Proof. The claim is proved by induction on lngth(B) . This works because every term appearing in a premise of a rule of aVclosed is shorter than the term appearing in its conclusion; the typing relations to be considered later do not have this property, and more subtle proofs will be required (section 5.2.1). By well-founded induction on lngth(B) , we have the goal Now using term structural induction on A , we have cases for sort, vari- able, parameter, binder and application (only case analysis is necessary here; we don't use the structural induction hypotheses). Consider the case for binder: we must show aVclosed( [q=v] hn:AiB ) , i.e. under the assumptions (i.e. aVclosed(hn:[p=v]Aiif(v=n; B; [p=v]B))) By aVclosed inversion applied to assumption vclp we also know By aVcl-bind, it suffices to show Noticing that [p=v] doesn't change lngth, the first of these holds by ih and h1. For the second, let r be an arbitrary parameter, and consider cases. If then we are done by h2; i.e. [q=v]B doesn't actually appear in the goal, and [p=v]B doesn't actually appear in h2. Finally the interesting case: if v 6= n we use a straightfoward lemma (alpha commutes alpha) to rewrite the goal to aVclosed( [q=v] [r=n]B ) . By ih it suffices to show aVclosed( [p=v] [r=n]B ) , which follows by h2 after again rewriting the order of substituting p and r . What have we gained? By defining aVclosed and showing it to be extensionally equivalent to Vclosed , we can view the induction principle of aVclosed as an induction principle for the extension of Vclosed , and this is clearly stronger than the induction principle of Vclosed . We insist on extension to point out that aVclosed-induction may be used to prove statements about the judgement Vclosed , but not about derivations of the judgement. Notice that we could directly prove the analogue of claim aVclosed alpha for Vclosed (the proof outlined above works), but it is not just the stronger premises of aVclosed we are after (i.e. the generation lemmas), it is the stronger induction hypotheses. 2.5 A Technical Digression: Renamings are sequential operations; we have not used a notion of simultaneous substitution, except in the following special case. A renaming is a finite function from parameters to parameters. Renamings are represented formally by their graphs as lists of ordered pairs. rp , PP \Theta PP (renaming pair) Renaming , list(rp) ae and oe range over renamings. The action of a renaming (renTrm) on parameters is by lookup in the representing list, and is extended compositionally to all terms. aep , (assoc p p ae) aeff , ff (ff 2 VV; SS) aehv:Aia , hv:aeAiaea ae(M N) , ae(M) ae(N) This is a "tricky" representation. First, if there is no pair (p; q) in ae , returns p , so the action of a renaming is always total, with finite support. Also, while there is no assumption that renamings are the graphs of functional relations, the action of a renaming is func- tional, because assoc finds the first matching pair. Conversely, consing a new pair to the front of a renaming will "shadow" any old pair with the same first component. We do not formalize these observations. Renamings commute with substitution in a natural way: Renaming is iterated substitution. We can analyse the action of a renaming in terms of substitution (renTrm is conjugated psub): From this lemma it is easy to show that renaming respects any relation that substitution of parameters respects (psub resp renTrm resp): Similar results hold for n -ary relations R . Injective and Surjective Renamings It is useful to have bijective renamings (e.g. in Section 5.2.1). The definitions are standard: It is surprisingly difficult to construct bijective renamings in general because of the trickiness of the representation mentioned above. However it's clear that any renaming that only swaps parameters is bijective (swap sur, swap inj), and this is enough for our purposes: 3 Reduction and Conversion In this section we outline the theory of reduction and conversion of Pure Languages. The main results are the Church-Rosser and standardization theorems. As in the definition of Vclosed (Section 2.4), the interesting point in defining reduction is how the relation goes under binders. To understand how reduction works, consider informally one-step beta-reduction of untyped lambda calculus. In our style the fi and - rules are: The substitution [N=x]M on the RHS of fi does not prevent capture, so some restriction is required. It is obvious that no capture can occur if N is closed in the usual informal sense, but because we distinguish between parameters and variables it is enough that N be Vclosed . This is no actual restriction: we will only want to reason about Vclosed terms anyway, as these are the "well-formed" terms. To use fi under a binder, as allowed by - , we must preserve the invariant that fi is only applied to Vclosed terms: we fill the "holes" left by stripping off the binder with a fresh parameter. Here is an instance of - where incorrect capture might occur (contracting the underlined redex): After removing the outer binder -x , replacing its bound instances by a fresh parameter, q , and contracting the Vclosed redex thus obtained, we must re-bind the hole now occupied by q . (Since q was fresh, all instances of q mark holes that should be re-bound). According to - , we require a variable, y , and a term, N , such that [q=y]N is the contractum of the Vclosed redex, -x:q in the example. Such a pair is y , -x:y (the one we have used above), as is z , -x:z for any z 6= x . However - does not derive the incorrect judgement because Thus incorrect capture is avoided. 3.1 Parallel Reduction Rather than use ordinary fi -reduction, we take parallel reduction ('a la Tait-Martin-L-of) as the basic reduction relation. Parallel reduction is convenient for the Church-Rosser and standardization theorems, as emphasised by Takahashi in her beautiful account [Tak95]. Our development follows that of [Tak95], with some refinements. 3.1.1 One-step parallel reduction This relation, ![ (par red1), is defined in Table 2. As in Vclosed above, the dependence of the congruence rule for binders on the choice of a parameter p is only apparent. However, something new arises here, namely the side conditions p 62 . These are eigenvariable pr1-atom ff ![ pr1-beta A ![ Vclosed(U) pr1-bind A ![ pr1-app A ![ Table 2: 1-Step Parallel Reduction conditions 5 , which ensure that the parameter p correctly indicates the position of the bound variables in the compound terms. Only Vclosed terms participate in ![ (par red1 Vclosed) and ![ is reflexive on Vclosed terms (par red1 refl) A stronger induction principle for ![ The rules pr1-atom, pr1- bind and pr1-app are the congruence rules for our language. As with Vclosed (section 2.4.1), we introduce a strong congruence rule for binders ~ pr1 -bind A ~ and prove that ![ and ~ are extensionally equivalent, giveing us stronger induction and inversion principles. Because of the eigenvariable conditions in pr1-bind, a technique using renamings is required 5 Kleene [Kle52, x78, on the notion of "pure variable" proof] explains how to treat such conditions; however, to do so he must explicitly consider operations on derivations, hence dependent elimination, whereas our methods require only rule induction, i.e. non-dependent elimination. The second author is grateful to N. Shankar for this reference. to show the equivalence. We omit the details, but a similar argument is used in section 5.2.1. The strong induction principle is used to show that ![ is closed under substitution (par red1 psub): Many-step parallel reduction ! (par redn), is the transitive closure of ![ . It inherits properties par redn Vclosed and par redn refl from the corresponding properties of ![ mentioned above. 3.1.2 Alpha-Conversion We define ff -conversion, ff - , to be the least congruence, i.e. ff - is exactly without the rule pr1-beta, so ff . This definition is symmetric, by inspection. To show that it is transitive requires the stronger induction principle for ff - , which we prove in the same way as above. Hence ff - is an equivalence relation 6 . It is decidable for Vclosed terms (decide alpha conv): ff with a straightfoward but messy proof, by double induction on and aVclosed(B) . Informally, alpha-conversion is used for changing the names of vari- ables. We do not have hx:AiB ff does not prevent capture. However, we do have (true alpha conv pi): Closure under ff -conversion One of Coquand's original motivations for distinguishing between variables and parameters was to avoid the need to reason about ff -conversion; many of the arguments below (Church-Rosser, standardisation, subject reduction) achieve this goal. Name-carrying syntax is regarded as an abbreviation for a quotient modulo ff -conversion, so that when we formalise a relation R such as parallel reduction above, we really intend R modulo the quotient structure, i.e. ff - . We say a relation R is: 6 This should be contrasted with Gallier's meticulous but long-winded treatment in [Gal90]. closed under ff if ffffiR ' Rffiff ; strongly closed, if ffffiR ' R ; full wrt ff if Rffiff ' ffffiR ; strongly full, if Rffiff ' R . Remark 3.1 ![ is strongly closed under ff -conversion: the proof is the same as that for transitivity of ff - , with the additional case of a redex handled by observing that an ff -variant of a redex is a redex. However ![ is not full w.r.t. ff -classes. For example but no ff -variant of the LHS ![ -reduces to ( [y 1 :q]y 1 although this is an ff -variant of the RHS. 3.1.3 A Church-Rosser Theorem Using the argument of Tait and Martin-L-of, as modernized in [Tak95], we prove the first CR theorem (par redn DP): by the usual strip lemma argument and the diamond property of ![ (comp dev par red1 DP). To do so, we introduce an inductive characterisation of complete development, \Gamma![ , (comp dev) 7 . It is given by the same rules as ![ except for the application rule: cd-app A \Gamma![ AB \Gamma![ A is not a lambda The side condition in cd-app forces contraction of all redexes: we have a deterministic sub-relation of ![ . The theorem on finiteness of developments now becomes the combination of: ffl induction on the definition of \Gamma![ , which we may think of as a partial correctness assertion; ffl the existence (for Vclosed terms) of complete developments, (comp dev exists), which we may think of as a termination argument 7 cf. the definition of \Gamma![ as a function by structural recursion on terms [Tak95] This separation of concerns gives us an advantage over Takahashi's informal proofs, in that we do not have to consider, in each proof about \Gamma![ , a subsidiary induction (case-analysis) to resolve the redex/non- redex distinction in the case of an application. This is handled once and for all in the existence proof, while induction on the definition of \Gamma![ already delineates the redex/non-redex distinction. The price we pay is that we no longer work with an object-level function, but rather a functional relation. Of course, we have simplified matters by considering developments of the entire set of redexes in a term: this is sufficient for our purposes, but a more refined analysis (e.g. [Hue94]) would take us beyond our simple datastructure of terms. The diamond property (comp dev par red1 DP) follows easily from the following "Takahashi" lemma (comp dev preCR): whose proof is by induction on M \Gamma![ M 0 , with inversion of M![ N . As usual, the interesting case, of a redex/redex, appeals to par red1 psub. Remark 3.2 These proofs do not make any appeal to ff -conversion. This is because both the ![ and \Gamma![ relations are strongly closed under ff -conversion. Indeed, we may show the following two properties, which strengthen comp dev exists, namely comp dev unique: and comp dev exists unique: 3.2 Conversion We define conversion, ' (conv), as the symmetric and transitive closure of ! . It inherits properties conv Vclosed and conv refl from those of ! mentioned above. The second Church-Rosser theorem is now straightfoward to prove for conversion (convCR): 3.3 Normal Forms A term is beta normal (beta norm), if it has no beta redexes . This may be defined with the same rules as aVclosed , except the rule for application, which is bn-app beta norm (A) beta norm (B) beta norm ( AB ) A is not a lambda All beta norm terms are Vclosed (beta norm Vclosed). A relation of reduction to normal form is defined: A # N , beta norm (N) - A ! (normal is reflexive, so there is reduction from a normal form, but every reduct of a normal form is a normal form (par-redn-bnorm-is-bnorm) Any reduct of a normal form alpha-converts with that normal form (par-redn-bnorm-is-alpha-conv) Hence, by Church-Rosser, normal forms of a term are unique up to alpha-conversion (nf-unique). Since ff , the converse also holds (nf-alpha-class) Thus the class of normal forms of a ( Vclosed ) term is either empty or exactly an alpha-conversion equivalence class. Deciding conversion Conversion is decidable for normalizing terms. The proof of this depends on Church-Rosser; since normal forms are unique only up to alpha-conversion, it also depends on decidability of alpha-conversion (Section 3.1.2). 3.4 The Standardization Theorem Our work on type-checking requires us to go beyond theorems such as Church-Rosser in the analysis of reduction. In particular, to talk of syntax-directed systems, we must consider deterministic reduction relations, of which weak-head reduction is the simplest. A typical A !wh A 0 Table 3: One step of weak-head reduction property required of such a relation is the following counterpart to the quasi-normalisation theorem (wh standardisation lemma): Takahashi showed how to approach such theorems with an analysis of parallel reduction into head reduction followed by internal reduction, a so-called semi-standardization lemma [Mit79]. We adapted her methods to the case of weak-head reduction, and the corresponding modified notion of internal reduction. In doing so we simplify them somewhat, in particular by removing the need for the complex invariant M ? N . Moreover, the arguments we employed can be replicated in the context of head reduction and internal reduction in their classical senses. Recently, we rounded off this line of development by proving a standardization theorem for pure languages. The main novelty here is the removal of any mention of residuals (so the reader may be unconvinced that we have formalised the standardization theorem). The other thing to observe is that all the desirable consequences of standardization, which we required to analyse type-checking, such as the lemma above, are already corollaries to the semi-standardization lemma. There are three main ingredients to the theorem: weak-head reduc- tion, internal parallel reduction, and standard reduction itself. 3.4.1 Weak-head reduction One step of weak-head reduction (wh red1) is shown in Table 3. By inversion, we see that there are no weak-head reducts of a lambda, so we may assume without loss of generality that A is not a lambda in the rule wh1-app. We have not built any Vclosed assumptions into the definition, as this will always be used in a context in which all terms are Vclosed . The reader may validate such a definition by considering the weak- head normal forms (Table 4), and various lemmas relating !wh and wh-atom whnf(ff) ff 2 PP [ SS wh-bind A is not a lambda Table 4: Weak-head normal forms ip1-atom ff![ ip1-bind Table 5: One step of internal parallel reduction whnf (wh red1 determin, wh nf is nf1, alpha conv resp wh nf). Many-step weak-head reduction, iwh , (wh redn) is defined as the reflexive transitive closure of !wh . It is closed under renam- ings, substitution (psub resp wh redn), and application on the right 3.4.2 Internal parallel reduction The classical notion of head reduction leads to a notion of "internal" redex, as any non-head redex. We adapt such a notion to weak-head re- duction, which gives us the definition of internal parallel reduction, ![ (ipar red1) as shown in Table 5. We allow arbitrary parallel reduction in each compound term, except in the rator position of applications, where we restrict to internal reduction. It is immediate by structural induction that internal parallel reductions are parallel reductions. We also have the important abstract property (ipar red1 refl wh nf) that ![ preserves and reflects weak- head normal forms, and a fortiori , the shape (outermost constructor) of a term. This reflection of weak-head normal forms, together with the lemma below, is the key to the proof of the quasi-normalisation result with which we opened this discussion. Semi-standardization (par red1 wh redn ipar red1) Proof. The proof proceeds by induction on M![ . The only tricky case is that of the parallel fi step. By inductive hypothesis, we obtain (introducing Skolem constants Aw ; Bw ) and we are required to show that there exists some Mw such that Since by psub resp wh redn and psub is vsub alpha, we may conclude the result by stitching weak-head reduction sequences together, provided we can establish the following claim, which is the easy base case of Lemma 2.4 in [Tak95] (wh ipar red1 psub): This is proved in the same way as we showed closure of parallel reduction under substitution, by induction on M![ . A detail to observe here is that we must explicitly assume that the reduction from N to N 0 is parallel. Takahashi builds this into her ? invariant, whereas in the use of wh ipar red1 psub, we obtain this assumption for free as the premise associated with reduction in A . We show the case of an application and B![ . By induction hypothesis, there exists some PA such that [N=p]A iwh PA![ . The proof of the claim, and hence of the whole lemma, is completed by taking P= def PA ( [N=p]B ) , and appealing once more to psub resp wh redn, ipar red1 app and par red1 psub. To establish the full semi-standardization result for ! , we must also show the commutation result (ipar wh redn commutes): This is a corollary (by induction on Mw iwh M 0 ) of the following lemma (ipar wh red1 commutes): Proof. Induction on M![ of the ancillary hypothesis All cases are trivial, except that of application . We show the case of a wh-redex, where we have: and We use the fact that ipar red1 reflects the weak-head normal form A to infer that A = [u:U ]A 000 . Moreover, since A![ by inversion that U![ V and 8p: [p=u]A 000 ![ [p=v]A 00 . Choosing p 62 A 00 ; p 62 A 000 , and applying pr1-sub, we obtain [B=u]A 000 ![ Now we appeal to Lemma 3.4.2, finally to conclude that for some P , we have as required. Throughout we used induction on the definitions of the various reduction relations to establish these lemmas. This is by contrast with Takahashi's treatment, where induction is on the term structure (with inversion of the relational hypotheses). This leads to slightly weaker arguments, and consequently to the need for stronger inductive in- variants. Such refinements to these arguments would be inconceivable without machine support. 3.4.3 Standard Reduction The property of being a standard reduction, \Gamma! s , is usually stated (e.g. by Mitschke [Mit79]) in terms of a highly intensional geometric definition on -terms. To formalise this definition directly, we would have to enrich the datatype of terms in order to be able to speak of std-atom ff\Gamma! s ff ff 2 PP [ SS std-bind std-app std-wh Table Standard reduction, standard (adapted from Plotkin) redex positions in a term. Such an approach has been taken in [Hue94]; we have chosen instead a presentation (table 6), adapted from Plotkin's notion of standard sequence [Plo75]. The essence of this presentation is to define standard reduction as the least congruence closed under prefixing by weak-head reductions. We leave implicit the sequence of redexes contracted, as this may be computed by recursion, and its "left-to-right" character, and thus avoid mention of residuals or redex positions. Remark 3.3 We have defined standard reductions of arbitrary length once and for all, and without any considerations of reduction to normal note furthermore, however, that we may easily achieve this end by adding the side condition A 0 6= - to the std-app rule, and whnf(B) to the std-wh rule. Also, this definition is both strongly closed under ff -conversion, and strongly full. Our final aim is the following standardization theorem (the standardisation lemma): Induction on A ! suffices if we can show that our notion of standard reduction absorbs single steps of parallel reduction: By our Lemma 3.4.2, and std-wh, it suffices to prove the following lemma, that standard reduction absorbs single steps of internal parallel reduction (standard absorbs ipar red1): Proof. The proof is by induction on B \Gamma! s C . This avoids reconsidering the tricky application case of the commutation lemma above, and we may exploit the fact preserves and reflects the shape of terms. The price we pay is the need for strong induction on B \Gamma! s C , as the ancillary hypothesis A![ then an expansion step. In each non-atomic case, we make a subsidiary appeal to semi-standardization, in order to be able to exploit the induction hypotheses. This gives a rather mechanical and "abstract nonsense" flavour to the argument, emphasising once more that the real complexity lies in the the proof of Lemma 3.4.2. We focus on the case of a binder, where we have as hypotheses: By inversion of H, we conclude that A and 8p: [p=w]B c ![ [p=u]B . But now by semi-standardization, A c iwh Aw![ . So by induction hypothesis ihA, we have Aw \Gamma! s A 0 , and using std-wh to fold back the reductions A c iwh Aw , we finally obtain A c \Gamma! s A 0 . In exactly the same way, modulo the choice of a parameter we obtain [p=w]B c \Gamma! s [p=v]B 0 . But now C \Gamma! s hv:A 0 iB 0 by construction This concludes the proof that standard reduction absorbs internal reduction, and hence parallel reduction, and so finally we may conclude that every (parallel) fi -reduction sequence may be standardized. 4 Pure Type Systems PTS is a class of type theories given by a set of derivation rules (table 7) parameterized by a PL , (PP; VV; SS) , and by two relations informally as ax(s 1 :s 2 ) Start Weak Lda App Table 7: The Typing Rules of PTS (formal name gts) ffl rules, rl ' SS \Theta SS \Theta SS , written informally as rl(s We usually intend ax and rl to be decidable, but this assumption is not used in the basic theory of PTS . When we are interested in algorithms for typechecking ([Pol94b, Pol95]), even stronger assumptions about decidability are needed. The typing judgement has the shape meaning that in has type A . (The formal name of this relation is gts, from the old name "Generalized Type Systems".) We call M (or (\Gamma; M) ) the subject and A the predicate of judgement. Contexts, ranged over by \Gamma , \Delta , bind parameters to types: Contexts are formalized as lists over PP \Theta Trm . If \Gamma participates in some derivable judgement, 4.1 A Generalization: Abstract Conversion We have further generalized PTS by parameterizing the rules of table 7 on another relation, - (called abstract conversion), occurring in the side condition of rule tConv. In conventional presentations of PTS [Bar92], the actual relation of beta-conversion ( ' ) is used for - . There are several reasons to be interested in parameterizing PTS on its conversion relation. For one thing, the type theory ECC, implemented in LEGO, is not actually a PTS because it uses a generalized notion of conversion called cumulativity. ECC is of special interest to us, so we formalize an extension of PTS which includes ECC. Our formal development includes a typechecking algorithm for ECC [Pol94b]. Even for PTS , there is a notorious open problem, the Expansion Postponement problem [vBJMP94, Pol94b], which asks if the conversion relation in table 7 can be replaced by beta-reduction without changing the typability of any terms. We know of one other work on PTS using an abstract conversion relation: [BM]. The only properties of - necessary to prove the substitution lemma (section 5.4) are reflexivity, transitivity, and invariance under substitution cnv trans 8A; To prove the subject reduction theorem (section 5.5), we finally need that conversion is related to contraction of redexes, and has an "inter- nal" Church-Rosser property: Notice the contravariance in the last property. It is easy to prove that ' has these properties, so we can formally instantiate - by ' (or by the cumulativity of ECC) and discharge all these assumptions; we are working abstractly, not making assumptions. ' is an equivalence relation, but - is a partial order (which is why we use an asymmetric symbol for - ). Other significant properties of ' that do not generally hold for - include: Some differences between abstract-conversion- PTS and fi -conversion- are detailed in [Pol94b]. This kind of analysis, of which properties are actually used in some body of work, is greatly aided by formalization 4.2 Are the rules a formalization of PTS? Leaving aside abstract conversion, the rules of table 7 differ from the standard informal presentation [Bar92] in several ways. First, the handling of parameters and variables in the Pi and Lda rules, is similar to that in the rules of table 2. Other differences are restriction of weakening to atomic subjects, and the Lda rule. Binding and substitution The treatment of operating under binders in table 7 is analogous to that in the reduction relations considered above. (See discussion of the rule Lda below.) As the substitution used in rule App may cause capture of variables, we must show that N is Vclosed . In fact we have (gts Vclosed lem) by structural induction. It also follows easily that a Valid context is Vclosed in an obvious sense. Atomic Weakening The standard presentation of weakening in PTS allows any term as subject (weakening) where we restrict to atomic terms, ff (see rule Weak). Our rules derive the same judgements (weakening is seen to be admissible in section 5.3), but allow fewer derivations (those derivations where weakening is pushed to the leaves). This gives a cleaner meta theory, as induction over derivations treats fewer cases. For example, given a with an application as its subject, there is no confusion whether it is derived by App or Weak. Thus, with atomic weakening, any judgement may only be derived by tConv or by exactly one of the remaining rules. The Lambda Rule For the rule for typing a lambda in informal presentations [Bar92, Geu93] is The conventional understanding is that the bound variable, x , doesn't really occur in the conclusion of rule - , as the notations " refer to alpha-equivalence classes 8 . Thus, in concrete notation, the subject and predicate of the lambda rule may bind different variables, which we formalize in our rule Lda. One might, instead, rule - as This was our first attempt, and surprisingly, this system derives the same judgements as the system of table 7 (lemmas rlts gts and gts rlts). However, using Lda', the subject reduction theorem is difficult to prove, and derivations are distorted by the need to use the conversion rule for alpha-conversion. See [Pol94b] for more details. 5 PTS With Abstract Conversion We survey the development leading to the subject reduction theo- rem. The main difference between this section and the presentation in [Bar92] is our use of the atomic weakening rules (section 4.2), and our simpler proof of subject reduction (section 5.5). 5.1 Some basic facts Here is a sample of the many small facts that had to be established, usually by simple structural induction. Parameter lemmas All parameter occurrences in a judgement are bound in the context, and the binding instances in a Valid context are distinct parameters: (free params lem1) 8 However, in the left premise of rule - , all free instances of the actual symbol are intended to refer to the context entry [x:A] . Thus the conventional reading of this rule doesn't make sense as concrete notation. Start Lemmas Every axiom is derivable in every valid context, and the global bindings of a valid context are all derivable: 5.2 A Better induction principle for PTS As for previous relations, we define an alternative typing relation, ' a (apts), that identifies all those derivations of each ' judgement that are inessentially different because of parameters occurring in the derivation but not in its conclusion. ' a differs from ' only in the right premise of the Pi rule and the left premise of the Lda rule. aLda In these premises we avoid choosing a particular parameter by requiring the premise to hold for all parameters for which there is no reason it should not hold, that is, for all "sufficiently fresh" parameters. As before, we will show that ' and ' a derive the same judgements. It is interesting to compare the side conditions of Pi with those of aPi. In Pi we need the side condition p 62 B so that no unintended occurrences of p (i.e. those not arising from occurrences of the variable x ) are bound in the right premise; we do not need p because the validity of \Gamma[p:A] is obvious from the right premise. In aPi, we cannot require the right premise for all p , but only for those such that \Gamma[p:A] remains valid, i.e. those p not occurring in \Gamma . However the condition not required because of "genericity", that is, the right premise of aPi must hold for the infinitely many parameters not occurring in only finitely many of these instances can occur in B . 5.2.1 ' a is equivalent to ' As with previous relations, this equivalence will give us a stronger induction principle, and stronger generation (inversion) lemmas for ' . (apts gts; gts apts) Proof. Direction ) is straightfoward by structural induction on To prove direction ( , first prove a lemma that bijective renamings respect ' a (bij ren resp apts) by ' a -structural induction 9 . Now we proceed to prove by structural induction on a derivation of All cases are trivial except for the rules Pi and Lda. Consider the case for Pi: we must prove under the assumptions l prem prem By rule aPi (using l ih) it suffices to show \Gamma[r:A] ' a arbitrary parameter r 62 \Gamma . Thus, using the free parameter lemmas of section 5.1, we know A from l prem and noccG prem prem Taking derivable using bij-ren-resp-apts to rename r ih. (Recall from section 2.5 that swap(p; q) is always bijective.) Thus we are finished if we can show It is clear that the first equation holds from nopG, noccG, norA and nopA. For the second equation, notice that if then we are done trivially, so assume r 6= p , and hence r 62 [p=n]B (from r prem and noccG). Now, using vsub renTrm commutes (section 2.5) we have 9 Actually injectivity of a renaming is enough for it to preserve ' a , but we cannot prove this until after we know ' a = ' . as required. 5.3 The Thinning Lemma and Weakening The Thinning Lemma is important to our formulation because it shows that full weakening (weakening) is admissible in our system, justifying our use of atomic weakening in the definition of ' (section 4.2). The subcontext relation is defined We also say \Delta extends \Gamma . This is the definition used informally in [Bar92, GN91, vBJ93]; a much more complicated definition is required to express this property in a representation using de Bruijn indices for global variables. Now we can state (thinning lemma): A naive attempt to prove the thinning lemma by structural induction encounters serious difficulties with parameter side conditions (see [MP93, Pol94b] for discussion), but a proof is straightfoward using ' a -induction, justified by the previous section. The full weakening rule is a corollary of thinning lemma. 5.4 Cut and type correctness The substitution lemma (substitution lemma) is proved by induction on the derivation of \Gamma[p : From this we get the commonly used case (cut rule) by instantiating \Delta to the empty context. Among the correctness criteria for type systems is that every type is itself well formed. In PTS we have the theorem (type correctness): The proof is by structural induction; the only non-trivial case is rule App, which uses the substitution lemma and vsub-is-psub-alpha (section 2.3). Subject Reduction Theorem: Closure Under Reductio An important property of type systems is that a term does not lose types under reduction, thus types are a classification of terms preserved by computation. In fact we will show entire ' -judgements are closed under reduction. We now need all five properties of abstract conversion (section 4.1). 5.5.1 Non-Overlapping Reduction Our goal is to prove usually called the subject reduction theorem. Our naive strategy is to show that one step of reduction preserves typing, by induction on . The critical case is rule App, when the application is actually a redex that is contracted. In order to simplify that case, we want to avoid overlapping redexes, as allowed in the fi -rule of ![ We want some reduction relation with no overlapping, whose transitive closure is equal to ! Another difficulty is that in rules Pi and Lda, a subterm of the subject of the conclusion (the type-label) appears in the context part of a premise; thus in these cases of an induction argument, a reduction in the subject of the conclusion may result in a reduction in the context of a premise. This suggests that the induction hypothesis should be strengthened to simultaneously treat reduction in the context and the subject, leading to the goal where ! is ordinary one-step fi -reduction. This approach is used in [GN91, Bar92], and produces a large number of case distinctions, based on which subterm of the subject contains the one redex which is contracted; all of these subcases are inessential except to isolate the one non-trivial case where the redex contracted is the application constructed by rule App. This simultaneous treatment of one reduction in either the context or the subject suggested to us that the proof would be smoother using a reduction relation that is congruent simultaneously in all branches, while forbidding overlapping of redexes. One step non-overlapping reduction, no (no red1), is defined by the same rules as ![ (table 2) except for the fi -rule, which is modified to prevent overlapping redexes: nor1-beta ( [u:U ]B ) A no no , so from the assumed property cnv conv (section 4.1), we have A no We extend no compositionally to contexts (red1Cxt), and to pairs of a context and a term (red1Subj), writing \Gamma no h\Gamma; Mi no We also define no (no redn), the transitive closure of no , and show no (no par redn, par no redn). 5.5.2 The Main Lemma (subject reduction lem) Proof. By structural induction on We show the interesting case, from rule App. Given l prem prem red we must show . By induction hypotheses By type correctness of gtsDM, for some s By the pi-generation lemma, for some s 2 and p 62 B By the cut rule on gtsDL and gtsDB (we also use vsub-is-psub-alpha (section 2.3) here, and several more times in this case) Now there are two subcases !M 0 and L no The goal is rule App and the induction hypotheses we easily have Use rule tCnv and gts- DBsub to expand L 0 in the predicate back to L as required (this uses cnv-conv). The goal is by the lambda-generation lemma, for some w , B 0 and s 0 By (cnvCR-pi c') (this is the only time it is used in this proof) By a generation lemma on gtsDpi' By (tCnv cnvA gtsDL gtsDA') By tCnv, cnvB and gtsDBsub, it suffices to show which follows by cut on gtsDL' and gtsDb. 5.5.3 Closure Under Reduction It is now easy to show the subject reduction theorem, gtsSR, and a useful corollary, predicate reduction Finally, extending ! compositionally to contexts 10 , ' is closed under beta-reduction (gtsAllRed) This relation is equivalent to transitively closing the compositional extension of of no ! to contexts. There is a trivial but useful lemma (predicate conv): Unlike rule tCnv, we don't ask for evidence that s has a type, but the side condition uses ' , not - . To prove such a lemma with - requires technical restrictions; e.g. ECC with its type hierarchy chopped off a finite level fails to have such a property because of the sort at the top of the hierarchy. Closure Under Alpha-Conversion From gtsAllRed and the fact that ff , it follows that ' judgements are preserved by ff (gts alpha closed). Hence an implementation may typecheck a judgement as stated by a user, rather than searching for an alpha- equivalent judgement which is derivable. See [Pol94b, Pol94a]. 5.6 Another Presentations of PTS In several rules of ' the context \Gamma occurs more then once in the list of premises; in order to build a complete derivation, \Gamma must be constructed (by Start and Weak) in each branch in which it appears. It is much more efficient to assume that we start with a valid context, and only check that when rules extend the context (i.e. the right premise of Pi and the left premise of Lda) they maintain validity. This is more in keeping with implementations which are actually used, where we work in a "current context" of mathematical assumptions. We present such a system in table 8, and show it is equivalent to ' . The idea is originally due to Martin-L-of [Mar71], and is used in [Hue89]. This system has two judgements, a type judgment of shape (lvtyp), and a validity judgement of shape \Gamma ' lv (lvcxt). Note that they are not mutually inductive: validity depends on typing, but not conversely. We have proved that ' lv characterizes ' (iff gts lvcxt lvtyp): Direction ( of the proof is subtle. Formally, it uses an auxiliary mutual inductive definition, and a well-founded induction requiring dependent elimination; this is the only place in the entire development that either mutual induction or dependent elimination are used. More abstractly, direction ( claims termination of a function that replaces all the proof annotations omitted by ' lv . As this is a fast-growing function, its termination is a strong result. The proof is described in [Pol94b]. lvLda lvCons Table 8: The system of locally valid contexts (lvtyp, lvcxt) 6 PTS with fi -conversion It is remarkable how little about - has been needed for the theory described so far (section 4.1). In [Pol94b] we pursue the theory of PTS with abstract conversion to a correct typechecking algorithm for cumulative PTS , including Luo's system ECC. Here, we point out a more standard theory of PTS with fi -conversion, i.e. we instantiate - in the preceeding with the actual relation ' . (In LEGO the command Cut executes the admissible rule substitution lemma of section 5.4.) This theory, leading to the strengthening theorem and typechecking algorithms for classes of PTS , is detailed in [vBJMP94]. 6.1 Strengthening Strengthening is a tricky result about PTS , first proved by Jut- ting [vBJ93]: \Gamma[q:C (gts strengthening) The development we formalize, in which strengthening is a corollary to work on typechecking, is described in detail in [vBJMP94]. We were particularly interested to prove strengthening because LEGO uses it in the Discharge command. 6.2 Functional PTS Functional PTS are well behaved and are, perhaps, the only ones that are interesting in practice. Functional , ae In a functional PTS , ax and rl are the graphs of partial functions, but we do not necessarily have procedures to compute these functions. Uniqueness of Types The definition of functional PTS makes sense for abstract-conversion PTS , and is useful in that setting, as it gives a kind of uniqueness: when building a derivation of a typing judgement guided by the syntax of its subject, it is deterministic which axiom to use at each instance of Ax, and which rule to use at each instance of Pi [Pol94b]. The idea behind the definition of functional is that the uniqueness just mentioned propagates through whole derivations to give a property that types are unique up to conversion: conv unique types , For fi -conversion PTS we prove Functional ) conv unique types (types unicity) by structural induction on However, this proof uses properties of ' , and cannot be modified to prove any similar property of abstract-conversion PTS . types unicity is too linear for - , which is only a partial order; the correct generalization is a principal types lemma, saying that any type is above some principal type, but we cannot hope that every two types are comparable [Pol94b]. Subject Expansion Any fi - PTS with uniqueness of types also has a subject expansion property (subject expansion): conv unique types ) While subject reduction says that terms don't lose types under reduc- tion, this lemma says terms don't gain types under reduction. In this A is the principal premise, and is a well-formedness premise. There are examples of two different ways subject expansion can fail for non-functional PTS in [vBJMP94]. --R A formalization of the strong normalization proof for System F in LEGO. Lambda calculi with types. Coq en Coq. Type Dependence and Constructive Mathemat- ics On the subject reduction property for algebraic type systems. F-sub, the system. An algorithm for testing conversion in type theory. Combinator shared reduction and infinite objects in type theory. An algorithm for type-checking dependent types Towards checking proof checkers. On Girard's The Collected Papers of Gerhard Gentzen. Logics and Type Systems. Five axioms of alpha con- version A modular proof of strong normalization for the calculus of constructions. The constructive engine. Residual theory in Incremental changes in LEGO: Introduction to Metamathematics. LEGO proof development sys- tem: User's manual. Technical Report ECS-LFCS-92-211, LFCS, Computer Science Dept., University of Edinburgh, The King's Buildings Program specification and data refinement in type theory. Computation and Reasoning: A Type Theory for Computer Science. Per Martin-L-of Inverting inductively defined predicates in LEGO. The standardisation theorem for Pure Type Systems for- malized More Church-Rosser proofs (in isabelle/hol) Coding binding and substitution explicitly in Is- abelle A proof of the Church-Rosser theorem and its representation in a logical framework Closure under alpha-conversion The Theory of LEGO: A Proof Checker for the Extended Calculus of Constructions. A verified typechecker. How to believe a machine-checked proof Natural Deduction Program Verification in Synthetic Domain The- ory Synthetic domain theory in type theory: Another logic of computable functions. Theory of symbolic expressions Auxiliary variables and recursive procedures. A mechanical proof of the Church-Rosser theorem Substitution revisited. Parallel reductions in Formulation of Martin-L-of's theory of types with explicit substitutions. Master's thesis --TR --CTR Michael Norrish, Mechanising Hankin and Barendregt using the Gordon-Melham axioms, Proceedings of the ACM SIGPLAN workshop on Mechanized reasoning about languages with variable binding, p.1-7, August 26, 2003, Uppsala, Sweden Christian Urban , Michael Norrish, A formal treatment of the barendregt variable convention in rule inductions, Proceedings of the 3rd ACM SIGPLAN workshop on Mechanized reasoning about languages with variable binding, p.25-32, September 30-30, 2005, Tallinn, Estonia Yasuhiko Minamide , Koji Okuma, Verifying CPS transformations in Isabelle/HOL, Proceedings of the ACM SIGPLAN workshop on Mechanized reasoning about languages with variable binding, p.1-8, August 26, 2003, Uppsala, Sweden Michael Norrish, Mechanising -calculus using a classical first order theory of terms with permutations, Higher-Order and Symbolic Computation, v.19 n.2-3, p.169-195, September 2006 S. J. Ambler , R. L. Crole , Alberto Momigliano, A definitional approach to primitivexs recursion over higher order abstract syntax, Proceedings of the ACM SIGPLAN workshop on Mechanized reasoning about languages with variable binding, p.1-11, August 26, 2003, Uppsala, Sweden Dimitri Hendriks, Proof Reflection in Coq, Journal of Automated Reasoning, v.29 n.3-4, p.277-307, 2002 Jonathan Ford , Ian A. Mason, Formal Foundations of Operational Semantics, Higher-Order and Symbolic Computation, v.16 n.3, p.161-202, September Conor McBride , James McKinna, Functional pearl: i am not a number--i am a free variable, Proceedings of the 2004 ACM SIGPLAN workshop on Haskell, September 22-22, 2004, Snowbird, Utah, USA James Cheney, Scrap your nameplate: (functional pearl), ACM SIGPLAN Notices, v.40 n.9, September 2005 Thierry Coquand , Randy Pollack , Makoto Takeyama, A Logical Framework with Dependently Typed Records, Fundamenta Informaticae, v.65 n.1-2, p.113-134, January 2005 Ren Vestergaard , James Brotherston, A formalised first-order confluence proof for the -calculus using one-sorted variable names, Information and Computation, v.183 n.2, p.212-244, 15 June Andrew M. Pitts, Nominal logic, a first order theory of names and binding, Information and Computation, v.186 n.2, p.165-193, 01 November Conor McBride , James McKinna, The view from the left, Journal of Functional Programming, v.14 n.1, p.69-111, January 2004 Riccardo Pucella, SIGACT news logic column 14, ACM SIGACT News, v.36 n.4, December 2005
type theory;LEGO proof checker;formal mathematics;lambda calculus;Pure Type Systems
594296
New Worst-Case Upper Bounds for SAT.
In 1980 Monien and Speckenmeyer proved that satisfiability of a propositional formula consisting of i>K clauses (of arbitrary length) can be checked in time of the order 2i>K / 3. Recently Kullmann and Luckhardt proved the worst-case upper bound 2i>L / 9, where i>L is the length of the input formula. The algorithms leading to these bounds are based on the i>splitting method, which goes back to the DavisPutnam procedure. i>Transformation rules (pure literal elimination, unit propagation, etc.) constitute a substantial part of this method. In this paper we present a new transformation rule and two algorithms using this rule. We prove that these algorithms have the worst-case upper bounds 20. 30897 i>K and 20. 10299 i>L, respectively.
Introduction . SAT (the problem of satisability of a propositional formula in conjunctive normal can be easily solved in time of the order 2 N , where N is the number of variables in the input formula. In the early 1980s this trivial bound was reduced for formulas in 3-CNF by Monien and Speckenmeyer [20] (see also [22]) and independently by Dantsin [1] (see also [4] and [2]). After that, many upper bounds for SAT and its subproblems were obtained [21, 17, 25, 13, 14, 26, 16, 8, 23, 24]. Most authors consider bounds w.r.t. three main parameters: the length L of the input formula, the number K of its clauses and the number N of the variables occurring in it. Before the conference proceedings version of this paper [9], the best known worst-case upper bounds for SAT were: p(L)2 L=9 [16], p(L)2 K=3 [21] (see also [16]), where p is a polynomial. Also, the upper bound p(L)2 L=4 for satisability problem for general Boolean formulas (i.e., not necessarily in CNF) is known [28]. Recently Paturi, Pudlak, Saks and Zane [24] proved that 3-SAT is checkable in the randomized time O(2 0:446N ), but it is still Some of the work described in this paper was presented at the 9th ACM-SIAM Symposium on Discrete Algorithms (SODA'98) [9]. y Steklov Institute of Mathematics at St.Petersburg, 27 Fontanka, 191011 St.Petersburg, Russia. Email: hirsch@pdmi.ras.ru, URL: http://logic.pdmi.ras.ru/~hirsch/index.html. Supported in part by grants from INTAS and RFBR. unknown whether the trivial upper bound 2 N for the general SAT problem can be improved. Kullmann and Luckhardt in [16] simplied the algorithm of Monien and Speckenmeyer from [21] and presented also an improved algorithm of the same complexity p(L)2 K=3 which works better when the ratio of the number of clauses to the number of variables is greater than 2:64557. In the conference proceedings version of this paper [9] we presented two algorithms which improve the bounds p(L)2 to p(L)2 0:30897K and p(L)2 0:10537L respectively. In this journal version we improve the second algorithm and prove the worst-case upper bound p(L)2 0:10299L . Our new algorithm uses the p(L)2 0:30897K -time algorithm as a subroutine; the proof of the corresponding upper bound is much simpler than the proof in the conference proceedings version [9]. The most popular methods for solving SAT are the local search method and the splitting method . Many experimental and average-case results show the signicant power of the local search method (for references, see the survey [7]). However, most best known worst-case upper bounds for SAT and its NP-complete subproblems are obtained using the splitting method [21, 16, 8] (for worst-case upper bounds for the local search method see [10, 11, 12]). The splitting method is also useful in proving worst-case upper bounds for exact and approximate MAXSAT solving [18, 3, 19]. The splitting method goes back to the Davis{Putnam procedure [6]. Let l be a literal occurring in a formula F . Let F [l] be the formula obtained from F by assigning the value T rue to the literal l, i.e. by removing all clauses containing l and deleting all occurrences of l from the other clauses. In short, the main idea of the Davis{Putnam procedure is that F is satisable i at least one of the formulas F [l] and F [ l ] is satisable. In a wider sense [5], a splitting algorithm constructs a tree by reducing satisability of a formula F to satisability of several formulas F [I 1 ]; F [I obtained from F by assignments I 1 ; I spectively. Then, a splitting algorithm simplies each of the formulas F [I 1 ]; F [I according to transformation rules which do not change their satisability. These transformations take a polynomial time. Their role is to reduce some of the parameters L, K or N of the formulas F [I 1 ]; F [I and to simplify these formulas by eliminating pure literals (a literal is pure if its negation does not occur in the formula), 1-clauses (1-clause is a clause consisting of one literal) and other \easy pieces". Recurrent equations are often used in complexity analysis of algorithms. Kullmann and Luckhardt in [16] described a very similar, but simpler technique which is very useful in the estimation of the running time of splitting algorithms. One can consider an execution of a splitting algorithm as a branching tree, i.e. a tree such that formulas labelling children are simpler than formulas labelling parents (leaves are labelled with the simplest formulas). With each node of this tree we associate a branching vector of non-negative numbers and a polynomial constructed from this vector. One can estimate the number of leaves in the tree using the largest of the positive roots of these polynomials. Precise denitions and formulations are given in Sect. 3. Transformation rules play an important role in splitting algorithms. Two simplest rules were proposed in the original paper of Davis and Putnam [6]: we can eliminate pure literals and 1-clauses. All transformation rules we use are described in Sect. 4. This paper introduces a new transformation rule which goes up to the following simple observation: if each clause of our formula contains at least one negative literal, then this formula is trivially satisable by the assignment in which all variables have the value F alse. Let P (l; F ) be a property of a literal and a formula, for example, \F contains exactly two occurrences of l and at least three occurrences of l ". In addition, we suppose that for each variable v in the formula F at most one of the literals v and v satises P . Given F , the literals that satisfy the property P will be referred as P -literals. The black and white literals principle. Let F be a formula in CNF. At least one of the following two alternatives holds. (1) There is a clause in F that contains a P -literal and does not contain the negations of any other P -literals. (2) Satisability of F is equivalent to satisability of the formula obtained from F by removing all clauses containing the negations of P -literals. A formal proof of the black and white literals principle is given in Lemma 4.3. Figure 1 illustrates this principle. White circles (-) denote P -literals, black circles () denote their negations, circles with dots (-) denote other literals (i.e., literals which are neither P -literals nor the negations of P -literals; note that the negation of such literal is again neither P -literal nor the negation of a P -literal). Columns correspond to clauses of the formula. The gure contains two formulas, the rst one satises the condition (1), the second one satises the condition (2). (1) | {z } (2) @ @ @ @ @ Figure 1: Two alternatives of the black and white literals principle. This principle is the key point of our algorithm corresponding to the upper bound p(L)2 0:30897K : In this algorithm we use the presence of clauses that contain a P -literal and do not contain the negations of other P -literals for a certain property P . The black and white literals principle is a kind of insurance: it guarantees that if we cannot nd a required clause, then we can replace F by a simpler formula, or F does not contain any P -literals at all. This principle is one of re-formulations of the following simple property: for any partial assignment A (i.e., a set of literals which does not contain simultaneously x and x for any variable x) for a formula F in CNF, either there is a clause in F containing only literals from A, or setting the values of all these literals to F alse does not change the satisability of F . Another known re-formulations of this property are the generalized sign principle [16] and the autarkness principle [20, 22, 17] (see also [16]). A comprehensive study of the autarkness principle can be found in [15]. The use of the black and white literals principle leads to two new bounds for SAT presented in this paper and also to several upper bounds for the satisability problem for formulas in CNF{(1,1) [8] (a formula is in CNF{(1,1) if each literal occurs in it positively at most once; the satisability problem for these formulas is NP-complete). In Sect. 2 we give basic denitions. Section 3 contains the technique that allows us to estimate the size of a branching tree. In Sect. 4 we explain transformation rules that we use in our algorithms. In Sect. 5 and Sect. 6 we describe the algorithms having the upper bounds p(L)2 0:30897K and p(L)2 0:10299L respectively, and the corresponding proofs. 2 Basic denitions. Let V be a set of Boolean variables. The negation of a variable v is denoted by v. Given a set U , we denote Ug. Literals are the members of the set Positive literals are the members of the set V . Negative literals are their negations. If w denotes a negative literal v, then w denotes the variable v. A clause is a nite set of literals that does not contain simultaneously any variable together with its negation. The empty clause is interpreted as F alse. A formula in CNF (CNF-formula) is a nite set of clauses. The empty formula is interpreted as T rue. The length of a clause is its cardinality, the length of a formula is the sum of the lengths of all its clauses. The length of a clause C is denoted by jCj. A k-clause is a clause of the length k. A k + -clause is a clause of the length at least k. A k -clause is a clause of the length at most k. If we say that a literal v occurs in a clause or in a formula, we mean that this clause or this formula contains the literal v. However, if we say that a variable v occurs in a clause or in a formula, we mean that this clause or this formula contains the literal v, or it contains the literal v. An assignment is a nite subset of W that does not contain any variable together with its negation. Informally speaking, if an assignment I contains a literal w, it means that w has the value T rue in I. To obtain F [I] from F and an assignment I remove from F all clauses containing the literals w i , delete all occurrences of the literals w i from the other clauses. For short, we write F [w An assignment I is satisfying for a formula F if F [I] is the empty formula. A formula is satisable if there exists a satisfying assignment for it. We say that two formulas F and G are equi-satisable if both are satisable, or both are unsatisable. Let w be a literal occurring in a formula F . This literal is a i-literal if F contains exactly i occurrences of w. It is a (i;j)-literal if F contains exactly i occurrences of w and exactly j occurrences of w. It is a (i;j + )-literal if F contains exactly i occurrences of w and at least j occurrences of w. It is a (i;j )-literal if F contains exactly i occurrences of w and at most occurrences of w. Similarly, we dene (i We denote the number of occurrences of a literal w in a formula F by #w the sum of the lengths of the clauses containing w by }w When the meaning of F is clear from the context, we omit F . A literal w is a }i-literal if Similarly, we dene 3 Estimation of the size of a branching tree. Kullmann and Luckhardt introduced in [16] a notion of a branching tree. It is intended for estimating the time complexity of splitting algorithms, since a tree of formulas which such an algorithm splits, is a branching tree. One can consider an execution of a splitting algorithm as a tree whose nodes are labelled with CNF-formulas such that if some node is labelled with a CNF-formula F , then its sons are labelled with (simplied) formulas F [I 1 ]; F [I 2 ]; for some assignments I Suppose we have a tree whose nodes are labelled with formulas in CNF. To each formula F we attach a non-negative integer The tree is a branching tree if, for each node, the complexity of the formula labelling this node is strictly greater than the complexity of each of the formulas labelling its sons. In this paper we use is the number of variables in F ; is the number of clauses in F ; is the length of F . We prove two upper bounds, w.r.t. K and w.r.t. L. However, in this section we do not x any concrete measure of complexity. Let us consider a node in our tree labelled with a formula F 0 . Suppose its sons are labelled with formulas F 1 , F 2 , . , Fm . A branching vector of a node is an m-tuple are positive numbers not exceeding The characteristic polynomial of a branching vector ~ t is dened by The characteristic polynomial h ~ t (x) is a monotone function of x on the interval (0; +1), and h ~ t exactly one positive root. We denote this root by ( ~ t ) and call it the branching number . We suppose leaves. We omit one pair of parentheses and write Example 3.1 For example, 0:69424::: is the golden ratio; The branching number of a tree T is the largest of the branching numbers ( ~ t ) of its nodes. We denote it by max;T . The following lemma proved by Kullmann and Luckhardt allows us to estimate the number of leaves in a branching tree using its branching number. Lemma 3.1 (Kullmann, Luckhardt, [16]) Let T be a branching tree, let its root be labelled with a formula F 0 . Then the number of leaves in T does not exceed ( max;T This lemma already allows us to estimate the running time of a splitting algorithm if we know the branching number of the splitting tree and the algorithm processes each its leaf in a polynomial time. However, our algorithm corresponding to the upper bound calls the algorithm corresponding to the upper bound p(L)(6; 7; 6; 7) K as a subroutine and thus processes some of the leaves in an exponential time. To estimate the overall running time, we use the following simple generalization of Lemma 3.1. Lemma 3.2 [8] Let T be a branching tree, let its root be labelled with a formula F 0 . Let G l denote the object labelling a leaf l of the tree T . Let l is a leaf of T Proof is the induction on the construction of the tree. Base. The tree consisting of the unique node. In this case Step. Consider the tree T presented in Fig. 2. Let ~ be the branching tuple in its root. l is a leaf of T s l is a leaf of T j f((G l ))A s s (by denition of the tuple ~ t) Figure 2: A splitting tree s s (by denition of h ~ t ) (max(; max;T (by monotonicity of h ~ t ) ut Now if we know the branching number of the tree corresponding to a splitting algorithm, then we can estimate its running time. We explicitly require our algorithms to perform splittings with branching numbers not greater than we wish (and we only need to prove that there always exists a splitting satisfying this condition). For this purpose, the algorithm has to compare branching numbers corresponding to dierent vectors. Of course, this can be done just by the examination of a constant number of cases and the use of the monotonicity of . However, a more general statement holds. Lemma 3.3 (Kullmann, Luckhardt, [16]) Let m, k be natural constants, x 1 positive rational numbers. The problem whether than (y 1 ;y solvable in time polynomial of max(x 1 ;x In the following, when estimating the running time of our algorithms, we use frequently inequalities like (7; 15) < (5; 17) or (5; 17) < (6; 7; 6; 7) 1=3 without proofs. One can check these inequalities by approximate calculation of the branching numbers. However, there are several simple observations that may help to prove some of this inequalities easier. Lemma 3.4 (Kullmann, Luckhardt, [16]) (1) Permutations of the components of a branching vector do not aect the corresponding branching number. (2) The branching number strictly decreases as one or more components of the branching vector increase. 4 Transformation rules. In this section F denotes a formula in CNF. Lemmas 3.1 and 3.2 allows us to take into consideration only the dierences between the complexity of an input formula and the complexities of the formulas obtained from it by splitting. The higher these dierences are, the smaller the number of leaves is. Thus, to obtain a good algorithm, we should reduce as much as possible the complexities of the formulas obtained by splitting. Here we explain transformation rules that allow us to do it. More precisely, we explain that in certain cases we can nd a simpler formula equi-satisable with F . In the following we use these rules so that they never increase the parameter under consideration (the number of clauses in a formula or the length of a formula). Elimination of 1-clauses. If F contains a 1-clause fag, then the formulas F and F [a] are equi-satisable since all assignments that contain a are unsatisfying. Subsumption. If F contains two clauses C and D such that C D, then F and F nfDg are equi-satisable since each assignment that satises the clause C, satises also the clause D. Resolution with subsumption. Suppose we are given a literal a and clauses C and D such that a is the only literal satisfying both conditions a 2 C and a 2 D. In this case, the clause (C [ D) n fa; ag is called the resolvent by the literal a of the clauses C and D. We denote it by R(C; D). Let now F contain such clauses C and D. It is clear that adding R(C; D) to the formula does not change its satisability. However, it increases its size. To avoid this eect, we use this rule only if R(C; D) D. In this case we reduce the satisability problem for the formula F to the satisability problem for D)g. Elimination of a variable by resolution. Given a literal a, we construct the formula DP a adding to F all resolvents by a; removing from F all clauses containing a or a. Lemma 4.1 (Davis, Putnam, [6]) The formulas F and DP a are equi-satisable. This transformation can increase the size of a formula or/and the number of clauses in it, but we use this rule only if it does not increase the parameter under consideration (the number of clauses in a formula or the length of a formula). Note that in particular, this transformation does not increase any of the parameters of F if a is a pure literal, and thus it eliminates pure literals. Elimination of blocked clauses. The clause C is blocked for a literal a w.r.t. F if C contains the literal a, and the literal a occurs only in the clauses of F that contain the negation of at least one of the literals occurring in C n fag. (Note that F may or may not contain C.) In other words, there are no resolvents by a of the clause C and any other clause of the formula F . For a CNF-formula F and a literal a occurring in it, we dene the assignment ag j the clause fa; xg is blocked for a w.r.t. Fg: When the meaning of F is clear from the context, we omit F and write I(a). The notion of a blocked clause was introduced and investigated by Kullmann in [13, 14]. We use the following two facts about blocked clauses. Lemma 4.2 (Kullmann, [13, 14]) (1) If a clause C is blocked for a literal a w.r.t. F , then F , F n fCg and F [ fCg are equi-satisable. (2) Given a literal a, the formula F is satisable i at least one of the formulas F [a] and F [I(a)] is satisable. Application of the black and white literals principle. Let P be a binary relation between literals and formulas in CNF, such that for a variable v and a formula F , at most one of P (v; F ) and P (v; F ) holds. Lemma 4.3 Suppose that each clause of F that contains a literal w satisfying P (w; F ) contains also at least one literal b satisfying P (b; F ). Then F and F [fl j P (l; F )g] are equi- Proof. This is a particular case of the autarkness principle [20, 22, 17, 16]. We denote Suppose G is satisable. Consider a satisfying assignment I for the formula G. It is clear that the assignment I [ . On the other hand, each assignment satisfying F satises also G. ut 5 A bound w.r.t. the number of clauses. In this section we present Algorithm 5.1 which checks satisability of formula F in the time is a polynomial). This algorithm has two subroutines, Function REDUCEK and Function SPLITK . Function REDUCEK simplies the input formula using the transformation rules (see Sect. 4). Function SPLITK is intended for reducing the satisability problem for the input formula to the satisability problem for several simpler formulas. The execution of this algorithm can be viewed as follows: Function REDUCEK simplies the input formula, then SPLITK splits it into several formulas, REDUCEK sim- plies each of these, and so on. In the following, we denote by REDUCEK the formula that Function REDUCEK outputs on input F . Similarly we dene SPLITK Function REDUCEK . Input: A formula F in CNF. Output: A (simplied) formula in CNF. Method. (KR1) Elimination of 1-clauses. If F contains a 1-clause Repeat this step while F contains 1-clauses. (KR2) Application of the black and white literals principle. If each clause C in F that contains a (2;3 contains also a (3 F := F [fa j a is a (3 Elimination of a variable by resolution. Choose a literal a such that a or a occurs in F and maximal. If there are several such literals, then choose a literal with # a minimal. If 0, then F := DP a Repeat this step while F satises this condition. If F has been changed at steps (KR1){(KR3), then go to step (KR1), otherwise return F . ut Function SPLITK . Input: A formula F in CNF. Output: If F is satisable, then T rue, otherwise F alse. Method. (KS1) The empty formula. If containing the empty clause. If ; 2 F , then return F alse. Splitting into two subproblems. For each literal a occurring in F , construct two If there exists a literal a such that (K(F then choose the formulas F 1 and F 2 corresponding to this literal. If SPLITK returns T rue for at least one of these, then return T rue, otherwise return F alse. Splitting into four subproblems. Choose a literal a occurring in F . For each two literals b and c occurring in F [a] and F [a] respectively, construct four formulas F 11 =REDUCEK F 21 =REDUCEK F 22 =REDUCEK If there exist literals b and c such that then choose the formulas F 11 , F 12 , F 21 and F 22 corresponding to these literals. If returns T rue for at least one of these, then return T rue, otherwise return F alse. ut Algorithm 5.1. Input: A formula F in CNF. Output: If F is satisable, then T rue, otherwise F alse. Method. Return SPLITK (REDUCEK ut In Sect. 4 we explained that none of the steps of REDUCEK changes the satisability of the formula. The steps (KR1){(KR4) cannot be repeated more than K(F since none of them increases the number of clauses or the number of variables, and during each iteration (KR1){(KR4){(KR1) at least one of these quantities decreases. Each of these steps takes a polynomial time. Thus, REDUCEK does not change the satisability of the formula and returns the answer in a polynomial time. We now construct a tree which re ects the behaviour of Algorithm 5.1 (together with Functions REDUCEK and SPLITK ). Its internal nodes are labelled with formulas that Algorithm 5.1 splits at steps (KS3) and (KS4). Its leaves are labelled with formulas that satisfy the conditions of the steps (KS1) and (KS2). Each internal node labelled with F has two sons labelled with F 1 and F 2 or four sons labelled with F 11 , F 12 , F 21 and F 22 . Since the conditions of the steps (KS3) and (KS4) guarantee that the corresponding branching numbers do not exceed (6; 7; 6; 7), by Lemma 3.1 the number of leaves in the tree is at most is the input formula. Thus, the Algorithm 5.1 returns the answer in the time p(L(F ))(6; 7; 6; 7) K(F ) , where p is some polynomial. It remains to prove that Algorithm 5.1 performs correctly, i.e., each formula in CNF reduced by REDUCEK satises at least one of the conditions of the steps (KS1){(KS4). To prove this statement, we need two simple lemmas concerning the output of REDUCEK . Lemma 5.1 Let F 1 be the value of F at step (KR1) of Function REDUCEK at some point of time (maybe after eliminating several 1-clauses), F 2 be the corresponding output formula of Function REDUCEK , d and e be literals occurring in F 1 . (1) If d is a (1 ;1 (2) If d and e are (1 ;1 and there are no clauses in F that contain d and e simultaneously, then 2. Proof. (1). If REDUCEK modies once more the formula at steps (KR1) or (KR2), then at least one clause will be deleted from it. Otherwise, at least one clause will be eliminated at step (KR3) since K(F 1 ) K(DP d (2). W.l.o.g we suppose that F 1 contains at most one 1-clause, this clause (if any) contains a 1-literal and after eliminating this clause the formula F 1 does not contain 1-clauses at all. If F 1 contains exactly one 1-clause, then it will be eliminated at step (KR1), and one more clause will be eliminated by (1) since at least one of the literals d and e remains a Now let F 1 contain no 1-clauses. If REDUCEK modies this formula at step (KR2), then at least one clause, i.e., at least two occurrences will be eliminated. Now F has the value F 1 before the step (KR3). W.l.o.g. we suppose that of the step (KR3). At step (KR3) a 1 -literal will be chosen. After the rst application of DP , at least one of the literals d and e remains a 1 -literal and its negation or the literal itself remains in the formula. Hence, by (1) at least one more clause will be eliminated. ut Lemma 5.2 Let F be a formula in CNF. Then the formula REDUCEK does not contain 1-clauses, 1 -literals and (2;2)-literals. Proof. Function REDUCEK eliminates 1-clauses at step (KR1). If d is a 1 -literal or a (2;2)-literal, then K(DP d eliminates such literals at step (KR3). ut Theorem 5.1 Algorithm 5.1 performs correctly and stops in the time p(L)(6; 7; 6; 7) K < p(L)2 0:30897K , where L is the length of the input formula, K is the number of clauses in it, and p is a polynomial. Proof. We have shown above that it su-ces to prove that each formula F in CNF reduced by REDUCEK satises at least one of the conditions of the steps (KS1){(KS4). Suppose F does not satisfy the conditions of the steps (KS1){(KS4). We now consider all possible cases. We prove that each of this cases is impossible. All symbols have the same meaning as they have at the corresponding steps of Function SPLITK . For The key observation is that if the formula F contains few \frequent" variables, then after the transformation of F into F [a]; F [a] for some literal a we can apply Lemma 5.1. We now prove this more formally. Case 1: The formula F contains a (3 All clauses of the formula F containing the literal a disappear in F [a], all its clauses containing the literal a disappear in F [a]. Since a is a (3 4. Thus, in terms of the step (KS3), i.e., the condition of the step (KS3) is Case 2: The formula F contains a clause that contains a (3;3)-literal a and a (2;3 b. All three clauses of the formula F containing a disappear in F [a], all three clauses containing a disappear in F [a]. In addition, b becomes a (1 ;1 + )-literal in F [a]. Thus, in terms of the step (KS3) we have 1 1 by Lemma 5.1. Hence, i.e., the condition of the step (KS3) is satised. Case 3: The formula F contains a 2-clause consisting of 2-literals. All clauses of the formula F containing a disappear in F [a], all its clauses containing a disappear in F [a]. By Lemma 5.2, a is a (2;3 3. Similarly to Case 2, in terms of the step (KS3) we have 1 1. In addition, C becomes a 1-clause in F [a] and will be eliminated at step (KR1), i.e. 2 1. We now have 4, i.e., the condition of the step (KS3) is satised. Case 4: The formula F contains a 3 consisting of 2-literals. sg. All clauses of the formula F containing a i disappear in F [a i ], all its clauses containing a i disappear in F [a i ]. By Lemma 5.2, a i is a (2;3 In addition, a 2 and a 3 become in F [a 1 ]. If C is the only clause that contains a 2 and a 3 simultaneously, then by Lemma 5.1 we have K(F [a 1 ]) K(REDUCEK 2. Otherwise, a 3 becomes a (0;3 in F [a 2 ], i.e. K(F [a 2 ]) K(REDUCEK 2. Depending on which of these two alternatives takes place, we choose a 1 or a 2 as a in terms of the step (KS3). We now have i.e., the condition of the step (KS3) is satised. Case 5: The conditions of Cases 1{4 are not satised. Since the step (KR2) does not change F , it contains only 3-literals. Since F does not satisfy the condition of the step (KS3), for each literal a, 3: Let a be a literal occurring in F . We now prove that there exist literals b and c such that the condition of the step (KS4) is satised. We consider three sub-cases (the rst two of them are similar to Cases 2{4). Case 5.1: There exists a clause in F [a], such that a 1 is a (3;3)-literal and a 2 is a 2-literal. We suppose in terms of the step (KS4). Similarly to Case 2, 11 4, 12 3. Case 5.2: There exists a clause in F [a], such that a 1 is a (2;3)-literal and a are 2-literals. We suppose in terms of the step (KS4). Similarly to Cases 3{4, 11 3, 12 4, or 11 4, 12 3. (Note that by (5.1) the literals a i and a j cannot occur together in two clauses.) Case 5.3: The conditions of Cases 5.1{5.2 are not satised. Since by (5.1) the steps (KR1){(KR3) do not reduce the number of clauses in the formula F [a] and the conditions of Cases 5.1{5.2 are not satised, this formula consists of clauses that contain only (2;2)-literals and clauses that contain only (3;3)-literals. Since exactly three clauses of the formula F disappear in F [a], all occurrences of any (3;3)-literal cannot disappear in F [a]. Hence, it contains at least one (2;2)-literal b 1 . Let b 1 be a 2-literal. The formula F does not contain 2-clauses by (5.1). Similarly, it cannot contain two clauses, each containing the literals a and b 1 simultaneously. Thus, F [a] contains a 3 are 2-literals. We choose the literal b 1 as b. Both clauses of the formula F [a] that contain b 1 disappear in F [a; b 1 ]. In addition, these clauses contain only (2;2)-literals which become literals. Thus, by Lemma 5.1 3: On the other hand, two clauses of F [a] containing b 1 disappear in F [a; b 1 ], and two more clauses disappear by Lemma 5.1 since b 2 and b 3 become )-literals. We note that by (5.1) the formula F contains only one clause that contains b 2 and b 3 simultaneously. Thus, We now have that in each of Cases 5.1{5.3 there exists a literal b in F [a] such that 7. A literal c can be chosen similarly. Hence, the condition of the step (KS4) is satised. ut 6 A bound w.r.t. the length of a formula. In this section we present Algorithm 6.1 which checks satisability of formula F in the time p(L(F ))2 0:10299L(F ) (where p is a polynomial). As in the previous section, we dene two subroutines, Function REDUCE L and Function SPLIT L . We use them similarly to REDUCEK and SPLITK : Function REDUCE L simplies the input formula, then SPLIT L splits it into several formulas, REDUCE L simplies each of these, and so on. Function REDUCE L . Input: A formula F in CNF. Output: A (simplied) formula in CNF. Method. (LR1) Elimination of 1-clauses. If F contains a 1-clause Repeat this step while F contains such a clause. Subsumption. If F contains two clauses C and D such that C D, then F := fDg. Repeat this step while F contains such clauses. Elimination of blocked clauses. If F contains a blocked clause C, then F := fCg. Repeat this step while F contains blocked clauses. Resolution with subsumption. If F contains two clauses C and D such that R(C; D)D, then F := [fR(C;D)g. Repeat this step while F contains such clauses. Elimination of a variable by resolution. Choose a literal a such that a or a occurs in F and Repeat this step while F satises this condition. If F has been changed at steps (LR1){(LR5), then go to step (LR1), otherwise return F . ut Function SPLIT L . Input: A formula F in CNF. Output: If F is satisable, then T rue, otherwise F alse. Method. (LS1) The empty formula. If containing the empty clause. If ; 2 F , then return F alse. containing no 2-clauses. If F does not contain 2-clauses, apply Algorithm 5.1 to F and return its answer. Splitting into two subproblems. For each literal a occurring in F , construct two If there exists a literal a such that (L(F ) choose the formulas F 1 and F 2 corresponding to this literal. If SPLIT L returns T rue for at least one of these, then return T rue, otherwise return F alse. ut Algorithm 6.1. Input: A formula F in CNF. Output: If F is satisable, then T rue, otherwise F alse. Method. Return SPLIT L (REDUCE L ut Similarly to REDUCEK , Function REDUCE L does not change the satisability of the formula and returns the answer in a polynomial time. We now construct a tree which re ects the behaviour of Algorithm 6.1 (together with Functions REDUCE L and SPLIT L ). Its leaves are labelled with formulas that satisfy the conditions of the steps (LS1){(LS3). Algorithm 6.1 processes in a polynomial time formulas satisfying the conditions of the steps (LS1) and (LS2), and passes formulas satisfying the condition of the step (LS3) to Algorithm 5.1 which processes such a formula F in the time is a polynomial (note that F contains no 2 -clauses since 1-clauses are eliminated at step (LR1)). Internal nodes of our tree are labelled with formulas that Algorithm 6.1 splits at step (LS4). Each internal node labelled with F has two sons labelled with F 1 and F 2 . Since the condition of the step (LS4) guarantees that the corresponding branching number does not exceed 3.2 the running time of Algorithm 6.1 is upper bounded by q(L(G))(6; 7; 6; 7) L(G)=3 for input formula G, where q is a polynomial. It remains to prove that Algorithm 6.1 performs correctly, i.e., each formula in CNF reduced by REDUCE L satises at least one of the conditions of the steps (LS1){(LS4). We need three simple lemmas concerning the output of REDUCE L to prove this statement. Lemma 6.1 Let F be a formula in CNF, d be a literal occurring in it. (1) If d is a (1;2 )-literal occurring in a 3 -clause, then L(DP d (2) If d is a (1;1 )-literal occurring in a 2 -clause, then L(DP d 2: (3) If d is a (1;1)-literal, then L(DP d 2: Proof. By straightforward calculations. ut Lemma 6.2 Let F 1 be the value of F at one of the steps (LR1){(LR5) of Function REDUCE L , F 2 be the corresponding output formula of Function REDUCE L , d be a literal occurring in (1) If d is a (1;2 )-literal occurring in a 3 -clause, then (2) If d is a (1;1 )-literal occurring in a 2 -clause, then 2: Proof. (1). Note that any change of the formula at steps (LR1){(LR4) is a removal of some clauses and/or a change of other clauses by their subsets, and thus results in decreasing the length of the formula by at least one occurrence. If the formula does not change any more before the step (LR5), then at least one occurrence will be removed by Lemma 6.1(1). (2). Suppose that after F gets the value F 1 , Function REDUCE L modies the formula before the step (LR4). This modifying cannot result in increasing the length of a clause or in increasing the number of occurrences of a literal. Moreover, if the formula changes at these steps, at least one occurrence is removed. If only one occurrence is removed, then at least one of the literals d, d remains in the formula. If this is the occurrence of d, then by Lemma 6.1(1) at least one more occurrence will be removed at step (LR5), and (2) holds. Otherwise, d becomes a pure literal and will be removed at step (LR5), thus (2) holds again. Suppose now that REDUCE L does not modify the formula before the step (LR5). Then, (2) follows from Lemma 6.1(2). ut We use the following simple properties of the formula reduced by Function REDUCE L in the proof of 6.1 without explicit mentioning. Lemma 6.3 Let G. Then (1) F does not satisfy any of the conditions of the steps (LR1){(LR5); (2) there are no 1-clauses in F ; (3) there are no pure literals and (1;1)-literals in F ; (4) there are no 2-clauses in F containing a 1-literal; (5) for any literal d occurring in F , } d for any 1-literal d occurring in F , } d +# d 7; (7) for any literal d occurring in F , } d +# d 5. Proof. (1) is trivial; (2){(7) since F does not satisfy the conditions of the step (LR1)and see also Lemma 6.1. ut Theorem 6.1 Algorithm 6.1 performs correctly and stops in the time p(L)(6; 7; 6; 7) L=3 < p(L)2 0:10299L , where L is the length of the input formula and p is a polynomial. Proof. We have shown above that it is su-cient to prove that each formula F in CNF reduced by REDUCE L satises at least one of the conditions of the steps (LS1){(LS4). Suppose F does not satisfy the conditions of the steps (LS1){(LS3). We now consider the two possible cases: when F contains at least one 1-literal a, and when it contains no 1-literals. We prove that in each of these cases F satises the condition of the step (LS4). Informally, this is done as follows. In the rst case we show that the assignment I(a) contains many literals. The second case is handled by careful examination of sub-cases: we choose a 2-clause fc; dg in F and examine the following sub-cases: there are many occurrences in the clauses containing the literal c (or d); F contains many occurrences of the literal c (or d). In these two sub-cases we show that many occurrences are eliminated during the transformation of F into F [c]; F [c] (or F [d]; F [d]) and the subsequent elimination of 1-clauses obtained from the clauses containing the literals c; c (or d; d). The third sub-case is: none of the previous sub-cases holds. In this sub-case one of c and d is a (2;2)-literal occuring only in 2- and 3-clauses which allows us to use Lemma 6.1 after one occurrence of this literal is eliminated during the splitting by another literal. In the following we denote F 2 and a literal a have the same meaning as they have at step (LS4) of Algorithm 6.1. Case 1: The formula F contains a 1-literal a. be the only clause containing a. Let and } a +# a 7 by Lemma 6:3, we have # a max(7 } a ; be the number of 3 -clauses among all clauses that contain a. During the transformation of the formula F into F 2 , all occurrences of the literal a and all clauses containing a will be eliminated. Thus, +# a 5: During the transformation of the formula F into F 1 , the clause D and all occurrences of the literal a will be eliminated; all 2-clauses containing a will be eliminated too (at step (LR1)). Since D is not blocked for any a i w.r.t. F , for each there exists clause D i in F such that D i \ fa; a g. All these clauses will be eliminated since all clauses fa; a i g are blocked for a w.r.t. F (i.e. fa; a clauses containing a, all clauses D i and the clause D are distinct. Thus, we have r+(7 r)+(# a We now have i.e., the condition of the step (LS4) is satised 1 . Case 2: The formula F does not contain 1-literals. Since F does not satisfy the condition of the step (LS3), it contains a 2-clause dg. Let us note that The formula F does not contain other clauses that contain the literal d and the literal c, or the literal d and the literal c, or the literal d and the literal c, simultaneously. This proposition is true since F does not satisfy the conditions of the steps (LR2) and (LR4). Also, in this case I(c) = fcg and I(d) = fdg by (6.2) and the denition of I(: : :). Let us denote one of the literals c, d by a (we shall choose later which one). We denote the remaining literal by b. Let be the number of 3 -clauses among all clauses that contain a (note that # a 1), let be the number of 3 -clauses among all clauses that contain a. We now count the occurrences that disappear during the transformation of the formula F into F [I(a)] (= F [a]) and F [a] and subsequent elimination of 1-clauses resulted from the 2-clauses containing the literals a and a. During the transformation of the formula F into F 1 , all occurrences of the literal a and all clauses containing a will be eliminated; all 2-clauses that contain a will be eliminated too (at step (LR1)), there are (# a clauses. 1 Kullmann and Luckhardt [16] prove for a very similar algorithm that in this case thus, (5; 17) in the condition of the step (LS4) can be replaced by (8; 12). However, this would not improve our upper bound while would make the proof longer; thus, we present here this simpler proof. The formula F [a] contains the 1-clause fbg obtained from the 2-clause D. It will be eliminated at step (LR1). If F contains other 2-clauses containing a, the clauses obtained from them will be eliminated at step (LR1) too. The number of such clauses is (# a 1 ). During the transformation of the formula F into F 2 , all occurrences of the literal a and all clauses containing a or b will be eliminated. By (6.2), D is the only clause containing b in which the literals a; a can occur. Thus, even if we count only occurrences that disappear in clauses containing a or a before the rst application of the step (LR2), we have We consider several sub-cases. Case 2.1: } c 6 and } d 6. From (6.3) and (6.4) we have Thus, i.e., the condition of the step (LS4) is satised. Case 2.2: # c 3 or # d 3. Let a 2 fc; dg be the literal for which # a 3 holds, let b be the remaining literal in fc; dg. Thus, we have a +# a 22: Thus, i.e., the condition of the step (LS4) is satised. Case 2.3: # or } d 5. Let a 2 fc; dg be a literal such that } a be the remaining literal in fc; dg. Then 4 } b 5 and } a } b . In this case b is a 2-literal, it occurs in D and in one other 2- or 3-clause. We denote this clause by C b . The literal b becomes a (1;2)-literal in F [I(a)] (= F [a]). We now complete the proof by showing that several literal occurrences will be eliminated in addition to the occurrences counted by (6.3). There are three sub-cases. Case 2.3.1: At least two occurrences are eliminated before the rst application of step (LR2) in addition to the occurrences counted by (6.3). In this case, +(# a )9; a +# a 20: Case 2.3.2: Exactly one occurrence is eliminated before the rst application of step (LR2) in addition to the occurrences counted by (6.3). We remind that (6.3) counts only occurrences to the clauses containing one of the literals a,a. By (6.2) the occurrence of the literal b to the clause C b or any occurrence of the literal b cannot be eliminated during the transformation of the formula F into the formula F [a]. If we counted any of these occurrences in the # a term of (6.3), then it was an elimination of a 1-clause fbg obtained from a 2-clause fa; bg. Since it is impossible to have two distinct identical clauses fa; bg in the formula, the other two occurrences (another occurrence of the literal b and the occurrence of the literal b in C b ) eliminated at this moment are not counted by (6.3); this fact contradicts the assumption of Case 2.3.2. Thus in this case b remains a (1;2)-literal before the rst application of the step (LR2). By Lemma 6.2(1) at least one more occurrence will be eliminated after that and thus (6.5){(6.7) hold. Case 2.3.3: No occurrences are eliminated before the rst application of step (LR2) in addition to the occurrences counted by (6.3). Similarly to Case 2.3.2, b remains a (1;2)-literal before the rst application of the step (LR2) during transformations of the formula F [a] by Function REDUCE L . Thus, if jC b two more occurrences are eliminated by Lemma 6.2(2), and (6.5){(6.7) hold. If, however, jC then we can apply only Lemma 6.2(1), but in this case } a } and thus 10; Hence, in all three sub-cases 2.3.1{2.3.3, i.e., the condition of the step (LS4) is satised. ut 7 Conclusion and further work In this paper we have improved the existing upper bounds for SAT with respect to K (the number of clauses) and L (the length of a formula). The key point in our algorithms and proofs is the black and white literals principle, a new transformation rule which can be viewed as a re-formulation of two previously known principles: the autarkness [20, 22, 17, 15] and the generalized sign principle [16]. Our proofs (as well as the proofs of the previous upper bounds [21, 16]) are neither short nor elegant. It would be a kind of breakthrough to nd a compact way to present proofs of upper bounds for splitting algorithms. It is believable that such a way could lead to even better bounds, since currently our possibilities to create new heuristics and prove the corresponding upper bounds are limited by the length of a comprehensible proof. On the other hand, it is a challenging problem to prove (more) tight lower bounds for (some class splitting algorithms: currently, the exponential lower bounds for resolution proofs (see, e.g., [27]) are far enough from the known upper bounds for splitting algorithms. (Most of splitting algorithms can be viewed as resolution proofs and vice versa.) Another direction for work is to nd randomized algorithms that give better upper bounds for SAT (or to nd a way how to apply modern randomized algorithms that have already been invented for 3-SAT in recent breakthrough papers [23, 24]). Also, it remains a challenging problem to nd a less-than-2 N upper bound where N is the number of variables. --R Tautology proof systems based on the splitting method Less than 2 n satis Approximation algorithms for Max Sat: a better performance ratio at the cost of a longer running time Exponential upper bounds for the satis A machine program for theorem-proving A computing procedure for quanti Algorithms for Satis Separating the signs in the satis Two new upper bounds for SAT Local Search Algorithms for SAT: Worst-Case Analysis Hard formulas for SAT local search algorithms SAT local search algorithms: worst-case study New methods for 3-SAT decision and worst-case analysis Investigations on autark assignments Algorithms and their complexity Obere Komplexit Parametrizing above guaranteed values: MaxSat and Max-Cut New upper bounds for MaxSat Upper bounds for covering problems Solving satis An Improved Exponential-time Algorithm for k-SAT Solving 3-satis ability in less than 1:579 n steps Pure literal look ahead: An O(1:497 n The Complexity of Propositional Proofs A satis --TR --CTR Rainer Schuler, An algorithm for the satisfiability problem of formulas in conjunctive normal form, Journal of Algorithms, v.54 n.1, p.40-44, January 2005 Bolette Ammitzbll Madsen, An algorithm for exact satisfiability analysed with the number of clauses as parameter, Information Processing Letters, v.97 n.1, p.28-30, January 2006 Evgeny Dantsin , Andreas Goerdt , Edward A. Hirsch , Ravi Kannan , Jon Kleinberg , Christos Papadimitriou , Prabhakar Raghavan , Uwe Schning, A deterministic (2 - 2/(k+ 1))n algorithm for k-SAT based on local search, Theoretical Computer Science, v.289 n.1, p.69-83, 23 October 2002 Ramamohan Paturi , Pavel Pudlk , Michael E. Saks , Francis Zane, An improved exponential-time algorithm for k-SAT, Journal of the ACM (JACM), v.52 n.3, p.337-364, May 2005 Tobias Riege , Jrg Rothe , Holger Spakowski , Masaki Yamamoto, An improved exact algorithm for the domatic number problem, Information Processing Letters, v.101 n.3, p.101-106, February, 2007 Ryan Williams, Algorithms for quantified Boolean formulas, Proceedings of the thirteenth annual ACM-SIAM symposium on Discrete algorithms, p.299-307, January 06-08, 2002, San Francisco, California Haiou Shen , Hantao Zhang, Improving exact algorithms for MAX-2-SAT, Annals of Mathematics and Artificial Intelligence, v.44 n.4, p.419-436, August 2005 Jens Gramm , Edward A. Hirsch , Rolf Niedermeier , Peter Rossmanith, Worst-case upper bounds for MAX-2-SAT with an application to MAX-CUT, Discrete Applied Mathematics, v.130 n.2, p.139-155, 15 August Edward A. Hirsch, Worst-case study of local search for MAX-k-SAT, Discrete Applied Mathematics, v.130 n.2, p.173-184, 15 August Laurent Simon , Daniel Le Berre , Edward A. Hirsch, The SAT2002 Competition, Annals of Mathematics and Artificial Intelligence, v.43 n.1-4, p.307-342, January 2005 Oliver Kullmann, Lean clause-sets: generalizations of minimally unsatisfiable clause-sets, Discrete Applied Mathematics, v.130 n.2, p.209-249, 15 August Rolf Niedermeier , Peter Rossmanith, An efficient fixed-parameter algorithm for 3-hitting set, Journal of Discrete Algorithms, v.1 n.1, p.89-102, February
SAT;worst-case upper bounds
594333
Alternative Translation Techniques for Propositional and First-Order Modal Logics.
We describe and analyze techniques, other than the standard relational/functional methods, for translating validity problems of modal logics into first-order languages. For propositional modal logics we summarize the &squ;-as-Pow method, a complete and automatic translation into a weak set theory, and then describe an alternative method, which we call i>algebraic, that achieves the same full generality of &squ;-as-Pow but is simpler and computationally more attractive. We also discuss the relationships between the two methods, showing that &squ;-as-Pow generalizes to the first-order case. For first-order modal logics, we describe two extensions, of different degrees of generality, of &squ;-as-Pow to logics of rigid designators and constant domains.
Introduction Translation techniques have oered valuable aid for the study and development of theorem provers for non-classical logics, and their nave use can be ascribed to tradition. The rst systematic study of the relational translation was carried out within Correspondence Theory (van Benthem, 1983). In later years further studies of the relational and similar techniques for computational purposes were initiated ((Ohlbach, 1991; Ohlbach, 1993; Nonnengart, 1993); see also (Ohlbach et al., 2001) for a comprehensive survey). The eld has revealed itself to be fruitful and interesting. The relational translation of a modal the rst-order formula RT('; x) (it has also been called ST , for \stan- dard translation", but we will reserve this designation for \set-theoretic inductively dened by c 2001 Kluwer Academic Publishers. Printed in the Netherlands. Alberto Policriti, and Matteo Slanina Here, each propositional variable p i is mapped one-to-one to a unary predicate symbol P i . Let RT(') be If axiomatizes a frame-complete logic, then Here f denotes validity of ' in the logic axiomatized by , and denotes classical (in this case second-order) validity. The second-order quantiers in RT(') can be removed, since they are universal and appear only as prex to the consequent of an implication. However, the second-order character of RT( is, in general, not eliminable, since, in this case, the second-order (universal) quantication appears in the antecedent of an implication. This makes the technique inapplicable for computational purposes, since the target language is highly undecidable (non-arithmetic, cf. (van Benthem, 1983)). Whenever RT( can be proved equivalent to a rst-order formula, the logic axiomatized by is said to be rst-order denable and the modal validity problem can be solved by a rst-order theorem prover. Determining when RT( a rst-order equivalent is the main topic of correspondence theory. In some cases it can even be done automatically (e.g. Sahlqvist formulae, the SCAN procedure (Gabbay and Ohlbach, 1992)). Unfortunately, not all modal logics have a rst-order (frame) correspondent, and, even when they do, there is no complete algorithm to nd one. A lot of work has been done to avoid these di-culties, resulting in some extensions of the class of tractable logics and, especially, in improved performance (Nonnengart, 1993; Ohlbach, 1998). In particular, (Ohlbach, 1998) proposed a way to incorporate axioms and rules of Hilbert calculi into relational translations, thus creating a hybrid system able to deal with the axioms that cannot directly be captured by the translation. A dierent perspective on the problem is taken in the approach proposed in (D'Agostino et al., 1995), where the 2-as-Pow translation was rst introduced. The novelty of 2-as-Pow lies in the fact that, instead of determining which modal theories can be mapped into a rst- order language, a fragment of second-order logic is used as a framework for the translation. It turns out that the fragment required is very small and can be specied as a weak rst-order axiomatic set theory. Under this theory, the resulting translation is sound and complete for all frame-complete (not only with respect to a rst-order class of frames) Alternative Translation Techniques for Propositional and First-Order Modal Logics 3 logics. More specically, given a frame-complete logic with axiom and a formula ', the 2-as-Pow translation ( is such that where is the weak axiomatic set theory that will be dened in Section 2. Several extensions of this method were discussed in subsequent papers (Montanari and Policriti, 1997a; Montanari and Policriti, 1997b; van Benthem et al., 1998), including a variation (van Benthem et al., 1997) that allows for completeness for any logic (not necessarily frame- complete) and another with strong decidability results (D'Agostino et al., 1996). In this paper we systematically present set-theoretic and algebraic translation techniques that are less restrictive than RT and its vari- ations, and whose range includes propositional modal logics as well as an important class of rst-order modal logics. After a short review of the original 2-as-Pow translation (Section 2), we introduce a new, very natural, translation method, that we call algebraic. This algebraic translation, suggested by the algebraic semantics for modal logics, achieves completeness for all logics in a simpler way than 2- as-Pow. We point out strong relationships between 2-as-Pow and the algebraic method, showing a new way to obtain the former from an algebraic point of view (Section 3). We show that, in a certain sense to be made precise later, 2-as-Pow and the algebraic translation coincide at the propositional level. However, no such strong relationship holds when we try to extend these methods to handle rst-order modal log- ics: 2-as-Pow can be so extended, but for the algebraic method there seems to be a signicant stumbling block (unless we allow the target language of the translation to be second-order, which is unacceptable). The relational translation and its derivations can be extended, in their restricted setting, to rst-order modal logics in a fairly straightforward manner. In Section 4, we collect and organize some results (Montanari et al., 2000a; Montanari et al., 2000b) on the extension of 2-as-Pow to the important class of rst-order modal logics of rigid designators and constant domain, which is signicant given the aforementioned observations. Section 5 contains some lengthy proofs that would have broken the continuity of the exposition if put in the main text. We conclude with some discussion on open problems and possible future work. We have tried to use a commonly understood symbolism. Some less common notations are the ones for logical syntactic derivation in classical and modal logic; 4 Angelo Montanari, Alberto Policriti, and Matteo Slanina 'Ks syntactic derivation in the modal calculus K s , that is, K plus the Instantiation Rule; truth in a model, semantic consequence in classical logic; truth in a world or model, frame semantic consequence in modal logic. 2. 2-as-Pow Translation (Propositional Case) The 2-as-Pow translation, at the propositional level, can be described as an extension to the 2 operator of the set-theoretic interpretation of the propositional connectives ^, _, and :. The approach brie y described here was rst presented in (D'Agostino et al., 1995) and subsequently developed in (van Benthem et al., 1997) and (van Benthem et al., 1998). The key idea is simply to replace the accessibility relation R of the possible world semantics by the membership relation 2. Accordingly, a world y accessible from x becomes an element of x, and taking a further step from y via R amounts to moving into y to inspect one of its elements. This viewpoint allows us to identify a frame F with its support w. Moreover, since we clearly want all worlds y accessible from a given world x in a frame w to be themselves elements of w, it is natural to require that w be a transitive set: w P(w). Valuations for propositional variables, i.e. sets of worlds, can now be identied with worlds themselves. 2-as-Pow is designed in such a way that the translation ' of a compound formula ' turns out to be the set of all worlds where ' is true. For propositional connectives we simply follow the Boolean tradition and say that In order to complete the denition it is su-cient to calculate the Kripke semantics of the 2 operator replacing R by 2: which is to say: 2' holds true at x if and only if x is a subset of ' . This allows us to complete our denition by putting Alternative Translation Techniques for Propositional and First-Order Modal Logics 5 It is possible to show that the collection of the following set-theoretic principles su-ces for proving the soundness and completeness of the translation: z Dening Trans(w) as w P(w) it can be shown that When a modal logic is characterized by a nite collection of Hilbert axioms, whose conjunction is , these results generalize to where Axiom (w) stands for 8~x (w ), where ~x is the list of all the variables in distinct from w. Thus when is frame-complete an exact correspondence between modal and set-theoretic derivability is guaranteed It turns out that the theory is almost minimal in providing a formal counterpart to the idea underlying the 2-as-Pow translation. Let us call MM (an acronym for \Minimal Modal") the theory whose axioms result from those of by replacing x 2 P(y) $ x y with the following two theorems of 1 As a technical note, if we were to deal with a logic with an innite axiom set , we could reformulate the result as follows (use skolemization on w/c and the Deduction Theorem for rst-order logic to obtain the equivalence): 6 Angelo Montanari, Alberto Policriti, and Matteo Slanina where u \ v is an abbreviation for u n (u n v). Although MM is even weaker than it can still drive the 2-as-Pow translation. A further possible simplication would be to eliminate the binary union symbol, observing that, by exploiting complementation, it can be dened in terms of set dierence. Theorem 1 (Completeness of 2-as-Pow). If 'Ks ', then MM Theorem 2 (Soundness of 2-as-Pow). If '. 3. Algebraic Translation (Propositional Case) The same mental shift that proceeds from relational semantics to the relational translation enables one to discover the algebraic translation as a straightforward rewriting of algebraic semantics in syntactic terms. The basic objects playing the role of Kripke frames are Boolean algebras with one extra operator (Bull and Segerberg, 1984; Blackburn et al., 2001). Denition 3. Consider the rst-order language of Boolean algebras (with meet u, join t, complementation 0 , bottom element 0, and top element 1) and an additional operator ' . Dene the theory MA (for Modal Algebra) to be the rst-order theory of Boolean algebras with the two additional axioms: Denition 4. Let us map each propositional symbol p to a distinct individual variable x p of the target rst-order language and dene the algebraic translation ' # of a modal formula ' by the following structural induction: Alternative Translation Techniques for Propositional and First-Order Modal Logics 7 Based on this denition and following the truth denition of algebraic semantics, we put It is straightforward to prove that this translation is sound and complete for all (nitely axiomatizable) modal logics, since the algebraic semantics is complete for all logics (in fact, it is equivalent to the general frame semantics). Theorem 5. Given a modal logic with axiom and a formula ', The algebraic translation converts a validity problem in any modal logic to a validity problem of rst-order logic, thus permitting the use of any rst-order theorem prover. As a validity checking method, this is slower than the use of the relational or functional translation, when the latter is applicable, but it has the advantage of being a completely syntactic method and it can turn out to be extremely useful for experimenting with new logics. 2 A further advantage of this technique is that it also works, with no modication, for multi-modal logics: just introduce one ' operator for each of the modalities involved and augment the underlying theory with a pair of axioms, of the form of (1), for each of them. What we are left with after the algebraic translation is a deduction problem in equational rst-order logic. Tools that deal with such prob- lems, such as EQP, are available and can be used for modal theorem proving based on the algebraic translation. Equational problems can be solved particularly e-ciently when rewriting techniques can be applied. In the case of modal logic, Foret (Foret, 1992) discovered con uent (modulo associativity and commutativity of the Boolean operators) and terminating rewriting systems for the logics K, D, T, and S5. The rewriting systems are oriented versions of the equational axioms of modal algebras and can be seen as a specialized solver for the algebraic translation of the formulae. For example, Foret's system for K adds the oriented versions of (1) to Hsiang's system for propositional logic| an oriented version of the equational theory of Boolean rings with multiplicative identity (Hsiang, 1982; Hsiang, 1985): We did some simple experimentation with a raw algebraic translation given to EQP (McCune, 1997), and were able to prove 31 out of the 72 problems of (Gore et al., 1997) in seconds (at most a few minutes in a few cases) on a Pentium class computer. 8 Angelo Montanari, Alberto Policriti, and Matteo Slanina 3.1. Relations with 2-as-Pow. If one looks at the original formulation of the 2-as-Pow translation, there is one peculiar feature: the membership symbol 2 does not appear explicitly. In fact, the translation seems to be expressed in an almost equational language and looks very much like the algebraic translation. Moreover, the axioms of MM are strikingly similar to those of modal algebras. We will now prove that these remarks have indeed a deep under- pinning: there is an equational theory of sets, capturing the Boolean properties of and MM , that proves the propositional 2-as-Pow translation sound and complete. In other words, at the propositional level, the algebraic and the 2-as-Pow translation are essentially equivalent. However, as will be shown in the next section, there is no simple way to lift this equivalence to the rst-order level. One would be inclined to use, as an equational set theory, a theory of Boolean algebras. However, we cannot hope to have a complementation function, since there is no constant for a top element. Hence, we add a relativized complementation which mimics the behavior of set dierence. This turns out to be all we need and the following axioms are su-cient. Denition 6. We dene the theory by the universal closure of the following axioms: Intuitively, the rst ten axioms describe a theory of \Boolean algebras without maximum element". It can be shown that the models of these axioms are such that the cone of elements below any given element is a Boolean algebra. This kind of structure is exactly a Boolean ring in the sense of, e.g., (Jacobson, 1951), that is, an idempotent ring with no multiplicative null element. This allows us to establish a two-way correspondence between models of and models of MA. Using this correspondence and the 2-as-Pow translation, which can be applied directly to the language of , we can prove the following equivalence: Alternative Translation Techniques for Propositional and First-Order Modal Logics 9 Theorem 7. Consider a nitely axiomatizable modal logic, with axiom , and a modal Proof. See Section 5. Corollary 8. If axiomatizes a frame-complete modal logic, then Proof. 'Ks ' by soundness and completeness of 2- as-Pow. By soundness and completeness of the algebraic translation (Theorem 5), To conclude apply Theorem 7. 4. 2-as-Pow Translation (First-Order Case) The analogy between the 2-as-Pow and the algebraic translation does not lift from the propositional to the rst-order level. An extension of the algebraic method to rst-order modal logics would necessitate employing a theory of complete Boolean algebras, which is second- order, therefore leading again to problems of the same sort as those mentioned in Section 1. However, we have successfully extended 2- as-Pow to an important class of rst-order modal logics maintaining the target language of the translation at the rst-order level. Hence, 2-as-Pow and the semantics it is based on can be seen as genuine generalizations of the algebraic models; only at the lowest (propositional) level do the two happen to be equivalent. Here we describe two extensions of the 2-as-Pow translation method to the class of rst-order modal logics with rigid designators and xed domains, known as Q1 systems (see (Garson, 1984; Fitting, 1993; Fitting and Mendelsohn, 1998)). Informally, the language of such logics is a rst-order language together with a 2 modality. A model M consists of a frame hW; Ri (possibly with restrictions dened by an underlying propositional modal logic) plus a rst-order logic interpretation with a single domain D, an interpretation FC for constants and one F F for function symbols, and|the only part that varies with worlds|the interpretation F P of predicate symbols: There is a sound and complete calculus for any such logic, obtained by adding to the propositional calculus for the underlying modal logic Alberto Policriti, and Matteo Slanina the axioms and rules of rst-order logic|the Universal Quantier Axiom schema and the Universal Generalization Rule|and the Barcan for every formula '(x). Both translations will be targeted to a two-sorted rst-order logic. This is done essentially for better understanding and ease of implemen- tation: in principle everything could be done in a one-sorted language (see (Slanina, 2001) for details; for a general discussion of the reducibility of many-sorted to one-sorted logic see also (Davis, 1993)). If L is the rst-order modal language, we dene L 0 to be a two-sorted language where the sort term denotes L-terms and the sort set denotes objects of the translation (possible worlds and sets thereof). The underlying theory is modied accordingly, with the third axiom now becoming 8x8y We call this modied theory . The rst translation (Montanari et al., 2000a) works for the fragment of the Q1 language where modal operators cannot appear in the scope of quantiers (below: \locally quantied formulae"). This case appears often in practice and can be treated using the observation that sentences can be decomposed into two disjoint parts: one rst- order, but classical, and the other modal, but propositional. 3 More precisely, a sentence of this fragment can be written as are classical rst-order sentences and the formula we obtain by replacing each ' i by a propositional variable p i , for short), is a propositional modal formula. The simple, but crucial, observation at this point is that derivability can be split accordingly, as stated in the following lemma. Lemma 9. There is a nite set of classical propositional formulae (over the variables 'g for all 2 , and, for all , Proof (sketch). Clearly, 3 This is the case, for instance, of Manna and Pnueli's temporal frame-work (Manna and Pnueli, 1995), where most specications can be conned to this restricted form of quantication. Alternative Translation Techniques for Propositional and First-Order Modal Logics 11 Let be the set of propositional formulae on that are valid under the substitution fp 1 7! ' g. Then, it is easy to see that the right-hand side of (2) is equivalent to (p nite modulo logical equivalence. Using the above results, the translation is dened as follows. Denition 10. As an auxiliary notation, we introduce a transformation that adds an extra variable (representing a world in a Kripke frame) to the arguments of each predicate symbol ( ~ t are terms): Now x a variable w of sort set. The translation ( of a locally quantied respect to the logic axiomatized by , is expressed as follows: where Here Axiom 2 (w) and are the same as in the propositional case, with all bound variables of sort set. The adequateness theorems are similar to the ones for the propositional case: Theorem 11 (Completeness). Theorem 12 (Soundness). Q1 F . The proofs of these theorems can be found in (Montanari et al., 2000a). We omit them here, since our next translation is stronger. The second translation works for the full Q1 language and is developed along a dierent line. Alberto Policriti, and Matteo Slanina Denition 13. The set-theoretic translation ( with respect to the logic with axiom (a propositional modal is the formula Trans(w) and Axiom (w) are the same as in the propositional case, while ' (w) encodes the semantics of quantiers. The notation is as follows: variables occurring|free or not|in ' are denoted by ~x, which stands for x are fresh function symbols f for some subformulae of ' (to be specied below), with signature set term set. Each term t , with subformula of ', represents the set of worlds where is true, in the frame w and under variable assignment ~x. Formally, let be a superset of the set of the variables appearing in '. Then are the free variables in ', with 2. ' (w) is simply the conjunction, over all subformulae of ', of the formulae (w), where is dened as (w) => > > > < if is 8x i and x are the free variables in , with otherwise. In analogy to the propositional case, we now have the following theorems, whose proofs are given in Section 5. Theorem 14 (Completeness). Theorem 15 (Soundness). Example. Let ' be the formula The superscripts mark the main operator of each of the eight subformulae of '. Let be the axiom of seriality, Alternative Translation Techniques for Propositional and First-Order Modal Logics 13 some freedom of notation (e.g., we use 9 , whose direct translation is obtained patching the one for 8 in the obvious way), the translation is the following: Axiom 5. Proofs 5.1. Proof of Theorem 7. We only give a sketch of the proof. The interested reader is referred to (Slanina, 2001) for the details. \Only if". The proof is by contradiction: assume that i.e., by completeness of rst-order logic, . This means there exists a model A of E and an assignment on A such that, putting a A P A (a) ; Let B be the model for MA dened in the obvious way taking as domain the lower cone of A with supremum a (see (Slanina, 2001) for details). Each term t in the language of E such that any occurrence of n has w as its left argument can be mapped to a term t 0 in the language of MA, and the assignment can be mapped into the assignment 0 on B dened by 0 in such a way that and 14 Angelo Montanari, Alberto Policriti, and Matteo Slanina It is then straightforward to prove that B 6 ( \If". The argument is analogous to the \only if" case, proceeding by contradiction from a model A of MA such that A ~ 8( building a model B of E such that B 6 ( 5.2. Proof of Theorem 14. We need a preliminary lemma, whose omitted proof directly follows from Denition 13. Lemma 16. If there is no free occurrence of x i in ', then there is no occurrence of x i in t ' (w; ~x). The proof of the theorem is by induction on the derivation of '. Propositional Axioms. The proof is exactly the same as the one for the propositional 2-as-Pow translation (cf. (D'Agostino et al., 1995)). Universal Quantier Axioms. Let us assume, without loss of gener- ality, that we are instanciating the variable x 1 and that the other free variables in where the substitution fx 1 7! sg is free for x 1 in (x 1 ). Let be and Assume ' (w) (Trans(w) and Axiom (w) are not needed in the proof). We shall prove that w t ' (w; ~x) by showing that, whenever From Def. 13, we have that t that is, taking the negation out of the quantier, with ' (w) (the (w) conjunct), this implies y 62 f (w; x Modus Ponens and Necessitation Rule. Both proofs are basically the same as those for the propositional case. Refer, again, to (D'Agostino et al., 1995). Alternative Translation Techniques for Propositional and First-Order Modal Logics 15 Universal Generalization Rule. Assume the same notation as with the Universal Quantier Axiom schema. The rule has then the form does not occur free in '. Let be the subformula 8x 1 . From Def. 13, we get Assume '!8x 1 (w) (Trans(w) and Axiom (w) are not needed in the proof). We shall prove that w t '!8x 1 (w; ~x) by showing that, necessarily y 2 f (w; x Assume y 2 w and y 2 t ' (w; ~x). By induction hypothesis, w (w n contains no occurrence of x 1 ; hence, we can apply classical Universal Generalization to conclude that thus, from the conjunct (w) of the assumption '!8x 1 (w), that Barcan Formulae. Let us call and the formulae 8x 1 2 respectively, and let xm be the free variables in (or, equivalently, in ). Then Assume ' (w) and y 2 w. Then _ 8z: set (z 2 y _ 8z: set (z 2 y Alberto Policriti, and Matteo Slanina The rst equivalence is just the unfolding of the denition of t ' (w; ~x) by the axioms of 2 . In the second we dropped the subformula y 2 w from the conjunct because it is true by assumption, and then we rewrote the rst disjunct using (w) and the second using the In the last, we used the denition of in the left disjunct and (w) in the right one. Now, to show that w t ' (w; ~x), we take, as above, y 2 w and assume that the second disjunct of the last line is false, which is equivalent to This obviously implies the rst disjunct, thus proving the claim. 5.3. Proof of Theorem 15. The proof is an adaptation of the technique introduced in the soundness proof of the propositional case in (D'Agostino et al., 1995). More pre- cisely, we are going to follow the proof of Theorem 5 of that paper. The rst part is almost the same and we will state some lemmas without proof. Let U be a universe of hypersets satisfying all the axioms of ZF FA (ZF except the Foundation Axiom) and AFA (see (Aczel, 1988) for details). Lemma 17. Let be an ordinal, V be the set of all well-founded sets of rank less than , and U n V be the universe of all hypersets not belonging to V . Then any model for the language of 2 with domain set and interpretation function () 0 satisfying is a model of . Proof. See Lemma 6 of (D'Agostino et al., 1995). Given a frame (W; R; D), we want to embed it into the universe suitable . In (D'Agostino et al., 1995), the authors prove the existence of a labeled decoration such that the following lemma holds. Alternative Translation Techniques for Propositional and First-Order Modal Logics 17 Lemma 18. For each a; b 2 W , 1. a 2. a 62 V +1 and a n V aRbg. Proof. See (ii) and (iii) in Lemma 7 of (D'Agostino et al., 1995). Now consider an interpretation an assignment : f~xg ! D. Moreover, let M be the Q1 model A model for the language of 2 can be obtained as follows. First, take the interpretation for set symbols as dened by the previous lemmas and let D be the domain for the sort term. Then, interpret term constant and function symbols as in FC and F F , respectively, and, for each k-ary relational symbol Q, let f be equal to if this set is not empty, and to V +1 otherwise. Finally, for each universally quantied formula variables x j 1 be equal to if this set is not empty, and to V +1 otherwise. We call this model AM . Furthermore, let W be equal to fa j a 2 Wg. W 62 V +1 because V +1 is transitive and, for each a 2 W , a 62 V +1 (Lemma 18). Lemma 19. We have: 1. for each a 2 W , d if and only if a 2. AM ' (W ). Proof. 1. Let The proof is by induction on '. immediate from the denition of f induction hypothesis) a d) or a 2 0 t d) , (for AM is a model of d) d) , (by denition of t ' ) a 2 0 t d). Alberto Policriti, and Matteo Slanina d) , (for a 2 0 W ) a 2 0 W and a 62 0 t d) , (for AM is a model of d) , (by denition of t ' ) a d). aRb , (by i.h.) for every b 2 W such that aRb, b 2 0 t d) d) , (by denition of (by denition of t ' ) a d). immediate from the denitions of f ' and t ' . 2. Assume, without loss of generality, that be the free variables in '. We are going to prove ' (W ). Let a 2 W and d be arbitrary elements of D. Then a 2 0 f (by denition of f ' , for every d 1 2 D, a; [x 1 =d (for the rst part of this lemma) for every d 1 2 D, a Corollary 20. if and only if Proof. Follows directly from Lemma 19 recalling that M Q1 L ' means that M Q1 L 8~x '. Lemma 21. A modal propositional formula , over the propositional variables is valid in the frame (W; R) if and only if, for the corresponding hyperset W , holds in U n V +1 . Proof. See Lemma 9 of (D'Agostino et al., 1995). To conclude the proof of Theorem 15, let us suppose that 4 By Lemma 16, all the d's beyond the m-th do not count in the interpretation. Alternative Translation Techniques for Propositional and First-Order Modal Logics 19 be a model based on a frame in which is valid (a Q1-L model). From Lemma 21 it follows that Axiom 2 is true in AM , and from the second part of Lemma 19 the same follows for ' (W ). Furthermore, it is easy to prove that Trans(W ) holds as well. Since AM is a model of and, by hypothesis, the formula 8~x: term W 0 t is true in AM , it follows from Corollary 20 that M Q1 L '. 6. Conclusions and Open Questions All the results presented in this paper uniformly rewrite the deduction problem for various families of modal logics into deduction problems with some form of set-theoretic avor. The main contributions are the proposal of an equational theory capable of driving the translation at the propositional level and the extension of the 2-as-Pow translation technique to a signicant class of rst-order modal logics. A systematic study of set-theoretic translation techniques for different kinds of rst-order modal logics could be carried out starting from the ideas presented here. In this paper, we deliberately conned ourselves to one of the most signicant classes. Moreover, we did not consider the extension of the algebraic method to the rst-order case: although we did not see any particular obstacle in principle, its development would almost certainly have called for the introduction of a very ad-hoc (two-sorted, non-equational) extension of the theory of Boolean rings, thus stripping the method of its elegance (which was the main motivation for its use in the propositional case). All of the results mentioned call for specialized techniques explicitly designed for deduction in such theories as or, better yet, able to exploit the syntactic features of the formulae obtained from the translation. Preliminary work in this direction is presented in (Piazza and Policriti, 2000), where a tableaux-based decision procedure for a class of formulae in which the resulting formulae can be embedded is presented. Other related results can be found in (Cantone and Zarba, 2000; Cantone et al., 2001). A further step toward the applicability of the results presented could be obtained via a higher level of integration of the set-theoretic translation with more classical techniques. For example, it would be interesting to study the possibility of separating the rst-order denable portions of a logic (for which more e-cient techniques are often available) from contexts in which the 2-as-Pow is the only available translation. In this context, we mention as an open problem, the question of precisely char- 20 Angelo Montanari, Alberto Policriti, and Matteo Slanina acterizing the expressive power of locally quantied rst-order modal logics (cf. Lemma 9). Finally, another research task, related to the quest for specialized techniques designed to be coupled with the translations presented here, is the search for reduction orders and rewriting systems to be used in automated engines for testing derivability in the theory introduced in Section 3. Acknowledgments We wish to thank prof. Martin Davis for a careful proof-reading of the nal draft of this paper. --R Modal Logic and Classical Logic. Modal Logic Set Theory for Computing. From Decision Procedures to Declarative Programming with Sets Handbook of Philosophical Logic Journal of Arti Lectures in Abstract Algebra Temporal Veri 'EQP 0.9c User's Guide'. Journal of Logic and Computation 1(5) Bulletin of the IGPL 1(1) Technical Report 07/01 --TR
first-order modal logics;computable set theories;translation methods
594334
Checking Simple Properties of Transition Systems Defined by Thue Specifications.
In (possibly infinite) deterministic labeled transition systems defined by Thue congruences, labels are considered as functions of states into states. This paper provides a method for computing domains of such functions for a large class of transition systems. The latter are related to model checking of transition systems defined by Thue congruences.
Introduction Countable transition systems can model any software or digital hardware system and appear as one of the most fundamental structures in computer science. While the research on finite transition systems has a long tradition (see e.g. [1] for further references), infinite transition systems have only recently attracted attention of computer scientists. One of the basic problems concerning infinite transition systems is their effective description. We argue that Thue systems may be used for defining both finite and infinite transition systems and semi- Thue systems may be used as mechanization tools for reasoning about underlying transition systems. When a transition system models a real-life system, the properties of the latter correspond to the theory of the transition system within some suitable logic. The falsity of a sentence on a given transition system amounts to the emptiness of a certain set, namely the meaning of the sentence. Roughly speaking, this set is obtained as a combination of the elementary relations that interpret atomic formulae. Such a combination is computable whenever so are elementary relations and the operators of the logic are simple enough. In particular, this is the case for the Hennessy-Milner logic [9] where the elementary relations are unary and the operations are boolean ones. To establish the truth of a sentence, the central problem is then to compute the elementary relations. This is the subject of the present paper. Our work is related to the well established grounds of formal verification which are based on the connections between logic and formal languages (see e.g. [19]). In that approach, the truth of a sentence on a finite transition system is checked using automata-theoretic con- structions, and is decidable. The transition systems considered here are more general and possibly infinite. As a consequence, the relations interpreting atomic formulae are not always computable. In this situation, the truth is no more decidable. We thus fall into a common pattern of various areas of automated reasoning where deduction rules do not yield an algorithm but rather a not always terminating procedure. We follow a similar approach to the problem which is a first step towards formal verification of what we call Thue specifications, and may be summarized as follows. The semantics of a Thue specification is given as a labeled transition system, the states of which form a language over an alphabet \Sigma and labels belong to an alphabet \Gamma . Each label a in \Gamma corresponds to an event and may be seen as a partial function of states. As in a Hoare triple [']P [/], where the domain (resp. range) of a partial function associated to a program P is restricted by a precondition ' (resp. postcondition /), an event a 2 \Gamma may be restricted to a partial function of X into U where X and U are some subsets of the set of states. This is written a: X!U . Besides letters of \Gamma , the words over \Gamma may be seen as functions. This leads to the word-functions of the form u: X!U with . The problem studied in this paper is that of finding the maximal subset X 0 of X on which u is defined and u(X 0 ) ' U . However, the only solutions of interest are those which may readily be subject to boolean operations, and for which the emptiness problem is decidable. Within the Chomsky hierarchy, only rational languages satisfy both requirements. This leads us to the problem of determining the domain of u: X!U whenever it is rational, assuming that both X and U are also rational. To solve this problem, we introduce a procedure that is given under the form of three deduction rules that derive a system of left-linear (language) equations from a word-function u: X!U . If the procedure terminates, the least solution of the resulting system of equations is the domain of u: X!U . In deduction rule-based automated reasoning, the most fundamental issues are soundness and completeness [20]. While the soundness is an obvious requirement, the completeness may be out of range in certain situations. Unlike the set of consequences of a finite set of axioms, the theory of a single structure need not be recursively enumerable. For instance, since the theory of the standard model of arithmetics is undecidable (see e.g. [17]), by the Turing-Post Theorem, one concludes that this theory is not recursively enumerable. A complete method for deciding the truth of a sentence cannot exist in such a case. This limitation is well known in the area of inductive theorem proving (see e.g. survey [4]). Similarly, the problem studied in this paper is not recursively enumerable, hence the procedure introduced here is not complete. In particular, this procedure is not guaranteed to terminate on each instance of the problem which has a rational solution. Nevertheless, we establish a reasonably weak sufficient condition for the termination and we show that this condition is satisfied by several well-known classes of semi-Thue systems for which the termination is expected. This paper is organized as follows. In Section 1 basic notions from language theory and string rewriting are recalled. Section 2 introduces Thue specifications and their models. The main problem of the paper is stated in Section 3 and is related to the Hennessy-Milner logic. In Section 4 a procedure that treats the problem is provided. Section 5 is devoted to the correctness of the procedure. In Section 6 it is established that the problem is not recursively enumerable and a reasonably weak condition for the termination of the procedure is provided. The techniques developed in the paper are applied in Section 7 to an example of a simple railway network. Brief conclusions close the paper. Assuming a minimum theoretical background from rational languages [16] and string-rewrit- ing [6], we recall in this section several notions from these topics and we set some notations for the sequel. The family of rational languages over an alphabet \Sigma is written Rat(\Sigma ) and the family of the rational subsets of a language L ' \Sigma is written Rat(L), viz stands for the powerset of L. Systems of equations We speak of rational expressions in the usual meaning but we also use rational expressions with variables. The variables are written are assumed to form an infinite, countable set. Subsequently we shall often replace variables in a given rational expression R (resp. set of equations E ) with some rational expressions (possibly with variables) by applying a substitution oe on it. The resulting expression (resp. set of equations) is written Roe (resp. E oe). A substitution of a single variable X by an expression P may be its application on an expression R (resp. set of equations E) is written A (left) linear system of equations is a set of equations such that, for each variable, there is at most one equation with this variable as lefthand side and each equation is of the form where Y is a rational expression, each Y i is a rational expression denoting a nonempty set and are variables. In such an equation, Y is called constant term. We say that the system of equations E is balanced if, for each variable X of the system, there is exactly one equation in E with X as lefthand side. Among all solutions of E with respect to (w.r.t.) a given variable X we distinguish the least solution (w.r.t. inclusion). It is well known that a balanced linear system of n equations has a least solution w.r.t. any of its variables. One may compute such a solution using the Gauss elimination method. For more details related to this topic the reader may consult e.g. [3]. Suffixes, prefixes, quotients and remainders Given u; v; w 2 \Sigma such that is called a prefix and v a suffix of w. When v 6= " (resp. u 6= "), u (resp. v) is a proper prefix suffix ) of w. We note pref (u) (resp. suff (u)) the set of all the prefixes (resp. suffixes) of a word u. These notations are extended to sets; for instance, given X ' \Sigma , pref (X) stands for u2X pref (u). When words x and z, then y is called a factor of w. When xz 6= ", y is an proper factor of w. The (right) quotient (resp. remainder) of M ' \Sigma by N ' \Sigma , written M=N (resp. MiN) is defined as follows resp. As we will see further, some steps of the procedure that is the main point of this paper depend on computations of quotients and remainders between rational languages. Remember that the family of rational languages is closed under both quotients and remainders [10]. String rewriting A semi-Thue system (sts) S over an alphabet \Sigma is a binary relation on di of S, also noted g ! d, is called a rule; l(r) (resp. r(r)) stands for the lefthand (resp. righthand) side of r, i.e. g (resp. d). The latter notation is extended on sets of pairs (e.g. binary relations, functions etc.), especially l(S) denotes the domain and r(S) the range of S. An sts S is 1. special if 2. monadic if r(S) ' A [ f"g and jlj ? jrj for each (l; r) 2 S, 3. left-basic if for all x; 2 r(S) and for all x; reduction relation "! " on \Sigma is associated with S in the usual way, namely ! and the rewrite relation is the reflexive-transitive closure of ! . A word is said to be reducible (resp. irreducible) w.r.t. S when it belongs (resp. does not belong) to l(! ). The set of all irreducible words over \Sigma is written \Sigma # S . It is well known that this set is rational whenever l(S) is so, since \Sigma # . A normal form of a word u w.r.t. S is an irreducible word v 2 \Sigma # S such that u v. 1 We say that an sts S is confluent when for all words u; v; v 0 2 \Sigma such that u v and u v 0 there exists a word w 2 \Sigma such that v w and v 0 w. If in addition S is terminating, i.e. there is no infinite chain is said to be convergent. It is well known that the convergence of an sts implies the existence and uniqueness of a normal form of each word. Throughout this paper we only use convergent semi-Thue systems and the unique normal form of a word u is written u# S . In notation ! usually omit the subscript S when there is no ambiguity about the sts we refer to. Notice that in our notations "#" has a lower precedence than the concatenation, thus uv# stands for the normal form of uv. Notations for rewriting of rational expressions are similar to those for individual words. For sake of simplicity, we identify every rational expression with the language it denotes. We may write etc. for any rational expressions X, Y that, in this context, stand for associated languages. Consequently, we speak of the normal form of a rational expression X and we write X# in the reference to the set fu# j u 2 Xg or possibly to some rational expression denoting this set. It may be useful to restrict applications of some rules of an sts so that they can only rewrite prefixes of words. For this aim we may assume that each alphabet \Sigma may be extended by an additional symbol the words to rewrite belong to $ \Sigma \Sigma and S is a subset of ($ \Sigma \Sigma \Theta $ \Sigma \Sigma ) [ (\Sigma \Theta \Sigma ). Note that if all rules are in $ \Sigma \Sigma \Theta $ \Sigma \Sigma then we are in the particular case of regular canonical systems [5] (see also [8,15]). The latter are known as generating rational languages. Convergent Thue Specifications, Their Languages and Models A convergent Thue specification cts is a triple h\Sigma is a state alphabet, \Gamma is an event alphabet, R is a rational subset of $ \Sigma \Sigma (of \Sigma if S " ($ \Sigma \Sigma \Theta $ \Sigma \Sigma is a finite convergent sts over \Sigma [ \Gamma and each word of R is irreducible, i.e. R 2 Rat($ \Sigma \Sigma # S ). The distinction between state alphabet and event alphabet has no importance from a theoretical point of view but is useful in practical situations. When simply write h\Sigma ; R; Si. Recall that the convergence is not decidable in general. On the contrary, for a rational language R, it is decidable whether R ' The aim of Thue specifications is to provide means for specifying and verifying processes. This article does not discuss that approach as a specification technique but focuses merely on We use a restricted version of the notion of normal form that is common in term-rewriting. In the general version a normal form of a word u w.r.t. S is an irreducible word v 2 \Sigma #S such that u v. a basic problem related to formal verification. In this field, semi-Thue systems are needed. Moreover since we are intersted in deterministic processes, we only consider confluent semi- Thue systems. The semantics of Thue specifications is given in terms of (a variant of) labeled transition systems. Formally, a labeled transition system (lts) A is a tuple hQA is a set of states, \Gamma is an alphabet, ' A 2 QA is an initial state and ffi A ' QA \Theta \Gamma \Theta QA a transition relation. When (q; a; q 0 we say that q (resp. q 0 ) is the origin (resp. target) of a transition labeled by a. A transition system is deterministic if for all q; q each a implies that q Each state q 2 QA recognizes some set of words of that is defined as the smallest set satisfying the following conditions: if The language of A written L is the set of all words recognized by the states of A , namely q2QA Definition 2.1. An lts is an initial model of cts and the following holds for all u; v 2 R and each a if and only if ua# The language L (cts) of a cts is the language of its initial model: L cag c c c c c \Deltaff A A A AU a c c c c c c c c \Deltaff A A A AU a c c \Gamma\Psi @ @ @ @R a b c \Gamma\Psi @ @ @ @R a b c c c c c c c \Deltaff A A A AU \Deltaff A A A AU ae ae ae ae ae ae ae ae= Z Z Z Z Z Z Z Z~ c a a c c a b Fig. 2.1. A convergent Thue specification and its initial model We emphasize that according to the above definition, the states QA of an initial model are identified with the words of R. It may be observed that the confluence of S is sufficient for an initial model to be deterministic. Consequently, convergent Thue specifications only model deterministic processes. Henceforth, for an initial model A of a cts, ffi A is considered as a function and we write ffi A (u; a) = v instead of (u; a; v) 2 ffi A . Another important observation is that the language of a cts is always closed under nonempty prefixes because the language of a transition system is so. We speak of an initial model because one may define a satisfaction relation labeled transition systems and Thue specifications. One may then consider the initial object in the category of the models of a cts. This initial object may be alternatively characterized as the initial model in the sense of the above definition. Obviously, the initial model is unique up to isomorphism and always exists. An example of a convergent Thue specification together with its initial model is provided in Fig. 2.1. In this example, the identification of the set of states with R 1 is left as an exercise to the reader. For sake of simplicity, from now on and up to Sect. 7, we neglect the distinction between state alphabet and event alphabet. Both alphabets are assumed equal and a cts is written 3 Domains of Word-functions and the Hennessy-Milner Logic In the previous section we explained the relationship between convergent Thue specifications and deterministic labeled transition systems. To any convergent Thue specification h\Sigma ; R; Si, one may associate its initial model, say A . Concerning A , a very basic problem consists in characterizing all transitions with some given label a 2 \Sigma . More precisely, one may wish to characterize all states that are origins of the transitions labeled by a. For this aim, it is convenient to view a as a function of states 2 into states by writing a: R!R. The problem consists therefore in characterizing the domain of this function. Instead of taking all the states into account, one may consider a restriction of the above function written a: X!U , for some subsets X and U of R. In the sequel, we restrict our attention to the case where both X and U are rational subsets of R. The main problem of this paper may be stated as follows. Problem. Given a convergent Thue specification cts = h\Sigma ; R; Si and a function a: X!U relative to cts and such that both X and U are rational subsets of R (X; U 2 Rat(R)), find Dom(a: X!U), viz the domain of the function, whenever it is rational. To better understand this problem, one has to reason about the initial model of the cts, say A . According to Definition 2.1, given a state w 2 X we have Hence, w 2 Dom(a: X!U) when wa# 2 U . Thus, Dom(a: X!U) is the largest subset X 0 of X such that X 0 a# ' U . In order to solve this problem, it is useful to consider a bit more general question that appears as a technical trick without specific semantics related to transition systems. With this aim in view, given u 2 \Sigma and rational subsets X;U of R, the word-function u: X!U is defined as follows: u(w) = wu# S , for every w 2 X such that wu# S 2 U . Thus Remember that R is the set of states of A . It is important to note that, without additional assumptions, a word-function needs not coincide with the corresponding composition of functions associated to labels. When unless 1g. Since the set of the states of a cts may be strictly included in \Sigma # S , the assumption about \Sigma # S being the source or a the target of a function is not compatible with the semantics in terms of transition systems. For this reason, the word-function are considered here as a technical tool without a special meaning. To close this section we point out that the problem we are interested in is related to the model checking, at least in the context of the Hennessy-Milner logic. Checking Properties Expressed in the Hennessy-Milner Logic The formulae of the Hennessy-Milner logic [9] are built up from the constants 1 and 0, using the classical connectives (:; ; "; etc.) and a unary operator hai for each letter a of the alphabet \Sigma . Given a labeled transition system A , the meaning of a formula ', viz the set of the states of A that satisfy ', written ['] A , is inductively defined as follows: where the map [hai] A : (QA defined by When a transition system A is an initial model of cts Si, we can express the meaning of a formula in terms of the domains of the word-functions. For this aim, the equalities (3.1) remain valid (we have precisely [:'] becomes [hai] A It is obvious from the above that the procedure developed in the sequel may be used for checking the properties, expressed in the Hennessy-Milner logic, of systems defined by convergent Thue specifications. This is relevant only for those cts's for which the procedure succeeds in computing Dom(a: R!X) for all a 2 \Sigma and all X 2 Rat(R) (see Subsection 6.2). Since in the case of success, Dom(a: R!X) is rational, the boolean operations stipulated in may readily be performed. 4 Introducing the procedure A method for solving the problem stated in the latter section is now developed. An instance of the problem consists of a cts h\Sigma ; R; Si and a word-function u: X!U relative to it, such that both X and U are rational subset of R. More precisely, X, U and R are rational expressions that, for sake of simplicity, are identified with corresponding languages. In addition, u is required to be irreducible. This is not restrictive since u: X!U and u#: X!U are equal due to the convergence of S. Hence, if u is reducible, u has to be normalized before applying the method. The solution of an instance of the problem is the subset X 0 of X such that X however that the method developed in the sequel can find X 0 only rational. The method consists of a procedure that, provided it terminates, computes a system of equations, the least solution of which is also the solution of the instance of the problem. Basic ideas underlying the procedure are developed in subsequent paragraphs. The task of finding Dom(u: X!U) becomes easier if X is partitioned into two parts Y and Z such that Y u is irreducible and each word of Zu is reducible. Further, Z will be partitioned into smaller parts, say Z . Indeed, such a "divide and conquer" strategy is based on the following obvious equality In order to characterize Z, assume that a rule g ! d of S is applicable to a word vu of Zu. Since both v and u are irreducible, g overlaps v and u, i.e. are such that x 6= " and w 6= ". This remark leads to the following definition. Definition 4.1. A word x 2 \Sigma is a left-overlap completion of u 2 \Sigma with respect to a semi-Thue system S if there exist w; w 0 2 \Sigma such that xw 2 l(S), ww left-overlap completion x of u is minimal if no left-overlap completion of u is a proper suffix of x. In view of the above definition, each word of Z has a suffix that belongs to the set, say L, of all left-overlap completions of u. The set Y , being the complement of Z in X, is characterized by X i L. In fact, one may characterize Y as X i L min where L min stands for the set of all minimal left-overlap completions of u. In order to partition Z into smaller parts, observe that, given a left-overlap completion x of u, the expression (X=x)xu denotes all the words of Xu that may be reduced by any rule d such that w is a prefix of u. Consequently, one may wish to partition Z according to L as x2L (X=x)x. However, this is not a partition because (X=x 1 )x 1 ' (X=x 2 )x 2 whenever x 2 is a suffix of x 1 . In fact, it is L min that provides a partition of Z, namely (X=x)x. In sum, the following partition of X is used as basis for the divide and conquer strategy: Now, according to (4.1), the domain corresponding to each part has to be found. With respect to the first member of the above partition, it is not too difficult to see that Indeed, since both U and (XiL min )u are irreducible. With respect to the other members of the partition described in (4.2), Dom(u: (X=x)x!U) has to be found for each x 2 L min . Observe first that Consider, now, various ways of computing (X=x)xu#. Obviously, one may compute xu# first and subsequently compute the normal form of (X=x)(xu#). Thus The maximal suffix of xu# that is not rewritten during any computation of the normal form of (X=x)(xu#) now has to be isolated. The concept introduced in the following definition is useful for this purpose. Definition 4.2. A word v of \Sigma # is a right irreducible divisor with respect to an sts S when f"g. The set of all right irreducible divisors w.r.t. S is written rid S . For any u 2 \Sigma such that rid S we say that w is a right irreducible divisor of w.r.t. S. The set of all right irreducible divisors of u is written rid S (u). A right irreducible divisor of u is maximal if it is not a proper suffix of a right irreducible divisor of u. It is easy to see that rid . Consequently, since S is finite and \Sigma # is rational, rid S is rational too. It is also obvious that any right-irreducible divisor of xu# is not rewritten during any computation of (X=x)(xu#: Proposition. Any u 2 rid S and w 2 \Sigma satisfy Let us go back to the problem of finding Dom(xu#: X=x!U) assuming that z is the maximal right irreducible divisor of xu#. Suppose also that for some V ' \Sigma . Then V in view of the above proposition, and V y#z ' U . Therefore V y# ' U=z. Consequently In sum, taking (4.4) and (4.5) into account, we have The next definition is not indispensable since it just combines Definitions 4.1 and 4.2. Nevertheless, it will allow a more concise presentation of the procedure. Definition 4.3. A left-overlap triple of a word u w.r.t. S is a triple of words hx; that x is a minimal left-overlap completion of u, z is the maximal right irreducible divisor of xu#. The set of all left-overlap triples of u is written lot S (u), or simply lot (u) if S is determined by context. A triple 2 lot (u) is written hx ; y ; z i. The procedure is given in the form of three deduction rules (see Table 4.1) Generate, and Eliminate, the first of which corresponds, roughly speaking, to the transformations discussed up to now in this section. Using (4.1), (4.2), (4.6) and Definition 4.3, those transformations may be summarized in the next lemma. Henceforth we speak of an irreducible expression to mean that the expression denotes a language, each word of which is irreducible. Lemma 4.4. Let X be an irreducible rational expression and u an irreducible word. Then U=z 6=? The proof of the lemma follows from the above discussion together with the following remarks. When X=x ?. For this reason, such cases are excluded from (4.7). Also the first member of the above sum is written in a way that slightly differs from the righthand side of (4.3), since Xi and x is the set of all minimal left-overlap completions of u (formerly L min ). Table 4.1. Deduction rules U=z 6=? U=z 6=? Xx is a set of fresh variables. We shall now describe the procedure by explaining its principles and related notations. As mentioned at the beginning of this section, in case of success the result of the procedure consists of a system of equations. Consequently, we still need some variables. In a derivation of such a system, an expression of the form X 7! u: X!U with a variable X represents a goal. This means that Dom(u: X!U) should be the least solution w.r.t. X of the resulting system of equations. Each rule of a procedure is applied downwards with reference to some goal. The goal is then said to have been treated ; otherwise it is called fresh. We denote by F (resp. T) the set of fresh (resp. treated) goals. Thus, the procedure manipulates sequents of the form T; F ' E where E is a system of equations. Let us now explain the r"ole of each rule of the procedure. First note that any rule is applied with reference to some fresh goal. Given a fresh goal X 7! u: X!U the rule Generate introduces a new equation based on (4.7) U=z 6=? and for each 2 lot(u) such that X=x 6= ? and U=z 6= ?, it introduces a fresh goal Each X is some new variable. The goal X 7! u: X!U is then no longer fresh and is transferred to T. Notice that if, in the new equation, we replace X by Dom(u:X!U) and each X by Dom(y : X=x !U=z ) we thereby obtain (4.7). The fresh introduced in this step correspond to the principle that the procedure should be applied recurrently in order to compute Dom(y : X=x !U=z ). The following idea underlies the rule Identify. If there is a treated goal X 7! u: X!U (the one for which an equation has been generated) and a fresh goal X 0 7! then we should not generate an equation from X 0 7! Instead, we identify the two variables X and X 0 . This is realized by renaming X 0 as X in the current system of equations. The goal X 0 7! simply removed from F. The rule Eliminate deals with goals X 7! u: X!U such that Xu is irreducible. A fresh goal satisfying this condition is moved to T and the equation is added to the systems of equations. The applicability of the above rules is restricted by certain conditions (written beside each rule). We insist on the fact that within such applicability conditions Xu be understood as a language and r(T) as a set of word-functions (the range of the assignment T). The notation F q fX 7! u: X!Ug is used for a partition of some F 0 such that X!Ug. Derivations and related concepts are defined as follows. A derivation is a (possibly finite) sequence of sequents s 0 for each there exists an instance of a rule, the upper and lower sequent of which are equal resp. to s i and s i+1 , and s i satisfies the applicability conditions of the rule. A derivation is called maximal when it is not a proper prefix of another derivation. We say that a sequent is pure when it is of the form ?; fX 0 7! stands for an irreducible rational expression and u 0 an irreducible word such that rid f"g. We say that a derivation is pure when it starts with a pure sequent. In order to compute Dom(u: X!U) the procedure starts with the pure sequent f"g. This is not restrictive since, if with z the maximal right irreducible divisor of u such that z 6= ", then X!U=z). The procedure may therefore start with the pure sequent ?; fX 7! y: X!U=zig ' ?. Next, the rules are applied as long as possible in order to produce a maximal derivation. The procedure stops if no rule may be applied. This happens only if there are no more fresh goals. It then remains to compute the solution of the resulting system of equations w.r.t. X . Let us illustrate this by applying the procedure to an example. Example 4.5. Let cts 1 be the convergent Thue specification of Fig. 2.1. We want to compute We check that both $(a are irreducible, and rid(c) = f"g. Taking into account that lot (c) = fha; "; bi; hb; c; aig, we have the following derivation. Generate ae oe Eliminate ae oe ae oe ae oe ae oe Finally, the least solution of the resulting system w.r.t. X 0 yields as expected, according to the minimal model of cts 1 depicted on Fig. 2.1. 5 Correctness of the procedure Usually, a sequent-like calculus is correct, when it transforms valid sequents into valid ones. We shall use the following notation in order to explain what we mean by "valid sequent". Notation 5.1. For any set G of goals, we denote by oe G the following substitution: The following definition is used as basis for the correctness. Definition 5.2. A sequent T; F ' E is valid, when Dom(v: Y !V ) is the least solution of goal Y 7! v: Y !V 2 T. We need to establish the following theorem. Theorem 5.3. Any sequent of a pure derivation is valid. Proof. Let be a derivation satisfying the assumptions of the theorem and let us prove the statement for any sequent s k by induction on k 2 N. Basis: The statement holds trivially for any pure sequent ?; fX 0 7! Induction step. We have to distinguish 3 cases depending on the rule that yields s k from s k\Gamma1 . By induction hypothesis s k\Gamma1 is valid and the validity of s k is obtained using the correctness of each rule, established in Appendix B. 2 The main result on the correctness is given as corollary of Theorem 5.3. Corollary (Correctness). Let D be a maximal finite derivation starting with a pure sequent be the last sequent of D. Then Dom(u: X!U) is the least solution of Proof. It is obvious that on the first sequent either Generate or Eliminate may by applied. Consequently, X 7! u: X!U 2 T. On the other hand, F is empty, since no rule can be applied on the last sequent of D. Hence oe F is empty and by Theorem 5.3 we conclude that Dom(u:X!U) is the least solution of 6 Incompleteness and termination Since the procedure computes only linear systems of equations, the rationality of the domain of a word-function is obviously a necessary condition for the termination. To see that this condition is not sufficient, consider the following specification: cts apply the procedure to the goal . The procedure generates the following infinite sequence of goals: This example points out that the procedure is not complete. In fact, as established in the next subsection, a complete procedure for this problem cannot exist. (By the completeness of the procedure we mean that it terminates on each instance of the problem that has a rational solution.) In Subsection 6.2, we therefore investigate a sufficient condition for the termination of our procedure. 6.1 Incompleteness of the problem The question as to whether a complete procedure exists for this problem is equivalent to the recursive enumerability of the following set of tuples: More precisely, we consider an even simpler version of the problem where the cts's are of the form h\Sigma ; \Sigma #; Si and the word functions are of the form a: X!U . We establish that the following set of pairs is not recursively enumerable: is a convergent sts over \Sigma , a 2 \Sigma ; To this end we reduce the complement of the halting problem for Turing machines. More precisely, given a deterministic Turing machine T and an input word w, we construct a convergent sts and a word-function the domain of which is rational if and only if T does not halt on w. be a deterministic single-tape Turing machine, where is the tape alphabet, denotes the blank symbol, Q is the set of states, is the transition function and q 0 is the initial (resp.final) state. Observe that without loss of generality it is implicitly assumed here that T cannot print . Consequently, each tape cell that has been visited by the tape head contains a letter of \Pi and the language accepted by T is stands for the reflexive-transitive closure of the single-step computation relation ' T Without loss of generality we assume in addition that is the set of "overlined" versions of letters of \Pi and 0; 0; are additional symbols. Let be an input word. Given the pair hT ; wi we construct the pair hS T ; %:Xw!U# i where and the sts ST over \Sigma is the union of several groups of rules below. When two rules differ only by 0 or 1, we abbreviate them by meta-variables ; 0 that rank over f0; 1g. The use of "overlined" ("underlined") symbols is not relevant for the understanding of the construction but aims at avoiding overlaps between lefthand sides of ST . At first reading the reader may assume that the "underlined" and "overlined" versions of any symbol and the symbol itself are identical. ffl Initialization rules ffl Rules that simulate the behaviour of T ffl Rules that apply immediately after each simulation step that does not result in a configuration where the tape head reads ffl Rules that apply before each simulation step (after (ii) or (vii)) ffl Terminal rules Before justifying this construction, we state several lemmas. The first one is obvious. Lemma 6.1. Both Xw and U# are irreducible w.r.t. ST . 2 Lemma 6.2. ST is confluent. Proof. To establish the confluence, it is enough to check that the lefthand sides of ST do not Lemma 6.3. ST is terminating. Proof. Consider the following precedence on \Sigma : 0 1 a # for all a 2 \Pi. Let 1 be the following (rewrite) ordering on (\Sigma r f%g) : where lex is the lexicographic extension of and e x (resp. e y) stands for the reversal of x (resp. y). Using 1 we define a (reduction) ordering 2 on (\Sigma r f%g) : It is easy to check that, except for (i), each rule of l ! r 2 ST satisfies l 2 r. Let then 3 be a (reduction) ordering on \Sigma defined, for all x; y 2 \Sigma , as follows: x 3 y if stands for the number of occurrences of % in x - or where lex stands for the lexicographic extension of 2 . Now, for Rule (i) we have % 3 % because (0; ") lex 1). For each other rule l ! r 2 ST we have also l 3 r because 2 and 3 coincide on (\Sigma r f%g) . 2 Justification of the construction Given a word q 0 wx 2 Xw , where some a below the rewrite steps of the word q 0 wx%. (The sequences of rules used in these steps are denoted by rational expressions.) Starting from q 0 wx% we get and by we end up with u n q n v n #, provided that Thus if T does not halt on w, Dom(%:Xw!U# Assume then that T halts after k steps in a configuration u k q f v k and let v for some b the length 2n of x may be arbitrarily big, without loss of generality we assume that n ? k l. Starting with q 0 wx% analogous rewrite steps lead to We then get (viii) (ix)(vii) Now, using the rules (x) and (xi) we end up with u k q f # if and only if the word belongs to the Dyck language D . Consequently Dom(%:Xw!U# which is not a rational language. We have therefore established the following lemma. Lemma 6.4. Dom(%:Xw!U# ) is rational if and only if T does not halt on w. 2 Putting together Lemmas 6.1, 6.2, 6.3 and 6.4 and taking into account the fact that the complement of the halting problem for Turing machines is not recursively enumerable, we conclude that the following holds. Theorem 6.5. DomRat is not recursively enumerable. 2 In view of this theorem a complete procedure for the problem of computing rational domains of (even single-letter) word-functions cannot exist. Otherwise, there would exist a Turing machine that "semi-decides" the complement of the halting problem. 6.2 Termination In this subsection we state a sufficient condition for the termination of the procedure. The following definition provides a tool for studying the termination. Definition 6.6. The generation set for a goal X 7! Dom(u:X!U) w.r.t. an sts S, written is the smallest subset of \Sigma \Theta (\Sigma )\Theta (\Sigma ) that contains hu; X;Ui and such that whenever hv; Y; W for each 2 lot (v). In the sequel, we shall consider the first, the second and the third projections of a generation set G S (u; X;U We now state two lemmas that lead to the final result of this section. The first one is obvious. Lemma 6.7. A maximal derivation starting with a pure sequent ?; fX 7! Dom(u: X!U)g ' ? is finite if and only if its generation set is finite. 2 Lemma 6.8. The second (resp. third) projection of a generation set G S (u; X;U) is finite if Proof. Note first that for a given rational set Z the cardinality of the set is not greater than 2 k , where k is the number of states of a minimal automaton accepting Z. In order to prove the statement, it is therefore enough to show that 2 (G S (u; X;U)) ' QX According to Definition 6.6, there exist hy and lot (y n ). Thus Hence The following theorem states a simple sufficient condition that guarantees that the procedure will terminate. Theorem 6.9. Starting with a pure sequent ?; fX 7! Dom(u: X!U)g ' ?, the procedure terminates if 1. S is finite, 2. for each minimal left-overlap completion x of u, the word xu belongs to l(S) and 3. for each rule r 2 S, any 2 lot (r(r)) and any minimal left-overlap completion x 0 of y , the word x 0 y belongs to l(S). Proof. According to Lemmas 6.7 and 6.8, the termination of the procedure is guaranteed if 1 (G S (u; X;U)) is finite. An easy induction allows us to establish that for any y in there exists a rule r 2 S such that xy and x is a minimal left-overlap completion of y. Hence 1 (G S (u; X;U)) is included in a finite set, namely suff (l(S)). 2 Note that, when used for computing the domain of a word-function that is a single letter (as e.g. in the case of the Hennessy-Milner logic), the procedure terminates provided that both the first and the third requirements of Theorem 6.9 hold. We may use this observation for more general word functions. In order to compute Dom(u: X!U) w.r.t. h\Sigma ; R; Si such that when the second requirement of Theorem 6.9 is not satisfied whereas the first and the third ones are, we may proceed as follows. 1. compute the ranges of the word functions: (For the computation of ranges we may use the procedure of [11]. This procedure always terminates for a single-letter word-function under conditions 1 and 3 of Theorem 6.9.) 2. Set D apply the procedure for computing the domains of the word functions: Now It is important to point out three classes of semi-Thue systems that may successfully be used for letterwise computation of domains of word-functions, namely left-basic sts's, monadic sts's and special sts's. Corollary 6.10. Starting with a pure sequent ?; fX 7! Dom(a: X!U)g ' ?, where a 2 \Sigma , the procedure terminates if S is finite and is special, monadic or left-basic. Proof. When S is monadic, for any r 2 S, the set lot (r(r)) may contain only elements of the form hx; a; "i or hx; "; ai such that xr(r) ! a is a rule of S. When S is left-basic or special, obviously for each rule r of S. Hence the sts's that are special, monadic or left-basic satisfy the third requirement of Theorem 6.9. 2 In view of the corollary, we may assert that the requirements of Theorem 6.9 provide a reasonably weak sufficient condition for termination. These requirements capture the well known classes of sts's, the models of which are related to the transition graphs of pushdown machines [7,12]. For the latter, the rationality of the domains of word-functions follows from the well known fact that the pushdown store languages are rational. In addition, the scope of Theorem 6.9 is not restricted to special, monadic or left-basic sts's. Indeed, S 1 (see Fig. 2.1) is neither special, monadic nor left-basic whereas it satisfies the requirements 1 and 3 of the theorem. Last but not least, the example of non termination that opened this section has one interesting property, namely that cts 0 equivalent to cts 2 , in the sense that cts 2 and cts 0 2 have the same initial model. But this time the procedure terminates for cts 0 2 on any single-letter word-function. 7 Final Example We consider a simple portion of a railway network. A single track connects two stations: left (L) and right (R), via a central one (C), as on the following picture. L C R All trains run from L to R or vice-versa with a stop at C, never being allowed to go in reverse. The central station has two platforms allowing train crossing. Each of these platforms may be occupied by at most one train. To model this situation we may consider the following state alphabet \Sigma a train has left L but has not yet arrived at C a train has left R but has not yet arrived at C \Gamma! D C a train has left C for R but has not yet arrived at R a train has left C for L but has not yet arrived at L \Gamma! A C a train coming from L is waiting on a platform of C a train coming from R is waiting on a platform of C The underlined versions of \Gamma! D C have no special meaning for the modeling but aim at making easier the definition of transitions by a semi-Thue system. For this reason, we distinguish the set \Sigma 0 of non-underlined letters of \Sigma : \Sigma g. We also consider the following event alphabet \Gamma . d L departure of a train from L departure of a train from R \Gamma! d C departure of a train from C for R departure of a train from C for L \Gamma! a C arrival of a train from L at C a C arrival of a train from R at C a L arrival of a train at L a R arrival of a train at R In our modeling a train "appears" when it leaves one of the stations L or R and "disap- pears" when it arrives at the opposite terminal station. Since the trains do not go in reverse, there is a deadlock, when two trains run towards one another on the same portion of the track (between L and C or between C and R). In order to define the set R of states, we need the following expressions, each of which describes, in fact, a subset of R satisfying some characteristic property (stated between paren- (D L trains run from L to R), trains run from R to L), (D L trains run towards C), trains run from C to L or R). Using the above expressions, the states of the system may be defined as follows: A word of R reflects the events in the reverse chronological order (a state alphabet letter corresponding to the most recent event is written immediately after $). For instance means that there is one train, say t running from R to C, a second one, say t running from L to C and a third one, say t 3 , on a platform of C coming from R. The events corresponding to this configuration occurred in the following has left R before t 2 left L and t 2 has left L before t 3 arrived at C. The transitions are defined by the following sts S over f$g [ xd L \Gamma! d L x for x 2 \Sigma 0 xd R \Gamma! d R x for x 2 \Sigma 0 $d L \Gamma! $D L \Gamma! $D R x \Gamma! a C \Gamma! \Gamma! a C x /\Gamma a C \Gamma! /\Gamma a C x for x 2 \Sigma 0 r f \Gamma! D C \Gamma! a C \Gamma! \Gamma! A C a C \Gamma! x \Gamma! A C \Gamma! \Gamma! A C \Gamma! x for x 2 \Sigma 0 \Gamma! \Gamma! x \Gamma! d C \Gamma! \Gamma! d C x for x 2 \Sigma 0 r f \Gamma! A C x /\Gamma d C \Gamma! \Gamma! A C \Gamma! d C \Gamma! \Gamma! D C \Gamma! x \Gamma! D C \Gamma! \Gamma! D C \Gamma! x for x 2 \Sigma 0 \Gamma! \Gamma! \Gamma! a R x for x 2 \Sigma 0 r f \Gamma! D C xa L \Gamma! a L \Gamma! D C a R a L When a meta-variable x is replaced by all possible values, we get a system of 62 rules with non-overlapping lefthand sides. This system is therefore finite and confluent. The termination of S may readily be established using standard reduction orderings, thus concluding that S is convergent. As an example of questions concerning this system, we may ask about the set of states in which a departure of a train may occur at one of the terminal stations, for instance at R. The set of such states, say P , is precisely given by P d R computing P d R it is useful to ask about the termination of the procedure in this case. Since the first condition of Theorem 6.9 is obviously satisfied by S, we check whether the second one is satified by d R Indeed, for each letter a and any left-overlap completion x of a, we have xa 2 l(S). Hence, it remains to verify the third condition of Theorem 6.9. To this end, we compute the set T of left-overlap triples of r(S): For the rules with a meta-variable, we obtain, for all x 2 \Sigma 0 lot(d L xig lot (d R xig lot ( \Gamma! a C lot ( /\Gamma a C lot( \Gamma! A C xig lot( xig lot ( \Gamma! d C lot ( /\Gamma d C lot ( \Gamma! D C xig lot ( /\Gamma D C xig lot(a R lot (a L For the rules with $, we have lot($D L \Gamma! A C and for other rules lot( \Gamma! A C ig lot( ig lot ( \Gamma! D C ig lot ( /\Gamma D C ig We observe that 2 (T ) f"g. Thus, for each u 2 2 (T ) and each left-overlap completion x of u we have xu 2 l(S), meaning that the third condition of Theorem 6.9 is satisfied. We may therefore conclude that the procedure terminates here for any single-letter word-function. Starting with a pure sequent ?; fX 7! Dom(d R the procedure terminates and yields a system of equations, the least solution of which is, as expected, ing Thus, the departures of trains from R are possible only when there are no trains running from C to R. Concluding remarks Semi-Thue systems and transition systems have been related here by sketching a formal specification technique named as Thue specifications. A particular emphasis has been placed on the simplest case of the problem of formal verification. A pragmatic point of view has been necessary in approaching the latter problem, since this has been shown not to be recursively enumerable, and has led to an incomplete procedure. This problem is perhaps the most elementary problem that one has to deal with in order to be able to approach the model checking of Thue specifications, at least in the special case of the Hennessy-Milner logic. However, since the expressive power of the latter is not very rich, our work needs be extended so as to handle more powerful logics. With this aim in view, our attention is currently focussed on the monadic second-order logic [13]. One of the most important problems concerns the specification of systems that consist of interacting processes. This problem may be handled via the concept of synchronized product [2]. Within that approach, each component of the system is specified separately and interactions between components are defined as synchronization constraints. For instance, the railway example may be specified in an even more elegant way as four interacting processes: the two portions of the track and the two platforms of the central station. When a system is specified componentwise by several Thue specifications provided with a synchronization constraint, these specifications may be combined into a single one, (on wich the procedure may be applied for checking the properties of the system). Based on the ideas of [14], an algorithm that performs the latter task, has been developed [18]. Acknowledgments We wish to thank Marie-Catherine Daniel-Vatonne and ' Etienne Payet for careful reading of a preliminary version of this paper, and Adrian Bondy for correcting the English text. We are grateful to Damian Niwi'nski for numerous suggestions that helped improving the presentation of this paper. Special thanks are due to Serge Burckel whose comments have influenced Subsection 6.1 and to an anonymous referee that has suggested the current proof of Lemma 6.3 which is more elegant than our original proof. --R Finite Transition Systems. Comportements de processus. Theorem proving in hierarchical clausal specifications. Regular canonical systems. A string-rewriting characterization of context-free graphs On the regular structure of prefix rewriting. Algebraic laws for nondeterminism and concurrency. Introduction to Automata Theory A procedure for computing normal forms of certain rational sets. The graphs of finite monadic semi-Thue systems have a decidable monadic second-order theory Thue specifications and their monadic second-order properties A method for enumerating cosets of a group presented by a canonical system. Finite automata. Mathematical Logic. An automata-theoretic approach to automatic program verification Automated reasoning: Introduction and Applications. first occur all self-Identify steps and other Identify steps occur after. --TR
automated deduction;infinite-state systems;formal verification;string-rewriting
594344
SAT-Based Decision Procedures for Classical Modal Logics.
We present a set of SAT-based decision procedures for various classical modal logics. By SAT based, we mean built on top of a SAT solver. We show how the SAT-based approach allows for a modular implementation for these logics. For some of the logics we deal with, we are not aware of any other implementation. For the others, we define a testing methodology that generalizes the 3CNFi>K methodology by Giunchiglia and Sebastiani. The experimental evaluation shows that our decision procedures perform better than or as well as other state-of-the-art decision procedures.
Introduction Propositional reasoning is a fundamental problem in many areas of Computer Science. Many researchers have put, and still put, years of eort in the design and implementation of new and more powerful SAT solvers. Most importantly, the source code of many of these implementations is freely available and can be used as a basis for the development of decision procedures for more expressive logics (see, e.g., (Giunchiglia and Sebastiani, 1996a; Cadoli et al., 1998)). In this paper, we present a set of SAT-based decision procedures for various classical modal logics (Montague, 1968; Segerberg, 1971). By SAT-based, we mean built on top of a SAT solver. We show how the SAT-based approach allows for a modular implementation for these logics. For some of the logics we deal with, we are not aware of any other implementation. For the others, we dene a testing methodology which generalizes the 3CNFK methodology by Giunchiglia and (1996a). The experimental evaluation shows that our deci- 1999 Kluwer Academic Publishers. Printed in the Netherlands. A. Tacchella sion procedures perform better than or as well as other state-of-the-art decision procedures. Our approach, rst suggested by Giunchiglia and Sebastiani (1998), consists of two steps: 1. take o the shelf one of the fastest SAT procedures available, and 2. use it as the basis for your modal decider. Our approach diers from most of the previous works in decision procedures for modal logics, which are either tableau-based, (i.e., based on Smullyan's tableau for propositional logic (Smullyan, 1968): see, e.g., (Fitting, 1983; Massacci, 1994; Baader et al., 1994)), or translation-based (i.e., based on a reduction to rst order logic: see, e.g., (van Benthem, 1984; Ohlbach, 1988; Hustadt and Schmidt, 1997a)). Some of the dierences and advantages of the SAT-based approach with respect to the tableau based and translation based approaches have been already pointed out in (Giunchiglia and Sebastiani, 1996b; Giunchiglia et al., 1998). Here we show how the SAT-based approach provides for a simple and natural schema for the development of decision procedures for modal logics. Our approach diers also from the approaches by Horrocks (1998) and Patel-Schneider (1998). These authors developed their own SAT checker and used it as a basis for their modal decider. This approach allows for a better integration among the dierent modules of the system and for a ner tuning of the reasoning strategies. On the other hand, it does not exploit the great amount of ongoing work on SAT deciders. Each year, new and faster SAT engines are proposed. With some standard modications of the source code we can inherit in the modal setting all the benets of state-of-the-art propositional checkers. 1 Finally, our system, called *sat, improves also on previous works by Sebastiani and ourselves (1996a; 1996b; 1998). In particular, *sat is built on top of the SAT decider sato (Zhang, 1997). *sat inherits many of the sato congurable options, and allows for some more options (including early pruning, caching, two forms of backjumping, and the choice of various splitting heuristics); and *sat is able to deal with 8 modal logics, namely E, EM (=M), EN, EMN, EC, EMC (=R), ECN, EMCN (=K) (see (Chellas, 1980)). For EN, EMN, EC and ECN, we do not know of any similar point is made by Kautz and Selman (1998) as an explanation for the better performances of SATPLAN with respect to specialized engines for planning problems. SAT-Based Decision Procedures for Classical Modal Logics 3 other implemented decision procedure, nor of any reduction to a formalism for which a decision procedure is available. The paper is structured as follows. In Section 2 we review some basic denitions about classical modal logics. Then, in Section 3, we present the idea of the SAT-based approach to modal reasoning, and discuss in detail the algorithms for the 8 logics we consider. Section 4 is devoted to *sat: it discusses some of the motivations underlying its construction, and some of its features. The experimental analysis comparing *sat with other state-of-the-art decision procedures is done in Section 5. Our analysis is restricted to K (for which there are several systems available) and E (for which there exists a reduction to bimodal K). We end in Section 6 with the conclusions and the future work. 2. Classical modal logics Following (Chellas, 1980), a modal logic is a set of formulas (called theorems) closed under tautological consequence. Most modal logics are closed under the rule for certain values of n. If then the logic is said to be monotone. then the logic is said to be regular. If n 0 then the logic is said to be normal. The smallest monotone, regular and normal modal logics are called M, R and K respectively (see, e.g., (Chellas, 1980)). Classical modal logics (Montague, 1968; Segerberg, 1971) are weaker than normal modal logics. In fact, the only requirement is that the set of theorems is closed under the rule As a consequence, the schemas 2 The symbols > and ? are 0-ary connectives representing truth and falsity respectively. 4 E. Giunchiglia, F. Giunchiglia and A. Tacchella which are theorems in K do not necessarily hold in classical modal logics. The three principles N, M, and C enforce closure conditions on the set of provable formulas which are not always desirable, especially if the 2 operator has an epistemic (such as knowledge or belief) reading. If we interpret 2' as \a certain agent a believes '", then N enforces that a believes all the logical truths, M that a's beliefs are closed under logical consequence, and C that a's beliefs are closed under conjunction. These three closure properties are dierent forms of omniscience, and |as such| they are not appropriate for modeling the beliefs of a real agent (see, e.g., (Giunchiglia and Giunchiglia, 1997) and (Fagin et al., Chapter 9). We can easily imagine situations involving a's beliefs, where only an arbitrary subset of the above properties holds. There are eight possible ways to add the three schemas M, C, and N to the smallest classical modal logic E. The resulting modal logics are called E, EM (equivalent to the logic M), EN, EMN, EC, EMC (equivalent to R), ECN, EMCN (equivalent to K), where EX denotes the logic obtained adding the schemas in X to E. 3. SAT-based procedures for classical modal logics We say that a conjunction of propositional literals and formulas of the form 2' or :2' is an assignment if, for any pair ; 0 of conjuncts in , it is not the case An assignment satises a formula ' if entails ' by propositional reasoning. A formula ' is consistent in a logic L (or L-consistent) if :' is not a theorem of L, i.e., if :' 62 L. Consider a formula '. Let S be a set of assignments satisfying ', and let L be a modal logic. As noticed by Sebastiani and Giunchiglia (1997), the following two facts hold: If at least one assignment in S is L-consistent then ' is L-consistent. If no assignment in S is L-consistent then ' is not L-consistent as long as the set S is complete for ', i.e., as long as the disjunction of the assignments in S is propositionally equivalent to '. From these facts, it follows that the problem of determining whether ' is L-consistent can be decomposed in two steps: generate an assignment satisfying ', and test whether is L-consistent. In all the logics we consider, testing the consistency of an assignment amounts to determining the consistency of other formulas whose SAT-Based Decision Procedures for Classical Modal Logics 5 depth (i.e., the maximum number of nested 2 operators) is strictly minor than the depth of . This implies that we can check the consistency of these other formulas by recursively applying the above methodology, at the same time ensuring the termination of the overall process. The above methodology can be implemented by two mutually recursive procedures: Lsat(') for the generation of assignments satisfying ', and Lconsist() for testing the L-consistency of each generated assignment . In order to simplify the presentation, we rst present Lconsist(), and then Lsat('). 3.1. Lconsist() Whether an assignment is consistent, depends on the particular logic being considered. Furthermore, depending on the logic L considered, the consistency problem for L (i.e., determining whether a formula is consistent in L) belongs to dierent complexity classes. In particular, the consistency problem for E, EM, EN, EMN is NP-complete, while for EC, ECN, EMC, EMCN it is PSPACE-complete (see (Vardi, 1989; Fagin et al., 1995)). Here, to save space, we divide these eight logics in two groups. We present the algorithms for checking the L-consistency of an assignment rst in the case in which L is one of E, EM, EN, EMN, and then in the case in which L is one of the others. 3.1.1. Logics E, EM, EN, EMN The following Proposition is an easy consequence of the results presented in (Vardi, 1989). PROPOSITION 1. Let be an assignment in which is a propositional formula. Let L be one of the logics E, EM, EN, EMN. is consistent in L if for each conjunct :2 j in one of the following conditions is satised: L-consistent for each conjunct 2 i in , and L=E; L-consistent for each conjunct 2 i in , and L=EM; are L-consistent for each conjunct 2 i in , and L=EN; are L-consistent for each conjunct 2 i in , and L=EMN. 6 E. Giunchiglia, F. Giunchiglia and A. Tacchella function Lconsist( foreach conjunct 2 j do foreach conjunct 2 i do if M [i; if M [j; if M [j; False then return False if M [j; if M [j; False then return False return True. Figure 1. Lconsist for E, EM, EN, EMN When implementing the above conditions, care must be taken in order to avoid repetitions of consistency checks. In fact, while an exponential number of assignments satisfying the input formula can be generated by Lsat, at most n 2 checks are possible in L, where n is the number of \2" in the input formula. Given this upper bound, for each new consistency check, we can cache the result for a future possible re-utilization in a This ensures that at most n 2 consistency checks will be performed. In more detail, given an enumeration of the boxed subformulas of the input formula, M[i,j], with i 6= j, stores the result of the consistency check for (' i ^:' j ). M[i,i] stores the result of the consistency check for :' i . Initially, each element of the matrix M has value Undef (meaning that the corresponding test has not been done yet). The result is the procedure Lconsist in Figure 1. Consider Figure 1 and assume that L=E or L=EN. Given a pair of conjuncts 2 i and :2 j , we split the consistency test for ( in two simpler sub-tests: rst, we test whether only if this test gives False, we test whether (: SAT-Based Decision Procedures for Classical Modal Logics 7 Notice also that, in case L=EN or L=EMN, if we know that, e.g., consistent and we store this result in M[j,j]. PROPOSITION 2. Let be an assignment in which is a propositional formula. Let L be one of the logics E, EM, EN, EMN. Assume that, for any formula ' whose depth is less than the depth of , Lsat(') returns True if ' is L-consistent, and False otherwise. returns True if is L-consistent, and False otherwise. Proof: The Proposition is an easy consequence of the hypotheses and Proposition 1. 3.1.2. Logics EC, ECN, EMC, EMCN The following Proposition is an easy consequence of the results presented in (Vardi, 1989). PROPOSITION 3. Let be an assignment in which is a propositional formula. Let be the set of formulas i such that 2 i is a conjunct of . Let L be one of logics EC, ECN, EMC, EMCN. is consistent in L if for each conjunct :2 j in one of the following conditions is satised: L-consistent for each non empty subset 0 of , and L=EC; L-consistent for each subset 0 of , and is empty or (( Assume that L=EC or L=ECN. The straightforward implementation of the corresponding condition may lead to an exponential number of checks in the cardinality jj of . More carefully, for each conjunct in , we can perform at most jj 1. for each formula i in , we rst check whether (: consistent in L. Let 0 be the set of formulas for which the above test fails. Then, 8 E. Giunchiglia, F. Giunchiglia and A. Tacchella function Lconsist( is a conjunct of g; foreach conjunct 2 j do foreach conjunct 2 i do if M [j; if M [j; if not Lsat( return True. Figure 2. Lconsist for EC, ECN, EMC (R), EMCN (K) 2. in case L=ECN or 0 6= ;, we perform the last test, checking whether (( consistent in L. Furthermore, the result of the consistency checks performed in the rst step can be cached in a matrix M analogous to the one used in the previous subsection. If L=EC or L=ECN, the procedure Lconsist in Figure 2 implements the above ideas. Otherwise, it is a straightforward implementation of the conditions in Proposition 3. PROPOSITION 4. Let be an assignment in which is a propositional formula. Let L be one of logics EC, ECN, EMC, EMCN. Assume that, for any formula ' whose depth is less than the depth of , Lsat(') returns True if ' is L-consistent, and False otherwise. returns True if is L-consistent, and False otherwise. Proof: The Proposition is an easy consequence of the hypotheses and Proposition 3. SAT-Based Decision Procedures for Classical Modal Logics 9 function Lsat(') return Lsatdp (cnf('), >). function Lsatdp (', ) if f a unit clause (l) occurs in ' g then return Lsat dp (assign(l; l := choose-literal('; ); return Lsat dp (assign(l; '), ^ l) or base */ /* backtrack */ /* unit */ /* split */ Figure 3. Lsat and Lsat dp 3.2. Lsat(') Consider a formula '. Let L be a modal logic. The generation of assignments satisfying ' is independent of the particular logic L being considered. Furthermore, it can be based on any procedure for SAT: if the SAT decider is complete, then we can generate a nite and complete set of assignments for ' as follows: at step 0, ask for an assignment satisfying ', and at step i for an assignment satisfying ' and the negation of the assignments generated in the previous steps. By checking the L-consistency of each assignment, we obtain a correct and complete decider for L. if the SAT decider is correct but incomplete, then we cannot generate a complete set of assignments for ', but we can still build a correct but incomplete decider for L by checking each generated assignment. Of course, whether an incomplete eective procedure for SAT can be turned into an eective incomplete procedure for a modal logic L, is still an open point. The above method for generating a complete set of assignments for ' has the advantage that the SAT decider is used as a blackbox. A. Tacchella The obvious disadvantage is that the size of the input formula checked by the SAT solver may become exponentially bigger than the original one. A better solution is to invoke the test for L-consistency inside the procedure whenever an assignment satisfying the input formula is found. In the case of the Davis-Putnam (DP) procedure (Davis and Putnam, 1960), we get the procedure Lsat represented in Figure 3. In the gure, cnf(') is a set of clauses |possibly with newly introduced propositional variables| such that, for any assignment in the extended language, the following two properties are satised: 3 1. if satises cnf(') then the restriction of to the language of 2. if satises ' then there exists an assignment in the language of cnf(') which (i) extends and (ii) satises cnf('). Examples of such conversions are the \classical conversion" (which given a formula in negative normal form recursively distribute conjunctions over disjunctions), and the conversions based on \re- naming", such as those described in (Tseitin, 1970; Plaisted and Greenbaum, 1986; de la Tour, 1990). choose-literal('; ) returns a literal occurring in ' and chosen according to some heuristic criterion. if l is a literal, l stands for A if l = :A, and for A if l = :A; for any literal l and formula ', assign(l; ') is the formula obtained from ' by (i) deleting the clauses in which l occurs as a disjunct, and (ii) eliminating l from the others. As can be observed, the procedure Lsat dp in Figure 3 is the DP- procedure modulo the call to Lconsist() when it nds an assignment satisfying the input formula. Notice that in this procedure the pure literal rule, used by some DP implementations, is not implemented. In fact, our main interest is in correct and complete modal deciders. With the pure literal rule, the set of assignments checked by Lconsist |assuming that each of such call returns False| is not ensured to be complete. 4 3 Let be an assignment in a language L. Let L 0 L be a language. The restriction of to L 0 is the assignment obtained from by deleting the conjuncts not in L 0 . Let 0 an assignment. 0 extends if each conjunct of is also a conjunct of 0 . 4 According to some authors (see, e.g., Freeman (1995)) the pure literal rule only helps for a few classes of SAT problems. Furthermore, when this rule does help, it does not result in a substantial reduction in search tree size. SAT-Based Decision Procedures for Classical Modal Logics 11 In the following, for any formula ' and for any assignment in the (possibly extended) language of cnf('), ' is the restriction of to the language of '. Analogously, for any set of assignments , g. PROPOSITION 5. Let ' be any modal formula. Assume that, for any assignment , Lconsist() prints and returns False. Let be the set of assignments printed as the result of invoking Lsat('). The disjunction of the assignments in ' is propositionally equivalent to '. Proof: In the hypotheses of the Proposition, what we need to prove is that _ First, it is clear that cnf(') 2 . For the thesis, the right-to-left implication is an easy consequence of the rst of the two properties of cnf('). Assume that the left-to-right implication is false. This means that there exists an assignment in the language of ' such that (1) for any propositional atom A, either A or :A is a conjunct of , (2) for any formula 2 in the language of ', either 2 or :2 is a conjunct of , (3) satises ', and does not satisfy . Given (3) and the second of the properties of cnf('), there must exist an assignment 0 in the language of cnf(') which extends , thus also . means that 0 entails 2 by propositional reasoning. Thus, there must exist an assignment 00 in such that |given (1), (2), (5)| each conjunct of 00 ' is also a conjunct of 0 ' and thus of . This implies that entails 00 ' . Since 00 , contradicting (4). We can now state and prove soundness and completeness results for our procedures. Theorem [Soundness and Completeness] Let L be one of the logics E, EM, EN, EMN, EC, ECN, EMC, EMCN. Lsat is sound and 12 E. Giunchiglia, F. Giunchiglia and A. Tacchella complete for L, i.e., for any modal formula ', Lsat(') returns True if ' is L-consistent, and False otherwise. Proof: First observe that Lsat(') returns True if there exists a call to Lconsist() which returns True, and False if each call to Lconsist() returns False. The proof is by induction on the depth d of '. If 0 the Theorem is trivial. Assume that 1. By induction hypothesis, for any formula whose depth is less than d, Lsat( ) returns True if is L- consistent, and False otherwise. In the following, we use the following two facts: (1) For any two assignments and 0 diering only for propositional conjuncts, is L-consistent i 0 is L-consistent. This is a trivial consequence of Proposition 1 and Proposition 3. (2) Let be an assignment such that Lconsist() is invoked during the execution of Lsat('). Lconsist() returns True if is L- consistent and False otherwise. Given that the depth of is less than or equal to d, this is a consequence of the the induction hypothesis, Proposition 2 and Proposition 4. There are two cases: 1. ' is L-consistent. From Proposition 5, it follows that there exists a call to Lconsist() in which the assignment (i) satises cnf('), and (ii) is such that ' is L-consistent. Since possibly extends ' in that it may assign some newly introduced propositional variables, from (1) it follows that is L-consistent, and from (2) it follows that Lconsist() |and thus also Lsat(')| returns True. 2. ' is not L-consistent. From Proposition 5, it follows that for each call Lconsist(), the assignment ' is not L-consistent. As above, from (1) it follows that is not L-consistent, and from (2) it follows that Lconsist() returns False. 4. *sat *sat is built on top of sato ver. 3.2 (Zhang, 1997). The choice of adopting sato as the basis for our system has been driven by the following motivations: SAT-Based Decision Procedures for Classical Modal Logics 13 sato is fast. We have not performed and are not aware of any up-to-date extensive comparison among the dierent SAT solvers publicly available. 5 However, according to some experiments we have done and to the results presented in (Zhang, 1997), sato seems to behave better than most of the currently available SAT solvers. sato has many options, including various splitting heuristics and backjumping. We have inherited some of these options, and they are available for experimentation. sato has been written using some Software Engineering conventions which have made and will make much easier to tune it for our goals. Besides the options inherited from sato, our system allows for other possibilities that we have developed while implementing the system. It is out of the goals of this paper to describe *sat structure, optimizations and congurable options. For a more detailed presentation, see (Tacchella, 1999) and the manual distributed with *sat. For our goals, it suces to say that the core of *sat is a C implementation of the procedures Lsat and Lconsist in Figures 1, 2, 3. In particular, with reference to Figure 3, in *sat cnf(') is the set of clauses obtained from ' by applying a conversion based on renaming, such as those described in (Tseitin, 1970; Plaisted and Greenbaum, 1986). choose-literal('; ) returns a literal according to a MOMS heuristic (Maximum Occurrences in clauses of Minimum Size) (Freeman, 1995). assign(l; ') is a highly optimized procedure which takes time linear in the number of occurrences of l in '. *sat also implements two important optimizations which have been used in the tests presented in the next Section: Early-pruning. Before each splitting step in Lsat, the L-consistency of the assignment generated so far is checked by a call to Lconsist. As in KsatC, care is taken to avoid the repetition of L-consistency checks on the same branch of the propositional search tree. 6 Early for a list of publicly available SAT solvers and more. 6 This is obtained through a pointer in the assignment stack which keeps track of the portion of the current assignment which has been already veried to be L-consistent. 14 E. Giunchiglia, F. Giunchiglia and A. Tacchella pruning has proved to be very eective at least on some of the tests presented in the next Section (see (Giunchiglia et al., 1998)). Caching for K. In the case of the logic K, *sat uses an additional data structure which allows to associate to any formula ' the result of Lsat('). Before invoking Lsat on a formula , Lconsist checks whether the K-consistency of has already been determined. As in caching introduces some additional costs, but it may produce dramatic speedups (see (Horrocks and Patel-Schneider, 1999b)). We have not yet conducted an exhaustive experimental analysis to see, for each class of formulas, which combination of *sat options leads to the best results (see (Horrocks and Patel-Schneider, 1999b) for a similar study of Dlp options). In all the tests in the next Section, we used *sat options which seemed more reasonable to us. In particular, we have set *sat as to use early pruning on all tests, and caching (for K) when the depth of the input formula is greater than 1. 5. A comparative analysis The availability of decision procedures for the logics we consider varies signicantly. For EMCN, that we recall is equivalent to K, there are many implemented decision procedures available, see, e.g., et al., 1998; de Swart, 1998). For E, EM and EMC, Gasquet and Herzig (1996) provide a reduction to normal modal logics: by implementing this reduction we indirectly obtain decision procedures for these logics. Fitting (1983) calls U the logic EM, and denes a tableau system for it. More recently, Governatori and Luppi (1999) dene a tableau-like proof system for classical, monotonic and regular modal logics. We are not aware of any implementation of these tableau sys- tems. For EN, EC, EMN and ECN we are not aware of any other implemented decision procedure, nor of any reduction into a formalism for which a decision procedure is available. Our comparative analysis is restricted to K and E. In fact, both our decision procedures for E and EM, and Gasquet and Herzig's reductions for E and EM, are similar. We expect that the experimental analysis for EM would lead to results similar to the ones we have for E. For SAT-Based Decision Procedures for Classical Modal Logics 15 EMC, Gasquet and Herzig's reduction is to a normal modal logic for which we do not have a system available. 7 5.1. Modal K As we said, there are several systems able to solve the consistency problem for K. In our comparative experimental analysis, we consider the four systems *sat, KsatC (Giunchiglia et al., 1998), Dlp (Patel- Schneider, 1998) and among the fastest solvers for K. We remember that TA, given a modal rst determines a corresponding rst order formula ' and then it performs conventional rst-order theorem proving. In our tests, as in (Hustadt and Schmidt, 1997a), uses Flotter to convert ' in a set of clauses Cl(' ), and then the theorem prover Spass to solve Cl(' ). For a brief description of Flotter and Spass, see (Weidenbach et al., 1996). 8 We test these systems on three problem sets of randomly generated formulas. A 3CNFK formula is a conjunction of 3CNFK clauses, each with three disjuncts. Each disjunct in a 3CNFK clause is either a propositional literal or a formula having the form 2C or :2C, where C is a 3CNFK clause. See (Giunchiglia and Sebastiani, 1996a) for a more detailed presentation. We only remark that for any formula ' there exist a 3CNFK formula which is K-consistent i ' is K-consistent. Sets of 3CNFK formulas can be randomly generated according to the following parameters: (i) the modal depth d; (ii) the number L of clauses at depth 7 The reduction maps the consistency problem for a formula ' in EMC, into the consistency problem for a formula ' 0 in the smallest normal modal logic with two modal operators 21 , 22 and augmented with the schema See (Gasquet and Herzig, 1996) for more details. 8 The experimental results have been obtained with Dlp ver. 3.1, ver. 1.4 (Spass/Flotter ver. 0.55), KsatC ver. 1.0, and *sat ver. 1.2. To compile the systems we have used sml-nj 110.0.3, sicstus prolog 3, ACL 5.0, and gcc 2.7.2.3. The tests have been run on several Intel PCs, whose conguration varies from P200MHz with 64MbRAM, up to a PII350MHz with 256 MbRAM. All platforms are running Linux 5.x RedHat. Dlp is available at http://www-db.research.bell-labs.com/user/pfps. at http://www.doc.mmu.ac.uk/STAFF/U.Hustadt/mdp. KsatC is available at ftp://ftp.mrg.dist.unige.it/ in pub/mrg-systems. *sat is available at the WEB page http://www.mrg.dist.unige.it/~tac/StarSAT.html. A. Tacchella (iii) the number N of propositional variables; (iv) the probability p with which a disjunct occurring in a clause at depth < d is purely propositional. Care is taken in order to avoid multiple occurrences of a formula in a clause, at the same time ensuring that the modal vs. the propositional structure of each generated formula only depends on p. In more detail, a clause is generated by randomly generating its disjuncts. When generating a disjunct, we rst decide whether it has to be a propositional literal or not. Then a disjunct of the proper type is repeatedly generated as long as it does not occur in the clause generated so far. In the tests we consider, a problem set is characterized by N and p: d is xed to 1, while L is varied in such a way to empirically cover the \100% satisable { 100% unsatisable" transition. For each pair in a problem set, 100 3CNFK formulas are randomly generated, and the resulting formulas are given in input to the procedure under test. Then, for each run, we consider the time the systems take for the main processing of the formula, thus excluding the negligible time the systems take to read and somehow normalize the input formula. (In particular, for this means that we take into account only the time needed by Spass to solve the formula generated at the end of the translation process.) For practical reasons, a timeout mechanism stops the execution of the system on a formula after 1000 seconds of CPU time. Even more, for any pair N; p, the execution of *sat, KsatC and stopped after the system exceeds the timeout on 51 of the 100 samples corresponding to the xed N; p. Dlp instead, stops its execution on the 100 samples corresponding to a pair N; p if a supermajority of the rst n tests timeout. 9 When this happens, it is assumed that Dlp exceeds the timeout for more than 50% of the tests with that The rst three problem sets we consider have xed to 0%: according to Hustadt and Schmidt (1997b), xing corresponds to particularly dicult tests. We call these problem sets PKN4p0, PKN5p0, and PKN6p0 respectively. PKN4p0 and PKN6p0 are called PS12 and PS13 respectively in (Hustadt and Schmidt, 1997b). In order to better highlight the behavior of *sat and KsatC, we also run these systems on a problem set (called PKN7p0) having In Figure 4, satisability percentages and the median of the CPU times for the four systems are plotted against the number 9 In more detail, Dlp stops its execution if there are at least 5 tests so far and more than 90% of them timeout, or if there are at least 10 tests so far and more than 75% of them timeout, or if there are at least 20 tests so far and more than 55% of them timeout (Patel-Schneider, 1999). SAT-Based Decision Procedures for Classical Modal Logics 17 Systems comparison - N=4, d=1, %p=0 CPU KsatC Systems comparison - N=5, d=1, %p=0 CPU KsatC Systems comparison - N=6, d=1, %p=0 CPU KsatC Systems comparison - N=7, d=1, %p=0 CPU KsatC Figure 4. Logic K. *sat, KsatC, Dlp, and 7. samples/point. Background: satisability percentage. of clauses L. Notice the logarithmic scale on the vertical axis, which causes that values equal to 0:00 do not get plotted. Consider Figure 4. The rst observation is that *sat and KsatC are the fastest. The two systems perform roughly in the same way, with *sat performing better when 5. For 7, the two systems have similar performances, one system performing better than the other for some values of L, but worse for other values of L. This comes at no surprise: the two systems have the same underlying structure, both use early pruning and a MOMS heuristic to select the splitting literal. The two systems do not have an identical behavior because they use dierent data-structures and implement slightly dierent MOMS strategies. Considering the other systems, for PKN4p0 the gap between *sat/KsatC and Dlp [resp. TA] is of more than one order of magnitude at the cross-over point of 50% of satisable formulas, and goes up to almost [resp. 4] orders of magnitude at the right end side of the horizontal axis. For PKN5p0 and PKN6p0, values exceed the timeout for respectively, while the corresponding values of A. Tacchella *sat [resp. KsatC] are 1.22 [resp. 1.83] and 3.32 [resp. 5.64] seconds. keeps exceeding the timeout for all the successive values. The gap between *sat/KsatC and Dlp on PKN5p0 is of more than one order of magnitude at the cross-over points of 50% satisable formula, and goes up to almost 3 orders at the very right of the plot. When Dlp median values exceed the timeout for does not terminate gracefully. 10 Comparing *sat and Dlp on PKN4p0 and PKN5p0, we see that the gap between *sat and Dlp seems to increase with L. Both for PKN4p0 and PKN5p0, the dierence in logarithmic scale between *sat and Dlp is (almost always) monotonically increasing. At the crossover point of 50% satisable formulas, the gap between *sat and Dlp [resp. TA] is roughly 1s [resp. 60s] for PKN4p0; and 220s [resp. 1000s] for PKN5p0. Such good performances by *sat and KsatC are due to early pruning, which has revealed to be very eective on these problem sets. For example, if we disable early pruning in *sat and rerun it on PKN5p0, the system keeps exceeding the time limit for When the better behavior of *sat and KsatC than Dlp and TA, is conrmed by the Q%-percentile graphs in Figure 5, corresponding to 4. Formally, the Q%-percentile of a set S of values is the value V such that Q% of the values in S are smaller or equal to . The median value of a set thus corresponds to the 50% percentile of the set. Figure 5 reports the 50%, 60%, 70%, 80%, 90% and 100% percentile values of the CPU-times when *sat (top left), KsatC (top right), Dlp (bottom left) and (bottom right) are run on PKN4p0. The percentile plots for PKN5p0 and PKN6p0 look similar to the plots in Figure 5. This means that all the systems perform in roughly the same way on the 50 most dicult samples of the 100 tests corresponding to a xed N and L. We also run the four systems on problems with p is xed to 50%. We call these problem sets PKN4p50, PKN5p50, PKN6p50, and PKN7p50. In Figure 6 the satisability percentages and the median of the CPU times for the four systems are plotted against the number of clauses L. As for roughly in the same way and are the fastest. Dierently from the tests in which 0%, at the transition point of 50% of satisable formulas, the gap between *sat/KsatC and Dlp seems to diminish when the number of variables increases. Horrocks and Patel-Schneider (1999a) ver. 3.2 is able to successfully handle formulas with SAT-Based Decision Procedures for Classical Modal Logics 19206010050709010 -2 percentiles CPU KsatC CPU TIME - N=4, d=1, %p=0 percentiles CPU percentiles CPU percentiles CPU Figure 5. Logic K. 50%{100% percentile CPU times of *sat, KsatC, Dlp, and TA. samples/point. show that for values of L bigger than 7, Dlp performances are superior to those of KsatC when that Dlp performs better than KsatC when p is high and worse when p is low. We have not yet done such a broad comparison using *sat instead of KsatC. However, we believe that Horrocks and Patel- Schneider's conclusions should extend also to *sat, if *sat is used with the parameter settings we have currently used, i.e., those which make *sat most similar to KsatC. Of course, a big role can be played by *sat already available congurable options, and this will be the issue of future research. In any case, both *sat and KsatC perform better than the other systems for large values of L, when the formulas are trivially unsatisable. This is due to the fact that for large values of L, formulas become propositionally unsatisable, and thus both *sat and KsatC mostly take advantage of their SAT-based nature, e.g., of their optimized data structures for handling large formulas. As for the percentile plots of the timings of the systems on PKN7p50, do not show big dierences with respect to the plots of 20 E. Giunchiglia, F. Giunchiglia and A. Tacchella Systems comparison - N=4, d=1, %p=50 CPU KsatC Systems comparison - N=5, d=1, %p=50 CPU KsatC Systems comparison - N=6, d=1, %p=50 CPU KsatC Systems comparison - N=7, d=1, %p=50 CPU KsatC Figure 6. Logic K. *sat, KsatC, Dlp, and 7. samples/point. Background: satisability percentage. the medians. Considering Dlp and *sat 100% percentile plots, it is interesting to observe that Dlp has a lower maximum than *sat; on the other hand, *sat values decrease more rapidly than those of Dlp. Finally, we see that for most of the problem sets we consider, all the systems seem to have an easy-hard-easy pattern, whose peak roughly correspond to the 50% of satisable formulas. When phenomenon is best evident for *sat and KsatC. We also consider the benchmarks formulas for K used at the Comparison of Theorem Provers for Modal Logics at Tableaux'98 (see (de Swart, 1998)). These consist of nine provable parameterized formulas (ending with \ p") and nine unprovable parameterized formulas (ending with \ n"). For each parameterized formula A(n), the test consists in determining the greatest natural number n 21 satisfying the following two conditions: 1. the prover returns the correct result for the formulas A(1); in less than 100 seconds, and SAT-Based Decision Procedures for Classical Modal Logics 21 percentiles CPU KsatC CPU TIME - N=7, d=1, %p=50 percentiles CPU percentiles CPU percentiles CPU Figure 7. Logic K. 50%{100% percentile CPU times of *sat, KsatC, Dlp, and TA. samples/point. 2. the prover cannot do the formula A(n+ 1) in less than 100 seconds or Even though it has been proved that most of these tests can be easily solved by current solvers, these are still interesting because they are not 3CNFK formulas, and some of these tests have not been solved yet. The results for *sat, and Dlp are reported in Table I. KsatC has not been tested since KsatC is able to deal with 3CNFK formulas only. We also show the CPU time requested by the system to solve the last instance A(n). Notice that *sat has been run with caching enabled, since the depth of all the formulas in the Table is greater than 1. For TA, we do not take into account the time needed to compute the rst corresponding to A(n) (which is negligible), but we do take into account the time requested by Flotter to convert A (n) 22 E. Giunchiglia, F. Giunchiglia and A. Tacchella Table I. Logic K. *sat, Dlp and performances on Tableaux'98 benchmarks *sat Dlp Test Size Time Size Time Size Spass Flotter branch p 21 0.21 lin n 21 47.80 21 0.70 21 16.07 63.94 lin path n 21 0.96 path p 21 0.72 ph n 12 0.60 9 40.16 9 45.21 9.92 ph p 8 48.54 6 11.34 6 42.19 0.97 poly n 21 2.25 poly p 21 1.73 into a set Cl(A (n)) of clauses (reported in the Flotter column), and the time requested by Spass to determine the consistency or inconsistency of (reported in the Spass column). Furthermore, we stopped on A(n) with n < 21, either because Flotter does not terminate gracefully when computing or Flotter exceed the 100 seconds time limit. In the table, these three cases correspond to the rows in which the value for n/Spass/Flotter respectively is underlined. As can be observed from Table I, the three systems are able to solve all the instances of a formula in four cases. *sat and Dlp are able to solve all the instances except for k branch n, k branch p, k ph n, ph p. Except for the rst of these four parameterized formulas, *sat is able to solve more instances than Dlp. For k branch n, both *sat and Dlp are able to solve the 12th instance, with Dlp taking less time than *sat to solve it. SAT-Based Decision Procedures for Classical Modal Logics 23 5.2. Modal E Gasquet and Herzig (1996) provide a translation which maps any formula ' into a formula 'GH in K 2 , i.e., the smallest normal modal logic with two modal . The translation is such that ' is satisable in E i 'GH is satisable in K 2 . This translation is dened in the following way: is a propositional variable, and homomorphic for the cases of the propositional connectives. Consider a formula '. We compare *sat performances on ' with respect to *sat, Dlp and performances on 'GH . We could not run KsatC on 'GH , since KsatC accepts only 3CNFK formulas with at most one modality. To make evident when a system is run using Gasquet and Herzig's translation, we append the string \+GH" to the name of the system. Therefore, in the following, we will have the systems *sat, *sat+GH, Dlp+GH, and TA+GH. In E, the 3CNFK test methodology is not suited. Indeed, it is no longer the case that for any modal formula ' there exists a 3CNFK formula which is E-satisable i ' is E-satisable. Furthermore, checking the consistency of an assignment in E amounts to determine the consistency of ( :) for each pair of conjuncts 2 and :2 in : most of these tests, in case and are 3CNFK clauses, can be trivially We therefore consider sets of 3CNFE formulas. A 3CNFE formula is a conjunction of 3CNFE clauses, each with three disjuncts. Each disjunct in a 3CNFE clause is either a propositional literal or a formula having the form 2C or :2C, where C is a 3CNFE formula. For example, where each l i (1 i 8) is a propositional literal, is a 3CNFE formula. For any formula ', there exist a 3CNFE formula logically equivalent to ' in E. Sets of 3CNFE formulas can be randomly generated according to the parameters used to generate 3CNFK formulas, and a new parameter C representing the number of clauses at depth d > 0. A 3CNFK is thus a 3CNFE formula in which C = 1. As in the previous subsection, a problem set is characterized by N and p: d and C are xed to 1 and L respectively; L is given increasing values in such a way to empirically cover the \100% satisable { 100% unsatisable" transition. We also A. Tacchella Systems comparison - N=4, d=1, %p=0 CPU TA+GH Systems comparison - N=5, d=1, %p=0 CPU Systems comparison - N=6, d=1, %p=0 CPU Systems comparison - N=7, d=1, %p=0 CPU TA+GH Figure 8. Logic E. *sat, *sat+GH, Dlp+GH, and TA+GH median CPU time. samples/point. Background: satisability percentage. check that in each sample there are no multiple occurrences of a formula in a clause, at the same ensuring that the propositional vs. the modal structure of the formula only depends on p. Notice that while increasing L also C is increased. As a consequence, for each pair of in an assignment satisfying a 3CNFE formula, the recursive E-consistency check for ( has itself a phase transition from 100% satisable to 100% unsatisable when increasing L. Overall, for low [resp. high] values of L we expect that each satisfying assignment should be trivially determined to be E-consistent [resp. not E-consistent]. As before, for each value N in a problem set, 100 3CNFE formulas are randomly generated, and the resulting formulas are given in input to the procedure under test. A timeout stops the execution of the system on a formula after 1000 seconds of CPU time. We consider the following problems sets: SAT-Based Decision Procedures for Classical Modal Logics 25 percentiles CPU percentiles CPU percentiles CPU percentiles Figure 9. Logic E. 50%{100% percentile CPU times of *sat, *sat+GH, Dlp+GH, and TA+GH. samples/point. PEN4p0, PEN5p0, PEN6p0, PEN7p0 in which respectively, and PEN4p50, PEN5p50, PEN6p50, PEN7p50 in which Given the huge amount of time that Flotter takes to prepare the formula for Spass, we run TA+GH only on the problems sets PEN4p0 and PEN4p50. For PEN7p0 and PEN7p50, we run TA+GH only on the initial points. As in the preceding subsection, we only take into account the time the systems take for the main processing of the formula. In particular, for each system, we do not take into account the time needed to perform the Gasquet and Herzig's conversion; and for TA+GH we take into account only the time taken by Spass. The median and the percentile plots of the systems on PEN4p0, PEN5p0, PEN6p0, PEN7p0 are shown in Figure 8 and Figure 9 respectively. Consider Figure 8. As can be observed, *sat is the fastest: the gap with the other systems is of more than one order of magnitude for 26 E. Giunchiglia, F. Giunchiglia and A. Tacchella Systems comparison - N=4, d=1, %p=50 CPU TA+GH Systems comparison - N=5, d=1, %p=50 CPU Systems comparison - N=6, d=1, %p=50 CPU Systems comparison - N=7, d=1, %p=50 CPU Figure 10. Logic E. *sat, *sat+GH, Dlp+GH, and TA+GH median CPU time. samples/point. Background: satisability percentage. certain values of L. However, both *sat+GH and Dlp+GH perform quite well, better than one could have imagined given that the consistency problem for E and K 2 belongs to two dierent complexity classes. However, a closer look to Gasquet and Herzig's reduction reveals that, considering a 3CNFE formula ', and an assignment (as usual we assume that is a propositional formula) in the language of ', 1. satises ' i GH satises 'GH . 2. for checking the E-consistency of , *sat performs at most 2mn consistency checks involving the formulas 3. for checking the K 2 -consistency of GH , both *sat+GH and Dlp+GH perform at most 2mn consistency checks involving the formulas SAT-Based Decision Procedures for Classical Modal Logics 27 percentiles CPU percentiles CPU percentiles CPU percentiles CPU Figure 11. Logic E. 50%{100% percentile CPU times of *sat, *sat+GH, Dlp+GH, and TA+GH. samples/point. The rst two points are obvious. To understand the last, it suces to notice that GH is propositionally equivalent to Given that both *sat+GH and Dlp+GH use caching, these procedures will perform at most a quadratic number of checks in the number of subformulas of 'GH . This is not the case for TA+GH, since Spass does not have any caching mechanism. This explains the good behavior of *sat+GH and Dlp+GH, and the bad behavior of TA+GH. It is interesting to compare *sat+GH and Dlp+GH performances. As can be observed, Dlp+GH performs better than *sat+GH for low values of L, but worse for high values of L. This behavior re ects the dierent mechanisms used by *sat+GH and Dlp+GH to prune the search space when checking the K-consistency of an assignment. As we said in Section 4, in all the tests we have set *sat+GH as to 28 E. Giunchiglia, F. Giunchiglia and A. Tacchella use the early pruning strategy. Dlp+GH instead implements a backjumping schema in the spirit of (Baker, 1995): when an assignment is discovered to be not K-consistent, backtracking to a point which does not lead to the same contradiction is enforced. While implementing early pruning does not introduce overheads, this is not the case for backjumping, where a dependency set of each derived clause has to be maintained (see (Patel-Schneider, 1998) for more details). Despite the additional costs introduced, backjumping clearly wins if compared to early pruning in logic K, for low values of L. In this case, almost each assignment is K-consistent and early pruning may cause additional (i.e., not performed by a backjumping strategy) checks. On the other hand, for high values of L, when the formula under test is not K-consistent but there are still assignments satisfying it, *sat+GH is able to greatly cut the search by checking the inconsistency of the assignment generated so far. Dlp+GH instead checks the consistency of an assignment only when it satises the current formula. Dlp+GH may therefore generate many assignments which, even though they satisfy the input formula, are not K-consistent. When L is so high that the input formulas become propositionally unsatisable, *sat+GH may still perform additional K-consistency checks, but these get compensated by (i) *sat+GH SAT-based nature and (ii) the costs Dlp+GH has because of back- jumping. Considering Figure 9, we see that Dlp+GH has a better behavior than the other systems on some of the hardest instances. Evidently, on these tests, backjumping leads to a more uniform behavior than early pruning. Horrocks and Patel-Schneider (1999a) show that for some randomly generate 3CNFK formulas, early pruning leads to a more uniform behavior than backjumping. For TA+GH median and percentile times are plotted in Figure 10 and Figure respectively. As it can be observed, the situation is very similar to the case in which 0%. The only dierence is that now *sat+GH performs better than Dlp+GH for a lower value of L. This is reasonable, since for each L, the number of consistency checks performed by *sat+GH because of early pruning, diminishes when p increases. Finally, notice the easy-hard-easy pattern of *sat. To better appreciate it, Figure 12 shows the number of calls to Lsat done by Lconsist on PEN4p0-PEN7p0 (left), PEN4p50-PEN7p50 (right) against the ratio between L and N . As can be observed, *sat performs a number of Lsat calls whose maximum roughly correspond to the 50% of satisable formulas. This transition happens when L N is close to 5 for to 6 for This behavior re ects the above stated intuition SAT-Based Decision Procedures for Classical Modal Logics 29 checks d=1 %p=0 OF CHECKS 7 variables 6 variables 5 variables 4 variables Consistency checks d=1 %p=50 OF CHECKS 7 variables 6 variables 5 variables 4 variables Figure 12. Logic E. *sat median number of calls to Lsat. 7. samples/point. Background: satisability percentages. according to which for low [resp. high] values of L all the formulas should be easily determined to be E-consistent [resp. not E-consistent]. 6. Conclusions and future work We have presented a set of SAT-based decision procedures for eight classical modal logics. We have shown how the SAT-based approach allows for ecient and modular implementations for these logics. We have presented *sat. *sat is the only system that is able to deal with EN, EC, ECN and EMN. In the case of the logic E, we have dened a testing methodology which generalizes the 3CNFK methodology by Giunchiglia and Sebastiani (1996a), and which is suitable for testing systems for non-normal modal logics. The experimental evaluation shows that *sat performances are superior to or comparable to the performances of other state-of-the-art systems. In the future, we plan to conduct an extensive experimental analysis (similar to that presented in (Horrocks and Patel-Schneider, 1998; Horrocks and Patel-Schneider, 1999b)) to understand, for each class of formulas, which combination of *sat options leads to the best results. We also plan to extend *sat in order to handle more expressive decidable logics. We will also consider logics, like S4, for which more sophisticated methods than that described in Section 4 have to be employed in order to ensure the termination of the decision procedure. A. Tacchella Acknowledgments We are grateful to Hullrich Hustadt, Peter F. Patel-Schneider, and Hantao Zhang for the assistance they provided on their systems. Special thanks to Roberto Sebastiani for the many useful discussions related to the subject of this paper. Thanks also to the anonymous reviewers for their helpful comments and suggestions. The rst and last authors are partially supported by the Italian Spatial Agency. --R Intelligent backtracking on the hardest constraint problems. An algorithm to evaluate quanti Modal Logic - an Introduction A computing procedure for quanti Swart, editor. Automated Reasoning with Analytic Tableaux and Related Methods: International Conference Tableaux'98 Reasoning about knowledge. Proof Methods for Modal and Intuitionistic Logics. Improvements to propositional satis PhD thesis From classical to normal modal logics. Ideal and Real Belief about Belief. Building decision procedures for modal logics from propositional decision procedures - the case study of modal of the 5th International Conference on Principles of Knowledge Representation and Reasoning - KR'96 More evaluation of decision procedures for modal logics. Labelled tableaux for non-normal modal logics Comparing subsumption optimizations. Advances in propositional modal Optimising description logic subsumption. Journal of Logic and Computation Using an expressive description logic: FaCT or On evaluating decision procedures for modal logic. On evaluating decision procedures for modal logic. BLACKBOX: A new approach to the application of theorem proving to problem solving. Strongly analytic tableaux for normal modal logics. Translation methods for non-classical logics - an overview DLP system description. Personal communication A Structure-preserving Clause Form Translation Journal of Symbolic Computation From Tableau-based to SAT-based procedures - preliminary report An Essay in Classical Modal Logic. On the complexity of proofs in propositional logics. Correspondence theory. On the complexity of epistemic reasoning. --TR --CTR Andrei Voronkov, How to optimize proof-search in modal logics: new methods of proving redundancy criteria for sequent calculi, ACM Transactions on Computational Logic (TOCL), v.2 n.2, p.182-215, April 2001 Enrico Giunchiglia , Armando Tacchella, A Subset-Matching Size-Bounded Cache for Testing Satisfiability in Modal Logics, Annals of Mathematics and Artificial Intelligence, v.33 n.1, p.39-67, September 2001 SAT-based planning in complex domains: concurrency, constraints and nondeterminism, Artificial Intelligence, v.147 n.1-2, p.85-117, July Enrico Giunchiglia , Yuliya Lierler , Marco Maratea, Answer Set Programming Based on Propositional Satisfiability, Journal of Automated Reasoning, v.36 n.4, p.345-377, April 2006 Alessandro Armando , Claudio Castellini , Enrico Giunchiglia , Marco Maratea, The SAT-based Approach to Separation Logic, Journal of Automated Reasoning, v.35 n.1-3, p.237-263, October 2005 Franz Baader , Diego Calvanese , Deborah L. McGuinness , Daniele Nardi , Peter F. Patel-Schneider, Bibliography, The description logic handbook: theory, implementation, and applications, Cambridge University Press, New York, NY,
modal logic;SAT-based decision procedure
594390
Fundamental Laws and Assumptions of Software Maintenance.
Researchers must pay far more attention to discovering and validating the principles that underlie software maintenance and evolution. This was one of the major conclusions reached during the International Workshop on Empirical Studies of Software Maintenance. This workshop, held in November 1996 in Monterey, California, brought together an international group of researchers to discuss the successes, challenges and open issues in software maintenance and evolution.This article documents the discussion of the subgroup on fundamental laws and assumptions of software maintenance. The participants of this group included researchers in software engineering, the behavioral sciences, information systems and statistics. Their main conclusion was that insufficient effort has been paid to synthesizing research conjectures into validated theories and that this problem has slowed progress in software maintenance. To help support this vision they identified the following supporting conclusions. (1) need to develop a genuine scientific approachmore than just using empirical methods, (2) other disciplines can help, (3) need to explore a wider range of approaches to empirical studies, (4) need to study both evolution and maintenance.
Introduction All software systems evolve [1]. As they evolve they undergo numerous, successive changes - to fix errors, improve performance or other attributes, and to adapt to new environments. The longer a system stays in service, the larger maintenance costs are, and, therefore, it's not surprising that maintenance costs often dominate initial development costs [2]. Clearly, improved tools, techniques, and processes can save a great deal of time and money throughout the software industry. To realize these savings, researchers are studying maintenance from many perspectives. Making individual modifications. To make changes developers have to understand the existing system, evaluate the effect of proposed changes, and then implement and validate it. To address these issues, researchers study topics such as design recovery, program comprehension, impact analysis and regression testing. Coping with evolution. Because systems undergo many changes, not just one, other researchers focus on the nature of change and how to limit its impact. Research in this area includes structured programming, object-oriented programming languages, software architectures and configuration management. Managing the maintenance process. Maintenance is costly and somewhat unpredictable. Thus, management tools for controlling maintenance are in great demand. Many researchers have developed metrics and models to predict when components are likely to need changing and how much that will cost. These are only a few examples of maintenance research. There are many others. Clearly, software maintenance is a significant problem that continues to foster considerable activity. However, as with other software engineering areas, there is a concern that our research efforts lack hard evidence and critical evaluation, and that without these, we can't develop a deep understanding of what tools and processes work, when they work, or why. Consequently, many people believe that rigorous empirical methods must be one of the cornerstones of research in our field. This belief brought a large group of researchers to the International Workshop on Empirical Studies of Software Maintenance. The workshop was help on November 8, 1996 in Monterey, California and involved researchers from software engineering, the behavioral sciences, information systems and management, and statistics. The goal of this meeting was to discuss the strengths, weaknesses, and open issues in empirical methods and determine how they could be profitably applied to improve software maintenance research. The workshop attendees were divided into four groups. 1. Study and assessment of (new) technologies, 2. Methodologies for performing empirical studies, 3. Studies of the maintenance process: fundamental laws and assumptions, and 4. Maintenance process improvement in practice. This article summarizes the findings of subgroup #3: Studies of the maintenance process: fundamental laws and assumptions. This subgroup is distinguished from others because their studies are not specific to particular tools or methods. Instead it tries to identify general principles. The group's participants have expertise in empirical methods, but approach their studies in different ways. Their backgrounds are diverse: software engineering, behavioral science, business, statistics, and others. 2. Summary of Discussion Subgroup #3 had two working sessions. During the first session, each participant presented their research and took questions from the other participants. Before the workshop each participant received the following list of questions about their work. 1. What were the study's goal and questions? 2. What empirical methods did you use (e.g., controlled experiment, case study, survey)? 3. Describe your study's design. 4. What threats to validity did you consider (construct, internal, external)? 5. What did you do well and what will you change in the future? 6. How much did it cost to perform the study? 7. What new questions do your results raise? 8. Can other researchers replicate your study? 2.1 Synopsis of Presentations. This section summarizes each presentation. Of course, the few paragraphs given to each presentation cannot capture the careful thought that goes into designing and conducting empirical research. Please see the complete proceedings. Any mistakes or misrepresentations are unintentional, and are the sole responsibility of the session chair, not the workshop participants. . Maintenance of Reuse-Based Domain-Specific Software Product Lines[3]. Presented by William Thomas: Reuse saves money and time. Therefore, greater levels of reuse promise even greater savings. Although this seems to be correct, the authors urge us to look more closely. A common way to get high reuse levels is to split systems into two parts: one specific to the domain and the other to the application. The domain-specific part provides the infrastructure: interconnection support, common functions, and system architecture. Everything else is the application-specific part. The authors strongly agree that this approach reduces coding effort, but warn us that it may also redefine maintenance. Traditionally, applications are developed and maintained separately. However, when applications share code, they may need to be maintained as a unit. Of course, developers can still choose to maintain all applications individually, but may lose the benefit of domain-specific assets. They might also choose to maintain all related applications as a unit, but that may constrain individual applications. Since the best way to manage this situation is unclear, the authors will conduct empirical studies using historical data. They will explore such issues as measuring the cost of evolving applications and domains, coupling between applications and domain, patterns of change, and the effect of domain maturity on maintenance strategies. This work reminds us that new technology can challenge, even invalidate, our basic assumptions. . Using a Didactic Model to Measure Software Comprehension[4]. Presented by Elisabeth Dusink One of the first steps in changing a system is understanding it. Documentation is supposed to help understanding because it collects system knowledge and presents it in a structured way. The authors of this article claim that developers learn about a system by reading its documentation. Furthermore, they suspect that qualities of the learning process affect the quality of maintenance. Based on a model of learning by Bloom[5], they argued that learning progresses from gaining knowledge, to understanding it, applying it, analyzing it, and, finally, to synthesizing new knowledge from it. The authors discussed a controlled experiment to link these levels of learning to maintenance success. The experimental subjects, second-year CS students, will be placed into three- and four-person groups. Each group will be given a 7,000 line email system written in C, and asked to modify it. They expect the change to take about 80 hours per person to complete. For each modification the experimenters will determine whether the group made the correct changes. Each person's understanding level will be measured via a questionnaire. One potential outcome of this research is a low-cost index for maintenance success. That is, we may be able to test new technology by measuring its effect on understanding levels rather than by measuring its effect on maintenance tasks themselves. . Change-Prone Modules, Limited Resources and Maintenance[6]. Presented by Warren Harrison Maintainers often work under tight deadlines with meager resources. So sometimes they take short cuts even though they'd prefer to design and implement each change carefully. This kind of change is "ad hoc". It can be made rapidly, but often degrades structure. In contrast "planned, structure preserving" changes may preserve structure, but are more expensive. Harrison argues that ad hoc changes aren't as damaging as has been claimed. He discussed data that showed that, in practice, changes tend to be confined to a small part of the system. His conclusion is that modules that rarely change cannot degrade and, thus, can tolerate ad hoc changes. The author also outlined a model to help decide when to make ad hoc changes and when to restructure. In the model, the cost of a maintenance request is a function of the request's complexity, the module's state, and the kind of change made. Although the authors are still developing the model, it captures two important notions: (1) ad hoc patches degrade structure more than planned, structure preserving changes do, and (2) the effect of ad hoc changes will be compounded in modules that change often. One interesting feature of this research is that it uses mathematical techniques to model long-term evolution. Since, by definition, evolution takes time, researchers must find ways to reason about it. Tools like these are a first step in that direction. . An Empirical Exploration of Code Decay[7]. Presented by Adam Porter Software systems must tolerate numerous, successive changes. As this happens, systems deteriorate and changes become increasingly difficult. Eventually, new functionality cannot be added and the system must be redesigned. Since solutions to this problem would be valuable, these researchers are conducting a long-term, multidisciplinary project to examine the fundamental causes, symptoms, and remedies for code decay. The project team contains researchers in Statistics, Experimentation, Organizational Theory, Programming Languages, Software Engineering, and Visualization. Their primary data source is the AT&T 5ESS^TM switching system and its development data. This includes the switch's source code and change control history (-700,000 changes) covering the last 15 years, its planned and actual development milestones, effort and testing data, organizational history, development policies, and coding standards. To help the project achieve its goals quickly, they have constructed a set of model code artifacts that they call the "code decay testbed". To do this they developed several small systems, attempted to induce code decay in them, and used the resulting systems to evaluate ideas about code decay. This research is a good example of how laboratory and field studies can work together to strengthen results. . On Increasing Our Knowledge of Large-Scale Software Comprehension[8]. Presented by Anneliese von Mayrhauser Far too often researchers create solutions (tools) without understanding the problem. Thus, there's a tendency to focus on a tool's novelty or performance rather than on its efficacy. The authors believe that program comprehension is a fundamental maintenance problem. Consequently, they have begun to study how software developers understand programs and how maintenance tools can make it easier. They presented a case study in which they analyzed how 11 professional programmers performed maintenance. Then they tried to correlate what they saw with known theories of program comprehension. These theories suggest that several factors may affect maintenance: task type, prior exposure to the system, domain expertise, and language expertise. The authors felt that limited sample size was one of the biggest problems that they and other experimenters face. Therefore, they argue that we will have to combine data from multiple studies. Further, this will require better methods for comparing studies, sharing data and terms, and analyzing aggregate data. One implication of this is that experimental results must generalize beyond the environment from which they were taken. This may be one of the most important open issues in empirical software engineering. . The Software Maintenance Process in Portugal[9]. Presented by Helena Mendes-Moreira Maintenance costs are the dues of success. In this work the authors surveyed 37 software development organizations in Portugal to characterize their maintenance processes. They selected companies with more than 500 developers and revenues in excess of 13 million dollars. After they examined the survey responses they conducted interviews with several of the respondents. They found that most of the maintenance performed in these organizations involved quick fixes, rather than planned enhancements. Also, in many cases, support activities such as updating documentation were not done. More importantly, they found that the ratio of maintenance effort to new development effort was growing rapidly. This might be expected in companies that enter the software market, develop new products, and now, for the first time, must maintain them. Several participants noted similar patterns in US industry. This work raises the possibility of comparing historical data from established organizations with current data from emerging ones. Would they find common patterns of organizational growth despite the changes in technology? . Predicting Aspects of Maintainability from Software Characteristics[10]. Presented by Jarrett Rosenberg We have to change the way we study change. The author presented several studies that tried to predict the likelihood and cost of certain maintenance tasks. His results cast doubt on the soundness of current metric-based modeling approaches. First he found little correlation between static complexity metrics and repair activity. Second, he found considerable and unexplained variation in the metric values, suggesting that other, unknown, factors are driving maintenance. The author stressed two flaws in current metrics research. Metrics tend to be purely syntactic, which excludes vital semantic information. They also tend to be static, ignoring information about a system's development and evolution (e.g., testing history). The solution, however, is not to define new measures. Instead, he argued, we will need to focus on creating deep theories of maintenance-related factors. This presentation stressed two important research goals: to continue raising our scientific standards, and to borrow wisely from the ideas and approaches of other research areas. . Operational System Complexity[11]. Presented by Scott Schneberger Maintenance bottlenecks can shift as technology changes. The industry is moving more and more from centralized to distributed architectures. Will this shift change our assumptions about maintenance? Schneberger suggests that it might. Since, he claims, there is little data on the maintenance of distributed applications, he performed an exploratory study. In this work he modeled a system's complexity in terms of its components, their internal complexity, and their interactions. To test this idea he conducted a survey. His initial results were that as distribution increased (i.e., components spread over more processors), system complexity increased, while component complexity decreased. One of his conclusions is that distribution shifts complexity out of individual components and into their interconnections. The author also points out that, in some domains, there is a trend back toward centralization. These results suggest that the benefits of solving certain maintenance problems can change over time. . The Impact of Documentation Availability on Software Maintenance Productivity[12]. Presented by Eirik Tryggeseth One way to evaluate technology is to ask what would happen if it didn't exist. The author takes this approach to see how documentation affects maintenance success. He also asks whether this effect is different for more- or less-skilled programmers. To do this he performed a controlled experiment with 34 undergraduate students in computer science as subjects. Before the experiment he also measured the students' skill in reading and writing C++ programs. During the experiment the participants were asked to modify a 2700 line C++ program. Half of them, group B, were given documentation, the rest, group A, were not. The experimenters measured the amount of time that the participants spent understanding and modifying the system, and they measured the quality of the modification. They found people who had documentation, group B, spent less time understanding the system and made higher quality changes. They also found that the performance measures for group B were correlated with the skill measures, while those of group A were not. One explanation might be that Group B understood the system quickly, so programming skill became the limiting factor for task performance. Group A, however, found it hard to understand the system, and, therefore, were unable to profit from their programming skill. One implication of this result might be that system-level knowledge has a greater (or, at least, more immediate) effect on maintenance quality than individual programming skill does. . Assessing Maintenance Processes Through Controlled Experiment[13]. Presented by Guiseppe Visaggio Sometimes the simplest approach is the most cost-effective. In this work the author explores the cost-benefit tradeoffs between quick-fixes and a more thorough change process, called iterative enhancement. The goal of this work is similar to Harrison's (described earlier), but uses a controlled experiment rather than mathematical modeling. The participants in this experiment were asked to modify two different systems, once using the quick-fix approach, once using an iterative enhancement approach. The author measured the correctness, completeness, effort, and traceability of each change. They found that quick fixes were less reliable and degraded structure more than iterative enhancements did. One other interesting result was that quick- fixes were done faster only when modification affected fewer than ten modules. One fascinating observation is that these results appear to agree with those discussed by Harrison even though the researchers used very different research methods. 2.2 Organizing the Current Literature In the second session the group analyzed their presentations and other existing research to see what, if any, fundamental laws of software maintenance are known. Because maintenance has so many facets, we first developed a very rough taxonomy of the factors that might affect maintenance. Then for each category we tried to synthesize common results, hoping to identify potential laws. We thought this might also indicate open areas that should be, but are not being, addressed. The categories represent factors related to product, people, process, and task. 1. Product. These factors relate to how attributes of system artifacts affect maintenance. We divided product factors into two subcategories: complexity and structure. We assumed that complexity affects our ability to understand the product. Research in this area includes software complexity metrics and structured programming. Structure refers to how system components are organized. We assumed that structure affects how changes impact a system. Research in this area includes studies of ripple-effect, development of design patterns, and studies of software architecture and domain-specific languages. 2. People. These factors relate to how the attributes of individuals and groups affect maintenance. We discussed three levels of this factor: Individual, Team, and Organizations. We assumed that abilities, group dynamics, and organizational constraints affect maintenance. Research in these areas include program comprehension, groupware, and cycle-time reduction studies. 3. Process. These factors relate to how the activities that individuals and teams carry out routinely affect maintenance. We assumed that process factors affect an organization's ability to predictably achieve their development goals. Research in this area includes documentation approaches, and configuration management tools. 4. Task. These factors relate to the viewpoint from which maintenance is studied. Some research focuses on individual changes while other focuses on longer-term evolution processes. Within each of these categories, we tried to extract common findings. One topic for which we had some success was modularization. Several people stated that Parnas' early work on information hiding [14] illustrated some benefits of modularization, and that the effect of this and other research can be seen in today's object-oriented programming languages. See Kemerer [15] for a survey of results in this area. Unfortunately, we were unable to find too many more. Certainly some exist. But they didn't spring to mind quickly. Although the group members knew of many studies, we couldn't distill their messages. Sometimes group members disagreed about the interpretation of single paper. Sometimes one paper's findings were in conflict with another paper's. Which one, if either, was correct? In many other cases we didn't think the work proposed any general findings. Our interpretation is that as a field, we've asked many questions and taken many measurements, but what we've learned is unclear. Thus, our first conclusion is that we need to review the many studies published in the literature to synthesize potential theories. 2.3 The Next Steps Given more time, we probably could have found more common results, but the difficulties we had were startling. One problem is that there hasn't been enough emphasis on synthesizing individual results into theories. But another, more fundamental, problem is that many studies aren't designed to produce general results. Consider the following kinds of empirical research. Feasibility studies. These studies are meant to validate new technology. Typically, the experimenter exercises his or her tool or method to show that it performs better than some other method or tool. These studies compare performances, but they rarely focus on the properties that make one tool better than another. Statistical Modeling. Many researchers have modeled the relationships between various software metrics and maintenance attributes, such as change effort, severity, and locality. Since these models fit data without understanding it (correlation vs. causality), there's a very little reason to believe that they will apply to other data sets. Observational studies. We also see many studies that document the behavior of a single project or organization. These studies may be useful as benchmarks of typical behavior, but they aren't intended to test any hypotheses. In fact, their authors rarely draw any actual conclusions. Each of these types of studies serves a purpose and can further our understanding of software engineering. However, they are not designed to and are not likely to produce general theories. Thus, our second conclusion is that theory building must be designed into our empirical research. 3. Future Challenges: In the previous sections we reviewed some of the subgroup's initial discussions. The group's activity took place in two sessions. In the first session, the participants presented their studies and discussed their goals, strengths, and weaknesses. During the second session we created a scheme for classifying maintenance studies, classified studies using this scheme, and then looked for common research findings in each class. We found some common findings, but not as many as we thought we should have. This led us to look more carefully at the field and ask why common results were so hard to find. Our main conclusion is that a critical part of scientific activity has been neglected in software maintenance research. That part is theory building - which can be done either by synthesizing individual results, or by proposing initial hypotheses and testing and refining them. The last issue we wrestled with was how to remedy this situation. In this section we discuss some of our recommendations for dealing with this problem. We divided these recommendations into three groups: rethinking the goal of empirical methods, supporting interdisciplinary research, and expanding the use of empirical methods. 3.1 Rethinking the Goals of Empirical Methods The quality of empirical research has improved tremendously in the last few decades. But it must continue to improve. In particular, we cannot forget that measurement is only one part of scientific inquiry. We routinely use measurement to describe, predict, and test. This is important, but, by itself, it does not give us the deep understanding we need to control software development. . To gain control over software development we need to have validated theories that are (1) general, (2) causal, and (3) suggestive of control strategies. General theories hold across several environments. Relationships that only hold in one environment are still important (for instance, for process improvement), but as scientists we should not be satisfied with them. Thus, we should strive to draw theories from our studies (even if they turn out to be wrong) so that they can be tested in other environments. Theories should also be causal. Although the literature describes many reasonably-good predictive models, they only capture correlations. We shouldn't confuse correlations with the underlying principles we really care about. One of the key challenges to developing causal theories is to focus more on discovering underlying principles and less on measuring high-level performance. Finally, if we have multiple candidate theories, we should prefer theories that suggest practical control strategies. . Another issue to consider is human variation. Differences in natural ability affect every empirical study. Some researchers are looking at ways to account for these differences, but much more work needs to be done. As we discuss later, other fields have this problem as well and may have some insight that will help us. . Although we are ultimately concerned with professional programmers who build industrial software, cost concerns lead us to use student subjects. Since we do not entirely understand the relationship between the student programmers and professionals, these studies are often discounted. Consequently, we need to develop models of the differences between student and professional populations. . Science must be a public activity. There have been calls to make data public, but we must also share artifacts, procedures, and terminology as well. Repositories and web sites should be set up and greater efforts should be made to conduct collaborative research. . Science is iterative. As we build and test theories, we will find almost all of them to be incorrect or imprecise. As authors and as reviewers, we will have to change how we think about and how we present our findings. 3.2 Supporting interdisciplinary research Too often software engineering researchers think that our field is entirely unique. Software development involves a web of individuals, groups and organizations working to build a complex array of products. Therefore, it's certain that other fields have tools and techniques that we will find useful. For example, . Behavioral scientists study how people work together (among other things). These have theories about how people learn and understand, and how they work together. One of the group members, Jarrett Rosenberg, coined the term "theory reuse" to describe this kind of interdisciplinary collaboration. . Cognitive scientists study how people think. Consequently, they have experience defining instruments to measure aspects of human skill. As we discussed earlier, such instruments could play a large role in factoring human variation out of our empirical studies. . Statistics already plays a big role in data analysis, experimental design, and hypothesis testing. We can also benefit from their knowledge of simulation methods, visualization methods, and mathematical modeling. . Business disciplines such as organization theory and information systems management have a great deal of knowledge about how process, organizational structure, and business strategies affect people's ability to get work done. Understanding these effects may help us to reconcile studies from multiple organizations. Also economists have considerable expertise in modeling complicated phenomena. 3.3 Expanding the Use of Empirical Methods Controlled experiments are the standard method for verifying theories. However, they are not perfect. They are expensive, have limited external validity, and it may be very difficult or, even unethical, to use them in certain situations. Therefore, in addition to traditional analysis techniques, we should consider other methods for generating and testing theories. . Qualitative analysis. Refers to the analysis of data that is represented in words and pictures rather than as numbers [16]. These approaches do not have as many supporting analyses as traditional quantitative methods do, but may provide a richer description of the phenomena being studied. . Meta-analysis. Gathering enough data to draw sound conclusions is a major problem in empirical research. One way to solve this problem may be to integrate the data and results of multiple studies. This can be done on an ad hoc basis, but there also approaches for statistically integrating the data. These approaches are called meta-analysis. . Mathematical modeling and simulation. This approach has been widely used in computer science, but not in software engineering. This is unfortunate because mathematical models can be powerful and much less expensive then experimenting with an actual system or process. They can also be very effective as a compliment to other approaches. . Case studies. Many case studies are simply retrospective descriptions of a project. Sometimes we call these "lessons learned" articles or "experience reports". We should try to do more case studies in which a hypotheses is stated and data is analyzed to see whether it is consistent with the hypothesis. . Surveys. Surveys can be a very inexpensive way to acquire lots of information at low cost. They have problems, but again, there is a large community of researchers that use them and understand their limitations. One final point to make here is that each of these approaches has strengths and weaknesses. Sometimes the best thing to do will be to combine two or more approaches. 4. Summary Because software vendors must respond frequently and correctly to changing user expectations, hardware platforms, and alternative products, software systems must tolerate numerous, successive changes. In practice, however, systems deteriorate and changes become increasingly difficult to implement. Thus it's easy to see why the costs of maintenance often dominate the costs of initial development. Research that makes maintenance easier will save considerable time and money throughout the software industry. Therefore, researchers have identified some sources of maintenance problems and developed potential remedies for it. For example, some researchers believe that repeated changes complicate a system's internal structure, so they have developed code metrics to model structural complexity. Others think that some software designs are inherently less flexible than others so they have focused on design patterns and software architecture. Still others argue that changes become more difficult because, over time, developers lose their understanding of the system They have developed reverse engineering tools, and formal documentation. The exact manner and degree in which different factors affect maintenance is unclear. It is clear, however, that this information is vital if our research is to have sustained, predictable improvement. Because many researchers share this belief, they attended a workshop on the topic of empirical studies of software maintenance. This article tries to summarize the ideas of the subgroup on Studies of the maintenance process: fundamental laws and assumptions. This group's main conclusion was that research community in software maintenance has produced many interesting results, but failed to consolidate them into useful theories. To help remedy this situation the group made three recommendations. 1. Researchers should design studies whose goal is to generate, test, and refine useful theories, not simply to describe behavior. This implies that we must use much more sophisticated empirical methods than we currently do. 2. Researchers should look to other disciplines for helpful insights, tools, and theories. However, they should borrow wisely. 3. Empirical research is severely limited by our inability to find adequately-sized samples. Researchers need to consider non-traditional techniques for combining data from multiple sources, and need to use multiple data collection and analysis approaches to make better use of small data sets. 5. List of participants The session chair would like to thank the subgroup participants for helping to make the workshop a success: Thomas, Helena Mendes-Moreira, Elisabeth Dusink, Warren Harrison, Jarrett Rosenberg, Giuseppe Visaggio, Marie Vans, Anneliese von Mayrhauser, Eirik Tryggeseth, Scott Schneberger. 6. --R A Model of Large Program Development. System Structure and Software Maintenance. Maintenance of Reuse-Based Domain-Specific Software Product Lines Using and Didactic Model to Measure Software Comprehension. Taxonomy of Educational Objects. An Empirical Exploration of Code Evolution. On Increasing our Knowledge of Large-Scale Software Comprehension The Software Maintenance Process in Financial Organizations. Problems and Prospects in Quantifying Software Maintainability. Position Paper for the International Workshop on Empirical Studies of Software Maintenance. The Impact of Documentation Availability on Software Maintenance Productivity. Assessing Maintenance Processes Through Controlled Experiment. On the Criteria for Decomposing Systems into Modules. Software Complexity and Software Maintenance: A Survey of Empirical Research. Methods in Qualitative Family Research --TR
theory development;software maintenance;empirical studies;interdisciplinary research
594421
Further Experiences with Scenarios and Checklists.
Software inspection is one of the best methods of verifying software documents. Software inspection is a complex process, with many possible variations, most of which have received little or no evaluation. This paper reports on the evaluation of one component of the inspection process, detection aids, specifically using Scenario or Checklist approaches. The evaluation is by subject-based experimentation, and is currently one of three independent experiments on the same hypothesis. The paper describes the experimental process, the resulting analysis of the experimental data, and attempts to compare the results in this experiment with the other experiments. This replication is broadly supportive of the results from the original experiment, namely, that the Scenario approach is superior to the Checklist approach; and that the meeting component of a software inspection is not an effective defect detection mechanism. This experiment also tentatively proposes additional relationships between general academic performance and individual inspection performance; and between meeting loss and group inspection performance.
Introduction Software inspection is a method for statically verifying documents. It was first described by Michael Fagan in 1976 [6]. Since then there have been many variations and experiences described, but a typical inspection involves a team of three to five people reviewing and understanding a document to find defects. The benefits of inspection are generally accepted, with success stories regularly published. In addition to Fagan's papers describing his experiences [6, 7], there are many other favourable reports. For example, Doolan [4] reports industrial experience indicating a times return on investment for every hour devoted to inspection of software requirement specifications. Russell [20] reports a similar return of 33 hours of maintenance saved for every hour of inspection invested. This benefit is derived from applying inspection early in the lifecycle. By inspecting products as early as possible, major defects are caught sooner and are not propagated through to the final product, where the cost of removal is far greater. One point of disagreement between inspection advocates is the role of detection aids. Many practitioners do not use any (The Ad Hoc approach), while others use Checklists to aid and regularise their process. Recently, Parnas and Weiss[16] have argued that even Checklists are not sufficiently focused to reliably aid inspectors, and have suggested that inspectors should be allocated specific responsibilities. Parnas argues that the Checklist approach results in an inspector repeatedly covering the same ground, while ignoring other ar- eas, resulting in reduced coverage. This idea has been further developed by 19], who has developed the Scenario approach in an attempt to address this problem. As well as developing the idea, Porter has attempted to show, via subject-based experimentation, that Scenarios are indeed superior to both Ad Hoc and Checklist approaches. To prove any hypothesis, results of experimental studies must be reproducible and hence, much to their credit, Porter et al. have made their experimental material available. This paper reports on our experiences and results in undertaking a partial replication of Porter's experiment. It will also draw upon experiences and results from another partial replication undertaken by Lanubile and Visaggio[13]. The paper will attempt to 'mirror' the organisation of Porter et al.[19] for ease of comparison. Hence the hypothesis of this replicated experiment is the same as the hypothesis in the original experiment: . that nonsystematic techniques with general reviewer responsibility and no reviewer coordination, lead to overlap and gaps, thereby lowering the overall inspection effectiveness; while systematic approaches with specific coordinated responsibilities reduce gaps, thereby increasing the overall effectiveness of the inspection. (Porter et 2 The experiment To further evaluate this hypothesis, the authors have replicated the original experiment. As with all subject-based experiments the design of Porter's experiment can be criticised, and hence a modified experimental design was undertaken to explore the hypothesis from a slightly different viewpoint. The alterations to the experiment will be discussed as they arise within the paper. As with the original experiment, the study uses students taking a formal university course 1 and commenced with a training phase in which the subjects were taught inspection techniques and the experimental procedures, during which they inspected two sample software requirements specification documents 2 ; this was followed by the experimental phase. 2.1 Experimental design This partial replication will focus on only 2 detection methods, Checklists (non- systematic) and Scenarios (systematic). This will simplify the design allowing greater focusing at the cost of being unable to generalise any results to include the Ad Hoc approach (nonsystematic), and hence our real hypothesis is less generic than the original experiment. 1 The original experiment used graduate students, this study used undergraduate students (3rd Year). 2 The original only used one sample document. All the documents inspected in all three experiments are software requirements specification documents. Variables The experiment manipulates three independent variables 3 : ffl The detection technique - Checklist or Scenario (Treatment Variable). ffl The specification to be inspected - two were used during the experiment (CRUISE, WLMS - see Section 2.2 for a brief description of these specifications ffl The subjects or groups of subjects - 50 subjects organised into 16 groups. The original experiment included another independent variable, the order in which the specifications were inspected. It was decided that it was unfeasible to allow variation in the order in which the specifications were inspected; this was because the inspections were part of an assessed class and hence the possibility of subject plagiarism corrupting the experimental data could not be discounted. The original experiment included this variation, but does not discuss how it avoids threats from plagiarism. The original experiment also measured four dependent variables: ffl The individual fault detection rate, ffl The team fault detection rate, ffl The percentage of faults first identified at the collection meeting (meeting gain rate), and ffl The percentage of faults first identified by an individual, but never reported at the collection meeting (meeting loss rate). The replication also measured the above variables, plus 3 additional variables ffl The time spent by each participant for the individual inspection, ffl The time spent by each group for the collection meeting, and ffl Average academic performance of each subject. Also the replication collected additional data from 4 debriefing questionnaires per specification - one individual and one group). Design Unfortunately the traditional fractional factorial design is unsuitable for this experiment because, as stated by Porter[19]: 3 The original experiment included an internal replication and thus included another independent variable. Round 1 Round 2 WLMS CRUISE WLMS CRUISE Ad Hoc 1B, 1D, 1G 1A, 1C, 1E 1A 1D, 2B 1H, 2A 1F, 2D Checklist 2B 2E, 2G 1E, 2D, 2G 1B, 1H Scenario 2C, 2F 2H 1F, 1C, 2E 1G, 2A, 2C Table 1: Subject allocation in the original experiment . they require some teams to use the Checklist method after they have used the Scenario method. Since the Checklist reviewers create their own fault detection techniques during the inspection (based on their understanding of the Checklist), our concern was that using the Scenario method in an early round might imperceptibly distort the use of the other methods in later rounds. Such influences would be undetectable because, unlike the Scenario method, the Checklist method does not require reviewers to perform specific, auditable tasks. Hence it was decided that a very straightforward design was the only alternative available. Each subject would participate in two inspections using the same technique. This design suffers from the problem that any ability effect between the two groups can distort any results applied to the treatment vari- able. Unfortunately, we have no direct way of measuring the subject's ability at software inspection, and hence it was decided that a random allocation of subjects to each group was the best alternative. In an attempt to estimate any effect, the average academic performance for each group was calculated. This indicated that the groups were extremely well balanced in terms of academic performance further post-analysis has revealed a strong linear association between average academic performance and average inspection performance (Pearson, hence it is believed that the experimental design has had little or no impact on the results affecting the treatment variable. The original experiment used a partial factorial design in which each group participated in two inspections, using some combination of the three detection methods, provided the above rule is not broken. Table 1 details the group allocation (A-H), where a 1 denotes the first run of the experiment and a 2 denotes the second run (or internal replication). Threats to internal validity All experiments suffer from the problem that some factor may affect the dependent variables without the experimenter's knowledge. This possibility must be minimised. In the original experiment Porter discusses five such threats: ffl Selection Effects The replication's approach is described in the previous section. The original experiment employed the following allocation strategy, in conjunction with the partial factorial design detailed above: . Create teams with equal skills. For example, rate each par- ticipant's background knowledge and experience as either low, medium, or high and then form teams of three by selecting one individual at random from each experience category. ffl Maturation Effects In the replication, each subject uses the same detection method through-out the experiment and the order of attempting the specifications is fixed, hence any learning effect should either be symmetrical across the two groups (i.e. derived from a common source, e.g. experimentation proce- dure) or directly related to the treatment variable (i.e. increased understanding of, or ability with, the detection method). The situation with the original experiment is more complicated, due to the asymmetrical allocation of subjects and the variation in the order of both applying the detection methods and inspecting the specifications, see Table 1. ffl Replication Effects Since this external replication does not contain an internal replication component, this potential threat is of no concern. ffl Instrumentation Effects This variation is impossible to avoid and is controlled by having each subject inspect both specifications. ffl Presentation Effects In the replication, the CRUISE specification is presented before the WLMS specification. Hence it is possible that the subjects can apply knowledge from their CRUISE inspection experience to their WLMS inspection. (See below for an explanation of why a fixed order of presentation was chosen - Plagiarism.) It is believed that any such effect should be symmetric between the groups and presents a far lesser risk than the risk of plagia- rism. The original experiment varies the order to control this effect, as described in Table 1. In considering the experiment and conducting initial trials with the experimental material, it was found that a larger list of internal threats exist than discussed in the original. ffl Plagiarism It was decided that it was unfeasible to allow variation in the order in which the specifications were inspected; this was because the inspections were part of an assessed class and hence the possibility of subject plagiarism corrupting the experimental data could not be discounted and was considered to be a greater risk than any asymmetric maturation effects. ffl Time Effects The original experiment allocated two hours for an individual inspection. It is believed that this fixed time period could have introduced a large bias into the results. The two specification documents are 24 and 31 pages long. Hence an inspection rate of 15.5 pages per hour is required for the longer document. This is well above the figure quoted in many inspection texts, e.g. Ebenau and Strauss[5] recommend 5 to 7 pages for specification documents, and Gilb and Graham[8] are even more conservative recommending 0.5 to 1.5 pages for software development documents. These figures by themselves do not show that the inspection rate in the original experiment was incorrect because of variations in the assumptions - the specifications are in a semi-formal notation, which the texts don't explicitly address and no precise definition of what constitutes a page are given. But they are a cause for concern. Also in their replication Lanubile and Visaggio[13] state that "the time limit was too short" and during our training of the subjects, it was found that the subjects required more than 2 hours to inspect a 16 page specification document. Hence to minimise this threat to validity, the original experiments strict time limit was abandoned and the subjects were given as much time as they required. Post-analysis shows that the average subject took 374 and 430 minutes to inspect the 24 and 31 page specifications respectively 4 . ffl Effort Effect The original experiment asked the subjects to either inspect the document using a Checklist or 2) inspect the document using one of the three Scenarios. These two tasks require different amounts of time and effort, the former requiring significantly more effort than the latter. This problem is compounded, as in the original, when this division of labour is used in conjunction with the fixed time limit. Asymmetric effort requirements in the application of the different detection methods leads to results parameterised by the percentage of task completed rather than a dimensionless quantity. This viewpoint is also supported by Cheng and Jeffery[2], who have carried out independent studies investigating the relationship between Scenarios and Checklists, and is consistent with the work of Kelly et al.[11], who found that increasing the amount to be inspected caused a decrease in inspector's performance. Hence to minimise this internal threat, the subjects using the Scenario technique were asked to apply all three Scenarios to balance the effort. Post-analysis shows that this alteration does balance the effort. A comparison of times taken can be found in Table 2. ffl Measuring multiple events The hypothesis of the original and this replication compares the various detection methods. The original experiment waited until after the group 4 Note: These times will have a small error associated with them. Given the length of the study, the subjects had to self-time parts of the study. 5 ignoring the Ad Hoc approach. Checklist Scenario Checklist Scenario Mean 460 403 379 370 St. Dev. 180 168 158 137 F ratio 1.34 0.04 Prob. Table 2: Average time taken to complete each inspection collection meeting to collect data for this comparison. Data collected at this point will be distorted by meeting gains and losses which are independent of the method used. Hence to minimise this threat, it was decided to collect data from before the collection meeting. ffl Statistical power All experiments should consider this most important design factor. Although statistical power is difficult to estimate, it is important to consider it, often in conjunction with effect size, in any well designed experiment. Given that the sample size was fixed (size of class undertaking the course) and assuming a standard setting for the power level (0.8), an experimenter can find out what size of effect must exist for their experiment to be considered well designed 6 . In the case of the original experiment, a large 7 effect size must be found, i.e. a large difference must be expected for the experiment to have an appropriate design. Unfortunately we do not know the estimated effect size of the original experiment, as it is not contained within the paper. Hence to minimise this threat, it was decided to reduce the required effect size by increasing the sample size. This was achieved in two ways: first increased subject numbers and second by measuring for the effect at the individual rather than the group level. This is especially appropriate given the decision on "Measuring multiple events". Hence with the increased sample size the replication can now reliably find significant results for small to medium 8 effect sizes. ffl Group dynamics Obviously the effectiveness of the group meeting component cannot be divorced from the dynamics of the group undertaking the inspections, and since the experiment has a repeated measure design, concern exists that the group dynamics will alter between the two inspection phases. Hence to minimise this threat, it was decided to have different group allocations for each phase, and therefore the group component resembles 6 The real situation is slightly more complicated than this, see Miller et al.[15] for a discussion on the impact of Statistical power on empirical research. 7 about 0.75, for a definition of large, medium and small power, see Cohen[3]. an experiment with an internal replication. There also exists some internal threats which we were unable to minimise fully. ffl Degree of application of the detection aids It is impossible to strictly enforce application of the detection aids all of the time. In fact, that situation is unnatural. Inspectors start their inspection by reading the document and obviously may well find defects during this initial preparation. Hence it is reasonable to assume that a proportion of the reported defects are not directly attributable to the application of the detection aid, even if the subject using it applies it fully. But which proportion? Should the defects for the initial read of the specification be discounted? The inspector may already be implicitly applying the detection aid while reading. Applying the detection aid to check a particular fact on a particular page leads the inspector to uncover an indirectly related fact, should this be discounted? What about an inspector applying an aid in a slightly non-standard fashion again leading to an error not directly attributable to the detection aid. Further imagine an inspector who chooses to augment their detection aided work with some additional Ad Hoc based inspection, during which they find errors directly related to the detection aids, how could these defects to reliably found and discounted? The authors believe that this is a very grey area with few distinguishing boundaries, and hence to attempt to subdivide the reported errors is highly dangerous and unadvisable, without large data sets to help to account for these potential variations. The experiment has relatively small data sets, especially with regard to defect coverage of individual Scenarios, e.g. as reported in the original experiment, the missing or ambiguous functionalities Scenario has a strict coverage of one defect in the CRUISE specification, which was not discovered by any subject, and hence all defects reported by application of this aid are at best indirect or secondary effects of applying the aid. Hence, rather than attempting to subdivide the recorded defects, the subjects will be asked to answer questions in the debriefing questionnaire about their level of application of the detection aid. Answers to these and other related questions should provide evidence for the validity of the undertaken experiment. ffl Typographical errors in specifications The specifications from the original experiment contained a significant number of typographical errors. Despite removing many of these errors, some still existed during the experiment and may have caused confusion for some of the subjects. Also these typographical errors preclude any analysis of the false positives reported. This is a major deficiency in the original and subsequent replications. ffl Initial Data Analysis Initial data analysis consists of the experimenter deciding on which reported defects are correct and which are incorrect. Defects are unstructured English descriptions of what the subjects' believe to be incorrect, and they are required to supply an accurate description of the fault. This is a inexact process and has potentially a large degree of variability. To illustrate, the experimenter checks possible defects against the master defect list (a list of all the defects contained within the document); Lanubile and Visaggio[13] state that they believe that this list from the original experiment is in error and provide several additional defects which they believe are missing from the original list, i.e. the researchers can't even agree on the perfect solution. It is concluded that this experiment has a natural in-built risk which is difficult, if not impossible, to safeguard against, unless the experimenter has the resources to employ multiple researchers cross-checking the initial analysis. ffl Learning Curve - Novice Subjects All the subjects had to learn both the inspection process and the specification language. It is quite possible that the experiment has experienced effects due to subjects lack of understanding of either of these topics. Apart from careful teaching of this material, it is difficult to employ further measures to safegaurd against this effect. Instead any effect will try to be 'measured' via the debriefing questionnaire and may subsequently alter the analysis of the independent variables. Threats to External Validity Threats to external validity limit the ability to generalise the results from an experiment to a wider population, under different conditions. The threats to external validity for the replication are the same as for the original experiment: ffl The subjects in our replication may not be representative of the general software engineering population, e.g. this study used students rather than software professionals. This threat is always a problem, because of the lack of sampling frame, and hence even studies using professionals will be exposed to this threat. ffl The specification documents may not be representative of industrial prob- lems. The documents used in this study are smaller and less complex than industrial specifications. ffl The inspection process may not be representative of industrial software development practice. Despite using a well known and widely used inspection technique[8], many other inspection processes exist in industry which pose a threat to the ability to generalise from this experiment. 2.2 Experiment Instrumentation This experiment has re-used, except as described above, the material from the original experiment, see Porter et al.[19] for a full description of the materials. The following is an abridged description of the materials to enhance the reader's appreciation of the experimental results. Software Requirements Specifications The specifications used in this experiment are: an automatic cruise control system[12] (CRUISE) and a water level monitoring system[22] (WLMS). The CRUISE specification is 31 pages long and the WLMS specification is 24 pages long. Each specification has four sections: overview, specific function require- ments, external interfaces, and a glossary. The overview is written in En- glish, whereas the other sections are specified using a tabular requirements All the faults in both documents are natural, with the CRUISE specification believed to contain 26 technical faults and the WLMS specification believed to contain 42 technical faults. Fault detection methods The two detection methods are designed to search for a well-defined population of faults. The original experiment used a general fault taxonomy to define the responsibilities of Ad Hoc reviewers and consequently directly derived a set of Checklist responsibilities from the Ad Hoc responsibilities. The individual Checklist questions were selected from several industrial checklists. Appendix A gives a complete description of the Checklist. The Scenarios were then derived from the Checklist by replacing individual Checklist items with procedures designed to implement them. The Scenario procedures have been further grouped into three distinct sub-groups: data type inconsistencies, incorrect functionalities, missing or ambiguous functionalities. Hence the Scenarios search for an exact subset of the faults defined by the Checklist's responsibili- ties. A post-experiment investigation by the original experimenters estimated that the Scenarios include about half of the faults covered by the Checklist. Appendix contains a complete list of all the Scenarios. 3 Data and Analysis Due to the alterations of the design of the experiment, the collected data cannot be analysed using the same strategy as the original experiment. Having said this, comparisons will be drawn whenever possible. 3.1 Verification of experimental validity Table 3 shows the average defect group detection rates for both the original, the first external replication 9 and this subsequent replication. This table shows that the subjects in this experiment performed at least as well as the subject's in the other experiments, and subsequently if we assume that the alterations in the experimental procedure have not increased the defect detection rate, this verifies these subjects suitability for this study. As stated earlier, it is believed that this group defect detection rate represents a composite result of using the detection method plus a collection meeting component. Hence rather than 9 Figures for this experiment are approximate, as Lanubile and Visaggio only give a graphical representation of this result. Checklist Scenario Checklist Scenario Original 0.24 0.45 0.41 0.57 1st. Rep. 0.24 0.20 0.32 0.33 This Rep. 0.39 0.42 0.48 0.49 Table 3: Average proportion of defects found in the three experiments, at the group level. 1. Do you think you understand Software Inspections? (a) Completely (b) Well (c) Reasonably Well (d) Not too sure (e) Not at all 2. Do you think you understand the notation? (a) Completely (b) Well (c) Reasonably Well (d) Not too sure (e) Not at all Figure 1: Questions regarding validity of experiment examining the above figures to analyse the treatment variable, the individual defect detection rates will be considered. Neither of the other experiments report these figures. This analysis is described in the following section. The experiment has one further source of data for analysis: the debriefing questionnaire. Each individual was asked two questions, one about their understanding of the inspection process and the other about the semi-formal specification language; Figure 1 details these questions and Table 4 summarises the results. The results are sufficiently positive, to verify these aspects of the experiment, i.e. the subjects have gained sufficient ability in these two areas to limit the impact of the learning curve. Further questions with a direct impact on assessing validity are discussed in future sections. It is believed that all these questions yield a sufficiently positive picture of the experimental procedures as to add further weight to the belief that no unknown external factor or factors have made the experiment invalid. 3.2 Individual Inspection Performance The following two figures (Figures 2 and the two specifications) inspection performance given a detection methodology. As can be seen, on average the Scenario technology out performs the Checklist Table 4: Summary of answers to questions regarding validity Normalised Defects Number of Subjects S. D.11 Figure 2: Histogram of average subject performance using Scenarios technology, but taking a closer look (Table 5), despite a noticeable difference in the means the experiment falls just short of providing a statistically significant result. Although not directly comparable, it is worth noting at this point, the treatment variable analysis from the other experiments: the original, and the first replication, Hence it is reasonable to conclude that the result from this experiment is more supportive of the original experiment than the first replication. We will return to the question of whose treatment variable results (at the group level) later in this paper. The only other variable to provide a significant result in either of the experiments was the specification variable. Unfortunately this experiment's specification variable could also hide a maturation effect, but given the experiences of the other two experiments and the fact that nothing has been altered which might directly impact on this effect, it is believed that this is unlikely. The specification variable shows a statistically significant effect, in line with the other Normalised Defects Number of Subjects S. D.11 Figure 3: Histogram of average subject performance using Checklists Detection Method Scenario Checklist Subjects 26 24 Mean 0.36 0.31 St. Dev. 0.11 0.11 St. F Ratio 2.80 F Prob. 0.10 Table 5: Basic detection method analysis two experiments (original first replication: Table 6. What is also of interest, is that this replication experienced an asymmetric effect with respect to detection method, again see Table 6, whereas the original experiment did not experience any 11 . These results suggest that a closer investigation is required and suggest that the treatment variable should be investigated against each specification sepa- rately. Table 7 shows this information. As can be seen, a large difference exists between the relative performance of the detection aids across the two specifications - with the first phase of the experiment (using the CRUISE specification) Remember these figures include results from using the Ad hoc approach and hence are not directly comparable. 11 The first replication does not quote any figure. Both Detection Method Methods Scenario Checklist Subjects 50 26 24 Mean -0.05 -0.01 -0.08 St. Dev. 0.11 0.12 0.12 St. F Ratio 7.81 4.85 Table Specification effect: independent and dependent of detection method; the negative means indicate a preference for the WLMS specification Scenario Checklist Scenario Checklist Subjects 26 24 26 24 Mean 0.35 0.27 0.37 0.35 St. Dev. 0.14 0.13 0.10 0.12 St. F Prob. 0.03 0.63 Table 7: Detection method analysis by specification showing a strong effect, while the second phase (using the WLMS specification) shows an extremely weak effect. The same analysis can be conducted for the original experiment, albeit at the group level. This analysis is shown in Table 8, and is derived from Table 3 (pp. 569) in the paper by Porter et al.[19] Again we see a strong performance bias towards one specification, further analysis is not possible given the reported data 12 . Hence further analysis was undertaken to attempt to explain this effect more fully. One possibility is an asymmetric change in the effort or commitment between the two groups. Obviously it is impossible to measure these effects di- rectly, and hence we must rely on the related measure of time spent when looking for an explanation. Table 9 summarises the time spent on individual inspections, it analyses the difference in time taken by each subject (CRUISE - WLMS), grouped by detection method. Note three subjects using the Scenario method failed to complete their time estimates, and are hence excluded from this analysis. This table shows that a small asymmetric effect does exist, but it is believed that it is too small to account for the effect. Taking the difference in means (31.3) and comparing it with the average time to complete (403), shows that the effect accounts for 7.5% of the average time spent on inspecting, and hence by itself is rejected as having sufficient likely impact to explain all of the difference. 12 The first experiment does not supply sufficient figures to allow a complementary analysis. Scenario Checklist Scenario Checklist Subjects Mean 0.45 0.24 0.57 0.41 St. Dev. 0.07 St. F Ratio 22.06 4.92 F Prob. !!0.01 0.06 Table 8: Detection method analysis by specification of the original experiment at the group level Scenario Checklist Subjects Mean 50.2 81.5 St. Dev. 152.6 162.7 St. F Prob. 0.50 Table 9: Difference in time spent between inspecting each specification Two further possible explanations were explored 1) that particular faults or particular types of faults are easier to find using one technique or the other; and differences in coverage, between the two detection aids. The original experiment states that in effect due to differences in coverage have been minimised by: " . deriving the Scenarios from the Checklist by replacing individual Checklist items with Scenario procedures designed to implement them." Unfortunately the experiment has only transposed a subset of the Checklist items into Scenario procedures and hence the possibility of a coverage effect cannot be ignored. For the CRUISE specification which has 26 potential defects, the Checklist items cover 24 of the defects, whereas the Scenario procedures only cover 14 (10 Data type inconsistencies, 1 Incorrect functionalities and 3 Missing or Ambiguous functionalities) defects; and for the WLMS specification which has 42 potential defects, the Checklist items cover 38 of the defects, whereas the Scenario procedures only cover 24 (14 Data type inconsistencies, 5 Incorrect functionalities and 5 Missing or Ambiguous functionalities) defects. These facts could be considered to add extra weight to the claim that the Scenario procedures are superior - increased performance, for lower coverage. This is probably true given the decision to give the subjects as much time as they required. It is more difficult to estimate the potential impact on the original with its fixed time limit. Mean Difference (Scenario Checklist)Defect Figure 4: Difference between technologies against individual defects in CRUISE-8Mean Difference (Scenario Checklist) Defect Figure 5: Difference between technologies against individual defects in WLMS The following figures (Figures 4 and 5) shows the difference in performance between the two technologies (Scenario - Checklist) against the individual de- fects. The first figure illustrates the behaviour in the CRUISE specification. As can be seen, subjects using the Scenario technique perform better than Check-list users for nearly every defect and by a relatively consistent amount, with little difference between the areas explicitly covered by the Scenario procedures and those which are not. Turning our attention to the second figure, detailing the results from the WLMS specification, we see that this consistent picture disappears, and that the Checklist users out-perform the Scenario users in 36% of the defects. Here many of the items covered by the Checklist but not by the Scenario procedures, show results in favour of the Checklist technique. But again there is no consistency in the interaction between the detection aids and the specifications. A further possibility that exists is that these defects fall into a single type of 3. Did you use your technique (Scenario or Checklist)? (a) Always (b) Most of the time (c) Sometimes (d) Occasionally Never 4. Do you understand your technique? (a) Completely (b) Well (c) Sort of (d) Not to sure (e) Not at all 5. Do you think your technique was better or worse than the Ad Hoc approach previously employed? (a) Definitely Better (b) Probably Better (c) About the same (d) Probably Worse Figure Questions regarding technique used. defect. To investigate this, the defects were characterised using the taxonomy developed in the original experiment, which is a composite of two schemes developed by Schneider et al.[21] and Basili and Weiss[1]. Unfortunately this classification failed to reveal any insights into the effects, as the defects are spread across the taxonomy. Finally, the questionnaires were investigated to address this issue. The questionnaires contain three questions directly relevant to this point (Figure 6). Question 3 attempts to find the rate of use of each technique; Question 4 how well each technique is understood by the relevant subjects; and Question 5 elicits the subject's opinion on each techniques performance against a common control technique. Table 10 summarises the subjects responses. Analysing table yields two relevant facts. Firstly the more regular use of the Scenario technique, compared with the Checklist technique; in fact the latter might be more correctly named Checklist plus Ad Hoc. It is impossible to estimate the impact this would have on the previous calculations, but any impact should be symmetrical with regard to the specifications. The second observation is the drop in usage of the Scenario technique between inspecting the CRUISE and WLMS specifications, while the Checklist technique does not experience any decline in application. This could well lead to an asymmetric effect, especially if the Scenario technique is superior to simply adopting an Ad Hoc approach to inspection. To summarise the asymmetric effect between the two specifications is difficult to explain. A number of factors have been found each of which could have Q. 3 Q. 4 Q. 5 Q. 3 Q. 4 Q. 5 Checklist C% Table 10: Detection method analysis by questionnaire (Figure 6), all percentages rounded to 5%. Scenario Checklist Scenario Checklist No. of groups 8 8 8 8 Mean 0.42 0.39 0.49 0.48 St. Dev. 0.08 0.09 0.07 0.11 St. F Prob. 0.49 0.75 Table 11: Initial analysis of group performance a small asymmetric impact on the experiment. The most likely explanation of the asymmetric effect is a combination of these factors, namely: ffl Asymmetric effort effect ffl Differences in coverage ffl Decline in application of the Scenario technique ffl Natural variation within the experiment ffl The performance of the detection aids is dependent on the nature of the specification 3.3 Analysis of Group Meeting Data Earlier it is stated that this experiment deviated from the original design because of concern of multiple events obscuring the basic hypothesis. Table 11 investigates the original hypothesis at the group level. Comparing this table with the same analysis at the individual level (Tables 5 and 7) clearly shows Scenario Checklist Scenario Checklist No. of groups 8 8 8 8 Mean 0.62 0.48 0.64 0.59 St. Dev. 0.09 0.11 0.08 0.09 St. F Ratio 7.46 1.14 F Prob. 0.02 0.30 Table 12: Analysis of Potential maximum group results Scenario Checklist Scenario Checklist No. of groups 8 8 8 8 Mean -5.13 -2.38 -6.00 -4.75 St. Dev. 2.30 3.46 4.20 3.69 St. F Prob. 0.29 0.60 Table 13: Analysis of the relationship between meeting effects and detection aid. Figures are quoted as per defect, rather than as percentages, due to the small numbers involved. that the group component has had a large effect upon the analysis of the treatment variable, especially with regard to analysing the results of the CRUISE specification. Recalling the results for the original experiment (p ! 0.01) and the first replication (p ! 0.92), the results at the group level are now more supportive of the first replication rather than the original experiment. This change of support is directly attributable to (some aspects of) the group inspection. This can be clearly seen by comparing Table 11 with Table 12, which estimates the group score by forming a union of the three individual scores. In Table 12, the familiar pattern from the individual analysis re-establishes itself, if anything the effect (pro-Scenario) is even stronger than at the individual level. So why the change? It is believed that the meeting component is introducing an effect unrelated to the detection aids, and this effect is demonstrated in the difference between Tables 11 and 12. The difference can be characterised by two variables: meeting loss and meeting gain. Meeting gain represents new defects found at the meeting, whereas meeting loss represents those defects found during the individual phase which are 'lost' during the meeting session. (Meeting loss occurs for a number of reasons which will be discussed later in the paper.) Table 13 explores this possible relationship further, directly exploring the possibility of a relationship between the meeting component (meeting gain - meeting loss) and the detection aid. It shows that no significant result exists between these two concepts and the larger negative results experienced by the Meeting Gain Meeting Loss Meeting Gain Meeting Loss No. of groups Total Defects 26 26 42 42 Mean 1.13 4.88 1.63 7.00 St. Dev. 0.81 2.80 1.20 3.27 Table 14: Meeting Gain and Loss by Specification Score Maximum Loss p 0.14 !!0.01 WLMS r -0.51 0.52 Loss Table 15: Correlation (Pearson) between Meeting Loss, Group Score and Maximum Possible Group Score Scenario subjects is believed to be directly related to the fact that these subjects, on average, found more defects than any indirect association with the detection aid. This conjuncture is discussed more fully in the following section and Table 15 shows the correlation between the meeting effects and the number of defects found. Examining Tables 11-13 and 15 shows that a strong case can be made that the meeting component is independent of the detection aid used, and hence this experiments choice to explore the treatment variable at the individual level rather than at the group level, as conducted in the original, is the more appropriate experimental design. 3.4 Analysis of Meeting Losses and Gains Another outcome of the original experiment was the rejection of the meeting component as a fault detection technique. The original experiment reports that the number of meeting losses outweighs the number of meeting gains, this viewpoint is also supported by the first replication. The results here are no different, with meeting losses comfortably outweighing meeting gains. Given the low resolution in these values, complex statistical evaluation is inappropri- ate, for example the median value of the meeting gain is one defect for both specifications, i.e. one unit of resolution. Table 14 gives a brief summary of meeting gains and meeting losses for each specification. The figures in this table broadly agree with the original and the first replication, that on average meeting losses are greater than meeting gains, and are also in line with the results reported by McCarthy et al.[14], who reported that inspection techniques which rely more on the individual component were more productive than those which relied more upon the meeting component. Further, with regard to the meeting gain component, these results are in line with the results reported by Votta[10]. Votta reported on his experiences observing a series of professional inspection meetings and found that the meeting component only contributed 4%, on av- erage, to the number of defects found. This experiment produced an average meeting gain of 9% from our novice inspectors, i.e. over twice the professional rate. It is clearly highly unlikely that our subjects are able to perform at twice the level of seasoned professionals. Much more likely, and certainly the view of the authors, is that this increase is due to an experimental process, and shows the danger of extrapolating from such small scale, low resolution data. The meeting loss figures broadly follow the trend of the other experiments, but at a more extreme level, i.e. this experiment has larger meeting loss values than the other two experiments. So why the large meeting loss? Table 15 shows that the loss is a systematic effect, and that the meeting loss correlates with the meeting score, and correlates extremely strongly with the potential maximum group score, i.e. the greater the diversity of the defects discovered by the individuals in the group, the greater the number of defects lost. One aspect of the experimental material which may be influencing this is the typographical errors contained within the specification. The subjects were told these errors were not defects and to discard them. The typographical errors don't cause any problem with the correctness of the specifications, but they do contribute an additional difficultly in understanding the document. Given that the experiment uses novice inspectors, it is believed that many groups had some difficulty distinguishing between defects and typographical errors, leading to some genuine defects being discarded at the meeting. Unfortunately it is impossible to measure the impact of the typographical errors on the meeting component, and even on the individual component, but it is believed that the greatest potential impact will be on the meeting loss estimation, especially given the lower sample size at the group level. This numerical concern of the effectiveness of the meeting component is not shared by the subjects. The subjects were asked if they thought that the meeting component was worthwhile, via the questions in Figure 7. In general, their responses must be categorised as positive (Table 16). Looking in more detail we can see that the responses to the second questionnaire are significantly less positive than the first. That is, there exists a large negative trend from the first exercise to the second exercise in the subjects opinions about the effectiveness of the meeting component as a vehicle for defect detec- tion. This trend, rather than the individual responses, is more in line with the quantifiable data from the inspection meeting- unfortunately there is no way of predicting if this trend would continue into subsequent phases of such an experiment. Conclusions Software inspection are undoubtedly one of the best ways to verify software documents. Having said this, the technique remains relatively unexplored, in terms of which variations upon the basic theme work best. This paper explores one such variation - detection aids, specifically it attempts to compare Scenarios 6. In considering the effective use of an inspector's time, how would you rate the collection meeting against the individual component? A. Far superior B. Useful C. Of Similar Worth D. Of limited use E. A waste of time 7. Which (you may answer yes to more than one) of the following objectives were achieved by your collection meeting? A. Defect Merging B. Extra Defect Detection C. Group Bonding/Team Spirit D. Education of Weak group Members E. Ensure all individuals adequately prepare their defect lists F. Ensure common practices amongst individual inspectors Figure 7: Questions regarding technique used. and Checklists, to see which approach is better, if any. Importantly this paper is not a one-off study, but is part of a large piece of work involving several other researchers investigating the same hypothesis. Multiple independent studies of the same hypothesis are essential if software engineering is going to produce empirically evaluated theories and procedures. The paper attempts to compare its results with the other studies whenever possible. With regard to the hypothesis of the experiment, the results are ambiguous, but on balance are generally supportive of the results in the original experiment. Hence, given the current weight of evidence, from the three experiments, there seems to be emerging support for the conjuncture that the Scenario approach is better than the Checklist approach. More work is required to finally confirm this conjecture. As normal when conducting an experiment several other effects, or more accurately possible effects, were revealed. These additional effects include: ffl a strong correlation between general academic ability and the ability to successfully inspect software. ffl an interaction between the detection aids and the specifications. ffl that the subjects were more willing to use Scenarios compared with Check-list ffl that the effectiveness of the meeting component is independent of the detection aid used. ffl that meeting losses outweigh meeting gains, suggesting the meeting component is not an effective defect detection mechanism. Q. 6 Q. 7 Q. 6 Q. 7 A% B% C% Table Analysis of Questions 6 and 7 (Figure 7) ffl that meeting losses correlate strongly with various measures of the number of defects found by a group. These new potential effects also require further investigation to verify their validity. Looking ahead and assuming that they exist, then the above results suggested several new lines of research within the software inspection area, e.g. inspection models with alternatives to the meeting component. A Checklist Method ffl General - Are the goals of the system defined? - Are the requirements clear and unambiguous? - Is a functional overview of the system provided? - Is an overview of the operational modes provided? - Have the software and hardware environments been specified? - If assumptions that affect implementation have been made, are they stated? - Have the requirements been stated in terms of input, output, and processing for each function? - Are all functions, devices, constraints traced to requirements and vice versa? - Are the required attributes, assumptions and constraints of the system completely listed? ffl Omission Missing Functionality * Are the described functions sufficient to meet the system objectives * Are all inputs to a function sufficient to perform the required * Are undesired events considered and their required responses specified * Are the initial and special states considered (e.g. system initia- tion, abnormal termination)? Missing Performance * Can the system be tested, demonstrated, analyzed, or inspected to show that it satisfies the requirements? * Have the data type, rate, units, accuracy, resolution, limits, range and critical values for all internal data items been specified? * Have the accuracy, precision, range, type, rate, units, frequency, and volume of inputs and outputs been specified for each function Missing Interface * Are the inputs and outputs for all interfaces sufficient? * Are the interface requirements between hardware, software, per- sonnel, and procedures included? Missing Environment * Have the functionality of hardware or software interacting with the system been properly specified? ffl Comission Ambiguous Information * Are the individual requirements stated so that they are discrete, unambiguous, and testable? * Are the transitions specified deterministicly? Inconsistent Information * Are the requirements mutually consistent? * Are the functional requirements consistent with the overview? * Are the functional requirements consistent with the actual operating environment? - Incorrect or Extra Functionality * Are all the described functions necessary to meet the system objectives * Are all inputs to a function necessary to perform the required * Are the inputs and outputs for all interfaces necessary? * Are all the outputs produced by a function used by another function or transferred across an external interface? Wrong Section * Are all the requirements, interfaces, constraints, etc. listed in the appropriate sections. B.1 Data Type Consistency Scenario 1. data object's mentioned in the overview (e.g. hardware com- ponents, application variable, abbreviated term or function) (a) Are all data objects mentioned in the overview listed in the external interface section? 2. For each data object appearing in the external interface section determine the following information: ffl Object name: ffl Class: (e.g. input port, output port, application variable, abbreviated ffl Data type: (e.g. integer, time, boolean, enumeration) ffl Acceptable values: Are there any constraints, ranges, limits for the values of this object ffl Failure value: Does the object have a special failure value? ffl Units or rates ffl Initial value: (a) Is the object's specification consistent with its description in the overview? (b) If object represents a physical quantity, are its units properly specified (c) If the object's value is computed, can that computation generate a non-acceptable value? 3. For each functional requirement identify all data object references: (a) Do all data object references obey formatting conventions? (b) Are all data objects referenced in this requirement listed in the input or output sections? (c) Can any data object use be inconsistent with the data object's type, acceptable values, failure values, etc.? (d) Can any data object definition with the data object's type, acceptable values, failure value, etc.? B.2 Incorrect Functionality Scenario 1. For each functional requirement identify all input/output data objects: (a) Are all values written to each output data object consistent with its intended function? (b) Identify at least one function that uses each output data object. 2. For each functional requirement identify all specified system events: (a) Is the specification of these events consistent with their intended 3. Develop an invariant for each system mode (i.e. Under what conditions must the system exit or remain in a given mode)? (a) Can the system's initial conditions fail to satisfy the initial mode's invariant? (b) Identify a sequence of events that allows the system to enter a mode without satisfying the mode's invariant. (c) Identify a sequence of events that allows the system to enter a mode, but never leave (deadlock). B.3 Ambiguities Or Missing Functionality Scenario 1. Identify the required precision, response time, etc. for each functional requirement. (a) Are all required precisions indicated? 2. For each requirement, identify all monitored events. (a) Does a sequence of events exist for which multiple output values can be computed? (b) Does a sequence of events exist for which no output value will be 3. For each system mode, identify all monitored events. (a) Does a sequence of events exist for which transitions into two or more modes is allowed? --R Evaluation of software requirements document by analysis of change data. Comparing inspection strategies for software requirement specifications. Statistical Power Analysis for the Behavioral Sciences. Experience with Fagan's inspection method. Software Inspection Process. Design and code inspections to reduce errors in program development. Advances in software inspection. Software Inspection. Specifying software requirements for complex systems. Votta Jr. An analysis of defect densities found during software inspections. Example NRL/SCR software requirements for an automobile cruise control and monitoring system. Assessing defect detection methods for software requirements inspections through external replication. An experiment to assess cost-benefits of inspection meetings and their alternatives Statistical power and its subcomponents - missing and misunderstood concepts in emprical software enginerring research Active design reviews: Principles and practices. An experiment to assess different defect detection methods for software requirements inspections. Comparing detection methods for software requirements inspections: A replicated experiment. Comparing detection methods for software requirements inspections: a replicated experiment. Experience with inspection in ultralarge-scale develop- ments An experimental study of fault detection in user requirements. --TR --CTR Stefan Biffl , Bernd Freimut , Oliver Laitenberger, Investigating the cost-effectiveness of reinspections in software development, Proceedings of the 23rd International Conference on Software Engineering, p.155-164, May 12-19, 2001, Toronto, Ontario, Canada J. Miller, On the independence of software inspectors, Journal of Systems and Software, v.60 n.1, p.5-10, 15 January 2002 Stefan Biffl , Wilfried Grossmann, Evaluating the accuracy of defect estimation models based on inspection data from two inspection cycles, Proceedings of the 23rd International Conference on Software Engineering, p.145-154, May 12-19, 2001, Toronto, Ontario, Canada Bente Anda , Dag I. K. Sjberg, Towards an inspection technique for use case models, Proceedings of the 14th international conference on Software engineering and knowledge engineering, July 15-19, 2002, Ischia, Italy Bjrn Regnell , Per Runeson , Thomas Thelin, Are the Perspectives Really Different? FurtherExperimentation on Scenario-Based Reading of Requirements, Empirical Software Engineering, v.5 n.4, p.331-356, December 2000 Hidetake Uwano , Masahide Nakamura , Akito Monden , Ken-ichi Matsumoto, Analyzing individual performance of source code review using reviewers' eye movement, Proceedings of the 2006 symposium on Eye tracking research & applications, March 27-29, 2006, San Diego, California Stefan Biffl, Using Inspection Data for Defect Estimation, IEEE Software, v.17 n.6, p.36-43, November 2000 Stefan Biffl , Michael Halling, Investigating the Defect Detection Effectiveness and Cost Benefit of Nominal Inspection Teams, IEEE Transactions on Software Engineering, v.29 n.5, p.385-397, May Stefan Biffl , Walter J. Gutjahr, Using a Reliability Growth Model to Control Software Inspection, Empirical Software Engineering, v.7 n.3, p.257-284, September 2002 Oliver Laitenberger , Thomas Beil , Thilo Schwinn, An Industrial Case Study to Examine a Non-Traditional Inspection Implementation for Requirements Specifications, Empirical Software Engineering, v.7 n.4, p.345-374, December 2002 Susan S. Brilliant , John C. Knight, Empirical research in software engineering: a workshop, ACM SIGSOFT Software Engineering Notes, v.24 n.3, p.44-52, May 1999 F. MacDonald , J. Miller, A Comparison of Tool-Based and Paper-Based Software Inspection, Empirical Software Engineering, v.3 n.3, p.233-253, September 1998 James Miller , Fraser Macdonald , John Ferguson, ASSISTing Management Decisions in the Software Inspection Process, Information Technology and Management, v.3 n.1-2, p.67-83, January 2002 Thomas Thelin , Per Runeson , Claes Wohlin, An Experimental Comparison of Usage-Based and Checklist-Based Reading, IEEE Transactions on Software Engineering, v.29 n.8, p.687-704, August Victor R. Basili , Forrest Shull , Filippo Lanubile, Building Knowledge through Families of Experiments, IEEE Transactions on Software Engineering, v.25 n.4, p.456-473, July 1999 Andreas Zendler, A Preliminary Software Engineering Theory as Investigated by Published Experiments, Empirical Software Engineering, v.6 n.2, p.161-180, June 2001 Oliver Laitenberger , Khaled El Emam , Thomas G. Harbich, An Internally Replicated Quasi-Experimental Comparison of Checklist and Perspective-Based Reading of Code Documents, IEEE Transactions on Software Engineering, v.27 n.5, p.387-421, May 2001 Oliver Laitenberger , Dieter Rombach, (Quasi-)experimental studies in industrial settings, Lecture notes on empirical software engineering, World Scientific Publishing Co., Inc., River Edge, NJ,
partial replication;experiments;software inspection;defect detection aids
594454
A Function Point-Like Measure for Object-Oriented Software.
We present a method for estimating the size, and consequently effort and duration, of object oriented software development projects. Different estimates may be made in different phases of the development process, according to the available information. We define an adaptation of traditional function points, called Object Oriented Function Points, to enable the measurement of object oriented analysis and design specifications. Tools have been constructed to automate the counting method. The novel aspect of our method is its flexibility. An organization can experiment with different counting policies, to find the most accurate predictors of size, effort, etc. in its environment. The method and preliminary results of its application in an industrial environment are presented and discussed.
Introduction Cost and effort estimation is an important aspect of the management of software development projects. Experience shows that accurate estimation is difficult: an average error of 100% may be considered "good" and an average error of 32% "outstanding" (Vicinanza, Mukhopadhyay and Prietula 1991). Most methods for estimating effort require an estimate of the size of the software. Once a size estimate is available, models can be used that relate size to effort. Cost estimation is not a one-time activity at project initiation. Estimates should be refined continually throughout a project (DeMarco 1982). Thus, it is necessary to estimate size repeatedly throughout development. 2 Giuliano Antoniol et al. Accurate estimation of size is vital. Unfortunately it has proved to be very difficult, especially early in development when the estimates are of most use. Most research on estimating size has dealt with traditional applications and traditional software development practices. Few methods have been proposed for object oriented software development. This paper presents a method for estimating the size of object oriented software development projects. It is based on an adaptation of the classical Function Point method to object oriented software. In the following sections, we present our approach to measurement. We explain how we map the concepts of function points to object oriented software. We describe the process of counting Object Oriented Function Points (OOFPs), and give an example. Results from a pilot study are presented and discussed. 2. Measurement Perspective Figure 1 shows the main phases of an object oriented (OO) development process, and measurements that can be obtained at different points in development. The purpose of these measures is to give the project manager something from which to estimate the size, effort, and duration of a project. These estimates can be revised as new artifacts become available during development. OO Analysis OO Design System Requirements Definition Implementation USER DESIGNER CUSTOMER/ Figure 1. Perspectives and measures in the software development process. As we move through the phases of the process, the measurement perspective changes from that of the user to the designer. At the end of the requirement specification phase, the classical Function Point counting method is applied on the requirements doc- ument. The function point method takes the perspective of the end user. What is actually measured are the functions of the system that main.tex; 23/06/1999; 14:27; no v.; p.2 A Function Point-like Measure for Object-Oriented Software 3 are visible to the end user. This measure is generally considered to be independent of any particular implementation. Some aspects of a system (e.g. a graphical user interface) are not included in the classical function point count. Nevertheless, they contribute to the final size of the system. If the objective of measuring functionality is to estimate the size of an implementation of a sys- tem, and from that the effort and duration of a software project, these aspects should be taken into account. This changes the perspective from that of the user to that of the customer i.e. the organization that acquires the system, accepts it and pays the development costs; and of the system designer, who has to produce an implementation of the given specifications. Once object oriented modelling begins (i.e. from the OO analysis phase on), measurements can be obtained from the object models; OOFP are used in place of FP. As development proceeds, this gives a natural shift of perspective. In the OO analysis phase, most of the elements in the object models are still related to the application domain, so the perspective is still that of the user. At the OO design and later phases, the object models reflect implementation choices. This includes aspects of the system that are not specified in the requirements documents. The count of OOFP on these models will thus include such functionalities. The measurement perspective is now that of the designer. Different size estimation models can be developed for different stages of development. More detailed information is available, as the system evolves from an abstract specification to a concrete implementation. It should be possible to refine a size estimate repeatedly, as long as the estimation process is not too difficult. Since we have developed tools to automate the counting of OOFPs, re-calculation is easy at any time. 3. Object Oriented Function Points Practitioners have found function points (Albrecht and Gaffney 1983, IFPUG 1994) to be very useful within the data processing domain, for which they were invented. We aim to exploit the experience that has been obtained with function points in traditional software development, in the OO paradigm. In adapting function points to OO, we need to map function point concepts to object oriented concepts, and must decide how to handle OO-specific concepts such as inheritance and polymorphism. 4 Giuliano Antoniol et al. Our presentation uses notations from the OMT method (Rumbaugh, Blaha, Premerlani, Eddy and Lorensen 1991). It would not be much different if the Booch notation (Booch 1991) or Unified Modeling Language (Rational Software Corporation 1997b) was used instead, since the main models and views in the different methodologies carry similar information. The OMT method uses three different, orthogonal views to describe a software system: Object Model: a static representation of the classes and objects in a system, and their relationships. Functional Model: data flow diagrams provide a functional decomposition of the activities of the system. Dynamic Model: state machines represent the dynamic and control aspects of a system. Although all three models provide important information about an object-oriented system, the object model is the most important for our purposes. It is usually the first to be developed, and so can be measured earliest. It is the one that represents what is actually to be built. In a sense the other models help in completing the object model: the functional model helps in identifying and designing some of the methods; the control model helps in identifying attributes that are needed to maintain state information, and events that must be implemented as classes or methods. There is, however, an ongoing discussion in the practitioners community on the content and role of those models. The functional model seems to have fallen into disuse and is not required any more by some methodologies (Rational Software Corporation 1997b). The dynamic model is often replaced with a list of use cases and scenarios. The object model is the only one that is present in all methodologies and describes the system using specifically object-oriented concepts. For these reasons, we decided to restrict our attention to object models. In traditional developments, the central concepts used in counting function points are logical files and transactions that operate on those files. In OO systems, the core concept is no longer related to files or data bases; instead the central concept is the "object". The central analogy used to map function points to OO software relates logical files and transactions to classes and their methods. A logical file in the function point approach is a collection of related user- identifiable data; a class in an object model encapsulates a collection of data items. A class is the natural candidate for mapping logical files into main.tex; 23/06/1999; 14:27; no v.; p.4 A Function Point-like Measure for Object-Oriented Software 5 the OO paradigm. Objects that are instances of a class in the OO world correspond to records of a logical file in data processing applications. In the FP method, logical files (LF) are divided into internal logical files (ILFs) and external interface files (EIFs). Internal files are those logical files that are maintained by the application; external files are those referenced by the application but maintained by other applica- tions. This division clearly identifies the application boundary. In the OO counterpart, the application boundary is an imaginary line in an object model, which divides the classes belonging to the application from the classes outside the application. External classes encapsulate non-system components, such as other applications, external services, and reused library classes (both directly instantiated and subclassed and parameterized classes). Classes within the application boundary correspond to ILFs. Classes outside the application boundary correspond to EIFs. Transactions in the FP method are classified as inputs, outputs and inquiries. This categorization is not easily applicable outside the data processing domain. In the OO paradigm the locus of operation are class methods, which are usually at a more fine-grained level than transactions. Since object models rarely contain the information needed to tell whether a method performs an input, an output or is dealing with an enquiry, we do not attempt to distinguish the three categories. We simply treat them as generic Service Requests (SRs), issued by objects to other objects to delegate to them some operations. In short, we map logical files to classes, and transactions to methods. Issues such as inheritance and polymorphism affect the structure of the object model, and how the model should be counted. They are addressed in Section 4. 3.1. Related Work Other authors have proposed methods for adapting function points to object oriented software. They too generally map classes to files, and services or messages to transactions. Whitmire (1993) considers each class as an internal file. Messages sent across the system boundary are treated as transactions. Schoon- eveldt, Hastings, Mocek and Fountain (1995) treat classes as files, and consider services delivered by objects to clients as transactions. This method gives a similar count to traditional function points for one system. A draft proposal by the International Function Point Users Group ("IFPUG") treats classes as files, and methods as transactions 1995). 6 Giuliano Antoniol et al. Fetcke, Abran and Nguyen (1997) define rules for mapping a use case model (Jacobson, Christerson, Jonsson and Overgaard 1992) to concepts from the IFPUG Counting Practices Manual (IFPUG 1994). Three case studies have confirmed that the rules can be applied consis- tently. No attempt has been made to relate the results to other metrics, such as traditional function points, lines of code, or effort. Sneed (1995) proposed object points as a measure of size for OO soft- ware. Object points are derived from the class structures, the messages and the processes or use cases, weighted by complexity adjustment factors The closest analogue to our method is Predictive Object Points (POPs), proposed by Minkiewicz (1997). POPs are based on counts of classes and weighted methods per class, with adjustments for the average depth of the inheritance tree and the average number of children per class. Methods are weighted by considering their type (con- structor, destructor, modifier, selector, iterator) and complexity (low, average, high), giving a number of POPs in a way analogous to traditional FPs. POPs have been incorporated into a commercial tool for project estimation. Our work differs from Minkiewicz in several ways. In two respects we consider more information: we count the data in a class as well as the methods; and we consider aggregation and inheritance in detail, instead of as averages. We consider less information when counting methods, since we do not distinguish between method types. Information about method type is seldom available at the design stage. Automatic tools would need to gather that information from the designer, which might be a tedious task for very large systems. For that reason we do not attempt to base our method complexity weighting on method type; instead we try to exploit information about a method's signature, which is most likely to be present in a design, at least at the detailed design stage. The key thing which is new about our method is its flexibility, with much scope for experimentation. For example, Fetcke et al. (1997) define that aggregation and inheritance should be handled in a particular way. As discussed below in Section 4.1, we define several options (one of which is Fetcke's approach) and leave it to the user to choose. We have written programs to count OOFPs automatically, with several parameters to govern counting decisions. An organization can experiment to determine which parameter settings produce the most accurate predictors of size in its environment. Thus we have a method which can be tailored to different organizations or environments. Moreover, the measurement is not affected by subjective ratings of complexity factors, like those introduced in classical function point analysis. A Function Point-like Measure for Object-Oriented Software 7 4. Measurement Process OOFPs are assumed to be a function of the objects in a given object model D. D might be produced at the design stage, or extracted from the source code. OOFPs can be calculated as: where: WELF (DET A denotes the set of objects belonging to the application, and o is a generic object in D. DETs, RETs and FTRs are elementary measures, calculated on LFs and SRs and used to determine their complexity through the complexity matrices W . Details are given in Sections 4.1- 4.3. Figure 2 shows the phases of the OOFP computation process: 1. The object model is analyzed to identify the units that are to be counted as logical files. There are four ways in which this might be done; which to use is a parameter of the counting process. This step is described in Section 4.1. 2. The complexity of each logical file and service request is deter- mined. W tables are used to map counts of structural items (DETs, RETs and FTRs) to complexity levels of low, average, or high. These tables can be varied, and represent another parameter of the counting process. This step is described in Sections 4.2 and 4.3. 3. The complexity values are translated to numbers, using another table. These numbers are the OOFP values for the individual logical files. The table used here can also be varied, and so is yet another parameter of the counting process. 8 Giuliano Antoniol et al. 4. If a logical file is a class which is annotated as "reused" (ie developed by reuse of another class), its OOFP value is multiplied by a scale factor (1.0 or less). The scale factor is another parameter of the counting process. This step is discussed in Section 4.4. 5. The OOFP values are summed to produce the total OOFP value. Det Ret Det Ftr Det Ret OOFP OOFP OOFP ILF L,A,H L,A,H L,A,H GB Module OO Design Figure 2. OOFP computation process. 4.1. Identifying logical files Conceptually, classes are mapped to logical files. It may not always be appropriate to count each class simply as a single logical file, however. Relationships between classes (aggregations and generalization / specializations in particular) can sometimes make it appropriate to count a group of classes together as a logical file. Aggregation and inheritance relationships pertain mostly to implementation aspects (internal organization, reuse). There tend to be few of them in an analysis object model. There may be many of them in a design or implementation model, as whole-part assemblies and inheritance hierarchies are identified. How these relationships affect the boundaries around logical files depends on the perspective chosen, and the artifact on which the OOF- Ps are computed. At the analysis phase, the user's perspective is the important one. It is too early to take the designer's point of view. At this stage, most of the classes in an object model represent entities in the application and user domain. There are few aggregation and inheritance relationships to complicate things. Counting each single class as a logical file is usually appropriate. At this stage, the origin of a class does not matter. The scale factor used in step 4 of the counting process shouild be set to 1.0, so the class is counted with its full inherent value. A Function Point-like Measure for Object-Oriented Software 9 At the design phase, the object models contain much more information related to the implementation. From a designer's perspective, considering each single class as a logical file will again be the correct choice. From a designer's or implementer's point of view, reuse makes classes easier to develop. If the OOFP count is intended now to help predict the effort or duration needed to build the system, a scale factor of less than 1.0 should be used in step 4 of the counting process. Counting a design object model from the user's perspective is more complicated. To count what can actually be perceived by the user of the system, the original abstractions present in the requirements and analysis models have to be recovered. Implementation details should not affect the count. There might no longer be a strict mapping of single classes to logical files; collections of classes may sometimes need to be counted together as a single logical file. There may be many different ways to identify logical files. We consider four, which are defined by different choices of how to deal with aggregations and generalization / specialization relationships: 1. Single Class: count each separate class as a logical file, regardless of its aggregation and inheritance relationships; 2. Aggregations: count an entire aggregation structure as a single logical file, recursively joining lower level aggregations. 3. Generalization/Specialization: given an inheritance hierarchy, consider as a different logical file the collection of classes comprised in the entire path from the root superclass to each leaf subclass, i.e. inheritance hierarchies are merged down to the leaves of the hierarchy. 4. Mixed: combination of options 2 and 3. Insert Initialize Delete Top-of-pile Bottom-of-pile Location Visibility Collection Of Cards Rank Display Discard Card Deck Shuffle Deal Hand Initial State Draw Pile Draw Discard Pile Draw Figure 3. Single class ILFs. Insert Initialize Delete Top-of-pile Bottom-of-pile Location Visibility Collection Of Cards Rank Display Discard Card Deck Shuffle Deal Hand Initial State Draw Pile Draw Discard Pile Draw Figure 4. Aggregations ILFs. Insert Initialize Delete Top-of-pile Bottom-of-pile Location Visibility Collection Of Cards Rank Display Discard Card Deck Shuffle Deal Hand Initial State Draw Pile Draw Discard Pile Draw Figure 5. Generalization/Specialization ILFs. For example, Figures 3-6 show the different counting boundaries that result from these four strategies, on a sample object model 1 . Aggregation merging decreases the number of classes in the object model from 6 to 5 as CollectionOfCards is merged with Card; the resulting logical file contains all the data members and methods of the two classes. Generalization/specialization merging projects the superclass CollectionOfCards onto its subclasses, again reducing the number of logical files from 6 to 5. Finally, combining Aggregation and Gener- alization/Specialization merging first aggregates CollectionOfCards with Card and then projects the result onto the subclasses of CollectionOfCards, resulting in 4 logical files. Conceptually, it makes sense to merge superclasses into subclasses for OOFP counting. It seems right to count the leaf classes, with their full inherited structure, since this is how they are instantiated. (The non-leaf classes of a hierarchy usually are not instantiated - they are created for subsequent re-use by deriving subclasses, and for exploiting polymorphism.) Also, two classes linked by a generalization / specialization relationship are intuitively less complex than two separate classes, because the subclass represents a refinement of the superclass. 1 The model is drawn from Rumbaugh et al. (1991). A Function Point-like Measure for Object-Oriented Software 11 Insert Initialize Delete Top-of-pile Bottom-of-pile Location Visibility Collection Of Cards Rank Display Discard Card Hand Initial State Draw Pile Draw Discard Pile Draw Deck Shuffle Deal Figure 6. Mixed ILFs. Associations may present a problem. If non-leaf classes of an inheritance hierarchy participate in associations, replicating the superclass association into each subclass would increase artificially the number of associations. In fact, the original superclass association contributes to complexity only in the superclass, and in code it will only be implemented once. Merging aggregations into a single entity for OOFP counting seems less intuitive. The objects that form aggregations are separate objects, that exist independently of each other and have their own methods and data members. At run-time, different objects will be instantiated for each class in the aggregation. However, it can be argued that dividing a user-identifiable class into an aggregation of sub-classes is an implementation choice. From the point of view of the end user, and of the function point measurement philosophy, the OOFP value should not be affected. From this perspec- tive, the aggregation structure should be merged into a single class and counted as a single logical file. Whether or not it is right to merge aggregations seems to depend on whether the user's or designer's perspective is chosen. A hybrid solution can be adopted, in which the treatment of aggregations is considered as a parameter of the OOFP counting process. Three options can be identified, with the choice left to the measurer: 1. merge aggregations; 2. do not merge aggregations; 3. flag on the design which aggregations should be considered as a unique entity and so must be merged. 4.2. Logical Files What is considered an ILF/EIF varies, according to the particular ILF/EIF identification strategy used. Merging aggregations or general- 12 Giuliano Antoniol et al. izations can generate ILFs or EIFs that correspond to sets of classes in the design. We call these composite ILFs / EIFs, to distinguish them from those consisting of a single class, called simple. For each ILF/EIF it is necessary to compute the number of DETs (Data Element Types) and RETs (Record Element Types). The rules for DET/RET computation are slightly different for simple or composite ILFs / EIFs. In both cases, one RET is associated to each ILF/EIF, because it represents a "user recognizable group of logically related data" (IFPUG 1994). When the DETs and RETs of an ILF or EIF have been counted, tables are used to classify the ILF/EIF as having low, average, or high complexity. We base these tables on those given in the IFPUG Counting Practices Manual Release 4.0 (IFPUG 1994). 4.2.1. Simple ILFs/EIFs Simple attributes, such as integers and strings, are considered as DETs, since they are a "unique user recognizable, non-recursive field of the ILF or EIF" (IFPUG 1994). A complex attribute in the OO paradigm is an attribute whose type is a class (this models the analogy of a complex attribute with a RET, i.e. "a user recognizable subgroup of data elements within an ILF or EIF" (IFPUG 1994)) or a reference to another class. Associations need to be counted as well, since they contribute to the functionality/complexity of an object. An association is usually implemented as a data member referencing the associated objects; this reference is used in methods to invoke the associated object's services. Associations are counted as DETs or RETs according to their car- dinality. A single-valued association is considered as a DET (IFPUG suggests counting a DET for each piece of data that exists because the user requires a relationship with another ILF or EIF to be maintained 1994)). A multiple-valued association is considered as a RET, because an entire group of references to objects is maintained in one attribute. Aggregations are a special case of associations. For simple ILFs / EIFs, they are treated as normal associations. 4.2.2. Composite ILFs/EIFs DETs and RETs are counted for each class within the composite, and summed to give the overall total for the composite ILF/EIF. DETs and RETs are counted using the same rules as for simple ILFs / EIFs, except for aggregations. Aggregations are dealt with in a special way because in a composite ILF/EIF they represent a subgroup. A Function Point-like Measure for Object-Oriented Software 13 One RET is counted for each aggregation, whatever its cardinality. The RET is assigned to the container class. In practice, the values of DET and RET for any ILF/EIF are computed by counting DETs and RETs for each component class on its own (this is trivial for a simple ILF/EIF), and just adding them up 2 . 4.3. Service Requests Each service request (method) in each class in the system is examined. Abstract methods are not counted. Concrete methods are only counted once (in the class in which they are declared), even if they are inherited by several subclasses, because they are only coded once. If a method is to be counted, the data types referenced in it are classified: simple items (analogous to DETs in traditional function points) are simple data items referenced as arguments of the method, and simple global variables referenced in the method; items (analogous to File Types Referenced - FTRs - in traditional function points) are complex arguments, objects or complex global variables referenced by the method. Several approaches are possible to distinguish complex items from simple ones. For example, compiler built-in types might be considered as simple and all other types as complex. This choice might not be appropriate, since all user-defined types would be counted as com- plex, even if they were scalar types or aliases of built-in types. Another approach is to regard a complex item as one whose type is a class or a reference to another class. This approach is used here. When the DETs and FTRs of a method have been counted, tables are used to classify the method as having low, average, or high com- plexity. We base these tables on those given in the IFPUG Counting Practices Manual Release 4.0 (IFPUG 1994) for external inputs and queries. Most of the time, the signature of the method provides the only information on DETs and FTRs. Sometimes, especially early on, even that is not known. In such a case, the method is assumed to have average complexity. 2 The counting rules defined make DET-RET additive. The only exception is the aggregation relation, which is handled differently in simple and composite ILFs. However, in practice, the contribution of aggregation in composite ILFs corresponds to considering one RET for each class involved in the aggregation structure, which becomes equivalent to summing the RETs of each component class separately. 14 Giuliano Antoniol et al. 4.4. Allowing for reuse In an early count, where the main aim is to capture user-oriented func- tionality, the scale factor used in step 4 of the counting process should be set to 1.0. A user doesn't care where a class comes from, so the class should be counted with its full inherent value. From a designer's or implementer's point of view, reuse makes classes easier to develop. In a later count, in which the OOFP count may be intended to help predict the effort or duration needed to build the system, a scale factor of less than 1.0 would be appropriate. 4.5. An example Figures 3-6 show four different ways that classes in an object model might be merged, according to which of the four different LF identification strategies is used. Here we show the OOFPs that are computed for each variant. Service Requests: Service requests (methods) can be counted immediately. Since they are only counted once anyway, it does not matter how the classes are aggregated into logical files. Because the signatures are unknown for the methods in the example, each method is assumed to have average complexity. They each receive the four OOFPs that are scored for an average service request. As there are 12 concrete methods in the model, service requests contribute 12 \Theta OOFPs. Logical files: The counting procedure for each individual class gives the DETs and RETs shown in Figure 7. The class Card has three DETs (two due to the two data items and one due to the many-to-one association with CollectionOfCards) and one RET (since the class itself is a collection of related data items). CollectionOfCards has two DETs due to its two data items, one RET due to the one-to-many aggregation with Card, and one RET for its own structure. Each other class has one RET and as many DETs as it has data items. Depending on which ILF identification strategy is used, there are four different ILF variants. Each variant merges classes together in different ways, resulting in different total DET and RET counts. Table I shows the result of applying IFPUG 4.0 complexity tables with each variant. The value Low is rated as 7 OOFP, according to the IFPUG tables. A Function Point-like Measure for Object-Oriented Software 15 Insert Initialize Delete Top-of-pile Bottom-of-pile Location Visibility Collection Of Cards Rank Display Discard Card Deck Shuffle Deal Hand Initial State Draw Pile Draw Discard Pile Draw Figure 7. DET/RET computation for LFs on the example system. Table I. ILF and SR complexity contribution (S Collection of Cards Low Low - Card Low - Low - Deck Low Low Low Low Hand Low Low Low Low Discard Pile Low Low Low Low Draw Pile Low Low Low Low 28 SR OOFP 48 48 48 48 Total OOFP 90 83 83 76 The highest OOFP count comes when each class is counted as a single ILF. All the other variants have the effect of reducing the OOFP value, as they reduce the number of ILFs. Although there is an increase in DETs / RETs in the merged ILFs, it is not enough to raise the ILF complexity to higher values. 5. Tools for Counting OOFPs The process for computing OOFPs has been automated, as shown in Figure 8. Object models produced with CASE tools are translated to an intermediate representation. The intermediate representation is parsed, Giuliano Antoniol et al. OOFP Translator Tool-AOL Other Tool Tool Output AST AOL Parser AOL Specification OOFP Counting Rules OOFP Counter OMT/STP Rational ROSE STP Output Petal Output OMT/STP-AOL Translator Translator Petal-AOL Figure 8. OOFP Computation Process producing an Abstract Syntax Tree (AST), to which the OOFP counting process is applied. In order to be independent of the specific CASE tool used, an intermediate language, called Abstract Object Language (AOL), has been devised. The language is a general-purpose design description language, capable of expressing all concepts available at the design stage of object oriented software development. This language is based on the Unified Modeling Language (Rational Software Corporation 1997b), a superset of the OMT notation that is becoming the standard in object oriented design. Since UML is a visual description language with some limited textual specifications, we had to design from scratch many parts of the language, while remaining adherent to UML where textual specifications where available. Figure 9 shows an excerpt from the AOL description of the object model depicted in Figure 7. The output of the specific CASE tool used is translated automatically into an equivalent AOL specification. One translator has been imple- mented, to convert the output from OMT/STP (Interactive Development Environments 1996) to an AOL specification. Other translators main.tex; 23/06/1999; 14:27; no v.; p.16 A Function Point-like Measure for Object-Oriented Software 17 class Deck operations class Hand attributes operations aggregation container class CollectionOfCards mult one parts class Card mult many; generalization CollectionOfCards subclasses Deck, Hand, DiscardPile, DrawPile Figure 9. Excerpt of the AOL specification for the example object model. could be implemented for other CASE tools, such as Rational Rose (Rational Software Corporation 1997a) which fully supports UML and represents its output using a language called Petal. The AOL specification is then parsed by an AOL parser, producing an AST representing the object model. The parser also resolves references to identifiers, and performs some simple consistency checking (e.g. names referenced in associations have been defined). The OOFP Counter implements the OOFP Counting Rules described in Section 4. The OOFP Counter is very different from other measurement and counting tools, because instead of assuming a specific counting strategy it allows one of several strategies to be chosen. This makes it suitable for experimentation. The tool is very flexible, being parameterizable with respect to the rules used in the counting process. The AOL parser and the OOFP counter have been implemented in both Refine (Reasoning Systems 1990) and Lex/Yacc. 6. Pilot Study The described methodology has been applied in an industrial environment producing software for telecommunications. Our first study is of the relationship between the OOFP measure of a system and its final main.tex; 23/06/1999; 14:27; no v.; p.17 Giuliano Antoniol et al. size in lines of code (LOC), measured as the number of non-blank lines, including comments. Eight sub-systems of a completed application were measured. These eight systems were chosen for study because all were developed by the same poeple, in the same environment, using the same language (C++). Design documents and the final source code were both avail- able. Measurements of design characteristics were taken from the design documents, not "reverse engineered" from the source code. Table II shows the numbers of various design elements in each system Table III shows the size of each system, spreading from about 5,000 to 50,000 lines of code. Table III also shows the OOFP count for each system, using each of the four different strategies for identifying logical files. Table II. Design Characteristics (Atr= Attributes, Aggregation, es). System Atr Ope Ass Agg Inh Cls The four OOFP series are strongly correlated with each other. The lowest Pearson correlation, between the Single Class (S) and Mixed (M) strategies, is .992. Other correlations range up to .998. As shown in table III, differences between the methods become appreciable only for the projects with large LOC values. The high correlation between the four OOFP series suggests that they are essentially linear transformations of each other. In that case, changing the strategy for identifying logical files might not make much difference to the accuracy of size estimation models. 6.1. Model Evaluation A leave-one-out cross-validation procedure (Stone 1974) was used to measure model performance. Each model was trained on A Function Point-like Measure for Object-Oriented Software 19 table III. System sizes and OOFPs. System LOC S A G M A 5807 63 63 D 19863 1071 1057 1057 1043 F 31011 518 403 483 368 G 47057 1142 1100 1124 1072 of the data set L (sample size is currently tested on the withheld datum. The step was repeated for each point in L and accuracy measures averaged over n. This method gives an unbiased estimate of future performance on new data, and enables quite different models to be compared directly. Model error was estimated as the cross-validation version of the normalized mean squared error (NMSE). This is the mean squared error, normalized over the variance of the sample: Regression based on least square minimization assumes that the distribution of errors is Gaussian. Statistical tests for skewness and kurtosis do not cause the hypothesis of normality to be rejected for any of the five distributions (LOC and the four OOFP values). But given the small size of our data set, it is not clear that the distributions really are normal. The least squares approach is sensitive to outliers (data points far removed from other data points) in the data, because it minimizes squared deviations. An outlier can have almost as much influence on the regression results as all other points combined. Standard box-and- whisker plots do not identify any of our systems as outliers, although system H is right on the edge of being considered an outlier. The impact of such influential points can be lessened by reducing the weight given to large residuals - for example, by minimizing the sum of absolute residuals rather than the sum of squared residuals. Thus another measure for errors, based on absolute values, was also 20 Giuliano Antoniol et al. considered to check inconsistencies due to possible influential or outlier points. This measure is normalised mean absolute error (NMAE): jy k \Gammamed y j where are the mean and median of the observed values in the sample L. Where available, the cross-validation estimates of the standard error oe of the residuals y and of the r-squared R 2 of the fit were also computed. Even with cross-validation, care is needed in interpreting the results. The small sample means that any observations must be regarded as indicative rather than conclusive. 6.2. Models considered Several regression techniques were considered to model the relationships of LOC with OOFP. Other predictors of LOC, based on direct indicators of OO size such as the number of classes or methods in the design, were considered for comparison. First, linear models (lms in table IV) based on minimizing the sum of squares of the residuals were developed for each LF selection method. Least absolute deviation, based on L 1 error, was also applied (l1s in table IV). This method minimizes the sum of the absolute values of the residuals, to reduce the effect of large error values. Robust regression techniques were also investigated, to handle non-obvious outliers. A family of M-estimators (see the Appendix) was considered (rregs and rlms in table IV). The basic idea of M-smoothers is to control the influence of outliers by the use of a non-quadratic local loss function which gives less weight to "extreme" observations. Examples of smoothers are Andrews, bisquare, fair, Hampel, Huber, and logistic (Venables and Ripley 1994). Each corresponds to a different weight function. Finally, multivariate linear models were developed that predict LOC directly from the numbers of OO design elements shown in table II. 6.3. Results table IV shows each of the models, parameterized over LF selection methods and the type of regressor. The model coefficients b 0 and b 1 indicated were computed from the full data set. The estimated model errors (NMSE and NMAE) are shown for each model. The estimated R-squared measure is also included for the linear models. A point of concern is whether an intercept term b 0 should be included in the model. It is reasonable to suppose the existence of support code main.tex; 23/06/1999; 14:27; no v.; p.20 A Function Point-like Measure for Object-Oriented Software 21 table IV. Model performance for linear regressors (lms and l1s) and robust methods (rregs and rlms). Method lm-A 0.43 0.66 0.69 8505 23.8 rreg-A 0.43 0.66 - 8255 24.0 not directly related to the functionalities being counted; and prediction is improved with the term. However, the intercept term is not significant in a non-predictive fit of the data. More importantly, the fact that the intercept term is always larger than our smallest system might indicate a poor fit for small OOFP values. It would be interesting to apply a Bayesian procedure to select the intercept from given priors. The results summarized in table IV are encouraging. For example, the lm-G model has an NMSE of 38 %, meaning that the square error variance is less than half of the sample variance. From another point of view, models based on the OOFPs counted using the Generalization strategy achieve a cross validation average error of 47 %, which is very good. The best model in table IV is rreg-G. Further investigation of rreg-G was done, with the results shown in table V. The best predictive accuracy (NMSE=0.337) was achieved by the rreg-logistic-G model with tuning parameter This corresponds main.tex; 23/06/1999; 14:27; no v.; p.21 22 Giuliano Antoniol et al. table V. Model performances for different weighting functions of the M-estimator rreg, for the Generalization selection method. Method NMSE Comments rreg-fair-G 0.48 converged after 50 steps) to the linear predictor (This model is very close to the basic linear model lm-G, whose equation is table IV suggests that in this data set the Generalization strategy is consistently best. This is not proven statistically, though. A non-parametric bootstrap approach (Efron and Tibshirani 1993) was used to assess the models. The null hypothesis that there are no differences between the errors from the lm-S and lm-G models cannot be rejected; similar results were obtained for the other models. Thus it is not clear that any counting strategy should be preferred over any other. For comparison, multivariate linear models were developed that predict LOC directly from the numbers of OO design elements (shown in table II). Poorer results are obtained from such models. For example, a model based on classes and methods has Models based on OOFP perform much better. This pilot study was conducted in a specific project and environ- ment, in a specific organization. The results are encouraging for size estimation in this context. The issue of external validity must be addressed by more extensive studies, targeting multiple organizations and different projects. We have taken the first step in empirical investigation; more needs to be done. A Function Point-like Measure for Object-Oriented Software 23 7. Discussion of Results As can be seen in table I, the complexity of each LF is always determined to be low, even when several classes are merged together. The same is true for service requests. The tables used to determine complexity are based on those from the IFPUG Counting Practices Manual (IFPUG 1994), in which quite large numbers of RETs and DETs are needed to reach average or high complexity (for example, to obtain an average complexity weight an LF needs a DET value between 20 and 50 and a RET value between 2 and 5). This is due to the data processing origins of the function points method, and doesn't seem to apply as well to all kinds of systems. Therefore the complexity tables should be recalibrated, to provide more discrimination. The implicit assumption in the use of these tables is that the complexity of a class, and hence the size of its implementation and the effort required to implement it, increases as the number and complexity of its attributes increases. Similarly, the complexity of a method (and hence its size and development effort) is assumed implicitly to increase as the number and complexity of its parameters increases 3 . Whether these assumptions are true needs to be determined experimentally. The assumption seems reasonable for classes as a whole, but perhaps not for methods. What works for transactions in traditional function points may not work for methods in an object model, because transactions tend to be much more coarse grained than methods. At the analysis and design stages, we often have no more information about a method than its signature. If it turns out that this is unrelated to complexity and size, we have nothing to go by in counting OOFPs. One possibility would be to permit the designer to annotate a method with a complexity rating. This would introduce a subjective element to the process, however, which we want to avoid. Another approach would be simply to count the methods, making no attempt to classify them by complexity. A promising approach would take advantage of the information available in use cases and scenarios to derive a complexity rating for methods. On the data available to us so far, it seems that recalibration of the OOFP tables for logical files might improve the accuracy of OOFP as a predictor of size; recalibration of the table for methods might not. Further experimentation is needed on this topic, with data from more systems. In order to support such experimentation, the tool used to 3 These assumptions are fairly common. They underlie the philosophy of the classical function point method. They also feature in the design metrics work of Card and Glass (1990). Giuliano Antoniol et al. count OOFPs is designed to consider the table entries as parameters that can be modified at any time. The pilot study suggests that for this organization there is no reason to prefer any of the four strategies for indentifying LFs over any other. Other organizations may find differently. Although for this organization the best size predictions appear to be obtained with the Generalization strategy, its superiority is not proven statistically and may be an accident of the data. Modifying the complexity tables might make a difference in determining the best strategy for selecting LFs. Once a counting scheme has been chosen, it is important that it be applied consistently. Consistent counting is straightforward for us, since tools are used to automate the process. 8. Conclusions We have presented a method for estimating the size of object oriented software. The method is based on an adaptation of function points, to apply them to object models. The proposed method takes full advantage of the information contained in the object model and eliminates the ambiguities of the traditional function points method. It can be parameterized in order to take into account more closely the characteristics of a specific design environment or of a particular problem. We have defined the mapping from FP concepts to OO concepts, and described the counting process. Tools have been developed that automate the process. Preliminary results from a pilot study in an industrial environment have been reported. The results from the pilot study show promise for size estimation. This is important, since an estimate of size is needed for many effort estimation models. In summary, we have shown that we can apply the concepts of function points to object oriented software and that the results are accurate and useful in an industrial environment. Future work will take several directions. One is to investigate the effect of recalibrating the complexity tables. Other relationships, beyond just OOFPs and code size, will be studied; those between OOFPs and traditional FPs, and OOFPs versus effort, are of particular interest. Another avenue is to consider the impact of using design patterns (Gamma, Helm, Johnson and Vlissides 1995) on the structure within object models; this may lead to other strategies for identifying logical ILFs. A Function Point-like Measure for Object-Oriented Software 25 Acknowledgements We thank the referees for their constructive comments. This research was funded by SODALIA Spa, Trento, Italy under Contract n. 346 between SODALIA and Istituto Trentino di Cultura, Trento, Italy. When this work was undertaken, G. Caldiera and C. Lokan were with the Experimental Software Engineering Group at the University of Maryland, and G. Antoniol was with the ITC-IRST, Istituto per la Ricerca Scientifica e Tecnologica, I-38050 Povo (Trento), Italy. --R Software function Measuring Software Design Quality Controlling Software Projects An Introduction to the Bootstrap Mapping the OO-Jacobson approach to function point analysis Design Patterns: Elements of Reusable Object Oriented Software Applied Nonparametric Regression Software Through Pictures Manuals. Estimating size for object-oriented software Measuring object-oriented software with predictive object points Unified Modeling Language Reasoning Systems: Measuring the size of object-oriented systems Estimating the costs of object-oriented software Modern Applied Statistics with S-Plus Applying function points to object oriented software --TR --CTR Maurizio Morisio , Daniele Romano , Ioannis Stamelos, Quality, Productivity, and Learning in Framework-Based Development: An Exploratory Case Study, IEEE Transactions on Software Engineering, v.28 n.9, p.876-888, September 2002 Marjan Heriko , Ivan Rozman , Ale ivkovi, A formal representation of functional size measurement methods, Journal of Systems and Software, v.79 n.9, p.1341-1358, September 2006 Gennaro Costagliola , Filomena Ferrucci , Genoveffa Tortora , Giuliana Vitiello, Class Point: An Approach for the Size Estimation of Object-Oriented Systems, IEEE Transactions on Software Engineering, v.31 n.1, p.52-74, January 2005 G. Antoniol , R. Fiutem , C. Lokan, Object-Oriented Function Points: An Empirical Validation, Empirical Software Engineering, v.8 n.3, p.225-254, September
design metrics;function points;object oriented;size estimation
594739
Efficient Window Block Retrieval in Quadtree-Based Spatial Databases.
An algorithm is presented to answer window queries in a quadtree-based spatial database environment by retrieving all of the quadtree blocks in the underlying spatial database that cover the quadtree blocks that comprise the window. It works by decomposing the window operation into sub-operations over smaller window partitions. These partitions are the quadtree blocks corresponding to the window. Although a block b in the underlying spatial database may cover several of the smaller window partitions, b<math> is only retrieved once rather than multiple times. This is achieved by using an auxiliary main memory data structure called the active border which requires <math>O\left(n\right)<math> additional storage for a window query of size <math>n\times n<math>. As a result, the algorithm generates an optimal number of disk I/O requests to answer a window query (i.e., one request per covering quadtree block). A proof of correctness and an analysis of the algorithms execution time and space requirements are given, as are some experimental results.
Introduction Spatial data consists of spatial objects made up of points, lines, regions, rectangles, sur- faces, volumes, and even data of higher dimension which includes time. Examples of spatial data range from locations of cities, rivers, roads, to the areas that are spanned by counties, states, crop coverages, mountain ranges, etc. They are increasingly finding their way into adaptations of conventional databases for use in applications in geographic information systems (GIS), resource management, space, urban planning, etc. [9, 22]. There are many different representations of spatial data (see [20, 21] for an overview). We are interested in representations that are based on spatial occupancy. Spatial occupancy methods decompose the space from which the data is drawn (e.g., the two-dimensional space containing the lines) into regions called buckets. They are also commonly known as bucketing methods. Traditionally, bucketing methods such as the grid file [19], BANG file [13], LSD trees [17], Buddy trees [25], etc. have usually been applied to points, although they can be applied to the other types as well. There are four principal approaches to decomposing the space from which the data is drawn. One approach buckets the data based on the concept of a minimum bounding (or enclosing) rectangle. In this case, the minimum bounding rectangles of the objects are grouped (hopefully by proximity) into hierarchies, and then stored in another structure such as a B-tree [7]. The R-tree [16] (as well as its variants such as the R -tree [6]) is an example of this approach. The drawback of these hierarchies of objects is that they do not result in a disjoint decomposition of the underlying space. The problem is that each object is only associated with one bounding rectangle even though it may also overlap a portion of the bounding rectangle of another object. In the worst case, this means that when we wish to determine which object is associated with a particular point in the two-dimensional space from which the objects are drawn (e.g., the containing rectangle in a rectangle database, or an intersecting line in a line segment database), we may have to search the entire database. The other approaches are based on a decomposition of space into disjoint cells, which are mapped into buckets. Their common property is that the objects are decomposed into disjoint subobjects such that each of the subobjects is associated with a different cell. They differ in the degree of regularity imposed by their underlying decomposition rules and by the way in which the cells are aggregated. The price paid for the disjointness is that in order to determine the area covered by a particular object, we have to retrieve all the cells that it occupies. Moreover, if we wish to report all the objects that overlap a particular area, then we may have to report an object as many times as its subobjects appear in the area. The first method based on disjointness partitions the objects into arbitrary disjoint subobjects and then groups the subobjects in another structure such as a B-tree. The partition and the subsequent groupings are such that the bounding rectangles are disjoint at each level of the structure. The R + -tree [26] and the cell tree [15] are examples of this approach. Their drawback (as well as the R-tree variants) is that the decomposition is data-dependent. This means that it is difficult to perform tasks that require composition of different operations and data sets (e.g., set-theoretic operations such as overlay). In contrast, the remaining two methods, while also yielding a disjoint decomposition, have a greater degree of data-independence. They are based on a regular decomposition. The space can be decomposed either into blocks of uniform size (e.g., the uniform grid [12]) or adapt the decomposition to the distribution of the data (e.g., a quadtree-based approach that makes use of regular decomposition such as [24]). In the former case, all the blocks are of the same size, while in the latter case, the widths of the blocks are maximal subject to being restricted to be powers of two, and a restriction on their positions. For example, Figure 1 shows the quadtree block decomposition of two square regions of space each of which contains a rectangular subregion (termed a window) delimited by heavy lines. The blocks are obtained by applying regular decomposition to the square regions thereby repeatedly breaking them up into four congruent blocks until each block is either completely within the window or completely outside the window. window w window w underlying spatial database underlying spatial database Figure 1: The decomposition of (a) a 12 \Theta 12 window, and (b) a 13 \Theta 13 window into maximal quadtree blocks. The uniform grid is ideal for uniformly distributed data, while quadtree-based approaches are suited for arbitrarily distributed data. In the case of uniformly distributed data, quadtree-based approaches degenerate to a uniform grid, albeit they have a higher overhead. Both the uniform grid and the quadtree-based approaches lend themselves to set-theoretic operations as the positions of the decomposition lines are restricted and thus there is much less variation between the operands of the operations. Thus they are ideal for tasks which require the composition of different operations and data sets. In gen- eral, since spatial data is not usually uniformly distributed, the quadtree-based regular decomposition approach is more flexible. A window query is the spatial analog of a range query in that it retrieves all objects that overlap the space covered by a range of x and y (and possibly z in three-dimensions) coordinate values which form the window. In this paper we focus on performing a variant of a window query using a regular decomposition quadtree. Both the window and the underlying database are represented by a quadtree. In particular, the quadtree blocks BW that make up the window are used to guide the retrieval process. The variant of the query is one that retrieves all of the quadtree blocks BU of the underlying database that cover the blocks that make up the window (i.e., BW ). This query differs from the classical window operation described above which retrieves the objects in the underlying database that cover the window instead of the blocks in the underlying database as we do here. The rationale for using the quadtree blocks of the window is to match the quadtree decomposition of the underlying spatial database. This makes it more straight-forward to answer the window query since there is a direct correspondence between each window block and some overlapping quadtree block(s) in the underlying spatial database. The answer to the window query is the union of all the answers generated by querying the underlying spatial database with the maximal quadtree blocks comprising the window serving as the individual queries. Our variant can be viewed as a preliminary step to the retrieval of the objects in that it retrieves the blocks in the underlying database (i.e., BU ) that correspond to the window. The next step would process blocks BU and extract the relevant objects from them. When the underlying database is a quadtree where the objects have been decomposed so that the blocks which contain them are disjoint, the step that extracts the relevant objects from the blocks may in fact encounter some of the objects more than once (e.g., when a region or line object has been decomposed into several blocks each of which contains a part of the region or line object). In this case, this step would have to eliminate the duplicates which is not a simple matter (but see [2, 4]). The rationale for our variant is that we may wish to use these blocks (i.e., BU ) as input to a subsequent operation whose underlying representation is also a quadtree thereby facilitating the composition of several operations. Another way to characterize our variant is that it is somewhat like a clipping operation where we are using the quadtree blocks that make up the query window (i.e., BW ) to clip the blocks that make up the underlying database (i.e, BU ). In this paper we show how to retrieve the quadtree blocks from the underlying database that cover the quadtree blocks that comprise the window. In particular, we describe a method that retrieves each block b in BU just once even though b may cover several blocks in BW . The rest of this paper is organized as follows. Section 2 gives an overview of our approach. Section 3 describes our algorithm. Section 4 contains an informal proof of correctness for the algorithm's block retrieval process, while an analysis of its worst-case execution time and space complexity is given in Section 5. Section 6 presents empirical results of the disk I/O behavior of the algorithm, while concluding remarks are drawn in Section 7. 2 Overview of our Approach A window decomposition algorithm is given in [3] which decomposes a two-dimensional window of size n \Theta n in a feature space (e.g., an image) of size T \Theta T into its maximal quadtree blocks in O(n log log T ) time. Once the set BW has been determined, we simply retrieve the elements of the underlying spatial database S that overlap each of its elements. The drawback of this algorithm is that many of the elements of S may be retrieved more than once. For example, in Figure 2, the algorithm would retrieve block p of the underlying spatial database four times (once for each of the maximal window blocks 1, 4, 8, and 10). We assume that the underlying spatial database is disk-resident, and we often speak of the operation of retrieving a block of the underlying spatial database as a disk I/O request. This means that redundant disk I/O requests will result. 2 One solution is to keep track of all blocks that have already been retrieved. This is not easy without additional storage (see [2] for a discussion of the similar issue of uniquely reporting answers in a spatial database). s r Underlying spatial database window w Figure 2: Examples where more than one window block retrieves the same block of the underlying spatial database. The problem with using the algorithm in [3] is that the process of generating the maximal blocks that comprise the window only depends on the query window and does not take into consideration the decomposition of space induced by the underlying spatial database. We overcome this problem by generating and retrieving each covering block in the underlying spatial database just once. This is achieved by controlling the window decomposition procedure through the use of information about blocks of the underlying spatial database that have already been retrieved. We use an approach based on active borders [23], at the expense of some extra storage. The algorithm that we present performs this task with the same worst-case CPU execution-time complexity as the one in [3] (i.e., O(n log log T )). The difference is in the I/O cost where the new algorithm makes just M requests to access the underlying spatial database instead of max(N; M) as in [3], where M is the number of quadtree blocks in the underlying spatial database that overlap the This problem can be overcome via appropriate use of buffering techniques. However, in this paper we show how to avoid the problem by retrieving each block of the underlying spatial database just once without relying on buffering techniques. window and N is the number of maximal quadtree blocks in the window. A general significance of both our algorithm and the one in [3] is that although the window contains n 2 pixel elements, the worst-case CPU execution-time complexity of the algorithms is almost linearly proportional (and not quadratic) to the window diameter, and is independent of other factors. It is important to note that we retrieve blocks in the underlying spatial database by use of information (partial) about their relationship to other blocks (e.g., containment, overlap, subset, etc. We do not retrieve a block of the underlying database by its identifier. If we could do this, then we could keep track of which blocks are retrieved via a hash table, for example, and avoid retrieving them again. Instead, we are given the spatial description of a window block, say b. The spatial description of b is used to retrieve all the blocks of the underlying spatial database that are spatially related to b (e.g., the blocks that contain, or are contained in, b). Blocks in the underlying spatial database can be retrieved more than once if they satisfy some spatial relationship with respect to different window blocks. In order to avoid retrieving the same block more than once when a different window block is processed, we maintain a spatial analog to the hash-table mechanism above. This is achieved through the usage of some spatial data structure, namely the active border, tailored to match the needs of this type of spatial retrieval. The active border can also be viewed as simulating the spatial equivalent of a sort-merge list of pages which is used in database query processing when accessing data through secondary indexes [10]. 3 Algorithm Answering a window query by first computing the maximal quadtree blocks comprising it, and then retrieving the corresponding covering blocks in the underlying spatial database proceeds as follows. Assume a query window W, a spatial database S, a query function F that performs the appropriate variant of a window query test (e.g., a containment test) and a record of type answer set that accumulates the answer to the window query. answer-set procedure Algorithm-1(S,W,F); begin reference spatial-database value window W; value function F; block B; block set C; spatial-object set T; answer-set RESULT; decompose W into its maximal quadtree blocks; foreach block B in W do BEGIN C:=blocks in S THAT cover B; foreach block Q in C do apply F to spatial objects associated with Q */ By varying the function F and the data type answer set, many window operations can be implemented using Algorithm-1. For example, to answer the report query (i.e., reporting the identity of all the features that exist inside a window), the function F simply identifies all the spatial objects inside the block of the underlying spatial database, and the data type answer set is just a set of spatial object identifiers for the qualifying objects. To answer the exist query (i.e., determining if feature f exists in w), the function F tests whether or not f (or f 's identifier) exists inside the block of the underlying spatial database, and the data type answer set is the type Boolean while U is a logical or operation. To answer the select query (i.e., reporting the locations of all instances of feature f in the window), the function F simply tests whether or not f (or f 's identifier) exists inside the block of the underlying spatial database, and the data type answer set is a quadtree that stores in it the location of these blocks. There is one principal issue in implementing this algorithm. This was discussed in Section 2 and corresponds to the situation that block q in the underlying spatial database covers more than one maximal quadtree block in the window. In this case, q will be retrieved several times. This is what happens in the algorithm reported in [3]. This could be overcome by avoiding the invocation of the retrieval step for some of the maximal quadtree window blocks. The issue is how do we skip some of the maximal quadtree window blocks. In order to understand this issue, we briefly focus on the relation between the maximal quadtree blocks of the window decomposition and the quadtree blocks in the underlying spatial database. Assume that b is a maximal window block that is generated by the window decomposition algorithm. Due to the quadtree decomposition of both the window and the underlying spatial database, b can either be contained in, or contain, one or more quadtree blocks of the underlying spatial database. In particular, there are three possible cases as illustrated by Figure 2. Case 1 is demonstrated in the figure by window block 2 which contains more than one quadtree block of the underlying spatial database. All of these blocks have to be retrieved (e.g., from the disk), and processed by the algorithm (e.g., the spatial objects associated with these blocks will be reported as intersecting the window). The second case is illustrated by window block 9 of Figure 2. Block 9 contains exactly one block of the underlying spatial database which will have to be retrieved (e.g., from the disk) as well. The third case is demonstrated by window blocks 1, 4, 8, and 10 of Figure 2 which all require retrieving (e.g., from the disk) the same quadtree block (i.e., block p of the underlying spatial database). Case 3 arises frequently in any typical window query, as shown by the experiments conducted in Section 6, thereby resulting in a large number of redundant disk I/O requests. Our algorithm is an improvement over Algorithm-1 and is based on the following observation (it is restated as Lemma 1, as well as proved, in Section 4): Observation 1: Assume that a block, say b, is a maximal block that lies inside the window w and overlaps with a block of the underlying spatial database, say q. If q is of greater size than b, then q must intersect with at least one of the boundaries of the window w (refer to Figure 3 for illustration). In other words, there cannot be database blocks that are bigger than the intersecting window blocks which are in the middle of the query window. These big database blocks have to intersect the boundary of the query window. Our window retrieval algorithm is based on this observation which we illustrate further later in this section. The new algorithm consists of procedures WINDOW RETRIEVE, GEN SOUTHERN MAXIMAL, and MAX BLOCK. They are described below, while their detailed code is given in the Ap- pendix. The algorithm works for an arbitrary rectangular window (i.e., it need not be square). We avoid generating non-maximal quadtree blocks in the window (or at least generate a bounded number of them) by using the same technique as in [3], which we outline below. Note that there are O(n 2 ) non-maximal blocks inside an n \Theta n window. Also, each maximal quadtree block in the window is processed only once (i.e., as a neighbor of another node) regardless of its size. We make use of an active border data structure [23] which is a separator between the window regions that have already been processed and the rest of the window. Note that the active border in our case will differ from the conventional one (which looks like a staircase) because of the nature of the block traversal process. In particular, we traverse the blocks in the window in a row-by-row manner rather than in quadrant order (i.e., NW, NE, SW, SE). Figures 4-8 represent the first five steps of the execution of the algorithm for the query window w. The heavy lines in Figure 4 represent the active border for window w at the initial stage of the algorithm. In generating a new block, the window decomposer has to consult the active border in order to avoid generating a disk I/O request for a window region that has already been processed by a block of the underlying spatial database that has already been retrieved. The active border is maintained as follows. First, a window block, say b, is generated by the window decomposer and a disk I/O request is issued to access the region of the underlying spatial database corresponding to b. Assume that b overlaps in space with block u in the underlying spatial database. Therefore, u is retrieved as a result of the disk I/O request corresponding to b. The spatial objects inside u are processed and thus there is no need to retrieve u again. As a result, the active border needs to be updated by block b or u depending on which one provides more coverage of the window region. Figures 4-8 illustrate the updating process of the active border. If u has a larger overlap with the unprocessed portion of the window than b (e.g. window block 1 and block p of the underlying spatial database in Figure 4, as well as window block 3 and block q of the underlying spatial database), then the active border is expanded using u's region Figure 5). If u is contained in b (e.g., window block 2 and block r of the underlying spatial database in Figure 4), then all the other blocks in the underlying spatial database have to be retrieved as well, and the active border is expanded by b's region (Figure 6). If the sizes of b and u are the same (e.g., window block 12 and block s of the underlying spatial database in Figure 4), then the active border is expanded by either one of them Figure 8). Notice that, if we were using Algorithm-1, window blocks 4, 8, 10, and 7 would still be processed and hence would generate four redundant disk I/O requests to retrieve blocks p and q. The generation of the maximal quadtree blocks inside a given window is controlled by procedure WINDOW RETRIEVE whose basic structure is given in Figure 9. WINDOW RETRIEVE scans the window row-by-row (in the block domain rather than in the pixel domain), and visits the blocks within it that have not been visited in previous scans 3 . For each visited window block, say b, the underlying spatial database is queried and a corresponding quadtree block, say q, is retrieved from the database. Procedures GEN SOUTHERN MAXIMAL and MAX BLOCK generate b's or q's maximal southern neighboring blocks (in fact, only the 3 Observe that we could have chosen to scan the window in a column-by-column fashion instead of row-by-row. The result is unchanged as long as the data structures for keeping track of the active border are reoriented appropriately. portion of q that lies inside the window will be used) according to the three cases presented earlier in this section, that relate the location and size of both b and q with respect to the query window. WINDOW RETRIEVE also makes sure that any of the remaining columns of row r that lie within b or q are skipped. For example, consider Figure 2, where five scans are needed to cover the 12 \Theta 12 window with maximal blocks. The first scan visits blocks 1, 2, and 3; the second scan visits blocks 12, 5, 6, and 9; the remaining scans visit blocks 14 and 11; 13; and 15. Notice that once blocks 5 and 6 have been visited, their columns (i.e., 2-5 in the window) have been completely processed. Also, observe that when block 1 is generated, block p of the underlying spatial database, which overlaps with block 1, is retrieved. As a result, window blocks, 4, 8, and 10 are skipped. This way, the algorithm can avoid reaccessing p by skipping all the window blocks that overlap with p. As a consequence, the southern neighbors of p (and not those of block 1) are generated by the algorithm. Procedure GEN SOUTHERN MAXIMAL generates the southern neighbors (maximal blocks) for each maximal block B generated by WINDOW RETRIEVE and that is not contained in another maximal block. There are a number of possible cases illustrated in Figure greater than or equal to B. Otherwise, the total width of blocks N 1 through Nm is equal to that of B. It is impossible for the total length to exceed that of B unless there is only one neighbor (see Figure 10b). Procedure MAX BLOCK takes as its input a window, say w, and the values of the x and y coordinates of a pixel, say (col,row), and returns the maximal block in w with (col,row) as its upper-leftmost corner. The resulting block has width 2 s , where s is the maximum value of i (0 - i - log T , where \Theta T is the size of the image space) such that row mod and the point (row lies inside w. Figure 11a gives the active border's most general form. The active border does not contain any holes (see Lemma 1 in Section 4 and thus Figure 11b corresponds to an impossible situation). When a block of the underlying spatial database, say q, is retrieved, the algorithm checks its size against the corresponding window block, say b. If q's size is larger than that of b, then the algorithm knows that q has to intersect one of the window's boundaries (see Lemma 1 in Section 4). We make use of this property here. Figure 3 shows the four possible cases where the block retrieved from the underlying spatial database intersects with one of the window boundaries. Each of the four cases must be treated separately by the algorithm. There is no need to maintain any data structures to explicitly store the northern portion of the active border since WINDOW RETRIEVE can handle this portion directly. During the first row-by-row scan of the window by WINDOW RETRIEVE, if a block of the underlying spatial database, say q, is retrieved that happens to intersect the northern boundary of the window (Figure 3a), then WINDOW RETRIEVE skips the window blocks in the current row scan that overlap with q. The portion of the southern boundary of q that lies inside the window is used to generate the southern neighboring blocks to be processed in the next scan. When block q of the underlying spatial database intersects only the southern boundary of the window (Figure 3d), then it also suffices for WINDOW RETRIEVE to skip all the window blocks that are adjacent to the window block that initiated q's retrieval. Although this seems intuitive, it is not straight-forward to see that all of the processing of block q by WINDOW RETRIEVE is localized in one part of the algorithm. In particular, although true, it is not directly obvious that all the blocks that overlap with q will be processed by WINDOW RETRIEVE at the same time so that they can be skipped. Thus as a result of this localized processing, there is no need to maintain any explicit data structures in this case either. If q intersects the western or eastern boundaries (Figures 3b and 3c), its overlap with the window creates a pocket-like region that needs to be stored in two separate lists, WestList or EastList, respectively. Each time a window block is generated, it has to be checked against the active border in order to make sure that the block is not covered by a previously retrieved block of the underlying spatial database. Below, we show how to perform this check in constant time. To facilitate our presentation, we represent both WestList and EastList as two one-dimensional arrays, each of length equal to the height of the window: WestList[r : r+n\Gamma1] where the height of the window is n and (r; c) corresponds to the x and y coordinate values of its upper-left corner. Figure 12b shows the border represented by each of the two arrays as a result of extracting an 8 \Theta 12 window from the underlying spatial database in Figure 12a. Let (r q ; c q ) be the location of the upper-left corner of q. If q intersects the west boundary of the window, then WestList[r q ] is set to the where the first component of the pair denotes the x coordinate value of q's east boundary while the second component (i.e., s q ) denotes the size of q. The pair represents the pocket-like region resulting from the intersection of q with w. Similarly, if q intersects the east boundary of the window, then EastList[r q ] is set to the Each time a window block is generated it has to be checked against the active border in order to make sure that the block is not covered by a previously retrieved block of the underlying spatial database. Notice that updating the active border only requires one array access (either updating WestList or EastList depending on whether q intersects the west or east boundaries of the window, respectively), while checking a window block against the active border takes only two array accesses (one access to each of WestList and EastList). Therefore, maintaining the active border, whether updating or checking, takes O(1) time. Observe that WINDOW RETRIEVE always generates maximal neighboring blocks, and a bounded number of non-maximal blocks. An example of this situation arises when processing blocks A-J in the first row of the window in Figure 13. Each of blocks B, can generate at most one non-maximal neighboring block. Even though these non-maximal blocks are generated, procedure WINDOW RETRIEVE skips them in the next scan since they are subsumed (i.e., contained) in the previously processed maximal block in the scan. For example, when scanning block K in Figure 13, blocks L, M, and N are skipped since they are contained in it. This is easy to detect because for each block we know the x and y coordinate values of its upper-left corner and its size. Proving that the algorithm is correct involves showing that every block of the underlying spatial database that overlaps with the query window is retrieved and processed by the algorithm. In order to prove this, we can structure our algorithm in the following way. The algorithm consists of two mechanisms: one for generating maximal quadtree blocks inside the window (also termed the window decomposition algorithm), and the other for retrieving blocks from the underlying spatial database and maintaining the active border. The active border keeps track of the blocks on the boundary of the window that have already been retrieved. This guarantees that each block in the underlying spatial database is not retrieved more than once. Our strategy for proving that the algorithm is correct is to separate these two mechanisms, show that each one is correct, and then prove that they interact properly. The algorithm has two cases. The first case arises when all the quadtree blocks of the underlying spatial database that overlap the window are smaller than or equal to the size of the smallest quadtree block in the window. The second case arises when this size criterion is not satisfied. In the first case the window decomposition algorithm will have to generate all of the maximal quadtree blocks inside the window and none will be skipped - i.e., each one causes a block of the underlying spatial database to be retrieved. In other words, there are no pockets and thus the arrays WestList and EastList are never updated or accessed. This means that the algorithm reduces to the window decomposition algorithm given in [3]. The window decomposition algorithm is proved correct in [3] and thus we will not address it here. However, we only state that proving that the window decomposition algorithm is correct involves showing that the execution of the algorithm generates a list of maximal blocks that lie entirely inside the window and that cover each point inside the window. In other words, each point inside the window is covered by one maximal block that is generated through the execution of the algorithm. The following two theorems are proved in [3]: Theorem 1: Each point inside a window is covered by one and only one maximal block generated by the algorithm. Theorem 2: The window decomposition algorithm generates all the maximal blocks inside the window and only maximal blocks, and hence is correct. We now address the second case where some of the blocks in the underlying spatial database are larger than the smallest block in the window - i.e., blocks of the underlying spatial database whose sizes are larger than the overlapping window blocks. We need to show that the interaction and maintenance of the active border with the window decomposition algorithm (1) guarantees that every block of the underlying spatial database that overlaps with the query window is retrieved and processed by the algorithm, and (2) does not interfere negatively with the window decomposition algorithm. From the complexity point of view, we prove, in Section 5, that every block of the underlying spatial database that overlaps with the window is retrieved only once. First, we use the concept of a maximal zone [3] to facilitate the presentation of the proofs. Assume a window having (c; r) as the x and y coordinate values of its upper-left corner with height w h (i.e., in the y direction) and width ww (i.e., in the x direction). First, let us look at the x direction. Processing along the width ww , we subdivide the window into p vertical strips with coordinate values of their upper-left corner . p is defined so that c An example of such a decomposition into vertical strips is shown in Figure 14a. The vertical strips are termed maximal columns. We now subdivide the window into horizontal strips in the same way. In particular, we have q horizontal strips with (c; r i as the x and y coordinate values of their upper-left corner where r r . q is defined so that r An example of such a decomposition into horizontal strips is shown in Figure 14b. The horizontal strips are termed maximal rows. Now we define the term maximal zones as follows. A maximal zone, say Z ij , is the region between the vertical strips (i.e., maximal columns) having c i and c i+1 as the x-coordinate values of their upper-left corner and the horizontal strips (i.e., maximal rows) having r j and r j+1 as the y-coordinate values of their upper-left corner where Figure 14c gives an example of decomposing a window into its maximal zones. Below, we state some propositions dealing with properties of maximal zones. Their proofs are straightforward, and we omit them in the interest of brevity. They are illustrated in Figure 14d. Proposition 1: Each maximal block inside the window is entirely contained in one and only one maximal zone. Proposition 2: All the maximal blocks inside a maximal zone are of the same size. Proposition 3: A maximal zone contains either one maximal block, or one row of maximal blocks, or one column of maximal blocks. Proposition 4: All the southern neighbors of a block lie in one maximal zone. Proposition 5: There exists a maximal column, say c k , inside the window such that, In other words, the sequence of distances between (width of) the maximal columns forms a monotonically increasing sequence followed by a monotonically decreasing sequence. An equivalent property exists for maximal rows. A useful invariant that holds during the execution of the window decomposition algorithm that also relates to maximal columns is stated below. Invariant 1: Each maximal window block and its southern neighbor window blocks that are both generated by the window decomposition algorithm always lie inside the same maximal column. In other words, the blocks inside a maximal column are processed independently of the blocks in other maximal columns inside the window. Put differently, although the algorithm scans the window row-by-row (in the block domain) and generates the maximal neighboring blocks to the south of each block encountered, there is no interaction between blocks of different maximal columns. We make use of this invariant to prove the lemmas below. Lemma 1: Assume that a block, say b, is a maximal block that lies inside the window w and overlaps with a block of the underlying spatial database, say q. If q is of greater size than b, then q must intersect with at least one of the boundaries of the window w Figure 3). Proof by Contradiction: Since b overlaps with q and b is smaller than q, then b is contained in q (by the definition of a quadtree decomposition of space). Assume to the contrary that the database block q lies entirely inside w. If q is of greater size than the window block b that overlaps with it, then b is not a maximal block since we can use a window block b 1 that contains b and that coincides with q as our new maximal block, which leads to a contradiction. 2 As a result, we deal with three categories of blocks of the underlying spatial database that intersect the window boundary: blocks that intersect the north boundary, blocks that intersect the east (west) boundary, and blocks that intersect the south boundary. Notice that the algorithm treats blocks that intersect both the west (east) and the south boundaries of the window as if they just intersect the west (east) boundary. On the other hand, it treats blocks that intersect both the north and west (east) boundaries of the window as if they just intersect the north boundary. Blocks intersecting the east or west boundary of the window receive the same type of processing and hence are considered as one group. We prove the correctness of the interaction of each category separately. Lemma 1 means that the active border does not contain any holes (see Figure 11b) since the query window is scanned row-by-row, and large-sized blocks of the underlying spatial database intersect only the window boundary. Therefore, storing only the outer boundary of the active border is enough. Lemma 2a: If a block of the underlying spatial database, say q, intersects the west (east) window boundary, then the east (west) boundary of q that lies inside the window must coincide with a boundary of one of the maximal columns of the window. Proof: We prove the lemma for the case when q intersects the west boundary of the window. The other case is similar. Assume the lemma does not hold - i.e., that q intersects the window boundary but that the eastern boundary of q that lies inside the window does not coincide with a maximal column of the window. Therefore, one of two possible cases must occur. These are illustrated in Figure 15. Both of the cases cannot happen since, by the definition of a quadtree decomposition, blocks cannot overlap in this manner. 2 An analogous lemma can be stated for blocks intersecting the north or south boundary of the window. Lemma 2b: If a block of the underlying spatial database, say q, intersects the north window boundary, then the south (north) boundary of q that lies inside the window must coincide with a boundary of one of the maximal rows of the window. Lemma 3: If a block of the underlying spatial database, say q, intersects the west (east) window boundary, then the part, if any, of the south boundary of q, say s, that lies inside the window must coincide with the north boundary of a maximal block inside the window. Proof: Assume that q intersects the west boundary of the window. From Lemma 2a, q's east boundary coincides with a boundary of a maximal column of the window, say c. However, other maximal columns to the west of c may intersect q as well (for example, in Figure 12, maximal column C 1 intersects block p of the underlying spatial database). If q intersects with no maximal columns other than c, then only two cases are possible (as illustrated in Figures 16a and 16b). Figure 16a cannot occur in a quadtree decomposition, while Figure 16b satisfies the Lemma. If q intersects with one or more maximal columns other than c, then s must coincide with a maximal row inside the window (Figure 16c) as the other case cannot exist in a quadtree decomposition (Figure 16d). Since a maximal row coincides with the north boundary of maximal blocks across the whole window, then this applies to q as well. 2 Lemma 4: If a block of the underlying spatial database, say q, intersects the west (east) window boundary, then the window decomposition strategy will only skip the window blocks covered by q while maintaining normal processing otherwise. In other words, up-dating the active border with q does not adversely affect the mechanism used for window decomposition. Proof: Assume that q intersects the west border of the window. By Lemma 2a, the east boundary of q coincides with a maximal column of the window. Therefore, the window decomposition mechanism will function properly to the east of q since, by Invariant 1, the block generation process works independently inside each maximal column. The portion of the south boundary of q, say s, that lies inside the window, is used by the algorithm to generate the new window blocks to the south of q. However, from Lemma 3, all parts of s coincide with the north boundary of a maximal block inside the window. Therefore, by applying a maximal block computation at s, the algorithm would still generate maximal blocks of the window to the south of q after skipping the ones inside q (and hence avoid retrieving q more than once by the overlapping window blocks). If the south boundary of lies outside the window, then the Lemma holds since no further processing to the south of q is needed. In addition, by Invariant 1, the window decomposition process to the east of q is not affected by q since the east boundary of q coincides with a maximal column.We now study the case where a block of the underlying spatial database intersects the south boundary of the window. We make use of the following lemma. Its proof is given in [3] (where it is Lemma 4). Lemma 5: All the maximal blocks arranged in a row inside a maximal zone are processed in the same iteration of the main loop of procedure WINDOW RETRIEVE. Lemma a block of the underlying spatial database, say q, intersects only the south boundary of the window, then q lies entirely inside one maximal column of the window. Proof: By Proposition 5, if q overlaps with more than one maximal column of the window, then either the size of q is not a power of two (a contradiction) or q must intersect with the east or west boundary of the window (a contradiction). Therefore, q lies inside one maximal column. 2 Combining Lemmas 5 and 6, we get the following result: Lemma 7: If a block of the underlying spatial database, say q, intersects only the south boundary of the window, then the window decomposition strategy will only skip the window blocks covered by q, while maintaining normal processing otherwise. Proof: By Lemma 6, q lies inside only one maximal column of the window. By Lemma 5, if one maximal window block, say b, results in retrieving q, then the rest of the window blocks in the maximal zone that lie in the same row as b, will exist in the same iteration of the main loop of procedure WINDOW RETRIEVE. Therefore, all of them can be automatically skipped by the algorithm once q is retrieved, and hence no additional data structure is needed to record q's retrieval. Since the south boundary of q is already outside the window, no further processing is needed to the south of q. The effect of this is that it results in skipping all the window blocks that overlap with q and that lie to the south of b up to the south boundary of the window. Also, by Invariant 1, q lies inside only one maximal column and hence does not affect other portions of the window decomposition mechanism.Lemma 8: If a block of the underlying spatial database, say q, intersects the north boundary of the window, then the window decomposition strategy will only skip the window blocks covered by q while maintaining normal processing otherwise. Proof: Since q intersects the north boundary of the window, q will be retrieved when the algorithm scans the first row in the window. In addition, q will be retrieved by the leftmost maximal window block, say b, that overlaps with q since scanning is from left to right. Therefore, all the window blocks to the right of b and that overlap with q are automatically skipped by the algorithm since all of them immediately follow b in TopList, the list of blocks to be processed. Processing of the algorithm resumes at the first window block to the right of q in the current row scan. By Lemma 2b, the part of q's south border, say s, that lies inside the window will coincide with a maximal row of the window. Since a maximal row coincides with the north boundary of maximal blocks across the whole window, this applies to s as well. Therefore, using s to generate maximal blocks to the south of q will resume regular processing of the decomposition algorithm as it results in generating legitimate maximal blocks of the window after skipping the window blocks that overlap with q. Therefore, q is retrieved just once by the algorithm without affecting the normal processing of the algorithm. 2 Combining Theorems 1 and 2 and Lemmas 4, 7, and 8 we get the following theorem: Theorem 3; Every block of the underlying spatial database that overlaps with the query window is retrieved by procedure WINDOW RETRIEVE and hence the algorithm is correct. Proof: By Theorem 1, maximal blocks of the window cover every point inside the window (without overlap). Therefore, if blocks of the underlying spatial database are smaller than the window blocks, then, by Theorem 2, the window decomposition algorithm will generate all the maximal blocks inside the window, and hence all the blocks of the underlying spatial database overlapping with the window blocks are retrieved. If some of the blocks of the underlying database, say D, are larger than the corresponding maximal window blocks, then by Lemma 1, each block, say q in D, has to intersect with some of the window boundaries. By Lemmas 4, 7, and 8, the algorithm will skip all but one of the maximal blocks of the window that overlap with q (this is because when one of the maximal blocks has to retrieve q, then the rest of the overlapping window blocks are skipped). Lemmas 4, 7, and 8 also show that the normal window decomposition mechanism is resumed after processing each block of the underlying spatial database that overlaps the 5 Complexity Analysis Based on Observation 1 (restated as Lemma 1) that relates the size of the query window blocks to the size of the underlying database blocks that they intersect, we are able to restrict the size of the active border, so that it has a worst-case space complexity of O(n) instead of O(n 2 ) for an n \Theta n query window. Analyzing the time complexity of our algorithm is a bit complex as there are two processes going on, and hence two ways of measuring it. The first is in terms of the blocks of the underlying spatial database that are retrieved (the I/O cost), while the second is in terms of the maximal blocks in the window, i.e., the window decomposition mechanism and the maintenance of the active border (the CPU cost). The CPU cost of the process of generating the maximal quadtree blocks in the window is computed as follows. First, we find the number of maximal quadtree blocks, say N , inside the window, and then compute the cost of generating each one of the maximal quadtree blocks, say T gen . The overall CPU cost T cpu is the product of these two terms, i.e., It is important to note that, usually, not all of the maximal blocks inside the window are generated. However, in the worst case, when none of the blocks in the underlying spatial database intersect the border of the window, all the maximal blocks inside the window are generated. It is known that the number of maximal quadtree blocks inside a square window of size n \Theta n is in the worst case ([8, 11, 27]). It remains to compute the cost of generating each maximal quadtree block comprising the window, i.e., T gen . This consists of the work, say Tm , to generate a maximal quadtree block, say B, and the work that is wasted, say Tw , in generating southern neighboring blocks of B that are non-maximal. Therefore, the total CPU execution time of the window decomposition algorithm is Given a point (x,y) in a T \Theta T space, there can be at most log T different blocks of size 2 i (0 - i - log T ) with (x,y) as their upper-left corner. We use a binary search through this set of blocks to determine the maximal quadtree block inside the window [3]. Thus Tm is O(log log T ). To compute Tw , we need to show that each maximal quadtree block inside the window is generated once, and that only a limited number of non-maximal blocks are generated. We say that the work required to generate blocks that are not maximal with respect to a particular window is wasted. Such blocks are ignored (i.e., bypassed) in subsequent processing. For example, the work in generating the southern neighbors of blocks B, C, D, F, G, H, and J (i.e., L, M, N, P, Q, R, and T, respectively) in Figure 13 is wasted. This is formulated and proved in the following two Lemmas. Lemma 9: Each maximal quadtree block inside window w is generated at most once. Proof: In Theorem 2, we proved that every maximal quadtree block inside window w is generated by the algorithm. To show that it is generated only once we observe that each window block processed by the algorithm generates only its southern neighbors. The facts that non-maximal window blocks are bypassed by the algorithm, and that maximal blocks do not overlap, mean that each maximal window block, say B, is generated as the southern neighbor of only one other maximal window block, say C. Note that this worst case only arises if WINDOW RETRIEVE generates all of the maximal blocks (i.e., none are Lemma 10: Each window block visited by the algorithm can waste at most O(log log T ) work in generating intermediate non-maximal window blocks. Proof: Assume that window block B generates wasted work. We show that this work takes O(log log T ) time. B can generate neighboring southern maximal blocks that are either smaller or larger. When the size of the neighboring block is greater than or equal to the size of B, then the algorithm takes O(log log T of whether or not it is wasted and the Lemma holds. When more than one southern neighboring block is generated (this number can be of the same order as the size of B), we need to show that all the generated southern blocks are maximal, and cannot be bypassed, i.e, they are not wasted work. We shall prove this by contradiction. Assume that B generates more than one southern neighboring block and that all of them are bypassed (i.e., not visited) in subsequent processing. It should be clear that due to the nature of the quadtree decomposition of space, either all of them are visited, or all are bypassed. Our assumption means that there exists a block C whose width is greater than the total width of B's southern neighbors. Let (B x ,B y ) and (C x ,C y ) be the locations of the upper-leftmost pixels of blocks B and C, respectively. Also, let B s and C s be the widths of blocks B and C, respectively. It is easy to see that the fact that B and C are maximal blocks that are southern neighbors of other visited maximal blocks means that C . The fact that C s ? B s means that the lower-rightmost pixel of C is at (C x which is in the window. Therefore, which is the lower-rightmost pixel of B's southern neighbor of equal size, say D, is also in the window. This means that D is B's neighboring southern maximal block. However, this contradicts the existence of more than one such block. Thus the assumption that all of the southern neighboring blocks of B are bypassed is invalid. Therefore, no work is wasted in generating B's southern neighbors in this case, and the Lemma holds. 2 Combining the results for Tm and Tw , and Theorem 3 means that we have proven the following theorem. Theorem 4: Given an n \Theta n window in a T \Theta T image, the worst-case CPU execution time for the algorithm is O(n log log T In order to compute the disk I/O execution time of the algorithm, say T io , we need to prove the following theorem. Theorem 5: Every block of the underlying spatial database that overlaps the query window is retrieved once, and only once, by WINDOW RETRIEVE. Proof: By Lemma 9, each maximal block is generated at most once. Let q be a block in the underlying spatial database and suppose that q overlaps the window. If q lies inside the window and is of equal or smaller size than the overlapping window block, say b, then q will be retrieved once by the algorithm when b is generated, and hence the theorem holds (notice that maximal blocks do not overlap). If q overlaps the window and if q contains more than one window block, then q will be retrieved by the first window block, say b, that encounters q. However, from that point onwards, all the window blocks that overlap q will be skipped and block q will not be retrieved again. By Lemma 1, q has to intersect one of the window boundaries. If q intersects the east or west boundaries of the window, then by Lemma 4 the active border (i.e., WestList and EastList) prevents block q from being retrieved again by the remaining window blocks that overlap q. Otherwise, if q intersects the north or south boundaries of the window, then by Lemmas 7 and 8 the algorithm skips the remaining window blocks that overlap q. Therefore q will be retrieved once, and only once. Hence the theorem also holds when q is of larger size than the overlapping window blocks. 2 Note that there is an onto relation between the set of blocks of the underlying spatial database that are retrieved by the algorithm and the set of maximal window blocks generated by the algorithm. This relation is only onto, rather than one-to-one onto, because a window block, say b, may overlap more than one block in the underlying spatial database (i.e., the overlapped blocks are smaller than b), in which case several blocks in the underlying spatial database will be retrieved. However, they will only be retrieved once. The actual disk I/O cost of the algorithm depends on how the quadtree is implemented. Assume that the underlying database consists of a total of K quadtree blocks and that M of these blocks are retrieved by the window query. Assume further that it spans a space of size T \Theta T . A pointer-based quadtree implementation may have an overall I/O cost as high as M log T as we must traverse at most log T pointers to access the relevant block in the quadtree. Using a pointerless quadtree representation such as a linear quadtree (e.g., [14]) where each leaf block is represented by a unique number which is stored in a -tree, the overall I/O cost is O(M log K) as the cost to retrieve each block is O(log K). 6 Empirical Results In this section, we study the performance of the two algorithms Algorithm-1 (which is based on the window decomposition algorithm [3]) and WINDOW RETRIEVE, which is proposed in this paper. The window decomposition part of the two algorithms has the same worst-case execution time complexity (i.e., O(n log log T )) as shown in Section 5. As a result, we only focus on comparing the I/O cost of the two algorithms. Figure 17 shows the results of experiments comparing the number of disk I/O requests (i.e., blocks retrieved) to answer a window query using Algorithm-1(labeled Old Alg)) with the number of disk I/O requests generated by WINDOW RETRIEVE (labeled New Alg). Our data consists of maps of the road network of the US provided by the Bureau of the Census. A sample map corresponding to Falls Church containing 640 line segments is given in Figure 18. The maps are represented using the PMR-quadtree [18, 21], a variant of a quadtree for storing vector data. The x-axis corresponds to the ratio between the window area and the area of the underlying spatial database which spanned a 512 \Theta 512 image. Experiments were run for the ratios .01, .001, .0001, and .00001. For example, the ratio .00001 corresponds to a 5 \Theta 5 window, while the ratio .01 corresponds to a 50 \Theta 50 window. For each such ratio, a set of 500 randomly positioned rectangles were generated. A window query is processed for each rectangle using both algorithms. The y-axis corresponds to the average of the disk I/O requests for each set of rectangles plotted on a logarithmic scale. Not surprisingly, use of WINDOW RETRIEVE does not lead to a great reduction in disk I/O requests for small window sizes (about 25%) since for both the window and the corresponding area in the underlying database the number of blocks is relatively small. However, for larger size windows, the reduction is much more pronounced, and, in fact, use of WINDOW RETRIEVE leads to an improvement of over one order of magnitude (e.g., a factor of 10). 7 Concluding Remarks An algorithm was presented for retrieving the blocks in a quadtree-base spatial database environment that overlap a given window. It is based on decomposing a window into its maximal quadtree blocks, and performing simpler sub-queries to the underlying spatial database. Each block in the underlying spatial database is only retrieved once. The algorithm is proven (analytically and experimentally) to lead to an improvement in disk I/O performance. The algorithm requires some extra space (on the order of the width of the window), to store the active border. It remains to consider how the algorithm can be adapted to handle spatial databases with non-disjoint objects (i.e., overlapping). Performance can be enhanced further by selecting a suitable buffering strategy for the underlying B-tree [1, 5]. In particular, if we can adjust the scan order of the window algorithm so that the window quadtree blocks are visited in Morton order, and accompany this with a most-recently-used buffer replacement policy, then this would guarantee that B-tree pages, both leaf and non-leaf pages, would be requested by the algorithm only once, and hence no redundant disk I/O requests would result. For more clarification on this issue, see [1, 5]. --R An algorithm and a cost model for window queries in spatial databases. Uniquely reporting spatial objects: yet another operation for comparing spatial data structures. Decomposing a window into maximal quadtree blocks. Hashing by proximity to process duplicates in spatial databases. Multidimensional window retrieval in large spatial databases. The ubiquitous B-tree The space efficiency of quadtrees. Advances in Spatial Databases - Fourth International Symposium Fundamentals of Database Systems. Analytical results on the quadtree decomposition of arbitrary rectangles. Adaptive grids for geometric operations. The BANG file: a new kind of grid file. An effective way to represent quadtrees. Efficient structures for geometric data management. R-trees: a dynamic index structure for spatial searching. The LSD tree: spatial access to multidimensional point and non-point data A consistent hierarchical representation for vector data. The grid file: an adaptable Applications of Spatial Data Structures: Computer Graphics The Design and Analysis of Spatial Data Structures. Spatial data models and query processing. Computing geometric properties of images represented by linear quadtrees. Storing a collection of polygons using quadtrees. The buddy-tree: an efficient and robust access method for spatial data base systems A formula for computing the number of quadtree node fragments created by a shift. --TR --CTR Ashraf Aboulnaga , Walid G. Aref, Window Query Processing in Linear Quadtrees, Distributed and Parallel Databases, v.10 n.2, p.111-126, September 2001
window block retrieval;active border;spatial databases;range query;design of algorithms;data structures;databases;quadtree space decomposition;clipping
594780
A Multi-Resolution Content-Based Retrieval Approach for Geographic Images.
Current retrieval methods in geographic image databases use only pixel-by-pixel spectral information. Texture is an important property of geographical images that can improve retrieval effectiveness and efficiency. In this paper, we present a content-based retrieval approach that utilizes the texture features of geographical images. Various texture features are extracted using wavelet transforms. Based on the texture features, we design a hierarchical approach to cluster geographical images for effective and efficient retrieval, measuring distances between feature vectors in the feature space. Using wavelet-based multi-resolution decomposition, two different sets of texture features are formulated for clustering. For each feature set, different distance measurement techniques are designed and experimented for clustering images in a database. The experimental results demonstrate that the retrieval efficiency and effectiveness improve when our clustering approach is used.
Introduction Satellite remote sensing development generated huge amount of image data during the last two decades. With the launch of the new breed of high spatial resolution satellite systems, the volume This research is partially supported by Xerox Corporation and NCGIA at Buffalo. of digital image data will significantly increase. Paralleling with the remote image sensing is the availability of Geographic Information System (GIS) data such as digital orthophotographs and digital elevation models that are available in pixel format and in sheer volumes. Thus, there is a great demand for effective retrieval mechanisms. In processing these geographic images, the ultimate interest is to recognize and locate specific contents. A content-based retrieval approach is paramount to organizing geographic images. In the design of such retrieval approaches, significant features must first be extracted from geographical data in their pixel format. The features are numerical measurements to characterize an image. These features can then be used to cluster and index the images for efficient retrieval. Methods have been developed for indexing and accessing alpha-numerical data in traditional databases. However, the traditional approaches to indexing may not be appropriate in the context of content-based image retrieval [CYDA88, RS91, BPJ93]. In this context, a challenging problem arises with many image databases, within which queries are posed via visual or pictorial examples (termed visual queries). A typical visual query might entail the location of all images in a database that contain a subimage similar to a given query image. Such a query is well suited to and most appropriate for geographic images in identifying specific geographic contents such as land use/land cover types. Texture in images has been an important aspect for geographic image classification and retrieval [Har79, Joh94, RW96, WH90], due primarily to the heterogeneity nature of geographic images and the association between the texture patterns and the geographic contents. The variety and complexity of geographic texture present a challenge as well as an impetus for seeking more effective approaches in feature representation. We have conducted experimental research on image data retrieval based on texture feature extraction [SZ97, RSZSM97, SZB97]. In most existing content-based retrieval approaches, feature vectors of images are first constructed (e.g., from wavelet transforms), which are then used to distinguish images through measuring distances between feature vectors. Our experimental results demonstrate that this type of approaches can be effectively used to perform content-based retrieval based on similarity comparison between images. However, we encounter a critical problem that the feature vectors of some semantically irrelevant images may be located very close in the feature space. Figure 1 presents two images of water and stone between which the distance of the feature vectors is very small, but these images are semantically not similar. Given a query such that its feature vector is located in the neighborhood of the feature vectors of the two given images, both images are highly possible to be retrieved together in response to the query. Thus, indexing alone, using R-tree or its variants based on the closeness of feature vectors in the feature space, sometimes may not provide satisfactory solutions. An effective clustering approach needs to be integrated into indexing techniques for efficient and effective retrieval of image data. Once the query is narrowed to a specific category, image retrieval can then proceed efficiently. (a) (b) Figure 1: Semantically different Images with similar feature vectors. Figure shows two arbitrary shape semantic clusters C 1 and C 2 in a two-dimensional feature space. We may assume that C 1 and C 2 represent the images of stone and water in the feature space. Consider the query image q which belongs to the cluster C 1 . If we only consider the closeness of feature vectors in the feature space to return relevant images, we may retrieve many images from cluster C 2 that are semantically irrelevant. Indexing trees such as R-trees are very efficient in retrieving the nearest neighbors of the query image, but since they do not consider semantics of images, the retrieved images might be semantically irrelevant. If we successfully classify the database images into different semantic clusters, a visual query can be quickly narrowed to a specific category. Therefore, combining clustering with R-tree related indexing, image retrieval can proceed both effectively and efficiently. x x x x x x x x x x x xx x x x x x x x x x x x x x x x x x x x x Figure 2: Two semantic clusters. In this paper, we present an approach to effective and efficient retrieval of images in large geographical image database systems. The goal is, given a query image of size n \Theta n pixels, to retrieve all the database images of size N \Theta N (where n - N ), that contain a portion similar to the query image. Database images have been decomposed into smaller segments (down to n \Theta n segments) offline. The query image will be compared to the corresponding segments (as will be discussed later). A clustering approach is designed to be incorporated into the existing indexing methods [SRF87, BKSS90, LJF94] to achieve the effectiveness and efficiency. Similar to most of other current clustering methods, our approach cannot distinguish overlapping clusters 1 . Thus, 1 The proposed method in [SCZ98] handles the overlapping clusters using heterogeneous features. we assume that different clusters do not overlap in the feature space. The semantic clusters can have arbitrary shapes or can be disconnected. We first choose a set of training samples to identify predefined semantics clusters (for example, residential or agriculture). These clusters include the semantically similar images. Then using a hierarchical method we find template images representing each cluster. This approach helps consider both semantics and feature vectors in classifying images. Using the multi-resolution property of wavelet transforms, we extract the features at different scales where only the coarse features are used in clustering. In our clustering approach, distribution of feature vectors in the feature space is also considered. Combining the proposed clustering with any existing indexing approach, our experiments demonstrate that retrieval from clustered database outperforms the methods that are based only on nearest neighbor retrieval. In addition, since the retrieval of images is narrowed to a specific category, image retrieval can proceed more efficiently. Since geographical images (specially airphoto images) are rich in texture, and have relatively well-defined clusters, it makes our approach, which extracts multi-resolution texture features and clusters the images, a good candidate to be used. The remainder of this paper is organized as follows. Section 2 presents the approach to multi-resolution texture feature extraction. Section 3 discusses the application of the feature extraction approach on image clustering and retrieval. Experiments are presented in Section 4. Concluding remarks are offered in Section 5. In this section, we will first discuss the importance of texture features in retrieval. We will then introduce wavelet transforms and the wavelet-based texture features extracted from images. 2.1 Texture Features In the current methods, spectral features (pixel values) are usually used to retrieve images from a geographical image database. However, spectral features alone may not be sufficient to characterize a geographical image because they fail to provide information about spatial pattern of pixels. One way of obtaining better classifications and retrieval is the use of spatial pattern recognition techniques that are sensitive to factors such as image texture and pixel content. In addition, using texture features of a group of pixels instead of individual pixels can reduce the information to be processed. Hence it decreases the required memory storage and computation for image retrieval and clustering. Because of the spatial nature of geographic images, texture has been used to distinguish among the images [RW96, Lan80, HK69, MM96, Joh94, GT83]. However, compared to purely spectral based procedures, texture-based retrieval systems have received limited attention in remote sensing in the past [LK94]. Different methods defining texture were proposed to model natural texture feature or to be used in classification. These approaches have included the use of Fourier transforms [SF86], fractals [Man77], random mosaic models [AR81, JMV81], mathematical morphology [DP89], syntactic methods, and linear models [CJ83]. However, most of these approaches have been used in the field of pattern recognition and have not been applied to the analysis of remotely sensed data [RW96]. One method in texture feature extraction that has been used for remotely sensed data is to pass a filter over the images and use the variance of the pixels covered by the filter as the value of the pixel in the texture image [AN91, BN91, Jen79]. Woodcock and Ryherd found that using the above local variance texturing technique in an adaptive windowing procedure resulted in an improved image for certain applications [WR89]. The location of the adaptively located window used to calculate local variance is the window with the lowest local variance of all windows that include the pixel. They then apply a multi-pass, pair-wise, region-growing algorithm to find spatially cohesive units, or "regions", in the image [RW96]. Johnson used a knowledge-based approach to classification of built-up land from multi-spectral data. The approach involves spectral classification, determination of size and neighbor relations for the segments in the spectrally classified image, and rule-based classification of the image segments into land-use categories [Joh94]. The implementation is slow, and performance problems were encountered due to the large number of objects being processed [Joh94]. Gray-level co-occurrence matrix, a matrix of second order probabilities, has been used in texture analysis [BL91, HTC86]. The element s(i; j; v) of the co-occurrence matrix is the estimated probability of going from gray-level i to j, given the displacement vector \Deltay). A set of statistical values computed from co-occurrence matrices, such as entropy, angular second momen- tum, and inverse difference moment have also been used [HSD73, PF91]. In practice, there are two problems regarding this approach. The first is that the co-occurrence matrix depends not only on the spatial relationship of gray-levels, but also on regional intensity background variation within the image. The second issue is the choice of displacement vector because the co-occurrence matrix characterizes the spatial relationship between pixels for a given vector v [WH90]. Wang and He presented a statistical method for texture analysis, where they define texture unit as the smallest complete unit that best characterizes the local texture in eight directions from a pixel. The distribution of texture units results in texture spectrum which can be used in texture analysis [WH90]. Since the local texture information for a pixel is extracted from a neighborhood of 3 \Theta 3 pixels, instead of one displacement vector used in the gray-level co-occurrence matrix, their method is more complete for the characterization of texture. They provided limited experimental results. The textural structures that can be described within a neighborhood are naturally limited to those which are observable within the size of neighborhood. Hence, features that are extracted based on measurements within a fixed size neighborhood (for example 3 \Theta 3 pixels window) have poor discrimination power when applied to textures not observable within the neighborhood because of wrong scale [BGW91]. All the above texture analysis methods share this common problem. This issue motivates the use of a multi-resolution method. Bigun et al formulated a general n-D least mean square problem whose solution facilitates the extraction of dominant orientation [BGW91]. In 2-D space, they used the linear symmetry along with the Laplacian pyramids approach, to produce feature images. The linear symmetry features directly measure the dominant orientation information in the frequency channels [BGW91]. QBIC is a content-based image retrieval system which retrieves images based on their color, shape, or texture In QBIC, texture features are based on modified versions of the coarseness, contrast, and directionality features proposed in [TMY78]. The distance between the three-elements feature vectors is measured by weighted Euclidean distance. However, new mathematical representations of image attributes are needed in QBIC [FSN 95]. Features that describe new image properties such as alternate texture measures or that are based on fractals or wavelet representations, for example, may offer advantages of representations, indexability, and ease of similarity matching [FSN Jacobs et al presented a searching algorithm for query images in an image database [JFS95]. They apply wavelet transform on query and database images. The wavelet coefficients are then truncated and only the coefficients with largest magnitudes are used in comparing images. They developed an image query metric which uses those truncated and quantized wavelet coefficients. According to them, wavelet coefficients provide information that is independent of the original resolution. Thus, a wavelet scheme allows queries to be specified to any resolution (potentially different to that of database images) [JFS95]. Hence they call their method multi-resolution image querying. Manjunath and Ma used Gabor wavelet transform to extract texture features and showed that it performed better than other texture classification methods [MM96]. They used mean and variance of frequency sub-bands to represent texture. In our method, we take advantage of multi-resolution property of wavelet explicitly. In our approach, as will be explained later, we apply wavelet transform to extract texture features. Based on the property that was mentioned in [JFS95], a query image can have a different size from that of a database image, which we can use. In addition to that, by applying wavelet transform, we can have information about images at different scales (resolutions) from fine to coarse at the same time. Our interpretation of the term (multi)resolution is different from that of [JFS95]. In their case, resolution indicates number of pixels in the image, and since number of pixels of the query image and the database image do not necessarily have to be the same, they call their method multi-resolution image querying. In our method, since we extract the image features at different scales or levels of details (based on number of levels in wavelet transform), we have a multi-resolution representation of the image. Having information about images at different scales (resolutions), we can control the desired level of accuracy in comparing images. For example, in clustering images where we roughly categorize images, we use only the coarse features, whereas in image retrieval, all the fine details of images will be used. However, the method proposed in [JFS95] uses a distance metric with fixed accuracy. Moreover, their method directly compares the m largest wavelet coefficients of images one by one. Thus, if the image is translated, so will the corresponding wavelet coefficients. Hence the result of comparison will be very different. In our method, we compare derived values (mean, variance, or number of edge pixels) from each of the corresponding frequency sub-bands, resulting in less sensitive comparison with respect to translation. The method by Jacobs et al has the advantage of being fast. However, it has the common problem of the nearest neighbor retrieval approaches that just rely on the closeness of the feature vectors without considering the semantics of images. As we explained in the previous section, such methods sometimes may not provide satisfactory solutions. In our approach, we consider the semantics of images through clustering which results in more effective retrieval. Feature extraction may be performed on either a whole image or the segments of images. We assume that images are decomposed into segments properly. Some decomposition approaches can be found in [SC94a, SC94b, RSZSM97, SRGZ97]. Upon the proper decomposition of images, the system extracts the features of all the segments of images. The query image will be compared to the corresponding segments of database images with the proper size. In our implementation, we used nona-tree decomposition [RSZSM97, SRGZ97]. Nona-tree decomposition is a block-oriented approach in which images (and their segments) are hierarchically decomposed into nine quadrants. The nine overlapping quadrants are uniformly distributed segments. If there are k levels in nona-tree, the size of its segments will be N 2 , N 2 , where N 2 is the size of the original image. A query image is compared to the smallest segments in nona-tree whose size is greater than or equal to the size of the query image. If the size of nona-tree's smallest segments is larger than that of the query image, the query image will not be compared to any of the segments. For example, if a nona-tree has segments of sizes 128\Theta128, 64\Theta64, and 32\Theta32, given a query image of size 64\Theta64, it will be compared to the 64\Theta64 segments. A query image of size 60\Theta60 will be compared to the 64\Theta64 segments but a 24\Theta24 query image will not be compared to any of the segments. In general, the segments generated by block-oriented decomposition methods such as quad-tree, quin-tree, or nona-tree may not perfectly align the portion which matches the query image. For example, in quad-tree, in the worst case, only 25% of the matching segment will be covered. However, in [RSZSM97, SRGZ97], we proved that nona-tree segments at least cover 56% of the matching segment. In average case, a larger portion of matching segment will be covered, and our experiments showed that nona-tree performs better than other block-oriented decomposition approaches [RSZSM97, SRGZ97]. 2.2 Wavelet Transform Wavelet transform is a type of signal representation that can give the frequency content of the signal at a particular instant of time. In this context, one row/column of image pixels can be considered as a signal. Applying a wavelet transform on such a signal decomposes the signal into different frequency sub-bands (for example, high frequency and low frequency sub-bands). Initially, regions of similar texture need to be separated out. This may be achieved by decomposing the image in the frequency domain into a full sub-band tree using filter banks [Vai93]. Each of the sub-bands obtained after filtering has uniform texture information. A filter bank based on wavelets could be used to decompose the image into low-pass and high-pass spatial-frequency bands [Mal89a]. We will now briefly review the wavelet-based multi-resolution decomposition. More details can be found in Mallat's paper [Mal89b]. To have the multi-resolution representation of signals we can use a discrete wavelet transform. We can compute a coarser approximation of input signal A 0 by convolving it with the low pass filter ~ H and down sampling the signal by two [Mal89b]. By down sampling, we mean skipping every other signal sample (for example a pixel in an image). All the discrete approximations A j , is the maximum possible scale), can thus be computed from A 0 by repeating this process. Scales become coarser with increasing j. Figure 3 illustrates the method. A A A G G Figure 3: Block diagram of multi-resolution wavelet transform. We can extract the difference of information between the approximation of signal at scale and j. D j denotes this difference of information and is called detail signal at the scale j. We can compute the detail signal D j by convolving A j \Gamma1 with the high pass filter ~ G and returning every other sample of output. The wavelet representation of a discrete signal A 0 can therefore be computed by successively decomposing A j into A j+1 and D j+1 for This representation provides information about signal approximation and detail signals at different scales. We denote wavelet representation of signal A 0 after K levels as . The idea of using multi-resolution property of wavelets in clustering is to use the features of the wavelet coefficients at the coarse scale levels. Corresponding to the lowpass filter, there is a continuous-time scaling function OE(t), and corresponding to the highpass filter, there is a wavelet !(t). The dilation equation produces OE(t), and the wavelet equation produces !(t) [SN96]. For example, for Haar wavelet transform with ~ [1= 2], and ~ 2], the dilation equation is and the wavelet equation is Figure 4 shows the Haar wavelet !(t).1t w(t) Figure 4: The Haar wavelet !(t). We can easily generalize wavelet model to 2 dimensions for images, in which we can apply 2 separate one-dimensional transforms [HJS94]. The image is first filtered along the horizontal (x) dimension, resulting in a lowpass image L and a highpass image H. We then down sample each of the filtered images in the x dimension by 2. Both L and H are then filtered along the vertical (y) dimension, resulting in four subimages: LL, LH, HL, and HH. Once again, we down sample the subimages by 2, this time along the y dimension. The two-dimensional filtering decomposes an image into an average signal (LL) and three detail signals which are directionally sensitive: LH emphasizes the horizontal image features, HL the vertical features, and HH the diagonal features. Figure 5-a shows a sample airphoto image. Figures 5-b,c, and d show the wavelet representation of the image at three scales from fine to coarse. At each level, sub-band LL (the wavelet approximation of the original image) is shown in the upper left quadrant. Sub-band LH (horizontal edges) is shown in the upper right quadrant, sub-band HL (vertical edges) is displayed in the lower left quadrant, and sub-band HH (corners) is in the lower right quadrant. Our feature extraction and clustering methods can use any appropriate wavelet transforms such as Haar, Daubechies, Cohen-Daubechies-Feauveau ((2,2) and (4,2)), or Gabor wavelet transforms a) b) c) d) Figure 5: Multi-resolution wavelet representation of an airphoto image: a)original image; b) wavelet representation at scale 1; c) wavelet representation at scale 2; d) wavelet representation at scale 3. [Vai93, SN96, URB97, MM96]. Applying wavelet transform on images results in wavelet coefficients corresponding to each sub-band. We can extract different features from wavelet coefficients of each of these sub-bands. Next subsection explains the features that we used in the experiments. 2.3 Image Features In clustering database images, different kinds of features such as shape, color, texture, layout, and position of various objects in images can be used. Texture in images has been recognized as an important aspect of human visual perception [TMY78, Har79, Joh94, RW96, WH90]. Tamura et al. [TMY78] have discussed various texture features including contrast, directionality, and coarseness. Contrast measures the vividness of the texture and is a function of gray-level distribution. Some factors influencing the contrast are dynamic range of gray-levels, polarization of the distribution of black and white on the gray-level histogram, and sharpness of edges. Directionality measures the "peakedness" of the distribution of gradient directions in the image. Coarseness measures the scale of texture. When two patterns differ only in scale, the magnified one is coarser. In a geographical image database we classify images into residential, agriculture, water, and grass clusters. These clusters have different textures. For example, their contrast is usually different; or, the cluster of water images does not have any directionality, whereas residential images are considered as directional images. These texture features of geographical images contain main visual characteristics of images that can be utilized in geographical image clustering and retrieval. Applying wavelet transform on images can provide information about the contrast of images. Also, different sub-bands generated by wavelet transform have information about horizontal, verti- cal, and diagonal edges in images [Mal89b], which helps in extracting features related to directionality of images. Moreover, multi-resolution aspect of wavelet transform provides information about images at different scales (from coarse to fine). These properties make wavelet transform a good candidate to extract texture features of geographical images. We calculated the mean and variance of wavelet coefficients to represent the contrast of the image. We also count number of edge pixels in horizontal, vertical, and diagonal directions to have an approximation of directionality of the image. These features are extracted at different scales to use the multi-resolution property of wavelets. Edge pixels in the image have high amplitude in detail signals (LH; HL, and HH). It means that the wavelet coefficient of such pixels has high positive/negative value. Let Image(i; j) be a pixel in the image of size n \Theta m. Let WLH (i; j), WHL (i; j), and WHH (i; j) be the corresponding wavelet coefficients of Image(i; j), for LH;HL, and HH sub-bands, respectively. We define mLH as the maximum of absolute value of WLH (i; j). That is, To detect horizontal edge pixels, we first compute mLH . We consider Image(i; as a horizontal edge pixel if where ff is a factor that is determined by experiments. In our experiments, we chose ff as 0.35. Similarly, we define and Image(i; is considered as an edge pixel if Tables 1, 2, and 3 show the wavelet coefficients of Figure 5-a at scale 3 using Haar wavelet transform. The corresponding edge pixels are shown in Figure 5-d. Table 1: Wavelet coefficients WLH of Figure 5-a at scale 3. Table 2: Wavelet coefficients WHL of Figure 5-a at scale 3. 28 26 43 -40 -45 28 Table 3: Wavelet coefficients WHH of Figure 5-a at scale 3. Figure The seven sub-bands generated after applying two levels of wavelet transform. We tested our approach using two different sets of features. To obtain the features in feature set 1, we apply a two-level wavelet transform. For the sub-band LL of the coarsest scale of the wavelet transform, we consider mean and variance of absolute value of wavelet coefficients as its features ?. For the other sub-bands, in addition to mean and variance, the system counts number of edge pixels in the sub-band. So the feature vector for each of these sub-bands is number of edge pixels ?. In this feature set, we will have seven sub-bands with features. Figure 6 shows the seven sub-bands of image schematically. In clustering, we just consider the features of the sub-bands of the coarsest scale of wavelet transform (11 features) to compute the distance. Table 4 shows the extracted feature set 1 for Figure 5-a. Number of Sub-band Mean Variance edge pixels Table 4: Feature set 1 of Figure 5-a. In feature set 2, the two-level wavelet transform is applied giving seven sub-bands. The feature vector for each sub-band is ! mean; variance ?. Each subsegment of the image has 14 features. To compute the distance between feature vectors in clustering, we only consider the features of the second scale of the wavelet transform (4 sub-bands, 8 features). We have extracted the feature sets using both Haar and Daubechies wavelet transforms [SN96]. Feature set 2 for Figure 5-a is shown in Table 5 where Haar wavelet transform is applied. Sub-band Mean Variance Table 5: Feature set 2 of Figure 5-a. To give equal contributions to all the features in determining the similarity, and to remove the arbitrary affects of different measurement units, we normalize the feature values and recast them in dimensionless units. To normalize the features, we first compute the mean - f i and variance oe i of each feature f i using the training samples. Our training samples include residential, agriculture, water, and grass areas. Let a feature vector be normalized feature vector be - f n ). We normalized f i using Equation (8): f i is the normalized value of f i . Our experiments show that we should give more (controlled) contribution to the more important features. We can do so, by assigning weights to the features and giving more weights to the more important features. Let vector, where w i is the weight that is assigned to feature f i . The weighted feature vector ~ f of feature vector f will be: ~ 3 Clustering and Retrieval Applying the feature sets introduced in Section 2, we will now present template-based image clustering and retrieval approaches. We first describe how cluster templates are chosen. Given clustering templates and features, images in a database can be classified based on their similarity to the cluster templates. Based upon the resulting clusters, image retrieval can be efficiently supported. 3.1 Clustering - Related Work Given a set of feature vectors, the feature space is usually not uniformly occupied. Clustering the data identifies the sparse and the dense places, and hence discovers the overall distribution of patterns of the feature vectors. In many existing clustering algorithms, k-medoid methods have been used, where each cluster is represented by the center of gravity of the cluster. For example, PAM (Partitioning Around Medoids) [KR90] was proposed to determine the most centrally located object medoid for each cluster. Each non-selected object is grouped with the medoid to which it is the most similar. CLARA (Clustering LARge Applications) [KR90] draws a sample of data set, applies PAM on the sample, and finds the medoids of the sample. Ng and Han introduced CLARANS (Clustering Large Applications based on RANdomaized Search) which is an improved k-medoid method [NH94]. Ester et al presented a clustering algorithm DBSCAN relying on a density-based notion of clusters which is designed to discover clusters of arbitrary shapes [EKSX96]. The key idea in DBSCAN is that for each point of a cluster, the neighborhood of a given radius must contain at least a minimum number of points, i.e. the density in the neighborhood has to exceed some threshold. BIRCH (Balanced Iterative Reducing and Clustering using Hierarchies) incrementally and dynamically clusters incoming data points to produce the best clusters with the available resources (i.e. available memory and time constraints) [ZRL96]. Ma and Manjunath used a neural network to learn the similarity measure and partition the feature space into clusters representing visually similar patterns [MM95]. As mentioned in Section 1, retrieval only based on closeness of feature vectors may not necessarily give satisfactory results, because it is possible that semantically irrelevant images have close feature vectors. Indexing trees such as R-trees and its variants have been successfully used in the past and are very efficient in retrieving the nearest neighbors of the query image in the feature space. However, they retrieve all the nearest neighbors, whether or not they have the same semantic as that of the query image. Our goal in clustering is to group together images considering their semantics. In other words, we want to cluster images such that images with the same semantics belong to the same cluster. These semantic clusters can be used to narrow down the search to the areas in the feature space where images have similar semantics to that of the query image. Searching only the semantically relevant images makes the retrieval more effective, and also more efficient. Note that, our proposed approach is not an alternative to R-tree or its variants. These indexing trees can be integrated within each cluster to provide better efficiency, while the clusters preserve the similarity in semantics. Semantics of images is application dependent, for example, in a GIS visual database, we can consider residential, water, grass, and agriculture clusters. 3.2 Template-based Clustering To support efficient content-based image retrieval, methods have been proposed to first cluster database images based on their similarity to predefined image templates, which are termed cluster templates or cluster icons [WAL To retrieve database images which are similar to or contain a query image, the query image will first be compared to the cluster templates. Only database images in the clusters corresponding to the chosen cluster templates will be searched. The cluster templates and their associated information (as will be explained later) approximate the clusters. An effective clustering of database images can greatly reduce the number of images to be searched, resulting in a more efficient retrieval. Indexing techniques can be incorporated within each cluster to provide a faster retrieval. R-trees can be used to find the templates representing the clusters. But since in practice, we have high dimensional feature vectors (20 features), R-trees might suffer from "dimensionality curse" [BKK96]. Thus, we decided to apply hierarchical clustering to find the cluster templates. Using separate training data for each cluster, we classify images considering their semantics. The x x x x x x x x x x x x x x x x x (a) (b) Figure 7: a) arbitrary shaped clusters. b) circular sub-clusters of clusters. type of semantic clusters is application dependent and should be determined by the application experts. For example, in a geographical image database, we can define the clusters of residential, agricul- ture, water, grass, and a special cluster other. The cluster other contains all images which cannot be included in any of the residential, agriculture, water, or grass clusters. We represent clusters by cluster templates to organize database images. In the feature space, since semantically different image data may have their feature vectors close to each other, each cluster may have arbitrary shape. Figure 7-a shows an example in two dimensional space where clusters have arbitrary shapes. Our proposed method is a hybrid (both supervised and unsupervised) approach. Using supervised classification approach, we choose a set of training samples for each cluster. Then using unsupervised classification we find the sub-clusters within each cluster. The centroids of those sub-clusters are used as the cluster templates of the cluster. Figure 7-b shows such centroids within each cluster as black square dots. Cluster other should include all the objects in Figure 7-a shown by " ". In a hierarchical approach, sub-clusters are defined within each cluster at different levels. Suppose the training set of a cluster C c has n objects. Initially, we assume there are n sub-clusters corresponding to the n objects in the cluster. We compute the distances between all pairs of sub-clusters. We then group together the two closest sub-clusters, resulting in We repeat the process of computing the distances between all pairs of current sub-clusters, and grouping the two closest ones until having only one sub-cluster. This hierarchically nested set of sub-clusters can be represented by a tree-diagram, or a dendrogram [Gor81]. Figure 8 shows a group of 5 objects and their corresponding dendrogram. Each of n leaves of the tree represents the single object whose label appears below it. Each position up the tree at which branches join, has an associated numerical value d, which determines the distance between the two grouped sub-clusters. The smaller the value of d, the more similar two sub-clusters will be. We considered the distance between the centroids of sub-clusters as the distance between sub-clusters. We call this method as centroid method. There are different methods to define the distance between sub-clusters. In unweighted pair-group method using arithmetic averages (UPGMA), the distance between two sub-clusters is the arithmetic average of the distances between the objects d d d2 (a) (b) Figure 8: A group of objects and the corresponding dendrogram. in one cluster and the objects in the other [Rom84]. In single linkage clustering method (SLINK), the distance is the minimum of the distances between the objects in one cluster and the objects in the other one (distance between two closest objects). The distance between sub-clusters in complete linkage clustering method (CLINK), is defined as the maximum of the distances between objects in the two sub-clusters [Rom84]. Applying either of these methods may give a different dendrogram. We can cut the dendrogram of each cluster at different levels to obtain k number of sub-clusters (cluster templates). The choice of k and the corresponding cutting distance is one of the most difficult problems of cluster analysis, for which no unique solution exists. For example, if we cut the dendrogram in Figure 8-b, where the distance between sub-clusters is d 1 , we will have two sub-clusters, whereas if we cut at level d 2 , we will have three sub-clusters. Let be the sub-cluster containing objects e 1 . If we cut a dendrogram at the level in which the distance between the last two grouped sub-clusters is d i , then the maximum distance between the children of the remaining sub-clusters will be d i . For example, in Figure 8-b, if we cut the dendrogram at level d 1 , the distance between sub-clusters f1g and f2g, or f3g and f4g, or f3,4g and f5g, will be less than d 1 . A measure of the quality of clustering can be obtained by computing the average silhouette width of the n image samples for each k value. Given an image i, which is assigned to a cluster C be the average distance of i to all other images in cluster C j and b i be the average distance of i to all images in cluster C h , where C h is the nearest neighbor cluster of cluster C j (h 6= j). The s k is calculated as follows: 1:0. The value of s k can be used to select an optimal value of k such that s k will be as high as possible. The maximum value of s k is known as the silhouette coefficient. In general, a silhouette coefficient value of 0.70 or higher indicates a good clustering selection [KR90]. Assume that a cluster C c consists of m sub-clusters S c1 In clustering database im- ages, whenever a sub-cluster S ci is chosen, the database images will be assigned to the corresponding cluster C c . Similarly, in image retrieval, if a sub-cluster S ci is chosen, only the corresponding cluster C c will be searched. We also tested different methods in determining the distance between sub-clusters, to make the dendrograms. 3.3 Clustering To cluster database images, we first compute the distance between the cluster feature vectors of cluster templates and all database images. Using the multi-resolution property of the wavelet trans- form, we extract the image features at different scales. In comparing cluster feature vectors, only coarse scale features will be used; while in retrieving images all the features will be compared. In Section 2.2, we used the notation fAK as the wavelet representation of signal A 0 after K levels of the wavelet transform. We define clustering scale (K; c) as the number of coarsest scales that will be used in clustering after applying K levels of the wavelet trans- form. It means that in clustering only the features extracted from fAK ; DK will be used, where 1 - K - J , 1 - c - J , and c ! K. For example, clustering scale (5, 2) means that the 5-level wavelet transform will be applied and the features of the 2 coarsest scales (fA 5 will be used in clustering. We define retrieval scale (K; r) as the number of scales that will be used in retrieval after applying K levels of transform. In retrieving images, the features extracted from will be used, where 1 - K - J , usually r ? c. In retrieving, we will use all the features. Clustering scale and retrieval scale for both feature sets 1 and 2 are (2,1) and (2,2), respectively. Let X and Y be two undecomposed images (for example, a query image, a cluster template, or an undecomposed segment of a database image). Let CFX be their cluster feature vectors extracted from the subbands fAK based on the clustering scale (K; c). Let RFX be the retrieval feature vectors of X and Y respectively, extracted from the subbands fAK according to the retrieval scale (K; r). To measure the similarity between X and Y in clustering, we define the clustering distance d c as follows: where dist could be any appropriate distance function. Here, we use Euclidean distance for dist. To measure the similarity between X and Y in retrieval, we define the retrieval distance d r as follows: d r (X; Y A database image Y can be decomposed into a set of smaller segments fY g. To measure the distance between the database image Y and the undecomposed image X (a query image, or a cluster template) in clustering, we use D c : Similarly, in retrieval, we use D r to compute the distance between the query image X and the database image Y: In determining the correct cluster(s) for a database image, we use the distribution of objects (feature vectors) in each sub-cluster. For each sub-cluster we define a scope. Let - i and - i be the mean and variance of clustering distances d c of objects to the centroid of the sub-cluster i. The scope of sub-cluster i is part of the feature space whose distance to the centroid of sub-cluster is less than Parameter fi is a factor that is determined by experiments. In a two dimensional space, the scope of a sub-cluster is a circle with radius In our clustering approach, whenever a database image Y falls within the scope of any sub-cluster, it will be assigned to the cluster corresponding to the sub-cluster. The reason is that a database image may contain the features of several different cluster templates and it should be assigned to all the corresponding clusters. If the database image Y is not within the scope of any of the sub-clusters, but the distance between the database image and the closest cluster template t is less than a threshold T , it will be assigned to the corresponding cluster represented by the closest cluster template. If there is no cluster template for which the distance between the database image and the cluster template is less than T , the database image will be assigned to the special cluster other. The cluster other contains all the database images that cannot be assigned to the predefined clusters. As an example, Figure 9 shows the scope of 3 sub-clusters A, B, and C in a two dimensional space. In general, these sub-clusters do not necessarily belong to different clusters. The object denoted by point 1 falls in the scope of sub-clusters A and B and will be assigned to both. x x Cx x x x x x x x x x x x x x A x x x x x3 Figure 9: Three sample sub-clusters. However, some objects may not fall in any of the scopes considered above. For example, Point 2 in Figure 9 is not in the scope of any of sub-clusters. Our clustering approach assigns such an object to the closest sub-cluster. Although the distance d c between point 2 and the centroid of sub-cluster C is less than that of sub-clusters A and B, but based on the distribution of objects within clusters, it is more logical to assign object 2 to sub-cluster A. So, we first compute the distance between the object and the border of the sub-clusters' scope, and then choose the closest cluster. Hence, object 2 will be assigned to sub-cluster A. The object shown as point 3 in Figure 9 will be assigned to the cluster other, because it is far away from all clusters and its distance d c to the closest cluster is greater than threshold T . Algorithm 1 shows the steps in choosing the cluster(s) for database images. After generating the templates, for each image, the time complexity of choosing clusters (either for adding the image to the clusters or for retrieving images) would be O(t) where t is the number of templates. The required time to assign database images to the clusters is O(tN) where N is the number of images. Since the number of templates is generally far less than the number of images, that is t !! N , the time complexity will be O(N ). That is, the time complexity of generating clusters will be linear in terms of number of database images. So it can process large image databases efficiently. As it will be shown in the experiments, just by clustering database images, the system avoids searching more than 75% of the database images. Algorithm 1 Choose Cluster To Add(Image) /* Based on distance between Image and subclusters, */ adds Image to the corresponding cluster(s). */ f /* compute the distance D c between all */ /* the subclusters and database image */ to Number Of Subclusters Adjusted Distance /* distance to the border of subcluster's scope */ /* find the index of the closest subcluster */ Closest Find Min(Adjusted Distance); if (Adjusted DistanceClosest f/* add to special cluster other */ Add Image T return; /* add image to "all" clusters that are close enough */ to Number Of Subclusters fif (Distance f/* add to the corresponding cluster of subcluster i */ Add Image T if (Cluster Added == FALSE) /* image not added to any cluster */ /* add to the closest cluster */ Add Image T 3.4 Image Retrieval To retrieve database images similar to or containing the query icon, the distance d c between query icon and cluster templates will first be computed. If the query icon is within the scope of any cluster template, the corresponding cluster(s) will be searched. To measure the similarity of query icon and the database images in the cluster(s), the distance D r will be used. If the query icon does not fall in scope of any cluster template, but its distance to the closest cluster template is less than a threshold T , then the cluster corresponding to the closest one will be searched. Otherwise, the cluster other will be searched to retrieve the relevant images. Given a good set of features, comparing with the nearest-neighbor search used in R-tree and its variants, retrieval on the clustered image database can potentially eliminate those images which are close to the query in the feature space, but semantically not similar. 4 Experiments As test data we used 1032 air photo images. To find the cluster templates representing each of residential, water, grass, and agriculture clusters, we used between 50 to 100 training image data per cluster. The training images were chosen such that the clusters do not overlap in the feature space. We tried the four different methods to compute the distance between sub-clusters: centroid method, CLINK, SLINK, and UPGMA. Figure 10 shows the sample cluster templates found by the hierarchical approach using these methods. Residential Water Grass Agriculture a) c) d) Figure images: a) Centroid method; b) CLINK; c) SLINK; d) UPGMA. It should be noted if additional terrains (such as jungle, or connifer forest) need to be added as new clusters, a set of training samples should be chosen for them. By applying the clustering method given in Section 3.2, the corresponding templates (subclusters) and their scopes can be determined. Using the new templates and the previously chosen templates (for the current clusters), database images can be easily reclustered. In other words, given new clusters (that are separate from current clusters), our system can accommodate them. To evaluate the performance of our clustering approach, we first visually inspected all database images and assigned them to their corresponding clusters. We consider these clusters as ideal clusters. We then compared the result of our clustering approach to the ideal clusters. Let C c be the set of images in the system's output for cluster c. Let I c be the set of images in the ideal cluster for cluster c. We define P c , the precision of clustering for cluster c as The recall of cluster c, R c , is defined as The higher P c and R c are, the better the clustering will be. As mentioned in Section 3.3, scope of a subcluster S i is determined by - . The values - i and - i are calculated by the system while generating the dendrogram. In our experiments a value between -0.5 to -1.5 for fi gives reasonable results. The threshold T is used to assign a database image (or search) in the cluster other. Based on our experiments, T - 2 max(- i ) can give satisfactory results. To find appropriate values for ff, fi, and T , we can first choose a sample set of database images with known clusters. We then determine the ideal clusters I c for the images in the sample set. After applying the clustering method on the sample data, precision and recall of clustering (P c and R c ) can be calculated. We can tune the values of ff, fi, and T by trail and error to obtain the desired P c and R c for the sample set. The optimal values of these parameters can then be used to cluster the whole database. In our experiments we chose 150 as the value of threshold T . We then tried the system for different values of parameter fi. In the experiments, we used the original extracted features (without any processing), normalized features, and weighted-normalized features. For weighted-normalized features, we assigned a weight of 4 to the first feature of the feature vector, and equal weight of 1 to the rest of features. We chose these weights on a trail and error basis. Assigning weights is an important issue that affects the results. The tables report the results when Haar wavelet transform is applied. Table 6 shows P c and R c for different methods of finding distance between sub-clusters when weighted-normalized features are used. A good clustering should have high recall and precision. But recall is more important than precision in clustering, because in retrieval from a cluster, we want to miss as few images as possible comparing to retrieval from the whole database. In our experiments, the results obtained by CLINK (maximum distance between objects in two sub-cluster) and UPGMA (average of distances between objects) are better than those of others. They yield both very high precision and recall of clustering Water 0.90 0.98 0.86 1.00 0.85 1.00 0.84 1.00 Agriculture 0.98 0.88 0.98 0.88 0.97 0.90 0.96 0.93 Residential Grass 0.83 0.86 0.85 0.83 0.86 0.82 0.88 0.81 a) Centroid Water 0.95 0.98 0.94 0.97 0.94 0.97 0.94 0.97 Agriculture Residential 0.98 0.95 0.99 0.94 1.00 0.93 1.00 0.92 Grass 0.94 0.91 0.93 0.93 0.91 0.93 0.89 0.92 Water 0.92 0.97 0.89 0.98 0.87 0.99 0.78 1.00 Agriculture 0.96 0.84 0.96 0.84 0.96 0.84 0.96 0.84 Residential 0.98 0.90 0.98 0.90 0.98 0.90 0.98 0.90 Grass 0.79 0.87 0.81 0.85 0.82 0.83 0.82 0.77 c) SLINK Water 0.90 0.98 0.86 1.00 0.85 1.00 0.84 1.00 Agriculture Residential 0.97 0.98 0.97 0.98 0.97 0.98 0.97 0.98 Grass 0.97 0.86 0.98 0.84 0.98 0.83 0.98 0.81 d) UPGMA Table Precision and recall of clustering for different values of fi. Weighted-normalized feature set 1. (more than 0.90). As Table 6 shows, varying the value of fi makes slight changes in the performance of clustering. To study the impact of clustering on effectiveness of retrieval, we chose 10 query icons of size 32 \Theta pixels. Our system processes queries such as "retrieve top N (say 50) images that have the same texture as the query icon". We calculated the effectiveness of retrieval in terms of retrieval \Gamma rate. Retrieval rate for top n images is defined as retrieval number of relevant images in top n images A high retrieval rate means that most of retrieved images are what the user looks for, and thus it shows a better performance in retrieval. We calculated retrieval rates for values 10, 20, 30, 40, and 50 of n for the 10 query icons and then found their average value. We first computed retrieval rate without using clustering. In this context, we consider nearest neighbor search approaches (such as R-tree or serial search) as without clustering methods. We then computed the retrieval rate when the clustering approach is used (with value 150 for threshold T and values 0.0, -0.5, -1.0, and -1.5 for fi). Different feature sets form different feature spaces. By normalizing features or assigning weights to them, the feature space changes which results in altering the shape of clusters or the location of query images. Also, different distance methods (centroid, CLINK, SLINK, UPGMA) result in different shapes of clusters. As a consequence, we get different performances from these various feature spaces and clusters. Given the appropriate feature vectors and distance method our proposed approach can outperform nearest neighbor methods in retrieval. Table 7 shows the average retrieval rate of top n images set 1. Table 8 has the same information for feature set 2. Tables 7 and 8 show that retrieval from clustered database (using all different distance methods) outperforms the retrieval from the whole database without clustering. In general, when a query image is near the boundary of the cluster, our method avoids retrieving the images outside the cluster (the ones that are semantically irrelevant) resulting in better performance. When the query image is in the middle of clusters, retrieving nearest neighbors (without clustering), and retrieval using our clustering approach would be the same. The retrieval rates obtained from CLINK and UPGMA are higher than those of others. It confirms with our previous clustering results where these two methods had higher precision and recall of clustering. In our experiments, we achieved the highest improvement in retrieval rate (from 0.70 to 0.95) when using original features set 2 and CLINK as shown in Table 8-a. Original feature set 1 provides a higher retrieval rate than that of feature set 2. Normalizing and then assigning weights to the features improves the retrieval rate when clustering is not used for both feature sets 1 and 2. When clustering is applied, weighted-normalized features provide a higher retrieval rate for feature set 2 (comparing to the original feature set), but not for feature set 1. In the latter case, another set of weights should be chosen. In general, a better method should be used to find the best weights for the features. For the same distance method and feature set, different values of fi do not make major changes in the retrieval rate. DB Ret DB Ret DB Ret DB Ret Centroid Clustering 1.00 0.73 1.00 0.73 1.00 0.73 1.00 0.73 a) Original features DB Ret DB Ret DB Ret DB Ret Centroid UPGMA b) Normalized features DB Ret DB Ret DB Ret DB Ret Centroid Clustering 1.00 0.76 1.00 0.76 1.00 0.76 1.00 0.76 c) Weighted-normalized features Table 7: DB-Ratio DB, and retrieval-rate Ret for different values of fi. a) Original features; b) Normalized features; c) Weighted-normalized features. (feature set 1). The other issue that should be considered is the number of images that are searched in retrieval from a clustered database. We define the ratio of database search DB \Gamma ratio as number of images searched in database total number of images in database The lower DB \Gamma ratio is, the better clustering would be. The time to search the database is directly proportional to the number of images to be searched. Thus a low DB \Gamma ratio means that less time is used to search the database. Tables 7 and 8 show the average DB \Gamma ratio for the 10 query icons. In average our clustering approach yields a DB \Gamma ratio of about 0.25, that is, about 75% of irrelevant images will not be searched. We performed the experiments on a SUN SPARC workstation using 168 MHz UltraSparc CPU with SunOS operating system and 1024 MB memory. The average time to search the whole database was 1.5 seconds, whereas after clustering, it only took about 0.23 seconds to return the relevant images. Integrating current indexing methods such as R-tree or its variants can speed up the retrieval process more. Figure 11 shows the 10 closest retrieved images (ordered from left to right) for a sample grass query image without clustering and using clustering, respectively. For the query image in Figure 11, DB Ret DB Ret DB Ret DB Ret Centroid 0.22 0.94 0.23 0.95 0.23 0.96 0.23 0.96 Clustering 1.00 0.70 1.00 0.70 1.00 0.70 1.00 0.70 a) Original features DB Ret DB Ret DB Ret DB Ret Centroid Clustering 1.00 0.71 1.00 0.71 1.00 0.71 1.00 0.71 b) Normalized features DB Ret DB Ret DB Ret DB Ret Centroid Clustering 1.00 0.78 1.00 0.78 1.00 0.78 1.00 0.78 c) Weighted-normalized features Table 8: DB-Ratio DB, and retrieval-rate Ret for different values of fi. a) Original features; b) Normalized features; c) Weighted-normalized features. (feature set 2) the first 10 retrieved images without clustering include four cases (images 6 through 10) which are not grass images. As shown in Figure 11-c, those semantically irrelevant images are not retrieved when our clustering approach is applied, which demonstrates the advantage of our method. a) Query icon Retrieved images without clustering c) Retrieved images with clustering Figure 11: Retrieval results 5 Conclusion We have designed a content-based retrieval system that supports novel clustering and retrieval for geographical images. This system utilizes the rich texture features existing in the geographical images to assist content-based retrieval. A clustering approach is designed to automatically categorize images based on themes such as land use or land cover. Using the multi-resolution property of wavelet transforms, we extracted the features at different scales where only the coarse features were used in clustering. Since the semantic distribution of feature vectors in the feature space is considered in the clustering, the retrieval of the relevant images can be performed by narrowing down to the relevant clusters. Thus, more precise image searching can be supported rather than the plain nearest-neighbor search. Through the experimental analysis, we have shown that texture features can be effectively used to cluster geographical images. In addition, our experimental results demonstrated that retrieval from the clustered database can be more effective than that of the whole database, given the appropriate feature sets and distance methods. Acknowledgments We would like to thank the reviewers of this paper for their constructive criticism and useful comments throughout the revision of the paper. --R Comparisons between spectral mapping units derived from spot image texture and field soil map units. Mosaic models for texture. Multidimensional orientation estimation with applications to texture analysis and optical flow. Sar sea ice discrimination using texture statistics: A multivariate approach. Seasonal variation of heterogeneity in the tallgrass prairie: A quantitative measure using remote sensing. A Visual Information Management System for the Interactive Retrieval of Faces. Markov random field texture models. An Intelligent Image Database System. A Density-Based Algorithm for Discovering Clusters in Large Spatial Databases with Noise Query by image and video content: The QBIC system. Classification Methods for the Exploratory Analysis of Multivariate Data. The use of contextual information in the classification of remotely sensed data. Statistical and Structural Approaches to Texture. Compressing Still and Moving Images with Wavelets. Pattern recognition with measurement space and spatial cluster ing for multiple image. Textural features for image classi- fication Use of texture operators in segmen- tation Spectral and texture features to classify elusive land cover at the urban fringe. Fast multiresolution image querying. Finding Groups in Data: an Introduction to Cluster Analysis. The development of a spectral-spatial classifier for earth observational data Remote Sensing and Image Interpretation. Multiresolution approximation and wavelet orthonormal bases of L 2 (R). A theory for multiresolution signal decomposition: the wavelet representa- tion Image indexing using a texture dictionary. Texture Features for Browsing and Retrieval of Image Data. Efficient and Effective Clustering Methods for Spatial Data Mining. Image texture processing and data integration for surface pattern discrimination. Analysis for Researchers. Automatic Image Indexation and Retrieval. Supporting Content-Based Retrieval in Large Image Database Systems Combining spectural and texture data in the segmentation of remotely sensed images. Transform Features For Texture Classification and Discrimination in Large Image Databases. Semantic clustering and querying on heterogeneous features for visual data. A Fourier-based textural feature extraction procedure Image decomposition and representation in large image database systems. An Approach to Clustering Large Visual Databases Using Wavelet Transform. Geographical Data Classification and Re- trieval Texture Features Corresponding to Visual Perception. Wavelet transforms using lifting scheme. Multirate Systems And Facial Images Retrieval A new statistical approach to texture analysis. Generation of texture image using adaptive windows. BIRCH: An Efficient Data Clustering Method for Very Large Databases. --TR --CTR Wei Wang , Daoying Ma , Yimin Wu , Aidong Zhang , David M. Mark, Webview: a distributed geographical image retrieval system, Proceedings of the 2002 annual national conference on Digital government research, p.1-4, May 19-22, 2002, Los Angeles, California Gholamhosein Sheikholeslami , Wendy Chang , Aidong Zhang, SemQuery: Semantic Clustering and Querying on Heterogeneous Features for Visual Data, IEEE Transactions on Knowledge and Data Engineering, v.14 n.5, p.988-1002, September 2002 Tao Li , Qi Li , Shenghuo Zhu , Mitsunori Ogihara, A survey on wavelet applications in data mining, ACM SIGKDD Explorations Newsletter, v.4 n.2, p.49-68, December 2002
texture features;multi-resolution wavelet transform;hierarchical clustering;geographical image retrieval
594788
An Object Model of Direction and Its Implications.
Direction is an important spatial concept that is used in many fields such as geographic information systems(GIS) and image interpretation. It is also frequently used as a selection condition in spatial queries. Previous work has modeled direction as a relational predicate between spatial objects. Conversely, in this paper, we model direction as a new kind of spatial object using the concepts of vectors, points and angles. The basic approach is to model direction as a unit vector. This novel view of direction has several obvious advantages: Being modeled as a spatial object, a direction object can have its own attributes and operation set. Secondly, new spatial data types such as oriented spatial objects and open spatial objects can be defined at the abstract object level. Finally, the object view of direction makes direction reasoning easy and also reduces the need for a large number of inference rules. These features are important in spatial query processing and optimization. The applicability of the direction model is demonstrated by geographic query examples.
Introduction 1.1 Modeling Direction Direction is a common spatial concept that is used everywhere in daily life. When people communicate about a geographic space, for example, giving route descriptions, direction is necessary to convey the information. Direction is also frequently used as a selection condition in spatial queries[11] or used for similarity accessing in image databases[18]. Example queries used in army battlefield visualization[10] are "Is there anything over the ridge?," "List the swamps in front of the tank."and "Let's move to the north of the tree." The first example refers to a viewer-based orientation, the second can be defined on either the intrinsic orientation of the tank(object-based) or of a viewer, and the third example refers to the absolute direction with respect to the tree. In order to process queries involving direction constraints in the selection criteria, a spatial database system should provide a way for users to formalize directions. The common means of handling direction is to model direction as a spatial relationship between objects[4, 20, 8, 26, 3, 16, 9, 6, 22, 15]. In this paper, we formalize direction from a different perspective: as a spatial object. The basic approach is to model direction as a unit vector and orientation as a set of directions. As a spatial object, direction can have its own attributes, and its own operators, and a richer set of predicates and operators on direction and orientation can be defined. Second, new spatial data types such as oriented spatial objects and unbounded spatial objects can be easily defined in the abstract object level. The object view of direction also makes it easy to perform direction reasoning by using only simple vector algebra, which also reduces the large number of inference rules commonly needed. This is useful in the processing and optimization of spatial queries that contain direction constraints. We model the direction for three frames of reference: absolute, object and viewer-based orientation We will not discuss the philosophical issue of whether directions are object/entities or not in the strictest sense. Absolute directions such as North, South, East, or West can be defined using a coordinate system w.r.t. the earth, and thus these can be considered to be objects in their own right. The object view for relative directions (e.g. left, front) needs further thought, because these directions are defined with respect to coordinate systems attached to other objects. The issue needs to be explored further. 1.2 Related work and our contributions The research work on direction modeling has been carried out in several areas such as geographic information systems and image analysis. Most of the studies is on how to capture the semantics of direction relations, and further, how to do spatial reasoning on the direction [3, 4, 6]. There are two major direction reference frames used to model direction in 2D space: the cone-based model[16], and the projection-based model[4, 6]. Frank[1] compared these two models and found the projection-based reference frame to be better in many aspects. The most common way to model directions between extended objects is through the object's Minimum Bounding Rectangle(MBR), where direction relations are obtained by applying Allen's [2] interval relations along the x and y axis, in which case, 169 different relations[3] can be distinguished. Some work based on MBR has been proposed on picture indexing in pictorial databases[18, 26], and some work aligns each boundary box to the object's major axis[14], which makes it possible to satisfy different reference frames[9]. On the other hand, Freska [5] proposed an alternative method: semi-intervals to formalize the one-dimensional temporal relation based on incomplete knowledge of the object. Goyal and Egenhofer [8] introduced a Direction-Relation Matrix to represent cardinal directions. Based on the projection-based frame, it partitions the space around the reference object and records into which direction tiles a target object falls. But this model still has limitations in the modeling of line objects, and it is limited to 2D space. Little work has been done on directions in 3D space [7]. The previous work modeled direction as a spatial relationship between spatial objects. This seems to be a natural mapping of the directional relationship that is used in geographic space. But this modeling method has some limitations. Operations on direction are limited. Oriented(directed) objects and unbounded objects cannot be represented in the spatial data model. This paper models direction as a spatial object: a unit vector. This novel view of direction has several advantages over the binary boolean relationship view. Being modeled as a spatial object, a direction object can have its own attributes and operation set. The implementation of operators can use vector algebra, making a richer set of predicates and operators on direction feasible. Secondly, new spatial data types such as oriented objects and unbounded objects can be defined at the abstract object level. The object view of direction also makes directional reasoning easy. Basic vector algebra is sufficient for inferencing new directions, and no special or new qualitative rules are needed. This reduces the complexity of directional reasoning, and involving appropriate quantitative information is useful in spatial query processing and optimization in spatial databases. More discussion on direction objects, formal proofs, and their applications appear in [24], where an algebra on the equivalence classes of direction objects is defined. The algebra unifies different methods for directonal qualitative reasoning that have appeared in previous work[4]. These issues are outside the scope of this paper which focusses on defining the basic concepts and implications of treating directions as objects. 1.3 Scope and Outline of the Paper The organization of this paper is as follows: In section 2 , we define a mathematical framework and propose new ADTs for directions and orientations. A new spatial data type hierarchy is also proposed in this section. In section 3, an oriented spatial object is defined using the orientation object, and its application in spatial queries is discussed. Open geometry and its applications are illustrated in section4. Finally, the paper concludes with discussions and recommendations for future work. 2 Basic Concepts 2.1 Points, Vectors, Angles The basic concepts we use here are points, vectors and angles, as illustrated in figure 1. A point has a position in space, which can be described using coordinates in a coordinate system. The only characteristic that distinguishes one point from another is its position. A vector, in contrast, has both vectors u and v are same, and Points P and Q are different, is the angle between vector v and t Figure 1: Diagram of points, vectors and angles magnitude and direction but no fixed position in space. An angle between two vectors represents the direction deviation from one vector to the other. Here, by angle, we mean the smaller one between two vectors, like ' in diagram 1. In this paper, we denote points by capitalized letters, such as P, and denote vectors as lower case letters with an arrow above, such as ~v. For simplicity we will use the Cartesian coordinate system to represent points and vectors in this paper. The Cartesian coordinate system in 3D is a right-handed rectangular coordinate system, with three axes x, y and z perpendicular to each other and intersecting at the origin. A point P in space is represented as an ordered triple (x; are rectangular components of P on the x, y and z axes , respectively. Let ~ U y and ~ vectors which are the basis vectors of the coordinate system. A vector ~v can be defined as a linear combination of these basis vectors in the U z , where a, b and c are some numbers. Operands Operations Definition U z U z \Theta ~a \Theta ~ b = jajjbj sin(') ~u, where ~u is a unit vector, perpendicular to ~a and ~ b scale ~ ~ ~ U z , where m is a number z a ), a point away from P with distance j~aj along the direction of ~a Table 1: Operations on vectors and points Table summarizes representative operations on points and vectors. Here P are two points, ' is the angle between vectors ~a and ~ b, where ~ ~ ~ U z and ~ ~ ~ ~ U z . We can see that the subtraction operation(-) is applied to point pairs, resulting in a vector. The operations of addition(+), subtraction(-), scale multiplication, dot product(fi), and cross product(\Theta) are available between vector operands. The definitions are given in terms of the components in the Cartesian coordinate system. There is only one operation(+) available between a point and a vector which actually produces another point. The associativity, commutativity and distribution properties of these operators will be examined in the future work towards integrating vector objects in query languages. 2.2 Direction and Orientation Direction is defined as a unit vector, i.e., a vector with magnitude equal to 1. Table 2 defines the operations on directions. Operations Definition composition ~ d2 deviation reverse (\Gamma1) \Theta ~ between 1 ~ d between ~ d 1 and ~ among 2 ~ d among ~ d 2 and ~ ~ ~ ~ ~ Table 2: Operations on Directions The operations on directions can be classified into three categories. The first category contains the operations that produce new directions. Composition and reverse are operations in this category. The composition operation is actually achieved by vector-addition, and the resulting vector is scaled down by its magnitude to be a unit vector, which represents the new direction. The reverse operator produces the reverse direction vector. The second category of the operations is to calculate the deviation between two directions. Operator deviation calculates the cosine of the angle between two directions, and hence gives the deviation of one direction from the other. A pair of vectors are orthogonal if their dot-product returns zero, i.e., they have 90 0 deviation. The last category of the operations is to test the relationships among directions. The operators between and among belong to this category. In figure 2, ~ d is between ~ d 1 and ~ d 1 is not between ~ d and ~ d 2 . As we will see in later sections, these three categories d d2 Figure 2: between operator of direction operations make the modeling of direction concise and flexible. Orientation is modeled as a spatial object which consists of a point of origin and N pair-wise orthogonal directions, where N is the dimension of the embedding space. The origin point and the N directions form a Cartesian coordinate system. In 3D space, the three directions may be labeled the Back-Front, Left-Right, and Below-Above directions of the orientation. Formally, we can define orientation and its operations in 3D space as follows: This definition works well as long as vector ~ d1 is not parallel to vector ~ d2 . The parallel case can be handled in a user defined manner. d 3 are not in the same plane. Orientation is a quadruple O= hOP, ~ front, ~ right, ~ abovei, where OP is a point, and ~ front, ~ right, ~ above are three orthogonal directions. It has two operations: front, ~ right, ~ right aboven i, where right aboven )= ( ~ right; ~ above)fi rotationMatrix; An example of the rotation Matrix which rotates the orientation along the ~ above axis for an angle ~ above Table 3 gives an illustration of these operations. ~ above above front right front right above right_n above_n above front right Table 3: Operations on orientation Abstract Data Types for Modeling Direction As a summary of the above discussion, the definition of ADTs for vector, direction and orientation is given in Table 4. The C++ like syntax is used here. The column labeled attributes declares the member variables of each class, and the representative operations column declares the interfaces of operations for each class. The vector class has three member variables of the real type. These variables represent the three coefficients when the vector is written as the linear combination of basis vectors of the coordinate system. The constructor constructs a vector object, given three real arguments. The five vector operators are declared as member operator functions. Direction is defined as a subclass of the vector class. It can be constructed either from a vector or from an ordered triple. Besides inheriting attributes and operators from the vector class, direction also adds new operators such as between and among. The orientation class has four member variables which form a Cartesian coordinate system. Translate and rotate are declared as the member operator functions of the orientation class. 2.3 Frame of Reference: Absolute, Object/Viewer-based Directions In this section, we will deal with point-based objects for simplicity. There are three different perspective systems. We define absolute directions(e.g. north, south), object-based directions(e.g. left, above), and ADT attributes representative operations vector x-comp: float; vector(float, float, float); // constructor z-comp: float; vector operator vector operator scale(); float operator dot-product (vector); vector operator \Theta (vector); direction(float, float, float); //constructor direction(vector); // constructor Direction inherited from vector direction operator Composition(direction); subclass of vector constraint = unit magnitude direction operator float operator deviation(direction); boolean between(direction, direction); boolean among(direction, direction, direction); Orientation // constructor ~ ~ right : direction; orientation operator rotate(rotate-matrix); ~ above Table 4: Abstract Data Types for direction and orientation viewer-based directions. Absolute Direction Absolute direction in embedding space is defined as a relationship among objects based on their locations in embedding space. For simplicity here, we will ignore the elevation and earth's curvature and use a projection-based model to map the space to a 2D local coordinate system with north up and east right 1 . We model absolute directions in two ways: constant direction objects and directional predicates. Constant direction objects are created for frequently asked directions, e.g: ~ east, ~ west, ~ north, ~ south, ~ NE, ~ SE, ~ SW . Table 5 illustrates how we can define the constant directions in terms of coordinates in local embedding space. The unit vector is denoted by an ordered pair(a; b) which represents vector a ~ U north . unit vector (1,0) (0,1) (-1,0) (0,-1) ( 1 directions ~ east ~ north ~ west ~ south ~ Table 5: constant absolute direction Introducing constant direction objects provides much flexibility in describing and deciding any di- rection. By using the operator deviation, we can calculate how much deviation is from one direction to another direction. We can also produce a new direction, given the deviation from a specific constant direction. This is very useful in the layout of facilities. Second, we also provide corresponding directional predicates to make it convenient for users to 1 For a global view, we can extend to the ellipsoidal coordinate system[13] specify the direction relation between two objects. They are East, W est, South, North, Northwest, NorthEast, SouthW est, SouthEast. For any two objects, assuming their centroids are P 1 , P 2 , we represent the relationship "P 1 is to the east of predicate The predicate is true if and only if equation ~ holds. Other predicates can be defined similarly. The above constant directions and predicates give precise directional calculation, that is, only one angle matches each case. Sometimes users may be interested only in approximate direction, such as a direction range: between north and northwest. The operator between makes this approximate directional predicate possible. If we want to test whether P 1 is between north and northwest of P 2 , we can use the between operator in C++ notation as: ~ north; ~ Object-based direction Object-based direction is the direction of the target object with respect to the orientation of the reference object. The reference object is an oriented object, while the target object may or may not have orientation. In figure 3, the person and the desk are oriented objects; the person is behind the desk and the desk is behind the person also. We will use OB to represent the orientation of object B, and right and OB : ~ above are the three directions of B's orientation. By using the direction operator reverse , the directions of B's behind, left and below can be described as \GammaO right above, respectively. The flag is to the left of the desk The desk is to the right of the flag Figure 3: Viewer/Object-based direction Given a target point object A, and a reference point object B, the direction of A relative to B can be calculated by using the deviation operation of the direction objects. First, a direction object ~ BA is constructed from the vector A \Gamma B. Then the direction of point object A with respect to point object B's orientation is decided by the deviation of direction object ~ BA from the three directions front, right and OB : ~ above of B's orientation. The vector dot-product is the meta operation for direction deviation operation. Table 6 illustrates the calculation of the object-based direction. The first column consists of the directional predicates and the remaining three columns are the conditions needed to satisfy the predicates. Direction predicates ~ BA front ~ BA right ~ BA above A in front of A behind of A right to A left to A above A between front(B) and ~ right) ~ right) ~ ~ ~ ~ ~ ~ ~ ~ right; ~ ~ right; A is among front(B),above(B) and left(B)= ~ right) ~ right) ~ right) ~ right) ~ ~ ~ ~ Table Object-based orientation The predicate returns true if and only if all three corresponding conditions hold. The first six rows illustrate the calculation of exact directions such as front, left, above, etc. As in row 1, if the dot-product of vector ~ BA and direction OB : ~ front is greater than zero, and the dot-products between vector ~ BA and the other two directions of orientation OB are both zeros , then means that A is in front of B. These directional predicates are for precise direction checking, whereas the other two categories of predicates are for direction range checking. The second category of directional predicates is to test if A is located in the region between two directions. Here the operator between is used in C++ notation. For example, the relationship that A is between front(A) and right(B) is represented as predicate ~ right), which has value true if and only if ~ BA front and ~ BA right are both greater than zero and ~ BA above is equal to zero. The predicates in this category are for the directions in 2D space. The third category of predicates consists of eight directional predicates which test if A is among three directions in a 3D space. We represent the predicates using the operator among. As can be seen from the table, by using direction object ~ BA, the checking of directional predicates converts to the calculation of direction deviation which is easily performed by the dot-product operation. Viewer-based direction Viewer-based direction refers to the directional relationship which is measured from the viewer's per- spective. In figure 3, the flag is to the left of the desk from the viewer's perspective. There are three related components in this system: target object A, reference object B, and the viewer. The viewer has his/her own orientation, whereas objects A and B may or may not be oriented objects. Given these three components, we can estimate the direction of object A relative to object B from the viewer's perspective, in a way similar to that used with the object-based system. The similar predicates but w.r.t. the viewer are defined and are calculated by the dot-product between vector ~ BA and the three directions of the reference orientation. The reference orientation here is the viewer's orientation O V , and hence the three dot-products needed to be calculated are ~ front , ~ right, and ~ above. A similar table can be obtained as in Table 7. Direction predicates ~ front ~ right ~ above ~ right) ~ right) Table 7: Viewer-based orientation In the viewer-based direction system, the viewer may change his/her orientation or position(unlike objects, which are still). Using the operators rotate and translate on the orientation, we can easily generate new orientations. Viewer-based directions are heavily used in route navigation. The route instructions can be given in terms of absolute direction, i.e., north, south, etc. But in fact, since many people have no idea of the absolute direction, particularly in an area new to them, route instructions in absolute direction have little meaning. The directions would be much easier to understand if the instructions were given by viewer-based direction. Many web routing servers such as MapsOnUs use as many viewer-based directions as possible in their route instructions. Figure 4 is an example of a route map that starts from the Computer Science Building of the University of Minnesota to a UM apartment on the St. Paul campus. The turn-by-turn instructions are given in table 8. As can be seen from the table, viewer-based directions are used at every turn to make it easy for users to follow the descriptions intuitively. Como AVSE Como Av Washington Av Knapp Av E24-5 Raymond Av University Avenue ( US52) Figure 4: Route Planning Go And Then Total miles Start Head South on Union ST SE, From Start Union ST SE, Mpls, MN)1 Less than .1 mi Turn LEFT onto Washington Ave. SE 0.0 3 1.4 mi Turn LEFT onto RAYMOND Ave. 1.9 4 1.1 mi Continue onto N Cleveland Ave 3.0 End Less that .1 mi End Point(2051 Knapp Ave, St.Paul, Table 8: Turn-by-turn Directions 2.4 New Spatial Data Type Hierarchy Some work has been done on modeling space into a spatial geometry hierarchy. The Open GIS Consortium[12] proposed a spatial object hierarchy for incorporating 2D spatial ADTs in SQL, as illustrated in figure 5. This hierarchy consists of P oints, Curves, and Surface and the class GeometryCollection. The basic operations are confined to topological and metric operations. Since these data types model only bounded shape objects, the hierarchy has limitations in modeling objects; for example, maps, rivers, roads and buildings which have their own directions/orientations. For example, with few exceptions, printed maps are all oriented with north up. This is not only for better visualization, but for more flexibility with spatial queries. It allows queries such as "Which state is to the left of Minnesota?" to be asked by non-geographers who may not use geographic concepts like north, west, etc. Spatial queries such as " Is Minnesota in the north part of the USA?" can be answered by comparing the location of Minnesota with the center of the USA and its direction to the center, without the necessity of getting the information on the boundary of the USA. But in the OGIS hierarchy, spatial objects can only be modeled as bounded objects, so north part of USA should be represented in a polygon which needs accurate boundary information that actually has little effect on the query. A new spatial data type hierarchy is proposed in figure 6. This hierarchy includes oriented objects and unbounded(open) objects, which could be defined using the direction and orientation objects discussed MultiLineString MultiPolygon LinearRing Line MultiPoint MultiCurve MultiSurface Polygon LineString Curve Point SpatialReferenceSystem Geometry GeometryCollection Surface Figure 5: Spatial Data Type Hierarchy[12] above. In the new spatial hierarchy, maps and roads can be defined as oriented objects. New Spatial Data Type Hierarchy Figure 6 illustrates the extended spatial data type hierarchy which consists of shape objects, vectors and oriented(directed) objects. The spatial data types and their operators can be embedded in an extended query language to implement a spatial query language. Vector object Oriented Spatial Surface Curve Solid direction Geometry Orientation open geometry open line open rectangle oriented directed line oriented solid Point closed geometry point classical type Figure Extended Spatial Data Type Hierarchy Many new operators are available in the extended spatial object model. Between vector objects and shape objects, affine transformations can be performed. The addition of these operations makes some GIS applications easy. For example, Urban planning needs interactive placement/orientation of facilities, location-based query exploration can use affine operations, and using these operations can also improve viewer flexibility, so that we do not have to look down on earth from space. Instead, a viewer can view from any direction by performing rotation transformations. In the later sections, the new spatial data types are discussed in detail. Their applications in geographic space are also described. 3 Defining Oriented(directed) Spatial Object There are many objects in geographic space that have intrinsic directions/orientations in addition to their locations and shapes, such as buildings, rivers and roads. As we discussed above, modeling maps as oriented objects provides users more flexibility in query specification. It also allows users who have little knowledge of geographic space to query directions intuitively, based on their own orientations. As in the routing example given in figure 4, the turn-by-turn directions in table 8 mostly use viewer-based direction to make the information more meaningful to non-geographical users. But since there is no information on the orientation of the starting address, the first directional instruction is based on absolute direction SOUTH. If we know the orientation of the exit of the building, then we can use viewer-based direction instead, since in this case the viewer's orientation is aligned with the building exit's orientation. The instruction then may be "Turn Right onto Union Street SE", which is very easy for any user to follow. Table 9 shows some queries that can only be answered when the reference objects and/or viewers are modeled as oriented objects. Queries List the swamps in front of the tank. List the farm fields suitable for tank movement, and are left to Lake A as viewed by Viewer 1. Let's move back a little. How far is the next intersection down this road? Table 9: Queries Need Orientation Information Introducing the Orientation class into the spatial object hierarchy makes it easy to formalize oriented objects. In addition to the attributes such as location and shape for the geometric shape objects, an oriented object has an attribute orientation, which is an instance object of class orientation. The operators available for oriented objects include both topological operators and direction and orientation operators. The ADT for oriented objects can be specified in Oracle8.0[25]. We first define the direction object type and orientation object in Oracle 8.0. In the following, we assume Geometry is the spatial type defined by OGIS[12]. It could be a point, a line, a polygon , and etc. x NUMBER, y NUMBER, z NUMBER, MEMBER FUNCTION composition(aDirection IN DirectionType) RETURN DirectionType, MEMBER FUNCTION reverse RETURN DirectionType, MEMBER FUNCTION deviation(aDirection IN DirectionType) RETURN NUMBER, MEMBER FUNCTION between(aDirection IN DirectionType, bDirection IN DirectionType) RETURN NUMBER, PRAGMA RESTRICT REFERENCES(default, WNDS) op PointType, d1 DirectionType, d2 DirectionType, d3 DirectionType, MEMBER FUNCTION translate(aVector IN VectorType) RETURN OrientationType, MEMBER FUNCTION rotate(degree IN NUMBER) RETURN OrientationType, PRAGMA RESTRICT REFERENCES(default, WNDS) OrientedObject AS OBJECT ( shape Geometry, orientation OrientationType, MEMBER FUNCTION left(obj IN Geometry ) RETURN NUMBER, /* other functions can be defined similarily */ PRAGMA RESTRICT REFERENCES(default, WNDS) name VARCHAR2(20), orientation OrientationType, MEMBER FUNCTION left(obj1 IN Geometry, obj2 IN Geometry) RETURN NUMBER, /* other functions can be defined similarily */ PRAGMA RESTRICT REFERENCES(default, WNDS) We can use these data types to solve spatial queries involving oriented objects. As an example, we will show how to write the first two queries in table 9 using Oracle 8.0. The first query contains object-based direction and the second one focuses on viewer-based direction. First a database schema is created , which consists of four tables, FarmF ield, LandCover, Tank, and V iewer. The queries are performed on this database schema. extent Geometry); extent Geometry); extent OrientedObject); List the swamps in front of the tank. SELECT L.name, L.extent FROM Landcover L, Tank T ffl Explanation:front is an object-based direction predicate with respect to the oriented object Tank. Landcover is a partition of the space based on landcover type. We search for partitions which lie in front of the tank and check their type. Example of possible types include "swamps", "water", "hill". List the farm fields suitable for tank movement, and are left to Lake A as viewed by Viewer1. SELECT F.name, F.extent FROM FarmField F, Landcover L, Viewer V WHERE V.left(F.extent, L.extent)=1 AND ffl Explanation: A farm field is suitable for tank movement if the crop growing inside is soft and low. The type field in the FarmF ield table identify this property. left is a viewer-based direction predicate with respect to the orientation of a viewer. The database will search for FarmFields that are left of the lake from viewer1's perspective, and list all the satisfying fields suitable for tank movement. 4 Open Geometry Open objects mean those geometries whose boundaries are partially defined, or extending beyond the data window, or infinite. To think of a query example first: Query: Given the map of Minnesota as our data window, what's the length of the Mississippi river? Questions: What should the GIS system answer? (a) The length of the Mississippi river inside Minnesota. (b) At least the length of the Mississippi river inside Minnesota, but I don't know exactly. Since the Mississippi river extends outside Minnesota, we cannot compute the exact length of the river given only the Minnesota map, so we prefer answer (b). Modeling the river as a closed object can only return answer (a) which is not precise; however, if we model the river as an open object, then answer (b) will be obtained, which is more accurate. 4.1 Categories of Open Geometry Table categorizes the open objects based on two factors: the extent of the object and our knowledge of the object. Figure 7 gives examples of each category of open objects. Knowledge Open Object within window beyond window Extent Open-G complete none finite Open-I complete complete infinite Open-P partial partial Table 10: Defining open objects If the object is finite and its boundary is well defined within the data window, then this is a closed object. Type Open-G refers to the object that has a finite boundary but extends beyond the data window; and the portion of the object within the data window is fully known, but the portion outside the data window is only partially known. This type of open object is useful in GIS applications. For example, if we consider the map of Minnesota as our data window, then the Mississippi river falls into the category of Open-G. Open-I refers to the object that has an infinite boundary, but we have complete knowledge of the object either within the data window or outside. Even though it seems of little interest to Geographic science since the earth has a finite surface, the category is useful in modeling direction for extended objects, as we will discuss in a later section. Moreover, Open-I and Open-G have similar closed Open_I Open_G Open_P Data window Figure 7: Examples of open shapes topological properties. Type Open-P refers to an object of which we have only incomplete information. Dealing with this kind of object is out of the scope of our discussion. In this paper, we will focus on the category of Open-G and Open-I, and will use the term Open object in the places where difference between the two is unnecessary. to be differentiated. 4.2 Geometric and Topological Operations on Open Objects In this subsection, we will examine the effects of open objects on geometric and topological operations. Table 11 and table 12 give the results for basic geometric and topological operations, respectively. Here "Closed" represents a closed object and "Open" refers to an open object. The first column of each small table contains the first argument type, and the first row contains the second argument type. For example, Difference(Closed, Open. If the two arguments of difference are Open objects, the result then could be either a closed object(If the two open objects extending to the same space) or an open object. Intersection Closed Open Union Closed Open Difference Closed Open Closed Closed Closed Closed closed open Closed Closed Closed Open/ Open/ Open Closed Closed Open Open open Open open Closed Table 11: Domains and Ranges for Geometric Operations on Open Objects ETDOC Closed Open Within Closed Open Closed boolean boolean Closed boolean boolean boolean/ boolean/ Open boolean don't know Open False don't know Table 12: Domain and Ranges for Logical Operations on Open Object. ETDOC stands for one of Figure 8 illustrates the four cases of operation Difference with an open shape involved. The pattern parts represent the difference between two objects. It is worth to mentioning here that since the open shapes E and F extend to the same region, so Difference(E; F ) is a closed shape. A A-B C-D X-Y F Data window Figure 8: Operation Difference on open shapes In table 12, we combine operations that have the same results into one table. Here ETDOC represents one of the operations in the set of fEqual, Touches, Disjoint, Overlap, Crossesg. The result of the operations could be either boolean or don't know. don't know is a special value which has properties similar to the null value in a database. A useful observation is that topological operations involving closed objects always have boolean results. 4.3 Defining Open line, Open Rectangle Examples of open spatial objects are open lines, open planes, and open spaces. A common way to define these in 3D Cartesian geometry is in terms of equations such as the parametric equations [17]: ct for a line through the point and Ax+C=0 for a 3D space. It would be more convenient if there were corresponding object classes so that users could describe them at an abstract level and use them to describe the space. The direction object could be used for this purpose. We define open lines and open rectangle regions here using directions and points. Table 13 shows how to formalize each open object with examples in figure 9. (d) (c) (e) (b) (a) Figure 9: Examples of open shapes There are several subtypes in each type. For an open line, it could be 1-end open(figure 9a), which needs a start point and a direction to define it, or it could be 2-end open(figure 9b), which needs an intermediate point and a direction to describe it. The open rectangle also has three subtypes: a 1-side open rectangle (9c), described by two vertices of the rectangle and one direction which is perpendicular to the edge connected by the two vertices. A 2-side open rectangle is described using two 1-end open lines which start from the vertex of the rectangle. Also a 3-side open rectangle could be defined by a Type Subtype Descriptions Examples open line openline1(1-end open line) start-point, direction (a) openline2(2-end open line) intermediate-point, direction (b) openrect1:1-side open rectangle vertex 1 ,vertex 2 ,direction, where direction perpendicular to segment formed by the two vertices (c) open rectangle openrect2(2-side open region) two 1-end open lines from the only vertex (d) openrect3(3-side open rectangle) 2-end open line, direction (e) Table 13: Defining open objects 2-end open line and a direction. It is obvious that extending to an open cube is easy. Abstract Data Type for open lines and open rectangles The ADTs for open objects are given in table 14 using C++ notation. The entire open object class is defined as a subclass of geometry. Geometric and topological operations available for open objects include openUnion, openIntersection, openDifference, openWithin, openEquals, openTouches, openDis- joint, openOverlap, and OverCrosses. The definition of the operators should be overloaded for open objects, according to table 11 and table 12. Here, we only show the attributes and constructors of each open object class, and as an illustration, we represent the spatial objects in figure 9. ADT attributes constructors examples in figure 9 line Table 14: Abstract Data Types for Open Lines and Open Rectangles 4.4 Modeling Direction Between Non-point Objects using open objects We can use open shapes to model the directions between extended objects by converting the calculation of directional relationships to the calculation of topological relationships between objects. Given two objects TO and RO, we want to decide the direction of target object TO related to reference object RO. Using the approach in [8], first we obtain the MBR of object RO and partition the space around object RO into nine direction tiles based on the MBR of object RO, which is a rectangle ABCD as in figure 10. We represent each direction tile as a spatial object, as in table 15.0000000000001111111111111111 O TO RO Figure 10: Object TO and RO Rectangle representation for figure 10 O rectangle(A,B,C, D) Table 15: Direction tile Eight of the nine direction tiles are open rectangles. NW, NE, SE, and SW are 2-side open rectan- gles, and N, W, E, and S are 1-side open rectangles. In order to test the direction of object TO related to object RO, we can test into which direction tile TO falls. The calculation is then converted to the overlap relationship between object TO and the direction tiles. For example, The other directional relationship can be tested similarly. As noticed in table 12, the topological operations involving closed objects produce boolean values. The union of all the tiles that overlap with A is the region where the target object TO is located w.r.t. the object RO. For example, in figure 10, TO is North and Northwest and West of object RO. In this paper, we model direction as a new kind of spatial object using the concepts of vectors, points and angles. The basic approach here is to model direction as a unit vector. This novel object view of direction has several obvious advantages: It allows the user to define the orientation of spatial objects; it also gives a richer set of predicates and operators on direction and orientation. New spatial data types such as oriented spatial objects and unbounded spatial objects can be defined as well. Finally, direction reasoning, which is useful in spatial query processing and optimization, can be implemented simply by basic vector algebra. The benefits of the new viewpoint are obvious and promising. In future work, we would integrate vector abstract data types with an extensible query language to evaluate their efficiency in the context of various GIS applications[21, 23]. We would also like to explore the use of vectors to model values of spatial attributes such as wind-velocity, the magnetic field of the earth, etc. Acknowledgments This work is sponsored in part by the Army High Performance Computing Research Center under the auspices of the Department of the Army, Army Research Laboratory cooperative agreement number DAAH04-95-2-0003/contract number DAAH04-95-C-0008, the content of which does not necessarily reflect the position or the policy of the government, and no official endorsement should be inferred. This work was also supported in part by NSF grant #9631539 Thanks to to Christiane McCarthy for helping to improve the readability of the paper. --R Qualitative Spatial Reasoning: Cardinal Directions as an Example. Maintaining Knowledge about Temporal Intervals. Hierarchical Reasoning about Direction Relations. Qualitative Spatial Reasoning about Cardinal Directions. Temporal Reasoning Based on Semi-Intervals Using Orientation Information for Qualitative Spatial Reasoning. Basic Meanings of Spatial relations: Computation and Evaluation in 3D space. The Direction-Relation Matrix: A Representation of Direction Relations for Extended Spatial Objects Various Views on Spatial Prepositions. Composing Conceptual Structure for Spoken Natural Language in a Virtual Reality Environment An Introduction to Spatial Database Systems. Open GIS Consortium Inc. Open GIS Consortium Inc. The observer's point of view: An Extension of Symbolic Projections. Qualitative representation of Spatial Knowledge in Two-Dimensional Space An Algorithm to Determine the Directional Relationship Between Arbitrarily-shaped Polygons in the plane Modern Mathematical Analysis. Using 2D C Mathematical Elements for Computer Graphics. CCAM: A Connectivity-Clustered Access Method for Networks and Network Computations Equivalence Classes of Direction Objects and Applications. A First Course in Database Systems. A Prime-Number-Based Matrix Strategy for Efficient Iconic Indexing of Symbolic Pictures --TR
orientation;oriented spatial object;open object;direction
594792
Compressing Triangulated Irregular Networks.
We address the problem of designing compact data structures for encoding a Triangulated Irregular Network (TIN). In particular, we study the problem of compressing connectivity, i.e., the information describing the topological structure of the TIN, and we propose two new compression methods which have different purposes. The goal of the first method is to minimize the number of bits needed to encode connectivity information: it encodes each vertex once, and at most two bits of connectivity information for each edge of a TIN; algorithms for coding and decoding the corresponding bitstream are simple and efficient. A practical evaluation shows compression rates of about 4.2 bits per vertex, which are comparable with those achieved by more complex methods. The second method compresses a TIN at progressive levels of detail and it is based on a strategy which iteratively removes a vertex from a TIN according to an error-based criterion. Encoding and decoding algorithms are presented and compared with other approaches to progressive compression. Our method can encode more general types of triangulations, such as those constrained by topographic features, at the cost of a slightly longer bitstream.
Introduction Huge terrain datasets are increasingly available in GIS. This fact gives rise to challenging problems in storage, transmission, and visualization of terrain models. In particular, the design of compact structures for encoding a terrain model as a sequential bitstream is important to reduce the amount of time needed to transmit it over a communication line (e.g., through Internet), to save disk storage space or to load a terrain from disk into memory. While regular grids can be compressed through techniques similar to those used for compressing images [16], this is not true for Triangulated Irregular Networks (TINs). For TINs, the problem is different and more involved. While the connecting structure of a grid is fixed, the triangulation at the basis of a TIN is not uniquely defined by its vertices, and thus it must be encoded explicitly. Such triangulation is usually the result of some specialized algorithm which takes into account issues such as preservation of terrain features, interpolation of isolines, minimization of c 2000 Kluwer Academic Publishers. Printed in the Netherlands. the approximation error, etc., and typically uses extra knowledge beside the position of vertices. Thus, it is not reasonable to encode just the vertices of a TIN, and recompute the underlying triangulation each time the TIN is loaded or received. The study of techniques for producing compressed formats of triangle meshes (and, in particular, of TINs), generally referred to as geometry compression, has gained increasing attention in the last few years because of the widespread use of TINs as terrain models. The problem of TIN compression involves two complementary tasks, which can be studied independently: Compression of the vertices, i.e., of the numerical information attached to each vertex of the TIN (location, elevation and, possibly, attributes). Compression of connectivity, i.e., of information describing the triangles of the TIN as triplets of vertices, and of adjacency information between pairs of triangles. Compression of vertices and of connectivity involve different problems and different techniques. Thus, they can be treated separately. Vertex compression has been treated by some authors [3, 8, 18] by using combinations of lossy methods based on quantization and lossless methods based on entropy encoding. In this paper, we do not address vertex compression, while we focus on compression schemes for TIN connectivity. We propose two compression methods. The first method is based on traversing a TIN in a shelling order (i.e., radially around a seed triangle); it encodes each vertex exactly once, and guarantees less than two control bits for each edge to encode connectivity. The second method compresses a TIN at progressive levels of detail and is based on a strategy which iteratively removes a vertex from a TIN according to an error-based criterion. Both methods encode both the triangles of a TIN and their adjacency relations directly in the bitstream. This paper is organized as follows: in Section 2, we provide a brief survey of existing techniques for compressing connectivity; in Section 3, we describe our compression method based on shelling; in Section 4, we present our method for progressive compression. Section 5 contains some concluding remarks. Compressing Triangulated Irregular Networks 3 2. Previous Work Our review of existing literature is focused on methods for compression of connectivity. We do not address here methods for vertex compression. A TIN is usually maintained (both on disk and in main memory) in an indexed format: a list of vertices and a list of triangles are maintained and the connecting structure of the TIN is described by providing three vertex references for each triangle. Since the number of triangles in a TIN is roughly twice the number of vertices and since a vertex reference requires log n bits for a TIN with n vertices, such a scheme requires at least 6n log n bits. The indexed format can be extended to encode triangle-to-triangle adjacencies by adding, for each triangle, the references to its three adjacent triangles. The total number of bits in this case raises to 12n log n + 6n. In the following we will refer to such structure as the explicit encoding for a TIN, and we will use it as a reference for evaluating the performance of compression methods. Techniques for compression of connectivity information of a TIN can be roughly classified into: Direct methods, which are aimed at minimizing the number of bits needed to encode a given TIN "as it is". The bitstream must be read and processed completely by a decompression algorithm in order to obtain a description of the TIN. Progressive methods, which encode a coarse TIN representation and a sequence of details which must be added to it in order to recover the initial TIN. In this case, an interrupted bitstream provides an approximation of the terrain at a lower resolution. Several compression methods proposed in the literature are not restricted to TINs but they have been developed for triangle meshes representing free-form surfaces in three-dimensional space. Thus, in our literature survey we sometimes refer to triangle meshes rather than to TINs. 2.1. Direct Methods Classical direct methods proposed in the literature are based on the decomposition of a triangle mesh into triangle strips, or generalizations of triangle strips [3, 8, 9]. Such methods are especially well-suited for visualization, but less to TIN transmission or encoding on disk because they do not provide the full mesh topology. A triangle strip is a path of triangles with alternated left and right turns; it is simply specified by listing the vertices of its triangles in a 4 Leila De Floriani, Paola Magillo, Enrico Puppo13571 3 Figure 1. (a) A triangle strip; the corresponding vertex sequence is 0; 7. (b) A generalized triangle strip; dashed arrow correspond to turns in the same direction as the previous one in the sequence; the correseponding vertex sequence with vertex replication is 0; zig-zag order (see Figure 1 (a)); each vertex appears only once in the sequence. A drawback of this simple structure is that a triangle mesh in general cannot be decomposed into strips with alternating left and right turns [9] . A more general form of strip is given by so-called generalized triangle strips, which represent standard technology in graphics packages. In this case we do not have a zig-zag sequence, but each new vertex may correspond either to a right or to a left turn in the pattern (see Figure 1 (b)). The encoding bitstream either uses a binary code, called a swapcode, to specify how to connect each new vertex to the last encoded triangle, or it repeats the vertex in case of two successive turns with the same orientation occur (see Figure 1 (b)). A drawback is that several strips are necessary to cover a triangle mesh and vertices shared by more than one strip must be replicated. On average each vertex appears twice in the resulting bitstream. It has been shown [9] that the problem of generating a minimum number of strips is NP-complete. Evans et al. [9] give several heuristics in which they try to maximize the number of triangles per strip, to minimize swaps, to avoid strips formed by just one triangle. Other methods [1, 8] also allow reusing more of the past vertices than the last two ones (as in generalized triangle strips), by storing the last k vertices in a buffer; small codes indicating how to operate on such a buffer are interleaved in the bitstream. Deering [8] proposes a buffer of positions and a syntax for the encoding, giving rise to the so-called generalized meshes. Bar-Yehuda and Gotsman [1] studied the extent to which we can increase the buffer size to reduce vertex duplication. They also show that the problem of minimizing the buffer size for a given mesh is NP-hard. Chow [3] gave algorithms for building generalized Compressing Triangulated Irregular Networks 5 (a) (b) (c) Figure 2. (a) A TIN; (b) the same TIN with a vertex at infinity added, and the edges of the vertex spanning tree superimposed as thick lines; (c) the triangle spanning tree (the dashed lines denote triangles incident in the vertex at infinity). meshes which try to optimize the use of the buffer through heuristics. Compression rates are about 11 bits of connectivity information per vertex in [3, 8] and from 2.5 to 7 in [9]. Methods based on triangle strips and their variants provide just a description of triangles and not of their mutual topological relations. Thus, if the decoded TIN must undergo operations that require adjacency relations (e.g., computation of visibility information, paths, drainage networks), these should be reconstructed separately with an extra cost. Other methods allow reconstruction of adjacency relations from the encoding bitstream [7, 19]. Our method presented in Section 3 belongs to this second class. Topological surgery by Taubin and Rossignac [18] cuts a triangle mesh along a set of edges that correspond to a spanning tree of vertices; such tree is determined by selecting the minimum number of edges which connect all the vertices. If we imagine cutting the given TIN along the edges of the vertex spanning tree, we obtain another TIN which contains the same triangles as the original one, and whose dual graph is structured as tree, thus forming a spanning tree of triangles of the original TIN (see Figure 2). The triangle spanning tree can be decomposed into a collection of triangle strips, called runs. Runs are connected through branching tri- angles, each of which connects three runs. The bitstream produced by the method consists of the vertex and of the triangle spanning trees. A rather complex algorithm permits to reconstruct the original TIN. Through experiments on manifold meshes Taubin and Rossignac show a compression of 4 bits per vertex on av- erage. The method produces longer strips than the previously discussed ones; it requires that all vertices are kept in main memory during both compression and decompression; both the coding and the decoding al- 6 Leila De Floriani, Paola Magillo, Enrico Puppo gorithms perform consistent graph manipulation and are not easy to implement. The method by Touma and Gotsman [19], unlike the previous ones, encodes triangle adjacencies directly in the bitstream. The compression algorithm maintains an active list of edges bounding a (not necessarily simply-connected) polygon which separates triangles already encoded (located inside the polygon) from triangles that have still to be to coded (located outside). At each step, the polygon is enlarged by encoding all triangles lying outside it which are incident in one of its vertices. The algorithm works in a completely similar way. The resulting bitstream contains each vertex once, and three types of commands: ADD, SPLIT and MERGE, where the first two commands have one argument and the third command has two arguments. The argument of ADD is the degree d of a new vertex to be inserted (i.e., the number of triangles incident in such vertex), the arguments of SPLIT and are integer numbers with the meaning of indexes and offsets; such commands mark those points at which the number of boundary curves of the current polygon increases by splitting one curve into two, or decreases by merging two curves into one, respectively. With the exception of MERGE operations (which are rare), the compression and the decompression algorithms run in linear time in the number of mesh edges. Details about the complexity of MERGE are not provided in [19]. The method generates an ADD command for each vertex of the mesh. The number of SPLIT and MERGE commands depends on the mesh. The authors do not report any evaluation of the number of SPLIT or MERGE commands that can occur: they just observe that MERGE commands can occur only for meshes of positive genus (i.e., three-dimensional surfaces with through holes), and thus not for terrains. If we assume that the degree of any vertex is less than b, the compressed sequence requires at least 2 bits per vertex: two bits for the command (since we have three possible commands) and log b bits for the degree: for instance, 5 and 6 bits, respectively, if 16. The bits necessary to represent possible SPLIT and MERGE commands add to these. Touma and Gotsman observe that the sequence of codes can be compressed further. Since the average degree of a vertex is 6, there is a spread of degrees around this value; moreover, there are easily long subsequence repeating the same commands and the same arguments. Thus, they use entropy coding combined with run-length encoding, getting experimental compression rates from 0.2 to 2.4 bits per vertex for connectivity. Compressing Triangulated Irregular Networks 7 vertex split edge collapse v' Figure 3. Edge collapse and its reverse vertex split. 2.2. Progressive Methods Progressive methods encode a coarse approximation of a mesh and a sequence of modifications which incrementally refine it into the original mesh [12, 13, 17]. An interrupted bitstream provides an approximation of the whole terrain, which can be seen as a form of lossy compression. Progressive compression methods are especially useful for mesh transmission or as compact formats for secondary storage since they allow a trade-off between loading/transmission time and loss of accuracy. This additional feature is paid with compression rates that are usually lower than those of direct methods. The basic technique for building progressive representations of meshes come from iterative techniques for mesh simplification based on local modifications of a mesh (see, e.g., [14, 11]). A sequence of progressive resolutions for a mesh is generated by iteratively applying a destructive operator, which removes details somewhere in the mesh. A destructive operator has an inverse constructive operator, which recovers such details. A compressed bitstream contains the coarse mesh (i.e., the result of a sequence of destructive operations) along with the sequence of constructive operations necessary to recover the original mesh. Progressive Meshes, developed by Hoppe [12, 13], provide a progressive compression method where the destructive operator is edge collapse. Edge collapse replaces an edge e with one of its endpoints v, and the two triangles sharing e with two edges incident at v. The corresponding constructive operator is vertex split, which expands a vertex v into an and two edges e 1 and e 2 among those incident at v into two triangles (see Figure 3). Each vertex split in the sequence is compactly encoded by providing the new vertex v 0 , a reference to an old vertex v, and a code specifying the position of e 1 and e 2 around v. The number of bits necessary to encode connectivity is n(log n a mesh of n vertices, if b is the maximum degree of a vertex, since the position of e 1 is a number in the range and that of e 2 a number in the range one position has already been chosen for e 1 ). 8 Leila De Floriani, Paola Magillo, Enrico Puppo For instance, if we have 21 (23n) bits of connectivity information. Snoeyink and Van Kreveld [17] have developed a progressive method which applies to TINs based on Delaunay triangulations, and guarantees that all intermediate meshes are Delaunay triangulations. The destructive operator removes a vertex from a Delaunay triangulation; the corresponding constructive operator re-inserts such vertex. The compression algorithm performs several stages starting from the given each stage removes a maximal set of independent vertices (i.e. vertices not connected by an edge) from the current Delaunay triangulation. The process is guaranteed to terminate in a logarithmic number of stages. The bitstream contains the final coarse mesh plus, for each stage in reverse order, the sequence of removed vertices (which must be re-inserted by the decoding algorithm). The sequence of vertices removed in a stage is sorted in the same order in which the triangle containing each vertex is encountered in the mesh traversal algorithm of de Berg et al. [4]. Thus, the decoding algorithm at each stage can locate all the triangles containing the new vertices in linear time in the size of the current triangulation. The length of the resulting bitstream is O(log 2 n) where n is the number of vertices and the constant is not known. A disadvantage of this method is in requiring heavy numerical calculations (i.e., the incremental reconstruction of a Delaunay triangulation) for which may cause not only time overhead, but also numerical and inconsistency problems (e.g., different triangulation of co-circular points in the original and in the reconstructed mesh). No experimental results are available. Note that both methods described above assume a specific strategy for retriangulating the mesh after the insertion of each vertex (i.e., vertex split in [12, 13], and Delaunay update in [17]). In Section 4, we propose a progressive method which does not depend on the retriangulation strategy. 3. A Direct Compression Method Based on Shelling In this section we describe a new compression method based on shelling. This method is designed for mesh transmission or disk storage because it encodes both triangles and their adjacency relations. A shelling sequence of a TIN \Sigma is a sequence oe 1 angles, containing each triangle of \Sigma once, and such that, for any prefix subsequence n), the boundary of the union of the set of triangles foe is a simple polygon. A mesh is shellable if it admits a shelling sequence; it is extendably shellable if Compressing Triangulated Irregular Networks 9 it is shellable and every shelling sequence defined for a subset of its triangles can be extended to a shelling sequence for the whole mesh. Extendably shellable meshes include TINs and all the triangulations of surfaces homeomorphic to a sphere or a disk; surfaces with a non-null genus (i.e., with through holes) are not extendably shellable [2]. The method we present here applies to triangle meshes that are extendably shellable. The bitstream implicitly encodes a shelling sequence of the given TIN. It contains the vertices of the TIN interleaved with control codes providing connectivity information. There are four control codes: VER- TEX, SKIP, LEFT, and RIGHT, each represented on two bits. Coding and decoding algorithms are very simple and work in time linear in the size of the TIN without performing any numerical computation. Both the triangles and their adjacency links are recovered from the bitstream without extra processing. Moreover, only the boundary of the current polygon and triangles incident to its edges have to be kept in memory, while compressed / decompressed triangles lying completely inside the polygon can be discarded. The method starts from an arbitrary triangle and traverses the TIN according to a shelling order. At each iteration, it produces a two-bit code and, possibly, a vertex. Each vertex of the TIN is encoded exactly once. 3.1. Compression and Decompression Algorithms The compression and decompression algorithms maintain the simple polygon bounding the currently encoded / decoded portion of the triangulation in a doubly-linked list; they also keep a queue of active edges, which are those edges of the current boundary that have not been tested yet to compress / decompress the triangle externally adjacent to them; a control code is sent/read for each active edge. The compression algorithm starts from an arbitrary triangle of the TIN, whose edges define the initial polygon, and sends its three vertices to the output stream. Then, it examines each edge e of the current polygon in turn and tries to enlarge the polygon by including the triangle oe externally adjacent to e. It performs the following actions on e: 1. if the third vertex v of oe does not lie on the polygon, then include oe and send a VERTEX code followed by vertex v (see Figure 4 (a)); 2. if oe does not exist (i.e., edge e is on the boundary of the TIN), or the third vertex v of oe lies on the polygon, and it is not the other vertex of one of the edges adjacent to e along the polygon (i.e., oe cannot be RIGHT (a) (b) (c) (d) e e e" e VERTEX e Figure 4. The five possible situations for the current edge e in the TIN compression algorithm: (a) the triangle externally adjacent to e brings a new vertex V ; (b) the triangle externally adjacent to e does not exist; (c) the triangle externally adjacent to e exists but cannot be included because it violates shelling; (d,e) the triangle externally adjacent to e has another edge on the current polygon, and such edge is the one lying to the left and to to right of e, respectively. included since it makes the polygon not simple), then send a SKIP code (see Figure 4 (b) and (c)); edge e will not be examined again; 3. if the third vertex v of oe lies on the polygon, and it is the other vertex of the edge adjacent to the left of e, then include oe and send a LEFT code (see Figure 4 (d)); 4. if the third vertex v of oe lies on the polygon, and it is the other vertex of the edge adjacent to the right of e, then include oe and send a RIGHT code (see Figure 4 (e)). The loop terminates when all the triangles have been compressed. The decoding algorithm works in a similar way. It builds the first triangle from the three vertices found at the beginning of the bitstream. Then, it examines each edge e of the current polygon in turn, and tries to enlarge the current polygon by creating a new triangle adjacent to e, Compressing Triangulated Irregular Networks 11 according to the directions contained in the next control code read from the bitstream: 1. in case of a VERTEX code: read a vertex v from the bitstream, create a new triangle oe from e and v, and set an adjacency link between oe and the triangle adjacent to the other side of e; 2. in case of a SKIP code: no new triangle is created; edge e will not be examined again; 3. in case of a LEFT code: create a new triangle oe from e and the edge adjacent to the left of e along the polygon; set adjacency links between oe and the triangles adjacent to the other side of e and e 4. in case of a RIGHT code: create a new triangle oe from e and the adjacent to the right of e along the polygon; set adjacency links between oe and the triangles adjacent to the other side of e and e 0 . The loop terminates when the bitstream has been read completely. 3.2. Analysis Since a TIN is extendably shellable, at each step one of the following two conditions occurs: either all triangles have been processed, or there exists a triangle which can be added to the current polygon while maintaining it simple; at least one edge of such a triangle must lie on the current polygon. In the first case, the process terminates. In the latter case we show that the compression algorithm can find a new triangle to encode. Each step of the main loop examinates an edge e of the current polygon until it finds one such that the triangle externally adjacent to it can be included. It is easy to see that, if e generates a triangle (cases VERTEX, LEFT, and RIGHT) then such triangle preserves the simplicity of the current polygon and, if e does not generate a triangle (case SKIP) then either e is on the TIN boundary, or the triangle t e externally adjacent to e does not preserve simplicity because the vertex of t e opposite to e belongs to the boundary of the current polygon. Edges of this last type will not be examined again (they are "virtually removed" from the current polygon). Therefore, the algorithm will scan edges of the current polygon until a non-SKIP edge is found. Since the triangulation is extendably shellable, we are always sure that one such edge exists until the whole triangulation has been traversed. It remains to be shown that a triangle t e externally adjacent to a SKIP edge e at some stage will be generated at some later stage. Since all three vertices of t e already belong to the current polygon, it will be possible to include t e only after having another edge e 0 of t e on the polygon. Let t 0 be the neighbor of t e along e 0 . Since we are following a shelling order, t 0 , hence e 0 , will be included in the current polygon at some time. Thus, while examining e 0 , either a LEFT or a RIGHT code will be sent, and t e will be included. Thus, we have shown that the compression algorithm always termin- ates, after having traversed all the triangles of the given TIN in shelling order. At each step, the number of operations performed on the current edge is bounded by a constant. Each TIN edge is examined at most once in the main loop, hence the time complexity for both compression and decompression is linear in the number of edges, which is about 3n, i.e., O(n). It is easy to see that the actions performed by the decompression algorithm, on each code read from the bitstream, reproduce exactly the same triangle configuration from which that code has been generated by the compression algorithm. Thus, the original TIN is reconstructed correctly. Each vertex is sent exactly once by the compression algorithm, and a code is sent for every examined edge, i.e., at most one code for each TIN edge. Indeed, not all edges are examined: LEFT and RIGHT codes eliminate one edge each, which therefore does not generate a code; moreover, the vertices of the boundary remaining after having processed all the triangles do not generate any code. Since the number of VERTEX codes is fixed, the bitstream is shorter when there are many LEFT and RIGHT codes and few SKIP codes. Implementing the list of active edges as a queue results in a polygon that tends to grow in a breadth first order, thus providing many LEFT and RIGHT codes. We found that an implementation based on a stack, hence on a depth first order, produces much more SKIP codes, and, thus, longer bitstreams. In the worst case, the number of control codes is bounded from above by the number of edges in the given TIN, i.e., about 3n (equivalent to 6n bits), where n is the number of TIN vertices. Experimental results give better compression rates (see Section 3.3). 3.3. Experiments and Discussion Our compression method takes into account only the connectivity structure of a TIN, not its geometry. Thus, in order to prove the performance of our approach on a wide range of inputs, we show experiments on TINs having different kinds and sizes of triangulations, rather than on TINs representing different terrains. Compressing Triangulated Irregular Networks 13 Table I. Results of the compression algorithm based on shelling. U1-4 are TINs at uniform accuracy, A1-4 and B1-4 are TINs at variable accuracy. Times are in seconds. TIN vert tri code bits / encoding decoding bits vert time (tri/sec) time (tri/sec) U2 28510 56540 123086 4.3173 2.53 (22347) 1.55 (36477) A3 15515 30818 65210 4.2030 1.36 (22660) 0.84 (36688) All TINs used here are approximated models built by a surface simplification algorithm from an original grid dataset representing the area of Mt. Marcy (NY), courtesy of U.S. Geological Survey. Data U1-4 are TINs approximating original data at a uniform error value, equal to 1, 2, 5 and 10 units, respectively; all triangles in a single TIN have about the same area. Conversely, TINs A1-4 and B1-4 are characterized by a great difference in triangle density of different zones: in A1-4 (B1-4), one fourth (one 16th) of the area is represented with an elevation error equal to 0.5 meters, while the rest is represented at the coarsest possible detail. Experiments have been performed on a Pentium Pro (200 Mhz, 64 Mb RAM). Results are shown in Table I. Compression rates are lower than 4:5 bits per vertex. Execution times are about 22K and 37K triangles per second in compression and decompression, respectively. The main advantages of our method are those of being very simple conceptually, and easy to implement, and at the same time, of giving good compression rates. The method is as simple as (generalized) triangle strips [3, 8, 9] and achieves higher compression rates, comparable to those of more complicated algorithms [18]; moreover, such algorithms do not provide adjacency information. 14 Leila De Floriani, Paola Magillo, Enrico Puppo w153 (a) (b) (c) (d) (e) Figure 5. The triangulation after (a) and before (e) removing a vertex v, and the related sequence of edge flips (b,c,d), to be read left-to-right to restore situation (e) from situation (a). In (a), the edges around w, and in (b,c,d) the edges around v are Triangle tv is encoded as the pair (w;4) and the sequence of edge flips is 1,3,0. The algorithm proposed in [19] also encodes adjacency information, and is more complicated than ours since it manages a multiply-connected polygon instead of a simple one. The bitstream generated by such method (disregarding the postprocessing step that is applied in [19] to further compress it) requires at least 5n or 6n bits (depending on the maximum degree of a vertex in the TIN), plus a variable number of SPLIT and MERGE codes, which depends on the input mesh. Thus, the length of the two bitstreams is comparable on average. 4. A Progressive Compression Method based on Edge-Flips In this section we present a new method for TIN compression at progressive levels of detail. We use a destructive operator which consists of removing a vertex v of degree bounded by a constant b from the current TIN and of retriangulating the corresponding polygonal hole -. The inverse constructive operator inserts vertex v and recovers the previous triangulation of -. We recall that the degree of a vertex is the number of triangles incident in it. The old triangulation is recovered from the new one by first splitting the triangle t v , containing v, in the current TIN, into three triangles, and then applying a sequence of edge flips. An edge flip is an operation that replaces an edge, shared by two triangles whose union forms a convex polygon, with the other diagonal of such polygon. Our bitstream encodes, for each removed vertex v, both the starting triangle t v and the sequence of edge flips to be performed (see Figure \Gamma The starting triangle t v is given by a reference to a vertex w of t v and by the index of t v in the star of triangles incident at w. The reference to w requires log n bits; the index of t v requires log b bits. Compressing Triangulated Irregular Networks 15 operation is specified by an edge opposite to v (in its star of triangles), which is involved in the flipping operation. If triangles incident at v are numbered in a conventional order in the data structure encoding the TIN, such edge can be uniquely identified with an index. Note that there are three edges incident at v at the beginning and their number increases by one at each flip: there will incident edges after i flips, and incident edges before the last flip, where k is the degree of the removed vertex (k - b). Therefore, for k ? 3, the whole sequence of edge flips is identified by a sequence of k \Gamma 3 indices s is in the range [0; 3+i]. It is easy to see that for a removed vertex of degree k, there exist 3 possible sequences of flip indices. A given sequence can be encoded compactly with a unique number in the range [0; (k\Gamma1)! called a flip code, which requires dlog( (k\Gamma1)! bits and is given by the formula: This flip code can be unpacked with a standard method: s k\Gamma4 is given by c div (k\Gamma2)! (j 2)!where div denotes the integer division operation. Our bitstream encodes each removed vertex once and, for each vertex, a vertex reference, a triangle index, the number of flips to be performed and a flip code, as described above. Disregarding the cost of the initial TIN, which is always very small, our method produces n(dlog ne 2dlog be)+ bits of connectivity information, where k is the number of removed vertices of degree k. Since, on average, the degree of a vertex in a TIN is six, this can be roughly estimated as ne Experiments we made on real data confirm such an estimate. 4.1. Compression and Decompression Algorithms The compression algorithm works as follows: 1. Initialize the current TIN as the input TIN. 2. While the current TIN is not "small enough" select a vertex v of degree at most b according to some criterion (to be explained next): remove v and retriangulate the polygonal hole -; locate the triangle t v of the new triangulation of - which contains choose a vertex w of t v , having degree at most equal to b and compute the index of t v around w; simulate the reconstruction of the old triangulation of the hole (the one containing v) from the current one (the one without v), and encode the corresponding sequence of edge flips. 3. Send the current TIN, in a directly compressed format (e.g., as explained in Section 3), followed by the sequence of removed vertices in reverse order; for each vertex v, output also the index of the associated vertex w, the index of t v in the set of triangles incident in w, the number of flips to be performed and the flip code. The compression algorithm has two key points: the choice of the vertex v to be removed at each iteration of loop 2, and the retriangulation of the corresponding hole -. Any criterion can be used for both purposes, without affecting the format of the bitstream, or the algorithm. We will discuss such issues further in Section 4.2. The decompression algorithm performs the following steps: 1. Initialize the current triangulation with the coarse triangulation read from the bitstream. 2. Read the next vertex v with associated information. Get vertex w from the list of vertices in the current TIN and retrieve triangle t v in the star of w. Insert v into the current triangulation by splitting triangle t v at v; unpack the flip code and perform the given sequence of edge flips. 3. The process can be stopped either when the bitstream is finished (in this case the original TIN is completely reconstructed, and we have a lossless decompression), or when the current TIN has a given number vertices, or when a certain time has elapsed (in the two latter cases, a TIN approximating the original one is built, and we have a lossy decompression). During both compression and decompression, the current TIN is maintained in a data structure corresponding to the indexed format with triangle-to-triangle adjacencies (see Section 2), augmented with Compressing Triangulated Irregular Networks 17 partial vertex-to-triangle incidence information: for each vertex w in the TIN, a link to one of its incident triangles is maintained. Such a data structure is updated according to fixed conventional rules, which allow us to obtain exactly the same indexing scheme during both compression and decompression. This is fundamental to uniquely identify, at each step, triangle t v in the star of vertex w and each edge to be flipped through the corresponding flip index. 4.2. Analysis The correctness of the method follows from basic properties of triangulations well known from computational geometry: 1. if we remove a vertex, together with all its adjacent triangles, from the interior of a plane triangulation, we make a star-shaped polygonal hole that can be always filled with triangles without adding extra vertices (we have a slightly different situation if the vertex is removed from the boundary of the triangulation); 2. given a plane triangulation T and a new vertex v, if t v is the triangle of T containing v, then we can obtain a new triangulation by replacing t v with three new triangles incident at v (triangle split); 3. any triangulation of a given set V of vertices can be transformed into any other triangulation of the same set of vertices through a sequence of edge flips; 4. each edge flip applied to a boundary edge of the star of a vertex v increases the degree of v by one unit. ensures that at each step of the compression algorithm we obtain a new (simplified) triangulation. Let T be the current triangulation at a generic step, and T 0 be the triangulation obtained by eliminating vertex v from T . Property 2 ensures that locating t v in T 0 provides sufficient information to obtain a new triangulation T 00 containing v with degree three, while property 3 ansures that T can be obtained from T 00 through a sequence of edge flips. Since T differs from T 00 only in the portion corresponding to the hole left by removing v from T , then edge flips will affect only that portion. Each edge flip given by a flip index introduces an edge incident at v which belongs to T , and such an edge is never removed by a subsequent flip, because each flip affects an edge on the boundary of the current star of v. Hence, if the degree of v in T is k, then T will be recovered by exactly k \Gamma 3 flip operations, which are specified by the flip code. The core of the compression algorithm is the main loop in Step 2. As already mentioned, we have two degrees of freedom: in the selection of the vertex v to be removed and in the retriangulation of the resulting hole. Any criterion can be used to retriangulate the hole -. Since - has a bounded number b of edges, retriangulating it requires a constant number of operations in any case. For instance, an ear cutting procedure might be applied, which corrsponds to iteratively flipping edges incident at v, until just three incident edges are left. Cutting an ear of - corresponds to flipping an edge incident at v, thus the sequence of edge flips to be encoded is provided directly. Otherwise, no matter what algorithm has been used to triangulate -, the sequence of flips can be obtained as follows. Let T - be the triangulation of -. Vertex v is located first in T - and triangle t v is found. An influence polygon is initialized with t v , and its edges are scanned. Each time an edge e is found, which does not lie on the boundary of T - , then e is deleted and the influence polygon grows by incorporating the triangle externally adjacent to e. The index of e is encoded as a split index. This process continues until the influence polygon covers the whole - (i.e., all edges are on the boundary of T - ). Any criterion can also be used for selecting the vertex to be removed, under the constraint that its degree is not higher than b. For a sufficiently large value of b (e.g., this constraint has no impact on real world cases. The specific implementation choice affects both the complexity of the compression algorithm and the efficiency of the resulting bitstream for the purpose of progressive transmission. We suggest to select the vertex causing the least increase in the approximation error (as it was also done by other authors, e.g., [14, 11]). This requires that the error caused by removing a vertex of the current TIN is known at each iteration. The approximation error of the current TIN is the maximum approximation error of its triangles. For each triangle t, its approximation error is equal to the maximum vertical distance of a removed vertex, whose vertical projection falls in t, from the plane of t. In order to compute approximation errors of triangles we associate with each triangle t in the current triangulation the set of vertices already deleted whose vertical projections fall in t. The error introduced by removing a vertex v is evaluated by simulating the deletion of v: the hole left by v is retriangulated and the error is measured by locating v and all points lying inside its incident triangles with respect to the new triangles that fill the hole, computing the errors of such triangles, and choosing the maximum error. The error of each vertex v is stored together with v, and it is updated whenever a vertex adjacent to v is removed. Note that it is sufficient to evaluate the error only at vertices whose degree is smaller than b, because they are the only Compressing Triangulated Irregular Networks 19 Table II. General information and theoretical bound for the three TINs considered in our experiments. terrain #vertices log 2 n log 2 b bits/vert (theor.) Devil Peak 38115 Mount Marcy 16384 14 4 28 San Bernardino 16384 14 4 28 candidates for removal. The triangulation of a hole, necessary to evaluate the error, is performed only at vertices of bounded degree, hence it takes constant time at each vertex. Vertices of bounded degree are maintained in a priority queue based on their errors. At each deletion, error must be re-evaluated only for a constant number of vertices, i.e., the neighbors of v. Location of points inside a given triangle t occurs at most four times, namely for evaluating the cost of removing each of its three vertices, and when one of such vertices is actually removed (when t is removed as well). During the whole coarsening process, a removed vertex may fall, and need to be relocated, in O(n) different triangles in the worst case. However, following [10], it is easy to show through amortized analysis that the expected cost of point location is still O(log n). It follows that the total cost for computing approximation errors is O(n log n). This equals the cost of managing the priority queue, since each operation on the queue costs O(log n), and a constant number of operations are performed for each removed vertex. Therefore, the total time complexity of the compression algorithm in this case is O(n log n). The selection criterion affects the quality of the intermediate triangulations obtained from an interrupted bitstream (i.e., the ratio between accuracy and number of vertices used), and thus the efficiency of the sequence for progressive transmission. In other words, if the bitstream is interrupted after k vertices, the approximation error of the currently reconstructed TIN is either larger or smaller, depending on the policy used within the compression algorithm for selecting vertices that must be removed. One can trade-off the complexity of the selection procedure with the quality of the compressed sequence. Other possible techniques for vertex selection, different from the one described above, have been discussed and experimented in [5]. Such alternative methods include the selection of a maximal set of independent vertices of bounded degree at each time, still based on an error-driven criterion, and the random selection of a maximal set of independent vertices (which runs in O(n), but gives poor quality of intermediate triangulations in the bitstream). Table III. Length of the bitstream with the three encoding algorithms terrain algo #bits bits/vert max deg avg deg Devil Peak IR 1127k 29.58 11 5.64 Devil Peak IE 1125k 29.53 11 5.62 Devil Peak SE 1159k 30.42 11 5.97 Mount Marcy IR Mount Marcy IE 449k 27.43 11 5.58 Mount Marcy SE 462k 28.22 11 5.89 San Bernardino IR 450k 27.50 11 5.61 San Bernardino IE 450k 27.50 11 5.60 San Bernardino SE 465k 28.39 11 5.95 Table IV. Number of vertices which must be decoded to obtain a TIN within a certain approximation error (expressed as a percentage on the heigth range of the data set), with the three encoding algorithms. Devil Peak IR 33766 29473 22385 13515 4060 1349 Devil Peak IE 30845 26113 18525 10086 2830 925 Devil Peak SE 32566 29154 20910 10512 2483 805 Mount Marcy IR 12323 9049 5578 3205 1368 818 Mount Marcy IE 10540 7278 4168 2378 936 527 Mount Marcy SE 10022 7037 3915 2068 692 230 San Bernardino IR 13302 11836 9863 7389 3414 1296 San Bernardino IE 12083 10613 8660 6070 2579 1296 San Bernardino SE 11222 10613 7311 4849 1965 657 We experimented three different encoding algorithms on three terrains (datasets courtesy of the U.S. Geological Survey). The characteristics of the three TINs are summarized in Table II. Tables III and IV show the results. The method denoted as IR removes a maximal set of mutually Independent vertices chosen in a Random way; IE removes a maximal set of mutually Independent vertices chosen in an Error-driven removes a Single vertex chosen in an Error-driven way. In all cases, the bound on the degree of vertices to be removed is Compressing Triangulated Irregular Networks 21 All heuristics give a bitstream length close to the theoretical bound: it is slightly lower for methods that remove independent vertices, and slightly higher for the method that removes a single vertex. Method IR needs more vertices to be decoded in order to achieve a given precision in the resulting TIN, with respect to error-driven methods IE and SE. At low resolution, SE always outperforms the other two heuristics, and generally behaves better than IE in most cases. The decompression algorithm works in linear time, since n vertices are read, and the insertion of any vertex v into the current triangulation involves a search for t v in the star of a vertex w of degree bounded by b, and a number of flips bounded by b \Gamma 1, with b a predefined constant. Note that no numerical computations are needed in reconstructing the TIN. 4.3. Discussion The main advantage of our method is in being more general and flexible than other progressive methods [12, 13, 17]. In fact, it does not impose any specific criterion for removing vertices and for retriangulating after vertex removal. It can incorporate different criteria for choosing the vertices to remove, trading off between compression time and quality of the intermediate meshes. The only constraint is on the maximum degree of removed vertices (similar to Hoppe's). Hoppe does not take into account the approximation error when selecting the next edge to collapse. Also, Snoeyink and Van Kreveld do not select vertices with an error-based criterion; moreover, deleting independent sets of vertices at once puts a strong constraint, hence reducing adaptivity: the number of vertices necessary to achieve a given approximation error may be much higher. In our approach, the specific criterion used for retriangulation is encoded in the bitstream itself rather than being predefined. On the contrary, [17] is restricted to Delaunay triangulations and [12, 13] to triangulations obtainable through edge collapse (for instance, it cannot guarantee a Delaunay triangulation at each step). Our method is less compact than those in [12, 13] and [17]. Disregarding the cost of the initial triangulation, which is always very small, our method produces n(dlog ne bits of connectivity in- formation. Hoppe's method requires n(log n bits. For method costs 21n (23n) bits, while our bitstream requires about 28n (30n) bits. Thus our method has an overhead of about 30% with respect to Hoppe's with the advantage of encoding more general TINs (e.g., Delaunay, data-dependent, etc.) and 22 Leila De Floriani, Paola Magillo, Enrico Puppo of being based on vertex removal/insertion, which is probably the most common technique used in GIS for terrain simplification. The bitstream of [17] describes connectivity in just O(log 2 n) bits; no experimental results are known. However, their decoding algorithm involves heavy numerical calculations, which are time-consuming and subject to floating point inaccuracies; our method does not perform any numerical computation in decoding. 5. Concluding Remarks We have proposed two new methods for compressing TINs. The first method is a direct method based on traversing a TIN in a shelling order. It sends each vertex once, and produces less than 4:5 bits of connectivity information for each vertex, which is advantageous with respect to other methods [3, 8, 9, 19]. Moreover, the adjacency relations between triangles are reconstructed directly from the bitstream at no additional processing cost. Because of this property, it is better suited to TIN analysis than methods based on strips (which are essentially optimized for rendering purposes). The main advantages of this compression technique are the following: it provides triangle adjacency information, it is conceptually simple and easy to implement, and it gives good compression rates. The algorithm proposed in [19], although more complicated than ours, has similar features: it works expanding a polygon through a TIN and provides triangle adjacency information. In [19], Touma and Gotsman achieve improved compression rates by further compressing the bit-stream through entropy encoding and run-length encoding. A similar signal compression strategy can be applied to the sequence generated by our shelling-based algorithm as well. We plan to test such possibility in the near future, and we believe that this could lead to improvements in compression rates comparable to those of [19]. Here, we have proposed our compression method for extendably shellable triangle meshes and we have applied it to TINs. The method can be extended to triangle meshes that are not extendably shellable. If the input mesh, representing a free-form surface in space, is not extendably shellable, then it is automatically partitioned into shellable patches and each patch is encoded independently. This adds just a small overhead due to the replication of vertices that belong to more than one patch. Based on a similar principle, the method also extends to higher dimensions to compress tetrahedral meshes which, in general, are not extendably shellable. In this case, the number of possible control codes and, thus, the number of bits to represent them, increases: four bits are needed to Compressing Triangulated Irregular Networks 23 compress tetrahedral meshes. Some results on generic triangle meshes and on tetrahedral meshes can be found in [15]. The second method belongs to the class of compression methods providing progressive approximations for a TIN. It allows transmission of a TIN according to criteria related to the approximation error of the terrain representation. It is more flexible than Hoppe's method [12, 13] since it can deal with Delaunay triangulations, and it has a broader applicability than the one by Snoeyink and van Kreveld [17], since it can deal with data dependent and constrained Delaunay triangulations. On the other hand, it is just a little more expensive than Hoppe's method. From the encoded bitstream a multiresolution model of a TIN, the Multi- Triangulation, can be constructed, which makes it possible to perform terrain analysis at variable resolution [6]. It applies to generic surfaces in space, other than TINs. We are currently studying an alternative encoding scheme for the flip sequence that might be more compact and, furthermore, be easily reversed in order to manage dynamic processes in which a mesh can be either refined or coarsened on-line. We have shown experimental results for the first method, while the second one is currently under implementation. Compression algorithms with different error-driven selection techniques have been experimented already in [6]. Acknowledgements This work has been developed while Leila De Floriani was visiting the University of Maryland Institute for Advanced Computer Studies (UMI- her research has been supported by National Science Foundation (NSF) Grant "The Grand Challenge" under contract BIR9318183. This work has been partially supported by the Strategic Project of The Italian National Research Council "Symbolic Computation - Systems for Geometric Modeling" under contract number 97.04800.ST74. --R Time/space tradeoffs for polygon mesh rendering. Shellable decompositions of cells and spheres. Optimized geometry compression for real-time rendering Simple traversal of a subdivision without extra storage. Building and traversing a surface at variable resolution. Compressing TINs. Geometry compression. Optimizing triangle strips for fast rendering. Randomized incremental construction of the Delaunay and Voronoi diagrams. Survey of surface simplification algorithms. Progressive meshes. Efficient implementation of progressive meshes. A drop heuristic conversion method for extracting irregular networks from digital elevation models. Spatial Operations on Multiresolution Cell Complexes. JPEG, Still image compression standard. Geometric compression through topological surgery. Triangle mesh compression. --TR --CTR Sebastian Krivograd , Mladen Trlep , Borut alik, Compression of triangular meshes by simultaneously processing pairs of triangles, Proceedings of the 21st spring conference on Computer graphics, May 12-14, 2005, Budmerice, Slovakia
geometric compression;data structures;triangulated irregular networks;terrain modeling
594830
Progressive Transmission of Vector Map Data over the World Wide Web.
Within distributed computing environments, access to very large geospatial datasets often suffers from slow or unreliable network connections. To allow users to start working with a partially delivered dataset, progressive transmission methods are a viable solution. While incremental and progressive methods have been applied successfully to the transmission of raster images over the World Wide Web, and, in the form of prototypes, of triangular meshes, the transmission of vector map datasets has lacked a similar attention. This paper introduces a solution to the progressive transmission of vector map data that allows users to apply analytical GIS methods to partially transmitted data sets. The architecture follows a client-server model with multiple map representations at the server side, and a thin client that compiles transmitted increments into a topologically consistent format. This paper describes the concepts, develops an architecture, and discusses implementation concerns.
Figure 1: Roads belonging to different layers. In the second case, the file is composed of a single layer and entities having different semantic and thematic meaning coexist in it. In this case, it is assumed that the original composing set of entities has been overlayed, i.e., there are no intersections between pairs of entities stored in the file. paper mill river major road lake creek minor roadA minor roadB (a) paper mill river lake creek creek mouth major road bridge intersection AB minor roadA minor roadB (b) Figure 2: (a) A collection of spatial entities with the following pairwise intersections: minor road A and minor road B; creek and (the boundary of) lake; river and major road; (b) an overlay algorithm has been applied to the set of entities in (a). For example, in Figure 2(a) a map is shown where intersections between entities occur. In order to obtain an overlayed set, some entities must be added to the vector file storing the map, namely the intersection point between minor road A and minor road B (intersection AB), the intersection point between creek and lake (creek mouth), and the intersection region between river and major road (bridge). Such a file, shown in Figure 2(b), can be generated by applying an overlay algorithm to the set of entities in Figure 2(a) (Bentley and Ottman 1979, Chazelle and Edelsbrunner 1992). Overlay operations are always required when a query involving entities whose thematic information is stored over two or more different layers must be performed. For example, given a layer representing the hydrography of Italy and a layer describing the set of Italian regions, we may want to perform queries such as, ?Which regions are traversed by the river Po?? Map overlay is a complex operation and it is computationally expensive. Thus, if the purpose of the vector file utilization is querying, it is more convenient to keep an already overlayed set of entities instead of performing an overlay operation every time a query is formulated. In the case of progressive transmission, if the client requires the downloading for query purposes, he more likely expects to be provided with a ready-to- use file instead of having to perform expensive preprocessing operations on it, such as map overlay. On the other side, keeping separated thematic layers in the same file prevents segmentation of entities. For instance, each road in Figure 1 is fragmented into two branches after the detection of the intersection point, if the overlayed approach is adopted. This splitting is not necessary if two distinct layers, one for major roads and one for minor roads, are maintained. Using this second approach, a smaller amount of entities is stored. Furthermore, a higher degree of granularity in the transmission can be obtained. With reference to the above example, the major road network and the minor road network can be extracted from the file and transmitted separately, even if they are included in the same level of detail. In conclusion, the adoption of either the overlayed or the separated layers approach depends on whether the priority is mainly querying or visualization, respectively. In this paper, we consider vector files composed of a single layer of overlayed entities composing a map. In this case, the most commonly used mathematical model is the cell complex. We describe this concept in the next section. 2.3.2 Cell-Complexes Cell complexes are commonly used as underlying mathematical models for vector maps (Bertolotto et al. 1994; Egenhofer et al. 1989; Frank and Kuhn 1986; G ting and Schneider 1995; Worboys 1992) as they provide a sound theoretical basis for the formalization of spatial entities and relations contained in a map. More formally, we define a map as an overlayed set of points, simple lines and (possibly multiply connected) regions. A (Euclidean) cell complex is defined as a collection of cells that are homeomorphic to disks. Each cell has a dimension and its boundary is composed of cells of lower dimension. Pairs of cells in the complex are either completely disjoint or they meet along their boundary. A complex is said to be n-dimensinal if n is the maximum dimension of its cells. The correspondence between entities in a map and cells in a two-dimensional cell complex is straightforward: cells in a cell complex describe entities of a map. Points are represented as 0-cells, lines as 1-cells, and regions as 2-cells. The interior and boundary properties of cells in a cell complex model topological relations among entities in a map (see Bertolotto (1998) for a more detailed description). Several data structures for cell complexes have been defined in the computational geometry and GIS literature (De Floriani et al. 1993; Frank and Kuhn 1986; Preparata and Shamos 1985). In particular, topological data structures must encode both entities and their mutual topological relations (Egenhofer and Franzosa 1991). 3. Challenges in Progressive Vector Map Data Transmission As shown in the previous section, progressive transmission of raster data is relatively simple. For example, if the purpose of the transmission is visualization, just adding pixels to an incomplete image generates a more refined version of it. Available compression mechanisms are also very effective. When the purpose of the transmission is object manipulation, a vector representation is required. Except for the particular case of triangular meshes for which effective compression methods have been defined (see previous section), progressive vector transmission is still challenging. Our focus is on vector map data, for which effective compression mechanisms are currently lacking. Indeed, the set of spatial objects that need to be represented in a map depends on its level of detail. In this context, the term detail refers both to the amount of entities contained in the file and to the information stored about such entities, e.g., their geometry (Goodchild and Proctor 1997). Therefore, changes of detail can be caused by the addition or elimination of some entities, as well as by the refinement or coarsening of the representation of existing entities (e.g., change of dimension or shape). Decreasing the level of detail of a vector map is a complex and time-consuming process called generalization whose automation and on-line application are still open problems. In fact, it relies on cartographic principles whose complete formalization is still lacking (see Weibel and Dutton (1999) for a survey). Automated solutions have only been proposed for sub-problems and are based on heuristic methods (e.g., the Douglas-Peucker algorithm (1973)). Therefore, the process still requires interaction between semi-automatic solutions and expert cartographers. A critical issue in map generalization is preservation of consistency (e.g., constraints on overlappings) while decreasing the level of detail (Robinson et al. 1984; McMaster and Shea 1992; M?ller et al. 1995; Weibel and Dutton 1999). For example, random subsampling of objects to be eliminated or to be simplified does not usually generate a consistent representation at coarser detail. In Figure 3 and 4, examples are shown where random subsampling generates topological inconsistencies across different representations. paper mill river lake creek creek mouth major road bridge intersection AB minor roadA minor roadB paper mill river lake creek creek mouth major road bridge intersection AB minor roadA minor roadB Figure 3: Random elimination of some entities (e.g., bridge) results in topologically incorrect representations: the two branches of major road and river have incomplete boundaries. paper mill river lake creek creek mouth major road bridge intersection AB minor roadA minor roadB paper mill river lake creek creek mouth major road bridge intersection AB minor roadA minor roadB Figure 4: Selection of a random subset of entities, namely, major road and river (but not bridge), to coarsen their representation generates a topologically incorrect representation: topological relations between major road and bridge (and between river and bridge) are not consistent across levels. Consistency is an essential property for the usability of data (e.g., for analysis and querying purposes). Therefore, ad hoc techniques and operators need to be applied. Unfortunately, the majority of proposed methods do not intrinsically provide consistency and thus a posteriori checks are required in order to adjust the result when some inconsistency has been introduced. During progressive transmission, consistent datasets need to be transmitted to the client site, where there is no way to discover inconsistencies. Thus, consistency must be checked for on the server site either implicitly by adopting generalization operators that preserve it or by means of a posteriori checks on the generalized datasets. In the following section we propose a solution to progressive vector map data transmission. 4. A Client-Server Architecture for Progressive Vector Transmission In this section we present a distributed client-server architecture that enables progressive transmission of vector data. This work extends our previous research on progressive vector transmission presented in Bertolotto and Egenhofer (1999). The solution we propose meets the architectural requirements for the design of contemporary information systems, involving both heterogeneity of data and processes. In fact, as different tasks are better performed by different system components, distributed environments represent the natural direction for which future generation systems are heading. This direction is in accordance with the exponential network growth and the global connectivity reached in the last few years. As a consequence, and thanks to the technological improvements recently obtained, some of the old systems and, in general, newly developed systems, are being redesigned and embedded into distributed environments (see, for example, Hofmann (1999); Kramer et al. (1997); Kumar et al. (1999)). To enable the progressive transmission of vector files, we pre-compute and store a sequence of consistent representations at lower levels of detail on the server site. Such representations are transmitted in order of increasing detail to the client upon request Figure 5). level 4 level 4 level 3 level 3 level 2 level 1 client server Figure 5: Representations at different levels of detail are maintained on the server and transmitted to the client in order of increasing detail. Besides providing temporary versions of the data on which preliminary operations can be performed, in this way, the user can realize during the process that the detail of the currently displayed representation is good enough for her purpose and so she can decide to interrupt the downloading of more detailed representations (stored in larger files). Therefore, both time and disk space can be saved. An interesting property of some techniques for progressive transmission of raster data (e.g., interleaving techniques) is the fact that they only transmit increments (i.e., sets of new pixels) at each step. Such increments are added to the currently displayed image to improve its resolution without requiring the transmission or downloading of another complete image file. Since vector files can be very large, it would be useful to speed up the transmission in a similar way to reduce network traffic. However, increments between two consecutive levels of detail of a vector dataset can be complex sets of entities that are added to the level at finer detail to refine the representation of a set of entities at the lower level. The integration of such increments with the currently downloaded or displayed representation is a non-trivial task if consistency between different representations must be preserved. From a purely graphical point of view, a major requirement for spatial analysis consists of providing a method for combining the results of several queries into a unique graphical representation. Therefore, the system should allow the addition or elimination of representation layers from the current display. However, in the case of progressive transmission, the user may be interested in more than a visual inspection of the data. She may want to analyze and work with a consistent dataset. A vector file is usually composed not only of a set of points, lines, and polygons, but also of a set of spatial relations linking such entities. Thus, suitable overlaying and integration techniques must be applied not only at the graphical level, but also at the data level to include the computation of spatial relations between newly introduced entities and preserved entities (see Section 8.2). Our solution for progressive vector map transmission relies on a distributed client-server architecture, with a server provided with methods for building, manipulating, and transmitting a sequence of map representations at different levels of detail; and a thin client provided with a set of operations for visualizing as well as updating and integrating the transmitted levels. In summary, the following functionalities are required: ? on the server site: a preprocessing task must be performed in order to build a sequence of map representations at different levels of detail, as well as a progressive transmission technique must be developed to send the different levels one at a time; and ? on the client site: a mechanism must be built in order to have different graphic layers so that the displayed representation is complete at each step (corresponding to the transmission of a given level of detail), and an integration algorithm must be developed and implemented for reconstructing the dataset corresponding to the displayed representation. Thus, the major issues involved in the process are: ? the definition of a model for consistent multiple map representations; ? the definition of a transmission technique; ? the development of a graphic interface and a dynamic visualization tool for displaying different layers corresponding to different levels of detail (during transmission); and ? the development of algorithms to compile transmitted increments into a topologically consistent format for the reconstruction of intermediate levels of detail. In this paper, we deal with modeling aspects involved in progressive vector map transmission, disregarding the mainly graphical issues. In particular, in next section we describe a model for multiple map representations. 5. Multiple Map Representations for Progressive Transmission In this section we describe a model that was initially presented in Puppo and Dettori (1995) and then further formalized in Bertolotto (1998). The model consists of a sequence of map representations of a given area, each corresponding to a different level of detail. Such sequence can be stored on the server and sent progressively to the client upon request (as described in Section 4). In Bertolotto (1998) a set of generalization operators have been defined and shown to be minimal and sufficient to generate consistent transformations of maps. Only topological changes (e.g., changes of dimension, complexity, etc.) are possible by applying such operators: metric and semantic changes are not being taken into account. An example of topological changes in a map is shown in Figure 6. river river bridge Figure Topological changes between two different levels of detail in a map: at higher level (left) the river is represented as a region split into two branches by a bridge (represented as a line); at lower level (right) the bridge is not represented and the set of entities representing the river are reduced to a single line. The set of operators defined is the following: contraction of an open line (including its endpoints) to a point Figure 7(a)). Region contraction: contraction of a simply connected region (with its boundary) to a point (Figure 7(b)). Region thinning: a region (and its bounding lines) is reduced to a line (Figure 7(c)). ? Line merge: fusion of two lines sharing an endpoint into a single line (Figure 7(d)). Region merge: fusion of two regions sharing a boundary line into a single region Figure 7(e)). elimination of an isolated point inside a region (Figure 7(f)). elimination of a line inside a region (Figure 7(g)). (a) (b) (c) (d) (e) Figure 7: Generalization operators. These operators have been formalized by means of functions between the cell complexes describing the original and the resulting map (Bertolotto 1998). For example, in Figure 7(a), if l is the line on the left, p?and p??are its endpoints, and p is the point on the right, the line contraction operator is formalized by the function lc:C? C?, where C and C' are cell complexes, such that: lc(p?)=lc(p?)?l=c(l) =p (1) and "e?C,e?p?,p?,?l,lc (e)= e (2) Equation (2) states that all entities in C other than l, p?and p??are preserved. Furthermore, constraints on mappings of topological relations must be fulfilled. For instance, all lines in the map having an endpoint in either p?or p??are transformed into lines of C?having an endpoint in p. A complete description of such constraints can be found in Bertolotto (1998). Contractions and thinning operators correspond to a decrease of dimension for a group of entities, merge operators group two entities of the same dimension into a single one, while abstraction operators correspond to the elimination of some lower-dimensional entity from a region. These operators are called atomic, because they perform minimal changes, i.e., they modify minimal sets of entities and preserve the others. By composing functions corresponding to such operators, complex map transformations can be defined. As outilined in Section 3, in the case of progressive transmission between two remote sites, the preservation of consistency is a major issue, and it must be guaranteed by the server as the client has no way to control it. When the previously described operators are used, consistency is automatically guaranteed. In fact it has been shown that they represent a minimal and sufficient set of functions that generate (by composition) only consistent transformations for decreasing the level of detail of a map (Bertolotto 1998). In particular, they allow to generate all map transformations that preserve the boundary of entities and whose inverse image preserves connectivity (i.e., they do not allow to model aggregation of non-connected entities3). Topological relations are thus consistently transformed. Therefore, all multiple representation sequences built on the basis of such operators are intrinsically consistent. The model for a multiple representation sequence can be formalized as follows: let M bethe map at higher level of detail (stored on the server) and let f1, f2,., fk be an ordered sequence of map transformations, i.e., a sequence of functions obtained as compositions of atomic operators; let f1: M0?M1, f2: M1?M2,., fk: Mk-?1 Mkbe such that f=fkfk-1. f2f1: M0?Mk is a transformation of maps. The sequence M0 ,., Mk is composed of maps corresponding to less and less detailed representations of the same area, and it is called a multiple representation sequence (see Figure 8 for an example). Figure 8: A multiple representation sequence with three levels of detail. 3 The formalization of aggregation of disjoint entities is traditionally a difficult issue, as this operation might generate semantic and topological inconsistency. A tentative solution generates a fictitious ?shell? containing the entities to be aggregated. However, this method has the drawback of forcing an arbitrary partition (Dettori and Puppo 1996). 6. Encodings for Efficient Transmission The efficiency of the model presented in Section 5 can be improved by defining encodings that avoid redundancy between levels and by adding to it intra-level links to allow navigation and browsing across levels. These issues are discussed in Section 6.1 and Section 6.2. 6.1 Data Structure An important issue relates to the efficient storing of multiple map representations. Some entities are preserved throughout the levels. Such entities should be stored only once in the sequence instead of redundantly encoding them at each level. If such representations are to be transmitted through the Web, reducing the amount of information to be exchanged is highly useful. A straightforward way to avoid redundancy is storing the entire set of entities only for the coarsest level, while all subsequent levels just record newly introduced entities and more refined representations of entities present in previous levels (i.e., the increments). This way transmission is also sped up; however, by not storing the complete set of entities corresponding to each transmitted level, the problem arises of defining algorithms for reconstructing the vector file corresponding to an intermediate level, once the transmission has been terminated, to use, manipulate, and query it. The problem lies in the computation of the spatial relations between newly introduced entities and preserved entities. This problem does not occur if the purpose of the transmission is just visualization. In such a case, only techniques for overlaying graphical layers are necessary. In the following we describe a structure for encoding a sequence M0 ,., Mk of multiple representations in which entities are stored only once to avoid unnecessary duplications. Such a structure is defined inductively on the basis of the inverse image of entities inside each map (Puppo and Dettori 1995). More formally, for 1 ?i?k, given a submap Mi, j of a map Mi , we consider each entity e of Mi, j such that its inverse image through fi contains more than one entity in Mi-1 , i.e., e is not a preserved entity, its representation is more detailed in Mi-1 . Mi-1,p= fi-1(e) is a submap of Mi-1 . The set of submaps of Mi-1 corresponding to all inverse images of entities in Mi that contain more than one entity form level i-1. In this way, the map at the coarsest level (i.e., Mk ) is completely stored, while only increments are stored for intermediate levels. In Figure 9 we show an example of this encoding structure for the multiple representation sequence of Figure 8. Note that entities that are preserved across different levels are represented only at the coarsest level in which they appear (they are drawn as dashed in the intermediate levels in which they are not stored explicitly). Figure 9: An example of encoding structure for the multiple representation sequence of Figure 8. 6.2 Adding Intra-Level Links In this section we enrich a multiple representation sequence with links that connect different representations of the same entities at different levels. We call such links intra-level links or vertical links, since they provide the sequence with a hierarchical structure. Keeping intra-level links facilitates the reconstruction of a representation at a given level of detail and it allows us to know which are the entities that are being modified at subsequent levels and which ones are preserved.An example of intra-level links joining different representations of a river and a bridge at different levels in a multiple representation sequence is shown in Figure 10. level 3 river level 2 river bridge river level 1 bridge Figure 10: Intra-level links joining different representations of a river and a bridge in a multiple map representation sequence. Intra-level links also allow for hierarchical spatial reasoning. In particular, in a querying environment, the user can be interested in retrieving information at different detail. Although a lower level of detail can be sufficient for processing a given query, sometimes a more detailed answer is required. Maintaining intra-level links allows for efficient navigation and browsing across the levels without having to query entire levels. A query is initially performed at the lowest level and the result can be evaluated according to a given user-defined criteria. If the outcome of the evaluation is satisfactory, there is no need to query against a more detailed dataset. Otherwise, the intra-level links are followed in order to find a satisfactory answer. Usually only links between entities that belong to consecutive levels are kept. By combining them, also links between non-consecutive levels can be obtained. For the model described in Section 5, links between entities in consecutive levels are formalized by means of the transformation functions defining the sequence (obtained as compositions of atomic operators). Such functions perform generalizations of maps. The inverse operation, i.e., the refinement of maps, can be formalized by ?inverting? such functions. Although these functions are not injective, in Bertolotto (1998), the inverse image of an entity in the co-domain of atomic generalization functions between cell complexes representing maps has been characterized in such a way that an entity can be refined through a small number of operations. Let p, l, and r denote a point, a line and a region, respectively. It has been shown that the inverse image of a region can only be one of the following sets: ? {r,r?,l} (i.e., the domain of operator region merge), ? {r,p} (i.e., the domain of operator point abstraction), ? {r,l} (i.e., the domain of operator line abstraction). The inverse image of a line can only be one of the following sets: ? {p,l,l?} (i.e., the domain of operator line merge), (i.e., the domain of operator region thinning). Finally, the inverse image of a point can only be one of the following sets: ? {r,l,p} (i.e., the domain of operator region contraction), (i.e., the domain of operator line contraction). Therefore, we can define the inverse of generalization operators (defined in Section 5) as follows: ? Point-region expansion: expansion of a point into a region (with its boundary) p? {r,l,p?} (3) expansion of a point into a line (with its endpoints) p? {l,p?,p?}? (4) ? Line-region expansion: expansion of a line into a region (with its boundary) l? {r,l?,l?}? (5) splitting of a line into two lines sharing an endpoint l? {l?,l?,?p} (6) Region split: splitting of a region into two regions sharing a bounding line r? {r?,r?,?l} (7) insertion of an isolated point inside a region r? {r?,p} (8) insertion of a line inside a region Basically, the inverse operations of contractions and thinning are expansions of an entity into a set of entities including some higher dimensional entity. The inverse of merging is splitting an entity into two entities of the same dimension, while the inverse of abstracting an entity from a region is the insertion of an entity inside a region. As for generalization operators, these refinement operators only perform minimal changes. Thus, for example, in the line split operator, endpoints of line l are mapped onto the corresponding endpoints of lines l?and l??(Figure 11). p" p" l" l p l' p' p' Figure 11: Endpoints of line l are preserved by line split. On the basis of refinement operators, i.e., by considering inverse images of entities, we enrich the model defined in Section 5 and encoded by means of the structure defined in Section 6.1 with intra-level links. The resulting model has a tree-like organization. The root is the map at the coarsest level of detail Mk . Each submap Mi-1,p that corresponds to the inverse image of some entity e belonging to a submap Mi, j at level i in the sequence is stored as a node and is called a child of M . An arc between M and M is established and labeled e. The child relationship represents the refinement of a submap, while the parent relationship represents generalization of a submap. In Figure 12 the example of Figure 9 has been enriched with intra-level links. e e" Figure 12: Adding intra-level links results in a tree-like structure. Arcs are labeled with the corresponding refined entities. In the following section, an example of a multiple representation sequence and its efficient encoding is described. 7. An Example In this section, we describe an example of a multiple representation sequence with four levels of detail (Figure 13). The coarsest level (level 4) contains the hydrographic map of a given geographic area. This level includes the following set of entities {river, creek, paper mill, lake, creek mouth}, such that river and creek are represented by lines, paper mill and creek mouth (i.e., the intersection between creek and lake) are points, and lake is a region. In level 3, the road network covering the same area is also represented: major road, minor road A, minor road B, intersection AB, and bridge are added. The first three entities are represented by lines. Intersection AB is the intersection point between minor road A and minor road B, while bridge is the intersection point between major road and river. Therefore, minor road A, major road and river are represented at this level as lines split into two segments. Level 2 corresponds to a refinement in the representation of paper mill, that is now a region, river, for which the thickness is shown (in both the branches composing it), and bridge, that is now a line. All other entities are preserved, i.e., their representation at this level does not change. Finally, the fully detailed map at level 1 refines the shape of paper mill and shows the thickness of the two segments of line that comprise major road, thus implying a transformation of bridge from line to region. In order to build such a sequence of representations to be stored on the server site, we apply the generalization operators described in Section 5. For example, to obtain level 3 from level 2, the following operators are applied: region thinning of river, line contraction of bridge, and region contraction of paper mill. Figure 14 illustrates the structure described in Section 6.1 for the efficient encoding of the multiple representation sequence of Figure 13. The set of entities in level i (for corresponds to the increment with respect to level i+1. Entities that are present at both level i+1 and i, and whose representation is refined at level i, are highlighted at level i. An example is provided by river, that at level 4 is represented by a single line, while at level 3 is composed of two segments joined by bridge. Level 4 is completely represented, i.e., all entities together with their spatial relations are stored in the encoding structure. For instance, the information that creek mouth is an endpoint of creek is maintained. The set of spatial relations that are explicitly stored depends on the particular data structure adopted to encode each map in vector format (e.g., the extended DCEL structure defined in De Floriani et al. (1993)). Among the entities present at level 3, only river (fragmented into two segments) and the newly introduced entities (major road, bridge, minor road A, minor road B, and intersection are encoded, as well as their spatial relations. At level 2, the only changes are the refinement of paper mill, river and bridge. Only these entities and their relations are represented. Similarly, at level 1, the encoding structure includes just paper mill, bridge and the two segments composing major road. Intra-level links are also maintained between different representations of the same entities at consecutive levels. paper mill lake river creek creek mouth level 4 paper mill lake river creek creek mouth major road level 3 intersection AB bridge minor road A minor road B paper mill lake river creek creek mouth major road bridge intersection AB minor road A minor road B level 2 paper mill lake river creek creek mouth major road intersection AB bridge minor road A minor road B level 1 Figure 13: A multiple map representation sequence with four levels of detail. paper mill lake river creek creek mouth level 4 river level 3 major road intersection AB bridge minor road A minor road B paper mill river bridge level 2 paper mill major road bridge level 1 Figure 14: Efficient encoding of the sequence represented in Figure 13. 8. Discussion and Implementation Issues In Section 5 we have described a model for multiple map representations that can be used for progressive transmission of map data over the WWW. The model is built on the basis of a set of operators that perform topological changes on a map. Although such operators have the interesting property of preserving topological consistency, they only modify the topology of the map. However, in real implementations also changes in the geometric shape of objects are commonly required, such as simplifications of lines (Douglas and Peucker 1973). We are currently investigating the possibility of extending our model to include this kind of operations. Changes depending on the semantic of objects should also be allowed. The incorporation of semantic aspects is another goal of our current research. A model for generating multiple map representations based on the Douglas-Peucker algorithm and used for progressive transmission over the WWW has been described in Section 2.2 (Buttenfield 1999). However, such a model has the drawback of not guaranteeing the preservation of consistency. Furthermore, it only deals with line simplification operations and does not allow other kinds of commonly used map generalization operations (McMaster and Shea 1992). Finally, as underlined in Section 2, a well known question related to the generation of multiple map representations is the following: ?How many and what are the levels of detail to be generated and stored?? The answer to this question strongly depends on the application and on the particular task. For an implementation of our model, we plan to analyze typical queries from different application fields to detect what the best levels of detail are to provide satisfactory answers. Also, as underlined by Buttenfield (1989), it is possible to detect the range of resolution at which some geographic processes become evident. Such considerations will help decide what levels of detail need to be stored on the server site. In the following sections we discuss in more detail issues related to the implementation of the model for progressive transmission proposed in this paper. In particular, we consider the data transmission technique adopted (Section 8.1) and the reconstruction of datasets corresponding to intermediate levels of detail downloaded on the client site (Section 8.2). 8.1 Transmission of Different Levels of Detail In the solution we propose, the process of transmitting a multiple representation sequence between two remote sites presents a few critical points. We assume to have a dataset d on the server and a sequence of multiple representations of d corresponding to k levels of detail, such that the higher the level, the less detailed the representation. We denote such sequence by mr(d,k). We consider a transmission model in which the user is attempting to download progressively the dataset d. Each level of detail corresponds to a separate package. As described in Section 4, the process starts with the server transmitting the representation corresponding to the first level of detail (i.e., the coarsest). We can formalize such process by means of a binary operation t that takes a sequence mr(d,k) and an integer i, with 1 performs the transmission of the representation corresponding to level i in the sequence. Thus, the first step in the transmission process can be denoted by Upon completion of the transmission of the first package, the server continues the transmission of the subsequent representations in order of increasing detail until the user decides to stop the process. We indicate this sequence of steps with In order to avoid unnecessary repetitions for storage saving purposes, in the encoding structure for the model described in Section 6.1, given a representation corresponding to a level i (0<i<k) of detail in the sequence, only the entities that have been modified are stored at level i-1. Not only this model saves space on the server site, but it also speeds up the transmission to the client site, i.e., it minimizes network traffic. On the server site the coarsest representation mr((d,k),k) is completely stored, while only portions of the subsequent representations (increments) are stored in the subsequent levels. Thus, the size of the transmitted packages is reduced. On the client site, a new buffer is used to store the increments to be integrated with the previously transmitted representation. From a graphical point of view, such a buffer can be used as a graphic layer to be superimposed on the previous one in order to completely display the representation at the new level of detail. From a more operational point of view, ad hoc integration techniques must be developed (such as for reconstruction of the topology of the dataset, discussed in next section). A further way to increase the granularity of the transmission is to send subparts of the increments that are topologically disjoint (corresponding to nodes in the tree) as separate packages. This way, topological consistency is still preserved and even smaller packages are transmitted. Also, the user would have the possibility to decide to increase the detail of only some subparts of the map (i.e., to decide to download only the subtrees corresponding to the interesting areas), thus generating a representation at ?variable? detail (the so-called selective transmission). Entities belonging to different levels of detail would coexist in the representation reconstructed on the client site. We are currently investigating this interesting aspect. An important issue concerns the transmission of intra-level links. As discussed in Section 6.2, during query processing, multiple representation sequences benefit from the addition of intra-level links between different representations of the same entity at different levels of detail. A query is first performed at the coarsest level; if the result is unsatisfactory, intra-level links can be followed to find a better answer. In this way we do not need to query the entire following level. Furthermore, as discussed in next section, intra-level links facilitate the reconstruction of the intermediate levels. Thus, such links must be encapsulated in the transmitted packages, in the form of pointers, in addition to the new entities to be displayed in the overlayed buffers. Each transmitted entity maintains a reference to (the ID of) its parent in the previous level. 8.2 Reconstruction of a Dataset Once the user at the client site decides to stop the transmission process, because the level of detail of the currently displayed representation is good enough, the dataset corresponding to that intermediate representation must be reconstructed. Such a dataset is not stored explicitly (i.e., as an independent representation level) on the server either. Thus, on the client site, integration operations are needed to compile a consistent representation at a given level of detail from the collection of displayed layers. Operations for deleting entities in the previous layers as well as for adding new sets of entities (belonging to newer layers) to substitute them are needed on the client site. Furthermore, spatial relations between newly introduced entities and preserved entities must be computed. This operation can be very time consuming. For the hierarchical model we consider, i.e., the sequence of multiple representations described in Sections 5 and 6.1, enriched with vertical links (Section 6.2), a complete representation corresponding to a given level of detail can be obtained by means of a visit of the tree. The visit includes all nodes up to the desired level: entities contained in a given node are deleted and replaced by the set of entities corresponding to the child of such node (i.e., its refinement). The visit of the tree can be performed on the client site as vertical links are transmitted together with increments (i.e., nodes in the tree) in the form of pointers (as discussed in the previous section). Due to the intrinsic consistency properties of the model (in terms of preservation of topology), topological relations between two entities e and e?at level i can be reconstructed on the basis of relations between entities that are parents of e and e?at the preceding level. As an example, we consider the sequence of multiple map representations shown in Figure 13. Such a sequence is stored on the server and transmitted progressively. For each transmitted level, the client receives only the increments and the intra-level links. In this example, only level 4 is completely represented. Therefore, intermediate levels must be reconstructed on the client site. For instance, to reconstruct level 3 from level 4 and the increment between levels 4 and 3 (including: the refinement of river; the new entities bridge, major road, minor road A, minor road B, and intersection AB; and the set of spatial relations among them), the following operations must be performed: ? the line representing river at level 4 must be substituted by the two segments that share bridge as an endpoint; bridge, major road, minor road A, minor road B, and intersection AB must be added as new entities; and ? the complete set of spatial relations for level 3 are reconstructed in a straightforward way as union of the set of relations between preserved entities (stored at level 4) and the set of relations between newly introduced entities (included in the increment package). 9. Conclusions This paper discusses the need for progressive vector transmission, by which a user is provided with coarser versions of the data while she is downloading a large vector file from a remote site. This process presents inherent challenges. Extracting a consistent representation at a lower level of detail from a vector map dataset is a complex and time-consuming operation that cannot be performed on-line during progressive transmission. A solution is to pre-compute, on the basis of sound cartographic principles, a sequence of multiple representations of the data to be stored on the server site. Each representation, corresponding to a different level of detail, is transmitted separately. This solution has the obvious drawback of requiring preprocessing. However, multiple representations can be utilized to define structured digital libraries of summaries of geospatial data. Such organization enables efficient access and manipulation of such data over the WWW. Our contribution is a solution to vector map data transmission, based on a distributed architecture. A model generating multiple map representations (Puppo and Dettori 1995; Bertolotto 1998) that can be used for progressive transmission is described. The model is built by applying a set of generalization operators that perform atomic topological changes on vector maps. The defined operators guarantee the preservation of consistency (Bertolotto 1998), an essential property for the usability of generalized data. Our future goal is to extend such a model to include metric and semantic aspects as well as to develop alternative models and to compare them in terms of storage and transmission costs. The development of efficient encoding strategies plays a central role in this framework. In a more general setting, we are planning the design of a heterogeneous and distributed geographic information system integrating data in different formats (e.g., raster, vector, and textual) and retrieved from different sources. The system will include distributed Internet archives in the form of digital libraries of generalized geospatial data for fast and efficient use. Different formats of the same data will be automatically chosen by the system to perform different tasks on the basis of efficiency. For example, rasterized (scanned) versions of maps will be provided to the user for mainly visual purposes while the corresponding vector versions will be progressively transmitted if actual object manipulation is required. Several aspects will be integrated in the data including spatial, temporal and semantic aspects to allow a wide range of operations and analysis. Another essential requirement of contemporary information systems is related to the concept of interoperability, i.e., the capability of autonomous systems to exchange data and to handle processing requests by means of a common understanding of processes and requests (Sondheim et al. 1999). This can be achieved if common representation models are used (UCGIS 1996). In GIS, data modeling plays a central role in the framework of interoperability, as agreement at the representation level is critical for exchanging geospatial data. Therefore, the design and implementation of our comprehensive Internet-based archive, will meet interoperability and standardization requirements as defined by the OpenGIS Consortium (OGC 1996a and 1996b). 10. --R http://king. Fractals Everywhere Fractal Image Compression: Theory and Application to Digital Images Digital Image Processing Generalization in Digital Cartography (eds.), Taylor and Francis OGC (OpenGIS Consortium) Computational Geometry: An Introduction Elements of Cartography The Design and Analysis of Spatial Data Structures Applications of Spatial Data Structures UCGIS (University Consortium for Geographic Information Science) --TR --CTR Alexander Kolesnikov , Alexander Akimov, Distortion-constrained compression of vector maps, Proceedings of the 2007 ACM symposium on Applied computing, March 11-15, 2007, Seoul, Korea Qiang Han , Michela Bertolotto, A multi-level data structure for vector maps, Proceedings of the 12th annual ACM international workshop on Geographic information systems, November 12-13, 2004, Washington DC, USA
progressive transmission;multiple representations;vector map data
594869
Bayesian Methods for Efficient Genetic Programming.
A Bayesian framework for genetic programming (GP) is presented. This is motivated by the observation that genetic programming iteratively searches populations of fitter programs and thus the information gained in the previous generation can be used in the next generation. The Bayesian GP makes use of Bayes theorem to estimate the posterior distribution of programs from their prior distribution and likelihood for the fitness data observed. Offspring programs are then generated by sampling from the posterior distribution by genetic variation operators. We present two GP algorithms derived from the Bayesian GP framework. One is the genetic programming with the adaptive Occam's razor (AOR) designed to evolve parsimonious programs. The other is the genetic programming with incremental data inheritance (IDI) designed to accelerate evolution by active selection of fitness cases. A multiagent learning task is used to demonstrate the effectiveness of the presented methods. In a series of experiments, AOR reduced solution complexity by 20% and IDI doubled evolution speed, both without loss of solution accuracy.
Table 1. A summary of previous efforts for upscaling genetic programming Subjects Selected references Angeline w3x Banzhaf w7x Bloat Landgon w22x Nordin w27x Wu w39x Zhang w44x Blickle w10x Iba w18x Parsimony Kinnear w19x Rosca w31x Soule w36x Zhang w44, 45x Andre w1x Angeline w2x Modularity Koza w21x O'Reilly w28x Spector w37x Rosca w32x Angeline w4x Banzhaf w7x Operators Chellapilla w11x Koza w20x Luke w24x Poli w29x Gathercole w13, 14x Hillis w17x Subset Schoenauer w33x Siegel w34x Teller w38x Zhang w40, 42x This list contains references which are directly related to the present work, and thus it is far from complete. produced by the solution. Soule et al. w35xobserve that removing non-functional codes at every generation does not halt the program's growth. Instead, the programs generate code which, while functional, is never actually executed. Similar observations have been made in "biological" evolution. That is, introns or non-coding segments emerge in DNA as well. Wu and Lindsay w39xgive a recent review of biological introns. Banzhaf et al. w7xcharacterize introns as having two salient features: An intron is a segment of the genotype that emerges from the process of the evolution of variable length structures, and an intron does not affect the survivability of the individual directly. Introns in GP turned out to be a mixed blessing. On one hand, they may benefit evolution since they enable a genetic program to protect itself against the destructive effect of crossover, and allow the population to preserve highly-fit building blocks w27x. On the other hand, from the practical point of view, introns usually result in run stagnation, poor results, and a heavy drain on memory and CPU time w8, 35x. An explanation for the cause of bloat in GP was provided by Zhang and Muhlenbein w44x. Based on statistical learning theory, they show that the total error of genetic programs can be decomposed into two terms attributed to bias and variance. Since complex models are more expressive and thus better at reducing the bias error than simple models, GP programs tend to grow until they fit the fitness data perfectly unless complex models are penalized to reduce the variance error. Empirical evidence supporting the explanation was given by an analysis of the distributions of fitness vs. complexity of genetic programs generated by a large number of GP runs for solving 7-parity problems. Recently, Langdon and Poli w22xprovide a similar but more general explanation for bloat. They argue that any stochastic search techniques, including GP, will tend to find the most common programs in the search space of the current best fitness. Since in general there are more of these which are long than there are which are short but GP starts with the shorter ones.the population tends to be filled with longer and longer programs. Based on this argument, Langdon w23xpresents new crossover operators that carefully control variation in size and produce much less bloat. For the purposes of upscaling genetic programming, there are several reasons for generally preferring parsimonious programs to complex ones. Parsimonious programs typically require less time and less space to run. This is particularly important during the GP process which may need to store and evaluate populations of hundreds or thousands of programs. When hardware implementation of GP solutions is in mind, or when GP evolves hardware circuits online "evolvable hardware"., simple circuits require less hardware resource and execution time than complex circuits. In addition, statistical theory says that simpler models are likely to generalize better on unseen fitness cases. As shown in w44x, when biases are equal, a simple program has less variance in average than a complex program, resulting in smaller total error. This is the statistical background behind the principle of Occam's Razor w43xand the necessity for parsimony pressure in genetic programming. However, an effective control of program growth is a difficult task, since too much pressure on parsimony that reduces variance error.may lead to loss of diversity and thus result in low accuracy greater bias error. The adaptive Occam method w44xwas presented as a method for striking a balance between accuracy and parsimony of GP solutions. It adapts the complexity penalty in the fitness function during a run so that programs of minimal complexity are evolved without loss of their accuracy. Several researchers support that parsimony pressure is one of the easiest and effective methods for avoiding bloat w10, 18, 19, 31, 35, 36x. Controlling program growth is one method for reducing time complexity as well as space complexity of genetic programming. Another approach to upscaling GP is by increasing modularity and reusability of programs. Koza w21xintroduces the automatically defined functions or ADFs. An ADF is a subroutine that is evolved during a run of genetic programming and which may be called by the main program that is being simultaneously evolved during the same run. He reports that GP with ADFs produced more parsimonious solutions and required fewer fitness evaluation. Angeline and Pollack w2xdevelop an alternative method called module acquisition MA. Rosca and Ballard w32xpresent mechanisms for adaptive representation learning ARL.that creates new subroutines through discovery and generalization of blocks of code. Spector w37xpresents a method for evolving a collection of automatically defined macros ADMs. He shows that ADMs sometimes provide a greater benefit than ADFs. The search process of GP can be made more efficient by designing novel genetic operators. Traditionally, crossover has been considered as the primary operator and mutation as the secondary. This view is consistent with the notion that GP is best made by combining building blocks, as hypothesized in bitstring genetic algorithms. Koza w20xhas argued that mutations have little utility in GP because of the position-independence of GP subtrees. However, the roles of building blocks and of crossover have become increasingly controversial in recent years. Luke and BAYESIAN GENETIC PROGRAMMING 221 Spector w24xexperimentally demonstrate that mutation can in fact have utility and that crossover does not consistently have a considerable advantage over mutation. Banzhaf et al. w6xalso found that increasing the mutation rate can significantly improve the generalization performance of genetic programming. Angeline w4x shows that mutation operations can perform on par with subtree crossover and suggests that the building block hypothesis may not accurately describe the operational chracteristics of subtree crossover. Chellapilla w11xproposes a method for evolving computer programs without crossover. Poli and Langdon w29xpropose various crossover operators and compare their search properties. Haynes w16x how small changes in representation, decoding, and evaluation of genetic programs can increase the probability of destructive crossover and mutation while not changing the search space. More recently, several authors have shown that fitness evaluation in genetic programming can be significantly accelerated by selecting a subset of fitness cases. The basic idea is that, other things being equal, the evolution time can be minimized by reducing the effective data size for each generation. Siegel w34x, for example, describes a GP method for evolving decision trees using subsets of given training cases. Each fitness case has a fitness measure and the training cases which tend to be incorrectly classified by decision trees become more fit, and therefore selected more frequently. This method is motivated by competitive selection of fitness cases in the host-parasite model of Hillis w17x. Similar ideas of competitive selection of fitness cases have been refined by Teller and Andre w38xand by Gathercole and Ross w14x. Here, the fitness of a training case is evaluated only when the cost of evaluating another fitness case is outweighed by the expected utility that the new information will provide. Another class of methods for fitness case selection is based on incremental learning. Schoenauer et al. w33xpropose the successive optimization scheme that gradually refines the fitness. Evolution typically starts by considering a unique fitness case, and additional fitness cases are "gradually" taken into account when the current population meets some performance criterion. Though in a slightly different context, Zhang w40xpresents a selective incremental learning SEL. method in which an "incrementally" chosen subset of given training data is used for fitness evaluation of the model. The data sets are divided into two disjoint sets of candidate and training sets. Each candidate data point has a fitness value, called criticality, defined as the error made by the model. Fitter candidates are then incrementally selected into the training set as learning proceeds. SEL evolves a single data set for a single model. The method presented in Section 5 generalizes this to populations of multiple models and multiple data sets. 2.2. Eoling multiagent cooperation strategies using GP A multiagent learning task is used as a testbed for the Bayesian genetic programming methods. In an single table and four robotic agents are placed at random positions, as shown in Figure 1. A specific location is designated as the destination. The goal of the robots is to transport the table to the 222 ZHANG Figure 1. The environment for multiagent learning. On a grid world of there are four robots, obstacles, and a table. The task of the robots is to transport the table in group motion to the designated target position G. destination in group motion. The robots need to move in herd since the table is too heavy and large to be transported by single robots. The robots share a common control program A . best The objective of a GP run is to find a multi-robot algorithm A that, when best executed by the robots in parallel, causes efficient table transport behavior in group. To evolve the programs, the robots activate each candidate program A for in parallel to run a team trial. At the beginning of the trial, the robot locations are chosen at random in the arena. They have different positions and orientations. During a trial, each robot is granted a total of S elementary movements. The robot is allowed to stop in less than S steps if it reaches the goal. At the end of the trial, each robot i gets a fitness value which was measured by summing the contributions from various factors. The terminal and function symbols used for building GP trees to solve this problem are listed in Tables 2 and 3. The terminal set consists of six primitive actions: FORWARD, AVOID, RANDOM- MOVE, TURN- TABLE, TURN- GOAL, and STOP. The function set consists of six primitives: IF- OBSTACLE, IF- ROBOT, IF- TABLE, IF- GOAL, PROG2, and PROG3. Each fitness case represents a world of 32 by grid on which there are four robots, 64 obstacles, and the table to be transported see Figure 1 for an example of fitness cases. A set of 200 training cases that are generated at random is used for evolving the programs. An independent set of 200 random cases is used to test the performance of the evolved programs. All the robots use the same control program. To evaluate the fitness of robots, we made a complete run of the program for one robot before the fitness of another is measured. The raw fitness value, eg., of individual i at generation g against case c is computed by considering various factors. These include the distance between the target and the robot, the number of steps moved by the robot, the number of collisions made by the robot, the distance between starting and final position of the robot, and the penalty for moving away from other robots. BAYESIAN GENETIC PROGRAMMING 223 Table 2. GP terminal symbols for the multiagent task Symbol Description FORWARD Move one step forward in the current direction AVOID Check clockwise and make one step in the first direction that avoids collision RANDOM- MOVE Move one step in the random direction TURN- Make a clockwise turn to the nearest direction of the table TURN- GOAL Make a clockwise turn to the nearest direction of the goal STOP Stay at the same position Table 3. GP function symbols for the multiagent task Symbol Description IF- OBSTACLE Check collision with obstacles IF- ROBOT Check collision with other robots IF- Check if the table is nearby IF- GOAL Check if the table is nearby PROG2, PROG3 Evalute two or three.subtrees in sequence 3. Bayesian genetic programming In this section we present a theory of genetic programing that is based on Bayesian inference. The general theory is then applied to addressing two important issues in genetic programming, i.e., control of program growth and acceleration of fitness evaluation. This section aims to provide an outline of the Bayesian GP approach Figure 2. An example genetic program for the multiagent learning task. Non-terminal nodes denote checking sensor inputs of the robots, and terminals indicate actions to be taken. The meaning of the symbols is described in Tables 2 and 3. and some distinguishing features in different implementations. The algorithms are detailed in the following two sections. 3.1. Bayesian formulation of genetic programming Genetic programming works by initializing a population of programs and iteratively producing the next generation of fitter programs, i.e. Ag.s Ag4 M, where Ag denotes the ith program at generation g, and M is the population size. Genetic operators such as mutation and crossover are used to produce offspring programs from the parent programs. New generations are produced repeatedly until the maximum number of generations g is reached or some other termination condition is satisfied. The goodness or fitness of a program is measured in terms of a set D of fitness cases or training data and the programs can be considered as a model of the unknown process f generating the data. In the Bayesian GP, the best program is defined as the most probable model, given the data D plus the prior knowledge on the problem domain. Bayes theorem provides a direct method for calculating such probabilities w15x. It states that the posterior i.e. after observing the data D.probability of a program A is PDN A.PA. PAN D.s PD. PDN A.PA. s ,1. A where A is the space of all possible programs in case of A taking discrete values, the integral will be replaced by summation. Here PA.is the prior i.e. before observing the data.probability distribution for the programs, and PDN A.is the likelihood of the program for the data. The relationship between the states of models and their probabilities is established by borrowing the concept of "energy" from statistical physics w26x. We regard the GP system as a thermodynamic system. Every possible state s of the system has some definite energy Es. The system energy can fluctuate and it is assumed that the probability of a thermodynamic.system being in state s, given that the temperature is T is given asPs.s expyEs.rT .,2. Z where Es.is the energy of the system and Z is the normalization constant needed to make the distribution integrate or sum.to one. This distribution is known as the canonical or Boltzmann.distribution. BAYESIAN GENETIC PROGRAMMING 225 Under the canonical distribution, the prior distribution of programs A can be expressed asPA.s expyaFA .,3. ZAa . where Z a.is a normalizing constant, a s 1rT, and F is the "energy" of model AA A in the "equilibrium" state. For example, the energy function, F , can be chosen A as the total number of nodes in genetic program A. This choice of prior distribution says that we expect the program size to be small rather than large, thus implementing a parsimony pressure. Similarly, the likelihood of models A for the fitness data D can be expressed asPDN A.s expybFD .,4. ZD b . where F is an error function for the fitness set D, b controls the variance of the noise, and Z b .is a normalization factor. The likelihood factor gives preference to programs that fit better to have less error for.the fitness cases. Initially, the shape of the prior probability distribution of programs PA.is flat to reflect the fact that little is known in advance. Evolution is considered as an iterative process of revising the posterior distribution of models PAN D.by combining the prior PA.with the likelihood PDN A. In each generation, Bayes theorem 1.is used to estimate the posterior fitness of individuals from their prior fitness values. The posterior distribution PAN D.is then used to generate its offspring. The objective of Bayesian genetic programming Figure 3.is to find a program Ag that maximizes the posterior probability: best est s min arg max PgAig N D. 5. AiggAg. Figure 3. Outline of the Bayesian genetic programming procedure. 226 ZHANG where g is the maximum number of generations and Ag.is the populations of size M: Ag.s Aig 4Mis1.6. The posterior probability P Ag N D. of program Ag is computed with respect to the gth population: PDN Aig .PAig . PgAig N D.s Mg,7. where P DN Ag .is the likelihood and PAg .is the prior probability of or degree ii of belief in. Ag. Note that the posterior probability is approximated by a fixed-size population Ag.which is typically a small subset of the entire program space A. Genetic operators are applied to generate L offspring AX , k s 1,.,L. For- mally, this proceeds in two steps. First, candidates are generated by sampling from the proposal distribution: QgAXkN Aig .8. The specific form of Q ?N ?.is determined by variation operators. For example, for subtree crossover from two parents it is given as: QCAXkN Ai.s PSAjN Ai.PRAXkN Ai, Aj.9. AjgAg. Here PAN A .is the probability of individual A in Ag.being selected as a mate for A and P AX N A , A . is the probability of A and A producing AX by crossover. Then, each candidate generated by genetic operators is accepted with probability PgAXkN D. agAXkN Aig .s min 1, PgAig N D.5, 10. where P Ag N D.is computed by 7.and PAX N D. is the posterior probability of A estimated with respect to the current population: PDN AXk.PAXk. PgAXkN D.s Mg.1. If AX is rejected in 10., then Ag is retained, i.e., AX / Ag. Note that this kiki acceptance function does not exclude the case that A is generated by crossover from Agand another parent Ag Ag., j / i. BAYESIAN GENETIC PROGRAMMING 227 After L offspring AX , k s 1,.,L, are generated, M of them are selected to build the new population: Ag q 1.s Aigq14is1.12. This defines the posterior distribution P Ag N D. at the next generation. It should be mentioned that this formulation of offspring selection is intentionally very general so that it can accommodate various forms of existing selection schemes, such as m, l.selection w5, 25x. In effect, the evolutionary inference step from generation g to g q 1 is considered to induce a new fitness distribution P Ag N D.from priors PAg . through posterior distribution P Ag N D. following Bayes formula, using genetic operators. Based on this theoretical framework we present in the following subsections two examples of Bayesian genetic programming that employ specific techniques for effective control of evolutionary dynamics. Detailed procedures and experimental results are described in Sections 4 and 5. 3.2. GP with the adaptie Occam's razor The first Bayesian GP focuses on the fact that genetic programming iteratively searches populations of more probable more likely in terms of the data and the priors. programs and thus the information gained in the previous generation can be used in the next generation to revise the prior before seeing the data.belief in true programs. Thus, the posterior distribution can be written in the form PDN Aig .Pgy1Aig . PgAig N D.s ,13. Mjs1 PDN Ajg.Pgy1Ajg. where the priors P Ag . are now expressed explicitly as a function of generation rather than the fixed prior P Ag .as in 7. After computing the posterior probabilities for Ag, the priors PA.are revised to PA.by a belief update function u?, ? .: PgA.s uPgy1A., PgAig N D.14. An implementation of u?, ? .will be described in the next section. Basically, the Bayesian GP starts with generating programs according to an initial prior distribution PA.on the program sizes. Typically, the prior distribution is given as uniform, reflecting the fact that little is known in advance about the optimal size of genetic programs. Then, the fitness Fg.of the programs is measured on the training cases, which results in the estimation of the likelihood programs more details in Section 4. Combining the prior and the likelihood by Bayes formula, we get the posterior probabilities P Ag N D. The current prior is then updated to reflect the new information gained from the posterior distribution. Note that we assign prior distributions on the complexity of models programs. In addition, information theory w12xsays that the probability and code length complexity.of models are related as LA.sylog PA. By making use of this, the complexity of programs can be controlled to evolve parsimonious and accurate programs. In fact, we show in Section 4 that the adaptive Occam method presented in w44xis derived from the Bayesian genetic programming framework. 3.3. GP with incremental data inheritance In the second example of the Bayesian approach to GP, we make use of the fact that the Bayes formula suggests an incremental, evolutionary learning rule: infer programs of higher posterior probability from the existing programs by observing new fitness cases. This leads to writing the posterior distribution as PDig N Aig .Pgy1Aig N Digy1 . PgAigN Dig.s Mg ,15. where the data set Dg is now a variable of generation number g and specific to a single program Ag. The collection of Dgconstitutes the data population Dg. ii The observation of new data will lead to update of the prior distribution: PgAN Dig .s uPgy1AN Digy1 ., PgAig N Dig .16. The revision of prior distribution is the same as the process for equation 14., except that the data Dgy1 for the estimation of likelihood and thus the posterior probabilities of programs. is now a function of generation rather than fixed as D. The genetic programming with the incremental data inheritance IDI.method w42xis an example of this approach, where Dg specifically satisfies the following conditions: where D is the entire data set given for training. In this method, the fitness of programs is estimated on incrementally chosen data subsets Dg, rather than on the whole data set D, and thus the evolution is accelerated by reducing the effective number of fitness evaluations. More details are described in Section 5. 4. Bayesian GP for parsimonious solutions Statistical theories suggest that models which are too simple lack sufficient learning capability while models which are too complex may generalize poorly on unseen BAYESIAN GENETIC PROGRAMMING 229 data. As reviewed in Section 2.1, several researchers have observed that the program size tends to grow without bound or "bloat" see, for example, w22x. In this section we describe a Bayesian method for evolving parsimonious programs. 4.1. Algorithm description The GP algorithm for the adaptive Occam method Figure 4.is the same as that of the Bayesian GP procedure described in Figure 3, except three differences. The first is the raw fitness calculated in step 2. The derivation of the fitness function is given in Section 4.2. The second difference is in step 3, where the posterior probability 13.is substituted for 7.; in 13.the prior is revised each generation while in 7.it is constant. The third is the additional step 6 for the revision of priors. 4.2. Fitness ealuation For a convenient implementation of the Bayesian evolutionary algorithm we take the negative logarithm of the posterior probability P Ag N D. and use it as the fitness function Fig.sylog PgAig N D.,18. Figure 4. Outline of the Bayesian genetic programming with the adaptive Occam's razor AOR. where Ag g Ag. Then the evolutionary process is reformulated as a minimization process est s min arg min Fig.,19. AiggAg. where the fitness function is expressed as Fig.sylog PDN Aig .y log PAig .20. As described in the previous section, we can write the likelihood function in Bayes' theorem 1.in the form w9xPDN A.s expybFD . 21. ZD b . where F is an error function, b controls the variance of the noise, and Z b .is a normalization factor. If we assume that the data has additive zero-mean Gaussian noise, then the probability of observing a data value y for a given input vector x would be where Z b .is a normalizing constant. Provided the data points are drawn independently from this distribution, we have PDN Aig.s PycN xc, Aig. 23. cs1s expybFD .24. ZD b . where x , y .g D are training cases and F is given as cc D FDs EDN Aig.s f xc; Aig.y yc.2.25.cs1 If we also assume that a Gaussian prior on the architecture of program Ag,we ZAa . BAYESIAN GENETIC PROGRAMMING 231 where Z a.is a normalizing constant. For example, F can be chosen in the form AA FAs CAig .s uk2 27.ks1 where u are the parameters defining the program Ag. This choice of prior distribution says that we expect the complexity parameters to be small rather than large, thus implementing a parsimony pressure. Substituting 24.and 26.into 20., the fitness function can be expressed as Fig.s bFDq aFA 28. s bEDN Aig .q aCAig .,29. where the first term reflects the error and the second the model complexity. The exact calculation of the constants b and a in equation 29.requires the true probability distribution of underlying data structure, which is in most real situations unknown. Instead, we define an adaptive fitness function in its most general form as Fig.s Eig.q a g.Cig.,30. where Eg.and Cg.are the measures for error and complexity of the program, ii and the parameter b is absorbed into the adaptive parameter ag.which balances the error and complexity factors as follows: if Ebestgy 1.) e a g.s~ N 2 C^bestg. 31.N 2 Ebestgy 1.? C^bestg. otherwise. This is the adaptive Occam method w44x. User-defined constant e specifies the maximum training error allowed for the run. Egy 1.is the error of the best best program of generation g y 1. C^ g.is the size of the best program at generation best g estimated at generation g y 1. These are used to balance the error and complexity terms to obtain programs as parsimonious as possible while not sacrificing their accuracy. The procedure for estimating C^ g.is given in w44x. best 4.3. Discussion The necessity and difficulty of non-coding segments or introns in genetic programming have been studied by many authors w22, 27, 39x. The adaptive Occam method deals with the non-coding segment problem using a two-phase strategy by means of an adaptive fitness function. In the first stage during Egy 1.) e., the best growth of non-coding segments is encouraged to increase the diversity of partial solutions. In the second stage during Egy 1.F e., a strong parsimony best pressure is enforced to prefer compact solutions. The transfer from the first stage to the second is controlled by Bayesian inference under the constraint of the user-defined parameter e. Note that the posterior probability P Ag N D. can be computed from the fitness values Fg.by taking its exponential function: PDN Aig .Pgy1Aig . Mjs1 PDN Ajg.Pgy1Ajg. expyEig.y a g.Cig. s 32. Mjs1 expyEjg.y a g.Cjg. This shows that the revision of priors P Ag . is implemented as the update of ag.in the adaptive Occam method since the priors are reflected in C^ g.which best leads to revision of ag. We also note that minimization of Fg.is equivalent to the minimum description length MDL.principle w18, 30x: the best model is a model whose total code length for model description L AgN Dg.and data description LDgN Ag. are ii ii minimal. In information theory w12xthe optimal description length for a model is given as the negative logarithm of probability of the model: LAig .sylog PAig .3. Similarly, the code length for the data given the program Ag is given as: LDN Aig .sylog PDN Aig .4.34. .correspond to the two terms in 20. ii 4.4. Experimental results The adaptive Occam method for complexity control was applied to the multiagent learning task. Experiments have been performed using the parameter values listed in Table 4. The terminal set and function set consist of six primitives, respectively, as given in Tables 2 and 3. A set of 200 training cases was used for evolving the programs. An independent set of 200 cases was used for evaluating the generalization performance of evolved programs. The Eg.-values of program i at generation g are measured as the average of its raw fitness values eg.less is better.for the cases c: Eig.s ei, cg.,35. BAYESIAN GENETIC PROGRAMMING 233 Table 4. Parameters used in the experiments for GP with the adaptive Occam's razor Parameter Value Population size 100 Max generation Crossover rate 0.9 Mutation rate 0.1 Training cases 200 Test cases 200 where N is the number of fitness cases in the training set. Each eg.-value was computed by considering such factors as the distance between the target and the robot, the number of collisions made by the robot, and the distance between starting and final position of the robot. Note that in the multiagent task there is no target output given for the training case. The goodness of a program is measured by scores or penalties.it collects by running the robots. The complexity of a GP tree is defined as the size and depth of the tree: Cig.s where k is a constant. We used k s 2. Figure 5 compares the fitness values for GP with and without AOR, averaged runs. Shown are the Eg.-values of the best individuals in each generation for each method. A tendency can be observed that the GP with the adaptive Occam's razor converges slightly faster than the baseline GP, though there is no significant difference in their final fitness. Figure 6 compares the typical changes of program complexity for both methods. The tree complexity was measured in terms of the tree size plus Figure 5. Comparison of fitness values Eig.-component only.of the genetic programs for the multiagent task. A tendency can be observed that the GP with the adaptive Occam's razor converges slightly faster than the baseline GP, though there is no significant difference in their final fitness. Figure 6. Evolution of the complexity of GP trees for the multiagent task. The GP with the adaptive Occam's razor AOR.promotes the trees to grow when significant fitness error.reduction is required, while it prefers smaller trees to larger ones when their fitness error.is comparable. In contrast, the GP without the Occam factor tends to grow as generation goes on. More detailed results are summarized in Table 5. The GP with AOR achieved, on average, a 20% reduction of program complexity without loss of solution accuracy in fact, with a slight improvement both in training and test performances. It can be concluded that the adaptive Occam method evolves smaller programs without loss of generalization capability of the programs. 5. Bayesian GP for accelerated evolution 5.1. Algorithm description The algorithm for incremental data inheritance Figure 7.is the same as the GP with AOR, except that the training set Dg now increases with generation. The additional step for this modification is step 5, where the training set grows by inheritance. The next section details the data inheritance procedure. Table 5. Effects of the adaptive Occam's razor GP programs for the multiagent learning task Method Complexity Baseline 31.4 " 6.5 AOR.on the complexity and average fitness values of Average Fitness Training Test The sizes of training and test sets were 200, respectively. The values are averaged over ten runs. BAYESIAN GENETIC PROGRAMMING 235 Figure 7. Outline of the Bayesian genetic programming with incremental data inheritance IDI. 5.2. Data inheritance procedure The basic idea in genetic programming with incremental data inheritance is that programs and their data are evolved at the same time. With each program is associated a separate data set. We describe a variant of uniform crossover that we call uniform data crossover. A simplified example for illustrating this process is given in Figure 8. First, two parent data sets, Dg and Dg, are crossed to inherit their subsets to two offspring data sets, Dgq1 and Dgq1. Second, the data of parents' are mixed into a union set which is then redistributed to two offspring Dgq1 and Dgq1, where the size of offspring data sets is equal to N s N q l, where l G 1 is the data increment size. Thus, the size of data sets monotonically increases as generation goes on. To maintain the diversity of the training data during inheritance we import some portion of data from the base set. The import rate r is given as ris r ? 1 y di.,0F r F 1. 38. Figure 8. An illustrative example for data inheritance in IDI. Two parent data sets, Dig and Dig, are merged to form the union set, Digqj, which is then inherited to two offspring data sets, Digq1 and Djgq1. The off-spring data points with shaded circles are the examples imported from the base data set D to maintain the diversity of data sets. where r is a constant for import strength. The diversity d is measured as the ratio of distinctive examples in the union set: d sy1, 0F d F 1. 39. Figure 8 illustrates the process of data inheritance, where two parent data sets of size 6 each are unioned to form the genetic pool of size 10, from which two offspring data sets of size 8 each are inherited. Marked are the data imported from the basis data set to maintain the diversity. 5.3. Fitness ealuation As in the implementation of the Bayesian GP with the adaptive Occam's razor, we take the negative logarithm of P Ag N D. and use it as the fitness function Fig.sylog PgAig N Dig .,40. where Ag g Ag., Dg ; D, and P AgN Dg.are defined by 15.: ii gi gg PDig N Aig .Pgy1Aig N Digy1 . PgAiN Di.s Mg gy1 .41. Then the evolutionary process is reformulated as a minimization process est s min arg min Fig.,42. gFgmam Aig, Dig where the fitness function is expressed as Fig.sylog PDig N Aig .y log Pgy1Aig N Digy1 .43. BAYESIAN GENETIC PROGRAMMING 237 Using the similar arguments as in Section 4.2 we can write the likelihood function and the prior distribution of programs in the formPDigN Aig.s expybFD. 44. ZAa . where Z b .and Z a.are normalizing constants. D A We define an adaptive fitness function in its most general form as Fig.s Eig.q a g.Cig.,46. where Eg.and Cg.represents the error factor F and complexity factor F ii D A with a and b absorbed into ag. The parameter ag.balances these two factors as follows if Ebestgy 1.) e a g.s~ Ng2 C^bestg. 47.Ng2 Ebestgy 1.? C^bestg. otherwise. This is a generalization of the adaptive Occam method in that N is now a variable, rather than fixed value N, and scheduled to increase monotonically as a function of generation g. 5.4. Discussion The incremental data inheritance method has interesting properties that characterize the stability of the algorithm as generation goes on. It maintains a growing subset of given fitness cases for each program. The monotonic growth of the fitness subset ensures that the IDI method eventually achieves the same level of performance as that of the baseline algorithm. This is contrasted with many of existing methods for fitness case selection, including LEF w14xand RAT w38x. Maintaining a separate subset for each program allows the learned component of the program can be retained during evolution. At the same time, the import of new fitness cases from the base set allows the programs to learn new situations, thus leading to gradual improvement in performance. 5.5. Experimental results We compare the performance of the GP with incremental data inheritance IDI.to the baseline GP, i.e. one that uses the complete training cases from the outset. Experiments have been performed using the parameter values listed in Table 6. GP Table 6. Parameters used in the experiments for GP with incremental data inheritance Parameter Value Population size 100 Max generation Crossover rate 0.9 Mutation rate 0.1 Training cases 200 Test cases 200 Initial data size N . 20Data increment size l. 6 runs with IDI used 20 q 6g examples for each generation g selected from the given data set, i.e. N s 20, l s 6, for fitness evaluation. For all methods, a totalof 200 training cases were used for training and an independent set of 200 test cases was used for evaluating the generalization performance of evolved programs. Fitness factors Eg.and Cg.are the same as in the experiments for GP with the ii adaptive Occam's razor, except that the training set size for Eg.is now N ig instead of N. Results are shown in two different forms. Figure 9 shows the evolution of fitness as a function of the generation number, in which there is no significant difference in the performance. Since the GP with IDI uses variable data size, computing time at each generation should be measured by a product of the population size and the data size. This result is shown in Figure 10, where IDI achieved a speed-up factor of approximately four compared with the baseline GP. More detailed results are summarized in Table 7. The incremental data inheritance IDI.method used only 50% of the time required for the baseline GP. It is interesting to see that, despite Figure 9. Comparison of fitness values as a function of the generation number. The curves are mean values over ten runs. The GP with IDI shows no loss of fitness values compared to that of the baseline GP algorithm. BAYESIAN GENETIC PROGRAMMING 239 Figure 10. Comparison of fitness values as a function of the number of function evaluations. The curves are mean values over ten runs. The GP with IDI converges much faster than the baseline GP algorithm. Table 7. Effects of incremental data inheritance IDI.on the time and average fitness values lower is better.of GP programs for the multiagent learning task. The sizes of training and test sets were 200, respectively. The values are averaged over ten runs. Time is measured as the total number of fitness evaluations. Also shown are the standard deviations. Average Fitness Method Time Training Test Baseline 1,220,000 215.8 " 8.3 221.3 " 7.6 IDI the reduced data size, the generalization performance of the GP with IDI was slightly better than that of the baseline GP. 6. Concluding remarks We have presented a Bayesian framework for genetic programming. Two specific GP algorithms were derived from this framework for reducing the time and space complexity of genetic programming. Applied to the multiagent learning task, the first method, i.e. GP with the adaptive Occam's razor AOR., achieved approximately 20% reduction of program complexity without any loss in fitness values. The second Bayesian approach to genetic programming, i.e. the incremental data inheritance IDI.method, used only 50% of the time required for the baseline GP to achieve the same or a little better fitness level. Though this improvement is significant, it should be mentioned that there is still much room for further reduction of spacertime complexity of genetic program- ming. Future work should address the following issues, among others. One is incorporating better genetic operators. We have focused in the present work on dynamics at the phenotypic level. Further improvement can be achieved by finding more "intelligent" variation operators that adapt to the dynamics at the genotypic level. In terms of Bayesian genetic programming, this involves adapting proposal functions. The second issue is to improve modularity of genetic programs by automatically designing and adapting reusable submodules such as ADFs or libraries. This, combined with the Occam's razor, will further improve the comprehensibility and reusability of genetic programs as well as speed up the GP process. From the theoretical point of view, the Bayesian framework for genetic programming provides a number of important features. One is that, by formulating the GP process as Bayesian inference, principled techniques for driving evolutionary dynamics of conventional GPs can be developed. In addition to the methods presented in this paper, one can think of other methods within the framework. For example, decision-making can be made more robust by combining multiple programs instead of a single GP tree. The Bayesian GP framework provides a principled way to combine multiple programs to build a committee machine. Another important feature of Bayesian inference is that it allows background knowledge in the problem domain to be incorporated in a formal way. For instance, if we can guess the distribution of specific function symbols for good GP trees, this knowledge can be reflected in the application probabilities of genetic operators. Background knowledge is important especially for solving real-life problems of practical interest. Finally, the Bayesian analysis of genetic programming appears to be a useful tool for incorporating various genetic programming procedures into a uniform frame- work. A theoretical framework is crucial for the design and comparative analysis of various genetic programming algorithms. Acknowledgments This research was supported by the Korea Science and Engineering Foundation KOSEF.Granta 981-0920-350-2, by the Korea Research Foundation KRF. Granta 1998-001-E01025, and by the Korea Ministry of Science and Technology through KISTEP Granta BR-2-1-G-06. Thanks to Wolfgang Banzhaf and three anonymous reviewers for helpful comments that improved the readability of the paper. --R "Automatically defined features: The simultaneous evolution of 2-dimensional feature detectors and an algorithm for using them," "Genetic programming and emergent intelligence," An Introduction "Evolving compact solutions in genetic programming: A case study," "Evolutionary programming with tree mutations: Evolving computer programs without crossover," Elements of Information Theory "Dynamic training subset selection for supervised learning in genetic programming," "Small populations over many generations can beat large populations over few generations in genetic programming," Bayesian Data Analysis "Perturbing the representation, decoding, and evaluation of chromosomes," "Co-evolving parasites improves simulated evolution as an optimization procedure," "Genetic programming using a minimum description length principle," "Generality and difficulty in genetic programming: Evolving a sort," On the Programming of Computers by Means of Natural Selection "Scalable learning in genetic programming using automatic function definition," "Fitness causes bloat: Mutation," "Size fair and homologous tree crossovers," "A comparison of crossover and mutation in genetic programming," "The science of breeding and its application to the breeder genetic algorithm," "Probabilistic inference using Markov chain Monte Carlo methods," "Explicitly defined introns and destructive crossover in genetic programming," "On the search properties of different crossover operators in genetic programming," "Stochastic complexity and modeling," "Analysis of complexity drift in genetic programming," "Discovery of subroutines in genetic programming," "Evolutionary identification of macro-mechanical models," "Competitively evolving decision trees against fixed training cases for natural language processing," "Code growth in genetic programming," "Effects of code growth and parsimony pressure on populations in genetic programming," "Simultaneous evolution of programs and their control structures," "Automatically choosing the number of fitness cases: The rational allocation of trials," "Empirical studies of the genetic algorithm with noncoding segments," "Accelerated learning by active example selection," "A Bayesian framework for evolutionary computation," "Genetic programming with incremental data inheritance," "Genetic programming of minimal neural nets using Occam's razor," "Balancing accuracy and parsimony in genetic programming," "Evolutionary induction of sparse neural trees," --TR --CTR Byoung-Tak Zhang, A unified Bayesian framework for evolutionary learning and optimization, Advances in evolutionary computing: theory and applications, Springer-Verlag New York, Inc., New York, NY, Sean Luke, Modification point depth and genome growth in genetic programming, Evolutionary Computation, v.11 n.1, p.67-106, Spring
bayesian genetic programming;incremental data inheritance;adaptive Occam's razor;probabilistic evolution;parsimony pressure;data subset selection
594976
Graph Coloring with Adaptive Evolutionary Algorithms.
This paper presents the results of an experimental investigation on solving graph coloring problems with Evolutionary Algorithms (EAs). After testing different algorithm variants we conclude that the best option is an asexual EA using order-based representation and an adaptation mechanism that periodically changes the fitness function during the evolution. This adaptive EA is general, using no domain specific knowledge, except, of course, from the decoder (fitness function). We compare this adaptive EA to a powerful traditional graph coloring technique DSatur and the Grouping Genetic Algorithm (GGA) on a wide range of problem instances with different size, topology and edge density. The results show that the adaptive EA is superior to the Grouping (GA) and outperforms DSatur on the hardest problem instances. Furthermore, it scales up better with the problem size than the other two algorithms and indicates a linear computational complexity.
Introduction The main goal of this paper is to present the results of an experimental study on solving graph coloring problems by Evolutionary Algorithms (EAs). In particular we show the working of a new, problem independent adaptive mechanism for constraint handling in EAs. In Section 2 a brief overview of graph coloring problems and graph coloring techniques is given. We decide to investigate graph instances that are 3-colorable and identify the instances around the so-called phase transition as the most challenging ones. Two graph coloring algorithms are selected to serve as competitors to our EAs. Among the traditional graph coloring techniques we choose DSatur from Br'elaz for its reported high performance, (Br'elaz, 1979). Besides, we also test the Grouping GA (GGA) of Falkenauer, because graph coloring can be seen as a grouping problem and the GGA shows excellent performance on grouping problems, such as bin packing (Falkenauer, 1994). In Section 3 performance measures for comparison of different algorithms are discussed. Thereafter, in Section 4 the Grouping GA is described; in Section 5 genetic algorithms with integer representation and with order-based representation are compared. In Section 6 we present our adaptive mechanism that is modifying the penalty function during the evolution, and we show that this adaptive mechanism highly increases performance. In Section 7 a big comparison between our adaptive EA, the Grouping GA and DSatur is made. We observe that the Grouping GA exhibits the lowest performance and that the adaptive EA outperforms DSatur on the hardest problem instances, moreover it scales up better with the problem size. Finally, we draw conclusions in Section 8. Graph Coloring The main motivation of the present research is our interest in the applicability of genetic, or more generally evolutionary, algorithms to constraint satisfaction problems. In general, a constraint satisfaction problem (CSP) is a pair hS; OEi, where S is a free search space, i.e. a Cartesian product of sets and OE is a formula, a Boolean function on S, (Eiben and Ruttkay, 1997). A solution of a constraint satisfaction problem is an s 2 S with Usually a CSP is stated as a problem of finding an instantiation of variables within the finite domains such that constraints (relations) c prescribed for (some of the) variables hold. The formula OE is then the conjunction of the given constraints. It has been proved that every CSP can be equivalently transformed to a binary one, i.e. to a CSP where each constraint concerns exactly two variables, (Nudel, 1983). For graph coloring problems this property is natural, if we envision nodes as variables and edges as constraints. This suggests that the findings on using a genetic algorithm with no specific knowledge on the problem structure can be applicable to other CSPs as well. In the family of graph coloring problems an undirected graph E) is given and the problem is to color each node in such a way that no two nodes connected by an edge are colored with the same color. There exist several variations of this problem, like finding the least number of colors that is needed to color the graph, or to find the largest subgraph in G that can be colored with the given number of colors. All of these problems are known to be NP-complete (Garey and Johnson, 1979), so it is unlikely that a polynomial-time algorithm exists that solves any of these problems (Arora et al., 1992). A pure constraint satisfaction problem is the graph 3-coloring variant, where each node in the graph is to be colored with one of three given colors. In this problem variant, there is no optimization involved, such as minimizing the number of colors used for a conflict-free coloring. We have chosen to perform an in-depth investigation of this variant, restricting ourselves to one problem type, but studying three different graph topologies, several problem sizes and a whole range of edge connectivity values, yet keeping the number of experiments manageable. 2.1 Problem Instances In the literature there are not many benchmark 3-colorable graphs and therefore we create graphs to be tested with the graph generator 1 written by Joe Culberson. This generator creates various classes of k-colorable quasi-random graphs. The classes we want to investigate are the following. Arbitrary 3-colorable graphs where vertices are randomly assigned to one of the 3 partition elements (colors) uniformly and independently. This is a class that has widely been investigated. We use the term 'arbitrary' to distinguish these kind of graphs from the following ones. Equi-partite 3-colorable graphs where the three color sets are as nearly equal in size as possible (the smallest sets having one element less than the largest). Culberson reports that these graphs present the most difficulty in obtaining the specified coloring for a suite of various algorithms, (Culberson and Luo, 1996). Because the sets are almost equal in size, an algorithm has less information to make use of. In case of flat 3-colorable graphs also the variation in degree for each node is kept to a minimum, so this class is even tougher, since even less information is available to the algorithm. Creating test graphs happens by first pre-partitioning the vertices in three sets (3 colors) and then drawing edges randomly. For the first two types of graphs, once the pre-partitioning has been done, a vertex pair v; w is assigned an edge with fixed probability p, provided that v and w are not in the same color set. So there is some variation in the degree for each vertex. This measure p is called the edge connectivity of a graph instance. When creating the test graphs a random number generator is needed. This generator is fed with a random seed, which obviously influences the created graphs, and - as it turns out from the experiments - also effects the performance of the graph coloring algorithms. In our preliminary experiments comparing different genetic algorithms equi-partite graphs will be used (Section 5 and Section 6), while using all three classes in a final comparison with two other methods (Section 7). For specifying the investigated graph instances we use the notation G eq;n=200;p=0:08;s=5 , standing for an equi-partite 3-colorable graph with 200 vertices, edge probability 8% and seed 5 for the random generator. It is known that for many NP-complete problems, typical instances can be very easy to solve. Turner found that many k-colorable graphs are easy to color, so comparisons between algorithms based on these graphs are not very meaningful, (Turner, 1988). For an interesting comparison one should look for hard problem instances that pose some challenges for candidate algorithms. Cheeseman et al. (Cheeseman et al., 1991) found that NP-complete problems have an 'order Source code in C is available via ftp://ftp.cs.ualberta.ca/pub/GraphGenerator/generate.tar.gz parameter' and that the hard problems occur at a critical value or phase transition of such a parameter. For graph coloring, this order parameter is the edge probability or edge connectivity p. Using the cost function estimation of Clearwater and Hogg, (Clearwater and Hogg, 1996), one can determine the approximate location of the phase transition depending on the number of nodes n, in the graph. The phase transition for our type of problems will occur when the edge connectivity values are around 7=n - 8=n. These values are a bit different from the estimate in (Cheeseman et al., 1991), but our experiments confirm this range and indicate that the hardest graphs are those with an edge connectivity around 7=n - 8=n, independently from the applied graph coloring algorithm. In this investigation we concentrate on graphs in this region. 2.2 Graph Coloring Algorithms There are many traditional graph k-coloring techniques, based on heuristics. Some existing algorithms are: an O(n 0:4 )-approximation algorithm by Blum (Blum, 1989), the simple Greedy algorithm (Kucera, 1991), DSatur from Br'elaz (Br'elaz, 1979), Iterated Greedy(IG) from Culberson and Luo (Culberson and Luo, 1996), XRLF from Johnson et al. (Johnson et al., 1991). Probably the most simple and best known algorithm is the Greedy algorithm, which takes some ordering of the nodes of a graph and colors the nodes in this order with the smallest color 2 that is possible without violating constraints. Grimmet and McDiarmid (Grimmet and McDiarmid, have shown that for almost all random graphs the Greedy algorithm uses no more than about twice the optimal number of colors. Nevertheless, several studies showed that in practice and in theory the Greedy algorithm performs poorly (Kucera, 1991; Turner, 1988). We only mention it here because our GA with order-based representation uses it as a decoder. DSatur from Br'elaz (Br'elaz, 1979) uses a heuristic to dynamically change the ordering of the nodes and then applies the Greedy method to color the nodes. It works as follows. ffl A node with highest saturation degree (= number of differently colored neighbors) is chosen and given the smallest color that is still possible. ffl In case of a tie, the node with highest degree (= number of neighbors that are still in the uncolored subgraph) is chosen. ffl In case of a tie a random node is chosen. Because of the random tie breaking, DSatur becomes a stochastic algorithm and just like for the GA, results of several runs need to be averaged to obtain useful statistics. For our investigation the backtrack version of Turner (Turner, 1988) has been implemented, which backtracks to the lastly evaluated node that still has available colors to try. One search step is defined as expanding a node with a new color, including those done after a backtracking. The DSatur heuristics performs very well, therefore we use it as a heuristic technique to compare our EAs with. Quite some research has been done about graph coloring with genetic algorithms, like Fleurent and Ferland who have successfully considered various hybrid algorithms in (Fleurent and Ferland, 1996a), and who have extended their study into a general implementation of heuristic search 2 The colors are ordered just to make selection possible. methods in (Fleurent and Ferland, 1996b). Others include von Laszewski who has looked at structured operators and has used an adaption step to improve the convergence rate of a genetic algorithm (Laszewski, 1991). Davis' algorithm in (Davis, 1991) was designed to maximize the total of weights of nodes in a graph colored with a fixed amount of colors. This resembles the problem definition used in the SAW-ing algorithm where we also add weights to the nodes of the graph that is being colored. The difference is that the SAW-ing algorithm uses variable weights to guide its search, while Davis' algorithm sees the fixed weights as the problem instance and then tackles this problem as an optimization problem. Coll, Dur'an and Moscato discuss graph coloring and crossover operators in a more general context in (Coll et al., 1995). Performance Measures of Algorithms The comparisons between different GAs and comparisons of GAs and DSatur will be based on two different measures: success rate and computational effort. Probabilistic algorithms may find a solution in one run and may not find one in another run. To cope with this problem we execute a number of runs and monitor how many times a solution is found. The measure success rate (SR) is the percentage of runs in which a solution is found, it gives an estimation of the probability of finding a solution in one run. The most obvious measure of computational effort for evaluating algorithms is the time complexity, e.g., CPU time in seconds, (Kronsjo, 1987). A better option, however, is an implementation and hardware independent measure, computational complexity, i.e., the number of basic operations required to find a solution. For search algorithms this is the number of basic search steps a particular algorithm uses. Unfortunately, different search algorithms can use different search steps. For a GA a search step is the creation (and evaluation) of a new individual, in our case a new coloring. Thus, computational complexity will be measured by the average number of fitness evaluations in successful runs, denoted as AES (Average # of Evaluations to Solution). Fair as this measure may seem, even different GAs may not fully be comparable this way. For instance, a GA using a heuristic within the crossover operator performs 'hidden labor' with respect to a GA applying a standard `blind' crossover. In other words, the extra computational efforts in performing a heuristic crossover are not visible if we compare the AES values. A search step of DSatur is a backtracking step, i.e. giving a node a new color. Thus, the computational complexity of DSatur is measured differently than that of a GA - a problem that is rooted in the different nature of the algorithms and cannot be circumvented. Despite of these drawbacks we compare algorithms by AES in the first part of Section 7, because this measure is independent from implementational and executional issues, such as the processor, programming language, network load, etc. In the second part of Section 7, however, we compare the scaling-up behavior of the investigated algorithms, that is we show how the AES values change with growing problem sizes. Regardless of the different meaning of AES for different algorithms this comparison is by all means fair. Grouping Genetic Algorithm The Grouping Genetic Algorithm (GGA) was introduced by Falkenauer in (Falkenauer and Delchambre, 1992) and further studied in (Falkenauer, 1994; Falkenauer, 1996). The GGA tries to capture and exploit the structure of grouping problems by using an appropriate chromosomal representation and genetic operators. The outlines of the GGA are given in Figure 1, while the different steps are explained in the reminder of this section. Grouping GA Initialize population Evaluate population while not Stop-condition do Sort the population using 2-tournament selection Apply crossover to the best N c individuals Replace the worst N c with the offspring Mutate Nm randomly selected individuals in the population Apply inversion to N i randomly selected individuals in the population Evaluate population while Figure 1: Pseudo code of the Grouping Genetic Algorithm In general, the grouping representation consists of two parts: an object part and a group part. The object part consists of n genes, where n is the number of objects to be grouped, the group part consists of a permutation of the k group labels. An object gene can take any of the k group labels as allele, indicating that the object in question belongs to group of the given label. In a graph coloring context objects are nodes and groups are colors; an example of a chromosome for shown in Figure 2. The group part shows that three are colors A, B and C are used for coloring the graph. The object part discloses that node 2 and 6 are colored with color A, nodes 1, 3 and 4 are colored with color B and node 5 is colored with color C. objects groups Figure 2: Example of a chromosome in grouping representation. The GGA uses genetic operators on the group part of the chromosomes, and adjusts the object part to the corresponding changes in the group part. We will use the three operators described in (Falkenauer, 1994), which are crossover, mutation and inversion. The crossover is by far the most difficult operator of these three, it uses a number of steps to compute two new chromosomes from two parents. These steps are the following. 1: Copy parent 1 to child 1 and copy parent 2 to child 2. Select at random two crossing sites, delimiting the crossing section, in each of the two parents' group part. Parent Parent 2: Inject the contents of the crossing section of the first parent before the first crossing site of the second child. Because we are working with the group parts, this means we are injecting groups of the first parent into the second child. 3: Overwrite the object part of child 2 such that membership of the newly inserted groups is enforced (inherited from parent 1). 4a: Adapt the resulting groups to satisfy the constraints of the problem. Here we throw away groups that have become empty or lost an object in step 3. This can result in objects which are not assigned to a group. These objects, which are marked by an x, are put in a queue. 4b: We now have to reinsert the objects which reside in the queue. This can be done by any heuristic function, as long as the representation remains valid. We do this by looking at the node we want to insert and, if possible, assigning a color, by using a first-fit heuristic on a random order of the available colors, such that no constraints are violated. If this is not possible, we create a new group and assign the object to this group. Note, that this step can lead to an increase in the number of colors. 5: Execute the steps 3, 4 and 5 again, but with both parents exchanged. The mutation operator uses a similar procedure. When a chromosome is selected for mutation, a number of elements in the group part are deleted. When a group is deleted all nodes in the object part having that color will be temporarily uncolored. After the deletion of groups we will reinsert the uncolored nodes using the same heuristic that is used in the crossover operator. The third operator is inversion, which only operates on the group part of the chromosome, without affecting the object part. We randomly mark two points in the group part of the chromosome and then reverse the order of the groups between these two points. An example can be found in Figure 3. Figure 3: Example of inversion in grouping representation. The crossover and mutation operators work in such a way that the length of the group part of a chromosome can vary between the minimal number of colors needed to color a graph and the number of nodes of the graph. We apply the algorithm to minimize the number of colors needed to color the graph. For the 3-colorable graph instances we investigate the minimum is known to be three. Therefore, we let the fitness of a chromosome depend on the number of 'unnecessary' colors and the amount of nodes which are colored with an 'unnecessary' color, where the three colors that color the most nodes are considered as 'necessary' and the remaining colors are defined as 'unnecessary'. For a formal definition of the fitness function let us assume that k is the minimal number of colors needed for coloring the graph. Furthermore, let a chromosome x use l colors l with k l. Now let us define g(x; y) as the function that returns the amount of nodes colored with color y in chromosome x. Without loss of generality we can assume that the colors are ordered in such a way that Now the fitness function can be defined as : l To minimize function (1) the GGA must minimize the number of colors used for coloring the graph, it receives penalty points for each extra color and for every node colored with such a color. It is easy to see that it reaches its global minimum of zero, if and only if the amount of colors used is equal to the minimum amount of colors needed. Initialization of the population is done by coloring every individual in the population using the first-fit heuristic and starting the coloring at a random node. The algorithm contains a pseudo-sorting procedure using 2-tournament selection. A pseudo-sorted list of individuals is created by repeatedly selecting two different individuals uniform randomly from the population to compete with each other. The loser, i.e. the one with the worse fitness value, is removed from the population and inserted at the top of the list. The procedure stops when the population becomes a singleton and the last remaining individual is added to the list. The individuals that were inserted first have 'sank' to the bottom of the list and are seen as the worst individuals, while those inserted last are the best ones. Note that the algorithm replaces the offspring of the best N c individuals with the worst N c individuals, therefore the population size must be at least 2 \Delta N c . As for the parameter setting we compared the values recommended by Falkenauer in (Falkenauer, 1994) (a population of fifty individuals, N using an allele mutation probability, giving the probability that an allele from the group part is being deleted when the individual is undergoing mutation, of 10%) with other values. The best option turned out to be using more mutation and less crossover, in particular N During the present investigation these values will be used. 5 Standard Genetic Algorithms We experimented with several non-grouping genetic algorithms, varying different components. The common features of the GAs used are summarized in Table 1. In the integer representation each gene of an individual represents a node in the graph and its value can be one of the three colors. Thus, the chromosome length L equals the number of nodes n in the given graph. The fitness function is based on penalizing constraint violation. We consider each edge as a constraint and in case of m edges the penalty function f is type steady state selection 2-tournament deletion strategy worst-deletion crossover rate 1.0 mutation rate 1/chrom.length stop-criterion T max fitness evaluations Table 1: GA setup used in the experiments, except for the GGA where w i is the penalty, or weight, assigned to the i-th constraint (edge e i ) and It is obvious that when the minimum penalty of zero is reached, no constraints are violated and a solution is found. Here we use the same penalty w for each constraint (edge), thus f simply counts the violated constraints. It has been conjectured by Falkenauer that this representation and the corresponding genetic operators are not well-suited for grouping problems, such as graph coloring, mainly because of the 'blind disruption' of chromosomes, (Falkenauer, 1994). Our experiments confirmed this conjecture, but one unexpected result deserves special attention. Namely, increasing the disruptiveness of crossover operators increased GA performance, seemingly contradicting that crossover is harmfull. Detailed presentation of all results would consume too much space, a full overview can be found in (Eiben and Van der Hauw, 1996). Here we only give an illustration in Figure 4 that illustrates that increasing the number of crossover points in multi-point crossover, (De Jong and Spears, 1992), and increasing the number of parents in multi-parent crossovers, (Eiben et al., 1994), leads to better results. Nevertheless, the best GA variant within this representation turned out to be an asexual GA using only mutation and no crossover in a (1+1) scheme, thus with population size 1 and preservative selection. Figure 5 shows a comparison of the best asexual (mutation only) and sexual (using crossover and mutation) variants. In order-based GAs the individuals are permutations and special operators are used to recombine and mutate permutations, (Starkweather et al., 1991; Fox and McMahon, 1991). For graph coloring we define chromosomes as permutations of nodes and apply a decoder that constructs a coloring from a permutation. So just as for integer representation the chromosome length is As a decoder we have used the Greedy Algorithm which colors a node with the lowest color that does not violate constraints and leaves nodes uncolored when no colors are left. The simplest way of evaluating a permutation is then to use the number of uncolored nodes in the coloring belonging to it. Formally, we use a penalty function that concentrates on the nodes, instead of the edges. The function f is now defined as: AES Number of parents (diagonal) or crossover points (m-point) diagonal m-point 1-point120000160000200000240000 AES Number of parents scanning uniform Figure 4: Integer representation: effect of more crossover points and more parents on AES for G eq;n=1000;p=0:025;s=5 . T in each run, the results are averaged over 25 independent runs for each setting (nr. of crossover points, nr. of parents).0.20.61 Edge connectivity sexual AES Edge connectivity sexual asexual Figure 5: Integer representation: asexual (only mutation) vs. sexual reproduction (uniform incest prevention) for near the phase transition. T in each run, the results are averaged over 25 independent runs on each instance. where w i is now the penalty (or weight) assigned to node i and 1 if node x i is left uncolored because of a constraint violation Just like before, we use w but while for integer representation the fitness function counted the 'wrong' edges, here we count the uncolored nodes. Note that the search space for order-based representation is much bigger than for the integer representation: n! instead of 3 n . Because at most 3 n different possible colorings exist this coding is highly redundant. Additionally, the penalty function given in Equation 3 supplies less information than the one given in Equation 2. These considerations might suggest that this GA will be less successful, but the experiments show that the opposite is true. Using this representation we compared different operators, various population sizes and the effect of using only mutation. The outcomes are similar to those of integer representation, in the sense that an asexual GA using only SWAP mutation and no crossover in a (1+1) scheme outperforms the best sexual GA using OX2 crossover together with SWAP on large graphs, see Figure 6 for illustration. On small graphs this is only partly true, but the global conclusion is that the asexual GA has the best overall performance and that order based representation is superior to integer representation.0.20.61 Edge connectivity sexual AES Edge connectivity sexual asexual Figure Order-based representation: asexual (only SWAP mutation) vs. sexual reproduction near the phase transition. T in each run, the results are averaged over 25 independent runs on each instance. Summarizing our findings on using GAs with traditional representations we can note the following. The best performance is achieved with an algorithm without crossover, exclusively using mutation and having population size 1. The question arises whether such an algorithm still can be seen as genetic. A 'no' is supported by noting that the presence of crossover and a population with more than one elements is crucial for GAs. However, one could also argue that the representation and the mutation is standard, only the parameter setting is extreme. To avoid conflicts with conventions on terminology we percieve and name the winning algorithm variant as an evolutionary algorithm. To this end note, that in contemporary evolutionary computation the term evolutionary algorithm comprises among others genetic algorithms, evolution strategies (ES), and evolutionary programming (EP) (Back et al., 1997), and that (1+1) style alorithms are common in ES (Schwefel, 1995) and EP always operates without using crossover (Fogel, 1995). 6 Stepwise Adaptation of Weights Evolutionary algorithms are intrinsically dynamic and adaptive processes. It is thus rather unnatural to use static control parameters that remain constant during the evolution. Using constant parameters is, however, the traditional practice. In the meanwhile, there is an increasing number of papers that consider varying parameter settings. This accumulates knowledge on a research area that is thus becoming an emerging sub-technology within evolutionary computation. Looking at related work on varying parameters in EAs one can distinguish common features several approaches share. The classification in the paper by (Hinterding et al., 1997) distinguishes three different streams. Dynamic parameters obtain different values along the evolution prescribed by a user defined schedule. This schedule typically assigns new values to parameters depending on time, most commonly expressed by the number of generations or fitness evaluations. Adaptive parameters obtain new values by a feedback mechanism that monitors the evolution. The new values typically depend on the achieved progress. This progress measure is the input of a mechanism that resets the parameter. Self-adaptive parameters are encoded in the chromosomes and undergo evolution themselves. As opposed to the previous two techniques, here there is not even an indirect user control of this parameter. It is clear that varying parameters suits the general 'evolutionary spirit' better than static ones, furthermore, they have technical advantages. First of all, they often lead to increased performance. Second, adaptive and self-adaptive parameters free the user from determining these parameters by having the EA doing it. This reduces the chance for incorrect parameter settings. Our approach falls in the adaptive category in the above classification scheme. Technically it amounts to modifying the weights of components of the penalty function, hence modifying the fitness landscape during the search, based on feedback from the actual population. This technique was introduced for constraint solving with GAs in (Eiben et al., 1995b), where the constraints were weighted. The rationale behind it is clear: satisfying a constraint with a high penalty gives a relatively high reward to the algorithm, hence it will be 'more interested' in satisfying such constraints. Thus, using appropriate weights focuses the 'attention' of the algorithm, hence it can improve the performance. Appropriate in this case means that the constraint weights should reflect how important, or rather, how difficult a specific constraint is. This causes two problems. On the one hand, to determine relative hardness of constraints can require substantial knowledge on the problem. On the other hand, 'being hard' cannot be seen independently from the applied problem solver. Therefore, it is a natural idea to leave the decision on measures of hardness to the problem solver itself. Although there are well-founded arguments (Culberson, 1996; Wolpert and Macready, 1997) stating that no search technique can be superior in general, having an evolutionary algorithm adjusting its parameters itself has been proved to be powerful under many circumstances (Angeline, 1995; Hinterding et al., 1997). In the particular case of constraint satisfaction problems we expect that a mechanism enabling the GA to learn weights itself can circumvent difficulties of setting these weights by a human user. It is interesting to note that using adaptive, or learning features to improve search performance is done for other types of algorithms too. The breakout method of Morris (Morris, 1993) is the earliest example we know of. Adaptive memory features as advocated by Glover, (Glover, 1996), and applied by Lkketangen and Glover, (Lkketangen and Glover, 1996), are close to the 'evolutionary spirit' and work very well on constrained problems. Also, variants of the the GSAT algorithm for satisfiability problems by Selman and Kautz, (Selman and Kautz, 1993) and Frank, (Frank, 1996b; Frank, 1996a), apply mechanisms that re-evaluate weights of clauses during the search process. The best EA for our graph 3-coloring problem we found so far uses order-based representation penalizing uncolored nodes. We extend this variant with an adaptive mechanism, thus the basic idea is now implemented by monitoring which variables in the best individual violate constraints and raising the penalty w i belonging to these variables. Depending on when the weights are updated we can distinguish an off-line (after the run) and an on-line (during the run) version of this technique, see Figure 7 and Figure 8. Off-line saw set initial weights (thus fitness function f) for x test runs do run the GA with this f redefine f after termination end for Figure 7: Pseudo code of the off-line weight adaptation mechanism On-line saw set initial weights (thus fitness function f) while not termination do for the next Tp fitness evaluations do let the GA go with this f end for redefine f and recalculate fitness of individuals while Figure 8: Pseudo code of the on-line weight adaptation mechanism In (Eiben et al., 1995a) and (Eiben and Ruttkay, 1996) the off-line version was applied, here we will use the on-line version modifying the weights, hence modifying the fitness function, during the evolution. After a certain period, in particular T p fitness evaluations, the best individual in the population is colored and the weights of its uncolored nodes (that are considered hard for the are increased by 4w, i.e. w i is set to w i +4w. This implies that the EA has to search on a dynamically changing fitness landscape, consequently the population has to be re-evaluated after each period 3 . We call this mechanism Stepwise Adaptation of Weights (SAW). 3 Hereby the total number of fitness evaluations will not equal the total number of generated individuals. It could be argued that the number of re-evaluations should be included in the total number of evaluations. However, we want to count the search steps by the total number of generated colorings. Besides, re-evaluation is computationally It is very interesting to see the fitness curve of a run of the EA with the SAW mechanism. Figure shows a run when a solution is found. The left curve has a higher resolution, displaying the fitness of the best individual between 0 - 10000 evaluations, the right curve shows the range 80000. The higher resolution curve (left) shows that within each period the penalty drops as the EA is making progress and then sharply rises when the weights are updated, giving the image of a SAW! Note that the SAW mechanism introduces two new parameters, T p and 4w. It is thus important to check whether the performance of a SAW-ing EA is sensitive for the parameter values. To this end we performed an extensive test series that showed that the performance is pretty much independent from these parameters. As an illustration we present the outcomes for for the asexual as well as for the sexual order-based EA in Figure 9 and 10. The exact values for 4w and T p do not have a significant effect on the performance, as long as T p is significantly smaller than T max . From now on, we will use reason. Figure 9 and 10 also show that, also in the presence of the SAW mechanism the sexual EA using OX + SWAP is clearly inferior to the asexual variant.0.20.61 OX2+SWAP100000200000300000 AES Figure 9: Effect of varying 4w on G eq;n=1000;p=0:010;s=5 . T uses population size 1, OX2+SWAP uses population size 700. The effect of the SAW mechanism on the EA performance on graphs with can be seen in Table 2. The increase in performance after adding the SAW-ing mechanism is dramatic: the success rate averaged over all seeds raises from 9% to 92%, while the number of search steps drops from 205643 to 89277. The figures show not only that a SAW-ing EA highly outperforms the other techniques, but also that the performance is rather independent from the random seeds. Thus, the SAW mechanism is not only highly effective, obtaining much better success rates at lower costs, but also very robust. Somewhat surprising is the low performance of the GGA. On average it terminates with 5 colors and approximately 50 nodes in the 'unnecessary' colors. cheap: the individual needn't be decoded again, only the sum in Equation 3 has to be re-computed for its uncolored nodes. Therefore, the number of re-evaluations is not included in the total cost. 2000 4000 6000 8000 10000 OX2+SWAP100000200000300000 AES Figure 10: Effect of varying T p on G eq;n=1000;p=0:010;s=5 . uses population size 1, OX2+SWAP uses population size 700. 7 Comparing the GGA, the SAW-ing EA and DSatur The results summarized in Table 2 serve as an indication of the viability of our SAW-ing mechanism. For a solid conclusion on the performance of the SAW-ing EA we perform an extensive comparison between the Grouping GA as described in Section 4, DSatur with backtracking, and a SAW-ing EA with order-based representation, simple greedy decoder, OneSWAP mutation 4 and no crossover in a (1+1) selection scheme using T for the SAW mechanism. The comparison is performed on arbitrary 3-colorable, equi-partite 3-colorable and flat 3-colorable graphs for connectivities around the phase transition. The results are based on 100, 50 and 25 runs for For all instances, the graph is generated with seed 5. DSatur and both GAs are allowed search steps. The test results concerning SR and AES values are given in the Figures 11 - 13. The Grouping GA is inferior to the other two algorithms on all graph instances. On small graphs DSatur is better for all p's than the SAW-ing EA, except for flat graphs near the phase transition, see Figure 13 for On medium size graphs (n = 500) the SAW-ing EA is slightly better on arbitrary and equi-partite topologies. On flat graphs we see that the performance of the SAW-ing EA deteriorates much less at the phase transition than that of DSatur. On large graphs the SAW-ing EA is clearly better w.r.t. success rates, because it is often able to find solutions where DSatur does not find any. The AES curves are sometimes crossing, but in general the SAW-ing EA needs fewer steps. The reason for these differences could be that the instances with are small enough for DSatur to get out of local optima and find solutions, while this is not possible anymore for where the search space becomes too big. Evaluating the overall usefulness of the SAW-ing EA for graph coloring one has to consider two cases. On Additional tests (not reported here) showed that OneSWAP which always swaps exactly one pair of genes is slightly better for the (1+1) GA than usual SWAP. the easy problems, i.e. small graphs and far from the phase transition DSatur is better. On the hard instances near the phase transition and on large graphs the SAW-ing EA outperforms DSatur. These results are very good in the light of the fact that DSatur is a highly problem tailored graph coloring algorithm, whereas the SAW-ing EA is a general purpose algorithm for constraint satisfaction, using no specific domain knowledge on graph coloring. Another strong property of the SAW-ing EA is that it can take more advantage of extra time given for search. We increased the total number of evaluations to 1000000 and observed that DSatur still had SR=0.00 on G eq;n=1000;p=0:008;s=5 . The performance of the SAW-ing EA, however, increased from SR=0.00 to SR=0.44 (AES=407283), showing that it is able to benefit from more time given, where the extra time for backtracking is not enough to get out of the local optima. It is an important question how the performance of an algorithm changes if the problem size grows. This is the issue of scalability, which we will consider w.r.t. to the success rates and the computational complexity. Thus, while the figures 11 - 13 were drawn for fixed n's and varying p's, in the sequel we vary n (and keep related to it). We tested the three algorithms using the same parameters as before for 8:0=n at the phase transition. The results are given in Figure 14 showing that the Grouping GA could only solve the smallest problem instances and DSatur was not able to find solutions on larger problem instances. Therefore, we also made a comparison on easier instances for Figure 15 exhibits these results showing that the SAW-ing EA scales up much better than the other two methods. As discussed in Section 3 even the implementation independent number of search steps is not ideal for comparing different algorithms. Nevertheless, comparing the growth rate of the number of search steps when the problem size (the number of nodes in the graphs to be colored) grows gives a sound basis for judgment, even though the absolute meaning of the given numbers may differ. The scale-up curves for AES on the right hand side plots of Figure 14 (p = 8:0=n) and Figure show interesting results. On the easier case, for DSatur is faster up to but the SAW-ing EA scales up much better. On the really hard graphs, 8:0=n, the SAW-ing EA outperforms DSatur already from The curves for both edge connectivity values indicate that the SAW-ing EA scales up linearly with the problem size. Edge connectivity (n=200) DSatur GGA 0100000200000300000 AES Edge connectivity (n=200) DSatur Edge connectivity (n=500) DSatur GGA 0100000200000300000 AES Edge connectivity (n=500) DSatur Edge connectivity (n=1000) DSatur GGA 0100000200000300000 AES Edge connectivity (n=1000) DSatur GGA Figure 11: Comparison for arbitrary 3-colorable graphs, for Edge connectivity (n=200) DSatur GGA 0100000200000300000 AES Edge connectivity (n=200) DSatur Edge connectivity (n=500) DSatur GGA 0100000200000300000 AES Edge connectivity (n=500) DSatur Edge connectivity (n=1000) DSatur GGA 0100000200000300000 AES Edge connectivity (n=1000) DSatur GGA Figure 12: Comparison for equi-partite 3-colorable graphs, for Edge connectivity (n=200) DSatur GGA 0100000200000300000 AES Edge connectivity (n=200) DSatur Edge connectivity (n=500) DSatur GGA 0100000200000300000 AES Edge connectivity (n=500) DSatur Edge connectivity (n=1000) DSatur GGA 0100000200000300000 AES Edge connectivity (n=1000) DSatur GGA Figure 13: Comparison for flat 3-colorable graphs, for Number of nodes DSatur GGA2000006000001000000 50 250 500 750 1000 1250 1500 AES Number of nodes DSatur GGA Figure 14: Scale-up curves for SR and AES on equi-partite graphs with based on runs. Conclusions NP-complete problems, such as graph coloring, form a big challenge for designers of algorithms in general. For evolutionary algorithms, in particular, constraint satisfaction problems form a specific challenge. In this paper we investigated how EAs can be applied for graph 3-coloring. After trying several traditional genetic algorithm variants based on integer and order-based representation we concluded that using only mutation is better than using mutation and crossover. This fact is in accordance with the common opinion in evolutionary computation that traditional0.20.61 50 250 500 750 1000 1250 1500 Number of nodes DSatur AES Number of nodes DSatur GGA Figure 15: Scale-up curves for SR and AES on equi-partite graphs with based on runs. genetic crossovers are not appropriate for so-called grouping problems, because they 'blindly' disrupt chromosomes. However, it is interesting to note that in integer representation using more crossover points and more parents -which both imply heavier mixing than usual operators- increases the performance, cf. Section 5. As discussed at the end of Section 5, it can be argued that swithching off crossover and setting the population size to 1 in a genetic algorithm results in an algorithm that cannot be called anymore. Iterated stochastic hill-climbing could be an appropriate name, but we rather use the name evolutionary algorithm. This name reflects the origin of the method and there are other algorithms in evolutionary computation that use exclusively mutation, or population size 1 (Fogel, 1995; Schwefel, 1995). Following the recommendations that on grouping problems special representations and crossovers are advisable (Falkenauer, 1994), we have implemented and tested the grouping GA. Surprisingly, the GGA is inferior to a simple order-based asexual EA, see Table 2. This, suggests that using specific grouping operators is not the only solution for handling the graph coloring problem. Using only mutation and population size 1 works as well. The main research subject of this investigation is the technique called Stepwise Adaptation of Weights. This mechanism is problem independent and it amounts to repeatedly redefining the weights that are used in the definition of the fitness (penalty) function. Similar mechanisms have been applied in other search algorithms; here we use it in an EA. Adding the SAW mechanism to the simple order-based asexual EA increased its performance with a factor 10 (success rates increased from 9% to 92%), and made it the best algorithm on the problem instances we used for a preliminary comparison, cf. Table 2. SR AES SR AES SR AES SR AES SR AES DSatur Table 2: Comparing DSatur with backtracking, the Grouping GA, (1+1) order-based GA using SWAP and (1+1) order-based GA using SWAP and the SAW mechanism (T different seeds. Conducting an extensive series of experiments on three different types of graph, three different sizes and various edge connectivity values confirmed that the SAW-ing EA is a powerful graph coloring algorithm. The SAW-ing EA was better than the GGA and also outperformed DSatur by three criteria: 1) on the hardest graph instances it performs better than DSatur, 2) it is able to increase its performance when given more time, whereas DSatur is not, scales up much better, indicating a linear scale-up w.r.t. computational complexity. Especially nice about the SAW-ing EA is that it is not tailored to the problem of graph coloring. Our findings are thus relevant in the broader context of evolutionary constraint handling. The SAW mechanism helps to circumvent a major problem of penalty based evolutionary constraint handling techniques by letting the GA find the right constraint weights. Further research concerns application of SAW-ing to other (NP-complete) constrained problems and investigating variations of the SAW mechanism. Straightforward modifications are varying the value of 4w over the constraints, and allowing decreasing w i for constraints that are already satisfied. Especially interesting to study is the learning procedure based on combining frequency and recency based memory of Lkketangen and Glover, where the changes in weights are related to solution quality, (Lkketangen and Glover, 1996). Finally, let us make an additional note on the constraint weights the EA finds. The plots in Figure suggest that problem solving with the SAW mechanism happens in two phases. In the first phase the EA is learning a good setting for the weights. In this phase the penalty increases a lot because of the increased weights. In the second phase the EA is solving the problem, exploiting the knowledge (appropriate weights) learned in the first phase. In this phase the penalty drops sharply indicating that using the right weights (appropriate fitness function) in the second phase the problem becomes 'easy'. This interpretation of the fitness curves is plausible. We, however, do not want to suggest that the EA could learn universally good weights for a given graph instance. In the first place, another problem solver might need other weights to solve the same problem. Besides, in a control experiment we applied the SAW-ing EA to a graph, thus learning good weights, and then applied the EA to the same graph again using the learned weights non- adaptively, i.e. keeping them constant along the evolution. The results showed worse performance than in the first run when adaptive weights were used. Suggesting that the SAW mechanism works by enabling the problem solver to discover some hidden, universally good weights is, therefore, wrong. This seems to contradict the interpretation that distinguishes two phases of search. At the moment we tend to see the main advantage of SAW in allowing the EA to shift the focus of search (quasi) continuously and thus allowing implicit problem decomposition that guides the population through the search space.2006001000140018000 2000 4000 6000 8000 10000 Fitness Evaluations5001500250035000 20000 40000 60000 80000 Fitness Evaluations Figure Fitness curve for the SAW mechanism for SWAP, --R Adaptive and self-adaptive evolutionary computation Proof verification and hardness of approximation problems. Handbook of Evolutionary Computation. Institute of Physics Publishing Ltd An O(n 0:4 New methods to color vertices of a graph. Where the really hard problems are. Problem structure heuristics and scaling behavior for genetic algorithms. A discussion on some design principles for efficient crossover operators for graph coloring problems. On the futility of blind search. Exploring the k-colorable landscape with iterated greedy A formal analysis of the role of multi-point crossover in genetic algorithms Constraint satisfaction problems. Graph coloring with adaptive evolutionary algorithms. A new representation and operators for genetic algorithms applied to grouping problems. A hybrid grouping genetic algorithm for bin packing. A genetic algorithm for bin packing and line balancing. Evolutionary Computation. Genetic operators for sequencing problems. Computers and Intractability: A Guide to the Theory of NP-Completeness Tabu search and adaptive memory programming - advances On colouring random graphs. Adaptation in Evolutionary Computation: a survey. Optimization by simulated annealing: An experimental evaluation Operations Research Algorithms: Their Complexity and Efficiency. Intelligent structural operators for the k-way graph partitioning problem In (Belew and Booker Surrogate constraint methods with simple learning for satisfiability problems. The breakout method for escaping from local minima. Evolution and Optimum Seeking. A comparison of genetic sequenceing operators. Almost all k-colorable graphs are easy to color --TR --CTR Steven Prestwich, Coloration Neighbourhood Search With Forward Checking, Annals of Mathematics and Artificial Intelligence, v.34 n.4, p.327-340, April 2002 Rhyd Lewis, Metaheuristics can solve sudoku puzzles, Journal of Heuristics, v.13 n.4, p.387-401, August 2007 Mrk Jelasity , Pilar Martnez Ortigosa , Inmaculada Garca, UEGO, an Abstract Clustering Technique for Multimodal Global Optimization, Journal of Heuristics, v.7 n.3, p.215-233, May 2001 Domingo Ortiz-Boyer , Csar Hervs-Martnez , Nicols Garca-Pedrajas, Improving crossover operator for real-coded genetic algorithms using virtual parents, Journal of Heuristics, v.13 n.3, p.265-314, June 2007 Peter Ross , Emma Hart , Dave Corne, Genetic algorithms and timetabling, Advances in evolutionary computing: theory and applications, Springer-Verlag New York, Inc., New York, NY,
adaptive parameters;genetic algorithms;constraint satisfaction;penalty functions;evolutionary algorithms;graph coloring;grouping problem
594977
Tabu Search for Frequency Assignment in Mobile Radio Networks.
The main goal of the Frequency Assignment Problem in mobile radio networks consists of assigning a limited number of frequencies to each radio cell in a cellular network while minimizing electromagnetic interference due to the reuse of frequencies. This problem, known to be NP-hard, is of great importance in practice since better solutions will allow a telecommunications operator to manage larger cellular networks. This paper presents a new Tabu Search algorithm for this application. The algorithm is tested on realistic and large problem instances and compared with other methods based on simulated annealing, constraint programming and graph coloring algorithms. Empirical evidence shows that the Tabu algorithm is very competitive by giving the best solutions to the tested instances.
Introduction The Frequency Assignment Problem (FAP) is one of the key applications in mobile radio networks engineering 1 . Although different versions of the FAP can be defined, the main purpose is to assign a limited number of available frequencies to each cell in a mobile radio network while minimizing electro-magnetic interference due to the re-use of frequencies. The difficulty of this application comes from the fact that an acceptable solution of the FAP must satisfy a set of multiple constraints, which impose conflicting objectives. The most severe constraint concerns a very limited radio spectrum consisting of a small number of frequencies (or channels). Telecommunications operators such as France Telecom must cope with a maximum of 60 frequencies for their networks, whatever the traffic volume Work supported by the CNET (French National Research Center for Telecommunications) under the grant No.961B134. 1 The frequency assignment problem studied in this paper applies essentially to widely used cellular networks and does not necessarily apply to other wireless technologies, such as the CDMA. to be covered, and this, in agreement with national and international regulations. This constraint imposes a high degree of frequency re-use, which in turn increases the probability of frequency interference. Indeed, in addition to this constraint, there are frequency interference constraints which state that frequencies assigned to some cells must satisfy a given separation distance in the frequency domain. The basic FAP can be shown to be NP-hard in its simplest form because it is reduced to the graph coloring problem [13]. More generally, the problem is equivalent to the so-called "set T-coloring problem" [20]. Therefore, it is very unlikely to find any efficient algorithm for this problem. So far, many heuristic methods have been proposed to tackle the FAP in cellular networks, including graph coloring algorithms (GCA) [11], constraint programming (CP) [1], simulated annealing (SA) [8], artificial neural networks [18, 10], evolutionary algorithms (EAs) [14] and Tabu Search (TS) [15]. Similar techniques have been developed and experimented on a closely related problem called "radio link frequency assignment" in the setting of military application [6, 17, 3, 4]. In this paper, we present a new and highly effective TS algorithm tailored for the frequency assignment problem. This algorithm distinguishes itself from previous algorithms by some important features such as incremental evaluation of solutions, dynamic tabu tenure, candidate list strategy, and co-cell constraints handling. Experiments carried out on realistic and large problem instances (up to 300 cells, 30,000 constraints with show that this TS algorithm is very competitive and gives us the best results for the tested instances compared with other methods based on graph coloring algorithms, constraint programming, and simulated annealing. The paper is organized as follows. In Section 2, the FAP is modeled as an optimization problem. In Section 3, TS is briefly reviewed. In Section 4, the components of the TS algorithm are described. In Section 5, after a review of the test data, experimental results are presented and compared. Conclusions are given in the last section. 2 The Frequency Assignment Problem 2.1 Constraints in the FAP A cellular network is defined by a set fC of N cells, each cell C i requiring T i frequencies. The number T i , called the traffic of C i , is determined by an estimation of the maximum number of communications which can simultaneously arise within a cell. The basic FAP consists of assigning to each cell C i of the network T i frequencies taken from a set of available frequencies while respecting a set of frequency interference constraints. Since the number of available frequencies is very limited and usually much smaller than the sum of the total traffics of the network, frequencies must be re-used by different cells in an assignment. It is this frequency re-using that may lead to frequency interference. Interference occurs when two frequencies assigned to a same cell or two adjacent cells 2 are not sufficiently separated. Therefore, the frequency interference constraints over a network are divided into co-cell constraints and adjacent-cell constraints. ffl co-cell constraint: any pair of frequencies assigned to a radio cell must have a certain distance between them in the frequency domain. Two cells are adjacent if they emit within a common area even if they are not geographically adjacent. ffl adjacent-cell constraint: any pair of frequencies assigned to two adjacent cells must be sufficiently separated in the frequency domain. These interference constraints are conveniently represented by a symmetric compatibility matrix M [N; N ] where N is the number of cells in the network and each element of M is a non negative integer. Let f i;k denote the value of the kth frequency (k 2 f1::T i g) of denote the set of NF available frequency values, then the interference constraints are formulated as follows: is the minimum frequency separation necessary to satisfy the co-cell constraints for the cell C i . 8m, represents the minimum frequency separation required to satisfy the adjacent-cell constraints between two cells C i and C j . M [i; means there is no constraint between the cells C i and C j . 2.2 Optimization in FAP Given a network defined by N cells with T i traffic for each cell C i and the compatibility matrix M , two basic optimization (minimization) problems can be defined. The first problem consists of minimizing frequency interference with a fixed number of frequencies. In this case, the optimization problem is defined by a couple (S,f) where the search space S is defined by the set of all possible frequency assignments and the cost function f to be minimized is defined as follows. For any s in S, f(s) represents frequency interference, measured by the number of unsatisfied co-cell and adjacent-cell constraints. More formally, one has: ae ae The second problem consists of minimizing the number of frequencies required in an assignment 3 . This problem can be dealt with as a series of the first problems as follows. To minimize the number of frequencies NF used in an assignment, one fixes NF to a sufficiently high value 4 and seeks an interference-free assignment s the above mentioned search space S and the cost function f. If such an interference-free assignment is found, one proceeds with NF-1 frequencies and so on. This process continues until one can no longer find an interference-free assignment. This minimization process is similar to that used by Hertz and de Werra in their TS algorithm for graph coloring [16]. Let us notice that minimizing the number of frequencies used is of great importance in practice since free frequencies can easily be used for extensions to an existing network. 3 More generally, the interference-free requirement may be relaxed to allow any interference threshold. 4 This initial value can be determined with a greedy method. 3 A Brief Review of Tabu Search This section gives a brief review of Tabu Search, emphasizing the most important features which have been implemented in our TS algorithm. For a complete presentation of TS, the reader is invited to consult the recent book by Glover and Laguna [12]. Tabu Search is a meta-heuristic designed for tackling hard combinatorial optimization problems. Contrary to randomizing approaches such as SA where randomness is extensively used, TS is based on the belief that intelligent searching should embrace more systematic forms of guidance such as memorizing and learning. can be described as a form of neighborhood search with a set of critical and complementary components. For a given instance of an optimization problem characterized by a search space S and a cost function f, a function N: S ! 2 S is first introduced to determine a neighborhood. A typical TS algorithm begins then with an initial configuration s in S and then proceeds iteratively to visit a series of locally best configurations following the neighborhood function. At each iteration, a best neighbor s' 2 N(s) is sought to replace the current configuration even if s' does not improve the current configuration in terms of the cost function. To avoid the problem of possible cycling and to allow the search to go beyond local optima, TS introduces the notion of Tabu list, one of the most important components of the method. A tabu list is a special short term memory that maintains a selective history H, composed of previously encountered solutions or more generally pertinent attributes of such solutions. A simple TS strategy based on this short term memory H consists in preventing solutions of H from being reconsidered for next k iterations (k, called tabu tenure, is problem dependent). Now, at each iteration, TS searches for a best neighbor from this dynamically modified neighborhood N(H,s), instead of N(s) itself. Such a strategy prevents Tabu from being trapped in short term cycling and allows the search process to go beyond local optima. When attributes of solutions instead of solutions are recorded in tabu list, some non- visited, yet interesting solutions may be prevented from being considered. Aspiration criteria may be used to overcome this problem. One widely used aspiration criterion consists of removing a tabu classification from a move when the move leads to a solution better than the best obtained so far. uses an aggressive search strategy to exploit its neighborhood. Therefore, it is crucial to have special data structures and techniques which allow a fast updating of move evaluations, and reduce the effort of finding best moves. Moreover, candidate list strategies may be used to limit the neighbors to be considered at each iteration to a subset of the neighborhood. A typical strategy consists of identifying subsets of influential moves, such as those considered promising to lead to improved solutions. A good candidate list strategy combined with an efficient technique for move evaluations is essential for high solution speed and good quality. There are other techniques available in TS such as intensification and diversification. In this paper, we show that a TS algorithm based on the above mentioned elements may be very effective and robust. The following notations will be used in the presentation: N the number of cells of a network, NF the number of available frequencies, f1.NFg the set of NF frequency values, C i and T i a cell and its traffic, f i;k (f i;k 2 f1.NFg) the value of the kth frequency of C i . 4.1 Components Configuration, search space and cost function Given a FAP instance involving N cells, traffics T i (i 2 f1.Ng), a set of frequency values f1.NFg, and a compatibility matrix M [N; N ], a configuration s corresponds to a complete frequency assignment: In other words, a configuration is a frequency assignment satisfying the co-cell constraints of the given instance. The search space S of the instance is therefore composed of all such configurations. According to equation (1) (Section 2), for each solution s 2 S, f(s) corresponds to the total number of unsatisfied interference constraints. Since co-cell constraints are taken into account by the configuration s, f(s) is now simplified to be the number of unsatisfied adjacent-cell constraints. Note that for an interference-free assignment s, An alternative approach consists of not imposing the satisfaction of co-cell constraints. However, as explained later, directly integrating co-cell constraints into configurations is an important factor in improving the search efficiency for this application. Neighborhood and candidate list Given S, the neighborhood function N defined as follows: s and s' 2 S are neighbors if they are different at the value of a single frequency of a cell. More formally, let s(f i;k ) denote the value of the kth frequency of the cell C i in s, then s' 2 N(s) if and only if the following condition is verified: such that s(f i;m ) 6= s'(f i;m ) and Therefore, a neighbor of s can be obtained by changing the value f i;m of the mth frequency of a cell C i in s in such a way that the new value always satisfies the co-cell constraint. A move is thus characterized by a triplet ! being respectively a cell, a frequency of the cell and a frequency value. The members of N(s) are not equally interesting. Indeed, it will be less pertinent to change a frequency value which does not violate any interference constraint. This observation leads us to define the following candidate list strategy: are different at the value of a conflicting frequency 5 g 5 A frequency of a cell is said to be conflicting if its value violates some interference constraints. Clearly, the size of V varies during the search according to the number of conflicting frequencies and in general is much smaller than jN(s)j. This candidate list strategy helps the search to concentrate on influential moves and to avoid irrelevant ones. Moreover, it significantly reduces the number of neighbors to be considered at each iteration. Incremental evaluation and neighborhood examination For any solution s, its cost f(s) represents frequency interference, measured by the number of unsatisfied interference constraints. Since the number of neighbors to consider may be very high at each iteration, we adopt a fast evaluation method inspired by a technique proposed by Fleurent and Ferland [9]. Let us note . The main idea consists of maintaining incrementally in a NF W matrix ffi the move value or cost variation for each possible move of the current solution s. Each time a move is carried out, the elements of the matrix affected by the move are updated accordingly. Initializing After a move, the matrix can be updated in time O(NF W ) in the worst case. A best neighbor in V can be searched in time O(jV j). Given that jV j is smaller then NF W , each iteration takes thus time O(NF W ) in the worst case. Tabu list and tabu tenure When the value of a frequency of a cell C i in a solution s is changed to a new value, the classified tabu for k (tabu tenure) iterations. That is, the old value will not be allowed to be re-assigned to C i during this period. To implement the tabu list, we use an N NF matrix T. Each time ! cell; old value ? is added to the tabu list, the corresponding element of T is set to the current iteration number plus tabu tenure k. At any moment, it is easy to verify if a given move is tabu or not by simply comparing the current iteration number with that recorded in the tabu matrix T. The tabu tenure k is dynamically adjusted by a function defined over the size of the candidate list V . More precisely, ff is a value from (0,1] and k is bounded by two constants Min and Max. Since jV j varies during the search, so does k for any fixed ff. ff typically takes values from 0.1 to 0.5. To avoid too large values of k due to the presence of many conflicting cells in a configuration, a constant Max is used to limit k. Similarly, a lower bound Min is used to prevent k from becoming too small at late stages of the search. Min and Max are typically fixed at 10 NF and NF. Aspiration criteria A very simple aspiration criterion is used: the tabu status of a move is cancelled if the move leads to a solution better than the best solution s* encountered so far. 4.2 Tabu Algorithm Now that the components of the TS algorithm have been presented, we give below the general algorithm. Tabu algorithm begin initialize the iteration counter */ generate an initial solution */ record the best configuration found so far */ while Iter ! MAX do choose a best neighbor such that NB ? T [i; new v] or f(s') ! f update update tabu tenure k Figure 1. Tabu Search Algorithm for the Frequency Assignment Problem This algorithm can directly be used to deal with the first optimization version of the frequency assignment problem defined in Section 2, i.e. to minimize frequency interference with a fixed number of NF frequencies. The algorithm stops and returns the best assignment found after MAX iterations. This algorithm can also be used to minimize the number of frequencies NF required in an interference-free assignment as explained in Section 2.2: 1. apply the above TS algorithm to search an interference-free assignment s with NF frequencies; 2. decrement NF, i.e. NF / NF-1, and go to step 1 if such an assignment is found; 3. stop and return NF+1 if no interference-free solution is found within MAX iterations. Each time the TS algorithm is called with a decremented number of frequencies, an initial configuration using NF-1 frequencies must be generated. There are two possibilities to do this. First, one can generate a completely new configuration from scratch. Second, one can re-use the last assignment obtained with NF frequencies as follows: for each frequency having a value greater than the decremented NF, it is given a new value randomly taken from 1 to the decremented NF 6 . Note finally that this optimization process can be used to minimize the number of frequencies NF required by an assignment with any admissible interference threshold. 4.3 Comments on the TS Algorithm The current TS algorithm distinguishes itself from a previous TS algorithm [15] by some important features. The first feature concerns the evaluation method. The ffi matrix 6 Given the definition of a configuration, this new frequency value must verify co-cell constraints. introduced above allows a fast searching of best moves and an efficient updating of move values. Conversely, no data structure is used in [15] to record move values. The evaluation of neighbors is thus expensive and very time-consuming in the previous TS algorithm. The second point concerns the candidate list strategy. It is well-known that the number and nature of the neighbors visited at each iteration have a great influence on the performance of a TS algorithm. The conflict-based strategy is an informative way to isolate influential moves and is more efficient than the entire neighborhood. Experimental data are presented in Section 5.4. to support this point. The third point concerns the way co-cell constraints are handled. In the initial problem formulation, both co-cell and adjacent-cell constraints are used to define the cost function (see Section 2). Therefore, both of them can be treated exactly in the same way to evaluate the quality of a configuration. This is the approach taken in [15]. However, a close look at these two types of constraints makes it clear that they are in fact different. In some sense, co-cell constraints are harder to solve than adjacent-cell constraints since co-cell constraints, defined on each pair of frequencies of a same cell, form cliques. As shown in [7], it is more interesting to remove these hard constraints from the cost function and solve them first. This is realized by imposing the satisfaction of co-cell constraints by all configurations (feasibility with respect to these constraints). This technique, called co-cell handling (CCH) in [7], allows to reduce the search space and to simplify the cost function. Experimental data are presented in Section 5.4. to support this point. Finally, the dynamic tabu tenure is another factor contributing to the efficiency of the algorithm. Indeed, it seems more natural to allow the tabu tenure to adjust itself during the search than to impose a fixed value. 5 Experimentation and Results In this section, we present experimental results of the TS algorithm on two sets of real size FAP instances provided by the French National Research Center for Telecommunications. These instances are produced by a generator in such a way that they reflect various situations encountered in real networks or sub-networks of France Telecom. The main goal here consists of finding interference-free assignments with a minimal number of frequencies. Two sets of experiments were carried out. The first one aims to evaluate the performance of the Tabu algorithm by comparing it with a baseline algorithm. The purpose of the second one is to compare the results of the Tabu algorithm with the best known results on the same instances obtained with other methods. 5.1 Tests Test The first set of FAP instances has the following characteristics. i.e. each cell is assigned one frequency. Consequently, a co-cell constraint does not exist. ffl adjacent constraint: jf for two adjacent cells must be assigned different frequency values. It is easy to see that these instances correspond to the classic graph coloring problem. Finding an optimal frequency assignment is equivalent to coloring a graph using only a minimum number (the chromatic number) of colors. For this test set, the chromatic noted Opt for each instance, is known in advance. These instances are labeled as NF.N.d representing respectively the optimal number of frequencies needed for an interference-free assignment, the number of cells in the network and the density of interference constraints over the network. For example, 8.150.30 defines a network composed of 150 cells with 8 the number of frequencies required in an interference-free assignment and 30% of a total of 150*(150-1)/2 interference constraints. Test The second set of FAP instances has the following characteristics. i.e. each cell is assigned two frequencies. i.e. two frequencies assigned to the same cell must have a minimum distance of 3. ffl adjacent constraint: jf are adjacent cells; i.e. two frequencies assigned to two adjacent cells must have a minimum distance of 1 or 2 according to the cells. The instances of this set are labeled as q.N.d.p. The first parameter has no special meaning. The second and third parameters have similar meanings to those of the first set. The fourth parameter allows distinguishing between different instances of the same class. For this set of instances, the exact optimal number of frequencies for an interference-free assignment is no longer known due to the way the instances are generated, but it is bounded by a lower bound LB given by the generator. Compared with the first set, these instances are often larger in terms of the number of constraints due to the doubled traffic. Large instances of 300 cells (representing 600 integer variables) contain up to 30,000 interference constraints. As we will see in the next section, these instances are usually harder to solve than those of the first set. 5.2 Performance Evaluation 5.2.1 Baseline Algorithm In order to evaluate the performance of the Tabu algorithm, we need a reference for comparison. We take the standard steepest descent (SD) algorithm (with re-runs) as a baseline reference for several practical reasons. First, SD is a well known and simple algorithm. Second, SD requires no parameter to tune. Third, from the point of view of implementation, SD and TS may share many important data structures. Forth, the results of SD will give us some indication as to the difficulty of the instances tested. The SD algorithm consists of repeating the following SD operation: Steepest descent (SD): At each iteration, choose a best neighbor s' 2 V* such that ties randomly). We must point out that the SD algorithm uses the same neighborhood N(s), candidate list V* and evaluation method as those used by the TS algorithm. 5.2.2 Method of Experimentation and Evaluation Criteria The TS algorithm is run with initial random configurations. For a fixed number of frequencies, TS is run 10 times, each run being limited to a maximum of 100,000 iterations. Once an interference-free assignment is found with NF frequencies, the TS algorithm is called with a new random assignment using NF-1 frequencies. The parameter ff of tabu tenure determined according to the following process. Values from 0.1 to 0.5 were tested for each instance with limited iterations (50,000 in this study) and the best one is used in later experiments (100,000 in this study). In order to make the comparison between TS and SD as fair as possible, we give both algorithms the same number of iterations (100,000). However, while TS may improve its results with more iterations, this is not the case for SD. To avoid this problem, we give SD more re-runs. For each run of TS, SD is run 20 times, each being limited to 5,000 iterations in order to reach 100,000 iterations. The number 5,000 is empirically determined in such a way that it is sufficiently high for SD to reach local optima and move around with some side-walks. The following three criteria are used to carry out the evaluation. NF(S) the smallest number of frequencies found for an interference-free assignment, followed in parenthesis by the number of successful runs which find such an interference-free assignment with NF frequencies. This criterion reflects the quality of a solution. Iter the number of iterations averaged over successful runs for the smallest number of frequencies found for an interference-free assignment. One iteration corresponds to a move from one solution to one of its neighbors. This criterion reflects the machine-independent speed of an algorithm. T(sec) the average user time 7 in seconds averaged over successful runs for the smallest number of frequencies found. The timing is based on a SPARCstation 5 (75MHz, 5.2.3 Experimental Results Table 1 shows the results of TS and SD on 4 instances of the first set and 6 instances of the second set. These instances are taken from a total of 60 instances and represent some hard problems. For each instance, we give its name (first column), the optimal number of frequencies (Opt) for the first set or the lower bound (LB) for the second set (2nd column), followed by the results of TS (3rd-7th columns) and SD (8th-11th columns). For instance, the data for 8.150.20 have the following meanings. TS finds optimal solutions for each of its 10 runs with an average of 18,923 iterations and 123 seconds per (successful) run. SD finds 8 times (out of 200) an interference-free assignment with 10 frequencies after 347 iterations and 3 seconds while none of its 200 runs manages to find an interference-free assignment with 9 frequencies after 5,000 iterations per run. From the higher part of Table 1, we see that the 4 instances of the first set are too difficult for SD to solve to optimality. Indeed, SD requires 2 to 8 extra frequencies to find an interference-free assignment. On the contrary, TS solves all the instances to optimality at each run with a maximum of 10 minutes. Following the size and the difficulty of an Both TS and SD are programmed in C++. Problems Opt/LB TS SD Runs ff NF(S) Iter T[sec] runs NF(S) Iter T[sec] Table 1: Results of TS and SD, 100,000/5,000 iterations per run for TS/SD instance, TS does approximately 70 to 150 iterations per second. Since SD and TS gives solutions of highly different quality, it is difficult to compare their solving speeds. Concerning the instances of the second set (the lower part of Table 1), let us notice that they are more difficult than the previous ones (the two 15.300.25.x instances seem extremely hard). For these instances, the difference between the results of SD and TS is even larger. Indeed, to find an interference-free assignment, SD requires at least 3 extra frequencies above the lower bound. For the last two instances, this excess number reaches respectively 19 and 15. finds interference-free assignments with the lower bound frequency for 2 of 6 instances within the limit of 100,000 iterations. For the 4 other instances, 1 to 5 extra frequencies above the lower bound are necessary to have an interference-free assignment with success rates ranging from 20% to 100%. As shown in the next section, these results will be improved if more iterations is given to the TS algorithm. For these instances, TS achieves approximately to 120 iterations per second. Finally, TS needs about 5 to 50 minutes to find its solutions for these instances. These computing times are considered to be quite satisfactory since in practice up to 48 hours are allowed to find a frequency assignment. Moreover, minimizing the number of frequencies used is far more important than the solution speed. 5.3 Comparisons with Other Methods In this section, we contrast the results of the TS algorithm with the best results on the same instances obtained with other methods including algorithms based on Simulated Annealing (SA), Constraint Programming with ILOG-solver (CP) and Graph Coloring (GCA) [11]. The results of TS are those presented above, limited to 100,000 iterations. The results of SA, CP, and GCA are those reported in [1, 2, 19]. Timing for SA and CP (not available for GCA) is based on HP Stations (HP 735, 125 MHz, 128MB RAM) which are considered to be about 2 to 3 times faster than the SPARCstation 5 (75MHz, 32MB RAM) we used. SA and CP are respectively stopped after 2 and 10 hours on HP Stations, equivalent to about 6 and 30 hours on a SPARCstation 5. Table 2 gives a summary of the results of all these methods on the same set of instances. A minus "-" in the table means that the result is not available. From the data of Table 2, we notice first that TS and SA largely outperform CP and GCA on these instances in terms of the quality of solutions. Indeed, none of these instances is solved to optimality by CP or GCA (with the exception for 8.150.20 by GCA). Problems Opt/LB TS SA 8.75.25.1 8.75.25.3 8.150.15.3 8.150.25.6 15.300.25.6 Table 2: Comparison of TS with other methods (maximum iterations for Many extra frequencies (up to 17) are needed by CP and GCA to find an interference-free assignment. On the contrary, the results of TS and SA are much better on all the instances. In general, the harder the problem, the bigger the difference: for some instances, there is a difference of more than 12 frequencies. Let us now examine the results of TS and SA. In terms of the quality of solutions, we see that the TS algorithm outperforms the SA algorithm on 4 instances (in bold) and does as well as SA on the other instances. This is remarkable since reducing even one frequency for such hard problems makes the problem much harder to solve. In terms of computing time, TS is always faster than SA even for solutions of better quality. If we take into account the difference of computing power between the machines used, it should be clear that TS dominates SA. In order to see how far the TS algorithm can go, another experiment was carried out. In this experiment, we run the algorithm with more iterations (1,000,000 maximum) on 4 instances for which no interference-free assignment was found at 100,000 iterations with lower bound frequencies. Table 3 shows the results of this experiment. Problems Runs TS Table 3: More TS results, maximum iterations = 1,000,000 From Table 3, we see that 8.150.15.3 is solved with 16 frequencies (lower bound). The results of 15.300.25.6 and 15.300.25.9 are improved with a gain of 1 and 3 frequencies respectively. For 8.75.25.1, no interference-free solution is found with frequencies. To solve the last three instances, a large number of iterations are needed with a running time ranging from 6 to 10 hours. In summary, we see that the TS algorithm gives the best results for these instances in terms of solution quality and solving speed. This remains true for 50 other instances solved but not reported here. Due to insufficient information about the testing conditions and implementation details of some algorithms, it is difficult to give a really fair comparison. However, contrasting these results seems to give some clear indications about the performance as to these algorithms. Finally, we point out that in addition to the previous results, the TS algorithm was also applied to real data representing some largest mobile networks used in France to minimize interference for a fixed number of frequencies and to minimize the number of frequencies for assignments. Due to the confidential nature, we cannot report in details our experiments on these networks. Suffice to say however that the TS algorithm gives very competitive assignments for these real networks compared with other methods. 5.4 Discussions The TS algorithm presented in this paper offers a significant improvement over the previous best results on the set of tested instances. The high performance of the algorithm comes from the combination of different features. In this section, we study the influence of some of these features. Conflict-based candidate list strategy In order to see the influence of this strategy, experiments were carried out using the neighborhood N(s) and the conflict-based candidate list V . We re-run the TS algorithm on some instances with different fixed number of frequencies. Table 4 shows the comparative results on a particular instance (8.75.25.3). In the table, each line corresponds to the result of 5 runs and "-" means that no result is found after 100,000 iterations. Iter T[sec] Iter T[sec] 22 267 2 104 0.6 Table 4: Influence of candidate list strategy on 8.75.25.3, maximum iterations = 100,000 We observe from the table that none of the 5 runs (100,000 iterations each) using the neighborhood N(s) is able to find an interference-free assignment using a number of frequencies smaller than 21. On the contrary, the conflict-based strategy allows the algorithm to find interference-free assignments with only frequencies. Moreover, this strategy is much more efficient in terms of number of iterations and computing time. These remarks remain valid for other tested instances. Another popular candidate list strategy is random sampling, i.e. only a percentage of the neighborhood is randomly considered at each iteration. Similar tests show that once again the conflict-based strategy largely dominates this sampling technique. Co-cell constraint handling In order to see the influence of this technique for the TS algorithm, we re-run our times with the CCH option disabled on the second test set (no co-cell constraint exists in the first test set). Table 5 shows the comparative results of TS with and without CCH. Problems LB Runs TS with CCH TS without CCH Table 5: Influence of co-cell constraint handling, maximum iterations = 100,000 From the table, we observe that CCH gives better solutions by reducing the number of frequencies by 1 or 2 for the last 3 instances. Also, to obtain solutions of the same quality, CCH seems to be more robust since the solutions are found more often. More generally, CCH allows to reduce the search space and to accelerate the search process. Re-generation Recall that there are two different ways to generate an initial configuration with NF-1 frequencies when an interference-free assignment is found with NF frequencies. One may take a completely new random configuration or re-generate a configuration from the last interference-free assignment (see Section 4.2). Results represented above are obtained with random initial configurations. It is thus interesting to know if the "re-generation" technique can lead to different results. Limited experiments were carried out for this purpose. We observed that given the same number of iterations, the re-generation often makes the search faster. We also observed that the re-generation may effectively improve the results. In particular, it helps finding interference-free assignments with frequencies (lower bound) for the two largest instances 15.300.25.x. At the same time, we notice that the improvement is problem-dependent. More work is needed to know to which extent and in which cases the re-generation is beneficial in general. 6 Conclusions and Future Work In this paper, we have presented a powerful TS algorithm for the frequency assignment problem. The algorithm has some key features including incremental evaluation, co-cell constraint handling, candidate list strategy and dynamic tabu tenure. The algorithm was evaluated against a steepest descent algorithm on various realistic instances. Results of the TS algorithm were also compared with the best results obtained with other algorithms based on simulated annealing, constraint programming and graph coloring. Empirical evidence shows that the TS algorithm outperforms largely other algorithms by producing the best results for the set of tested instances both in terms of quality of solution and solving speed. Consequently, TS should be considered to be a very competitive and promising method for the frequency assignment problem. The FAP instances solved in this work may be considered to be large for today's networks. However, mobile radio networks grow extremely fast and will reach easily 1,000 to 2,000 cells (several thousands of variables) in the near future. Therefore, very powerful optimization methods are needed to deal with such networks. Parallel and distributed TS techniques may be considered since they are probably among the most promising methods for tackling large and challenging combinatorial problems [5]. Acknowledgments We would like to thank gratefully Dr. A. Caminada from CNET - France Telecom for his assistance. We also thank the referees of the paper for their useful comments. --R "R'esolution du Probl`eme de l'Affectation des Fr'equences Par Programmation Par Contraintes" Some Results of Constraint Programming for Frequency Assignment CALMAR Symposium on Combinatorial Algorithms for Military Applications "A Tabu Search Algorithm for Frequency Assignment" "Towards a Taxonomy of Parallel Tabu Search Heuristics" "A Parallel Genetic Algorithm for Frequency Assignment Problems" "Constraint Handling in Evolutionary Case Study on Frequency Assignment" "Channel Assignment for Cellular Radio Using Simulated Annealing" "Genetic and Hybrid Algorithms for Graph Coloring" "A Neural Network Parallel Algorithm for Channel Assignment Problems in Cellular Radio Network" "A Ressource Allocation Technique for FDMA Systems" "Frequency Assignment: Theory and Application" "Study of Genetic Search for the Frequency Assignment Problem" "Tabu Search for the Frequency Assignment Problem in Cellelar Radio Networks" "Using Tabu Search Techniques for Graph Coloring" "Using Genetic Algorithms to Solve the Radio Link Frequency Assignment Problem" "Channel Assignment for Cellular Radio Using Neural Networks" "Algorithmes de Coloration des Graphes et d'Affectation des Fr'equences: Approches G'eom'etrique et Analytiques, Heuristiques, Recuit Simul'e, Programmation par Contraintes, Algorithmes G'en'etiques" "T-colorings of Graphs: Recent Results and Open Problems" --TR --CTR Sancho Salcedo-Sanz , Carlos Bousoo-Calzn, A portable and scalable algorithm for a class of constrained combinatorial optimization problems, Computers and Operations Research, v.32 n.10, p.2671-2687, October 2005 Junghee Han, Frequency reassignment problem in mobile communication networks, Computers and Operations Research, v.34 n.10, p.2939-2948, October, 2007 J. Mark Ware , Ian D. Wilson , J. Andrew Ware , Christopher B. Jones, A tabu search approach to automated map generalisation, Proceedings of the 10th ACM international symposium on Advances in geographic information systems, November 08-09, 2002, McLean, Virginia, USA Michel Vasquez , Jin-Kao Hao, A Heuristic Approach for Antenna Positioning in Cellular Networks, Journal of Heuristics, v.7 n.5, p.443-472, September 2001 Albert Grf , Thomas McKenney, Ensemble planning for digital audio broadcasting, Handbook of wireless networks and mobile computing, John Wiley & Sons, Inc., New York, NY, 2002 Yehuda Ben-Shimol , Boaz Ben-Moshe , Yoav Ben-Yehezkel , Amit Dvir , Michael Segal, Automated antenna positioning algorithms for wireless fixed-access networks, Journal of Heuristics, v.13 n.3, p.243-263, June 2007 Carlos Bousoo-Calzn , Sancho Salcedo-Sanz, A Discrete-Time Quantized-State Hopfield Neural Network, Annals of Mathematics and Artificial Intelligence, v.42 n.4, p.345-367, December 2004 L. M. San Jos-Revuelta, A new adaptive genetic algorithm for fixed channel assignment, Information Sciences: an International Journal, v.177 n.13, p.2655-2678, July, 2007 Marco Chiarandini , Thomas Sttzle, Stochastic Local Search Algorithms for Graph Set T-colouring and Frequency Assignment, Constraints, v.12 n.3, p.371-403, September 2007 Harilaos G. Sandalidis , Peter Stavroulakis, Heuristics for solving fixed-channel assignment problems, Handbook of wireless networks and mobile computing, John Wiley & Sons, Inc., New York, NY, 2002 Celso C. Ribeiro , Simone L. Martins , Isabel Rosseti, Metaheuristics for optimization problems in computer communications, Computer Communications, v.30 n.4, p.656-669, February, 2007
constraints;optimization;frequency;tabu search;assignment
595008
A Constraint Programming Framework for Local Search Methods.
We propose in this paper a novel integration of local search algorithms within a constraint programming framework for combinatorial optimization problems, in an attempt to gain both the efficiency of local search methods and the flexibility of constraint programming while maintaining a clear separation between the constraints of the problem and the actual search procedure. Each neighborhood exploration is performed by branch-and-bound search, whose potential pruning capabilities open the door to more elaborate local moves, which could lead to even better approximate results. Two illustrations of this framework are provided, including computational results for the traveling salesman problem with time windows. These results indicate that it is one order of magnitude faster than the customary constraint programming approach to local search and that it is competitive with a specialized local search algorithm.
Introduction Local search methods in Operations Research (or) date back to over thirty years ago [13]. Applied to difficult combinatorial optimization problems, this heuristic approach yields high-quality solutions by iteratively considering small modifications (called local moves) of a good solution in the hope of finding a better one. Embedded in meta-heuristics designed to escape local optima such as simulated annealing and tabu search, it has been very successful in achieving near-optimal (and sometimes optimal) solutions to a variety of hard problems [8][23][1]. Constraint Programming (cp) features very flexible modeling capabilities to easily and closely reflect the various constraints of a problem. In solving real-life combinatorial optimization problems, it has to date almost invariably adopted the branch- and-bound strategy, a global and therefore complete search method. 1 When an exact algorithm proves too costly, simple approximate algorithms are often devised by heuristically discarding the least-promising branches in the branch-and-bound search tree (a form of beam search). * A preliminary version of this paper appeared as [18]. What can constraint programming bring to local search methods? The fact that heuristics tend to be specialized for a particular context can become a drawback when faced with an even just slightly different context, such as side constraints: adapting them may demand considerable time and effort. It is therefore tempting to gain both the efficiency of local search methods and the flexibility of constraint programming. In this paper we argue that local search is not so foreign to branch- and-bound search and that constraints can be more actively involved in the exploration of these local search spaces. We propose a clean integration of local search in constraint programming by keeping on doing what comes naturally, i.e. branch- and-bound, but on a different search space, though related to the original one. In contrast with or practice, the resulting framework maintains a clear separation between the constraints of the problem and the actual search procedure. For both or and cp, the potential pruning capabilities open the door to more elaborate local moves, which could lead to even better approximate results. This adaptation of local search to the constraint programming paradigm stems from its perception as a generalization of conventional cp branch-and-bound search. In the latter case, we branch on the variables of the model and the neighborhood degenerates into the whole solution space for which a single iteration becomes sufficient since the optimal solution will necessarily be found. In order to lift this approach to local search, the art then consists of choosing a representation for a particular neighborhood structure which cp branch-and-bound search can exploit. Each iteration of local search will simply be a branch-and-bound search, but on a different search space. The active role of modeling constraints (discarding infeasible neighbors) together with lower bounds on the cost of partial solutions (discarding unattractive neighbors) will help prune the tree and thus reduce the search effort over the whole neighborhood. The rest of the paper is organized as follows. Section 1 first gives an overview of local search methods and constraint programming. Our general framework for local search in cp is presented in section 2. The implementation of that framework is then illustrated on personnel scheduling (section 3) and single-vehicle routing (section problems. An empirical study is included for this latter type of application. Finally we address the flexibility offered by our framework in section 5. 1. Background 1.1. Local Search Methods in Operations Research Local search methods involve repeatedly going from one solution to another through a local move. What constitutes a valid local move varies according to the problem and even within it (see for example section 4.1). The set of all solutions reachable from a solution s through a local move is termed the neighborhood of s. The set of all feasible solutions in this neighborhood will be called its feasible neighborhood. Given this framework, a simple strategy called iterative improvement moves to the best feasible neighbor (i.e. of lowest cost) every time until it does not improve on the current solution, reaching a local optimum. Several ways of alleviating the obvious drawback of this strategy have been proposed. Multi-start iterative improvement achieves local optima from a pool of solutions and returns the best one. Genetic local search builds upon the previous by recombining the local optima (in the fashion of genetic algorithms [10]), applying iterative improvement, discarding the least-promising solutions and repeating the process until some stopping criterion is satisfied. Two very successful strategies try to escape local optima by allowing moves which temporarily increase the cost of the solution. Tabu search [7] moves to the best neighbor at each iteration, regardless of whether or not it improves on the current solution. To avoid cycling, a dynamic list of tabu solution attributes is kept. Typi- cally, such a list covers recently examined solutions, which will remain forbidden for a certain number of iterations. Simulated annealing [12] randomly selects a neighbor at each iteration. If it improves on the current solution, the move is performed; otherwise, it will be performed with a certain probability which depends on the cost difference and which also decreases over time according to a cooling schedule. Both strategies iterate until some stopping criterion is satisfied. One crucial aspect in all of these local search methods is obviously the choice of the neighborhood structure (see for example [9]). Ambitious neighborhoods increase the chances of success but are more expensive to explore for methods which need to do so. Small neighborhoods are both simple and fast to explore but may prevent us from ever reaching a particularly good solution: it could require a sequence of local moves (as opposed to a single one in a larger neighborhood), and every such move would have to be the one selected in its iteration. A neighborhood which strictly includes another induces a search which encounters fewer local optima and thus facilitates their avoidance. Because of the above, large neighborhoods are attractive. Several techniques have been developed to speed up their exploration. The size of the neighborhood can be somewhat reduced either by ignoring parts of it which are unlikely to produce good solutions or, in a more exact fashion, by interrupting a carefully engineered exploration when the remainder can only lead to worse (or infeasible) solutions. In addition, the feasibility of neighbors must be assessed. As an early example in routing problems, [24] describes a way to verify time-window constraints in constant time per neighbor, though assumptions about the neighborhood structure must be made. Others may perform approximate tests of feasibility to quickly identify promising neighbors which are then thoroughly investigated, with the potential risk of missing the best one. So, the challenge of expressive neighborhoods has been met with specialized techniques embedded in the local search and sometimes with compromises. 1.2. Constraint Programming Constraint programming solves combinatorial problems by actively using the constraints of the problem to implicitly eliminate infeasible regions of the solution space. The algorithm at the heart of this approach implements complex logical reasoning over the set of constraints. To every variable of a cp model is associated a domain: each value in that domain represents a possible value for the variable. Constraints on the variables forbid certain combinations of values. Picturing the model as a network whose vertices are the variables and whose (hyper)edges are the constraints provides insight into the basic algorithm used in cp. A vertex is labeled with the set of values in the domain of the corresponding variable and an edge is incident to those vertices representing the variables appearing in the associated constraint. Looking locally at a particular edge (constraint), the algorithm attempts to modify the label (reduce the domain) of the incident vertices (variables) by removing values which cannot be part of any solution because they would violate that individual constraint; this local consistency step can be performed efficiently. The modification of a vertex's label triggers the inspection of all incident edges, which in turn may modify other labels. This recursive process stops when either all label modifications have been dealt with or the empty label is obtained, in which case no solution exists. The overall behavior is called constraint propagation. Since constraint propagation may stop with indeterminate variables (i.e. whose domain still contains several values), the solution process requires search and its potentially exponential cost. It usually takes the form of a tree search in which branching corresponds to fixing a variable to a value in its domain, thus triggering more constraint propagation. We call variable-selection heuristic and value- selection heuristic the way one decides which variable to branch on and which value to try first, respectively. For combinatorial optimization problems, the tree search evolves into a branch-and-bound search in which one branches in the same way and lower bounds at tree nodes are obtained by looking at the smallest value left in the domain of a cost variable. 1.3. Local Search in Constraint Programming As we just saw, branch-and-bound search is by far the most popular (and natural) solution strategy in constraint programming to handle combinatorial optimization problems. A few exceptions are nevertheless found in the literature. A recent survey on applications of constraint programming [28] describes applications in which local search and cp are "loosely connected": the latter is used as a preprocessing step for the particular heuristic employed. Others offer a tighter connection but essentially consider individually each of the possible moves in a neighborhood to then assess their feasibility and cost [22][3][2] 2 . This can be a costly endeavor when the nature of the local moves is such that a great number of possibilities must be considered. To solve constraint satisfaction problems, which abound in Artificial Intelligence, repair-based methods (e.g. [25][15]) have had a good share of success. They can be viewed as local search methods which typically consider a very small neighborhood (changing the value of one variable) and use iterative improvement in order to minimize the number of violated constraints. 2. A New Combination of Local Search and CP As mentioned in the introduction, we will view local search as a sequence of cp branch-and-bound searches. Given a combinatorial optimization problem P , which we call the master problem, and some initial solution to it, we will solve a sequence of auxiliary problems using standard constraint programming techniques. P is described by a constraint programming model whose role is to ensure the feasibility of our solutions. Each auxiliary problem is described by another constraint programming model, called the neighborhood model, which describes the local search space. The two types of models are related by interface constraints. Definition 1. Neighborhood model. Let N denote some neighborhood structure for solutions to a combinatorial optimization problem P . A set of finite-domain variables usually distinct from the variables appearing in the model for P , together with a (possibly empty) set of constraints on f- is a neighborhood model for N if there is a one-to-one mapping between the set of feasible combinations of values for f- and the neighbors in N . Example: Let P be the traveling salesman problem on m cities and consider the exchange of two cities in a solution as our neighborhood structure N . We introduce variables I and J , both ranging over the interval [2; m], and the constraint I ! J between them. If a solution is represented as the sequence of cities forming the tour, a particular (feasible) combination of values for fI ; Jg is interpreted as exchanging entries c I and c J in that solution to obtain a neighbor. One easily verifies that this constitutes a neighborhood model. The requirement of a one-to-one mapping could be relaxed to a surjective mapping though this would mean that a neighbor may be examined more than once: without the constraint I ! J in the example above, symmetry would lead to identical solutions. However, surjectivity is crucial since otherwise we would miss some of the neighbors. The solution space defined by the neighborhood model can be explored by standard cp branch-and-bound search. We branch on f- and also bound the cost of partially constructed neighbors. As we saw in section 1.1, some heuristic algorithms based on local search, such as tabu search and variations on iterative improvement, are interested in acquiring the best solution in the neighborhood and hence do not require to manipulate the whole of it per se. We can therefore record the cost of the best neighbor found so far and derive lower bounds for partial neighbors to reduce the portion of the neighborhood that has to be explored, as in conventional branch-and-bound. Note that we do not prescribe the order in which we select variables for branching: in accordance with common practice in the constraint programming community, dynamic variable-selection and value-selection heuristics may be used to increase the efficiency of the search. Initially, the model for the master problem is stated and the variables of that model are only constrained to that extent - in other words, they are not bound to the current solution. Then the local search takes place in the form of a complete NEIGHBORHOOD MODEL local move restricting neighbors fixing main constraint checking interface constraints Figure 1. The interaction between the master and neighborhood models. branch-and-bound search on the neighborhood model. At different stages of the branch-and-bound search, valuable information can be exploited in order to restrict the set of allowable values for the neighborhood-model variables on which that search is performed and, more importantly, constrain as well the principal modeling variables of the master problem which may independently prune our search tree by propagating these changes through modeling constraints we need not know anything about (see figure 1). We automate this behavior by introducing a set of conditional constraints 3 called the interface constraints. Definition 2. Interface constraints. An interface constraint has the form stands for a subset of the variables in the model for P . It adds constraint once each of - fixed. The set of interface constraints establishes the link between the model for P and the neighborhood model: as variables of the latter model are fixed during the branch-and-bound search, constraints are generated on some variables of the neighborhood model and some of the master problem, reflecting the fact that we narrowed the search. We see two main advantages for such a framework: Flexibility, genericity. The modeling constraints for P (and often some of the are kept separate from the neighborhood model. This provides a flexible way to state instance-specific constraints, independently of the local search. It is especially useful when lots of different side constraints are present in the problem: they will not clutter the local search with explicit feasibility tests as is often the case in operations research heuristics. They will rather be indirectly involved when some of the modeling variables will be further constrained through the interface constraints as a result of branching on - . The end result is a generic neighborhood exploration method parameterized by the type of neighborhood structure but not by the nature of the modeling constraints for the master problem. Table 1. Availability of physicians A,B,C,D,E. (a=am, p=pm, n=night). Mon Tue Wed Thu Fri A p/n a/p n a/p/n p C p/n a a/p/n a/n n Search Economy. There is a strong relationship between the savings brought about by this branch-and-bound approach and how ambitious the neighborhood is. Typically, enlarging the neighborhood means increasing the degrees of freedom and so requires a greater number of variables to encode its structure. This translates into a greater depth of the neighborhood search tree and a potentially larger gain with every branch pruned 4 , either from the lower bound at the particular node or the modeling constraints. This constitutes an asset in view of the current trend toward more ambitious neighborhood structures. The freedom of choosing the order in which variables are fixed and values tried also contributes to that gain. On the other hand, there is of course a price to pay for that flexibility, namely the cost of constraint propagation and of built-in backtracking. We will attempt to empirically evaluate the overall efficiency of the approach in sections 4.4 and 4.5. 3. An Introductory Application: Physician Scheduling Here is a first illustration of this general framework. On a hospital ward, there must be a physician on duty at all time. Consequently, a ward schedule has to be established for the staff of physicians. As can be expected, the rules governing the construction of a valid schedule vary from one hospital to the next. We describe but one such context. Each day is divided into three shifts: am, pm and night. (For the night shift, the physician is simply on call.) The planning horizon is typically a few months. There are several rules about the frequency of shifts covered by the same physician: 1. no more than one shift per day; 2. no two consecutive shifts (for example, Monday-night then Tuesday-am); 3. no two consecutive night shifts. Physicians also have other duties elsewhere, especially the ones who have a part-time status. They hand in a form specifying the shifts for which they are available (table 1 gives a small example for five physicians over an horizon of five days). If two physicians share a responsibility outside the ward on a certain day, they may ask that at least one of them be off daytime duty on that day. Finally, a target workload (in number of shifts) is assigned to each physician, which may vary according to her status. A good schedule should remain close to the targets. Let the planning horizon range from day 1 to n and let be the set of shifts to assign in the schedule. Let MD represent the set of physicians on staff, A k , the set of pairs hi; ji such that physician k is available for the j shift on day i and R, the set of pairs hfk; k 0 g; ii such that physicians k and k 0 cannot be both on daytime duty on day i. A natural model for this problem associates a distinct variable to every shift being assigned: The constraints listed above can be expressed as follows, in the same order: stand for the target workload of physician k whereas ff k stands for her actual workload in the schedule. We adopt a least-squares objective function to be A local move will consist of changing the value of one slot in the schedule. Let oe represent the current schedule and oe(hi; ji) the physician currently covering the j shift on day i. A neighboring schedule of oe can be encoded through a variable S representing the slot to be modified and a variable V representing the new value for that slot. Our neighborhood model is then: The cost of a local move replacing physician k 0 by k in some slot is which simplifies to If the target workload is the same for all physicians, evaluating the relative merit of a move simply amounts to looking at ff k \Gamma ff k 0 . Table 2. A feasible schedule (left) and an optimal schedule (right). Mon Tue Wed Thu Fri Mon Tue Wed Thu Fri Table 3. The partial schedule after choosing one possible completion (right). Mon Tue Wed Thu Fri Mon Tue Wed Thu Fri Let ffXgg denote the current domain of finite-domain variable X . We introduce the following interface constraints: When S is fixed to some slot ha; bi: conditional constraint (7) fills all the other slots with the value in the current schedule since they will not change; conditional constraint (8) binds V to S a;b , thus restricting V to the possible values for that slot and forbidding the current value for S a;b , through constraint of the neighborhood model. When V is fixed to some physician k: conditional constraints (9) and (10) maintain the value of all the slots which cannot possibly change: those whose value is k already and those which cannot take that value; conditional constraint (11) restricts S to the slots which may admit value k. Example: Consider a simple instance with five physicians, a scheduling horizon of five days, physician availabilities as given in table 1 and a uniform target workload of three shifts per physician. In addition, physicians A and D cannot be both on daytime duty on Thursday and neither can B and C on Wednesday. Table 2 gives a feasible schedule in which A works four shifts and D only two, yielding a score of 2 on the objective function. We will investigate the possible local moves from that schedule. In a cp branch-and-bound search on our neighborhood model, a popular variable- selection heuristic would have us branch on the variable with smallest domain first. In our case that would be V . As far as trying the different values for V , it makes sense to order them according to a lower bound on the relative merit of the possible local moves introducing that value, and starting with the smallest bound. For value A, we look at (ff A \Gamma ff x similarly, for the other values we get \Gamma1; \Gamma1; \Gamma2; \Gamma1, in that order. We will therefore start by examining D, as illustrated in the search tree on the left in figure 2. This triggers constraints (9)-(11), the first two partially filling the schedule as illustrated on the left in table 3. Entries shown in bold result from more elaborate reasoning than a mere initial unavailability of D for that shift, and are achieved through constraint propagation. For example, once shift h3; nighti has been filled by D (constraint (9)), constraint (3) tells us that shift h4; ami cannot be filled by D as well and so should take its current value A (constraint (10)). Now then, because A and D cannot be both on daytime duty on Thursday (constraint (5)), shift h4; pmi should also take its current value E. The domain of S is made to correspond to the unassigned shifts in the partial schedule, fh1; nighti; h2; pmig through the combined effect of constraints (11) and V 6= oe(S). Branching now on S, we can again order the two branches according to the lower bound, though here they have identical value. The leftmost branch brings us to a terminal node, changing the value of slot h1; nighti to D, with a relative merit of \Gamma1. Conditional constraint (7) fills the rest of the schedule. Since that relative merit score is as good as any lower bound on unexplored branches in the search tree, the rest of the tree can be pruned and we are done (see figure 2(left)), yielding the schedule on the right in table 3. The actual cost of that move is which keeps the objective function at 2. A further iteration using the same strategy would produce the search tree on the right in figure 2 and a local move of cost \Gamma2, bringing the objective function down to 0. We would therefore have an optimal schedule, shown on the right in table 2 with the overall modifications in bold. This small example is already sufficient to bring forth the two advantages stressed at the end of section 2. Exploring the neighborhood by first fixing V then S runs contrary to what would naturally come to mind, that is, first choose a slot to modify and then pick a new value for it. Nevertheless, it is probably the better strategy of the two since the search trees generated tend to be more sparse. The combination of this variable-selection heuristic with a value-selection heuristic, constraint propagation and lower-bound pruning yields a neighborhood exploration which is very efficient in terms of the number of neighbors (terminal nodes) actually considered. The other advantage was a generic local search with the flexibility to handle instance-specific constraints. As we mentioned before, in this type of personnel scheduling problem the rules of the game depend on where you are. With this "C" "E" "A" "D" "A" Figure 2. Neighborhood search trees: first (left) and second (right) iterations. Branch labels are "value"/"lower bound". Terminal nodes are labeled with the relative merit of the move. approach, the different hospital contexts will be reflected and handled in the master model and not in the local search itself. 4. A Routing Application Using Tabu Search This section provides another instance of the general framework described in section 2. It addresses a well-known problem on which several local search heuristics have been applied in the past. The traveling salesman problem with time windows (tsptw) consists of finding a minimum cost (usually the total travel distance or total schedule time) tour of a set of cities where each city is visited exactly once and which starts and ends at a unique depot. In addition, each city must be visited within its own time window. Early arrival is allowed but implies a waiting time until the beginning of the window. [24] showed that simply deciding whether there exists a feasible solution to an instance of the tsptw is NP-complete. Nevertheless, the full problem has important applications in bank and postal deliveries, school-bus routing and scheduling, disjunctive scheduling with sequence-dependent processing times, automated manufacturing environments and as a subproblem of the vehicle routing problem with time windows (vrptw). A constraint programming model for our master problem, the tsptw, was presented in [19]. We sketch it below, emphasizing the main variables which will be referred to later on. Let ng represent the set of cities to visit and duplicate the unique depot into an origin-depot and a destination-depot, identified as 0 and tour thus becomes a Hamiltonian path starting at 0 and ending at n+ 1. At the heart of the model are variables S i associated to each of the cities (and the origin-depot) and which represent their successor in the tour. Their domain will therefore be an integer in the range valid tour assigns a distinct successor to each city and avoids sub-tours. We also define predecessor variables counterpart of the S i 's (S To account for the scheduling component of the problem, variables T i are introduced to represent the time at which we visit each city. These must take a value within their respective time window while being coherent with the order in which cities appear on the tour and taking into consideration the travel time between cities. Several redundant constraints are also present to increase the amount of propagation. 4.1. Some Neighborhoods for Routing Problems Given the inherent difficulty of the problem, several or heuristic algorithms have been designed and among them some based on local search. We briefly describe the most popular neighborhoods in the context of routing problems, which include the tsptw but may allow solutions consisting of a set of routes and involve other restrictions such as capacity constraints. Neighborhoods generated by modifications at the level of vertices include re-inserting a vertex or exchanging two vertices. GENI, a generalized insertion procedure [5], directs the re-insertion between some of its p (from 2 to 7) nearest neighbors. In the -interchange [17], subsets of vertices are selected each from a different route and then swapped. The possibility of an empty subset allows simply re-inserting vertices. For efficiency reasons, - rarely exceeds 2. At the level of edges, a k-interchange [13] replaces k edges in the solution by k others that reconnect the route(s). Lin-Kernighan moves [14] for the tsp increase k until a gain criterion fails and only consider nearest neighbors of vertices for the exchanges. An Or-opt move [16] relocates a string of one, two or three consecutive vertices. A 2-opt ? move [21] replaces two edges (v different routes with edges (v This exchanges the end portion of one route with the end of the other. The CROSS exchange [27] goes further by exchanging a middle portion of one route with a middle portion of the other, replacing four edges, and includes the two previous types of moves as special cases. 4.2. An Orientation-Preserving 3-Interchange Neighborhood We consider the orientation-preserving 3-interchange neighborhood and provide a neighborhood model for it. Let T be the current tour. If I represents a city on this tour, then I its successor (respectively predecessor) on T . We define the following binary relation on cities: I OE J holds if I appears before J on T . Without loss of generality, let I OE J OE K. A 3-interchange move from T deletes the three edges reconnects the tour by introducing three new ones on the vertices (cities) I , I orientation-preserving 3-interchange move reconnects the tour in the only possible way which does not reverse any of the original route segments, by adding edges I I Figure 3. The orientation-preserving 3-interchange. 3). This is desirable when a scheduling component is present in the problem, such as time windows. Two segments of the original route have been swapped but the ordering within a segment is kept. The neighborhood defined by orientation-preserving 3-interchange moves can be encoded through three finite-domain variables I ; J; K, ranging from 0 to n. Their interpretation is the one given above. Because of constraint I OE J OE K, each neighbor is in one-to-one correspondence with a 3-tuple of values describing it. We therefore have a neighborhood model: ng I OE J OE K: Associating a cost c ij to every edge (i; j), the cost of a local move is given by the total cost of the arcs we add minus the total cost of the ones we remove. Its smallest possible value will be used as our lower bound and will be refined automatically through constraint propagation as the domains of I ; J; K shrink (ultimately to a single value). Here are the interface constraints used during the branch-and-bound search (refer to figure 3 throughout): Conditional constraints (12)-(15) deal with information traveling from the neighborhood model (the box on the right in figure 1) to the master model (the box on the left in the same figure): in (12) for example, as we try a value for I , say i, we know that the initial tour segment up to i will remain in the new tour and so the appropriate successor variables in the routing model can be fixed. (13)-(15) respectively address tour segments hi which are unchanged in the neighboring tour. Conditional constraints (16) and (17) reflect the fact that having fixed I to i allows us to constrain further J and K: J must appear, in the current tour, just before the successor of i in the new tour (corresponding to the addition of edge must be the predecessor of i + in the new tour (corresponding to the addition of edge Similarly, (18) and (19) restrict I and K given a value for J while (20) and (21) restrict I and J given a value for K. In turn, this is information traveling from the master model to the neighborhood model: the successor and predecessor variables, constrained by the master model, restrict the possible values for I , J and K, which are used to identify the potential arcs to be added and removed in a neighboring tour. Suppose for example that the variable ordering in our search tree is I ; J; K throughout. Then at depth 1, with I fixed, conditional constraints (12), (16) and (17) will be triggered; at depth 2, with both I and J fixed, (13), (19) and (18) will be triggered but only the first two will be useful since the value of I is already known; finally at depth 3, with all the variables fixed, only (14) and (15) will be worth triggering. As we saw, only the S i 's and P j 's in the master model and the I ; J; K variables for the neighborhood model were involved in the interface. Note for instance that the time variables T i were not directly involved. Example: Consider the following example in which we wish to find a good tour of cities fb; c; d; e; f; gg starting at a and ending at h, the two copies of the depot. Aside from the usual constraints in the master model, we add two others, c OE g (in any feasible tour) and S d 6= g, to show how instance-specific constraints are taken into account when examining neighbors. We shall concentrate here on the implicit pruning of infeasible neighbors through the constraints and leave out that of unattractive neighbors through the lower bounds. We therefore ignore the cost of individual arcs in our exploration. Our variable-selection heuristic will again be the common "smallest-domain-first", with lexicographic order as a tie breaker. Let the current tour be ha; b; c; d; e; f; Initially, branching variables I , J and K have domains as indicated in figure 4-i; they are of equal size and so we first branch on I . At the root of our search tree we have five possible values for I : a, b, c, d or e. They should be tried in turn for a complete exploration of the neighborhood but let us jump to when b is considered (figure 4-ii). Several things happen: { c, d, e, f, g } { a, b, c, d, e } { b, c, d, e, f } c f d e a { c, d, e, f, g } { a, b, c, d, e } { b, c, d, e, f } c f d e a I c d e I b c d e a I d e I d e c f d e a { d, e, f } { c, d, e } I b c f d e a e d -i-ii-iii-iv- Figure 4. Exploring a neighborhood. ffl The neighborhood-model constraint I OE J OE K triggers the removal of value b from the domain of J and then c from the domain of K. ffl Conditional constraint (12) sets S a to b as the initial tour segment. ffl Conditional constraint (16) enforces could reduce the domain of J depending on what the possible values for S b are. ffl Conditional constraint (17) enforces . The combined action of instance-specific constraint c OE g removes g from K's domain, since g is certainly not part of the domain of P c . In addition, f is removed from J's domain, through J OE K. We are left with three possible values in each of J and K's domain and so move on to fix J . Suppose we try d (figure 4-iii): ffl The neighborhood-model constraint J OE K triggers the removal of value d from the domain of K. ffl Conditional constraint (13) sets S c to d as the current tour segment from i + to j. ffl Conditional constraint (19) enforces Now since S d 6= g was given as an instance-specific constraint, (S d so f is removed from K's domain. Finally we fix K to the only possible value, e (figure 4-iv): ffl Conditional constraint (14) does not result in any variable assignment since tour segment consists of the single city e. ffl Conditional constraint (15) sets S f to g and S g to h as the remaining tour segment. 4.3. A Tabu Search Algorithm In order to evaluate some of the efficiency issues related to this approach, we used the previous neighborhood structure within a tabu search algorithm. Our aim is twofold: to evaluate the efficiency of the branch-and-bound search in terms of the proportion of the neighborhood explicitly visited; to compare its execution time with that of a conventional, specialized tabu search algorithm with the objective of measuring the cost of constraint propagation and backtracking. We will therefore leave out ad hoc constraints and remain strictly within the reach of that specialized algorithm for the tsptw. This section describes the specifics of the algorithms. A random tabu tag is associated with each pair of cities (each arc) and initially set to zero. When a move is performed at iteration i, the three arcs being replaced in the tour are only allowed to be brought back into the tour after iteration i+', where ' is an integer randomly drawn from the interval [5; 10] (a ' is drawn separately for each arc): their tabu tag is thus updated to i '. The tabu status of a move is determined by examining the tabu tags of the three arcs it would bring into the tour: unless each tag is less than the index of the current iteration, the move is declared tabu. However, an aspiration criterion is included according to which the move will be performed anyway if its cost improves on the cost of the best tour met so far. To obtain our initial feasible solution, we use a "near-sighted" version (with of the generalized insertion heuristic geni-tw [6]. We then run 100 iterations of tabu search on it (this fixed number of iterations serves as our stopping criterion and simplifies comparisons in the next section). Two versions of the cp tabu search algorithm were implemented, coined "full" and "light". They differ in the extent of the master model used. Neither includes the structural tsp constraints (distinct successors; sub-tour avoidance) since the nature of a move guarantees a valid tour (ignoring time windows). The "full" version uses the rest of the model outlined at the beginning of section 4 while the "light" version omits the redundant constraints, keeping only what is necessary to ensure time-window compliance. A non-cp, specialized version only departs from the other two in the way the neighborhood exploration is carried out: three nested loops iterate over the whole neighborhood and the feasibility of each neighbor is explicitly assessed without resorting to constraint propagation. Finally, a "passive" cp version also proceeds by nested loops but uses constraint propagation (through the "light" model) to verify the feasibility of neighbors. It represents a popular approach to combining cp and local search (as indicated in section 1.3) and does not actively use the constraints to restrict the part of the neighborhood which must be explored. All four implementations are written in C++ (the cp versions use ilog Solver, a C++ library implementing constraint programming for combinatorial problems [11]) and even share some of the code. The seed for the generation of random tabu tags is the same so that all four executions follow identical sequences of moves on a given instance. Neighborhood exploration is therefore the only variable. As in the example at the end of the previous section, our variable-selection heuristic for the branch-and-bound search was "smallest-domain-first". When a tie-breaker was frequently needed (i.e. for the "light" version without lower-bound pruning, which reduces little the domain of branching variables), we found that the ordering I ; K; J performed significantly better than lexicographic ordering. It has the advantage of identifying early on the only backward arc introduced by the move, (K; I + ), which is most likely to cause a time-window violation. For the value- selection heuristic, we considered several strategies. Our first attempt was to imitate what we had done in the physician scheduling example and order values according to some measure of the cost of the related moves. For each value v in the domain, an inexpensive measure such as "\Gammac vv + ", the cost of one of the arcs being replaced, or refinements such as "minfc vSv proved no better and even worse than a simple lexicographic ordering, on our bench- marks. A possible explanation is that whereas in the physician scheduling example it seemed likely that the selected branch of the search tree could be extended to a feasible move, here such likelihood is doubtful as hinted in the next section by the constrainedness indicator. As a result, we may not find a feasible neighbor for a while and thus miss opportunities to prune unattractive subtrees. An orientation-preserving 3-interchange should more likely be feasible if it swaps two short tour segments, i.e. when I , J and K are close on the tour. Lexicographic order was successful because our initial solutions happened to almost follow that ordering and so lexicographic proximity was synonymous to proximity on the tour. Accordingly, we implemented a simple value-selection heuristic, favoring the earliest v on the tour, which outperformed on average all the previous strategies. 4.4. Experimental Results For our experiments, we considered a set of Euclidean problems with travel times between cities taken as the distance separating them. They are in fact subproblems of a well-known vrptw testbed from [26]. The original problem instances feature 100 cities distributed on the [0; 100] \Theta [0; 100] grid and require several vehicles to service them while obeying the side constraints. Some partition of the cities such that each group may be visited by a single vehicle yields our subproblems, 25 in all. We used the instances from the RC2 class which features a mixture of clustered and uniformly distributed cities and requires few vehicles, yielding larger routes. Their resulting size varies from 20 to 42 cities. In addition to execution time, we introduce two indicators to analyze the results: jfeasible neighborhoodj jneighborhoodj \Theta 100 Table 4. Constrainedness (fl) and search effort (ffl) for several tree-search strategies. problem fl ffl-full ffl-light pruning pruning pruning pruning rc203 3.4 36.4 12.9 40.2 77.1 20.8 74.0 rc204 10.1 54.5 11.6 72.2 94.3 12.7 88.4 rc206 1.1 17.9 8.3 17.1 31.4 14.1 34.7 rc207 2.5 36.4 11.1 37.1 61.6 13.3 58.9 rc208 4.1 64.3 18.8 79.7 100.0 17.3 93.4 will indicate how constrained a problem instance is by evaluating for a given solution the proportion of neighbors which are feasible; jsearch-tree leavesj jneighborhoodj \Theta 100 will measure the effort put into our neighborhood exploration by comparing the size of the search tree with the size of the neighborhood. Both fl and ffl are expressed as percentages. Table 4 evaluates fl and ffl on the benchmark problems for the versions which explore the neighborhood by tree search. Instances are aggregated according to the original vrptw instances (rc201 to rc208) since they tend to have similar time-window characteristics. Each entry thus represents the average of 2 to 4 tsptw instances, each averaged over 100 iterations of tabu search. The second column gives the percentage of feasible neighbors; the following two groups of three columns, the search effort for the "full" and "light" version, respectively. Within each group, the column on the left reports on a plain tree search, whose pruning comes from constraint propagation only and not from lower bounds on the cost at tree nodes; the middle column reflects the addition of lower-bound pruning to the tree search to obtain branch-and-bound; we leave out the column on the right for now. Our first observation is that practically all these instances exhibit very constrained neighborhoods for the solutions considered: fl averages 3.3%. Even rc204, with very few meaningful time windows, has an average fl of 10.1%. Avoiding the systematic exploration of the whole neighborhood therefore appears advantageous. Looking at the search effort for plain tree search, the "full" version performs much better than the "light" one in examining roughly half the number of leaves: as can be expected, more powerful propagation ends up producing smaller ffl's but also slows down the exploration as we will see in table 5. The gap in the search effort narrows for branch- and-bound search where the lower-bound pruning common to both versions has a significant impact. That impact is striking when we look at the difference between plain tree search and branch-and-bound search within each version, particularly the "light" one. The variation in the search effort between the eight groups of instances, especially visible in column six where we only rely on the time windows for Table 5. Computation times (in cpu seconds on a sun Sparc 10) for several tree-search strategies. problem full light pruning pruning pruning pruning rc206 4065.2 3488.4 4538.0 282.0 1767.9 323.4 rc207 8532.2 5103.5 8933.6 524.4 2936.7 686.3 rc208 40444.4 9454.2 28542.6 964.3 3475.7 1663.9 pruning, reflects the variation in typical time-window width between those groups of instances. The aspiration criterion included in our tabu search algorithm prevents missing out on the best solution found so far simply because the move leading to it is tabu. On the other hand, it requires evaluating every tabu move. In a tree-search setting such as ours, knowing at an internal node that a tabu arc is involved implies that every move through that node is necessarily tabu. Disregarding the aspiration criterion could allow pruning whole subtrees of tabu moves. We investigated what we call tabu pruning. Each time one of the branching variables is fixed to some value v, thus identifying a break in the current tour, tabu pruning eliminates tabu arcs from the two sets of candidates, f(v; w) to reconnect that break. This translates into removing some of the possible values for the remaining branching variables. Columns five and eight of table 4 report the result of adding tabu pruning to plain tree search (without lower-bound pruning). This often lowers the search effort though not necessarily (for example, rc208 in column five or rc206 in the last column) since it affects the variable-selection heuristic and so the topology of the search tree. Table 5 is organized as the previous table and gives the computation times. Turning first to the "full" version, we see that the addition of lower-bound pruning does significantly reduce the computation time, as hinted by the reduction in search effort in the previous table. The usefulness of tabu pruning is not proven here: it may decrease or increase the computation time, depending on the instance. Tabu pruning can also change the best solution found by the tabu search algorithm. Six of our 25 benchmark problems showed an increase of the solution cost ranging from 0.1% to 1.5%. On the other hand, one problem showed a decrease of 1.9%. It is not at all clear that a large and powerful constraint model with several redundant constraints, designed with a standard branch-and-bound exact (global search) algorithm in mind, is appropriate for local search. Indeed, the figures for the "light" version show a marked improvement over the other one. For plain tree search, this translates into a decrease in the area of one order of magnitude. What comes as a surprise is that the addition of lower-bound pruning significantly hampers the efficiency despite its great contribution in reducing the search effort. Table 6. Computation times (in cpu seconds on a sun Sparc 10) for four implementations of tabu search on benchmark problems and comparative ratios. problem non-cp passive cp ratio light ratio light ratio passive plain plain rc204 30.4 2615.4 86.0 1412.7 46.5 160.1 5.3 rc206 48.6 1458.7 30.0 282.0 5.8 108.8 2.2 rc207 52.2 2595.3 49.7 524.4 10.0 103.8 2.0 rc208 60.9 1993.2 32.7 964.3 15.8 134.7 2.2 average: 40.5 16.2 4.0 Runtime statistics reveal that the amount of backtracking information maintained for the branch-and-bound is about 10 times the amount for plain tree search. The sharp drop in efficiency is apparently due to the repeated storage and retrieval of all that information. It was not a major factor in the "full" version because that extra amount of backtracking information was added to an already significant amount associated to the full constraint model. Tabu pruning generally increases the computation time and is not attractive for this light-weight version. This setback for lower-bound pruning prompted us to implement the explicit computation of lower bounds at tree nodes instead of relying on constraint propagation to a cost variable. We subsequently refer to this implementation as lower-bound pruning ? . The improvement in speed was astonishing, as we will see in the next table. Table 6 compares the computation time of four tabu search implementations. Column two represents the specialized version, which is faster as expected. Next comes the computation time for the passive cp version followed by a comparison with the specialized version (passive cp / non-cp). That ratio reflects the sole cost of constraint propagation, here averaging 40:5. The other two groups of two columns represent the light-plain and light-plain with lower-bound pruning ? versions and the ratios are all taken with respect to the non-cp version. The ratios for the light-plain combination indicate how profitable it can be to proceed by tree search instead of systematically considering every neighbor as in passive cp: tempering the cost of constraint propagation with the elimination of subtrees of neighbors yields the average ratio of 16:2. The light-plain combination is thus 2:5 times faster than passive cp. It even outperforms the latter on rc208 for which there was no elimination of neighbors to speak of (see table 4): the triggering of interface constraints at various levels of the tree corresponds to factoring out common constraint propagation fragments among subsets of neighbors. The lower-bound pruning ? implementation now brings out the savings observed in search effort. It is on average four times faster than plain tree search and ten times faster than passive cp. It also comes within a factor of four of the specialized version. Figure 5. Performance ratio for larger neighborhood search trees.1357 CPU time ratio nb of cities 3-interchange 4-interchange These experimental results clearly indicate the superiority of our approach over a more passive use of constraint propagation in local search. Unfortunately, the cost of constraint propagation and built-in backtracking still prevented us from being competitive with the specialized version on these instances. We therefore decided to investigate larger neighborhood search trees in order to increase the natural payoff of branch-and-bound pruning. This was done in two ways: increasing the size of the instances (number of cities), which affects the breadth of the neighborhood search tree; and considering 4-interchanges instead of 3-interchanges, which increases the depth of the search tree by one. As our previous instances fell in the range 20-40 cities, we considered 20-, 40-, 60-, and 80-city instances taken from [4]. These additional results are summarized in figure 5 which plots the ratio of the execution time for our final approach (light lower-bound pruning ? ) over that for the specialized version, for both 3- and 4- interchange neighborhood structures and for increasing problem size. Each point is the average of over a dozen instances of varying time window width. As expected, our relative performance improves with the growth of the search tree, especially between 20 and 40 cities. (Note that these results are consistent with the average value and variability of the ratios previously reported in table 6 on different instances.) For 3-interchanges, the average ratio falls to 1.5 at 80 cities. Note in particular that for 4-interchanges it drops below one starting at 60 cities: we thus perform better than the specialized version from that point on. Even though a 4-interchange is seldom used in practice, a neighborhood search tree of depth four would not be uncommon: we have already mentioned the cross exchange for routing; for our physician scheduling application in section 3, changing the value of two slots at a time instead of one, a very plausible neighborhood structure, would also correspond to a depth of four. 4.5. Applying the Framework to the GENIUS heuristic The combination of a well-known tsp heuristic, genius [5], with the constraint programming routing model sketched at the beginning of the section is described in [20]. Though this heuristic, composed of a tour construction phase and a post-optimization phase, is not strictly speaking a local search algorithm, each of its two phases involves the exploration of a neighborhood very similar to what was used earlier. It was therefore a good candidate for our framework. The resulting algorithm, genius-cp, was able to solve variations on the tsp (with single or multiple time windows, precedence constraints, sequencing constraints, pick-up and delivery problems) by simply adding the appropriate constraints to the master model without any modification to the algorithmic part (genius). In contrast, a subsequent version of the original heuristic to handle single time win- dows, genius-tw [6], had required major changes: the addition of backtracking capabilities, special-purpose code to verify feasibility and adjustments of algorithmic control to reflect the temporal flavor of the problem. A comparison of genius-cp and genius-tw on benchmark problems with single time windows revealed that the flexibility of our approach cost us about a factor of four in speed 5 , but the former more often found the best solution whenever their solutions differed. As for the other variations on the tsp, no specialized heuristic based on genius existed and so we compared with a close relative on the other side of the family, an exact cp algorithm using the same routing model. It confirmed that genius-cp was more robust and several orders of magnitude faster. The heuristic solutions obtained were also often optimal and generally very good. 5. Addressing Flexibility The previous section provided evidence of the search economy achieved by our framework; the report on genius-cp also displayed some of its flexibility. We would now like to briefly return to that latter advantage. Again, the root of that flexibility is the separation of the constraints of a problem (the declarative part) from the heuristic search strategy (the algorithmic part). Because of this, a particular heuristic search strategy can serve in different contexts without any modification of its implementation. Such an approach presents clear advantages from a software engineering point of view: since one only needs to adapt the model to the problem, it makes it much easier and faster to tailor a piece of software to real-life situations while reusing the algorithmic component as is. It also has an impact on the maintainability of the code, especially if additional constraints suddenly arise. In such a situation, a non-cp approach would either have to modify the way valid neighbors are generated or add feasibility tests, to reflect the additional constraints. Both solutions alter the algorithmic component. In contrast, a cp approach will add the required constraints to the declarative component (the model), which often simply amounts to expressing them using variables already present in the model and standard constraints. For example, we give below some constraints and their expression in the context of section 4: sequencing constraints: city j must be visited last: must immediately follow city i: S precedence constraints: city j must be visited after city i: multiple time windows: city i may be visited from 9am to 5pm but not between noon and 1pm: The adaptation of the program would thus consist of one additional line of code per new constraint, leaving the rest of it unchanged. Conclusion We have proposed a novel framework for local search algorithms in constraint pro- gramming. By maintaining branch-and-bound search at their core, we believe that a cleaner and more natural integration is achieved. In addition, the familiar pruning which can take place yields substantial savings on the number of neighbors that actually need to be considered. A major advantage of bringing constraint programming into local search methods is the flexibility and ease with which additional constraints can be handled. Two examples of this framework were given: physician scheduling and single-vehicle routing. Computational experiments on the latter provided an estimate of the price of constraint programming in this setting: it showed that such a framework outperforms the simpler, more passive use of constraint propagation in local search and that it is competitive with a specialized, non-cp local search algorithm. It also stressed that a light-weight master model may be preferable and that the standard branch-and-bound approach comes with a computational cost which may outweigh its benefit: in our case, the explicit computation of lower bounds proved much more efficient. A tabu algorithm which is tailored and optimized for a particular context is expected to perform better. Our experimentation suggests that, given a large enough neighborhood (and not inordinately large), an implementation of our framework may run faster. The initial cost of using constraint propagation can turn up a profit when combined with a branch-and-bound exploration of the neighborhood search space. The small percentage of feasible neighbors which we encountered, even when the proportion of truly constraining time windows was very small, indicates that it is worth investing in the elimination of infeasible subsets of the neighborhood. A more ambitious study of different and especially larger neighborhoods as well as different combinatorial problems could serve to strengthen the previous claim and also to identify possible settings in which the "full" version or tabu pruning may be advantageous. Acknowledgments Financial support for this research was provided by the Natural Sciences and Engineering Research Council of Canada (NSERC). Notes 1. Though incomplete search methods have been successfully introduced in the constraint satisfaction community. 2. The neighborhood for the "shuffle" moves of [2] is explored through a tree search guided by constraint propagation. This lies one step closer to our approach. 3. A conditional constraint p ) q enforces the consequent q once the antecedent p is verified. We use the syntax "Var = value" for an antecedent to mean that it is verified when variable Var is fixed to some value. 4. Granted, such a tree remains relatively shallow but potentially very bushy. 5. This is for a version of genius-cp equipped with lower-bound pruning ? , an improvement over what is described in [20] --R Local Search in Combinatorial Optimization. Disjunctive Scheduling with Task Intervals. A Constraint Logic Programming Approach to the Vehicle-Fleet Scheduling Problem New Insertion and Postoptimization Procedures for the Traveling Salesman Problem. A Generalized Insertion Heuristic for the Traveling Salesman Problem with Time Windows. Heuristic for Integer Programming Using Surrogate Constraints. ' E. Taillard, and D. de Werra. A User's Guide to Tabu Search. Adaptation in Natural and Artificial Systems. Computer Solutions of the Traveling Salesman Problem. An Effective Heuristic Algorithm for the Traveling Salesman Problem. Solving Large-Scale Constraint Satisfaction and Scheduling Problems Using a Heuristic Repair Method Traveling salesman-type combinatorial problems and their relation to the logistics of regional blood banking Metastrategy simulated annealing and tabu search algorithms for the vehicle routing problem. A View of Local Search in Constraint Programming. An Exact Constraint Logic Programming Algorithm for the Traveling Salesman Problem with Time Windows. GENIUS-CP: a Generic Vehicle Routing Algorithm. An Exchange Heuristic for Routing Problems with Time Windows. Modern Heuristic Techniques for Combinatorial Problems. Local Search in Routing Problems with Time Windows. A New Method for Solving Hard Satisfiability Problems. Algorithms for the Vehicle Routing and Scheduling Problem with Time Window Constraints. Practical Applications of Constraint Programming. --TR --CTR Michel Gendreau, Constraint Programming and Operations Research: Comments from an Operations Researcher, Journal of Heuristics, v.8 n.1, p.19-24, January 2002 Louis-Martin Rousseau , Michel Gendreau , Gilles Pesant, Using Constraint-Based Operators to Solve the Vehicle Routing Problem with Time Windows, Journal of Heuristics, v.8 n.1, p.43-58, January 2002 Simon de Givry , Laurent Jeannin, A unified framework for partial and hybrid search methods in constraint programming, Computers and Operations Research, v.33 n.10, p.2805-2833, October 2006 Samir Loudni , Patrice Boizumault , Philippe David, On-line resources allocation for ATM networks with rerouting, Computers and Operations Research, v.33 n.10, p.2891-2917, October 2006 Christian Blum , Andrea Roli, Metaheuristics in combinatorial optimization: Overview and conceptual comparison, ACM Computing Surveys (CSUR), v.35 n.3, p.268-308, September Lucas Bordeaux , Youssef Hamadi , Lintao Zhang, Propositional Satisfiability and Constraint Programming: A comparative survey, ACM Computing Surveys (CSUR), v.38 n.4, p.12-es, 2006
personnel scheduling problem;interface constraints;constraint programming;traveling salesman problem with time windows;neighborhood model;local search;tabu search
595043
Solving Vehicle Routing Problems Using Constraint Programming and Metaheuristics.
Constraint Programming typically uses the technique of depth-first branch and bound as the method of solving optimization problems. Although this method can give the optimal solution, for large problems, the time needed to find the optimal can be prohibitive. This paper introduces a method for using local search techniques within a Constraint Programming framework, and applies this technique to vehicle routing problems. We introduce a Constraint Programming model for vehicle routing, and a system for integrating Constraint Programming and local search techniques. We then describe how the method can be accelerated by handling core constraints using fast local checks, while other more complex constraints are left to the constraint propagation system. We have coupled our local search method with a meta-heuristic to avoid the search being trapped in local minima. Several meta-heuristics are investigated ranging from a simple Tabu Search method to Guided Local Search. An empirical study over benchmark problems shows the relative merits of these techniques. Investigations indicate that the specific long-term memory technique used by Guided Local Search can be used as a diversification method for Tabu Search, resulting in significant benefit. Several new best solutions on the Solomon problems are found in relatively few iterations of our algorithm.
Introduction This paper describes the use of iterative improvement techniques and meta-heuristics as the search engine within a Constraint Programming framework (ILOG Solver) and its application to the Vehicle Routing Problem (VRP). The VRP is a practical problem that has been studied widely in the literature [10]. The problem is usually expressed as follows: given a set of customers requiring a visit, and a fleet of vehicles based at a depot that can perform the visits, construct a set of routes for the vehicles that minimises the costs of operation. The objective function is usually expressed as costs related to the number of vehicles and to distance travelled. Constraints include various capacity constraints on weight, volume, length, etc., time constraints on when the customer will accept a visit, and the total length of routes. However, in practical problems we may have many different kinds of additional constraints including legislative restrictions, established work practices and customer preferences, and a complex objective function reflecting complicated pay provisions. Constraint Programming (CP) is a paradigm for representing and solving a wide variety of problems. Problems are expressed in terms of variables, domains for those variables and constraints between the variables. The problems are then solved using complete search techniques such as depth-first search (for satisfaction) and branch and bound (for optimi- The richness of the language used to express problems in CP makes it an ideal candidate for VRPs. However, the VRP is an NP-hard problem [3], and for problems of practical size computing optimal solutions can be too time consuming. In this paper we investigate how iterative improvement techniques can be used in a CP framework to replace complete search. As simple local search procedures get trapped in local minima, the improvement techniques are used in conjunction with a meta-heuristic. We examine two meta-heuristics: a simple Tabu Search approach [5, 6], and Guided Local Search [21, 23]. The paper is organised as follows. In section 2 we describe the Constraint Programming paradigm, how it can be applied to the VRP, and how local search can be performed within this framework. In section 3 the solution technique using meta-heuristics is discussed. In section 4, computational tests on the two meta-heuristics are reported. Investigations reveal that the performance of simple Tabu Search can be significantly increased by incorporating specific long-term memory features of Guided Local Search. 2. Constraint-Based Modelling of VRPs Part of the attraction of Constraint Programming is the richness of the language for specifying the model. ILOG Solver [14] allows real and integer variables to be constrained by simple bounds and linear or non-linear constraints. Similarly the objective function can be complex. Very general expressions can be used as constraints. In addition to expressions involving the usual arithmetic and logical operators, complex symbolic constraints can be used to describe a problem. Examples include the all-different constraint, which is used to express (and efficiently propagate) the fact that each variable in a set must take a different value, and the ability to constrain the index of an array (by using a constrained variable). enhances search using constraint propagation. If bounds or constraints on a variable can be inferred, or are tentatively set, these changes are "propagated" through all the constraints to reduce the domains of constrained variables. ILOG Solver uses depth-first search with constraint propagation. At each node in the search tree, the propagation mechanism removes values from the domains of constrained variables that are inconsistent with other variables. If the propagation mechanism removes all values from any variable, then there cannot be a solution in this subtree and the search backtracks making a different decision at the backtrack point. Backtracking is chrono- logical: decisions can only be undone in the opposite order to which they were made. Moreover, the domains of constrained variables can only be reduced as search progresses down the tree. The domains of all constrained variables can be restored by backtracking to a previous node, but general enlargement (or relaxation) of domains is not supported. This has important implications for the way in which vehicle routing problems are solved in this paper, as discussed in section 2.4. 2.1. A Constraint Programming Model for the VRP A decision variable R i is associated with each customer visit i, representing the next visit performed by the same vehicle. For each vehicle k, there are additional visits S k marking the start of the route, and E k marking the end of the route. To read off the itinerary for a vehicle k, start at S k and follow the "next" pointers through to E k . The set of customer visits will be referred to as N , starting visits as S, and ending visits as E. visits. Regarding R as a function, R maps N [ S onto To model the VRP and other routing problems, it is desirable to have a special constraint type for dealing with constraints along paths. We use a path constraint [1, 8] that allows real-valued quantities accumulated along a route to be bounded. The quantities accumulated can relate to time, load, or other variables. The constraints are of the form Thus, if visit j immediately follows visit i, quantity Q is accumulated. For example, setting equal to the travel time between i and j means Q j would be the arrival time at j. Setting q ij to be the demand at i would accumulate load on the vehicle. This specification allows other real-world constraints to be expressed succinctly. In each case a fixed point (boundary condition) must be supplied, for example in the case of load constraints. A single constraint of this type which contains an upper bound also has the effect of eliminating subtours (routes that do not include a S k 2.2. Core Constraints The core constraints in vehicle routing problems are related to time and capacity. Time is restricted both by the working day and by time windows at customers. Capacity may be restricted in terms of weight, volume, number of pallet places, etc. A path constraint can be used to propagate the start time of the customer service along each vehicle route. In this case, q ij is the service time at i, plus the travel time from i to j. Single or multiple service time windows could be set by further constraining the start of service variable for each visit. The start and end of the working day for each vehicle can be represented by imposing time windows on visits in S and E. A method of modelling capacity constraints on vehicles is to propagate (via the path constraint) the free space in the vehicle along each vehicle route. To model F i as the free space on arrival at visit i, q ij is the change in free space associated with visit i. To indicate that all vehicles must not be overfilled at any time, constraints of the form F are imposed. For vehicle k with capacity c k , the fixed point F can be set. Simultaneous pickups and deliveries can be modelled in a similar way, but this is not discussed here. 2.3. Side Constraints One of the significant benefits of CP techniques is that side constraints can be incorporated into the model with comparative ease. For example, some visits may require special equip- ment, or some vehicles may be simply too large to enter the premises. Therefore, we need to be able to disallow a visit being made by particular vehicles. We model this by propagating a vehicle tag - k for each visit k along each route, using a constraint similar to the path constraint for which R S. Thus customer visits served by vehicle k will have tag k. Vehicle k is then forbidden from making visit i by imposing a constraint of the form - i 6= k. A conjunction of such expressions can be used to exclude more than one vehicle. The requirement that two visits i and j must be made by the same (different) vehicles can also be specified using tags: - Implementing a length capacity for vehicles provides one final example that demonstrates a powerful use of vehicle tags. A length restriction is different from a capacity constraint because it does not accumulate over a route. If visit i involves goods of length l, and L k is the length of vehicle k, then the constraint L - i l defines the length constraint. 2.4. Search Strategy Solutions to CP problems are usually found using complete methods such as depth-first search and branch and bound. However, for routing problems of practical size, complete search methods cannot produce solutions in a short and reliable time period. By contrast, iterative improvement methods have proved very successful in this regard. Iterative improvement methods operate by changing small parts of the solution, for instance moving a visit from one route to another. This type of operation involves retracting previous decisions and making new ones. In contrast to depth-first search, retraction can be done in any order, not simply in the opposite order the decisions were made. In general the overhead of implementing this type of non-chronological retraction mechanism in CP frameworks is very high. To overcome this problem, the CP system is only used to check the validity of solutions and determine the values of constrained variables, not to search for solutions. The search is performed by an iterative improvement procedure. When the procedure needs to check the validity of a potential solution, it is handed to the CP system. As a part of checking, constraint propagation using all constraints still takes place. This adds value to the constraint check, as the iterative improvement method can then take advantage of the reduced domains to speed up search by performing fast legality checks (see later in this section). The implementation relies on two representations of the solution. There is a passive rep- resentation, which is the template against which new solutions are constructed, and which holds the "current solution" (not necessarily the best solution). The second is an active representation that holds the constrained variables, and within which constraint propagation takes place. Variable states in the active representation are "checkpointed" (current domains are saved) before any changes are made, so that domains can be restored later. When the CP system performs propagation and validity checks it instantiates the set of decision variables R using the itinerary in the passive representation. The constraints then propagate and constrained variables (such as time and capacity variables) have their domains reduced. If any variable has no legal values, the solution is illegal. Improvement heuristics generally only modify a very small part of a routing plan. There- fore, testing the complete solution in the above manner can be inefficient. We have tried to avoid this inefficiency in two ways: by reducing the amount of work carried out by the CP system to perform each check, and bypassing the CP checks altogether. The first method applies if vehicle routes are independent, i.e. have no constraints acting between visits in them. This is the case in many routing problems. For these problems, only the routes involved in the modification of the plan need be instantiated in the active representation. Normally, this means just one or two routes will be instantiated. The second method of increasing efficiency is to perform tests on the core constraints in situ (using the domains of constrained variables such as capacity and time, preserved from the last propagation, as described earlier). Propagation precisely computes the remaining slack for each cumulative (Q) variable. This slack is reduced as much as possible using all constraints (i.e. all side constraints) in the problem. These in situ checks can all be done in constant time using a method derived from Savelsbergh [17]. Only if the constant time check is successful need the route set be passed to the CP system for full checking. The acceleration techniques just described can increase performance of a VRP search engine by a factor of twenty or more, such that up to 200,000 neighbourhood moves can be examined per second on a high-end PC [1]. 3. Solution Methods Local search operates by considering a set of possible moves. The moves considered in this paper are described in the following section. Local search repeatedly chooses a move which reduces the objective value, and then implements this move, until no further legal cost-reducing moves are available. Two alternatives in choosing a move to perform are "first accept": choose the first cost-reducing move found, or "best accept": choose the legal move that reduces the objective by the largest amount. The best accept method is used throughout this paper. Local search suffers from the problem that it will get caught in the first local minimum that it encounters. In order to escape a local minimum, a controlled method of accepting an up-hill move is required. Meta-heuristics provide the framework for accepting such moves. In this paper, we examine two meta-heuristics, tabu search [5, 6], and guided local search [21, 23]. These are described in more detail later. 3.1. Move Operators We have considered four move operators (see figure 1): a) 2-opt reverses a section of a route by deleting two arcs, and replacing them with two others to reform the route. b) Relocate moves a visit from its position in one route to another position in either the same or a different route. c) Exchange swaps two visits from either the same or different route. d) Cross swaps the end portions of two route. 2-opt is described in [11], the others in [16]. A A (a) The 2-opt heuristic A A (b) The relocate heuristic A A (c) The exchange heuristic A A (d) The cross heuristic Figure 1. Four move operators 3.2. Tabu Search Tabu Search is a meta-heuristic that has been widely applied to many combinatorial optimization problems ([7] gives numerous applications) with great success. In particular, Tabu Search been extensively applied to vehicle routing problems, with good results, for example [4, 12, 13, 15, 19]. Tabu Search is sufficiently well known that we do not give a general description, but descriptions can be found in [5, 6, 7]. Table 1. Thresholds for different move operators Operator arcs threshold cross 4 3 exchange 8 6 relocate 6 5 Our implementation of Tabu Search is a simple one involving tabu moves. It does not use long-term memory for explicit intensification and diversification phases. For the VRP, moves can be characterised by two sets of directed arcs: arcs added and arcs removed. For example, the relocate operator (figure 1(b)) removes three arcs and adds three new arcs. We use two tabu lists of fixed and equal size: one for storing the most recently added arcs, and the other for storing the most recently removed arcs. We use the following procedures: performs a local search, starting at solution S , and returns a new solution. We use a "best accept" algorithm: among all the legal moves defined by the move operators (section 3.1), choose the move that decreases the objective by the greatest amount. Repeat until there is no move that strictly decreases the objective. ffl RankMoves(S) returns a list of all legal moves from S , ordered according to ascending cost difference. That is, moves reducing the objective by the greatest amount will be near the beginning of the list. ffl IsNotTabu(m) is true if the move m is not tabu as dictated by the tabu lists. In our implementation, the tabu status of a move m is decided as follows: The arcs that are removed by m and occur in the "recently added" tabu list are counted. Likewise, the arcs that are added by m and occur in the "recently deleted" tabu list are counted. The sum of these is then compared to a threshold dependent on the move type (see table 1). If the threshold has been met or passed, the move is tabu, and is not tabu otherwise. Note that in table 1, the number of arcs modified in 2-opt is variable because the inverted part of the route may be of any length. In all experiments the length of both lists was 100 arcs. An aspiration criterion was used which could change the tabu status of a move. If the move would result in a solution better than the current best, then it was accepted even if tabu. ffl Perform(m) performs the move. ffl InsertInTabuList(m) inserts the added and deleted arcs into the appropriate tabu lists. The Tabu Search algorithm is described by Figure 2. O(S) is the objective function. The function head(l) returns the first element in the list l, and removes it from l. InitialSolu- tion() and StoppingCondition() are defined in section 4. S := InitialSolution() // S is current solution S := LocalSearch (S) // Make sure we start at a local min best solution while not StoppingCondition() do moves := RankMoves(S) moved := false while not moved do if IsNotTabu(m) then Perform(m) moved := true return S Figure 2. Tabu Search algorithm 3.3. Guided Local Search Guided Local Search (GLS) [21, 23] is a meta-heuristic based on penalties. The method works by adding a penalty factor to the objective function, based on the experience the search has gained. Roughly speaking, search is penalised if it strays too close to previously visited local minima, allowing escape from local minima. GLS has proved to be an effective meta-heuristic for a range of combinatorial optimization problems, for example see [20, 22, 23, 24]. In what follows, we describe the general algorithm, and then its application to the vehicle routing problem. 3.3.1. The Guided Local Search Algorithm Guided Local Search moves out of a local minimum by penalizing particular solution features that it considers should not occur in a good solution. It defines a modified objective function, augmented with a set of penalty terms on these features. The usual local search method is then invoked to improve the augmented objective function. The cycle of local search and penalty term update can be repeated as often as required. GLS requires the following components: ffl A set of features F . Let the indicator function for feature i 2 F be f i . f i the feature i is in solution S, 0 otherwise. ffl A cost vector c. c i is the "cost" of feature i. ffl A penalty factor -. GLS tracks penalties applied via a penalty vector p. p i is the (integer) number of times feature i has been penalised so far. Assuming O(S) is the the original objective function for the problem, GLS defines an augmented objective function: O and requires a local search procedure that minimises it. The user must provide procedure that performs a local search, starting at solution S , and returns a new solution. The improvement is carried out with respect to the augmented objective O 0 . GLS provides a function called ChoosePenaltyFeatures(S, p) which takes a solution and the current spread of penalties, and returns the set of features to be penalised. GLS penalises the most costly features in the current solution, weighted by the number of times the features have already been penalised. This has the effect of concentrating penalties on "bad" features, while tending to accept bad features if they keep appearing in local minima (i.e. if they seem inevitable). GLS chooses all features i 2 S for which c i =(p i +1) is largest amongst the features in S. Unless the cost values in c are equivalent or low multiples of each other, usually only one feature is chosen to penalise. Assuming the usual functions of InitialSolution() and StoppingCondition() exist, the GLS algorithm is described in figure 3. S := InitialSolution() // S is current solution S := LocalSearch (S) // Make sure we start at a local min best solution while not StoppingCondition() do f := ChoosePenaltyFeatures(S, p) forall g in f do return S Figure 3. GLS algorithm Parallels can be drawn between GLS and the use of frequency memory in Tabu Search [5]. Frequency memory records the number of times a particular feature has appeared in previous solutions. It is then used in a diversification phase to penalise the occurrence of features that have appeared often. However there are two main differences between the ideas. GLS uses its "memory" throughout the search, rather than in a separate phase. This eliminates the need for "tuning parameters" for the use and duration of a diversification phase. Second, the selection of features to penalise in GLS is based on the cost of the features, as well as the frequency of their occurrence. By penalising features that have appeared often, Tabu Search may attempt to drive out "good" features. GLS uses domain knowledge, in the form of the feature costing, to try to avoid this. In addition, the sense in which the frequency is used is different. In GLS the more often a feature appears and is penalised, the less likely it is to be penalised further. The is not the case in Tabu Search. 3.3.2. GLS for Vehicle Routing We define our implementation of GLS for the VRP with reference to the description above. ffl Feature set F : In [21, 23] for the Travelling Salesman Problem, arcs were chosen as the feature to penalise. We use the same method here, but use directed arcs, since for routing problems involving time, direction is important. ffl Feature costing: We assume that the cost c a of directed arc feature a is the length of the arc d a . ffl Penalty factor -: We have found that values of 0.1 to 0.3 work well, and have used 0.2 for the results in section 4. ffl We again use a "best accept" algorithm: among all the legal moves defined by the operators in section 3.1, choose the move that decreases the objective by the greatest amount. Stop when there is no move that strictly decreases the objective. To implement the objective O 0 , the local search is carried out using a modified distance matrix. In the modified matrix, each directed arc a has a penalised distance d a + -p a d a . The matrix is updated each time an element of p is changed. InitialSolution() and StoppingCondition() are defined in section 4. 4. Computational Experiments Computational experiments were performed using both the Tabu Search and Guided Local Search meta-heuristics. For both heuristics the StoppingCondition() function used was 1000 forward moves in the search space (these 1000 forward moves took under 10 minutes of CPU time). The InitialSolution() function used was that produced by the Savings method of [2], followed by search to a local minimum. The objective function was the total distance travelled by all vehicles. There was no incentive to reduce the number of vehicles. This means that although the results from different methods are comparable with each other, they are not directly comparable with previously reported best values which emphasise reducing the number of vehicles. After the 1000 moves, a final phase is entered by both Tabu Search and GLS where restrictions are switched off and a search is performed to a local minimum. For GLS, this corresponds to zeroing all penalties, and finding a local minimum. For Tabu Search, we search to a local minimum using no tabu list. This final improvement can be particularly useful in the case of GLS which may be prevented from reaching a true local minimum during search due to the penalty method employed. The data for comparison are as used by Solomon [18]. These problems are in 6 classes: C1, C2, R1, R2, RC1, RC2. Each class contains between 8 and 12 individual problem instances. All problems have 100 customers, a central depot, capacity constraints, time windows on the time of delivery, and a total route time constraint. The C problems have customers located in clusters. In the R problems the customers are at random positions. The RC problems contain a mix of both random and clustered customers. The C1, and RC1 problems sets have a short scheduling horizon, and require 9 to 19 vehicles. Problem classes C2, and RC2 are more representative of "long-haul" delivery with longer scheduling horizons and require fewer (2 - vehicles. Both time and distance are given by the Euclidean distance between points. 4.1. Performance of the Algorithms Figure 4 shows the rate of improvement of Tabu Search and GLS over the six problem classes. The graphs were produced as follows: at each forward move, the lowest travel distance found so far for each meta-heuristic on each problem was recorded. This describes a series of monotonically decreasing curves for each meta-heuristic applied to each problem. Then, all curves corresponding to a single problem class and single meta-heuristic were av- eraged, resulting in 12 curves. Finally, these curves were normalised against a baseline, which is the best average distance value over the class attained by the best meta-heuristic for the class. Thus, in the graphs, one of the meta-heuristics always attains the 0% point. In five of the six classes, GLS outperforms simple Tabu Search by a margin varying from 1.2% to 2.7%. In the remaining class, Tabu Search performs 0.5% better than GLS. 4.2. Inside the Search In an attempt to understand the better performance of GLS over simple Tabu Search, we examined the search process by observing the change in objective function. From this, we hope to see when search stagnates, and how aggressively the search attempts to move out of local minima. Figure 5 shows the best objective value found so far O(S ), and the current objective value O(S), as GLS and Tabu Search proceed through 1000 iterations on problem RC107. Problem RC107 is a typical problem, in that the starting solution can be improved. Problem RC107 demonstrates the gradual fall in the cost of S . It also shows how GLS diversifies the search after moving to a new best solution. This can be seen in the rise in the cost of S following a fall in the cost of S . After a period of diversification, GLS moves out of the current local minimum, and finds a path to a new local minimum. This trace also shows how the diversification period tends to become longer during search, suggesting that the search must roam farther from the previous best solution in order to reduce the best cost. In comparison, this simple Tabu Search appears to roam within a relatively small distance from S , and as such cannot escape from the local minimum it is trapped it. Figure 6 shows behaviour for problem C106. The initial solution to C106 is optimal, and search is futile. We conjecture that GLS investigated three distinct local minima, or clusters of local minima. It appears that the search repeatedly returns to the same or a nearby local minimum before escaping to a new part of the solution space with a new, deep local minimum. Simple Tabu Search appears to exhibit a harmonic behaviour, swinging away and then back to S with a frequency apparently related to the tabu tenure. Percent above Min Iterations GLS Tabu Search Problem class C11357 Percent above Min Iterations GLS Tabu Search Problem class C21357 Percent above Min Iterations GLS Tabu Search Problem class R11357 Percent above Min Iterations GLS Tabu Search Problem class R21357 Percent above Min Iterations GLS Tabu Search Problem class RC11357 Percent above Min Iterations GLS Tabu Search Problem class RC2 Figure 4. Rate of improvement of Tabu Search and GLS 4.3. Including a GLS-type long-term memory into Tabu Search The drawback of the simple Tabu Search approach presented is that it is only based on short-term memory. This limits its capability to explore various regions of the search Distance Iterations Current GLS on problem RC10712601300134013800 200 400 600 800 1000 Distance Iterations Current Tabu Search on problem RC107 Figure 5. Inside Search, problem RC107. On the left GLS, and on the right simple Tabu Search. We see that as search proceeds GLS tends to roam progressively further from S whereas Tabu Search appears to wander within a small distance from S .85095010500 200 400 600 800 1000 Distance Iterations Current GLS on problem C10685095010500 200 400 600 800 1000 Distance Iterations Current Tabu Search on problem C106 Figure 6. Inside Search, problem C106. The starting solution is optimal, and we see on the left, GLS again roams progesssively further from S and appears to find 2 or 3 new local minima. On the right, a harmonic appears within simple Tabu Search, the frequency possibly a function of the tabu tenure. space. To better diversify the search, various procedures have already been proposed in the literature including frequency memory as discussed in section 3.3. (For an alternative method that addresses the specific case of the VRP, see [15].) GLS, because of its simplicity and the excellent results it gives (see [9] for a comparison of GLS with best reported methods) seems to offer an alternative method of implementing long-term memory within Tabu Search with diversification. It has two advantages over other methods: incorporation of domain knowledge through the costing of arcs, and constant diversification without the need for a separate phase. The resulting Tabu Search still operates by selecting a non-tabu move from the available neighbourhood. However, after each move a GLS-type procedure is called to update the weights in the cost matrix. This method is called Guided Tabu Search (GTS) below. For the combined method, the length of the tabu lists remained at 100 arcs. However, very slightly better results were achieved by reducing the penalty factor from 0.2 to 0.15. The performance of GTS relative to GLS and simple Tabu Search is shown in figure 7. Guided Tabu Search outperforms the other two methods on the "long haul" problems RC2), performing on average about 1.7% better. For the remaining problem classes, GTS is not quite as convincing, but competitive. 5. Conclusion We have described a method for combining Constraint Programming with iterative improvement techniques. The combination makes full use of the propagation facilities of the Constraint Programming framework, but does not use a backtracking procedure. The method was accelerated using a system whereby certain core constraints are maintained within the search engine, while other side constraints are handled by the Constraint Programming system. This method was applied to the vehicle routing problem. Search engines using two different meta-heuristics were examined: a simple Tabu Search technique, and Guided Local Search. These techniques were compared over a set of standard benchmark problems. Guided Local Search was found to be superior to the simple Tabu Search method. Examination of the behaviour of both algorithms suggested that adding features of Guided Local Search to Tabu Search would improve its performance. A combined method, Guided Tabu Search, was able to outperform the other methods on most problems. In the future, we wish to examine other diversification methods in Tabu Search, and compare the results to those obtained with Guided Tabu Search. Intensification strategies in Guided Local and Guided Tabu Search are also of interest. Further testing on real-world problems is in progress. We hope this will give us further insights into ways of improving the methods presented. Acknowledgments The production of this paper was supported by the GreenTrip project, a research and development undertaking partially funded by the ESPRIT Programme of the Commission of the European Union as project number 20603. The partners in this project are Pirelli (I), Tollpost-Globe (N), and University of Strathclyde (UK). Percent above Min Iterations GLS Tabu Search Problem class C12610 Percent above Min Iterations GLS Tabu Search Problem class C22610 Percent above Min Iterations GLS Tabu Search Problem class R12610 Percent above Min Iterations GLS Tabu Search Problem class R22610 Percent above Min Iterations GLS Tabu Search Problem class RC12610 Percent above Min Iterations GLS Tabu Search Problem class RC2 Figure 7. Rate of improvement of basic Tabu Search, GLS, and Tabu Search with GLS long-term memory --R Scheduling of vehicles from a central depot to a number of delivery points. Computers and Intractability: A Guide to the Theory of NP-Completeness A tabu search heuristic for the vehicle routing problem. search search search. Guided local search for the vehicle routing problem. Routing problems: A bibliography. Computer solutions of the traveling salesman problem. Metastrategy simulated annealing and tabu search algorithms for the vehicle routing problem. A tabu search heuristic for the vehicle routing problem with time windows. Beyond the glass-box: Constraints as objects Probabilitic diversification and intensification in local search for vehicle routing. Computer aided routing. The vehicle routing problem with time windows: Minimizing route duration. Algorithms for the vehicle routing and scheduling problem with time window constraints. Hybrid genetic algorithm Fast local search and guided local search and their application to British Telecom's workforce scheduling problem. Guided Local Search for Combinatorial Problems. Function optimization using guided local search. Guided local search. Partial constraint satisfaction problems and guided local search. --TR --CTR Vincent Tam , K. T. Ma, Combining Meta-Heuristics to Effectively Solvethe Vehicle Routing Problems with Time Windows, Artificial Intelligence Review, v.21 n.2, p.87-112, April 2004 Andrew Lim , Brian Rodrigues , Lei Song, Manpower scheduling with time windows, Proceedings of the ACM symposium on Applied computing, March 09-12, 2003, Melbourne, Florida C. Voudouris, Guided Local Search an Illustrative Example in Function Optimisation, BT Technology Journal, v.16 n.3, p.46-50, July 1998 Pascal Hentenryck , Laurent Michel, Control Abstractions for Local Search, Constraints, v.10 n.2, p.137-157, April 2005 Alain Chabrier, Vehicle routing problem with elementary shortest path based column generation, Computers and Operations Research, v.33 n.10, p.2972-2990, October 2006 Laurent Michel , Pascal Van Hentenryck, Comet in context, Proceedings of the Paris C. Kanellakis memorial workshop on Principles of computing & knowledge: Paris C. Kanellakis memorial workshop on the occasion of his 50th birthday, p.95-107, June 08-08, 2003, San Diego, California, USA Filippo Focacci , Andrea Lodi , Michela Milano, Mathematical Programming Techniques in Constraint Programming: A Short Overview, Journal of Heuristics, v.8 n.1, p.7-17, January 2002 J. Christopher Beck , Patrick Prosser , Evgeny Selensky, A case study of mutual routing-scheduling reformulation, Journal of Scheduling, v.9 n.5, p.469-491, October 2006 Christian Blum , Andrea Roli, Metaheuristics in combinatorial optimization: Overview and conceptual comparison, ACM Computing Surveys (CSUR), v.35 n.3, p.268-308, September Haibing Li , Andrew Lim, Local search with annealing-like restarts to solve the vehicle routing problem with time windows, Proceedings of the 2002 ACM symposium on Applied computing, March 11-14, 2002, Madrid, Spain Beatrice Ombuki , Brian J. Ross , Franklin Hanshar, Multi-Objective Genetic Algorithms for Vehicle Routing Problem with Time Windows, Applied Intelligence, v.24 n.1, p.17-30, February 2006 K. C. Tan , Y. H. Chew , L. H. Lee, A Hybrid Multiobjective Evolutionary Algorithm for Solving Vehicle Routing Problem with Time Windows, Computational Optimization and Applications, v.34 n.1, p.115-151, May 2006 Alexandre Le Bouthillier , Teodor Gabriel Crainic, A cooperative parallel meta-heuristic for the vehicle routing problem with time windows, Computers and Operations Research, v.32 n.7, p.1685-1708, July 2005
vehicle routing;constraint programming;tabu search;Guided Local Search
595077
Constraint Logic Programming and Generalized Shortest Path Problems.
In this paper we study the relationship between Constraint Programming (CP) and Shortest Path (SP) problems. In particular, we show that classical, multicriteria, partially ordered, and modality-based SP problems can be naturally modeled and solved within the Soft Constraint Logic Programming (SCLP) framework, where logic programming is coupled with soft constraints. In this way we provide this large class of SP problems with a high-level and declarative linguistic support whose semantics takes care of both finding the cost of the shortest path(s) and also of actually finding the path(s). On the other hand, some efficient algorithms for certain classes of SP problems can be exploited to provide some classes of SCLP programs with an efficient way to compute their semantics.
Introduction Shortest Path (SP) problems [4, 11] are among the most studied network optimization problems. They are mainly used to represent and solve transportation problems, where the optimization may involve dierent criteria, says cost, time, resources, etc. Most interesting is the multi-criteria case, where the optimization involves a set of criteria to be all optimized [5]. In this paper, we propose the Soft Constraint Logic Programming (SCLP) frame-work [2, 7] as a linguistic support and a high-level and exible programming environment where to model SP problems naturally and solve them eciently. SCLP programs are logic programs [9] where each ground atom can be seen as an instantiated soft constraint [1, 3] and it can be associated with an element taken from a set. Formally, this set is a semiring, that is, a set plus two operations, and , which basically say how to combine constraints and how to compare them. The presence of these two operations allows to replace the usual boolean algebra for logic programming with a more general algebra where logical and and logical or are replaced by the two semiring operations. In this way, the underlying * Research partially supported by Italian MURST projects \Tecniche Formali per Sistemi Soft- ware" and "Concurrency, Higher-Order and Types". logic programming engine provides a natural tool to specify and solve combinatorial problems (like SP problems), while the soft constraint machinery provides greater expressivity and exibility. In particular, the exibility comes from the fact that the same SCLP programming environment, and operational semantic engine, can be used with dierent semirings (thus using dierent kinds of soft constraints), as we will see later in this paper. Here we consider several versions of SP problems, from the classical one to the multi-criteria case, from partially ordered SP problems to those that are based on modalities, and we show how to model and solve them via SCLP programs. The basic idea is that soft constraints allow to faithfully represent the optimization criteria, and CLP provides a declarative way to describe the given SP problem. Moreover, this way of modeling and solving SP problems allows to associate with such problems both a declarative and an operational semantics. The main results of the paper are as follows: Both classical, multi-criteria, partially-ordered, and modality-based SP problems are given a modelization as SCLP programs; such programs are able to both nd the cost of the shortest path(s) and also a shortest path itself. A general methodology is provided to nd a non-dominated path for both multi-criteria and partially-ordered SP problems; this methodology is based on a change of semiring, but does not require any change in the underlying SCLP syntax and semantics. A new algorithm to obtain the semantics for a particular class of SCLP programs is given, which is obtained by using a generalized version of the Floyd-Warshall algorithm [6] for SP problems. The paper is organized as follows: Section 2 gives the basic notions of soft constraints and SCLP programs, Section 3 shows the construction to pass from a classical SP problem to a CLP program, while Section 4 considers multi-criteria and partially-ordered SP problems, and Section 5 deals with modality-based SP problems. Then, in Section 6 we provide a class of SCLP programs with an efcient algorithm to compute their semantics, and nally Section 7 concludes the paper by summarizing its contributions and hinting at interesting directions for future work. 2. Soft Constraint Logic Programming In this section we describe the soft constraint logic programming (SCLP) framework, which we will later use to to describe SP problems in a declarative fashion. All the formal details that are skipped here can be found in [2, 3]. The SCLP framework is based on the notion of c-semiring introduced in [1, 3]. A c-semiring S is a tuple hA; A is a set with two special elements and with two operations and that satisfy certain properties: is closed, commutative, associative, idempotent,0 is its unit element and 1 is its a . 0.9 b . 0.5 a a . 0.8 a b . 0.2 b a . 0 b . 0.1 x y a . 0.9 Figure 1. A fuzzy CSP. absorbing element; is closed, associative, commutative, distributes over +, 1 is its unit element, and 0 is its absorbing element. The operation denes a partial order S over A such that a S b i a when a S b holds, we say that b is better than a. Other properties related to the two operations are: + and are monotone on S ; 0 is its minimum and 1 its maximum; is a complete lattice and + is its lub. Finally, if is idempotent distributive lattice and its glb. Semiring-based constraint satisfaction problems (SCSPs) are constraint problems where each variable instantiation is associated with an element of a c-semiring A (to be interpreted as a cost, level of preference, . ), and constraints are combined via the operation and compared via the S ordering. Varying the set A and the meaning of the + and operations, we can represent many dierent kinds of problems, having features like fuzziness (using the semiring optimization Moreover, the Cartesian product of two c-semirings is another c-semiring, and this can be fruitfully used to describe multi-criteria constraint satisfaction and optimization problems. Figure 1 shows an example of a fuzzy SCSP. In this gure, variables are circles, constraints are arcs, and semiring values are written to the right of the corresponding tuples. The solutions of this SCSP are pairs of values (that is, a value for each of the two variables) and, according to the c-semiring semiring value is obtained by looking at the smallest value for all the subtuples (as many as the constraints) forming the pair. The SCLP framework [2] extends the classical constraint logic programming formalism [8] in order to handle also SCSP [1, 3] problems. In passing from CLP to SCLP languages, we replaced classical constraints with the more general SCSP constraints where we are able to assign a level of preference to each instantiated constraint (i.e. a ground atom). To do this, we also modied the notions of inter- pretation, model, model intersection, and others, since we had to take into account the semiring operations and not the usual CLP operations. For example, while CLP interpretations associate a truth value (either true or false) with each ground atom, in SCLP ground atoms must be given one of the elements of the semiring. Also, while in CLP the value associated with an existentially quantied atom is the logical or among the truth values associated with each of its instantiations, here we Table 1. A simple example of an SCLP program p(a,c) :- r(a). q(a) :- t(a). t(a) :- 2. r(a) :- 3. had to replace the or with another operation which refers to one of the semiring operations (the +). The combination of atoms, which in CLP is modeled via logical and, in SCLP is instead handled via the operation of the semiring. Besides the model-theoretic semantics based on models and interpretations, SCLP programs come also with a xpoint and an operational semantics. These semantics are conservative extensions of the corresponding ones for logic programming (LP), since by choosing a particular semiring (the one with just two elements, true and false, and the logical and and or as the two semiring operations) we get exactly the semantics. The three semantics have been shown to be equivalent. In particular, we have proven that, given the set of all refutations starting from a given goal, it is possible to derive the declarative meaning of both the existential closure of the goal and its universal closure. A simple example of an SCLP program over the semiring hN; min; +; +1; 0i, where N is the set of non-negative integers and is represented in Table 1. The choice of this semiring allows to represent constraint optimization problems where the semiring elements are the costs for the instantiated atoms. Note that the ordering S in this semiring coincides with the ordering over integers. The intuitive meaning of a semiring value like 3 associated with the atom r(a) is that r(a) costs 3 units. Thus the set N contains all possible costs, and the choice of the two operations min and + implies that we intend to minimize the sum of the costs. This gives us the possibility to select the atom instantiation which gives the minimum cost overall. Given a goal like s(x) to this program, the semantics collects both a substitution for x (in this case, x = a) and also a semiring value (in this case, 2) which represents the minimum cost among the costs for all refutations for s(x). Figure 2 shows the tree structure representing the two possible refutations for the goal s(x). Each node in the tree represents a state of the computation, and the root represents the initial state. Each computation state contains both the current goal and a semiring value. At the beginning, the semiring value is the best one (0 in this example). Then, at each step, one clause or fact is used, and this changes the current goal, accumulates a substitution (written on the links of the tree), and combines the current semiring value with a new value given by the used clause. If a clause has no semiring value (like the rst four clauses of the example), the associated value is assumed to be the best one. The combination of these two semiring values is achieved via the multiplicative operation of the semiring, which in this example is the sum. All the semiring values obtained by the refutations of the same goal (and with the same substitution) are then combined together via the additive operation of the semiring, which in this example is the min operation. Therefore the two refutations in Figure 2 produce the semiring value 2. The formal denition of the operational semantics of SCLP programs can be found in [2]. Figure 2. The SCLP refutation tree for goal s(X). The presence of a possibly partial order among the elements of the semiring makes the operational semantics more complex to compute. In fact, there could be several refutations for a goal which lead to dierent semiring elements which are not comparable in the partial order. In this case, these elements have to be combined in order to get the solution corresponding to the given goal, and their combination could be not reachable by any refutation path in the search tree. The fact that we have to combine several refutation paths when we have a partial order (instead of a total one) can be fruitfully used in the context of this paper when we have an SP problem with incomparable costs in the arcs. In fact, in the case of a partial order, the solution of an SP problem should consists of all those paths which are not \dominated" by others. SCLP programming has been implemented. Its implementation has been described in [7], and the system clp(FD,S) is available for free. More information about the clp(FD,S) system, including also the instructions on how to down-load it, can be found at the URL http://pauillac.inria.fr/~georget/clp fds.html. 3. Classical SP Problems A shortest path (SP) problem can be represented as a directed graph where each arc e 2 E from node p to node q (p; q 2 N) has associated a label representing the cost of the arc from p to q. There are four versions of the problem: the single pair problem, the single source problem, the single sink problem, and the all pair problem. However, the single source and the single sink problems are directional duals of each other, the single pair problem at least partially solves the single source/sink problem, and one way to solve the all pair problem is to solve n single source/sink problems. For these reasons, the single sink (source) is fundamental and we concentrate on it. Given a set S of nodes and any node v in S (the sink), a solution of the problem consists of nding a path (or a set of paths) between any node of S and v, whose cost is minimal. Consider for example the SP problem represented in Figure 3: each arc has associated a label representing the cost (in money, time, space, . ) of that arc. In this example node v is the sink. Thus, given any node, we want to nd a path from this node to the sink v (if it exists) that minimizes the cost.33 s r Figure 3. An SP problem. To represent the classical version of SP problems, we consider SCLP programs over the semiring which, as noted above, is an appropriated framework to represent constraint problems where one wants to minimize the sum of the costs of the solutions. From any SP problem we can build an SCLP program as follows. For each arc we have two clauses: one describes the arc and the other one its cost. More precisely, the head of the rst clause represents the starting node, and its body contains both the nal node and a predicate, say c, representing the cost of the arc. Then, the second clause is a fact associating to predicate c its cost (which is a semiring element). For example, if we consider the arc from p to q with cost 2, we have the clause Table 2. The SCLP program representing the SP problem in Figure 3. 2. pr , r. c pr :- 3. r :- c rq , q. c rq :- 7. r :- c rt , t. c rt :- 1. r :- c ru , u. c ru :- 3. s :- c sp , p. c sp :- 1. 2. 2. ts , s. c ts :- 3. t. c ut :- 2. and the fact c pq :- 2. Finally, we must code that we want v to be the nal node of the path. This is done by adding a clause of the form v :- 0. Note also that any node can be required to be the nal one, not just those nodes without outgoing arcs (like v is in this example). The whole program corresponding to the SP problem in Figure 3 can be seen in Table 2. To compute a solution of the SP problem it is enough to perform a query in the SCLP framework; for example, if we want to compute the cost of the path from r to v we have to perform the query :- r. For this query, we obtain the value 6, that represents the cost of the best path(s) from r to v. Notice that to represent classical SP problems in SCLP, we do not need any variable. Thus the resulting program is propositional. However, this program, while giving us the cost of the shortest paths, does not give us any information about the arcs which form such paths. This information could be obtained by providing each predicate with an argument, which represents the arc chosen at each step. Figure 4 shows the same SP problem of Figure 3 where the arcs outgoing each node have been labeled with dierent labels to distinguish them. Such labels can then be coded into the corresponding SCLP program to \remember" the arcs traversed during the path corresponding to a solution. For example, clause would be rewritten as p(a) :- c pq , q(X). Here constant a represents one of the arcs going out of p: the one which goes to q. If all clauses are rewritten similarly, then the answer to a goal like :- r(X) will be both a semiring value (in our case 6) and a substitution for X . This substitution s r a c a a a c c a Figure 4. An SP problem with labeled arcs. will identify the rst arc of a shortest path from r to v. For example, if we have means that the rst arc is the one that goes from r to t. To nd a complete shortest path, we just need to compare the semiring values associated with each instantiated goal, starting from r and following the path. For example, in our case (of the goal 9X:r(X)) we have that the answer to the goal will be with semiring value 6. Thus we know that a shortest path from r to v can start with the arc from r to u. To nd the following arc of this path, we compare the semiring values of u(a), u(b), and u(c). The result is that u(c) has the smallest value, which is 3. Thus the second arc of the shortest path we are constructing is the one from u to v. The path is now nished because we reached v which is our sink. Notice that a shortest path could be found even if variables are not allowed in the program, but more work is needed. In fact, instead of comparing dierent instantiations of a predicate, we need to compare the values associated with the predicates that represent nodes reachable by alternative arcs starting from a certain node, and sum them to the cost of such arcs. For example, instead of comparing the values of p(a) and p(b) (Figure 4), we have to compare the values of q of r Figure 3). A third alternative to compute a shortest path, and not only its cost, is to use lists: by replacing each clause of the form with the clause during the computation we also build the list containing all arcs which constitute the corresponding path. Thus, by giving the goal :- p(L)., we would get both the cost of a shortest path and also the shortest path itself, represented by the list L. An alternative representation, probably more familiar for CLP-ers, of SP problems in SCLP is one where there are facts of the form c(p,q) :- 2. s r Figure 5. A multi-criteria SP problem. c(u,v) :- 3. to model the graph, and the two clauses path(X,Y) :- c(X,Y). path(X,Y) :- c(X,Z), path(Z,Y). to model paths of length one or more. In this representation the goal to be given to nd the cost of the shortest path from p to v is :- path(p,v). This representation is obviously more compact than the one in Table 2, and has equivalent results and properties. However, in this paper we will use with the simpler representation, used in Table 2, where all clauses have at most one predicate in the body. The possibility of representing SP problems with SCLP programs containing only such a kind of clauses is important, since it will allow us to use ecient algorithms to compute the semantics of such programs (see Section 6 for more details). 4. Partially-ordered SP Problems Sometimes, the costs of the arcs are not elements of a totally ordered set. A typical example is obtained when we consider multi-criteria SP problems. For example, if an arc represents a piece of highway between two cities, then we can label each arc both with the cost and with the time needed to follow this piece. Then, we want to nd the paths which have both the minimum overall cost and the minimum overall time. In this example, there may be cases in which the labels of two arcs are not compatible, like h$5; 20 0 i and h$7; 15 0 i. In general, when we have a partially ordered set of costs, it may be possible to have several paths, all of which are not dominated by others, but which have dierent incomparable costs. Consider for example the multi-criteria SP problem shown in Figure 5: each arc has associated a pair representing the weight of the arc in terms of cost and time. Given any node, we want to nd a path from this node to v (if it exists) that minimizes both criteria. Table 3. The SCLP program representing the multi-criteria SP problem in Figure 5. r :- c rq , q. c rq :- < 7,3 >. r :- c rt , t. c rt :- < 1,3 >. r :- c ru , u. c ru :- < 3,4 >. ts , s. c ts :- < 3,2 >. t. c ut :- < 2,1 >. We can translate this SP problem into the SCLP program in Table 3. This program works over the semiring are classical min and +, suitably extended to pairs. In practice, this semiring is obtained by putting together, via the Cartesian product, two instances of the semiring hN; min; +; +1; 0i (we recall that the Cartesian product of two c-semirings is a c-semiring as well [3]). On of the instances is used to deal with the cost criteria, the other one is for the time criteria. By working on the combined semiring, we can deal with both criteria simultaneously: the partial order will tell us when a pair h cost, time i is preferable to another one, and also when they are not comparable. To give an idea of another practical application of partially ordered SP problems, just think of network routing problems where we need to optimize according to the following criteria: minimize the time, minimize the cost, minimize the number of arcs traversed, and maximize the bandwidth. The rst three criteria correspond to the same semiring, which is hN; min; +; +1; 0i, while the fourth criteria can be characterized by the semiring h[0; U ]; max; min; 0; Ui, where U is the maximal bandwidth in an arc. In this example, we have to work on a semiring which is obtained by vectorizing all these four semirings. Each of the semirings is totally ordered but the resulting semiring, whose elements are four-tuples, is partially ordered. Notice that the only dierence between the structure of the program in Table 3 and the one in the previous section is that here we have costs represented by pairs of values, and, since we have a partial order, two such pairs may possibly be incomparable. This may lead to a strange situation while computing the semantics of a given goal. For example, if we want to compute the cost and time of a best path from p to v, by giving the query :- p., the resulting answer in this case is the value h7; 7i. While the semiring value obtained in totally ordered SCLP programs represented the cost of one of the shortest paths, here it is possible that there are no paths with this cost: the obtained semiring value is in fact in general the greatest lower bound of the costs of all the paths from p to v. This behavior comes from the fact that, if dierent refutations for the same goal have dierent semiring values, the SCLP framework combines them via the + operator of the semiring (which, in the case of our example, is the min 0 operator). If the semiring is partially ordered, it may be that a + b is dierent from both a and b. On the contrary, if we have a total order a + b is always either a or b. This of course is not satisfactory, because usually one does not want to nd the greatest lower bound of the costs of all paths from the given node to the sink, but rather prefers to have one of the non-dominated paths. To solve this problem, we can add variables to the SCLP program, as we did in the previous section, and also change the semiring. In fact, we now need a semiring which allows us to associate with each node the set of the costs of all non-dominated path from there to the sink. In other words, starting from the semiring (which, we recall, in our case is hN now have to work with the semiring P H is the Hoare Power Domain [10] of A, that is, In words, P H (A) is the set of all subsets of A which are downward closed under the ordering S . It is easy to show that such sets are isomorphic to those containing just the non-dominated values. Thus in the following we will use this more compact representation for eciency purposes. In this compact representation, each element of P H (A) will represent the costs of all non-dominated paths from a node to the sink; the top element of the semiring is the set A (its compact form is f1g, which in our example is fh0; 0ig); the bottom element is the empty set; the additive operation ] is the formal union [10] that takes two sets and obtains their union; the multiplicative operation takes two sets and produces another set obtained by multiplying (using the multiplicative operation of the original semiring, in our case +') each element of the rst set with each element of the second one; the partial order of this semiring is as follows: a P H (S) b i a b, that is for each element of a, there is an element in b which dominates it (in the partial order S of the original semiring). From the theoretical results in [10], adapted to consider c-semirings, we can prove that P H (S) and its more compact form are indeed isomorphic. Moreover, we can also prove that given a c-semiring S, the structure P H (S) is a c-semiring as well. Theorem 1 [P H (S) is a c-semiring] Given a c-semiring 1i, the structure obtained using the Power domain of Hoare operator is a c-semiring. Proof: It easily follows from the properties of the operator in the c-semiring S and from the properties (commutativity, associativity, and idempotence) of the formal union ] in P H (S). Note that in this theorem we do not need any assumption over the c-semiring S. Thus the construction of P H (S) can be done for any c-semiring S. Notice also that, if S is totally ordered, the c-semiring P H (S) does not give any additional information w.r.t. S. In fact, if we consider together the empty set (with the meaning that there are no paths) and the set containing only the bottom of A (with the meaning that there exists a path whose cost is 0), it is possible to build an isomorphism between S and P H (S) by mapping each element p (a set) of onto the element a of A such that a 2 p and a dominates all elements in the set p. The only change we need to make to the program with variables, in order to work with this new semiring, is that costs now have to be represented as singleton sets. For example, clause c pq :- < 2, 4 >. will become c pq :- f< 2, 4 >g. Let us now see what happens in our example if we move to this new semiring. First we give a goal like :- p(X). As the answer, we get a set of pairs, representing the costs of all non-dominated paths from p to v. All these costs are non-comparable in the partial order, thus the user is requested to make a choice. However, this choice could identify a single cost or also a set of them. In this second case, it means that the user does not want to commit to a single path from the beginning and rather prefers to maintain some alternatives. The choice of one cost of a specic non-dominated path will thus be delayed until later. If instead the user wants to commit to one specic cost at the beginning, say then proceeds to nd a path which costs exactly By comparing the answers for all goals of the form p(a), where a represents one of the arcs out of p, we can see which arc can start a path with cost In fact, such an arc will be labeled and will lead to a node with an associated set of costs Suppose it is the arc from p to q, which has cost h7; 3i. Now we do the same with goals of the form q(a), to see which is the next arc to follow. However, we now have to look for the presence of a pair Notice that each time we look for the next arc, we choose just one alternative and we disregard the others. If we used a xpoint (or any bottom-up) semantics to compute the answer of the initial goal :- p(X)., then all the other answers we need for this method have been already computed, thus the method does not require any additional computational eort to nd a non-dominated path. Notice also that the sets of costs associated with each node are non-dominated. Thus the size of these sets in the worst case is the size of the maximal "horizontal slice" of the partial order: if we can have at most N non-dominated element in the partial order, then each of such sets will have size at most N . Of course in the worst case N could be the size of the whole semiring (when the partial order is completely " at"). Most classical methods to handle multi-criteria SP problems nd the shortest paths by considering each criteria separately, while our method deals with all criteria at once. This allows to obtain optimal solutions which are not generated by looking at each single criteria. In fact, some optimal solutions could be non-optimal in each of the single criteria, but still are incomparable in the overall ordering. Thus we oer the user a greater set of non-comparable optimal solutions. For example, by using the time-cost multi-criteria scenario, the optimal solution w.r.t. time could be 1 minute (with cost of 10 dollars), while the optimal solution w.r.t. cost could be 1 dollar (with time of 10 minutes). By considering both criteria together, we could also obtain the solution with 2 minutes and 2 dollars! Finally, this method is applicable not only to the multi-criteria case, but to any partial order, giving us a general way to nd a non-dominated path in a partially-ordered SP problem. It is important to notice here the exibility of the semiring approach, which allows us to use the same syntax and computational engine, but on a dierent semiring, to compute dierent objects. 5. Modality-based SP Problems Until now we have considered situations in which an arc is labeled by its cost, be it one element or a tuple of elements as in the multi-criteria case. However, sometimes it may be useful to associate with each arc also information about the modality to be used to traverse the arc. For example, interpreting arcs as links between cities, we may want to model the fact that we can cover such an arc by car, or by train, or by plane. Another example of a modality could be the time of the day in which we cover the arc, like morning, afternoon, and night. In all these cases, the cost of an arc may depend on its modality. An important thing to notice is that a path could be made of arcs which not necessarily are all covered with the same modality. For example, the path between two cities can be made of arcs, some of which are covered in the morning and others in the afternoon. Moreover, it can be that dierent arcs have dierent sets of modalities. For example, from city A to city B we can use both the car or the train, and from city B to city C we can use only the plane. Thus modalities cannot be simply treated by selecting a subset of arcs (all those with the same modality). An example of an SP problem with the three modalities representing car (c), plane (p), and train (t) can be seen in Figure 6. Here the problem is to nd a shortest path from any node to v, and to know both its cost and also the modalities of its arcs. This SP problem with modalities can be modeled via the SCLP program in Table 4. In this program, the variables represent the modalities. If we ask the query :-p(c)., it means that we want to know the smallest cost for a trip from p to v using the car. The result of this query in our example is (using the path p r u v). c c c c c c c s r Figure 6. An SP problem with modalities. Table 4. The SCLP program representing the SP problem with modalities. 2. pr (X), r(X). c pq (p) :- 3. pr (c) :- 3. 7. 7. ts (X), s(X). c sr (t) :- 2. 2. ts (p) :- 3. ts (t) :- 3. c up (c) :- 3. c up (t) :- 1. c ut (t) :- 2. c uv (t) :- 3. c uv (c) :- 2. Notice that the formulation shown in Figure 4 puts some possibly undesired constraints on the shortest path to be found. In fact, by using the same variable in all the predicates of a rule, we make sure that the same modality (in our case the same transport mean) is used throughout the whole path. If instead we want to allow dierent modalities in dierent arcs of the path, then we just need to change the rules by putting a new variable on the last predicate of each rule. For example, rule would become Now we can use a modality for the arc from p to q, and another one for the next arc. In this new program, asking the query :-p(c). means that we want to know the smallest cost for a trip from p to v using the car in the rst arc. The same methods used in the previous sections to nd a shortest path, or a non-dominated path in the case of a partial order, can be used in this kind of SCLP programs as well. Thus we can put additional variables in the predicates to represents alternative arcs outgoing the corresponding nodes, and we can shift to the semiring containing sets of costs to nd a non-dominated path. In particular, a clause like would be rewritten as p(X,a) :- c pq (X), q(Y,Z). 6. An SP Algorithm for a Class of SCLP Programs Summarizing what we did in the previous sections, we can say that the general form of the SCLP programs we use to represent SP problems consists of several clauses for each predicate p i , where each clause body has one constraint and one other predicate p j , plus one special clause for the sink predicate (with a 0 in the body), plus the facts dening the costs of the arcs. Table 5 shows this general form (only the clauses and without variables). It is important to notice that these are not general SCLP programs, since there is always one predicate symbol in each clause body (since the constraint can be replaced by its cost). In the logic programming literature, these programs are called linear, because of the restriction on the number of predicates in the bodies of the clauses. Given any SCLP program of the form shown in Table 5, its semantics can be obtained using classical methodologies (for the CLP literature) as described in Section 2. However, we will now show an additional method, based on a classical algorithm for SP problems, suitable adapted to c-semirings, to obtain the same semantics. The algorithm we will use is the one developed by Floyd and Warshall [6]. To put the SCLP program into a shape which can be handled by this algorithm, we rst need to transform it into a system of linear equations, one for each predicate: the left hand side of the equation is the chosen predicate, while the right hand side is obtained by combining the bodies of all the clauses dening the predicates via the operator, and replacing the comma with the operator in each body. The Table 5. The general form of an SCLP program representing an SP problem. Table 6. The system of equations corresponding to the SCLP program form of Table 5. result of this transformation, applied to the program of Table 5, can be seen in Table 6. From this system of equations we can now build a graph with a node i for each predicate and where the cost of the arc between nodes i and j is given by predicate c ij . In this way, this system of equations can be interpreted also as the matrix showing, for each pair of nodes, the cost of the arc (if it exists) between such nodes. In fact, each column (and row) can be associated with one node. Given this matrix, the algorithm works by selecting each triple of nodes, say i, and k, and performing the following and operators in this assignment refer to the operations of the chosen semiring. The triples are chosen by rst selecting a value for index k (which is the intermediate node in the two-step path), and then varying the indices i and j in all possible ways (to consider all two-step paths from i to j). After considering all values for k, that is, after n 3 steps, the value of the element in row i and column v represents the cost of the shortest path from node i to v. In terms of SCLP semantics, this is the semantics of predicate p i . This method to obtain the semantics of an SCLP program of the form in Table 5 can be used independently of the semiring underlying the given program. In fact, as noted above, the version we use of the original Floyd-Warshall algorithm does not depend on the meaning of the + and operators. In particular, it can also be used for partially ordered semirings. However, if the semiring we consider is totally ordered, we can also use any other classical algorithm for solving SP problems. Summarizing, this section provides a class of SCLP programs, over any semiring, with a new method to obtain the semantics whose complexity is cubic in the number of predicates of the program. The form of the SCLP program in Table 5 does not contain any variable. However, our method works also when each predicate (both the p i and the c ij predicates) has some variables, provided that, in each clause, all the predicates have the same variables in the same order. Thus, for example, we can admit clauses of the form This kind of clauses can be handled by passing to a more complex semiring, where each element is a vector (or a matrix) of the elements of the original semiring, and the two operations are vectorizations of the original ones. 7. Conclusions We have investigated the relationship between shortest path problems and constraint programming, proposing the soft constraint logic programming framework to model and solve many versions of this problem in a declarative way. In particular, both classical, multi-criteria, partially-ordered, and modality-based SP problems are shown to be expressible in this programming formalism. We also have shown that certain algorithm to solve SP problems can be used to compute eciently the semantics of a certain class of SCLP programs. Such a class could be extended by using structures more complex than semirings, or by considering trees instead of paths. We plan to investigate these structures in the future. In particular, we plan to use SCLP programs also to model and solve best tree problems. We also intend to apply our methodology to specify and solve planning problems, since in many cases such problems can be cast as best tree problems over the graph describing all possible moves of the planning scenario. --R Constraint Solving over Semirings. An appraisal of some shortest-paths algorithms Algorithm 97: Shortest path. Compiling semiring-based constraints with clp(fd Foundations of Logic Programming. Power domains. --TR --CTR Stefano Bistarelli , Francesca Rossi, Semiring-based contstraint logic programming: syntax and semantics, ACM Transactions on Programming Languages and Systems (TOPLAS), v.23 n.1, p.1-29, Jan. 2001 Alberto Lluch-Lafuente , Ugo Montanari, Quantitative -calculus and CTL defined over constraint semirings, Theoretical Computer Science, v.346 n.1, p.135-160, 23 November 2005
shortest path problems;soft constraints;constraint logic programming
595085
Greedy and Local Search Heuristics for Unconstrained Binary Quadratic Programming.
In this paper, a greedy heuristic and two local search algorithms, 1-opt local search and k-opt local search, are proposed for the unconstrained binary quadratic programming problem (BQP). These heuristics are well suited for the incorporation into meta-heuristics such as evolutionary algorithms. Their performance is compared for 115 problem instances. All methods are capable of producing high quality solutions in short time. In particular, the greedy heuristic is able to find near optimum solutions a few percent below the best-known solutions, and the local search procedures are sufficient to find the best-known solutions of all problem instances with n 100. The k-opt local searches even find the best-known solutions for all problems of size n 250 and for 11 out of 15 instances of size n &equals; 500 in all runs. For larger problems (n &equals; 500, 1000, 2500), the heuristics appear to be capable of finding near optimum solutions quickly. Therefore, the proposed heuristicsespecially the k-opt local searchoffer a great potential for the incorporation in more sophisticated meta-heuristics.
is given, and a binary vector of length n is searched, such that the quantity is maximized. This problem is also known as the (unconstrained) quadratic bivalent programming problem, the (unconstrained) quadratic zero{one programming problem, or the (uncon- strained) quadratic (pseudo-) Boolean programming problem (Beasley, 1998). The BQP is known to be NP-hard (Garey & Johnson, 1979) and has a large number of applications, for example in capital budgeting and nancial analysis problems (Laughunn, 1970; McBride & Yormark, 1980), CAD problems (Krarup & Pruzan, 1978), tra-c message management problems (Gallo et al., 1980), machine scheduling (Alidaee et al., 1994), and molecular conformation (Phillips & Rosen, 1994). Furthermore, several other combinatorial optimization problems can be formulated as a BQP, such as the maximum cut problem, the maximum clique problem, the maximum vertex packing problem and the maximum independent set problem (Ivanescu, 1965; Pardalos & Rodgers, 1992; Pardalos & Xue, 1994). Several exact methods have been developed to solve the BQP (Pardalos & Rodgers, 1990; Barahona et al., 1989; Billionnet & Sutter, 1994; Helmberg & Rendl, 1998), but due to the computational complexity of the problem, heuristics have been proposed recently to nd solutions to large problem instances, including tabu search (Beasley, 1998; Glover et al., 1998b; Glover et al., 1998a), scatter search (Amini et al., 1999), simulated annealing (Beasley, 1998; Katayama & Narihisa, 1999) and evolutionary algorithms (Lodi et al., 1998; Merz & Freisleben, 1999b). In this paper, a greedy heuristic and two local search algorithms for the unconstrained binary quadratic programming problem (BQP) are presented. The greedy heuristic constructs a feasible solution in n steps where n denotes the problem size, i.e. the number of elements of a solution. In each step, the greedy heuristic selects an element and a value for it by making a most favorable choice. Additionally, a randomized greedy heuristic is described in which random components are added to the deterministic greedy heuristic. The local search heuristics iteratively search a better solution in the neighborhood of the current solution until no better solution exists and thus a local optimum is reached. The neighborhood of the rst local search algorithm {the 1-opt local search { is dened by the solutions that can be reached by changing a single element in the current solution. The second local search is a k-opt local search where k elements in the solution are changed simultaneously. Due to the computational complexity, only a small fraction of the k-opt neighborhood is searched by the local search procedure. The algorithms are tested on a set of 115 problem instances of the BQP. The experiments show that the local search heuristics are able to nd optimum or best-known solutions for instances with a problem size of up to 100 easily. The multi{start k-opt local search is capable of nding best-known solutions in all runs for all problem instances studied with a problem size less or equal to 250 and for 11 out of 15 instances of size 500. For larger instances the algorithms are shown to nd near optimum solutions in 10 to 60 seconds on a state-of-the-art workstation: in most cases, the multi{start randomized greedy heuristic, the multi{start 1-opt local search, and the multi{start k-opt local search nd solutions within 1%, 0.5%, and 0.1% below the best-known solution, respectively. The paper is organized as follows. In section 1, the BQP is described in more detail. Section 2 describes the greedy and local search heuristics for the BQP. The results obtained from various experiments conducted with these algorithms are discussed in section 3. Section 4 concludes the paper and outlines areas of future research. 1 The Binary Quadratic Programming Problem Throughout this paper, a solution vector x of the BQP is a boolean vector of length n: . Thus, the solution space X of the BQP is of size 2 n ; it grows exponentially with n. The density of the matrix Q is sometimes used to characterize a BQP instance. The density dens(Q) is dened as the number of non-zero entries divided by the number of total entries in the matrix Q. Thus, the density varies between 0 and 1. To demonstrate the relevance in various scenarios, some special cases and a generalization of the BQP are described in the following. 1.1 Special Cases of the BQP The BQP has been shown to be a generalization of other combinatorial optimization problems. For example, the maximum clique problem and the maximum independent set problem are special cases of the BQP. Let E) be an undirected graph and E) be the complement graph of G, where Eg. Furthermore, let be the adjacency matrix of G, I denote the identity matrix, and T g. Then, the maximum clique problem is min If x solves equation (2), the maximum clique of G is dened as Similarly, the maximum independent set problem is min If x solves equation (3), the maximum independent set of G is dened as In the maximum cut problem the objective function has to be maximized, where w ij denotes the weight of the edge (i; in the graph which the maximum cut is desired. The maximum cut problem can be formulated as a 0/1 quadratic programming problem by assigning: The maximum cut size c(x ) is equal to the objective f(x ) of the corresponding BQP, and the cut itself is 1g. Another application of the BQP arises in condensed matter physics. The calculation of ground states in Ising Spin Glasses is a combinatorial optimization problem in which a conguration of the spins with minimum energy is searched. The energy of an ising spin glass, in which the spins lie on a two dimensional grid, is given by the Hamiltonian: denotes the interaction between site i and j on the grid. By setting the solution of the BQP yields a conguration with minimum energy, where s and 1.2 A Generalization of the BQP There is a close relation between binary quadratic programming and NK-landscapes as dened by Kauman (Kauman & Levin, 1987) to model gene interaction in genetic evolution. This relation is shown in the following. The objective function of the BQP can be decomposed into functions. The tness of a BQP solution can thus be rewritten as a sum of functions for each site, called the tness contributions f i of component i in the bit vector: Similar to the NK-landscapes dened in (Kauman & Levin, 1987), the tness contribution f i of a site i depends on the gene value x i and of k(i) other genes x . While for NK-landscapes is constant for all i, in the BQP k(i) is dened as the number of non-zero entries in the i-th column of matrix Q. The mean k of the k(i) is given by dens(Q). The BQP can be viewed as a special class of NK-landscapes with a special tness contribution function f i . Alternatively, NK landscapes can be seen as a family of more general binary programming problems. Heuristics for the BQP Heuristics can be divided into construction heuristics and improvement heuristics. The former construct feasible solutions for a given optimization problem from scratch, while the latter take a feasible solution as their input and try to nd better solutions by stepwise transformations. Both types of heuristics can be implemented e-ciently and are often capable of producing near optimum solutions for combinatorial optimization problems. If, however, the results obtained are not su-cient, the algorithms can be combined or incorporated into meta{heuristics such as memetic algorithms (Moscato, 1989), as for example shown in (Merz & Freisleben, 2000) for the graph bipartitioning problem: A dierential greedy heuristic (Battiti & Bertossi, 1998) combined with Kernighan-Lin local search (Kernighan & Lin, 1972) has been shown to be su-cient for structured problems of small or medium sizes. If larger problems have to be solved or the problems are unstructured, a memetic algorithm { a genetic algorithm incorporating the local search and the greedy heuristic { has been shown to be more eective. Since greedy and local search heuristics are important (components of) algorithms for combinatorial optimization problems, corresponding approaches for the BQP are desired. Thus, greedy and local search algorithms for the BQP are described in the following. To the best of our knowledge, our proposals are the rst algorithms of this kind for the BQP. 2.1 Greedy Heuristics Greedy algorithms are intuitive heuristics in which greedy choices are made to achieve a certain goal. In combinatorial optimization, a solution to a given problem is searched which maximizes or minimizes an objective function. Greedy heuristics are constructive heuristics since they construct feasible solutions for optimization problems from scratch by making the most favorable choice in each step of construction. By adding an element to the (partial) solution which promises to deliver the highest gain, the heuristic acts as a \greedy constructor". 2.1.1 A Simple Greedy Algorithm In the BQP, a solution is constructed by assigning a binary value to an element in the solution vector x: a zero or a one. The general outline of the greedy algorithm is provided in Figure 1. In each step, the heuristic searches for an element k in the solution vector and a value l to procedure Greedy(x 2 X): X; begin repeat nd k; l with g l return x; Figure 1: A Greedy Heuristic for the BQP assign to it so that a gain function g l k is maximized. Afterwards, the value l is assigned to the vector component x k . To nd an appropriate gain function, we modify the problem by adding a third state: Let be a vector in which each component y i can have three values 0, 1, and 1. Starting with a vector ^ all i, the greedy heuristic can be viewed as a transformation algorithm that transforms ^ y into a vector x for which x i 2 f0; 1g for all i and thus x 2 X. The objective of the solution ^ y is Let ~ be a vector that is equal to y 2 Y except for component y k . Then Hence, the gain for changing y k from 0:5 to 0 (denoted g 0 k ) can be dened as Using this formula, the greedy heuristic as displayed in Figure 1 has a runtime complexity of O(n 3 ) since the calculation of all g i takes O(n 2 ) and this has to be done n times before the algorithm terminates. Thus, the greedy heuristic has an expected running time equal to an algorithm that calculates the objective function f(x) for n solutions. However, the greedy algorithm can be implemented more e-ciently. If the matrix Q is not stored in a two dimensional array, but instead, for each i in ng a list of j's for which q ij 6= 0 is maintained, the running time is reduced considerably for instances with a low density. To reduce the running time for nding a choice with a highest gain, the following formula can be used to calculate the new gain g i after y k has been set to 0 or 1: 1q ik if y q ik otherwise and g 0 1q ik if y Thus, once the gains g i have been calculated, they can be e-ciently updated after the component y k has been set. Since the gains g only change for those i with q ik 6= 0, the running time for updating the gains has a complexity in O(n). In gure 2, the pseudo code of the fast greedy heuristic is provided. The running time for the improved heuristic is in O(n 2 ) since the running time for nding the maximum gain g 0 k and g 1 k is in O(n) and has to be performed n times to construct a feasible solution. Thus, the greedy heuristic has the same computational complexity as an algorithm calculating f(x) for a single solution. The running time behavior of the greedy heuristic can be improved even further if additional data structures are used to nd the maximum gain in shorter than O(n) time. However, this has no in uence on the computational complexity of the algorithm since the time for the initial calculation of the gains dominates over the times for the greedy choices. procedure Greedy(x 2 X): X; begin for to n do x calculate gains g i for all i in repeat nd k 0 with g 0 nd then else endif update gains g i for all i 2 C; return x; Figure 2: A Greedy Heuristic for the BQP 2.1.2 A Randomized Greedy Algorithm The greedy heuristic described above is deterministic, since it always produces the same solution for a given problem instance. Often it is desired that a construction heuristic produces many dierent solutions, for example in hybrid algorithms. The above procedure can be randomized as follows. By making the rst choice randomly, i.e. selecting k and l randomly and setting x in the rst step, a random component is incorporated. Furthermore, the deterministic choice among x k 0 and x k 1 can be replaced by a random choice proportional to the gains g 0 and is set to 0 with probability g 0 and x k 1 is set to 1 with probability . The pseudo code of the randomized greedy heuristic is provided in Figure 3. 2.2 Local Search Local search (LS) algorithms are improvement heuristics that search in the neighborhood of the current solution for a better one until no further improvement can be made, i.e. there is no better solution in the neighborhood of the current solution. Local search algorithms can be categorized by the neighborhoods they consider. For example, the neighborhood of a solution represented by a binary vector can be dened by the solutions that can be obtained by ipping a single or multiple components in the binary vector simultaneously. procedure begin for to n do x calculate gains g i for all i in Select k; l randomly and set x k := l; repeat nd k 0 with g 0 nd set if randomNumber[0,1]< p then else endif update gains g i for all i 2 C; return x; Figure 3: A Randomized Greedy Heuristic for the BQP 2.2.1 1-opt Local Search The simplest form of local search for the BQP is the 1-opt local search: In each step, a new solution with a higher tness in the neighborhood of the current solution is searched. The neighborhood of the current solution is dened by the set of solutions that can be reached by ipping a single bit. Hence, the 1-opt neighborhood contains all solutions with a hamming distance of 1 to the current solution. In our implementation, we search for the solution with the highest tness, i.e. we search for a ip with the highest associated gain in tness new f old ). The gain g k of ipping bit k in the current solution can be calculated in linear time using the with The local search algorithm is given in pseudo code in Fig. 4. A straightforward implementation of the 1-opt local search displayed in the gure has a running time of O(n 2 ) per iteration. procedure begin repeat nd k with if until return x; Figure 4: 1-opt Local Search Analogous to the greedy heuristic, the e-ciency of the algorithm can be increased considerably. The gains g i do not have to be recalculated each time. Instead, it is su-cient to calculate the dierence of the gains g i . Assuming that all g i for a BQP solution have been calculated and the bit k is ipped, the new gains g 0 i can be calculated e-ciently by the formula: Thus, the update of the gains can be performed in linear time. This property has also been used in (Glover et al., 1998a) to speed up tabu search for the BQP. Furthermore, only the gains have to be updated. The fast 1-opt local search is displayed in Figure 5. The procedure begin calculate gains g i for all i in repeat nd k with if update gains endif until return x; Figure 5: Fast 1{opt Local Search for the BQP running time of this algorithm is O(n) per iteration. The initialization of the gains is in O(n 2 ). 2.2.2 k-opt Local Search The k-opt neighborhood N k opt of a binary vector of length n is dened by the binary vectors that can be reached by ipping one up to k bits in the vectors simultaneously. Hence, the neighborhood N k opt denotes the hamming distance between bit vectors) grows exponentially with k: jN k opt Since it is computationally too expensive to search the complete k-opt neighborhood, Lin and Kernighan have proposed heuristics for the traveling salesman problem (TSP) and the graph partitioning problem that e-ciently search a small fraction of the k-opt neighborhood. These algorithms, known as the Lin-Kernighan algorithm for the TSP (Lin & Kernighan, 1973), and the Kernighan-Lin algorithm for graph partitioning (Kernighan & Lin, 1972), belong to the best available heuristics for these two combinatorial optimization problems. In the following, a local search algorithm for the BQP is presented that is based on the ideas of Lin and Kernighan. The basic idea of the heuristic is to nd a solution by ipping a variable number of k bits in the solution vector per iteration. In each step, a sequence of n solutions is generated by ipping the bit with the highest associated gain. Analogous to the 1-opt local search procedure, a vector of gains is maintained and updated according to equation (15) after each ip. Furthermore, a candidate set is used to assure that each bit is ipped exactly once. The best solution in the sequence is accepted as the new solution for the next iteration. Thus, in each iteration of the algorithm a variable number of bits are ipped to nd a new solution in the neighborhood of the current solution. The pseudo code of the k-opt local search is presented in Figure 6. The runtime complexity for the initialization is in O(n 2 ) and the running time per iteration is also of complexity O(n 2 ). The reduce the running time, the termination condition of the inner repeat-loop can be modied so that the loop is terminated if there were no new x best for more than m iterations. Thus, the resulting fast k-opt procedure has a shorter running time for m n than the k-opt procedure described before. Performance Evaluation To evaluate the performance of the algorithms, we conducted several experiments on all 105 problem instances contained in ORLIB (Beasley, 1990). The sets glov-a, glov-b, glov-c and glov-d contain the instances of type a through d described in (Glover et al., 1998b). The set glov200 (glov500) consists of ve problem instances of size and is denoted as type e (f) in (Glover et al., 1998b). The six sets described in (Beasley, 1998) with a density procedure begin calculate gains g i for all i in repeat repeat nd j with G; x best := x; endif update gains g i for all x := x best ; else endif return x; Figure for the BQP dens(Q) of 0.1 and n 2 f50; 100; 250; 500; 1000; 2500g consist of 10 instances each. They are denoted beashni. In the experiments, the instances of type g used in (Glover et al., 1998b; Amini et al., 1999) with between 0:1 and 1 were also considered. They are denoted as kb-g. To nd a good parameter value for m in the k-opt procedure, experiments were performed for beas1000 and beas2500. It appeared that is a good trade-o between running time and solution quality. For larger values, the running time increased considerably with only small changes in solution quality. Therefore chosen in all subsequent k-opt local search experiments. In the experiments, the performance of the randomized greedy algorithm, the 1-opt local search applied to randomly generated solutions, the fast k-opt local search applied to randomly generated solutions, and the combination of the randomized greedy heuristic and fast k-opt local search was investigated. To enable a fair comparison, all algorithms implemented in C++ were run under the same conditions on a Pentium II PC (300 MHz) under the operating system Solaris 2.6.2060100140180 time in ms density of Q greedy 1-opt k-opt time in ms Problem size n greedy 1-opt k-opt greedy+k-opt Figure 7: Average running times of greedy and local search heuristics for the BQP In a rst experiment, the average running times of the four algorithms and the average solution quality was studied. In Figure 7, the average running times (in milliseconds) of 10000 runs of the algorithms are provided. In the left plot, the running times are provided for the ve instances of the set glov500 with and a density dens(Q) contained between 0.1 and 1. As expected, the running time of the algorithms grows linearly with the density of the matrix Q. The running times of the combination of the randomized greedy heuristic and k-opt local search are slightly lower than the running times of the k-opt local search applied to randomly generated solutions, since the number of iterations of the local search is reduced when it is applied to solutions produced by the greedy heuristic. In the right plot, the running times are provided for the six sets beas50 to beas2500. For all algorithms, the running times grow quadratically with n. The k-opt algorithms appear to be 2.5 times slower than the 1-opt algorithm, and for large n, the greedy heuristic is slower than 1-opt. Thus, the number of iterations of the 1-opt procedure grows at most linearly with the problem size for the instances studied. The average solution quality of the approaches is displayed in Table 1. The solution quality is measured by the average percentage excess (avg) (100 (1:0 f(x)=f best )) over the best-known solution for a set of up to 10 instances, and the standard deviation (sdev) is also provided for each algorithm. The greedy heuristic shows a better average performance on six of the sets than the 1-opt local search, while the latter performs better on the remaining six. The k-opt heuristic performs considerably better than the greedy and 1-opt heuristic: with one exception, the average percentage excess is below 1%. However, the combination of greedy and k-opt local Table 1: Average solution quality of greedy and local search heuristics for the BQP instances greedy 1-opt k-opt greedy-k-opt avg sdev avg sdev avg sdev avg sdev glov-d 2.81 % 0.42 2.71 % 0.73 0.71 % 0.35 0.42 % 0.27 search performs best on all but one instance with respect to the average quality of the solutions. Table 2: Time to reach the optimum instances 1-opt k-opt greedy-k-opt glov-d 52.11 beas250 { { { 20.00 0.09 0.77 10.30 In a second experiment, the heuristics were repeatedly applied (multi{start) to show their ability to reach the optimum or best-known solution. Each of the heuristics was started multiple times and the best solution found was returned. For each instance, runs were performed for each algorithm, and the times to reach the best-known solutions were recorded. In Table 2, the times for the algorithms that were capable of nding the optimum in all 30 out of runs for all instances in a set are displayed. The average number of repetitions needed (rep), the average time in seconds (avg t) to reach the best-known solution, as well as the maximum time in seconds (max t) to reach the best-known solution is provided. For problems up to a size of 200, the 1-opt local search is capable of nding the optimum in less than 2 seconds. For the set glov200 the average number of local searches is about 217. The k-opt heuristic needs only about 13 local searches on the average to nd the best-known solutions for the instances of this set. Both k-opt algorithms perform better on the instances up to need less than 0.23 seconds to nd the best-known solutions and are able to nd the best-known solutions of all the instances of the set beas250. On this set, the greedy k-opt combination performs slightly better than the k-opt on random solutions: less than 0.53 seconds are needed. Table 3: Comparison of greedy, 1-opt, k-opt, and greedy-k-opt on large BQP instances instances greedy 1-opt k-opt greedy-k-opt time rep avg rep avg rep avg rep avg (sec) The results show that for small instances (up to simple local search is su-cient to nd best-known solutions quickly. The more challenging problems have a size of higher. The third experiment concentrated on these instances. To enable a fair comparison, the four algorithms were repeatedly applied until a predened time limit was reached. Once again, runs were performed for each instance. The results are summarized in Table 3. For each instance and algorithm, the average number of repetitions (rep), and the average percentage excess (avg) is given. The time limit in seconds (time) used is provided in the last column of the table. The greedy heuristic shows to be inferior to the 1-opt local search: The average percentage excess for the set beas1000 and beas2500 is between 0.727% and 1.170% in case of the greedy heuristic; for the 1-opt local search the results are between 0.216% and 0.543%. The algorithms based on k-opt are considerably better. The worst performance lies 0.128% below the optimum for the beas instances. The kb-g instances appear to be harder; within a time limit of the average solution quality lies between 0.012 % and 0.489 %. A preference to one of the k-opt based algorithms can not be given, since their performance does not dier signicantly. Both are able to nd the best-known solution for the problems glov500-1 and glov500-2 in all but not for the other problems with and a density dens(Q) greater 0.25. This indicates that problems with high densities are slightly harder for the k-opt local search. However, as the results on the kb-g instances show, the average solution quality is not a simple function of the density: the average percentage access for the problem with density 0.8 (kb-g08) is better than for the problems with density 0.4 and 0.7 (kb-g04 and kb-g07). In comparison to the tabu search and simulated annealing for the BQP proposed in (Beasley, 1998), the best found solutions obtained with the greedy heuristic and k-opt local search for the 10 problems of size 2500 are in 7 out of 10 cases better than the best found solutions reported in (Beasley, 1998). For tabu search and simulated annealing, the running times on a Silicon Graphics Indigo workstation (R4000, 100MHz) are between 12721 and 51873 seconds compared to the 60 seconds for the local search on a Pentium II 300 MHz PC. Thus, the results produced by the k-opt local search appear to be superior or at least competitive to the other approaches in solution quality per time. However, a comparison of the methods under the same conditions (computing hardware, operating system, programming language, coding is required to support this claim. More sophisticated algorithms, however, are expected to produce better solutions: the tabu search utilizing critical event memory (Glover et al., 1998a) found the best-known solution for 7 out of the 10 instances of set kb-g, in a CPU time of four minutes on a Pentium 200 PC. The scatter search approach proposed in (Amini et al., 1999) found the best-known solutions of all problems in this set, but no CPU times were reported. The heuristics proposed in this paper are not intended to be competitive with these approaches. Instead, they have been developed with the idea of building powerful meta-heuristics called memetic algorithms (Moscato, 1999; { evolutionary algorithms incoporating greedy heuristics and local search. Conclusions In this paper, a randomized greedy heuristic and two local search algorithms based on the 1-opt and a k-opt neighborhood for the BQP have been proposed. The runtime behavior and the average performance of these heuristics has been investigated in experiments on 115 problem instances. Furthermore, four algorithms based on these heuristics have been studied: the multi{start randomized greedy algorithm, the multi{start 1-opt local search on randomly generated solutions, the multi{start k-opt local search on randomly generated solutions, and the multi{start k-opt local search on solutions generated by the randomized greedy heuristic. The results indicate that multi{start 1-opt local search is able to nd best-known solutions of the studied instances up to a size of n = 100 in all runs, and multi{start k-opt local search is capable of nding the best-known solutions in all runs for all studied instances up to Furthermore, for 11 out of 15 problem instances of size the best-known solution was found in all runs with multi{start k-opt LS. Applied to larger instances (n > 500), the multi{start randomized greedy heuristic produces solutions less than 1% below the best-known solutions in most cases, the multi{start 1-opt local search solutions below 0.5% in most cases and the multi{start k-opt algorithms solutions less than 0.5% below the best-known solution in most cases in a CPU time limit of 10 seconds up to a minute on a state-of-the-art workstation. Thus, the proposed k-opt local search appears to be highly eective, and the additional coding eort required in comparison to the 1-opt heuristic is denitely justied. Both greedy and local search heuristics are well suited as components for meta{heuristics since their average running time is below 700 milliseconds even for large problems of size best-known solutions are desired for large problems, the incorporation of the heuristics in genetic algorithms appears to be a promising approach (Merz & Freisleben, 1999b). Further studies have shown that the combination of evolutionary algorithms and k-opt local are highly eective applied to large problem instances (n 500) (Katayama & Narihisa, 2000; Merz, 2000). There are several areas of future research. First, the implementation specic parameter m of the k-opt local search should be studied in more detail; an interesting question is whether there is an instance independent optimum value for the parameter or whether the parameter has to be tuned for each instance separately. Second, there may be instances for which the greedy heuristic performs better than local search, as shown for the graph bipartitioning problem (Merz & Freisleben, 2000). Additional experiments are necessary to support this claim. Third, a comparison of the k-opt local search with the recently proposed tabu search algorithms for the BQP (Beasley, 1998) is another interesting are of study. Finally, the heuristics described in this paper may be extended to NK landscapes, a generalized unconstrained binary programming problem. In fact, we are currently working on a greedy and a k-opt heuristic for NK landscapes. Initial experiments indicate that both heuristics are highly e-cient. --R A Scatter Search Approach to Unconstrained Quadratic Binary Programs. Heuristic Algorithms for the Unconstrained Binary Quadratic Programming Problem. Minimization of a Quadratic Pseudo-Boolean Function European Journal of Operational Research Quadratic Knapsack Problems. Computers and Intractability: A Guide to the Theory of NP-Completeness Solving Quadratic (0 Performance of Simulated Annealing-based Heuristic for the Unconstrained Binary Quadratic Programming Problem Solving Large Binary Quadratic Programming Problems by E Quadratic Binary Programming. An Evolutionary Heuristic for Quadratic 0-1 Programming An Implicit Enumeration Algorithm for Quadratic Integer Programming. Memetic Algorithms for Combinatorial Optimization Problems: Fitness Landscapes and E Fitness Landscapes On Evolution Memetic Algorithms: A Short Introduction. Computational Aspects of a Branch and Bound Algorithm for Unconstrained Quadratic Zero-One Programming A Branch and Bound Algorithm for the Maximum Clique Problem. A Quadratic Assignment Formulation for the Molecular Conformation Problem. --TR --CTR Kengo Katayama , Akihiro Hamamoto , Hiroyuki Narihisa, Solving the maximum clique problem by k-opt local search, Proceedings of the 2004 ACM symposium on Applied computing, March 14-17, 2004, Nicosia, Cyprus Patrick E. Meyer , Kevin Kontos , Frederic Lafitte , Gianluca Bontempi, Information-theoretic inference of large transcriptional regulatory networks, EURASIP Journal on Bioinformatics and Systems Biology, v.2007 n.1, p.8-8, January 2007 Kengo Katayama , Akihiro Hamamoto , Hiroyuki Narihisa, An effective local search for the maximum clique problem, Information Processing Letters, v.95 n.5, p.503-511, 15 September 2005 Peter Merz, Advanced fitness landscape analysis and the performance of memetic algorithms, Evolutionary Computation, v.12 n.3, p.303-325, September 2004
local search;greedy heuristics;unconstrained quadratic binary programming
595086
Cooperative Strategies for Solving the Bicriteria Sparse Multiple Knapsack Problem.
For hard optimization problems, it is difficult to design heuristic algorithms which exhibit uniformly superior performance for all problem instances. As a result it becomes necessary to tailor the algorithms based on the problem instance. In this paper, we introduce the use of a cooperative problem solving team of heuristics that evolves algorithms for a given problem instance. The efficacy of this method is examined by solving six difficult instances of a bicriteria sparse multiple knapsack problem. Results indicate that such tailored algorithms uniformly improve solutions as compared to using predesigned heuristic algorithms.
Introduction One of the major goals of a production facility is to utilize its inventory in the best possible way to satisfy demand. In make-to-order production systems, such as the process industry, a surplus inventory accumulates due to cancellations of orders and rejection of production units for failing to satisfy quality requirements. Clearly, it is to the advantage of the production facility to utilize this surplus inventory before planning its production acitivites. This raises the problem of assigning a list of orders to the production units in the inventory. The objectives are both maximizing the total amount of orders that are assigned, and minimizing total waste of production units. Manufacturability considerations such as the compatibility of orders and production units in terms of quality, size, etc. impose additional assignment constraints. As production operations involve more complex processes and a larger product variety, the problem becomes more constrained. The bicriteria sparse multiple knapsack problem that we consider in this study is motivated by this application. In this paper we focus on the use of a team of heuristic algorithms which cooperate to generate non-dominated solutions for this problem in a short computation time. Although there exist several heuristic approaches for solving multiple knapsack problems there does not exist a single dominant algorithm. Moreover, the performance of the heuristics vary by problem instance and as a result a specific heuristic will often demonstrate poor aggregate performance over a set of problem instances. However, if the heuristics were allowed to cooperate with each other so that ffl the solution generated by one heuristic can be subsequently improved by another or ffl the most appropriate subset of heuristics can be used to construct solutions for a given problem instance then the aggregate performance of a collection of cooperating heuristics over a set of problem instances may be greatly improved. For this purpose we have developed a collection of fast heuristics and incorporated them in an A-team architecture, which provides a computational framework for implementing cooperation strategies among heuristics. We present results of an experimental analysis that compares the effectiveness of these heuristics working individually, and cooperating within an A-team framework. Additionally for calibration purposes we compare these results against feasible solutions derived using integer programming formulations. Since an important consideration in a real application is the computation time required to generate solu- tions, we also compare the performance of such a cooperative problem solving strategy against traditional integer programming techniques. The paper is organized as follows. We first formalize the problem in Section 2. In Section 3 we provide an overview of the related work in the literature on the two versions of the problem each with a single objective. We present our collection of heuristics and a discussion of the cooperative problem solving strategy used in Section 4. Section 5 is a summary of the results of the experimental analysis and finally we conclude in Section 6. Problem Definition We are given a set of items ng and a set of knapsacks Each item j 2 N has a weight w j and each knapsack has a capacity c i associated with it, where w j and c i are positive real numbers. In addition, for each item j 2 N a set A j ' M of knapsacks that can hold item j is specified. For convenience we also specify for each knapsack , the set B i ' N of items that can be assigned to the knapsack. For each knapsack we need to choose a subset S i of items in N to be assigned to knapsack i, such that: (1) All S i 's are disjoint. (Each item is assigned to at most one knapsack.) (2) Each S i is a subset of B i , for (Assignment restrictions are satisfied.) m. (Total weight of items assigned to a knapsack does not exceed the capacity of the knapsack.) w j is maximized. (Total weight of items assigned is maximized.) i2I denotes the set of indices of non empty due to the unused portion of each utilized knapsack is minimized.) We refer to this problem as the bicriteria sparse multiple knapsack problem (BSMK). Without loss of generality we assume that w can be removed from B i . The problem becomes trivial if all A j 's are disjoint, or if P M . In the case that all B i 's are disjoint, the problem decomposes into m single 0-1 knapsack problems. Thus, we exclude these cases from consideration. Note that the assignment restrictions can also be represented by a bipartite graph, where the two disjoint node sets of the graph correspond to the sets N and M . Let E) be the corresponding bipartite graph with . Then, there exist an edge (i; . With this representation the sparsity of the problem refers to the edge sparsity of the bipartite graph G. The bicriteria problem is more relevant for sparser problems because for more constrained problems, a solution with maximum assigned weight does not necessarily have small waste. 3 Related Problems in the Literature The single objective versions of the bicriteria problem are slightly modified forms of two well-known problems in the literature. If we consider the objective of maximizing total assigned weight alone, then the problem is a variation of the multiple knapsack problem, which we call the sparse multiple knapsack problem (SMK). For the objective of minimizing total waste alone, the problem reduces to a variation of the variable-size bin packing problem, which we refer to as the sparse bin packing problem (SBP). In the classical multiple knapsack problem, any item can go into any knapsack, hence the bipartite graph representing the problem is complete, whereas we generalize the problem by allowing any bipartite graph. On the other hand the multiple knapsack problem has a more general objective function: there exists a positive profit p j for assigning item j to any of the knapsacks and the objective function is to maximize the total profit of assigned items. In the application which motivated us, the profit of an assigned item can be assumed to be proportional to the weight of the item, hence we maximize total assigned weight. The multiple knapsack problem is known to be NP -hard in the strong sense [Kar72], [MT90]. The reduction from the 3-partition problem is still valid when the objective function coefficients equal to the weights of items, so any instance of SMK with a complete bipartite graph representation is also NP -hard in the strong sense. Thus, for the objective of maximizing total assigned weight, our problem is strongly NP -hard and there exists no fully polynomial time approximation scheme for SMK unless For the multiple knapsack problem, several exact and heuristic solution methods have been developed and tested in the literature (see Martello and Toth [MT90] for a survey). Exact solution methods consist of branch and bound, and cutting planes. The branch and bound methods use bounds based on either the Lagrangean relaxation 1 (Hung and Fisk [HF78]) or the surrogate relaxation 2 (Martello and Toth [MT80, MT81a]) of the problem. The cutting plane methods use minimal cover, (1,d)-configuration and multiple cover inequalities (Ferreira, Martin and Weismantel [FMW96]). Unfortunately, these exact solution methods cannot solve large instances arising in real applications in reasonable computation time. Heuristic methods include fast greedy algorithms followed by local exchange heuristics (Martello and Toth [MT81b]), as well as non-polynomial time approaches such as solving single knapsack problems successively (Martello and Toth [MT81a]), or obtaining a feasible solution from the surrogate relaxation (Hung and Fisk [HF78]). Considering the objective of minimizing waste alone produces an ill-posed problem with a trivial optimal solution that does not assign any items. Hence, we can consider 1 relaxing the assignment constraints decomposes the problem into m single knapsack problems. getting a linear combination of the capacity constraints results in a single knapsack problem. a version in which we impose the condition that all items in N or a specified subset of N must be assigned and the goal is to use knapsacks with minimum total capacity. Then, the problem is a generalization of the variable-size bin packing problem, where we allow assignment restrictions in addition. The bin packing problem is known to be NP -hard (i.e. [GJ79]), thus the more general problem SBP is NP-hard as well. The bin-packing problem has been extensively studied in the literature and it is one of the first problems for which efficient approximation algorithms have been developed. The recent survey by Cook, Garey and Johnson covers worst and average case analyses for online and offline algorithms. A previous survey by the same authors [CGJ84] considers also some variations on the problem. Exact algorithms have been developed by Martello and Toth [MT89]. The variable-size bin packing problem has been studied by Friesen and Langston [FL86] who provide modifications of well-known bin packing heuristics such as the next fit, first fit and best fit heuristics. Cooperative Problem Solving Given an NP-hard optimization problem, it is difficult to design heuristic algorithms which exhibit uniformly superior performance over all problem instances. An alternate approach to tackle difficult problems is to organize a collection of heuristic algorithms so that they can cooperate with each other and uniformly exhibit superior performance which might not have been possible if they were used separately. Such an approach is especially attractive when the collection of heuristic algorithms vary in their performance over problem instances in an unpredictable way. Another ingredient required for cooperative problem solving is an architecture that facilitates cooperation between the heuristic algorithms and a control strategy that defines the rules of collaboration among heuristics. In the following paragraphs we discuss in more detail the organization (i.e. the architecture and the control strategy) that we have used to build a cooperative problem solving team of heuristics for the multiple knapsack problem. We will also discuss in some detail the collection of heuristics that have been used to build the cooperative problem solving team. 4.1 The Asynchronous Teams Architecture An asynchronous team or A-Team [TdSM93] is an architecture that facilitates multiple heuristics to work together on a common problem. Cooperation between heuristics is allowed through a shared population of candidate solutions. Figure 1 provides a schematic of this architecture. The architecture is similar to a blackboard system in that the solutions are posted onto a blackboard which is shared by all the heuristic algorithms. Each heuristic has access to the entire population of solutions and can choose an appropriate (partial) solution to work on. The heuristics that are used in this architecture are usually classified into three categories based as follows: ffl Constructors are heuristics which are used to create initial solutions. ffl Improvers are heuristics which take existing partial solutions from the population and modify them to produce a new solution. The criteria used to decide whether a solution is added to the population depends on the choice of the control strategy. In purely hill climbing approaches (such as genetic algorithms) only solutions which are non-dominated would be added to the population. In variants such as simulated annealing we might allow the entry of dominated solutions into the population with the expectation that they might allow for better solutions to be created later on. ffl Destroyers are heuristics that remove redundant solutions from the population with the intent of managing the size of the population. The determination of whether a solution is redundant is difficult and usually a destroyer is designed to retain a redundant solution with a non-zero probability. Note that destroyers are typically used with control strategies that allow for the inclusion of dominated solutions into the population. Dominated Solution Constructors Destroyers Improvers Dominated Solution Non-Dominated Solution Figure 1: Schematic of an A-Team architecture The typical approach for generating solutions using this architecture involves creating an initial population of solutions using all the constructor heuristics and subsequently evolving the population of solutions by repeated application of the improver and destroyer heuristics. At any time the set of non-dominated solutions constitute a Pareto-frontier and provide a set of non-dominated solutions to the problem. The control strategy in this solution approach prescribes the rules of collaboration between the heuristics. The control strategy specifies two rules: ffl The first rule specifies how an improver heuristic picks a solution from the population to improve, and ffl the second rule specifies the criteria for incorporating a new solution created by an improver heuristic into the population. In our implementation we used a stochastic hill climbing approach for the control strat- egy. A stochastic hill climbing algorithm searches a space S with the aim of finding a state with optimal properties. The algorithm does this by making successive improvements to the current state ae 2 S. In the context of this paper, the state ae corresponds to a population of solutions to the bicriteria multiple knapsack problem. The algorithm attempts to improve the current state ae by making a transition to one of the neighbors - of ae in S. Within our A-Team implementation, this transition is made by randomly picking an improver heuristic and then randomly picking a solution from the population for the improver to work on. If the new solution generated by the improver is non-dominated then it is added to the population which corresponds to the new state - . Since the hill climbing approach does not allow non-improving moves, we do not explicitly need to worry about managing the size of the population. As a result our implementation of A-team for the bicriteria sparse multiple knapsack problem did not require destroyers. A-team architecture has been used to obtain good feasible solutions to various combinatorial optimization problems such as the TSP problem (Talukdar and de Souza[TdS93]), and scheduling problems arising in steel and paper manufacturing industry (Lee et al. [LMHM96], Murthy et al. [MARW97]). The reader is referred to Talukdar, de Souza and Murthy [TdSM93] for more on A-Team architecture. 4.2 Heuristics With the aim of generating non-dominated solutions for BSMK, we have developed a collection of constructor and improver heuristics. Most of these heuristics are simple greedy heuristics which are adapted from the heuristics in the literature used for multiple knapsack and variable-size bin packing problems. In addition, we have a few randomized heuristics. The construction heuristics are mainly greedy heuristics with various item and knapsack selection rules, in addition to a couple of heuristics that round the LP relaxation solution of SMK. Most of these constructors aim at maximizing total assigned weight. The improver heuristics are either local exchange heuristics which aim to improve both of the objectives, or heuristics which rearrange assigned items among knapsacks and unassign some items for the purpose of minimizing total waste. In the next two sub-sections we give a description of each heuristic. 4.3 Constructor Heuristics Simple Greedy Heuristics These heuristics first sort the items in non-increasing order of weight and the knapsacks in non-decreasing order of capacity. There are two versions. In the first one, the next knapsack in the order, say i, is picked, and then the next item, in the order, say j, is picked. If item j is allowed to go into knapsack j, i.e. if does not exceed the residual capacity of knapsack i, then item j is assigned to knapsack i. So, for each knapsack picked, as many items as possible are packed into it. In the second version, each item in the order is picked and assigned to the next feasible knapsack, if possible. Both of the heuristics have running time O(n log n +m log m+nm). We call these heuristics greedy-knapsack and greedy-item. Greedy Heuristics with Various Knapsack Selection Rules In these heuristics the decision to pick the next knapsack depends on the assignments made upto that point. The heuristics sort the items in non-increasing order of weight, pick the next item j and then pick a knapsack i 2 A j according to one of three rules. There are three versions based on picking a knapsack with 1) minimum residual capac- ity, residual capacity, or 3) minimum surplus demand, which are called greedy-minrc, greedy-maxrc, and greedy-minsd, respectively. Surplus demand of a knapsack i is the total weight of unassigned items in B i minus the residual capacity of the knapsack. The reason for picking a knapsack with minimum surplus demand is that the smaller the surplus demand, the more likely is the item to fit the knapsack. While greedy-maxrc and greedy-minsd heuristics tend to maximize assignments without considering waste, the greedy-minrc heuristic tends to minimize waste in addition. Successive Assignment Heuristic This is another greedy heuristic, where rather than picking a knapsack for each item or picking items for each knapsack one by one, items are matched to knapsacks successively with the objective of maximizing total weight of items matched. At each iteration a maximum weight bipartite matching (assignment) problem is solved on a bipartite graph in which edge (i; exists with weight w j only if w j does not exceed the residual capacity of knapsack i. Initially the bipartite graph G is used. The assignments given by the maximumweight bipartite matching solution are performed. Then, the bipartite graph is updated by deleting all assigned nodes, all edges (i; j) for which w j exceeds the residual capacity of knapsack j, and nodes with degree zero. The heuristic is repeated until the graph has no remaining edge. This heuristic is called successive-assign. Randomized Heuristics The greedy heuristics can be modified randomly in order to break the pattern of greedy choices. Suppose item j is picked in any of the greedy heuristics. The item will be considered for assignment with a probability p j . After running through all items, the heuristic is repeated with in order to assign the remaining items. There are two versions based on the choice of p j . In the first one p j is proportional to the weight of item j. That is, weight of an item)\LambdaC, where C is a constant factor, so that items with larger weight are more likely to be picked. We call this heuristic random-weight. In the second version, called random-degree, degree in G) C, where deg(j) is the degree of node j in G. The idea is to defer the assignment of items with more choices for knapsacks and to give priority to items that may go into only a few knapsacks. Heuristics Based on the LP Relaxation of SMK These heuristics solve the LP relaxation of the problem for the single objective of maximizing total assigned weight and then construct a feasible solution by rounding the fractional LP solution. The IP formulation of SMK is as follows. st P where the 0-1 variable x ij denotes whether item j is assigned to knapsack i. The LP relaxation corresponds to relaxing integrality of these variables. Although the LP relaxation of the multiple knapsack problem has a solution that can be constructed easily in O(n) time 3 , the LP relaxation solution of the sparse problem cannot be identified immediately. However, the relaxation can still be solved efficiently by a maximum flow algorithm. The continuous problem reduces to a maximum flow problem on a directed graph constructed from the bigraph G as follows. Each edge (j; i) of G is directed from the item node j to the knapsack node i and is assigned capacity . A source node s is connected to each item node j via an arc (s; In addition a sink node t is connected to each knapsack node i via an arc (i; t) with capacity c i . Then, the maximum flow from s to t equals the LP relaxation value and the amount of flow on arc (j; i) divided by w j gives the value of x ij . Thus, if flow on is assigned to knapsack i. If 0 1, the variable is said to be fractional (in the corresponding solution). There are two versions of the heuristic. In the first one, a simple greedy heuristic is used to assign items corresponding to the fractional variables and all the remaining 3 It amounts to solving the LP relaxation of a single 0-1 knapsack problem. items. In the second version, the fractional variables are sorted in non-increasing order of their values, and for each fractional variable in the order, the assignment is done if it is feasible. Then, the remaining items are assigned greedily as in the first version. These two heuristics are called lp-greedy and lp-round. 4.4 Improver Heuristics These local exchange heuristics aim at improving both of the objectives, and are repeated until no more improvement occurs. 1. Exchange Items Assigned to Different Knapsacks Consider all pairs of items assigned to different knapsacks. Swap the two items, if the swap is feasible and allows an unassigned item to be assigned to one of the two knapsacks. If the exchange of items is performed, pick the knapsack whose residual capacity has just increased and repetitively assign the item with maximumweight to it, if the assignment is feasible. This heuristic is called exchange. 2. Replace Assigned Items with Unassigned Items Replace an assigned item or a pair of assigned items, with a single unassigned item or a pair of unassigned items of larger weight. These heuristics are called replace-single and replace-pair based on whether a single item or a pair of items are replaced. 3. Rearrange Rearrange assigned items in different knapsacks to aggregate residual capacity into one knapsack and then use the aggregated capacity to assign new items. Swap two items assigned to two different knapsacks, if the exchange is feasible and the maximum residual capacity (over all knapsacks) increases. After all pairs of items have been considered, repeat assigning a new item with maximum weight to the knapsack with maximum residual capacity, if feasible. This heuristic is called rearrange. Heuristics to Reduce Waste 1. Empty Under-utilized Knapsacks: This is a randomized heuristic. Two parameters are picked randomly: minimum allowable utilization and maximum allowable percentage decrease in assigned weight. Cancel assignments in all knapsacks with utilization less than minimum allowable utilization, if the decrease in weight is less than the maximum allowable percentage. This heuristic is called empty. 2. Empty Knapsacks Randomly and Reassign Items For all knapsacks that are utilized, empty the knapsack with probability utilization of the knapsack). Then, reassign items by the greedy-minrc heuristic. This heuristic is called empty-and-reassign. 3. Pack Again These are variable-sized bin packing heuristics. Cancel assignments in all knap- sacks. Reassign originally assigned items by first fit decreasing or best fit decreasing heuristics (i.e. the greedy-item or greedy-minrc heuristics). These heuristics are called pack-again-ffd, and pack-again-bfd, respectively. Heuristics to Increase Assigned Weight Any of the constructor heuristics can be used as an improver to assign the remaining items, if feasible. In our implementations we used the simple greedy-item heuristic for this purpose and we refer to it as assign-remaining. Computational Experiments In this section we examine the performance of the cooperative problem solving strategy and compare its behavior against traditional integer programming based techniques for problem using computational experiments. We also compare the performance of the cooperative strategy against individual heuristics in an effort to quantify the improvements gained by cooperation. Finally we analyze the non-dominated solutions to identify concatenations of heuristics that generate good solutions. 5.1 Data We used real data from an inventory application problem in Steel Mill Industry [KDTL97]. For the instances available to us, the number of items vary between 111 and 439, while the number of knapsacks is between 18-43. The sparsity of the problems are in the range 10% - 28%. Size and sparsity of these instances are summarized in Table 1. 5.2 Implementation Individual heuristics We coded the heuristics presented in Section 4 in C++ language using the LEDA library [NU95] and performed the tests on an IBM RS4000 machine operating under AIX. First, we collected solutions output by each constructor and improver heuristic and then found the non-dominated solutions among all collected solutions. When running the improvers, we used the solution provided by the greedy-knapsack constructor Tot. Cap. Tot. Weight d3 393 d4 209 43 10.6 889.21 3528.04 d6 155 Table 1: Information on real-life data. Sparsity denotes the edge density of the bipartite graph representation in percentage of the number of edges of a complete bipartite graph. The last two columns denote the total capacity of knapsacks and the total weight of items. as input. The randomized heuristics were run times and the best solution among all runs was output. A-team We incorporated the individual heuristics into an A-team architecture. We set the parameters such as the stopping time and the probabilities for picking constructors and improvers by a few initial tests. The probabilities for some of the improvers, (such as replace-single, replace-pair, empty, and empty-and-reassign) which were more effective during initial runs were increased. We examined the convergence of the solution population by running the A-team code for a cycle of 100, 500, 1000, 1500, 2000 and 3000 heuristics. By observing the number of non-dominated solutions output, the maximum and average value of assigned weight minus waste of these non-dominated solutions, we decided on the cycle length of the A-team run for each data set. These statistics and the chosen cycle lengths are given in Table 5 in Appendix. IP based approaches In order to assess the performance of the heuristics and the computational difficulty of the problem, we tried to solve two IP's each with a single objective by a general branch-and-cut method. The first problem considered to compare our results is the sparse multiple knapsack problem, SMK, where the objective is to maximize assigned weight only. The IP formulation of SMK was given in Section 4. In the second problem considered, the objective function is the sum of the two objectives, that is we maximize assigned weight minus waste. We call this problem MKBP as it combines multiple knapsack and bin packing aspects. The IP formulation is as follows. st P where we introduce the 0-1 variable z i to denote whether any item is assigned to knapsack i. The objective function equals 2 P and the LP relaxation corresponds to relaxing the integrality of all variables. The LP relaxation of both SMK and MKBP problems have the same optimal value because an optimal solution to the LP relaxation of MKBP will have zero waste. Hence, we refer to the relaxation of both problems by "the LP relaxation". In order to obtain the best bounds possible in a reasonable computation time, we added the best lower bound obtained from our heuristics to the IP formulation of SMK and MKBP, and solved them by a general integer programming solver, CPLEX4:0:3 [CO94] using the default settings. After 1 hour of CPU time, we added the upper bound output by Cplex to the formulation and ran Cplex again (we waited longer for larger instances such as d1 and d3). We repeated this procedure until no better bounds were found in more than 2 hours. A comparison of the LP relaxation values and the upper bounds generated by a branch and bound method for data sets d1 to d6 are given in Table 2. The running times given in Table 2 also give a crude idea on the computational difficulty of solving the problems by exact methods. It took 2.4 hours to get the optimal solution to SMK for data set d2, and we could not obtain the optimal solution for any other problem. Data LP bound UB %Gap Time (hrs) UB %Gap Time (hrs) d1 641.85 639.70 0.34 9.2 639.70 0.34 2.6 d2 612.64 472.14 29.76 2.4 403.83 51.70 4.4 d3 385.37 384.99 0.10 6.2 381.14 1.11 8.0 d4 785.67 687.63 14.05 4.0 606.86 29.46 6.9 d6 444.01 424.20 4.67 8.2 414.37 7.15 5.0 Table 2: A comparison of the LP relaxation value and the best upper bound (UB) obtained by a branch and bound method for the SMK and MKBP problems. The column (% Gap) denotes the deviation in % of the LP bound from the best available bound. Time denotes the cpu time to obtain the given bounds. Here, we note that as the problem gets sparser, it gets harder in the sense that a solution that maximizes assigned weight does not necessarily have small waste, hence the choice of the knapsack to which an item is assigned becomes more critical. As the problem gets more relaxed (that is, the bipartite graph representation gets closer to a complete graph), the problem gets closer to the multiple knapsack problem and usually, maximizing assigned weight suffices to minimize waste at the same time. The sparsity of the problem plays a role also in determining the strength of the LP relaxation. As sparsity increases, the gap between the LP relaxation value and the optimal value gets larger for both problems. We also collected feasible solutions output by Cplex and recorded the cpu times to obtain the solutions. In these runs we did not provide any bounds to the objective function initially, but reran Cplex with previously obtained bounds whenever we had to stop due to memory problems. We stopped this procedure when no more improvement could be obtained till the computer ran out of memory. One may also consider using the best lower bound output by the individual heuristics in the IP formulation, as opposed to a cooperative strategy. However, using the heuristics in this way does not improve upon the performance of the branch-and-cut method of Cplex significantly. Even using the better bounds output by the cooperative strategy does not yield the optimal solution in reasonable computation time (as seen in Table 2). In addition, the cooperative strategy has the advantage of generating many non-dominated solutions for the two objectives while IP formulations specify one objective function. 5.3 Comparative Evaluation of the Cooperative Strategy Quality of Solutions The quality of the solutions generated by using a cooperative strategy are significantly better than the ones generated by individual runs, especially in the waste objective. A comparison of the solutions with maximum assigned weight generated by A-team implementation and individual runs is provided in Table 3. The waste of these solutions are also given in the table. We see that the cooperation of the heuristics have been useful to decrease the waste of the solutions that have the maximum assigned weight. We also note that we could not obtain any solutions with a better assigned weight by the exact solution method (using Cplex) for any of the problems except for d2. Solutions with maximum value of (assigned weight - waste), that are generated by the A-team implementation, individual heuristics, and Cplex are given in Table 4. We observe a significant difference in (assigned weight - waste) of the solutions generated by the cooperative strategy versus those generated by individual heuristics, especially for the harder instances such as d2 and d4. Slightly better solutions could be obtained by Cplex for only d2 and d4 in 4.4 and 6.9 hours, respectively. For larger instances such as d1 and d3, the feasible solutions output by Cplex are significantly inferior to Data AW Ratio Waste Waste % Cpu Time I 636.60 0.9952 5.25 0.82 4399.06 d1 II 617.69 0.9656 24.16 3.76 37.81 III I 470.98 0.9975 108.24 18.69 70.01 III 472.14 1.0000 110.04 18.90 8236.05 I 383.20 0.9954 5.64 1.45 9834.42 d3 II 382.33 0.9931 6.52 1.68 72.32 III 366.10 0.9509 22.74 5.85 46009.14 I 686.99 0.9990 110.97 13.91 318.17 d4 II 673.39 0.9793 155.54 18.76 3.44 III 686.99 0.9990 140.68 17.00 39923.20 I 592.33 0.9899 96.92 14.06 183.70 d5 II 590.23 0.9864 99.02 14.37 3.06 III 591.33 0.9882 97.92 14.21 39360.16 I 406.12 0.9574 30.95 7.08 76.46 d6 II 402.92 0.9498 34.15 7.81 1.63 III 402.97 0.9500 34.10 7.80 46570.32 Table 3: A comparison of the solution with maximum assigned weight obtained by I) the A-team implementation, II) individual runs of all heuristics, and III) branch-and-cut. AW is assigned weight. Ratio is the ratio of AW to the best available bound for the assigned weight objective (from Table 2). Waste % is the ratio of the unused capacity to the total capacity of utilized knapsacks in percentage. Cpu time is given in seconds. those output by the cooperative strategy with a difference of 10% and 7% of the best upper bound available, respectively. The collection of heuristics are able to generate significantly more non-dominated solutions when they cooperate in an A-team implementation. Few heuristics are able to effectively improve both the objectives at the same time. As a result, while the individual heuristics are good enough to maximize assigned weight, they are not capable of minimizing waste at the same time. On the other hand, using a cooperative organization, the heuristics which favor maximizing assigned weight and those which favor minimizing waste have the chance to take the output of one another as input. Therefore, they generate solutions with better values in both objectives. Plots of the non-dominated solutions generated by the A-team implementation and the individual heuristics are given in Figures 2, 3, and 4 in the Appendix. Note that for problems d1 and d3, which are less sparse compared to the other problems, solutions with large assigned weight have small waste at the same time, so not many non-dominated solutions were generated. This can be attributed to the fact that the two objectives are almost parallel for these problems. Data AW-Waste Ratio AW Waste Waste % Cpu Time I 631.35 0.9869 636.60 5.25 0.82 4399.06 d1 II 593.53 0.9278 617.69 24.16 3.76 37.81 III 565.62 0.8842 603.74 38.11 5.94 44543.33 I 396.72 0.9835 450.59 53.87 10.68 70.01 d2 II 326.48 0.8094 338.14 11.66 3.33 1.48 III 398.86 0.9888 471.31 72.45 13.32 15681.26 I 377.56 0.9906 383.20 5.64 1.45 9834.41 d3 II 375.81 0.9860 382.33 6.52 1.68 72.32 III 354.36 0.9297 371.60 17.24 4.43 28723.18 I 595.55 0.9814 657.99 62.45 8.67 318.17 d4 II 566.41 0.9333 649.01 82.61 11.29 3.44 I 497.14 0.9733 580.33 83.19 12.54 183.70 d5 II 491.21 0.9617 590.23 99.02 14.37 3.06 III 465.90 0.9121 564.62 99.08 14.93 23117.35 I 375.17 0.9054 406.12 30.95 7.08 76.46 d6 II 368.70 0.8898 402.92 34.15 7.81 1.63 III 365.83 0.8829 x 401.45 35.62 8.15 18203.67 Table 4: A comparison of the solution with maximum (assigned weight - waste) obtained by I) the A-team implementation, II) individual runs of all heuristics, and III) branch- and-cut. The ratio is obtained using the best available upper bound for maximizing assigned weight minus waste obtained from Table 2. Cpu time is given in seconds. Run Time Clearly, combining the heuristics by the cooperative strategy increases running time as a cycle of 1500 - 3000 heuristics are run. However, still the run times are in a reasonable range and are significantly smaller compared to that of the branch-and-cut method. A single run of each individual heuristic takes between seconds, depending on the size of the problem and the heuristic used. The running time of each individual heuristic is given in Table 6 in Appendix. The most time-consuming heuristic is the replace-single heuristic, which takes 45 seconds for d3 and only 0.24 seconds for d2. The constructor heuristics take very little time. While greedy heuristics run in less than a second, the lp-round and lp-greedy heuristics take slightly more time (0.17 - 1.31 seconds). The most time-consuming constructor successive-assign takes 0.32 to 4.25 seconds of cpu time. The A-team implementation takes a time of approximately 1 minute - 3 hours, depending on the size of the problem. Still, the run times are significantly small compared to that of the branch-and-cut approach of Cplex, which is in the order of 3 - 13 hours. All the run times are given in Tables 3 and 4. In these tables, the run time for individual heuristics is the total time over all heuristics, as the best solution was picked after running all the heuristics. Concatenation of Heuristics in the A-team Implementation By examining the non-dominated solutions on the Pareto frontier (shown in Figures 2- we can identify the heuristics and the sequence in which they were applied to yield a particular solution. Naturally the question arises whether it is really necessary to randomize the sequence in which these heuristics act on each others' solutions. If it is possible to identify one or more sequences (or concatenation of heuristics) which yield the Pareto-frontier for all the problem instances, then we might abandon a stochastic control strategy for constructing these solutions. In this section we show that the concatenation of heuristics used to generate the Pareto-frontier varies significantly by problem instance. This illustrates the need to tailor the solution strategy by problem instance which is automated by the stochastic control strategy adopted in this paper. For each non-dominated solution generated by the A-team implementation, we traced the heuristics whose output solutions were used to obtain the non-dominated solution. We can conclude that some heuristics were more effective and were repeated more. Nevertheless, we observed no regular patterns in the sequence of agents called across different problem instances. This justifies the use of an A-team approach for the concatenation of heuristics as opposed to identifying some effective patterns and using these patterns instead. For the A-team implementation, the total number of occurrences of each heuristic which participates in the generation of non-dominated solutions (for each problem instance) is given in Table 7 in the Appendix. We see that lp-round has been the most frequently used constructor. Local exchange heuristics replace-single and replace-pair, and waste reduction heuristics empty and empty-and-reassign were the most effective improvers in generating non-dominated solutions. We also give the heuristics that output the non-dominated solutions of the individual runs in Table 8 for comparison purposes. The statistics presented in these tables clearly indicate that the patterns used for constructing the Pareto-frontier varied significantly across problem instances. Figures 5-8 in the Appendix illustrate the construction of the non-dominated solutions generated by the A-team implementation. 6 Conclusions In this paper we introduced a bicriteria sparse multiple knapsack problem. We characterized the computational complexity of this problem experimentally by computing the gap of the LP relaxation to the best derivable bound and showed that increased sparsity makes the problem more difficult. Integer programming techniques were unable to generate solutions in a reasonable amount of time. A new cooperative problem solving technique was introduced which used a collection of fast heuristic algorithms embedded in a cooperative problem solving architecture called A-Teams. The solutions generated using this approach were shown to be superior to those derived from using these heuristics individually and to the feasible solutions derived from integer programming techniques. The cooperative problem solving strategy is shown as an alternative approach for automating the design of heuristics for hard combinatorial problems. --R Approximation algorithms for bin-packing: An updated survey Approximation algorithms for bin-packing: A survey Variable sized bin packing. Solving multiple knapsack problems by cutting planes. Computers and Intractibility:A Guide to the Theory of NP An algorithm for 0-1 multiple knapsack problems Reducibility among combinatorial problems. The surplus inventory matching problem in the process industry. Primary production scheduling at steel-making industries Solution of the zero-one multiple knapsack prob- lem A bound and bound algorithm for the zero-one multiple knapsack problem Heuristic algorithms for the multiple knapsack problem. Knapsack Problems. Lower bounds and reduction procedures for the bin packing problem. Leda user manual (version r 3.2). Asynchronous organizations for multi- algorithm problems Organizations for computer-based agents --TR --CTR Jayant R. Kalagnanam , Andrew J. Davenport , Ho S. Lee, Computational Aspects of Clearing Continuous Call Double Auctions with Assignment Constraints and Indivisible Demand, Electronic Commerce Research, v.1 n.3, p.221-238, July 2001
cooperation;bicriteria;multiple knapsack;multiple heuristics;asynchronous teams
595245
Lightweight Reasoning about Program Correctness.
Automated verification tools vary widely in the types of properties they are able to analyze, the complexity of their algorithms, and the amount of necessary user involvement. In this paper we propose a framework for step-wise automatic verification and describe a lightweight scalable program analysis tool that combines abstraction and model checking. The tool guarantees that its True and False answers are sound with respect to the original system. We also check the effectiveness of the tool on an implementation of the Safety-Injection System.
Introduction Recent years have seen an increasing interest in computer-supported techniques for analyzing correctness of software artifacts. In particular, this interest is caused by the effectiveness and potential of lightweight formal methods [18]. In this approach, verification consists of automated checking of an artifact against some critical properties (e.g., deadlock- security, fairness), often concentrating on debugging instead of assurance. Most often lightweight methods include model checking [5] - a technique for automatically verifying properties of a system. Given a system and a prop- erty, a model checker buildsthe reachability graph by exhaustively exploring the state-space of the system. A number of industrial model checkers have been developed, including SPIN [15], SMV [22], and MurOE [10]. Althoughmodel checking started as a technique for verifying hardware, it has been effectively applied in a variety of software projects. For ex- ample, SMV was used to verify correctness of mode logic in A-7 aircraft [25] and TCAS specifications [4]; SPIN was applied to the validation of the remote object invocation in CORBAGIOP [20], checking Java programs [14], and many others. Model checking became part of the routine V&V process during the development of Lucent's new server product [16], and has been applied to reasoning about user interfaces [12] and business processes [19]. Model checking offers a potential for push-button verifi- cation. However, this potential is not easily realizable, especially for checking correctness of programs, as opposed to specifications, protocols, or other software artifacts. First of all, model checking is mostly limited to finite-state systems (i.e., every variable in the system should have a finite domain). Several model checkers allow reasoning about infinite-state systems by "executing" all paths of the system up to a certain depth [13, 15]. However, such systems cannot guarantee that the system satisfies the desired property. To check pro- grams, an analyst has to utilize abstractions, computed either automatically or by hand [16]. And, although it is highly- desirable that properties hold on the abstracted model if and only if they hold in the original model [6], such assurance is difficult to obtain: a different abstraction has to be built for each class of properties under analysis [9]. Given a large number of available verification techniques and a potential complexity and expense of their application and interpretation of results, we propose a "layered" approach to automatic verification, depicted in Figure 1. Given a system S and a property P , we would like to know if P holds in S. We would like to start at verification level 1, which is fairly inexpensive, both in terms of the work required of the user, and in terms of computing resources required. This step will result in one of three conclusions: P is definitely true or definitely false in S, at which point the verification stops, or the analysis cannot yield any information. In the latter case, the analyst will apply a technique on verification property is true more expensive analysis property is false verification step 2 verification step 1 step N verification Figure 1: Framework for automatic verification. level 2. This technique is more expensive than that of verification level 1, but may help in determining whether or not holds. If it does not, the analyst proceeds in applying more and more complex and expensive techniques until 1) P is definitely proved or definitely disproved or 2) all levels are exhausted or all resources are exhausted. Note that no precision is lost at each level. All properties that have not been concluded to definitely hold or definitely not hold on S during the verification level have to proceed to level k. What is the benefit of the step-wise verification frame-work outlinedabove? It allows us to categorize existing tools based on their effectiveness in verifying properties and the complexity of application (this complexity metric includes the effort needed by a human and the effort needed by a com- puter). This also allows one to utilize verification efforts more effectively. In this paper, we discuss verification of sequential programs against fairly complex properties, involving temporal logic and arithmetic on values of variables, e.g. "a is never less than b", "immediately after 2a will be true". However, our approach is to build a "level 1" verifier. First, we compute the abstraction of behaviors of the program under analysis using abstract interpretation. This abstraction is not dependent on a choice of properties to verify and is computed automatically, even though the program may not be finite-state. Then we model check the abstracted system. If our analysis yields True, the property holds in the original system; if it yields False, the property does not hold, and if it yields Maybe, the analysis is inconclusive. For such cases, properties can be verified using more expensive techniques. The idea of verification with the presence of abstraction has been explored by several researchers. In particular, Jackson [17] proposed a model checking method to analyze infinite specifications expressed in Z or VDM. His approach defines an abstract state space where each abstract state represents a (possibly infinite) equivalence class of concrete states. Dams [9] demonstrated how to abstract reactive systems, so that the abstracted transition systems preserves certain forms of combined safety/liveness properties. Pardo [24] built the abstract and the concrete models of the system and conservatively verified properties expressed in -calculus on the abstracted model. If the formula is proved false, related states are then successively refined, until the given formula is verified or computational resources are exhausted. The rest of this paper is organized as follows: Section 2 gives an overview of model checking and abstract interpre- tation. Section 3 discusses the theoretical goals of this work and introduces new algorithms for our program verification system. Proofs of correctness of some of these algorithms are given in the Appendix. The design of this system is described in Section 4. Section 5 demonstrates the results of using our abstract model checker to analyze the Safety-Injection Sys- tem. We conclude the paper with the summary of this work and the outline of future research directions. Background In this section we recall the basic definitions of abstract interpretation and model checking. 2.1 Abstract Interpretation Abstract interpretation [8, 9] is a way of symbolically executing programs using abstract instead of concrete domain. Familiar data-flow analysis algorithms, e.g., constant propagation or live variables, are examples of abstract interpretation. Let D c and D a be the concrete and the abstract domains, re- spectively. The abstraction function ff : 2 D c ! D a maps a set of concrete values into an abstract value. ff has an inverse, the concretization function . An abstraction is valid if the pair of functions (ff, fl) forms a Galois connection For example, we can perform a "sign analysis" by replacing a set of integers (D signs ((\Gamma), (+), or (0)). Here, . We can execute the program on the abstract values. For example, (f\Gamma1345g \Theta f17g) ff However, the abstract values cannot always be determined exactly. Consider the following example: (f\Gamma1345g \Theta f17g can be represented as a set f(\Gamma); (+); (0)g with the interpretation that the result can be any of these values. When the abstract domain is finite, the abstract interpretor acts as a data-flow analyzer. However, we may also want to use abstract interpretationto reason about infinite-domain variables. In order to achieve tractability, we need to ensure that the abstraction is converging: 1. we have a finite representation of the infinite set of val- ues. One way is to abstract from a set to an interval by taking the minimum (maximum) value from the set as the left (right) bound of the interval. For example, 2. we ensure convergence in a finite number of steps. With a finite-domain abstraction, convergence is guaranteed. To achieve convergence for the infinite-domain abstrac- tion, [8] introduced an abstract binary operator widen- ing, denoted as r, which represents a "jump". For any abstract values i 0 and ri 1 . [8] defined widening as follows: For example, 2.2 Model Checking In this paper we will concern ourselves withCTL model checking - an automatic technique for verifyingproperties expressed in a propositionalbranching-time temporal logic called Computational Tree Logic (CTL) [5]. The system is defined by a Kripke structure, and properties are evaluated on a tree of infinite computations produced by the model of the system. The standard notation M; s indicates that a formula f holds in a state s of a model M . If a formula holds in the initial state, it is considered to hold in the model. A Kripke structure consists of a set of states S, a transition relation R ' S \Theta S, a set of initial states I, a set of atomic propositions P , and a labeling function L must be a total function, i.e, 8s 2 If a state s n has no successors, we add a self-loop to it, so that Intuitively, for each s 2 S, the labeling function provides a list of atomic propositionswhich are True in this state. We use CTL on boolean expressions that include variables and arithmetic operations on them. The set P of atomic propositions is defined as follows: If P 1 and P 2 are terms, then are atomic propositions. Terms are defined recursively as 1. Every identifier and every constant is a term. 2. If t is a term, so is (t). 3. If t 1 and t 2 are terms, so are t 1 mod and exp are the mod and the exponentiation functions, respectively. For example, the following are some atomic propositions in this version of CTL: CTL is then defined as follows: 1. Every atomic proposition 2. If ' and are CTL formulas, then so are :', ' - , The logic connectives :, - and - have their usual meanings. The existential (universal) quantifier E (A) is used to quantify over paths. The operator X means "at the next step", F represents "sometime in the future", and U is "until". There- fore, EX' (AX') means that ' holds in some (every) immediate successor of the current program state; EF' (AF') means that ' holds in the future along some (every) path emanating from the current state; E['U ] (A['U ]) means that for some (every) computation path starting from the current state, ' continuously holds until becomes true. Formally, exists some path s Note that AF (') j A[True U ']; and EF (') j E[True U '], that is, we are using a "strong until". We also use the abbreviations EG(') and AG(') to represent the property that ' holds at every state for some (every) path emanating from s 0 . EG and AG are defined as: 3 Lightweight Model Checking The goal of this work is to use abstract interpretation to alleviate the state explosion problem of model checking while ensuring that the properties verified on the abstract system can be properly interpreted in the original system. This goal is achieved by constructing an abstract model checker on our three-valued logic that returns values True, False and Maybe, such that the analysis that results in True and in False is sound. Using static analysis, we build the abstract system by associating each line of the program with an abstraction of the set of values that program variables can attain when the control reaches this point along any execution path. This abstraction, which reduces the state-space for finite-state and for infinite- state systems, is computed completely automatically. In this section we introduce the language for constructing programs, describe the process of building the labeled transition machine, and present a model checking algorithm on the three-valued logic. 3.1 The Input Language Our input language, called C-, is a sequential language with the syntax similar to C. The language includes the following constructs: booleanandinteger types; conditionalcon- trol structures (if, else); loops (while); input and output (print, fprint, scan, fscan); assignments; functions and procedures. Dynamic features such as recursion or pointers are not provided in this language. It also does not allow any user-defined (compound) data structures. A complete grammar of the language is available in [11]. Figure 2 gives an example program written in C-. 3.2 Construction of a Labeled Transition System Here, we describe the transformation of the program representation into a labeled transition system. We start with a (infinite-state) program is a (infinite) set of states, s 0 2 W is the initial state, R ' W \Theta W is the total accessibility rela- tion, and LT and LF are truth and falsity labeling functions, mapping each state to the set of propositionsthat are True and False, respectively, in this state (L T In C-, as in C, there is no one-to-one correspondence between assignments to variables and lines of code. In fact, before attempting to verify programs expressed in C-, we need to give it a well-defined formal semantics, i.e., describe the way in which each construct transforms the "program state" [23]. We define a state to be the mapping between the set of global variables and their values, and thus a state change occurs when at least one of the global variables changes its value. Our goal is to construct an abstract finite Kripke structure, in which every edge represents a state change in the program. In order to do that, we define a set of variables V and let Vw ' V be the set of variables which are accessible in the lexical scope associated with a state w 2 W , and G ' V be the set of variables which are accessible at every point of the pro- gram. Thus, G is the set of global variables. Each state w in the program is an n+1-tuple, (v )), in which ln corresponds to the line number of the state in the program, and here d i is a subset of D i - the values of the concrete domain of v i . CTL properties about such programs can be expressed only in terms of global variables. We start the analysis by parsing the program and building an Abstract Syntax Tree (AST). AST is an intermediate representation for the structure of the program under interpreta- tion. Next, we propagate information for all variables (global and local) in the current scope throughout the AST, until we reach a fixpoint. Abstractions are formed by mapping a concrete state w onto an abstract state w ff (w each d i 2 2 D i is mapped onto an abstract value D ff i . The result is an abstract state space W ff , in which each w ff 2 W ff is an n+1-tuple w Notice that line numbers and the set of variables are the same in the concrete and the abstract state space. ff is chosen so that W ff is finite, and an abstract state w ff can represent one or more or even an infinite number of concrete states due to the abstraction. Moreover, there is only one state associated with each line of code. Let R ff be a transition relation over the abstract state space, R ff ' W ff \Theta W ff . R ff is constructed as follows: (s Our labeling functions then become In the concrete domain, 8w 2 W , Under the assumptions of the Galois connection framework, an abstract system has at least as many behaviors as the corresponding concrete one. Typically, verification on abstracted systems is done either conservatively or optimistically. The former case provides "reliable negative" answers, with L ff T ' LT , and L ff F ' LF . The latter case provides "reliable positive" answers, with In either case, one side of the answer cannot be trusted. The goal of our work is to ensure that we get "reliable positive" and "reliable negative" answers, i.e., L ff T ' LT and L ff F ' LF . So, in our case, . The resulting abstract finite-state program is PG In order to construct an abstract Kripke structure in which every transition corresponds to a change to a global variable, we define a "global variable changed" predicate p on a state such that p(y) is True iff 9x; (x; y) 2 R ff , 9g 2 G, s.t. g's abstract value in state x (y). Now we construct an abstract aggregate state space S ff in which every element contains one state w ff which involves a change to a global variable, and other states that do not involve changes to global variables and can be reached from w ff via the transitive closure of R ff (denoted R ff ). s ff is defined recursively as follows: ffl if p(w ff ) is true, then w ff 2 s ff . 1: int b; 12: 2: int xy; 13: else 3: int main ( ) { 14: 4: int a; 15: if ( (a != 5: int c; 16: if ( (a != 2) && (a != 7: 8: print("xy is ", xy); 9: while { 20: print("b is ", b); Figure 2: A program fragment. Note that values of global variables within s ff are the same. We will refer to L ff T and L ff F as labeling functions that map each s ff 2 S ff to a set of atomic propositions on global variables that are true (false) in that state. Finally, the transitions between states in S ff , E ff ' S ff \Theta S ff , are thus defined as: Our abstract Kripke structure K ready. 3.3 Model Checking Algorithm We now present the algorithm that receives a Kripke structure K ff constructed above and a correctness propertyexpressed in the version of CTL described in Section 2, and determines whether or not the property holds in the system. As mentioned in the previous section, we want to ensure that our analysis yields "reliable positive" and "reliable negative" answers, i.e., if the analysis concluded that a property is True, then it holds in the original system, and if the analysis concluded that a property is False, then it does not hold in the original system. In order to do so, we introduce a third logical value Maybe. Thus, if the analysis concluded that a property Maybe holds in the system, then it is unknown whether or not the property holds in the concrete system. The algorithm recursively goes through the structure of the property under analysis, associating each subproperty ' with a pair of sets of states (Yes('), No(')). Yes(') ' S ff is a set of states in which ' is True, or, more formally, s ff 2 which represents a set of states in which ' is False, is defined similarly. We also define a predecessor function pred : 2 S ff which, given a set of states Q, returns all the states that can reach some state in Q in one transition: The algorithm, inspired by Bultan'ssymbolic model checker for infinite-state systems [3], is given in Figure 3. For exam- ple, a property p - q holds in state s ff if s ff is in Yes sets of both p and q. The same property does not hold in state s ff if s ff is in the No set of either p or q. When verifying EXp, we note that if p holds in some immediate successor of state s ff , then EXp holds in s ff ; any immediate successor in which p may hold (S ff \Gamma No(p)) should be excluded from No(EXp). A[pUq] is computed recursively as follows: A[pUq] is True in all states S 0 in which q holds; it is also True in predecessors of S 0 in which p holds and all of which successors are in S 0 . A[pUq] is False in a state s ff iff q does not hold in s ff and either p does not hold in s ff or one of its successors does not lead to q. Implementation Figure 4: Architecture of the Abstract Model Checker. We have implemented our Abstract Model Checker (AMC) as a 26,000-line C program. Figure 4 shows the architecture Procedure CHECK(') 2. Until 3. Return (Qm , 2. Until 3. Return (Qm , Figure 3: Model checking algorithm. of our software. The CTL formulas and the input language have been described in Sections 2 and 3, respectively. The Abstract Interpretor (AI) receives the program under analysis and builds the Kripke structure K using the process described in Section 3. This structure, together with a set of CTL formulas, becomes the input to the Model Checker which checks each property and returns True if the formula holds in the program, False if the formula does not hold, or Maybe if the validity of the formula cannot be established. In the latter two cases, the model checker also returns a counter-example. At the moment, the counter-example facility includes just the line number and the variable-value mappings of the states where the formula is not True. The AI receives a program and "interprets" it by starting with an input context that consists of a set of values that variables have before a program statement, executing the state- ment, and producing an output context. The output context is then stored as part of the state. The abstract values of finite- domain variables (boolean or enumerated types) consist of sets of (concrete) values these variables can attain, or UNDEF (undefined) 1 . However, values of infinite-domain variables such as integers should be abstracted further. In Section 2, we have briefly discussed how an abstraction function ff can be applied to a set to get an interval. However, for bet- 1 For brevity, we do not discuss the treatment of UNDEF here. For details, please refer to [11] ter precision, we will associate each infinite-domain variable with a (finite) set of intervals, with the following interpreta- tion In the current implementation of AI, each set can consist of up to 5 intervals. We define ff [ (union on the set of intervals) below. Let a be intervals and assume, without a loss of generality, that m - n: fag ff fag ff ff ff ng fag ff When we encounter two sets, each containing more than one interval, we first union elements of the set with less intervals (in this case, fa 1 ; :::; amg) into one interval, and then union the result with each interval of the other set. Interval operations unionand difference have their usual meaning, and widening on intervals is defined in Section 2. Other operations on sets of intervals, ff ff r (widening) are similar. Additional operations, including comparison and arithmetic, are defined formally in [11]. The algorithm used in our AI for analyzing conditional statements is depicted in Figure 5. Given an input abstract statements to execute when iexpr is True or False (stmt t and stmt f , respec- tively), we either execute stmt t (stmt f ) based on S i and then return the resulting abstract state, or call the Omega calculator to get abstract states that correspond to taking the If and the Else part (S i t and S i f , respectively), execute the state- ments, and compute the union of the resulting output con- texts. The Omega calculator is a set of C++ classes [21] for manipulating integer tuple relations and sets described by Presburger formulas. We use it for symbolically executing conditional expressions involving intervals. Procedure EVAL-IF Evaluate iexpr IF iexpr is True Execute stmt t starting with S i to get S Return S ELSE IF iexpr is False Execute stmt f starting with S i to get S Return S ELSE IF iexpr is Maybe Call Omega calculator to get S i f Execute stmt t starting with S i t to get S Execute stmt f starting with S i f to get S f Return S f Figure 5: Algorithm for analyzing conditional statements. For example, suppose we are running our AI on the program fragment depicted in Figure 2 (Figure 6 shows the control-flow graph for this fragment, with each state associated with the program line number). Let the input context before executing state 11 be ((xy; f[\Gamma20; 52]g); (a; f[\Gamma5; 8]g); (b; f[13; 2]g)). The condition xy == 0 evaluates to Maybe; therefore, we call the Omega calculator to determine that the value of xy in input contexts for states 12 and 14 should be f[0; 0]gand f[\Gamma20; \Gamma1]; [1; 52]g, respectively. The values of b in output contexts of these states are f[5; 5]g and f[26; 26]g; these are unioned to obtain f[5; 25]g in the input context to state 15. The values of a after executing state 15 and state 16 are f[\Gamma3; \Gamma1]; [1; 8]gand f[\Gamma3; \Gamma1]; [1; 1]; [3; 3]; respectively. At this point, a has reached its limit of five intervals, and further splitting cannot be done; instead, we union a's intervals to get f[\Gamma3; 8]g and proceed with the execution. This introduces a loss of information and precision, but it is strictly conservative [11]. The outputvalue for a after state 17 is f[\Gamma3; \Gamma3]; [\Gamma1; 8]g. Loops are executed until a fixpoint on values of all variables has been achieved. In order to ensure that this fixpoint occurs in a finite number of steps, we change values of variables in each loop at most 20 times, keeping track of whether they decrease or increase between iterations. If a fixpoint was not achieved, we widen values of non-converged variables, with the increase and the decrease leading to the values of +1 and \Gamma1, respectively. Afterwards, we proceed executing the loop again to ensure that dependencies between the variables are adequately captured. Table 1 lists several values that variables b and xy attain in the input context to state 9 as we execute the main while loop of the program in Figure 2. At the first iteration, these values are f[13; 13]g and \Gamma20]g, respectively. In the following19 iterations we note that the maximum values b and xy can attain are in- creasing, whereas their minimum values stay the same. Thus, the widening which occurs on the 20th iteration changes only the maximum values of these variables. The 21th iteration does not bring any further changes, thus achieving a fixpoint. Figure 7 shows the final Kripke structure built from the control-flow graph of Figure 6. Each state is associated with a line number of the statement that changes a global variable in the original program, and withthe abstract values that global variables have after the execution of this state. For example, state 10 of Figure 7 is an aggregation of states 10 and 11 of Figure 6. int xy int b 8: 10: True 11: xy == 0 14: False 15: a!=-2 True True 19: print xy int a int c 7: 9:12: 20: print b True False False False 3: MAIN END True False Figure Control-flow graph of the program in Figure 2. iteration b xy Table 1: Execution of the while loop of the program in Figure 2. The resulting Kripke structure becomes input to the model checker whose algorithm is described in Section 3. For ex- Figure 7: Kripke structure K ff built from the program fragment in Figure 2. ample, we can model check the structure depicted in Figure 7 against CTL properties and EF (b = 12). AMC returns False for the first property because it is violated in state corresponding to line 12 of the program. The second property is determined to be True because it is satisfied in the state corresponding to line 12. The third property is determined to be Maybe: it Maybe holds in the state corresponding to line 14 and does not definitely hold in any state. 4.1 Performance Given a programPG, let jV j be the total number of variables, global and local, and n be the number of statements in PG. The worst case of the AI algorithm occurs when the program has jV j loops, and each loop widens exactly one variable. We go through each loop at most 20 times; therefore, each statement in PG can be changed at most 20 \Theta jV j times, and there are n \Theta 20 \Theta jV j changes altogether. Furthermore, every state has at most predecessors. For each change of state, we union abstract values of variables of all the prede- cessors, which takes (n steps (m is a constant proportional to the number of intervals associated with each variable). Therefore, the entire computation of the abstract interpretor takes 20 \Theta jV j \Theta n \Theta m \Theta (n \Gamma 1) \Theta jV j steps, which is O(jV To compute the performance of our model checker, we let jP j be the length of a property P . Among all the CTL for- mulas, A['U/] is the most complex. For this algorithm, Q i can change value at most n times before a fixpoint is reached, and it takes steps to compute Q i 's predecessors each time. Verification of this property takes O(n \Theta (n \Gamma 1)) steps. Therefore, the total running time for our model checker to check a formula P is O(jP j \Theta n 2 ). 5 Case Study To determine the effectiveness of our abstract model checker, we analyzed the simplified version of a Safety-Injection System [7]. Safety-Injection is an embedded system that monitors the water pressure and injects the coolant into the reactor core when the pressure falls below a certain threshold. There is a manual control that the operator can use to prevent the system from injecting the coolant, which causes the system to be overridden. A reset switch prevents the system from being overridden. The system inputs the value of the water pressure and outputs a boolean condition signifying whether to inject the coolant. In addition, it maintains the internal state reflecting the water pressure. If the water pressure falls below a threshold Low, the system's pressure level becomes too low; if the water pressure raises above Permit, the system's pressure level becomes high; otherwise, this level is "within the permitted range". We have implemented the Safety-Injectionsystem as a 200- line C- program with 8 global variables closely reflecting those of the specification: WaterPres of type integer, Block and Reset of type boolean, Injection of type boolean, Overridden of type boolean, constants Low, Permit, TooLow, Permitted and High and Pressure of type integer (our system does not allow enumerated types, and the last three constants are used to indicate symbolic values of Pressure). The implementation also includes 7 functions and 8 local variables. The specification language of our system is expressive enough to capture complex properties of the Safety-Injectionsys- tem, such as 1. AG((Reset - Pressure 2. AG((Reset - 3. AG((Block - :Reset - Pressure 6= High) 4. Permitted - WaterPres - Permit) AX(WaterPres - Permit For example, Property 2 states that the system will inject the coolant if the pressure is too lowand the reset button is pressed, whereas Property 4 states that whenever the pressure is permitted and the water pressure raises above the allowed thresh- old, then the system will eventually transit into a state where the pressure is high. We verified the above properties on Sun UltraSPARC-II with 4 400 MHz processors and 4 GB of RAM. The entire verification effort, includingbuildingthe abstract Kripke structure and checking all the properties, took 3.92 seconds (user), 6.20 seconds (system). Our model-checker yielded True for each of the four properties. The final Kripke structure consisted of only abstract states. Safety-Injection has been verified by two other research groups. Bultan [2] built an infinite-state symbolic model checker that uses binary decision diagrams (BDDs) and a system of linear constraints to reason about models with Presburger arithmetic. This model checker also uses the Omega library together with abstract interpretation to achieve finite-time convergence of the analysis, but the properties are verified optimistically, allowing false positives. In addition, this procedure is partial, with the convergence dependent on the structure of the program and the formula to be verified. This approach does not utilize abstraction for state-space reduc- tion. Bultan verified Properties 1 and 2, and we were not able to determine the exact size of his models. Bharadwaj and Heitmeyer [1] analyzed SCR specifications using the SPIN [15] model checker. Their technique only allows finite-domain variables, including integer subranges and enumerated types. The size of the concrete state space is reduced by two meth- ods: eliminating variables which are not relevant to the property being verified (SCR ensures that dependencies between variables form a partial order), and by replacing input variables by predicates. The latter approach makes the verification conservative, allowing false negatives. The unabstracted system for Properties 1 and 2 (SPIN performs on-the-fly ver- ification, without building a complete state space) consists of over 1.7 million states, whereas the combination of the above abstractions brings the state space down to 650 states. 6 Summary and Future Work In this paper we proposed a framework for step-wise automatic verification and described an implementation of a very cheap and not particularly precise model checker. This model checker verifies infinite-state sequential programs written in a subset of C against CTL formulas containing arithmetic op- erations. It applies property-independent abstract interpretation to create an abstract Kripke structure, and then uses this extremely compact structure to verify properties in low-order polynomial time. No user-created abstractions are nec- essary. The verification always converges and is guaranteed to be sound: if the model checker yields True, the property holds in the concrete system, and if it yields False, the property does not hold. This approach is not limited to the analysis of programs; it can be applied to finite-state and infinite- state specifications equally well. We also believe that tightening up the code of our model checker and making the state encoding symbolic will further improve its running time. However, the results of our work are limited in several ways: (1) The implementation of the tool cannot handle complex constructs of the input language. These include recur- sion, user-defined data types, dynamic memory allocation, pointers, etc. We also currently limit our verification to sequential programs. (2) Our tool interacts with the Omega library, which can only handle operations on integer-valued variables. Thus, reasoning about floating-pointnumbers is currently not supported. (3) There is only one built-inlevel of abstraction provided in our system. (4) The input language, being a subset of C, does not have formal semantics; in particular, the notion of a state transition is poorly-defined. We chose to associate a state with values of global variables, and a state transition with changes of values of global variables. Perhaps a more flexible way to determine the granularity of state transitions is more appropriate (5) Our model checker returns Maybe if it cannot determine whether a property holds in the system. We believe we can reduce the number of cases for which the verification is inconclusive by improving the reasoning about abstract values and/or by choosing property-specific abstractions. In short-term future work we hope to extend our model-checker to reasoning about CTL [5] which combines branching-time and linear-time operations and is strictly more expressive than CTL. We would also like to address the issue of state granularity. We can do so by either asking users to specify which global variables constitute a "state" or to add language constructs for explicitly stating the beginning and the end of each state, either via begin-state/end-state or via adding the notion of time (time-tick), where each state occurs between consecutive time-ticks. Acknowledgments We would like to thank Ric Hehner and Radu Iosif for reading earlier versions of this paper, and Mark Pichora, Albert Yu and Daniel House for many interesting discussions. We acknowledge the financial support of NSERC Postgraduate Scholarship. Appendix In this section, we give proofs of correctness of algorithms for checking A['U/]. See [11] for proofs of correctness of other algorithms. Let (AM s ff )) indicate that our model checker returns True (False) when checking a formula OE in state s ff of the abstract model M ff . Assume: where M , s are the model and the state of the concrete pro- gram, respectively. The above expressions state that our model checker is sound with respect to Yes and No answers for ' and /. Prove: 1. 2. (AM 1. The proof is by induction on the length of the path from s ff to a state in which / holds. IH: Let T be the set of states from which can be reached in ! n steps. Assume 8t ff PROVE: The formula holds for paths of length - n. We will not consider the case where AM because it is covered by the base case. 2. (AM A['U/] does not hold in state s ff if either (1) / does not occur on some path emanating from s ff , or (2) on some path emanating from s ff , before the first occurrence of /, there is a state in which ' does not hold. The algorithm can be expanded as follows: Case (1): Until Case Case (1) starts from the set of states where :/ is true and recursively intersects it with those states that have successors in which :/ holds. The result is the set of states that can never reach /, i.e. M; s (2) results in the set of states in which neither ' nor / hold, i.e. M; s --R "Model Checking Complete Requirements Specifications Using Abstrac- tion" "Verifying Systems with Integer Constraints and Boolean Predicates: A Composite Approach" "Symbolic Model Checking of Infinite State Programs Using Presburger Arithmetic" "Decoupling Synchronization from Local Control for Efficient Symbolic Model Checking of StateCharts" "Automatic Verification of Finite-State Concurrent Systems Using Temporal Logic Specifications" "Model Checkingand Abstraction" "Documentation for Safety Critical Software" "Static Determination of Dynamic Properties of Programs" "Abstract Interpretation of Reactive Systems" "The MurOE Verification System" Analyzing infinite-state programs with abstract interpretation "Model Checking Graphical User Interfaces Using Abstractions" "VeriSoft: A Tool for the Automatic Analysis of Concurrent Reactive Software" "Model Checking Java Programs Using Java Pathfinder" "The Model Checker SPIN" "A Practical Method for Verifying Event-Driven Software" "Abstract Model Checking of Infinite Spec- ifications" "Lightweight Formal Meth- ods" "Model Checking for Managers" "Validation of Remote Object Invocation and Object Migration in CORBA GIOP using Promela/Spin" "The Omega Calculator and Library" Symbolic Model Checking. "An Abstract Dynamic Semantics for C" "Automatic Abstraction techniques for Propositional -calculus Model Check- ing" "Feasibility of Model Checking Software Requirements: A Case Study" --TR --CTR Steve Easterbrook , Marsha Chechik , Benet Devereux , Arie Gurfinkel , Albert Lai , Victor Petrovykh , Anya Tafliovich , Christopher Thompson-Walsh, Chek: a model checker for multi-valued reasoning, Proceedings of the 25th International Conference on Software Engineering, May 03-10, 2003, Portland, Oregon Marsha Chechik , Benet Devereux , Steve Easterbrook , Arie Gurfinkel, Multi-valued symbolic model-checking, ACM Transactions on Software Engineering and Methodology (TOSEM), v.12 n.4, p.371-408, October
program analysis;CTL;abstract interpretation;model checking
595282
A KDD framework to support database audit.
Understanding data semantics from real-life databases is considered following an audit perspective: it must help experts to analyse what properties actually hold in the data and support the comparison with desired properties. This is a typical problem of knowledge discovery in databases (KDD) and it is specified within the framework of Mannila and Toivonen where data mining consists in querying theories e.g., the theories of approximate inclusion dependencies. This formalization enables us to identify an important subtask to support database audit as well as a generic algorithm. Next, we consider the DREAM relational database reverse engineering method and DREAM heuristics are revisited within this new setting.
Introduction We are interested in understanding data semantics from real-life databases. This process is considered following an audit perspective in the following sense: it must help experts to analyse what properties actually hold in the data and support the comparison with desired properties. This research paper takes examples from relational database audit, assuming that inclusion and functional dependencies that (almost) hold in the data capture the so-called data seman- tics. This will be called hereafter the basic problem. However, our framework can be applied to other kinds of databases and/or properties. This work has been done while the author was on sabbatical year in the Department of Computer Science at the University of Helsinki (Finland). It is partly supported by AFFRST, Association Franco-Finlandaise pour la Recherche Scientifique et Technique. Auditing databases is an important topic. Integrity constraints that have been more or less explicited at the design time of a database may not always hold in a given instance. Indeed, only recent Data Base Management Systems enable to enforce important integrity constraints such as foreign keys. In most of the cases, it is assumed that application programs enforce desired integrity constraints and, obviously, it is not always done in real-life databases. Understanding data semantics in databases is of a crucial interest to support their maintenance and evolution. The fact that some property holds or not in an instance can be used by experts to fix some integrity violation in the data. It can also lead to an explicit definition of an integrity constraint for further use of built-in checking mechanisms. Improving our knowledge of encoded data semantics is also useful for semantic query optimization (see e.g., [2]). Last but not least, audit is an important preliminary step for a database reverse engineering process [5] or the design of federated databases. Indeed, solving the basic problem provides the raw knowledge that is needed to start a restructuring phase on a denormalized relational schema [13]. Auditing as querying multiple theories. Auditing databases is a typical problem of Knowledge Discovery in Databases (KDD). Discovering knowledge from databases can be seen as a process containing several steps: understanding the domain, preparing the data set, discovering patterns (e.g., dependencies), post-processing of discovered patterns (e.g., selecting dependencies that should become integrity constraints), and putting the results into use [8]. This is a semi-automatic and iterative process that is often described using ad-hoc formalisms and/or notations. However, a general KDD framework has been proposed by Mannila and Toivonen [11]. Data mining consists in querying the so-called theory of the database w.r.t. a class of patterns and a selection predicate that defines their interestingness. Audit can then be supported by queries over relevant theories. This approach emphasizes a human-centered process: expert users can precisely specify the theories they are interested in and formulate queries to learn about the properties that really hold in the data. Contribution. First, we specify auditing tasks within this general KDD frame- work. The basic problem is formalized as mining theories of approximate inclusion and functional dependencies (see Section 2). This enables us to identify an important subtask to support database audit i.e., querying an intensionally defined collection of dependencies. A generic algorithm, the "guess-and-correct" scheme introduced in [11], is a good starting point for the evaluation of such queries (Sec- tion 3). Finally, in Section 4, we revisit the heuristics that constitute the core of the DREAM reverse engineering method [16,17]. In this project, equi-joins that are performed in the application programs are used to support the discovery of "relevant" dependencies. 2. A formal framework for database audit Notations The reader is supposed to be familiar with relational database con- cepts. Suppose r is a relational database instance over the schema R. A relation belongs to R and is defined by a relation name R i and a set of attributes Each relation R i associated with a table r i which is a set of tuples. The database extension r represents the set of tables r i . r i [Y ] is the projection of the table r i on Y ' X i and t[Y ] is the projection of the tuple t following Y. Let Y and Z be two subsets of X i , a functional dependency denoted by R on R i [Z]. It can be be two relations associated with tables r i and r j respectively. Let Y (resp. Z) be a subset of attributes of X i dependency denoted by R i [Y ] ' R j [Z] is true in r i and It can be written r 2.1. Computing theories First, we introduce the KDD framework of Mannila and Toivonen [11]. Given a database instance r, assume the definition of a language L for expressing properties of the data and a selection predicate q. The predicate q is used for evaluating whether a sentence ' 2 L defines a potentially interesting property of r. Therefore, a mining task is to compute the theory of r with respect to L and q, i.e., the set Th(L; is trueg: It is possible to consider generic algorithms to compute such theories following the popular "learning as search" paradigm. A reasonnable collection of data mining tasks (association rules, sequential patterns, data dependencies, etc) have already been carried out using this approach (see [12] for a survey). Example 2.1. Consider the discovery of dependencies that hold in a database. Assume L 1 is the language of inclusion dependencies and consider q 1 as the satisfaction predicate: let r and s be the instances of R and S, and be the language of functional dependencies. Here again, the predicate q 2 is the satisfaction predicate. For instance, assume and the two following instances: Looking for a generic data mining technique, a key issue is to organize the search through the space of L sentences and get some safe pruning strategies. Here, to be safe means that we do not want to miss any interesting sentence w.r.t. the selection predicate. A simple idea is to define a specialization relation on L and then use a levelwise algorithm to compute Th(L; specialization relation - is a partial order: ' is more general than `, if ' - ' (' is more specific than '). It is a monotone specialization relation w.r.t. q if for all r and ', if In other words, if a sentence ' satisfies q, then also all more general sentences fl satisfy q. A simple but powerful "generate- and-test" algorithm can now be derived: start from the most general sentences and then try to generate and to evaluate more and more specific sentences, but do not evaluate those sentences that can not be interesting given the available information. Example 2.2. A monotone specialization relation w.r.t. inclusion dependencies is defined as follows: for have i: For instance, given i]. The most general sentences are all the potential unary inclusion dependencies and at each itera- tion, one consider "longer attribute sequences". Now, assuming the restriction to functional dependencies with a fixed right-hand side denoted as B, a monotone specialization relation w.r.t. functional dependencies is the reverse of set inclu- 5sion: for X , Y ' R and B 2 R we have . For instance, D. The most general sentences have the whole set R as the left-hand side. At each iteration, one consider shorter left-hand sides. The selection predicate q 1 (resp. It means that if R[hA; Bi] ' S[hE; F i] holds then R[hAi] ' S[hEi] holds. A safe pruning criteria is now obvious: if R[hAi] ' S[hEi] does not hold then R[hA; Bi] ' S[hE; F i] does not hold either and should be discarded at the candidate generation step. The same idea applies for functional dependencies: if AB ! D does not hold then A ! D doe not hold either and might be pruned. By alternating candidate generation and candidate evaluation, a levelwise algorithm moves gradually to the more specific interesting sentences. This has been already implemented for inclusion and functional dependency computation (see [11] for a complexity analysis and pointers to related work). It provides the best known algorithm for the discovery of all the inclusion dependencies. For functional dependencies, better algorithms are available (e.g., [9]). However, it is clear that functional dependency discovery is a very hard problem due to its inherent exponential complexity and the fact that functional dependencies with long left-hand sides are less likely to hold than functional dependencies with shorter ones. A problem with such a scheme is that the computation of the most interesting sentences in a theory can be quite slow if there are interesting statements that are far from the most general sentences (the typical case for functional de- pendencies). Furthermore, a framework designed to support (basic) audit task might consider some important specificities of this kind of application. First, one should not consider only exact dependencies: we want to study dependencies even in the case where some tuples violate these constraints. Next, the expert user is quite often interested in tightly specified subsets of the dependencies that hold. For instance, he/she just want dependencies that involve a given collection of attributes. Finally, quite often, dependencies do not have to be computed from scratch i.e., either the expert user has already a good knowledge of constraints that should hold and/or the computation of constraints has been already done on a previous state of the database. This motivates the definition of the theories of approximate dependencies, a querying approach over intensionally defined theories, and the use of a variation of the levelwise algorithm, the so-called "guess-and-correct" scheme. 2.2. Solving the basic problem Computing approximate dependencies Inconsistencies in the database can be allowed by defining q 0 ffi) to be true iff some error measure of the dependency is lower or equal to a user-defined threshold. Let us define an error measure g for the inclusion dependency where r is the instance of R. Using g enables to consider dependencies that almost hold since it gives the proportion of tuples that must be removed from r to get a true dependency. Among the several ways of defining approximate functional dependencies in an instance r of R, one can also consider the minimum number of rows that need to be removed from r for the dependency hold (the so-called g 3 error measure in [9]): Example 2.3. Assuming the instances of Example 2.1 for approximate inclusion and functional dependencies are given. Inclusion dependencies Error Functional dependencies Error The selection predicate q 1 (resp. q 2 ) can be modified to denote that all the approximate inclusion (resp. functional) dependencies whose error is lower or equal to a user-supplied threshold are desired. These selection predicates remain monotone w.r.t. their respective specialization relations. Now, one naive approach to solve the basic problem might be to compute theories of approximate dependencies for some error thresholds, store them in a "SQL3" table and then query such tables using available query languages. This is a typical approach in many KDD applications where interestingness of patterns is considered in a postprocessing phase while pattern discovery is mainly guided by simple criteria like statistical significance or error thresholds. This gives rise to several problems. The size of such theories can be huge while the expert user is interested in only a few dependencies. Not only it might be untractable to compute the whole theories but also it gives rise to tedious postprocessing phases (e.g., a posteriori elimination of redundancies). It motivates a flexible querying framework that support the analysis of tightly specified theories. Querying tightly specified theories It happens that, a priori, only small subsets of the languages of dependencies L 1 or L 2 are interesting. Restrictions of practical interest concern non trivial inclusion or functional dependencies, unary inclusion dependencies but also various selection criteria over attributes. Attribute data types and application domain semantics guide the definition of such restrictions. Example 2.4. Continuing Example 2.1, an inclusion dependency like S[hEi] ' R[hBi] can be considered as irrelevant if the corresponding domains for E and B are respectively, a collection of transaction identifiers and f1,2g to denote male or female. Also, in an application domain like relational schema restructuring, it is clear that not all the functional dependencies are interesting (see Section 4). In fact, only expert users can define such restrictions. It is possible to define them either as context-sensitive restrictions to the definition of L 1 and L 2 or by means of new selection predicates. These different views influence the computation process and its efficiency: it is more or less a "generate-and-test" scheme when the generation of candidate dependencies can make an active use of given restrictions. Notice that refining the language (re)definition is the basic technique for dependency discovery with inductive logic programming systems, the so-called declarative linguistic bias definition (see e.g., [7]). A typical audit process requires the computation of many related theories. Not only, several theories for the same dependency class are needed, depending of the dynamically evolving user's interest, but also different theories for different kind of dependencies might be useful. An obvious example is the audit of referential integrity constraints for which one must consider inclusion dependencies whose right-hand side sets of attributes are a key i.e., a special case of functional dependency. To cope with such a querying approach, the conceptual framework of inductive databases has recently emerged. It suggests an elegant approach to support audit or more generally data mining over multiple theories. 2.3. Towards inductive databases An inductive database, is a database that contains inductive generalizations about the data, in addition to the usual data [3]. The idea is that the user can then query the data, the properties in the data as well as the "behavior" of the data w.r.t. some properties. Theories are here defined intensionally. Indeed, it is not realistic to consider that querying properties can be carried out by means of queries over some materializations of every properties (e.g., data dependencies). The idea is that properties are computed only at query evaluation time, when the user is asking for some specific ones. However, during the formulation of the query, he can think that every property is just there. Formally, the schema of an inductive database is a where R is a database schema, QR is a collection of patterns, V is a set of result values, and e is the evaluation function that defines how patterns occur in the data. The function e maps each pair (r; ' i ) to an element of V , where r is a database over R and ' i is a pattern from QR . An instance (r; s) of an inductive database over the schema R consists of a database r over the schema R and a subset s ' QR . For our basic problem, we need two inductive databases that associate to a database all the inclusion dependencies and functional dependencies that can be built from its schema. We can choose that evaluation functions respectively return the g and g 3 error measures as previously defined. At each stage of manipulating an inductive database (r; s), the user can think that the value of e(r; ') is available for each pattern ' which is present in the set s i.e., that every dependency is actually in s. He/she send queries over the intensionally defined collections of all dependencies to select only dependencies fulfilling some constraints. Example 2.5. Continuing Example 2.1, a user might be interested in "select- ing" only inclusion dependencies between instances r and s that do not involve attribute R:A in their left-hand side and have a g error value lower than 0:3. One expects that a sentence like R[hC; Di] ' S[hE; F i] belongs to the answer. The definition of a concrete query language for inductive databases is out of the scope of this paper. However, let us take ideas from the SQL-like operator MINE RULE [15] and propose two queries. Example 2.6. The first part of a query specifies the kind of dependency one want to mine while the second one, that begins with the keyword FROM, defines the data in which mining is performed. The intuition is that all the power of SQL can be used for that selection of the data part. The query introduced in Example 2.5 can be written as follows: MINE INCLUSION DEPENDENCIES as IND SELECT 1.3 as LHS-IND 1.3 as RHS-IND FROM R,S The schema of the output table IND has three attributes LHS-IND, RHS-IND and ERROR that correspond respectively to the left-hand side, the right-hand side and the error measure of inclusion dependencies. Given the instances in Example 2.1, we expect that the tuple (hC; that denotes the approximate inclusion dependency R[hC; Di] ' S[hE; F i] belongs to IND. One can now search for functional dependencies in s whose left-hand sides are a right-hand side of a previously discovered inclusion dependency. Such a query might look like: MINE FUNCTIONAL DEPENDENCIES as FD SELECT 1.3 as LHS-FD 1.1 as RHS-FD The schema of the output table FD has also three attributes LHS-FD, RHS-FD and ERROR with the obvious meaning. Given the instances in Example 2.1, we expect that the tuple (hE; that denotes EF ! G belongs to FD. Evaluating this kind of query provides information about potential foreign keys Gg. Actual object-relational query languages can be used as a basis for inductive database query languages. However, non classical optimization schemes are needed since selections of properties lead to complex data mining phases. In- deed, implementing such query languages is difficult because selections of properties are not performed over previously materialized collections. Optimizing this kind of query remains an open problem for properties like functional dependen- cies. However, many inspiring ideas emerge from current research on association rule mining [3]. 3. The "guess-and-correct" generic algorithm [11] provides a generic algorithm that start the process of finding Th(L; from an initial guess S ' L. It appears as an interesting basis for query evalua- tion. Consider a set S ' L closed downwards under -, i.e., if ' 2 S and fl - ', then (by definition, this is true for Th(L; q)). The border Bd(S) of S consists of those sentences ' such that all generalizations of ' are in S, the so-called positive border Bd + (S), and none of the specializations of ' is in S, the so called negative border Bd \Gamma (S)). Bd consists of the most specific sentences in S, and Bd \Gamma (S) consists of the most general sentences that are not in S. Roughly speaking, the positive border is the collection of the sentences that are "just in" the theory while the negative border is the set of sentences that are "just off ". Example 3.1. Assume that the collection of maximal nontrivial inclusion dependencies i.e., the positive border of the theory is fR[hA; B; Di] ' S[hG; F; Ei]; R[hCi] ' S[hEi]; S[hEi] ' g. Its negative border contains many non-dependencies like R[hAi] ' S[hEi], or R[hBi] ' S[hEi]. Computing borders is not a simple task in general but it is tractable for sets of data dependencies in real-life business databases. Algorithm 3.1. The "guess-and-correct" algorithm [11]. Given, a database r, a language L with specialization relation -, a selection predicate q, and an initial guess S closed under generalizations, this algorithm outputs Th(L; 1. E := ;; // correct S downward 2. C := Bd 3. while C 6= ; do 4. 5. S := S n f' 2 C j q(r; ') is falseg; 7. od; 8. C 9. while C 6= ; do // evaluation: find which sentences of C i satisfy q: using S: 12. C := Bd \Gamma (S); 13. od; 14. output The algorithm first evaluates the sentences in the positive border Bd removes from S those that are not interesting. These steps are repeated until the positive border only contains sentences satisfying q, and thus S ' Th(L; Then the algorithm expands S upwards, it evaluates such sentences in the negative border that have not been evaluated yet, and adds those that satisfy q to S. Again, these steps are repeated until there are no sentences to evaluate. Finally, the output is q). Notice that, from the complexity point of view, the selection predicate has to be evaluated on every sentence that belongs to the border of a theory. If the initial guess and the first part of the algorithm is just skipped while the second part starts with the candidate set C containing the most general sentences of L. We get the simple levelwise algorithm that has been sketched in Section 2.1. The discovery of (approximate) functional and inclusion dependencies in a database can be solved by algorithm 3.1 given the specialization relations we introduced. Results about the complexity of such a scheme can be found in [11] and are not discussed here. However, it is clear that the better the guess is, the better the algorithm is efficient. How to obtain good original guesses S? One fairly widely applicable method is sampling: take a small sample s from r, compute Th(L; s; q) and use it as S. Another obvious situation where a guess is available is when a new audit is performed on the same database: most of the dependencies should have been preserved. Definitions at the schema level and application programs can also be used to produce a guess. 4. Revisiting DREAM heuristics This section revisits heuristics about dependency discovery for relational database reverse engineering. Given an operational database, the aim of a Database Reverse Engineering (DBRE) process is to improve the understanding of the data semantics and to support the (re)definition of a validated conceptual model. A DBRE process is naturally split into two major steps [17]: ffl Eliciting the data semantics from the existing system Various sources of information can be relevant for tackling this task, e.g., the physical schema or the dictionary, the data, the application programs, but especially expert users. Among other things, application programs might encode integrity constraints that have not been encoded at the schema level. ffl Expressing the extracted semantics with a high level data model This task consists in a schema translation activity and gives rise to several difficulties since the concepts of the original model (e.g., a relational schema) do not overlap those of the target model (e.g., an Entity-Relationship model). Many works have been done where a conceptual schema is more or less automatically derived from a hierarchical database, a network database or a relational database [1,4,16]. For relational databases, it is not realistic to assume that functional dependencies or foreign keys are available at the beginning of a DBRE process. Furthermore, the less we make assumptions on the knowledge a priori (normalization, attribute naming discipline, etc), the more we can cope with real-life databases. Several works [18,16,19] have proposed independently to fetch the needed information from the data manipulation statements embedded in application programs. In the DREAM project [16], we began to study the use of equi-joins to support 3NF schema reverse engineering. This work has been extended to cope with denormalized schema in [17]. In the spirit of [14], the DREAM approach consider the relational schemas that can be translated into conceptual schemas, by looking into the method which has been used to design them. The key problems can be resumed as follows: identifying the relevant objects of the application domain, recovering the structure of each of these objects and eliciting the links (or relationships) between these objects. This process is inherently iterative and interactive: only a part of dependency discovery can be done automatically. To cope with denormalized schemas, [17] propose a restructuring phase that leads to 3NF schemas where, according to the experts in charge of the validation, each relation maps exactly one object of the application domain. For that pur- pose, one has to find the functional dependencies which are meaningful for the application domain while they are not conceptualized as relations. Assuming that primary keys are known, the difficulty is to find out the non-key attributes that correspond to identifiers of objects of the application domain. These attributes constitute the left hand side of relevant functional dependencies and are involved in some (approximate) inclusion dependencies (foreign keys). As a matter of fact, one must support the discovery of hidden objects [10] that can even be encoded in 3NF schemas. The main contribution in the DREAM proposal has been to explore how the attributes on which equi-joins are performed help the discovery of interesting inclusion and functional dependencies for these restructuring purposes. The main result has been the following heuristics. ffl equi-joins between sets of attributes that are embedded in application programs can be used to discover "relevant" inclusion dependencies. ffl non key attributes of inclusion dependencies are good candidates for the left-hand side of "relevant" functional dependencies. These heuristics obviously reduce the number of dependencies to be considered and "relevancy" refers to the interestingness of discovered dependencies for the restructuring process. A complete scenario is considered in [17] though the following example carries out the intuition. Example 4.1. Given emp=fcode,name,tel,addg, dept=fdep,director,addg and the dependencies (1) dept[director] ' emp[code] (2) emp[add] ' dept[add] (3) code ! name, tel, add and (4) tel ! add. Dependencies (1) and (3) seem relevant for a restructuring phase while (2) and (4) are just integrity constraints. The DREAM heuristics rely on the assumption that dept.director ./ emp.code is probably performed in application programs (pointing out the potentially interesting dependency (1) and that code is a candidate for a left-hand side of a potentially interesting functional dependencies) while dept.add ./ emp.add probably do not occur. It is now clear that analyzing the set of equi-joins in application programs, enables to focus on interesting inclusion dependencies. Furthermore, it helps to fix integrity problems when we find that, e.g., R:C ./ S:E is performed while neither R[hCi] ' S[hEi] or S[hEi] ' R[hCi] hold. The collection of equi-joins can be considered as a theory Th(L; is the language of all the equi-joins between sets of attributes from r and q is the predicate that says if a given equi-join is performed on r. Computing such theories requires non trivial compilation techniques. Indeed, even if we consider only SQL queries, equi-joins can be performed in many ways, with nested or unnested queries, with a where clause or with an intersect operator, etc. Such a collection is a valuable source of information to support maintenance of application code as well as data semantics elicitation. DREAM heuristics can be encoded in queries over the inductive databases of inclusion and functional dependencies, using selection criteria derived from equi-join occurrences. The DREAM method does not claim any completness about discovered dependencies. However, it appears that by combining the different sources of information, we can compute guesses and speed up the discovery of the complete collections. 5. Conclusions We presented a framework for the audit of databases based on a KDD per- spective. It emphasizes that high-level querying tools are needed to support expert analysis of operational databases. It provides a nice application domain for an ongoing research on generic data mining tools (inductive database management systems) though it brings a solution to concrete problems of practical interest (e.g., mining approximate dependencies). We must now study typical audit tasks. For instance, supporting the elicitation of referencial integrity constraints can be useful when migrating from an old DBMS to a recent one. It needs not only to mine inclusion and functional dependencies but also to support the efficient search for erroneous data. Finally, it seems interesting to study the relationship between approximate inclusion dependency discovery and other measures of similarity between sets of attributes e.g., [6]. --R Conceptual database design: an Entity-Relationship ap- proach Discovering rules in relational databases for semantic query optimisation Querying inductive databases: a case study on the MINE RULE operator Reverse engineering of relational databases: extraction of an EER model from a relational database A Framework for the Design and Evaluation of Reverse Engineering Methods for Relational Databases Machine Learning 26 (2) Advances in Knowledge Discovery and Data Mining A method for translating relational schemas into conceptual structures Levelwise search and borders of theories in knowledge discovery Methods and problems in data mining extended entity-relationship object structures in relational schemas A new SQL-like operator for mining association rules Usign queries to improve database reverse engineering Towards the reverse engineering of denormalized relational databases An approach for reverse engineering of relational databases Reconstruction of ER schema from database applications: a cognitive approach --TR
reverse engineering;data mining;integrity constraint
595383
Robot Pose Estimation in Unknown Environments by Matching 2D Range Scans.
A mobile robot exploring an unknown environment has no absolute frame of reference for its position, other than features it detects through its sensors. Using distinguishable landmarks is one possible approach, but it requires solving the object recognition problem. In particular, when the robot uses two-dimensional laser range scans for localization, it is difficult to accurately detect and localize landmarks in the environment (such as corners and occlusions) from the range scans.In this paper, we develop two new iterative algorithms to register a range scan to a previous scan so as to compute relative robot positions in an unknown environment, that avoid the above problems. The first algorithm is based on matching data points with tangent directions in two scans and minimizing a distance function in order to solve the displacement between the scans. The second algorithm establishes correspondences between points in the two scans and then solves the point-to-point least-squares problem to compute the relative pose of the two scans. Our methods work in curved environments and can handle partial occlusions by rejecting outliers.
Introduction Previous research in mobile robotics has typically addressed several types of problems. The first is path planning in the case of a perfectly known environment and perfect sensing. The optimality criterion in this class of problems is the minimization of the cost for traversal between a start and an end position, while avoiding obstacles [18]. The second type of problem is exploration of an unknown world with perfect range sensing and odometry information [23]. Here the issues are primarily the complete coverage of the environment and the complexity of the algorithm as a function of the complexity of the environment (number of vertices and edges of objects). The third type of problem is path execution within a known real environment. The focus here is typically on the sensing required to accurately execute a preplanned path, and the key problem is robot self-localization (the "where am I" problem). One primary issue in solving this problem is how to match sensed data (vision, sonar, laser, infrared etc. ) against map information. A comprehensive collection of literature in this area can be found in [7,13]. The robot self-localization problem can be addressed either by metric means, by ensuring that the difference between the robot's actual position and the position where the robot thinks it is remains small and bounded, or by qualitative and topological means. It is well understood that odometry is not sufficient because it leads to unbounded position error [13]. A qualitative solution by tracking which side of landmark-defined lines the robot is on is proposed in [20]. If the robot is equipped with vision, then matching 3D models with 2D scene images is possible [9,16]. Because of the enormous computational requirements of using image data, using two-dimensional laser range scans has also been proposed and demonstrated successfully [5]. The robot "Blanche" assumes that a metric map of the environment consisting of polygonal obstacles is available, and it matches noisy range scans against the map frequently to keep the position error small [6]. At the heart of the method is an iterative least-squares algorithm that finds the congruence between a range scan and the map provided that the initial displacement is small. We can thus consider the problem of position estimation in a known (or partially known) two-dimensional polygonal environment as solved. Finally, exploration of an unknown world with imperfect range sensing and odometry information is addressed [17]. Here the self-localization of the robot is still an important issue. Distinct locations or landmarks detected from sonar data are used in [19]. Matching local models of line segments obtained from sonar data with a cumulative global model is adopted in [8,24]. In the context of the above literature, our work addresses the robot self-localization problem by using laser range scans, similar to the work in [5]. However, we consider more general cases as (1) an arbitrary two-dimensional world, not necessarily polygonal; and (2) an unknown environment. We intend to address the localization problem for both path execution and exploration of unknown world. The main issue in our problem is the consistent alignment of two data sets (range scans in particular). This is quite different from model-based matching [10]. Although a range scan represents a 2D shape (contour of the visible world), this shape is only represented by noisy discrete points instead of a high-quality model, which makes it very difficult to reliably define or extract features, as described in [10]. Another complication is due to the presence of occlusion, since the range scans from two different robot positions will only partially overlap. This implies that spatial correlation methods which use arclength to reference shape contours (such as [15]) can not be easily applied to scan-matching. The registration method based on fitting Spherical Attribute Images [12] also has difficulty in matching partial views. Our approach is to start with an approximate alignment of the two scans (obtained from odom- etry), and then iteratively improve the alignment by defining and minimizing some distance between the scans. We present two algorithms. The first algorithm is based on matching data points with tangent directions on the two scans. It is formulated as searching a distance function to estimate the relative rotation between the scans and using an embedded least-squares procedure to solve for the relative translation. Previous methods which iteratively match points to curves or surfaces have also been reported [1,4]. They typically solve both rotation and translation from a least-squares procedure and use fixed-point iterations to improve the solution. To ensure convergence, an initial small rotation and translation must be assumed. Our method differs from previous methods in that we do not include the non-linear rotation component in the least-squares solution and we use a search procedure rather than fixed-point iterations to minimize the distance function. Thus we are able to solve larger rotational error than the previous methods. Our second algorithm is based on iterative least-squares solutions using point to point correspon- dences, similar to the ICP algorithm proposed in [1]. But in addition to choosing the closest point on the model as the correspondence for a data point, we also choose the model point which has the same distance from the origin as the data point. Using two correspondence sets allows our method to estimate the rotation and translation very accurately and our method converges significantly faster than the ICP algorithm. The rest of the paper is organized as follows: Section 2 defines the problem as the alignment of two range scans so as to recover the relative translation and rotation of the two scans. Section 3 describes the tangent-based matching method. Section 4 describes the point-based matching method. Section 5 summarizes the integration of the two algorithms and shows experimental results from both simulated and real laser range data. Finally, section 6 concludes the paper. Problem Definition 2.1 Pose Estimation by Aligning Scans A range scan is a list of points corresponding to the intersection points of a laser beam with objects in the robot's environment. The laser beam rotates in a horizontal plane and emanates from a sensor mounted on the robot. Thus a range scan describes a 2D slice of the environment. Points of the range scan can be specified in a polar coordinate system whose origin is the location of the sensor, and the reference axis for the laser beam direction is the home orientation of the rotating range sensor. Each scan point is represented by the laser beam direction, and the range measurement along that direction. We refer to (O(x; y); ' 0 ) as the scan pose, where O is the position of the sensor in a global coordinate system and ' 0 is the sensor home orientation. Suppose that the robot starts at pose P ref (which is our reference pose) and takes a scan (call it the reference scan S ref ). The robot then moves through a static environment to a new pose P new and takes another scan (call it the new scan S new ). The approximate difference of pose P new from pose P ref (i.e. the relative translation and rotation) is usually known from odometry information. However, this information is often imperfect due to wheel slippage. Our task is to determine the exact difference of pose P new with respect to pose P ref , by aligning the two scans. The matching problem is formulated as the following: Initially assuming that the pose of S new is new , find a rotation ! and a translation T for S new such that, after the transformation, S new is aligned with S ref . 2.2 Criterion for Matching Scans A scan is a sequence of points which represent the contour curve of the local environment. Due to the existence of random sensing noise and self-occlusion, it may be impossible to align two scans perfectly. There exist two types of discrepancies between the scans. In the first type, there are small deviations of scan points from the true contour due to random sensing noise. These deviations can be characterized by an error distribution function which depends on the sensor. Another type of discrepancy is the gross difference between the scans caused by occlusion (i.e. some area is only visible in one scan but not in the other). We regard this kind of discrepancy as an outlier which cannot be described by the sensing error distribution. Naturally, we adopt a scan matching criterion as to find the best alignment of the overlapping part in the sense of minimum least-squares error, while ignoring the outlier parts. We can formulate this matching criterion as the minimization of a distance measure between the two scans, as a function of the rotation and translation. To exclude outliers in defining the distance measure, we can apply the concept of robust statistics [11,2]. 3 Search/Least-Squares Matching Algorithm 3.1 Method Overview Our approach to the scan matching problem is to define a distance measure between the two scans and search for an appropriate rigid transformation which minimizes the distance. Although the search space is essentially a three-dimensional one (rotation and two-dimensional translation), we try to reduce the search problem to an efficient one-dimensional search plus an embedded least-squares solution, by carefully formulating the distance measure. The idea of the matching method is briefly described below. First, we compute the tangent directions on the scans by fitting lines to sets of consecutive scan points. Then we associate approximate correspondences of scan points, assuming a known rotation angle ! (but not the translation T ). From each pair of corresponding points, we formulate a linear equation about the unknown translation T . Then, by using all correspondence pairs, we define a least-squares model for T which also represents a matching distance as a function of !. Outliers can be detected using gates and they contribute a constant cost to the distance measure. The final step is to search for a rotation ! which minimizes this distance function. The translation T is solved through the least-squares method. We refer to this method as rotation search/least-squares method to distinguish it from pure least-squares based fixed-point iteration type of methods. The steps of the algorithm are summarized as the following: 1. Project the reference scan S ref to the pose P 0 new so that the two scans are represented in the same coordinate system. Discard those points on S ref which are likely not to be visible from the new pose. 2. Compute the tangent directions on each scan by fitting lines to a neighborhood of sample points. Discard unreliable tangent lines (such as at corners or depth discontinuities). 3. Decide on a trial value of the rotation ! from a global search procedure. 4. For each point on S new , use the rotation angle ! to define an approximate corresponding point on S ref and compute the point through interpolation. Check the correspondence with some thresholds in order to reject outliers. 5. Use all the correspondence pairs to construct a least-squares model for T and find the least-squares solution. Define a matching distance as a function of ! from the least-squares residual and the outlier penalties. 6. Update the rigid transformation by the least-squares solution of T . 7. Repeat steps 3-6 and find the rotation ! which minimizes the matching distance function. Also obtain the overall translation by integrating the individual updates. The algorithm will be explained in detail in the following sections. 3.2 Projection of Reference Scan The reference scan S ref is originally defined at pose P ref . We would like to project it to the (approximate) new pose P 0 new to emulate the world as viewed from P 0 new . This is easily done by a change of coordinate systems for the points in S ref . After projection, we determine whether each point in S ref is visible from P 0 new based on the bounded obstacle assumption and opaque assumption. (1) Assume that initially the points in S ref are ordered by their polar angles (say counterclockwise). After projecting the points to the new pose, if the new polar angles of some points are in the wrong order (i.e. become clockwise), then the surface containing these points is facing away from the sensor and thus the points are not visible. (2) Along the rays from the new origin P 0 new to the points, if there are other points (either from S ref itself or from the new scan S new ) close enough to the rays, then the points further from the origin are considered hidden (see Fig. 1 for illustration). The points determined as nonvisible are discarded. (a) (b) occluded points new new Figure 1: Scan points are considered invisible after projection if (a) their angles are in the wrong order or (b) there are other points blocking the rays. Finally, we obtain a set of points in S ref which are ordered by angles. The projected scan can be readily used as a reference for registering S new . 3.3 Fitting Tangent Lines At each sample point on a scan, we compute an approximated tangent line by fitting to a neighborhood of sample points centered at that point. OE ae O Figure 2: Parameters of a line fit to a set of points. A line fit to a set of points in neighborhood of size n is defined by the parameters ae (normal distance from the origin to the line) and OE (direction of a normal to the line) (see Fig. 2) which minimize the following error: There exists a closed-form solution for ae, OE, and min (ae;OE) Figure 3: The points labeled by circles are discarded. Tangent lines are computed at all the other points. Only some of these tangent lines are shown for clarity. Although we fit a line at every sample point, some of the lines should not be considered as tangent lines if the world contour itself is not smooth enough near these points. Specifically, we want to exclude sample points near corners or occlusion boundaries. Two indicators can help us to recognize these non-smooth regions. First, we check the incidence angle is the polar angle of the sample point and OE is the computed normal direction, see Fig. 2). A high incidence angle indicates that either the sensing direction is nearly parallel to a surface so that the range measurements are unreliable, or there is a depth discontinuity (occlusion boundary). Another value we check is the fitting error min(E fit ) which indicates the co-linearity of the points or the smoothness of the local region. The error is usually small for straight lines or smooth curves, but large at high-curvature points or highly noisy regions. Therefore, a large fitting error usually means that the tangent line is poorly defined. We use the above two indicators to reject unreliable fitting results. Only when both values are within predefined limits, we consider the fitted line as the tangent line at the current sample point. Otherwise, we regard the tangent line to be undefined. Fig. 3 shows an example of using incidence angle and fitting error to detect points where tangent lines are unreliable. We compute tangent lines at all the other scan points. It is interesting to note that the points we discard for being unreliable in defining tangent (which are likely corners and occlusion boundaries) can be considered as features. It is possible to do feature-based matching using these points, provided that the correspondences between the features on the two scans can be determined. As the number of identifiable features in the scan is far less than the number of available data points, we believe that the feature-based solution is less accurate and less robust than our method which uses all the data points where a tangent is defined. 3.4 Correspondence Definition We define the correspondence of scan points and set up an equation based on the values of a pair of corresponding points. For the convenience of analysis, we initially regard the scans as continuous curves rather than sets of discrete points and we also ignore sensing noise or occlusion, until we instantiate the equation for the actual scan points. Once we have projected S ref to the pose P 0 new , we represent the two scans in the same coordinate system (defined by (T )). The two scans differ only by a rotation ! and a translation Let P 1 be a sample point on S new and let P 2 be the true corresponding point on S ref (i.e. they both represent the same physical point in the world). The two points are related by sin is the rotation matrix. Consider the tangent line defined on S new at P 1 and another tangent line on S new at P 2 . Let the normal directions of the two tangent lines be ~n 1 and ~n 2 , respectively. Then we can derive from Eq. 2 that Our strategy is to use Eq. 4 to estimate the translation T , given the rotation ! and the two scans. However, we note that for a point P 1 on S new , the exact correspondence point P 2 on S ref also depends on T . Therefore, we want to derive an approximated version of Eq. 4 which does not use the exact correspondence point P 2 . We will choose a point P on S ref which is close to P 2 . There are many choices of selecting P based on R ! P 1 (which is P 1 after correcting the rotation), such as the closest point from R ! P 1 or the intersection of the normal line at R ! P 1 with S ref . For the convenience of searching, we choose as the intersection of the extension of vector R ! P 1 with scan S ref (see Fig. 4 for illustration). Let ~n be the normal direction of the tangent line at P on S ref . If scan S ref is smooth and if P is not too far away from P 2 , we have the approximate relationships: (R new rotated S new O Figure 4: Illustration of point relationships. The approximation errors of the above relationships are of the order O(x 2 ) where x is related to jT j=jP 1 j (see Appendix B for derivation). Therefore, if jT j - jP 1 j and if the contour curve is smooth, the approximation error is small. We can combine the above two relationships to form a more accurate approximation whose error is of the order O(x 3 ) (see Appendix B): (R Eq. 7 is a linear equation about T in the form: It is established based on the point P 1 and the rotation angle !. Notice that, given P 1 and !, we can determine P without using T . The coefficients C x ; C y ; D can be computed from ! and the parameters of the tangent lines. Now if we instantiate Eq. 7 for n p points on S new , we can define an error function: This error indicates the "distance" between the two scans as a function of ! and T . For a given !, we can solve for T which minimizes the above distance, in terms of !. Moreover, the least-squares error min T E(!; T ) can be considered as a matching distance function which is primarily defined by the rotation !. Our final solution is based on searching the function for an optimal ! such that the least-squares error is minimum. 3.5 Correspondence Search Considering the fact that S ref is a discrete set of points, we can obtain the approximate correspondence point P for P 1 by interpolation. Let the polar angle of P 1 be ', then the point P is at direction ' !. Note that the points in S ref are ordered by their polar angles. We can locate two consecutive points on S ref whose angles enclose ' !. Then we linearly interpolate for both the range and the normal direction of P across the two points. This correspondence search is very efficient. We can implement the search for all correspondence pairs in linear time with respect to the number of points in the two scans. Correspondence points which differ greatly in their normal directions or positions are considered as outliers and they are not included in the least-squares solution. Typical causes of outliers are occlusion or poor tangent estimation. We empirically choose thresholds ff and H d . A correspondence pair is accepted only if both of the following conditions are met: (R Otherwise, we consider the pair as outlier and discard it. We also exclude the points if they are near a corner or a depth discontinuity (as detected by the line fitting procedure). 3.6 Optimization We define a total matching distance for a given rotation ! based on the least-squares error (Eq. and the number of outliers. Let n p be the number of matching pairs of points and n o be the number of outliers, the total matching distance is defined as: (min d d is the constant cost of an outlier (note that H d is the threshold that detects outliers). The effect of the thresholding and adding fixed cost for outliers to the least-squares error is approximately equivalent to using a truncated quadratic robust estimator to define the distance measure. The advantage of using a robust estimator is that a small number of arbitrarily bad outliers do not lead to an arbitrarily bad estimation [11,2]. The matching distance in Eq. 11 is a function of !. Implicitly, it is also a function of the translation T (because the least-squares solution T may be difference from the true translation). But the variable ! is dominant. We expect that the distance function should have a global minimum at the true rotation angle and that the function is unimodal near the minimum. From the plot of a typical E match (!) (Figure 5), we can see that the distance function has a single prominent valley. Furthermore, if there is little translation, it has a well-defined lowest point. An optimization procedure needs to be applied to minimize a distance function. Because the distance function is generally not a smooth one, gradient based search methods are not applicable. distance measure rotation angle (degree) Matching Distance Functions Figure 5: Matching distance measure as a function of rotation angle, in an example. The solid curve is the distance function when translation is zero. The dashed curve is the distance function when translation is (0.2, 0.1) meters (comparing with dimension of the environment of 10 meters). One possible way of searching for a minimum is to sample the search space. Stochastic methods (such as simulated annealing) may also be applied to minimize the non-smooth distance function (such as [3]). These methods are usually computationally expensive. We use the search by golden section method [14] to find the minimum along the rotation dimension in the distance function. We choose this method because of its efficiency in terms of required function evaluations. The search by the golden section method is stated as follows. Assume that the global minimum is enclosed in an interval [! The first trial point is set at x every step, a new point is chosen at x depending on whether the function has a lower value at x 1 or x 2 , the better point becomes the new x 1 and the worse one replaces ! 1 or ! 2 to form a new interval which still encloses the minimum. After sufficient iterations, the final interval will be narrow enough to localize the minimum point. Due to the residue in the translation, the optimal rotation found by the one-dimensional search may be biased. To address this problem and reduce the bias, we correct the translation by the least-squares solution every time we evaluate the distance function. As the search narrows down the rotation, the translation residue is also getting smaller and so is the bias. From the plot, we can see that the valley of the curve is within an interval of about 0.5 radians degrees) in width. Beyond this interval, the curve is mostly flat. We require this initial interval to start the search procedure. However, in case the initial rotation error is very large or completely unknown, we can determine an initial interval by coarsely sampling the function at some rotation values. For example, we can sample the function at every 15 degrees (resulting in a total of 24 samples) and choose two adjacent points which have the lowest function values to form an interval. This effectively allows our method to handle arbitrarily large rotations. The amount of translational error handled by our method depends on the threshold H d . Usually the algorithm is good enough to handle the residuals of the odometry estimates. 4 Point Correspondence Based Matching 4.1 General Approach We now present another method to align two range scans. It is an iterative algorithm and is based on associating point to point correspondence. The idea of the method is the following. For each point P i on S new , we use a simple rule (independent of the actual rotation and translation) to determine a corresponding point P 0 on S ref . Then from all the correspondence pairs of points, we compute a least-squares solution of the relative rotation and translation. This solution is applied to reduce the pose error between the two scans. We repeat this process until it converges. The least-squares solution is derived by minimizing the following distance function which is defined on n pairs of correspondence points: dist (!; T Closed-form solutions for ! and T can be derived (see Appendix C). The central issue of the algorithm is then to define a sensible rule to determine correspondences without knowing the actual rotation and translation. 4.2 Rules for Correspondence We describe the correspondence rules in the following sections. For convenience, we will regard the reference scan S ref as a continuous curve and refer to it as the model. We will consider the correspondence search on a discrete scan later. a b Figure Use rules to determine correspondences (which are joined by line segments in the figure). (a) Use the closest-point rule; (b) use the matching-range-point rule. The model is the ellipse. The data points are labeled by small circles. The big circle in the center is the origin. 4.2.1 Closest-Point Rule A commonly used rule is to choose the closest point on the model as the correspondence for a data point. We refer to this rule as the closest-point rule. Fig. 6(a) shows an example of finding the correspondences for a set of points in an elliptic model, using the closest-point rule. Besl and McKay described a general-purpose iterative closest point (ICP) algorithm for shape registration based on this rule and they proved that the ICP algorithm always converges monotonically to a local minimum with respect to the least-squares distance function [1]. We observe from experiments that, if the rotation is small, the ICP algorithm is good at solving the translation. One disadvantage of the ICP algorithm is that it converges very slowly. Especially, when the model is curved, the correspondences found by the closest-point rule may contain little information about the rotation. As seen from Fig. 6(a), the vectors joining the correspondence pairs have very inconsistent directions and they tend to cancel out each other when used together to solve for the rotation. Moreover, regardless of the type of the model, the convergence speed of the algorithm is always very slow when the distance function approachs a local minimum. To accelerate the ICP algorithm, Besl and McKay used a line search method to heuristically determine the transformation variables based on their values in two or three recent iterations. Although this improves the convergence speed near a local minimum, the problem of obtaining a poor solution for the rotation component still exists. Therefore, the improvement in convergence speed is limited. Moreover, in order to apply the line search method, an implicit assumption is made about the smoothness of the least-square distance function. But this is typically not true if the number of correspondence pairs changes during the iterations (as a result of rejecting outliers). O model Figure 7: Matching-range-point rule: For a point P , the corresponding point P 0 on the scan lies within the sector and jOP 0 j is closest to jOP j. 4.2.2 Matching-Range-Point Rule We propose a different rule which finds the correspondences in such a way that they significantly reveal the rotation component. Consider a data point P and its corresponding point . If we ignore the translation, we have jP j. On the other hand, the polar angle ' of P and the polar angle - ' of P 0 are related by - This implies that the correspondence of P under a rotation is a point which has the same polar range as that of P , and the polar angles of the corresponding points differ by the rotation angle !. Now in the presence of a small translation, we can still expect that the point P 0 with the same range as that of P is possibly a good approximation to the true correspondence of P , and this approximate correspondence provides rich information about the rotation angle !. To ensure that the rule finds a unique and reliable correspondence, we only search for the matching-range point within a local region of the model near P . Suppose that we can estimate a bound B ! for the rotation !, i.e. j!j means that P 0 should lie within the sector bounded by ' . Therefore, we propose the matching- range-point rule as the following: For a data point P , its corresponding point is P 0 on the model is closest to jP j. The rule is illustrated in Fig. 7. Fig. 6(b) shows an example of using the matching-range-point rule to find correspondences in the elliptic model. We can clearly see that the vectors joining the correspondence pairs consistently indicate the rotation direction. Therefore, the least-squares solution should be a good approximation to the true transformation, especially in the rotation component. Based on this new rule, we design an iterative matching-range-point (IMRP) algorithm. In this algorithm, the parameter B ! controls the size of the neighborhood to be searched for a correspondence and also the maximum rotation possibly solved in one iteration. Thus it will be best to choose B ! to be close to the rotation residual at every iteration. We empirically generate using an exponentially decreasing function of the (0)e \Gammafft . It is also possible to use the least-squares solution of ! in the current iteration to determine the value of B ! for the next iteration. A comparison of the performance of the ICP algorithm and the IMRP algorithm is illustrated in Fig. 8. We notice that the IMRP algorithm converges faster than the ICP algorithm in estimating the rotation. For the translation part, the IMRP algorithm is initially slower but it eventually converges faster than the ICP algorithm. The reason that the translation residuals from the IMRP algorithm are initially reduced slowly is that the matching-range-point rule only tends to influence the translation component when the sector becomes small enough. In the early iteration when B ! is large, the algorithm tends to explain the displacement between the correspondence pairs by rotation rather than translation. This phenomenon may present a potential problem to the stability of the algorithm if we need to dynamically reject outliers using a threshold, as good correspondences may be falsely rejected as outliers due to the incorrect translation. We will study more about outlier detection later. 4.2.3 Combining the Two Rules It is desirable to combine the two rules in order to achieve both the convergence speed of the matching-range-point rule and the stability of the closest-point rule. We propose an iterative dual correspondence (IDC) algorithm which uses both rules, as the following: 1. In each iteration do the following steps: 2. For each data point (a) apply the closest-point rule to determine a correspondence point P 0 for (b) apply the matching-range-point rule to determine a correspondence point P 00 for 3. Compute the least-squares solution (! from the set of correspondence pairs (which are obtained using the closest-point rule). 4. Compute the least-squares solution (! from the set of correspondence pairs (which are obtained using the matching-range point rule). 5. Form (! as the solution for the transformation in the current iteration. The basic idea of the above algorithm is to take the translation component from the closest-point rule solution and the rotation component from the matching-range-point rule solution to form the current solution for the transformation. The combination of the two rules appears to achieve significantly better results than each of the individual rules. It does not only ensure the stability of the iterations, but also increases the convergence speed significantly. In fact, the two rules reinforce each other by reducing different components of the transformation, so that each of them can be more effective in the following iteration. We notice that the IDC algorithm is insensitive to the choices of parameter B ! . Fig. 8 illustrates the residuals of the transformation components during the iterations of the three algorithms. The model and the data points in this example are the ones given in Fig. 6, where the initial rotation is \Gamma6 degrees; the initial translation is (-5, 5) units (the width of the elliptic model is 1000 units). Clearly, the IDC algorithm reduces the residuals much more quickly than the other two single-rule algorithms. We experimentally estimate the rate of convergence for each of the three algorithms using the above example. The iterative closest point algorithm appears to have a sublinear convergence rate. The ratio of error residuals larger and larger and it approaches to 1. In fact, after iterations, C is at 0.998. For both the iterative matching-range point algorithm and the iterative dual correspondence algorithm, the rate of convergence seems to be linear. The error ratio for the iterative matching-range point algorithm is 0:875; the error ratio for the iterative dual correspondence algorithm is 4.3 Correspondence Search We will use the IDC algorithm to register a new scan S new to the reference scan (model) S ref . Considering that S ref is discrete, we need to interpolate in order to locate the correspondence point for each of the two rules. First, we consider the matching-range-point rule. Let P (r; ') be a point on S new . We want to find another point according to the matching-range rule. Since the points in scan S ref are parameterized in a polar coordinate system, we choose to linearly interpolate 1=r from ' between two points. be two adjacent points, the interpolation function - r( - is: This interpolation scheme is approximately equivalent to connecting the two points with a line degree Iterations Rotation Residuals IMRP algorithm IDC algorithm2610 cm Iterations Translation Residuals IMRP algorithm IDC algorithm a b Figure 8: Comparison of iterative algorithms. (a) Rotation residuals from the three algorithms over iterations (note the curve for the IDC algorithm is barely visible as it is almost zero after three iterations). The initial rotation angle is 6 degrees. (b) Magnitudes of translation residuals. The magnitude of initial translation is 7.07 units. The model and data set are given in Figure 6. segment if j' small. This can be seen by comparing Equation 13 with the equation of a straight line passing through P 1 and The interpolation in Equation 13 also has the advantage that the interpolated - r is a monotonic function of - '; it is easy to inversely compute - ' for a given - r. The region where we need to search for P 0 consists of pairs of points whose angular intervals overlap with be one such interval. We find the intersection [ - we use Eq. 13 to interpolate two points P 0 at the ends of the new interval. If - one of P 0 1 and P 0(whose polar range is closer to r) is a candidate for P 0 . In this case, we still need to check other intervals for a potentially better candidate. But if - r 1 - r - r 2 , or - , we can (inversely) interpolate for a P 0 which has the same range as P . Now we consider the closest-point rule. For consistency with the matching-range rule, we slightly modify the closest-point rule such that it chooses the closest point P 0 within the sector [' \Gamma only as the correspondence of P . The interpolation here is simply connecting two adjacent points with a line segment. In a similar way as in the search for a matching-range correspondence, we search all the angular intervals and find a P 0 by checking the distance from P to the line segments (or end-points). The search required to determine all the correspondence pairs can be efficiently implemented in linear time with respect to the number of points on the two scans, by considering the fact that all the points in the scans are ordered by their angles. Due to occlusion, we may not find appropriate correspondence points for some data points. We need to identify these outliers and exclude them from the least-squares distance function. Assume that we have a bound B r for the maximum displacement resulting from translation for a pair of corresponding points P and P 0 , i.e. jjP . We use this condition to test for outliers. For a pair of points found by one of the rules, we accept them as a correspondence pair only if the above condition is met. The threshold B r can be selected as the kth largest distance among all the correspondence pairs, according to a predetermined constant fraction (p-tile). 4.4 Convergence of Iterative Algorithm The iterative algorithm terminates when the change in the least-squares errors in sufficiently small. In practice, we find it sufficient to fix the number of iterations to about 15-20. The iterative process in the algorithm resembles fixed-point iterations. Since the zero pose error is a stable fixed-point for the process, we expect that once the iteration converges that the final solution should be very accurate. To ensure convergence a good initial estimate (small initial pose error) is required by the algorithm. Usually, we can use odometry to provide this initial estimate. Another possibility is to first apply the rotation search/least-squares method which we presented in Section 3 to obtain a relatively good registration before applying the point-based method. We find from experiments that this strategy usually guarantees the convergence of the point-based method and the point-based method also usually improves the accuracy of the solution. 5 Scan Matching Experiments 5.1 Combining The Two Algorithms We have presented two algorithms, the rotation search/least-squares algorithm and the iterative point correspondence algorithm, for matching range scans. The rotation search/least-squares uses a search procedure in the optimization process. This allows the algorithm to robustly solve for the transformation even in the presence of a large initial pose error. However, because the matching distance function is not smooth and we only use a simple one-dimensional search procedure to do the minimization, the solution given by this algorithm may not be highly accurate. On the other hand, the iterative point correspondence based algorithm gives a more accurate solution through its fix-point iterations, as long as it converges. This natually leads to a two-staged algorithm to solve a general matching problem by sequentially combining the two algorithms. We apply this algorithm in the following experiments. 5.2 Sensing Strategy If we have control over the sensing directions (as is the case with the ARK sensor [25]), we can choose to take a range scan in such a way that the sampling points are evenly distributed in space. In other words, we want the distances between adjacent points to be approximately equal. This is more efficient than a uniform sampling of orientation where the interval width is proportional to the ranges of the points from the sensor. Assume that the current sample point is P ('; r). We will decide the direction of the next shot, \Delta', such that the new point is a fixed distance d away from P . If we approximate the contour curve near P by its tangent line (which has parameter (ae; OE)), the increment in sensing direction is: where the angle ' \Gamma OE can be computed by the following: dr d' Note that the derivative dr d' can be estimated from recent range measurements. Combine the above two equations and take a first order approximation (for a small d), we get: \Delta' - d d' To avoid infinitely small increments, we can predefine a minimum value for \Delta'. Now ' + \Delta' is the direction to take the next measurement. This sensing strategy is used to generate both simulated and real scan data in our experiments. 5.3 Experiments with Simulated Data Range measurements are simulated for various environments in order to test the matching algo- rithms. We model the environments with lines and spline curves. Sensing noises are modeled as additive random variables. For a given environment model, we can simulate a range scan from any robot pose. We present a series of experiments using simulated range data. In each experiment, we choose an environment model and two poses, P ref and P new , from which to take the scans S ref and S new . The pose error is generated randomly where ! is uniformly distributed in [-0.25, 0.25] radians uniformly distributed in a disk of radius 50 centimeters (comparing with the dimension of the environment of 10 meters). Sensing noise is assumed to be uniformly distributed. We vary the maximum sensing noise in each experiment. In each experiment setting, the matching process was run 1000 times (with randomly generated initial pose error and sensing noise as described above). We computed the standard deviations (from the theoretical zero means) of the residuals of the variables (!, T x , and T y ) resulting from each matching process. The standard deviations of the residuals after the first stage (rotation search algorithm) and after the second stage (iterative point correspondence algorithm) were recorded. In the matching process, the number of iterations was fixed at 15 in each of the two stages. The results from six experiments are listed in Table 1. In each experiment, we list the environment model, the maximum sensing noise, and the standard deviations of the residuals. We also plot all the translation residuals from each of the six sets of experiments (Fig. 9). Each translation residual is shown as a dot in the x-y plane. The dots form a cluster around the origin. We have some observations from the statistical results. (1) The pose error residuals are very small compared with the sensing noise. The standard deviation of translation residuals is about 15% to 25% of the standard deviation of sensing noise. The rotation residuals are well within one degree. With a typical sensor accuracy of 5cm, the rotation residuals are within one tenth of a degree. (2) The algorithm behaved robustly over thousands of runs. (3) The second algorithm (iterative point correspondence algorithm) gave significantly more accurate estimates of rotation than the first algorithm. But the two methods are equally good in finding the translations (the second algorithm is slightly better when the sensing noise is relatively small or when the environment is mostly linear (experiments 1, 2, 3). (4) The residuals appear to be normally distributed with approximately zero means. (5) When the sensing noise becomes large (experiments 5 and 6), the algorithm degrades gracefully. In experiment 6 (where the sensing noise is set to 20cm), it appears that the limit of the algorithm is reached. We experienced a few failure cases where the iterative algorithm did not converge. There were about 10 such cases among the 1000 runs. The failure is partly caused by the step of rejecting outliers using thresholds. Raising the threshold level may improve convergence. But it may result in bias in the estimation because of outliers. High sensing noise also leads to bad estimation of local contour directions. Thus the correspondence association step (which involves interpolation) is subject to large error. We plot three typical matching examples from the above experiments. In each example, we show the two scans which are misplaced with respect to each other. Then the same two scans after alignment are shown. In the figures, the reference scan is plotted with x's. The big X at the center is the actual pose P ref of the reference scan. The new scan to be registered is plotted with circles. The big circles with arrows are the poses of the new scan before and after registration (that is, new and P new ). Example 1 (Fig. 10) is a typical case from experiment 3. Here the rotation is 0:25 radians (14:3 2.1811 cm 2.1961 cm 2.5478 cm 2.2832 cm 1.1438 cm 1.1269 cm 1.2604 cm 1.2268 cm After Stage 2 After Stage 1 noise: noise: noise: noise: noise: noise: oe x oe y oe x oe x oe y oe x oe y oe x oe y oe x oe y oe y 1.5336 cm 0.8532 cm 0.7836 cm 0.8446 cm Residual Standard Deviations Simulated Environments and Maximum Sensing Noise Table 1: Statistics of experiments in simulated environments. Maximum initial rotation and translation are set at \Sigma14:3 ffi and 50cm, respectively. cm cm Translation Residuals in Experiment #1 cm cm Translation Residuals in Experiment #2 a b cm cm Translation Residuals in Experiment #3 cm cm Translation Residuals in Experiment #4 c d cm cm Translation Residuals in Experiment #5 cm cm Translation Residuals in Experiment #6 Figure 9: Translation residuals after apply the matching algorithm. (a) to (f) correspond to experiments 1 to 6 in table 1. There are 1000 samples in each experiment. Larger spread of dots in the plot corresponds to higher levels of sensing noise. Figure 10: Example 1 (from experiment 3, maximum sensing noise is 10cm). The robot went from pose A to pose B in reality, but due to odometry errors, it thinks it went to pose B 0 . Points on scans from pose A and B are labeled as x's and small circles respectively. The poses are indicated by arrows. Part (a) shows the alignment error due to the difference between pose B 0 and true pose B. Part (b) shows the result of aligning the two scans. The pose B 0 is corrected to the true pose B by the same transformation. Notice the large occlusion in the lower right corner. Figure (from experiment 4, maximum sensing noise is 10cm). Part (a) shows the pose error (indicated by the difference from B 0 to B) and the resulted alignment error of the scans. Part (b) shows the result of correcting the pose error and aligning the two scans. The large rotation and translation are corrected successfully. There is occlusion in the scan. A A a Figure 12: Example 3 (from experiment 6, maximum sensing noise is 20cm). Part (a) shows the pose error (indicated by the difference from B 0 to B) and the resulted alignment error of the scans. Part (b) shows the result of correcting the pose error and aligning the two scans. This level of sensing noise is about the limit of algorithm capacity. and the translation is (\Gamma30; \Gamma20) centimeters. Notice that there is a substantial portion of the world (about one third) which is visible only in one scan but not in the other. Our algorithm is able to ignore the unmatched parts but successfully perform the matching based on the remaining part of the world which is visible in both scans. Example 2 (Fig. 11) is from experiment 4, in which the environment consists of mostly curves (rotation 0:5 radians (28:6 ffi ), translation (\Gamma60; 40) centimeters). There also exist occlusions in the scans. Our matching method is successful. Example 3 (Fig. 12) is chosen from experiment 6, in which the maximum sensing noise is as large as \Sigma20 centimeters. Rotation is \Gamma0:2 radians (\Gamma11:5 degrees) and translation is (\Gamma40; 30) centimeters. The correct alignment is found by the algorithm in this example. Note this level of sensing noise is close to the limit of algorithm capacity. In the above experiments, the second algorithm (point correspondence based method) is not fairly tested as its input (from the output of the first algorithm) is too good. Here we present another experiment which examines the iterative point correspondence algorithm without first applying the rotation search method. We use the same environment as the one in experiment 2, but set the maximum sensing noise to \Sigma10cm. The initial pose error is randomly generated so that ! is uniformly distributed over [-0.1, 0.1] radians and T is uniformly distributed in a disk of radius centimeters. We ran this experiment 1000 times. The standard deviations of the residuals are listed in Table 2. For comparison, we also run the rotation search algorithm using the same input set. We can see that the second algorithm is still successful despite of the relatively larger initial pose errors. Again, the second algorithm performs significantly better than the first one in finding the rotation, but only slightly better in finding the translation. Residual First Algorithm Second Algorithm Rotation Translation x oe x 0.7652 cm 0.7827 cm Translation y oe y 0.7998 cm 0.6514 cm Table 2: Standard deviations of the residuals from the two algorithms when they are run independently with the same set of input. Maximum sensing noise is \Sigma10cm. Maximum initial rotation and translation are \Sigma5:7 ffi and 20cm, respectively. 5.4 Experiments with Real Data In this section, we present the experimental results using real range sensor data in two different environments. In the first experiment, we used the ARK robot [25] to collect range scans in the Vision Lab at the Department of Computer Science, University of Toronto. The ARK robot is equipped with a laser rangefinder (model Optech G150) mounted on a pan and tilt unit on the robot platform. The accuracy of the rangefinder is 4 centimeters. We present two examples of matching scans from this environment (Figures 13 and 14). Relatively large pose errors between the two scans were intentionally introduced. Unfortunately, the ground truth for this experiment is not available as the robot was driven by hand. Our matching algorithm successfully aligned the scans in both examples. Notice that in performing the matching the algorithm is able to ignore the cluttered areas at the right-side wall and the bottom left corner (which are bookshelves). The algorithm also handled the occluded area well. Another testing environment is the cafeteria and nearby corridor in FAW, Ulm, Germany. The laser rangefinder used was a Ladar 2D IBEO Lasertechnik mounted on the AMOS robot. The laser sensor has a maximum viewing angle of 220 degrees, and an accuracy of \Sigma20mm. We obtained scans which were taken by the robot along its path at a pace of approximately 2 meters between scan poses. We then selected 84 pairs of scans which have sufficient overlap and applied the scan matching algorithm to each pair. Our matching algorithm successfully aligned the scans in all 84 cases. All of these scan matching results were used by a global scan registration algorithm which is discussed in [22,21]. Two examples of the matching results are shown in Figures 15 and 16. In both examples, we exaggerated the pose errors for testing the algorithm. The odometry error from the actual robot is much smaller. Note that this robot samples orientation uniformly when taking the scans, as opposed to our early strategy of making the sample points uniformly dense. But the matching algorithm still works well with this kind of scan. We also include another example using real range data from a hallway (Figure 17). Notice that the constraints along the hallway direction is weak (the only reference are several shallow doorways). Our algorithm is able to derive the correct matching. It appears that the real sensor data are more accurate than our typical simulated data. After thoroughly testing our algorithm with simulations, we believe that it should also work well in real indoor environments of this type. Moreover, in a typical indoor robot course where range measurements are taken at short discrete steps, the pose errors are usually much smaller than the ones we simulated. As a conclusion from the experiments, we believe that our algorithm should work robustly in real-life robotic applications. 6 Conclusion We have developed two methods for matching two range scans and correcting the relative pose error. The two methods can either be used individually or applied in sequence. Our matching methods have the following desirable properties: Figure 13: Example of matching scans taken by the ARK robot at U of T. Figure 14: Another example of matching scans taken by the ARK robot. Figure 15: Example of matching real scans at FAW, Ulm, Germany. The scans are shown in dots. Part (a) shows the pose error (indicated by the difference from B 0 to B) and the resulted alignment error of the scans. Part (b) shows the result of correcting the pose error and aligning the two scans. Figure Another example of matching real scans at FAW, Ulm, Germany. Figure 17: Example using real data from a hallway. Notice that the constraints are weak along the hallway direction. Alignment still can be made using our algorithm. 1. Both methods can handle sensor noise and partial occlusion. By checking against thresholds, the methods can effectively reject possible outliers. The methods are not very sensitive to the choice of threshold values. 2. The methods do not rely on distinguishable features in the environment. Thus we avoid the difficult process of feature detection and feature correspondence. The availability of the number of features should not concern us either. In fact, we use most of the sensor data in the matching process. This suggests that our algorithms are robust. 3. We do not require an a priori world model. We can match a scan directly to another scan. Therefore our method can be used for exploration and map building in unknown environments. 4. Our method can work in a general curved environment. We do not require the world contours to be linear, as many other methods do. 5. The algorithm is fast (linear time in the number of sensor data points per iteration, for a small fixed number of iterations). There is no exponential searching needed in finding the primitive correspondence. 6. Our algorithm has the ability to correct arbitrarily large errors in rotation. 7. Our algorithms are demonstrated to be superior over similar existing methods in robustness and efficiency. We have tested our algorithm thoroughly with simulated data under various adverse conditions. We also tested the algorithm with real data from two different environments using different sensors. Acknowledgements Funding for this work was provided by NSERC Canada and by the ARK project which receives its funding from PRECARN Associates Inc., the Department of Industry, Science and Technology, NRC Canada, the Ontario Technology Fund, Ontario Hydro, and AECL. The authors would like to thank Steffen Gutmann, Joerg Illmann, Thomas Kaempke, Manfred Knick, Erwin Prassler, and Christian Schlegel from FAW, Ulm for collecting range scans and making the data available for our experiments. --R A method for registration of 3-D shapes Robust Incremental Optical Flow. Registering multiple range data to create 3d computer objects. Object modelling by registration of multiple range images. Blanche: Position estimation for an autonomous robot vehicle. Blanche: An experiment in guidance and navigation of an autonomous robot vehicle. On the congruence of noisy images to line segments models. World modeling and position estimation for a mobile robot using ultrasonic ranging. Object Recognition by Com- puter: The Role of Geometric Constraints Robust Statistics: The Approach Based on Influence Functions. Building 3-D models from unregistered range images Autonomous Mobile Robots Numerical Methods and Software. Fast vision-guided mobile robot navigation using model-based reasoning and prediction of uncertainties A robot exploration and mapping strategy based on a semantic hierarchy of spatial representations. Robot Motion Planning. Dynamic map building for an autonomous mobile robot. Qualitative navigation for mobile robots. Shape Registration Using Optimization for Mobile Robot Navigation. Optimal global pose estimation for consistent sensor data registration. A comparative study of the path length performance of maze-searching and robot motion algorithms positioning using model-based maps Design of ARK --TR --CTR Annalisa Milella , Carmine Dimiccoli , Grazia Cicirelli , Arcangelo Distante, Laser-based people-following for human-augmented mapping of indoor environments, Proceedings of the 25th conference on Proceedings of the 25th IASTED International Multi-Conference: artificial intelligence and applications, p.151-155, February 12-14, 2007, Innsbruck, Austria Thomas Kmpke, Convex Translation Estimation, Journal of Intelligent and Robotic Systems, v.21 n.3, p.287-300, March 1998 Tom Duckett , Stephen Marsland , Jonathan Shapiro, Fast, On-Line Learning of Globally Consistent Maps, Autonomous Robots, v.12 n.3, p.287-300, May 2002 Amalia Foka , Panos Trahanias, Real-time hierarchical POMDPs for autonomous robot navigation, Robotics and Autonomous Systems, v.55 n.7, p.561-571, July, 2007 F. Lu , E. Milios, Globally Consistent Range Scan Alignment for Environment Mapping, Autonomous Robots, v.4 n.4, p.333-349, October 1997 Juan Nieto , Tim Bailey , Eduardo Nebot, Recursive scan-matching SLAM, Robotics and Autonomous Systems, v.55 n.1, p.39-49, January, 2007 Thomas Kmpke , Matthias Strobel, Polygonal Model Fitting, Journal of Intelligent and Robotic Systems, v.30 n.3, p.279-310, March 2001 Stergios I. Roumeliotis , Ioannis M. Rekleitis, Propagation of Uncertainty in Cooperative Multirobot Localization: Analysis and Experimental Results, Autonomous Robots, v.17 n.1, p.41-54, July 2004 O. Horn , A. Courcelle, Interpretation of Ultrasonic Readings for Autonomous Robot Localization, Journal of Intelligent and Robotic Systems, v.39 n.3, p.265-285, March 2004 Wolfgang Hbner , Hanspeter A. Mallot, Metric embedding of view-graphs, Autonomous Robots, v.23 n.3, p.183-196, October 2007 Christian Frh , Avideh Zakhor, An Automated Method for Large-Scale, Ground-Based City Model Acquisition, International Journal of Computer Vision, v.60 n.1, p.5-24, October 2004 John Mullane , Ebi Jose , Martin D. Adams , Wijerupage Sardha Wijesoma, Including probabilistic target detection attributes into map representations, Robotics and Autonomous Systems, v.55 n.1, p.72-85, January, 2007 Kin Leong Ho , Paul Newman, Detecting Loop Closure with Scene Sequences, International Journal of Computer Vision, v.74 n.3, p.261-286, September 2007 Urbano Nunes , Jos Alberto Fonseca , Lus Almeida , Rui Arajo , Rodrigo Maia, Using distributed systems in real-time control of autonomous vehicles, Robotica, v.21 n.3, p.271-281, June Ioannis Rekleitis , Gregory Dudek , Evangelos Milios, Multi-robot collaboration for robust exploration, Annals of Mathematics and Artificial Intelligence, v.31 n.1-4, p.7-40, 2001 Frank Dellaert , Michael Kaess, Square Root SAM: Simultaneous Localization and Mapping via Square Root Information Smoothing, International Journal of Robotics Research, v.25 n.12, p.1181-1203, December 2006
rotation search;robot pose estimation;reference scan;aligning scans
595863
Temporalizing Epistemic Default Logic.
We present an epistemic default logic, based on the metaphore of a meta-level architecture. Upward reflection is formalized by a nonmonotonic entailment relation, based on the objective facts that are either known or unknown at the object level. Then, the meta (monotonic) reasoning process generates a number of default-beliefs of object-level formulas. We extend this framework by proposing a mechanism to reflect these defaults down. Such a reflection is seen as essentially having a temporal flavour: defaults derived at the meta-level are projected as facts in a next object level state. In this way, we obtain temporal models for default reasoning in meta-level formalisms which can be conceived as labeled branching trees. Thus, descending the tree corresponds to shifts in time that model downward reflection, whereas the branching of the tree corresponds to ways of combining possible defaults. All together, this yields an operational or procedural semantics of reasoning by default, which admits one to reason about it by means of branching-time temporal logic. Finally, we define sceptical and credulous entailment relations based on these temporal models and we characterize Reiter extensions in our semantics.
Introduction In [MH93a,b] a nonmonotonic logic was introduced, called Epistemic Default Logic (EDL). It is based on the metaphor of a meta-level architecture. It already has been established in how upward reflection can be formalized by a nonmonotonic entailment based on states, and the meta-level process by a (monotonic) epistemic logic. The meta-level reasoning can be viewed as the part of the reasoning pattern where it is determined what the possibilities are for default assumptions to be made, based on which information is available and (especially) which is not. The outcome at the meta-level concerns conclusions of the form Pj, where j is an object-level formula. In EDL, default conclusions are kept separate from the object level knowledge (they remain at the meta-level), by means of this explicit default operator (just like in NML3, see [Doh91]). If one wants to draw further conclusions from them using object level knowledge this should be done at the meta-level. Compared to a meta-level architecture, what was still missing was the step where the default assumptions are actually made, i.e., where such formulas j are added to the object level knowledge, in order to be able to reason further with them at the object level. Here we actually "jump (down) to conclusions". This is what should be achieved by the downward reflection step. In the current paper we will introduce a formalization of the downward reflection step in the reasoning pattern as well. Thus a formalization is obtained of the reasoning pattern as a whole consisting of a process of generating possible default assumptions and actually assuming them (a similar pattern as generated by the so-called BMS-architecture introduced in [TT91]). The formalization of downward reflection is inspired by [Tre94, HMT94], where it is pointed out how temporal models can provide an adequate semantics for meta-level architectures in general, and [ET93, ET94] where these ideas have been worked out to obtain a linear time temporal semantics for default logic. The general idea is that conclusions derived at the meta- This work is partially supported by ESPRIT III Basic Research Action No. 6156 (DRUMS II). a Utrecht University, Dept. of Computer Science, P.O. Box 80.089, 3508 Utrecht. Email: wiebe@cs.ruu.nl. Also partially supported by the Free University of Amsterdam. b Like footnote a, but also partially supported by the University of Nijmegen. Email: jj@cs.ruu.nl. c Free University of Amsterdam, Dept. of Mathematics and Computer Science, De Boelelaan 1081a, 1081 HV Amsterdam, The Netherlands. Email: treur@cs.vu.nl. level essentially are statements about the state of the object level reasoning at the next moment of time. So we can define downward reflection as a shift of time in a (reasoning) process that is described by temporal logic. In this paper in the Sections 2 to 6 the logic EDL is presented. As compared with earlier publications ([MH91a, MH92a, MH93a,b]), the logic is slightly extended in order to cater for downward reflection in subsequent sections. In Section 7 we define a labeled branching time temporalization of this logic, in spirit of the approach of [FG92]. In Section 8 we define sceptical and credulous entailment relations based on temporal models. 2. Epistemic logic To start with, our language consists of objective formulas together with epistemic formulas. 2.1. DEFINITION (objective and epistemic formulas). Let P be a set of propositional constants I is either a finite or countably infinite set. The set OBJ of objective formulas is the smallest set containing P and closed under the classical propositional connecitves. The formula - is shorthand for (p k . -p k ), for an arbitrary p k - P . The set FORM of epistemic formulas j, y,. is the smallest set containing OBJ, closed under the operator K , where Kj means that j is known. Moreover, we use Mj as an abbreviation for -K-j, meaning that j is epistemically possible. For G - FORM, we denote by Prop(G) the set of objective formulas in G. Objective formulas are interpreted on ordinary valuations: 2.2. DEFINITION (propositional models, classical consequence). A valuation is a function s from P to {t, f}. The set of all valuations is denoted Mod(PC). By determining s' behaviour under the Boolean connectives, such a valuation is straightforwardly extended to a truth function on arbitrary objective formulas a (cf. Definition 2.5): we then say that s is a (propositional model) for a. For any set of valuations M and objective formula a, we write M . a if any s - M is a model for a. To interpret the whole set FORM of epistemic formulas, we need richer structures: 2.3. DEFINITION (simple S5-Kripke models). A (simple) S5-model M is a non-empty subset of Mod(PC). The elements of M are called worlds. We denote the class of simple S5-models by Mod(S5). Note that for any M # Mod(PC) we may interpret arbitrary objective formulas, but in order to interpret epistemic formulas (cf. Definition 2.5) we have to conceive M as an S5- model M. The more general definition of an S5-model requires M to be of the format where, for each w - W, p(w) is a valuation (thus, different elements of W may verify the same objective formulas) and R is an equivalence relation (needed to interpret the operator K), but for our purposes we will assume S5-models to be simple in the sense as we defined above. One can show that our assumption does not change the validities of the logic (cf. Definition 2.6 below), but it facilitates a number of technical issues such as the definition of submodels and union of models. However, some caution is in order here: although the restriction to simple S5- models does not affect the notion of validity, some of the propositions on S5-models below do depend on the simplicity of the models, and do not hold for general S5-models. The set of worlds in an S5-model represents a collection of alternative worlds that are considered (equally) possible on the basis of (lack of) knowledge. In the next section we shall use S5- models as representations of the reasoner's objective knowledge and epistemic meta-knowledge (i.e., what he knows that he knows or does not know). 2.4. REMARK (submodels and union of S5-models). Note that under our notion of S5-models, which are just sets of valuations, one easily defines a subset relation on S5-models, and an operation of union upon them. 2.5. DEFINITION (interpretation of epistemic formulas). Given the S5-model M # Mod(PC), we define the relation (M, s) . j , with s -M by induction on the structure of the epistemic (M, s) . p - (M, s) . y (M, s) . -y - (M, s) - y (M, s) . Ky - (M, s') . y for all s' - M (M, s) . My - (M, s') . y for some s' - M. Note, that the definition of the K-operator states that j is known precisely when j holds in the whole set of epistemic alternatives; that for the M-operator states that j is considered epistemically possible iff there is at least one epistemic alternative that satisfies j. In the sequel, when we write M # Mod(PC) we want to interpret it as a collection of propositional formulas, whereas, if we write M for the same collection, we intend that M -Mod(S5), on which the formula Kj has an interpretation. We then say that M is based on M. 2.6. DEFINITION (validity and satisfiability). (i) j is valid in an S5-model M - Mod(S5), denoted M . j, if for all m - M: (M, m) . j. (ii) j is valid, notation Mod(S5) . j, if M . j for all S5-models M. (iii) j is satisfiable if there is an S5-model M and a world m - M such that (M, m) . j. w.r.t. S5-models can be axiomatized by the system S5: 2.7. DEFINITION (system S5). The logic S5 consists of the following: Axioms: propositional tautologies Knowledge is closed under logical consequence. Known facts are true. One knows that one knows something. One knows that one does not know something. Derivation rules: (R1) y Modus Ponens (R2) Kj Necessitation That j is a theorem derived by the system S5 is denoted by S5 - j. 2.8. THEOREM (Soundness and completeness of S5). S5 - j - Mod(S5) . j 2.9. DEFINITION. We say a formula y is in normal form if it is a disjunction of conjunctions of the form d = a . Kb 1 . Kb 2 . Kb n . Mg 1 . Mg 2 . Mg k , where a, b i and g j (i - n, - are all objective formulas. The following holds for the logic S5; 2.10. THEOREM ([HC68]) In S5 every formula j is equivalent to a formula y in normal form. In particular, this theorem implies that every epistemic formula can be represented equivalently without nested epistemic modalities. In S5 one thus can do without nestings. 3. Epistemic states and stable sets In this paper we simply define an epistemic state as an S5-model. The idea behind this was already touched upon after Definition 2.3. The worlds in an S5-model represent the states of the (real) world that the reasoner considers possible. Thus the S5-model as a whole delimits the ways the real world is like as far as the reasoner is concerned. In other words, it determines what he knows about the world and what he does not (what he has doubts about by considering contradictory possibilities). Therefore, an S5-model represents truly the epistemic state of the reasoner. 3.1. DEFINITION. An epistemic state is an S5-model M - Mod(S5). The set M of worlds in M is the set of epistemic alternatives allowed by the epistemic state M. For such a model M, we define K(M) as the set of facts known in M: . Kj}. We call K(M) the theory of M or knowledge in M. We mention here that the knowledge in M are exactly the validities in M: 3.2. LEMMA. For any S5-models M 1 and 3.3. REMARK. The converse relation of - on Kripke models (Cf. Definition 2.4), will play an important role in the sequel. M 1 - M 2 means that the model M 2 , viewed as a representation of the knowledge of a reasoner, involves a refinement of the knowledge associated with model . This has to be understood as follows: in the model M 2 less (or the same) worlds are considered possible by the reasoner as compared by the model M 1 . So, in the former case the reasoner has less doubts about the true nature of the world. By Lemma 3.2 it turns out that this means that with respect to model M 2 the reasoner has at least the objective knowledge associated with model M 1 , and possibly more. So in a transition of M 1 to M 2 we may say that objective knowledge is gained by the reasoner. Thus the relation '-' acts as an information ordering on the set of S5-models. Finally we remark that Lemma 3.2 is a typical example of a property that holds for simple S5-models only; with respect to general S5-models we would only have the 'only if' part of the Lemma (cf. the discussion following Def. 2.3). 3.4. PROPOSITION (Moore [Moo85]). (i) The theory of an epistemic state M is a so-called stable set, i.e., satisfies the following properties: (St 1) all instances of propositional tautologies are elements of (St (St 5) S is propositionally consistent. (ii) Every stable set S of epistemic formulas determines an S5-Kripke model M S for which it holds that is a finite set, then M S is the unique S5-Kripke model with this property. The (proof of) the following proposition, which determines stable sets in the realm of classical propositional logic (cf. St 1 & 5), is also to be found in [Moo85]. 3.5. PROPOSITION. A stable set is uniquely determined by the objective formulas it contains. 3.6. REMARK. Thus stable sets act as the epistemic contents of an epistemic state (viz. an S5- model): a stable set describes exactly the formulas known by the reasoner when he is is in epistemic state M. From Proposition 3.4(i) we see that this knowledge is closed under classical propositional reasoning and under positive and negative introspection: if some formula is known also the fact that this is known is in its turn known by the reasoner, and if some formula is not known, the fact that it is not known is known itself. This reflects the (perhaps rather strong notion of) rationality of the reasoner. On the other hand, 3.4(ii) says that a stable set of epistemic formulas determines an S5-model, the epistemic state associated with (knowing) this stable set. Proposition 3.5 states that a stable set, and thus the associated state, is completely determined if one specifies exactly what objective knowledge is present at the reasoner. The rest of the known set of formulas (the stable set) then follows. 4. Entailment based on epistemic states: upward reflection On the basis of epistemic states, Halpern & Moses define an entailment relation |~ with which one can infer what is known, and, more importantly, what is unknown in such epistemic states. Recall that with every set of valuations M we can associate its epistemic state M in a straightforward way (cf. Definition 2.5). Intuitively, M associates with M the epistemic state given by considering (only) the set M of classical propositional models to be the possible states of the world: this set, so to speak, gives a partial description of the real world. In general, more than just one classical model is considered possible so that only those facts are known that are true in all of them. We can exploit this idea further and define an entailment relation based on the premise that we only know some objective formula j. To this end we need some additional 4.1. DEFINITION. Given some objective formula j, we define M j as the set of valuations satisfying j}. The epistemic state that is associated with M j is denoted by M j . We have the following alternative characterizations of 4.2. PROPOSITION. M Proposition 4.2 says that in order to get M j , we can also consider all S5-models of j and take their union to obtain one 'big' S5-model. The result immediately follows from the fact that we have simple S5-models, and from the fact that, for objective formulas j and arbitrary M - Mod(S5), M . j - (since j is objective-adopting the view that M - Mod(PC)) M . j - Now we are ready to define an entailment relation based on this "big" model. Keeping in mind that '#' acts as an information order on models (Cf. 3.3), we define what are the consequences of knowing only the objective formula j as those formulas that hold in the "big" model 4.3. DEFINITION (Nonmonotonic epistemic entailment). For FORM: Informally, this means that y is entailed by j, if y is contained in the theory (knowledge) of the "largest S5-model" M j of j. Halpern & Moses showed in [HM84] that this "largest model" need not always be a model of j itself if we allow j to contain epistemic operators. However, in our case where we only use objective formulas j, M j is always the largest model for j. This is obvious from our construction of M j , since M j . j and therefore M j . j (since j is objective). Moreover, in this case the theory K(M j ) of this largest model is a stable set that contains j and such that for all stable sets S containing j it holds that (by Lemma 3.2 and Proposition 3.4), thus K(M j ) is the "propositionally least" stable set that contains j. So |~ can also be viewed as a preferential entailment relation in the sense of Shoham [Sho87, 88], where the preferred models of j are the largest ones, viz. M j , where the least objective knowledge is available. We denote the mapping j - K(M j ) by k: the stable set associated with knowing only j. Alternatively viewed, k(j) is the |~-closure of j. Note that since stable set, it is also propositionally closed. We give a few examples to show how the entailment |~ works: Let p and q be two distinct primitive propositions. Then: Obviously, the entailment relation |~ is nonmonotonic, justifying the name we have given it. (For instance, we have p |~ M-q, while not p - q |~ M-q; it even holds that p - q |~ -M-q . ) The nonmonotonic epistemic entailment |~ enables us to derive from an objective formula j, characterizing the exact epistemic state of the reasoner (viz., technically, the epistemic state exactly what is known and, even more importantly, what is unknown in this epistemic state. This latter property renders the entailment relation context-sensitive and nonmonotonic, so that the relation |~ goes beyond an entailment expressible in ordinary epistemic logic: with respect to its premise the relation |~ involves a kind of epistemic closure. Finally we state a property of that we shall need in the sequel: 4.4. PROPOSITION. The entailment relation |~ enjoys the property: PROOF. Directly from Definition 4.3 and the fact that the stable set K(M j ) is closed under conjunction (which in its turn follows from the fact that a stable set is closed under propositional reasoning and contains the tautology y 5. The epistemic preference logic S5P The "upward reflection" entailment relation |~ enables us to derive information about what is known and what is not known. In this section we show how we can use this information to perform default reasoning. To this end we extend our language with operators that indicate that something is a default belief and thus has a different epistemic status than a certain fact. In this way the proverbial "jump to conclusions" is not made directly in the logic, but a somewhat more cautious approach is taken. The "jump" itself will be part of a next operation, the "downward reflection", which will be discussed in the next section when we incorporate a temporal element into our approach. Let I be a finite set of indexes. The logic S5P of epistemic default logic is an extension of the logic S5 by means of special modal operators P i denoting default belief (w.r.t. situation or frame of mind i), for i - I, and also generalisations P t , for t - I. Informally, read as "j is a default belief (within frame of reference i)". As we shall see below, a frame of reference (or mind) refers to a preferred subset of the whole set S of alternatives. This operator is very close to the PA (possible assumption) of [TT91] and the D (default) operator of [Doh91]. The generalisation P t j is then read as a default belief with respect to the (intersection of the) frames of reference occurring in t. Formally, S5P-formulas are interpreted on Kripke-structures (called S5P-models) of the form is a collection of worlds, M i # M (i - I) are sets ('frames') of preferred worlds, and R mean that the set of worlds of M 1 is a subset of those of M 2 . Again we have identified worlds s and their truth assignments. We let Mod(S5P) denote the collection of Kripke-structures of the above form. Given an S5P-model M we call the S5- model M based on M the S5-reduct of M'. 5.1. DEFINITION (interpretation of S5P-formulas). Given a model we extend the truth definition on S5-models by the following clauses for the P i - and (M, m) . P (M, m) . P t j iff (M, m') . j for all m' with R t (m, m'), where R Thus the former clause states that P i j is true if j is a default belief w.r.t. subframe M i , whereas the latter says that P t j is true if j is a default belief w.r.t. the intersection of the subframes t. We will sometimes denote this intersection by M t . In the clause for P t , we assume that, for M. So in this special case we get that the P t modality coincides with the knowledge operator K . Validity and satisfiability is defined analogously as before. It is possible to axiomatise (the theory of) Mod(S5P) as follows (cf. [MH91a, MH93a]): take the S5 system for the modality K (and dual M) and use K45 for the P-modalities (both the P i and the P t ), together with relating axioms, resulting in the system: 5.2. DEFINITION (system S5P). In the following, i ranges over I, and t over subsets of I . Moreover, - is a variable over {K, P propositional tautologies; (R1) Modus Ponens (R2) Necessitation for K: - j - Kj. Axiom (B1) says that we are dealing with an extension of propositional logic; (B2) says that all the operators K, P i , and P t are 'normal'; (B3 and B4) express that the relations R, R i and R t are transitive and Euclidean, respectively; (B5) says that R is reflexive; (B6) and (B7) help us to get rid of nested modalities: a nested modality is always referring to the frame corresponding to the innermost one. (B8) provides us with a bridge between the modalities with indices from I and those of P(I); it also shows that in fact we could do without the P i 's. (B9) says that if t # t', then R t' # R t . Finally, (B10) is the syntactical counterpart of our definition that We call the resulting system S5P. In the sequel we will write G - S5P j or j - Th S5P (G) to indicate that j is an S5P-consequence of G. We mean this in the more liberal sense: it is allowed to derive j from the assertions in G by means of the axioms and rules of the system S5P , including the necessitation rule. (So, in effect we consider the assertions in G as additional axioms: G - S5P j iff - S5P-G j.) As for S5, one can again prove a normal form theorem for S5P implying that all formulas can be represented equivalently as formulas without nestings. 5.3. THEOREM G - S5P j - (for all M - Mod(S5P): M . G - M . PROOF. In the Appendix. 6. Epistemic Default Logic (EDL) In the language of S5P we express defaults of the (using Reiter's notation, Cf. are objective formulas. The reading of such a formula is "if j is true and y is (considered) possible, then c is preferred (within frame M i )". Usually we consider cases where y is syntactically equal to c (the so-called normal defaults). By combining the formal apparatus of S5P with Halpern & Moses' nonmonotonic epistemic entailment we obtain a framework in which we can perform default reasoning. In this paper we call this framework Epistemic Default Logic (EDL). 6.1. DEFINITION (default theory). A default theory Q is a pair (W, D), where W is a finite, consistent set of objective (i.e. non-modal) formulas describing (necessary) facts about the world, and D is a finite set of defaults of the form j - My- R i c, where j , y and c are again objective formulas, and i - I. The sets W and D are to be considered as sets of axioms, and we may apply necessitation to the formulas in them. In principle, it would be possible to use the index i of the P-operator in a completely arbitrary way, to be chosen by the knowledge engineer. However, in order to be able to treat the various defaults in a default theory separately from each other, in the sequel we shall assume all P- operators in a default theory to be distinct. This will allow for a generic way to look at the possibilities of combining default beliefs by using the P t -operators. So, for example rather than to look at a default theory with in which for both defaults the same operator P 1 is used, we consider the set D R 2 s}, in which both defaults are represented by means of different operators P 1 and P 2 , and consider the combined operator P {1, 2} . 6.2. DEFINITION (default entailment). Let a default theory D) be given. Let W* be the conjunction of the formulas in W. Note that W is a finite set, and moreover that W only consists of objective formulas. Furthermore, let j be an objective formula. Then we define the entailment relation |~ Q w.r.t. default theory Q as follows: (Recall that k(j) stands for the |~-closure of j, where |~ is the nonmonotonic epistemic entailment of section 4.) y. This definition states that, given a default theory Q, y is a default consequence of j iff y follows in the S5P-logic from the defaults together with what is implied by knowing only the conjunction of j with the background information W. Notice that the following equivalence holds, which states something about a modularity of the entailment relations. 6.3 PROPOSITION there exists a S5-formula y' - FORM such that j-W* |~ y' and {y'} - D - S5P y PROOF. This follows directly from the definition of |~ Q and Proposition 4.4, as is seen as follows. Suppose we have j |~ Q y, i.e, k(j-W*) - D - S5P y. The logic S5P is compact, hence there are a 1 , ., a k -k(j-W*) and d 1 , ., d n -D such that a 1 , . a k , d 1 , ., d n y. By definition of k, since for each i - k, a i - k(j-W*) we have j-W* |~ a i for all i by Proposition 4.4, y' satisfies the right-hand side of 6.3. - Instead of true |~ Q y, we simply write |~ Q y. Note that, for y, with Furthermore, if G is a finite set of epistemic formulas, and y an S5P-formula, then we define G |~ Q y as G* |~ Q y, where G* stands for the conjunction of the formulas in G. 6.4. EXAMPLE (Tweety). Consider the following default theory -f} and representing that penguins do not fly, and that by default birds fly. (For convenience we omit the subscript of the P-operator.) Now consider the following (i). b |~ b -K-f (1) meaning that from the mere fact that Tweety is a bird, we conclude that Tweety is assumed to fly. (Here |~ stands for Halpern & Moses' nonmonotonic epistemic entailment.) This must be contrasted to the inference: (ii). b - p |~ Kp - S5P K-f (1) i.e, not b - p |~ Q Pf, meaning that in case Tweety is a penguin, we cannot infer that Tweety is assumed to fly, but instead we can derive that we know for certain that Tweety does not fly: b - p |~ Q K-f . 7. A temporal formalization of downward reflection In the previous section it has been described how upward reflection can be formalized by a nonmonotonic inference based on epistemic states, and the meta-level process by a (monotonic) logic. In the current section we will introduce a formalization of the downward reflection step in the reasoning pattern. The meta-level reasoning can be viewed as the part of the reasoning pattern where it is determined what the possibilities are for default assumptions to be made, based on which information is available and (especially) which is not. The outcome at the meta-level concerns conclusions of the form Pj, where j is an object-level formula. What is still missing is the step where the default assumptions are actually made, i.e., where such are added to the object level knowledge, in order to be able to reason further with them at the object level. Here we actually "jump (down) to conclusions". This is what should be achieved by the downward reflection step. Thus the reasoning pattern as a whole consists of a process of generating possible default assumptions and actually assuming them. By these downward reflections at the object level a hypothetical world description is created and refined. This means that in principle not all knowledge available at the object level can be derived from the object level theory W: downward reflection creates an essential extension to the object level theory. Therefore it is excluded to model downward reflection according to reflection rules as sometimes can be found in the literature, e.g., "If at the meta-level it is provable that Provable(j) then at the object level j is provable" (e.g., see [Wey80]): A reflection rule like this can only be used in a correct manner if the meta-theory about provability gives a sincere axiomatisation of the object level proof system, and in that case by downward reflection nothing can be added to the object level that was not already derivable from the object level theory. In the above rule the meta-theory is only concerned with one fixed object level theory. Since we essentially extend the object level theory, and consequently want to move to another object level theory, an approach as taken in the rule above cannot serve our purposes here. In fact, in our case a line of reasoning at the object level is modelled by inferences from subsequently chosen theories instead of inferences from one fixed theory, where in the choice of the next object-level theory the upward and downward reflections play a role. Another difference with the above "traditional" approach to meta-level reasoning is that we shall treat these reflection steps model-theoretically instead of proof-theoretically, as shifts from one model to another, which results in a temporal "super"-model indicating how these shifts may take place over time. Actually, to keep in perfect line with the above, we should explain by means of these models how object knowledge (modelled by a set M of valuations) is reflected upwards to get a model for epistemic meta-knowledge (modelled by an S5-model M), how this meta-knowledge is then extended with meta-knowledge about default beliefs (modelled by S5P- models), and how, subsequently, this meta-knowledge is reflected downwards to object knowledge (modelled by a new set M' of valuations) again. However, representing such shifts in a temporal model would involve (sequences of) three different models (sets of valuations, S5-models and S5P-models), which would be quite involved and cumbersome. Therefore, in order not to complicate our temporal models too much, in our present setting we shall not represent the above three steps separately, but treat them "in one blow", so to speak, i.e., as one "super"-step. Now we are able to represent these "super"-steps as shifts from S5P-models to S5P-models, so that our temporal "super"-model only involves sequences of one kind of model, viz. S5P-models, of which the S5-reducts represent the reasoner's objective knowledge and epistemic meta-knowledge, and the frames ('P-') parts represent meta-knowledge in the form of default (or preferred) beliefs. We shall still refer to these supersteps as downward reflection, although as we stated above, apart from a reflection of the "meta-knowledge" in the form of default beliefs contained in the S5P-model at hand to the object level by converting some of these beliefs into object knowledge, it also includes an immediate upward reflection from the next object level theory (or rather model) resulting from this to the next meta-level theory as represented by an S5P-model again. In [GTG93] such a shift between theories is formalized by using an explicit parameter referring to the specific theory (called 'context' in their terms) that is concerned, and by specifying relations between theories. In their case reflection rules ('bridge rules' in their terms) may have the form: Here it is assumed that at the meta-level, knowledge is available to derive conclusions about provability relations concerning a variety of object level theories OT. So, if at the object level from a (current) theory OT some conclusions have been derived, and these conclusions have been transformed to the meta-level, then the meta-level may derive conclusions about provability from another object level theory OT'. Subsequently one can continue the object level reasoning from this new object level theory OT'. The shift from OT to OT' is introduced by use of the above reflection rule. As we said above, in the approach as adopted here we give a temporal interpretation to these shifts between theories. This can be accomplished by formalizing downward reflection by temporal logic (as in [Tre94]). In a simplified case, where no branching is taken into account, the following temporal axioms can be used to formalize downward reflection: for every objective formula j. Here X is the temporal operator asserting that in the next (epistemic) state its argument is true. In the general case we want to take into account branching and the role to be played by an index t in P t j. We will use this index t to label branches in the set of time points. By combining EDL with the temporal logic obtained in this manner we obtain a formalization of the whole reasoning pattern. We start (following [FG92]) by defining the temporalized models associated to any class of models and apply it to the classes of models as previously discussed. In contrast to the reference as mentioned we use labeled flows of time. We use one fixed set L of labels, viz. I , the powerset of the index set I. However, in most definitions we do not use this fact, but only refer to (elements t of) L. 7.1 Flows of time 7.1. DEFINITION (discrete labeled flow of time). Suppose L is a set of labels. A (discrete) labeled flow of time (or lft) , labeled by L is a pair consisting of a nonempty (countable) set T of time points and a collection of binary relations < t on T. Here for s, t in T and t in L the expression s < t t denotes that t is a (immediate) successor of s with respect to an arc labeled by t. Sometimes it is convenient to leave the indices out of consideration and use just the binary relation s < t denoting that s < t t for some t (for some label t they are connected). Thus we have that We also use the (nonreflexive) transitive closure - of this binary relation: We will make additional assumptions on the flow of time; for instance that it describes a discrete tree structure, with one root and in which time branches in the direction of the future. 7.2. DEFINITION (labeled time tree) An called a labeled time tree (ltt) if the following additional conditions are satisfied (recall that (i) the graph (T, <) is a directed rooted tree. (ii) (successor existence) Every time point has at least one successor: for all s - T there exists a t and a t - T such that s < t t (iii) (label-deterministic) For every label t there is at most one t-successor: for all s, t, t' - T it holds: s < t t, s < t t' - There are still some additional properties that sometimes are required. 7.3. DEFINITION (sub-lft and (maximal) branch) a) An lft (T', (<' t called a sub-lft of an lft (T, (< t holds It is also called the sub-lft of (T, (< t defined by T', or the restriction of (T, (< t b) A branch in an lft T is a sub-lft (i) - T' x T' is a total ordering on T' x T' (ii) Every t' in T' with a <-successor in T also has a <-successor in T': for all t' - T', t there is a s' - T' : t' < s' (iii) Every element of T that is in between elements of T' is itself in T': for all s' - T', t - T, u' c) A branch is called maximal if every t' in T' with a t-predecessor in T also has a t ' - predecessor in T' for some t': for all t - T, t' there is a t ' and an s' - T' : s' < t' t'. (Note that if T is a labeled time tree, this condition can be simplified to: the root r of T is also the root of T'.) We now immediately have the following: 7.4. PROPOSITION Any branch of an ltt T is an ltt. 7.5. DEFINITION a) An lft is called successor-complete if for every label t every time point has at least one t- successor: for all s and t there exists a t such that s < t t. b) A path is a finite sequence of successors: s 0 , ., s n such We call s 0 the starting point and s n the end point of the path. 7.6. DEFINITION (time stamps). Given an ltt (T, (< t is called a time stamp mapping if for the root r it holds that and for all time points s, t it holds s < t 1. (Note that this time stamp mapping is unique.) Note that an ltt is infinitely deep, i.e., for every k - N there is a time point t - T with This is a direct consequence of the following proposition: 7.7. PROPOSITION If B is a maximal branch in an ltt, then any time stamp mapping is an isomorphism between B and N. 7.2 Temporal models We first define our temporal formulas: 7.8. DEFINITION (temporal formulas). a) Given a logic L, temporal formulas over (the language of) L are defined as follows: of L then Cj is a temporal formula (also called a temporal atom) ii if j and y are temporal formulas, then so are: -j, j . y, j ' y, X $,t j, X b) Below, our main concern will be temporal formulas over S5P; we will also refer to them as TEDL-formulas (temporal epistemic default logic formulas). The above temporal operators are fairly standard in branching-time temporal logic: X refers to next-time, F to sometimes in the future and G to always in the future; the subscripts refer to whether one considers only some ($) possible path (possibly with a fixed label t) or all (") of these. The C-operator is the least standard: it means 'currently'. Usually in temporal logic this operator is not really useful, since it just states that its argument, say j, holds in the current state, which is normally represented by the formula j itself, without a C in front. Here, however, we have a logic where we mix temporal and epistemic logic, and the C-operator acts as a kind of separator between the epistemic and temporal part: in Cj, its argument j is an itself is a temporal formula. This also facilitates the semantic definition below. 7.9. DEFINITION (temporal models) a) Let MOD be a class of models, and labeled flow of time. A temporal MOD-model over T is a mapping M: T - MOD. If M is a temporal MOD- model for any class MOD we call M a temporal model. For (the snapshot at time point t) by M t . The temporal model can alternatively be denoted by (M t ) t-T . b) If we apply a) to the classes of models Mod(PC), Mod(S5) and Mod(S5P), we call these temporalized models temporal PC-models, temporal S5-models and temporal S5P- models over T, respectively. c) Given an lft T, the temporal formulas are interpreted on MOD-models over T as follows: conjunction and implication are defined as expected; moreover M, s . -j iff not M, s . j; ii The temporal operators are interpreted as follows: means that in the current state j is true, i.e. M, s . Cj iff M s . j means that j is true in some t-successor state i.e., M, s . X $,t j iff there exists a time point t with s < t t such that M, t . j means that there is a t with some t-successor in which j is true. there exists a time point t with s < t such that M, t . j meaning that j is true in all t-successor states, i.e., M, s . X ",t j iff for all time points t with s < t t it holds M, t . j means that j is true in all immediate successors: means that j is true in some future state, i.e., there exists a time point t with s - t such that M, t . j means that for all future paths there is a time point where j is true, i.e., starting in s there is a t in B with M, t . j. means that j is true along some future path, i.e., there exists a branch B starting in s with M, t . j for all t in B. means that j is true all future states i.e. M, s . G " j iff for all time points t with s - t it holds M, t . j. Note that the operator C enforces a shift in the evaluation of formulas; taking us from a temporal model M and a time point t to an S5P-model M t . During the reasoning process we assume to gradually extend the information we have at the object level, and consequently to shrink the set of possible worlds by means of reflecting default beliefs downwards to object knowledge. In terms of temporal S5-models we can formulate this property as follows: 7.10. DEFINITION. A temporal S5P-model obeys downward reflection, if the following holds for any s and the frame M t in M s is non-empty - there is a t with s < t t and for all such t the set of worlds of M t equals M t The above property expresses that the possible worlds in frame M t (representing the P t -default beliefs) are taken to be the whole set of possible worlds (representing the objective knowledge) in some (t-) successor epistemic state (or, rather, S5P-model). Finally, we are ready to zoom in into the models we like to consider here, the temporal models. 7.11. DEFINITION (TEDL-models) TEDL-model M is a temporal S5P-model over an lft T such that: 1) T is a labeled time tree; For every time point s, there is exactly one t with s <- t; reflection. The following notion is a crucial one for the sequel: 7.12. DEFINITION (conservativity and limit models). Let labeled flow of time. A temporal PC-model M over T is conservative if for every two time points s and t with s < t it holds that M(s) # M(t). Suppose M is a conservative temporal PC-model. The intersection of the models M(s) for all s in a given (maximal) branch of the lft T is called the limit model of the branch, denoted lim B M. The set of limit models for all (maximal) branches is called the set of limit models of M. These definitions straightforwardly extend to S5- and S5P-models, by identifying M with its set of worlds, M. 7.13. THEOREM. TEDL-models are conservative. PROOF. Suppose s < t, that is, for some label t, s < t t. Since TEDL-models obey downward reflection and the underlying lft is an ltt, this means that the (now unique) set of worlds M t of M t is the frame M t of M s . Since M t # M s , the set of worlds of M s , this gives M t # M s , which proves the theorem. - Since we know that TEDL-models are conservative, we have that once we have obtained that somewhere in a time point s along some path of such a model some objective formula is known, it remains known in all successor points of s. (This is in fact the rational behind the name "conservative models".) We can make this more precise if we introduce so-called cko- formulas, which are formulas of the form CKj, where j is objective. In the sequel we will denote cko-formulas by a. Then we have that: 7.14. PROPOSITION. Let M be a TEDL-model, and a a cko-formula. If, for some s, M, s . a, then M, t . a for all t with s - t. PROOF. Suppose M, s . a, for a = CKj. We show that for t with s < t we have that M, t . a. (Then we can finish the proof by induction on the number of time point between s and t.) By Theorem 7.13, we have that M t # M s . This yields: M, s . a - M, s . CKj - M s . is objective and M t # M s we then must have M t , . j for all m in M t and hence M t . Kj. Finally, we have the equivalences M t . Kj - Furthermore, TEDL-models enjoy a number of properties that can be expressed as validities in our logic: 7.15. THEOREM. TEDL-models satisfy the following validities: T0 All the operators of {X ",t , X" , F" , G" } satisfy the K-axiom (C too) and generalisation; (introduction of C) existence & label-deterministic) is union of < t ) (dual of T6) objective (allowing downward reflection) objective (trivial downward reflection) objective (conservativity) transitive closure of <) (from conservativity and induction) 7.17. REMARK. The Theorem above says that the formulas are at least sound; until now we have not been concerned by designing a logic that is complete for TEDL-models. 8. TEDL models of default theories and entailment relations We could formulate definitions of entailment of objective formulae related to any model, or any model based on the standard tree. But it may well happen that there are branches in such models, for instance labeled by the empty set only, that contain no additional information as compared to the background knowledge. It is not realistic to base entailment on such informationally poor branches in a model. Therefore we define: 8.1. DEFINITION (informationally maximal) We define for TEDL-models M 1 and M 2 over the same flow of time that M 2 is informationally larger We call M 1 informationally maximal in a class M of models if it is itself the only model that is informationally larger. For a given model M we will apply this definition to the set B of all maximal branches, with T flow of time. 8.2. DEFINITION (regular model) TEDL-model M is called regular if all maximal branches are informationally maximal in the class of maximal branches. The submodel based on all time points t included in at least one maximal, informationally maximal branch is called the regular core of M, denoted by reg(M) . 8.3. REMARK. In general the regular core will not be label-complete, because branches may be cut off. The idea behind taking informationally complete branches is that we want to maximalise the effect of applying defaults in order to obtain as much (default) knowledge as possible. 8.4. DEFINITION. Let M be a TEDL-model over We define for k - N where M'(t) stands for the S5-reduct of the S5P-model M(t). 8.5. PROPOSITION. Let M be a TEDL-model. a) For k - k' it holds M (k') # M (k) branch of reg(M PROOF. a) follows directly from Theorem 7.13, b) follows from the distributivity of intersection over union (and vice versa). - 8.6. DEFINITION (sceptical entailment) Let M be a TEDL-model and a a cko-formula. We define the sceptical entailment relation by: M | - scep a - for every maximal branch B in reg( M) there is a t in B such that M, t . a. An immediate consequence of this definition is: 8.7. PROPOSITION. If, for some k - N, M (k) . Kj, then M | - scep CKj. PROOF. Suppose M (k) . Kj, for some k - N. This means that M'(t) . Kj for all S5-reducts M'(t) with t in reg(M) and k. So for every maximal branch B in reg( M) we can find a t in B such that M, t . Kj (viz. take t in B such that |t| = k. By Proposition 7.7 we can always do this). Hence, M | - scep CKj. - 8.8. PROPOSITION. Let M be a TEDL-model with root r. Let a = CKj be a cko-formula. The following are equivalent: reg(M), r . F " a . Kj for every maximal branch B of the regular core of M. PROOF. Clearly (i) and (ii) are equivalent.By Prop. 8.5 also (iii) and (iv) are equivalent. We now show the equivalence of (i) and (iv). First we prove (i) - (iv): Suppose M | - scep CKj, i.e. for every maximal branch B in reg( M) there is a t in B such that M, t . CKj. Now consider a maximal branch B in reg( M). Then there is a t in B with M, t . CKj, i.e. M(t) . Kj. By the conservativity of M we have that M(u) . Kj for all u - t, i.e., all worlds of the models M(u) with u - t satisfy j. But then also all worlds of the models M(u) with u - t and u in B satisfy j. Consequently, lim B Next we prove (iv) - (i): Suppose lim B M . Kj for every maximal branch B of the regular core of M. We have to prove that for every maximal branch B in reg( M) there is a t in B such that M, t . CKj. Take some maximal branch B in reg( M). Then we know that lim B M . Kj, i.e., - u in B M(u) . Kj. By the conservativity of the model M (Theorem 7.13) we know that the sequence <M(u) | u in B> is monotonically decreasing (with respect to -). Let P the set of propositional atoms occurring in the formula a (or j). Clearly, P 0 is finite. Now, for any S5-model M with set S of worlds (truth assignment functions) we let M 0 denote the model with set S 0 of worlds: S for some s - S}. Clearly, M . Kj iff M 0 . Kj. Now consider the sequence <M(u) 0 | u in B>. This is a monotonically decreasing sequence with intersection - u in B M(u) 0 . Since P 0 is finite, and we identified worlds with truth assignment functions, we have that all the models M(u) 0 contain only a finite number of worlds. Together with the fact that the sequence is monotonically decreasing this yields that the sequence must be stable from some point s in B on: M(u) obviously the intersection - u in B M(u) . So now M(s) 0 . Kj, and thus M(s) . Kj, i.e. M, s . CKj. - For our definition of credulous entailment we can be less restrictive. Especially, too less information in one branch can always be overcome by another, informationally larger branch. 8.9. DEFINITION (credulous entailment) Let M be a TEDL-model. We define: cred CKj - there exists an s with M | - cred CKj - there exists an s with M(s) . Kj 8.10. PROPOSITION. Let M be a TEDL-model, and a = CKj a cko-formula. The following are equivalent: cred a cred a for some k - N. . Kj for some maximal branch B PROOF. Since clearly (i), (ii) and (iii) are equivalent, we concentrate on the equivalence of (ii) and (iv): we first prove (ii) - (iv): Let M | - k cred CKj for some k - N. This means that there exists an s with and M(s) . Kj, i.e., M, s . CKj. By Prop. 7.14, we then have that M, t . CKj for all t with s - t. Now consider any maximal branch B through s. Clearly, this branch has the property that the models M(t) with s - t all satisfy Kj, i.e., j holds in every world of every model M(t). But then j holds in the intersection of these models as well. This proves . Kj. Next we prove (iv) - (ii): Suppose that lim B M . Kj for some maximal branch B . This means that Kj holds in the intersection of the models M(t) along B . By the conservativity of the model M (Theorem 7.13) we know that the sequence <M(t) | t in B> is monotonically decreasing (with respect to -). Let P the set of propositional atoms occurring in the formula a (or j). Clearly, P 0 is finite. As above, we denote, for any S5-model M with set S of worlds (truth assignment functions), the model with set S 0 of worlds by M 0 where S for some s - S}. Again, M . Kj iff M 0 . Kj. Now consider the sequence <M(u) 0 | u in B>. This is a monotonically decreasing sequence with intersection - u in B M(u) 0 . Since P 0 is finite, and we identified worlds with truth assignment functions, we have that all the models only a finite number of worlds. Together with the fact that the sequence is monotonically decreasing this yields that the sequence must be stable from some point s on: obviously the intersection - t in B M(t) the intersection satisfies Kj, we obtain that M(s) 0 . Kj, and thus M(s) . Kj. Hence M | - cred CKj. - We can now associate TEDL-models with default theories as follows. 8.11. DEFINITION (TEDL-model of a default theory). D) be a default theory. Then we define a TEDL-model of Q as a TEDL-model M Q such (i) (basis: the root) M Q r is an S5P-model such that (a) the S5-reduct of M Q r is the S5-model W* , as defined in Section 4, and (b) M Q r satisfies the set of defaults, i.e., M Q r . D. (ii) (induction step) Suppose that we are given an S5P-model at snapshot M Q s . Then we have that for a(n S5P-) model M Q t with s < t t, it holds that: (a) the S5-reduct of M Q t is the S5- model M t as it appeared as a frame in M Q s , and (b) M Q t satisfies the set of defaults again, i.e., . D. Note that, in general, there are multiple TEDL-models of a default theory Q. Furthermore, note that clause (ii)(a) reflects the downward reflection operation with respect to the P t -defaults. This definition enables us to finally give the definitions of sceptical and credulous entailment from a default theory. 8.12. DEFINITION (entailment from a default theory). D) be a default theory, and j an objective formula. Then: models M of Q it holds M | - scep CKj cred j iff for all models M of Q it holds M | - cred CKj Of course, we have that what is entailed sceptically, also is entailed credulously: 8.13. PROPOSITION. For default theory Q and objective formula j: Q | - scep j - Q | - cred j Sceptical and credulous entailment are not the only approaches to define semantics for default logic. Sceptical entailment can be considered a lower bound of the conclusions that can be drawn, credulous entailment as an upper bound. A more differentiated approach bases semantics of the default theory on the set of all alternative extensions that actually can be constructed out of a default theory using the defaults as (non-monotonic, context-dependent) inference rules. The original approach by Reiter [Rei80] defines the extensions of a default theory by means of a pseudo-construction or fixed point definition; we will call these extensions Reiter extensions. In Constructive Default Logic, introduced in [TT92], the notion of Reiter extension is replaced by the weaker notion of constructive extension. In [TT92] it was proven that the Reiter extensions of a default theory are among the constructive extensions and they were characterised. Constructive extensions give a broader perspective on the semantics of a default theory. The formalisation in terms of branching time models introduced in the current paper, in the first place is the temporal formalisation of the construction process of a constructive extension. Just like in the case of Constructive Default Logic, in the applicability condition of a point s, the justification My only refers to the current state of information (and not to the whole (future) context of the reasoning pattern as in Reiter's case is required). Because of this similarity for every (maximal) branch B in a TEDL-model M of a default theory D), the theory of its limit Th(lim B M) is a constructive extension of Q and the branch itself represents a construction chain of this constructive extension. As the Reiter extensions are among the constructive extensions, a natural question is how in a TEDL-model of a default theory the branches that lead to a Reiter extension can be characterised. Based on variants of the notions self-defeatingness and exhaustiveness introduced in [TT92], it is not difficult to find such a characterisation. 8.14. DEFINITION Let M be a TEDL-model of the default theory branch in M. The branch B is called self-defeating if there is a time point s in B and a default in D such that the following holds: for all time points t, t' in B and labels t with t < t t' and every default j ' - My ' - R j c in D with j in t, there is some time point u in B with B u . o/oo y '. Otherwise the branch B is called non-self-defeating. The notion of non-self-defeatingness quarantees that any default conclusion that is drawn will remain justified in the strong Reiter sense during the whole process. However, this is only one side of the characterisation. It may well be the case that the defaults applied in branch B are justified in the strong sense, but not all applicable defaults actually have been applied in B . Therefore the notion of exhaustiveness is required: 8.15. DEFINITION Let M be a TEDL-model of the default theory branch in M. The branch B is called exhaustive if for every time point s in B and every default in D such that B s . j and for all time points t in B it holds y, then there is a time point s' in B and a label t with s< t s' and i is in t. Based on these two conditions we can characterise the branches leading to Reiter extensions: 8.16. THEOREM Let M be a TEDL-model of the default theory D). The following are equivalent: (i) E is a Reiter extension of Q (ii) there exists a non-self-defeating and exhaustive branch B in M with The proof is similar to the proof of the main result in [TT92]. The following example default theory illustrates our a r1: a - Mc - R 1 b r3: a - Mo/oo c - R 3 o/oo c In Figure 1 the different reasoning paths are depicted. At each node the literals that are true at the object level are depicted, and, between [.], the applicable default rules. At each arc the label is depicted and, between (.), the justification it is based on. a [r1, r3] a a, b a, o/oo c a, b, o/oo c [r1, r3] [r2, r3] a, b a, b, o/oo c [r2, r3] (c) ( o/oo c) (c, o/oo c) (c) ( o/oo c) Figure Reasoning paths for an example default theory The branch leading to a, b, c is non-self defeating and exhaustive, so it leads to the Reiter extension Cn({a, b, c}). The same holds for the branch leading to a, o/oo c. The two branches leading to a, b, o/oo c, however, are self-defeating: the occurrence of o/oo c leads to the defeat of the earlier justification c. Therefore Cn({a, b, o/oo c}) is a constructive extension, but not a Reiter extension. 9. Conclusions In [MH93a,b] an Epistemic Default Logic (EDL) was introduced inspired by the notion of meta-level architecture that also was the basis for the approach introduced in [TT91]. In EDL drawing a default conclusion has no other semantics than that of adding a modal formula to the meta-level. No downward reflection takes place to be able to reason with the default conclusions at the object level (by means of which default assumptions actually can be made). In [TT91] downward reflection takes place, but no semantical formalization was given: it was defined only in a procedural manner. In principle downward reflection disturbs the object level semantics, since facts are added that are not logically entailed by the available knowledge. Adding a temporal dimension (in the spirit of [FG92]) to EDL enabled us to obtain formal semantics of drawing a default conclusion in a dynamic sense: as a transition from the current object level theory to a next one (where the default conclusion has been added). The approach as developed provides a formalisation of default reasoning using labeled branching time temporal models. Within these models various types of semantics can be defined: semantics based on sceptical or credulous entailment, and semantics based on constructive extensions or Reiter extensions. [TT92] Y.H. Tan, J. Treur, Constructive Default Logic and the Control of Defeasible Reasoning, in: B. Neumann (ed.), Proc. of the European Conference on Artificial Intelligence, ECAI'92, Wiley & Sons, 1992, pp. 299-303. Acknowledgements Discussions about the subject with Joeri Engelfriet have played a stimulating role for the development of the material presented here. Appendix In this appendix, we provide a proof for theorem 5.3. First, we introduce an alternative class of Kripke models. Definition A1. We define the class A(S5P) of alternative models for S5P . Elements M of A(S5P) are tuples --R A Temporal Model Theory for Default Logic Adding a temporal dimension to a logic system An Introduction to Modal logic Towards a Theory of Knowledge and Ignorance Formal Semantics of Temporal Epistemic Reflection A. Modal Logic for An Epistemic Logic for Defeasible Reasoning using a Meta-Level Architecture Metaphor A Default Logic Based on Epistemic States. Epistemic Logic for AI and Computer Science Semantical Considerations on Nonmonotonic Logic A Logic for Default Reasoning Annual Reviews of Computer Science 2 A Semantical Approach to Nonmonotonic Logics Constructive Default Logic and the Control of Defeasible Reasoning Temporal Semantics of Meta-level Architectures for the Control of Reasoning Prolegomena to a Theory of Mechanized Formal Reasoning --TR
default reasoning;temporal models;meta-level architectures;epistemic states
595884
Definability of Polyadic Lifts of Generalized Quantifiers.
We study generalized quantifiers on finite structures. With every function f : we associate a quantifier Q_f by letting Q_fx &phis; say there are at least f(n) elementsx satisfying &phis;, where n is the size of the universe. This is the general form of what is known as a monotone quantifier of type 1 . We study so called polyadic lifts of such quantifiers. The particular lifts we consider are Ramseyfication, branching and resumption. In each case we get exact criteria for definability of the lift in terms of simpler quantifiers.
to a given logic? Such questions can be quite hard. On finite models, Partially supported by a grant from the University of Helsinki. This research was initiated while the first author was a Junior Researcher at the Academy of Finland. answering some of these questions would mean solving notoriously difficult open problems about computational complexity. The results on (un)definability in the literature usually concern particular quan- tifiers; few general definability results for interesting classes of quantifiers are known. For monadic quantifiers, however, the definability question above has recently been given a complete answer in algebraic terms (cf. [18]), an answer which also permits complete classifications with respect to expressive power of various classes of monadic quantifiers. Nothing similar is in sight for polyadic quantifiers, not even for the simplest kind whose signature consists of one binary relation. A rough way of measuring complexity of a quantifier is by its arity, i.e., the largest arity of the relations occurring in the corresponding signature. So monadic quantifiers are unary, and an early result by Lindstr-om [14] is that the binary well-ordering quantifier is not definable by means of any monadic quantifiers added to first-order logic. This leads to the question of the existence of arity hierarchies of quantifiers: sequences hQ n i n2! where Q n is n-ary and not definable in terms of Q k for k ! n, or even in terms of any quantifiers of lower arity. Several such hierarchies are known; a systematic treatment is given in [6]. These hierarchy results make essential use of infinite models. Recently, some hierarchy results have also been obtained for finite models. One result by Luosto [15] is that the binary quantifier " R contains at least half of the ordered pairs of elements of the universe" is not definable from any monadic quantifiers over finite models. The proof uses advanced finite combinatorics. Another recent result, the Hierarchy Theorem of [8], states that, with respect to a more fine-grained complexity ordering than the arity ordering, there exists at every step quantifiers not definable in terms of quantifiers of lower complexity, over finite models. However, the existence of these quantifiers is proved by probabilistic methods, not by explicit construction. In this paper we shall give definability characterizations over finite models for certain classes of polyadic quantifiers, more precisely for polyadic quantifiers that are obtained from monadic ones by means of certain operations. We call these operations polyadic lifts. Three such lifts are studied: branching, Ramseyfication and resumption. Let Q be a simple unary quantifier (one whose signature contains just one unary predicate symbol), which is also monotone. Then there is a function f from natural numbers to natural numbers such that on a universe with n elements "at least f(n) ". The branching of Q (with itself) says of a binary relation R that there are subsets X;Y of the universe with at least f(n) elements such that the Y is contained in R . Similarly we can branch two distinct monotone simple unary quantifiers (requiring then that jXj - f(n) and jY j - g(n)), or k such quantifiers. The k-ary Ramseyfication of Q says that there is a set X of cardinality at least f(n) which is homogeneous for R , i.e., any k distinct elements of X stand in the relation R to each other. A resumption of a monadic quantifier, finally, says the same thing about tuples of individuals that the monadic quantifier says about individuals. So the k-ary resumption of Q says, on a universe with n elements, of a k-place relation that it contains at least f(n k ) k -tuples. These polyadic lifts turn up in various contexts. The Ramseyfication operation is familiar from model theory. The ability to quantify over k-tuples is sometimes assumed in finite model theory in order to be able to describe certain behavior of Turing machines. And all three lifts have been proposed in natural language semantics. Actually, this is the context where the idea of polyadic lifts and their definability arose, even though the definability issues themselves are standard mathematical- logical questions. The most immediate question to ask about a polyadic lift is whether it really gives anything new, i.e.: Is definable in terms of Q is the question we deal with in this paper. We prove that for the three lifts under consideration, the lifted quantifiers are usually not definable in first-order logic augmented with the argument quantifiers. More precisely, we show that a certain condition of unboundedness of the functions associated with monotone simple unary quantifiers is sufficient for undefinability of branching and Ramseyfication. Actually, our results are stronger. We show that, if Q is unbounded, the 1)-ary branching of Q is not definable from any monadic quantifiers added to first-order logic 1). And for Ramseyfication we obtain undefinability even if any quantifiers of arity at most k are allowed, and even if first-order logic is replaced by the stronger logic 1! (a logic which extends fixed point logic). Moreover, we also show that unboundedness is a necessary condition for undefinability. That is, if Q is bounded, branching and Ramseyfication of Q is definable in terms of Q. We have a similar characterization for the definability of the branching of k For resumption, our results are similar but a little weaker. We show that a slight strengthening of unboundedness is necessary and sufficient for the 1)-ary resumption of Q to be undefinable from any monotone simple unary quantifiers added to first-order logic. We also identify another necessary condition for the +1)-ary resumption of Q to be undefinable from Q, roughly that the behavior of Q on universes of size n does not determine its behavior of universes of size n k+1 . In natural language semantics, the basic quantifiers are not the simple unary ones but instead those of type h1; 1i, whose signature has two unary predicate symbols. English determiners like all, some, no, at least five, most, all but three, more than two thirds of, etc., denote such quantifiers. The lift operations arise from the way noun phrases containing determiners are combined with each other in sentences. The standard lifts here give polyadic quantifiers which are definable from the determiner denotations, but some sentences appear to use the three lifts studied in this paper. We show that our results about branching and Ramseyfication generalize to type h1; 1i quantifiers, provided these satisfy two conditions. These conditions are believed to hold (almost) universally for determiner denotations. We restrict attention to finite models, a constraint which is common in natural language semantics and model theory for computer science. The undefinability results of course hold for arbitrary models - it's just that the counter-examples we give to definability are finite. But our results that certain quantifiers are definable in a certain logics require the assumption of finiteness. In the remainder of this section we first recall the relevant notions pertaining to logics with generalized quantifiers, and then introduce the three lift operations in the context of natural language semantics. The brief linguistic excursion there is included for background and motivation, but not used in the rest of the paper. In Section 2 we state some facts about monotone quantifiers and establish that boundedness implies definability (of branching and Ramseyfication). Sections 3,4, and 5 deal with branching, Ramseyfication and resumption, respectively. In each case, undefinability is established by means of a suitable Ehrenfeucht-Fra-iss'e game, although the three proofs are quite different. Section 6 ends the paper with some questions and open problems. 1.1 Generalized quantifiers and definability We begin by recalling the concept of a generalized (Lindstr-om) quantifier, and of logics with such quantifiers (cf. [13]). A (generalized) quantifier is a class Q of structures of a finite relational signature which is closed under isomorphism. The type of Q can be identified with a finite sequence positive natural numbers (n ? 0). Equivalently, Q can be defined as a functional assigning to each non-empty set A a quantifier QA of a subset of relation over A). Instead of simply . The arity of Q is maxfk g. Q is monadic if it is unary polyadic. We let Q k be the class of all quantifiers of A quantifier Q of type hki is called simple. It is monotone if for all A and all R; S ' A k , QAR and R ' S implies QAS . The notion of a type hk quantifier being monotone in the i:th argument is defined similarly. With many familiar logics L one can uniquely associate a logic L(Q) obtained by adding the quantifier Q: a new formation rule yields formulas are formulas, x i is a string of k i distinct variables, and the semantics is given by the clause A Here A is the universe of A, a is a finite sequence of elements in A, and ' A b)g. Similarly one defines the logic L(Q) when Q is a class of quantifiers. In this paper L will be either first-order logic L !! , or L k 1! (which is like L1! except that it only has k variables), or L ! 1! . As usual, means that the same L(Q)-sentences are true in A and in B. When ' is an its quantifier rank qr(') is the maximal number of nestings of quantifier symbols (including 8 and in ', and means that the same L !! (Q)-sentences of quantifier rank at most q are true in A and in B. The relativization of a type hk quantifier Q rel defined by A (X; R This can be extended to formulas of L(Q): if P is a (new) unary predicate symbol and ' such a formula one defines inductively a formula ' (P ) of L(Q rel ) which says the same thing about (A; X) as ' says about the substructure of A generated by X . The quantifier Q is definable in L(Q) if the sentence is logically equivalent to some L(Q)-sentence in those predicate symbols. A logic extends another logic if every sentence in the latter is equivalent to some sentence in the former. In particular, L(Q 0 ) extends L(Q) iff each Q 2 Q is definable in 1.2 Polyadic lifts and natural language quantification Before defining the polyadic lifts we briefly explain how they turn up in natural language semantics. More details can be found in [10]. Monadic generalized quantifiers provide the most straightforward way to give the semantics of noun phrases (NPs) in natural languages. Consider a sentence like It consists of a noun phrase no students, in turn constituted by the determiner no and the noun students, and a verb phrase smoke, in this case a simple intransitive verb. In a model A the noun and the verb phrase denote subsets of A. Thus, the determiner is naturally taken to denote a binary relation between subsets of A, i.e., with varying A, a type h1; 1i generalized quantifier. In (1) the quantifier used is noAXY , (using the same symbol for the determiner as for its denotation). Changing the determiner in (1) we obtain other quantifiers such as all, some, at least five, most, all but three, more than two thirds of, etc., where, for example, and all but three AXY , The first argument in these quantifiers is called the noun argument, and the second the verb argument. Some determiners take more than one noun argument, as in (2) More students than teachers smoke where more. than denotes the type h1; Zj. The role of type h1i quantifiers in this context is as NP denota- tions. For example, the denotation of most students is obtained by fixing the noun argument of most to the denotation of students. Other NPs do not involve deter- miners, like something and everything which denote 9 (9AX , X 6= ;) and 8 respectively, and phrases like John or Mary, which denotes the set of X ' A such that note that whereas the determiner denotations above are all closed under isomorphism, most NP denotations are not, and thus do not qualify as quantifiers in the present sense. Determiners usually denote type h1; 1i quantifiers, but not all such quantifiers are determiner denotations. Indeed, the following two constraints have been found to hold (almost) universally: CONS QAXY () QAX X "Y CONS predicts, for example, that the H-artig quantifier I AXY () is not the denotation of any English determiner. And EXT implies that no determiner can mean, say, some on universes with less than 10 elements and all on other universes. Nevertheless, the determiner denotations form a rich class of quantifiers: as further examples we may take the proportional quantifiers and more than (at least) m n:ths of the, and Boolean combinations of these, such as less than half of the, between ten and twenty percent of the, not more than ten or at least twenty percent of the, etc. Together, CONS and EXT express a characteristic asymmetry between the two arguments of determiner denotations: the role of the noun argument is to restrict the domain of quantification. A precise statement is the following 1.1 Fact. A type h1; 1i quantifier is CONS and EXT iff it is the relativization of some simple unary quantifier. Clearly, Q rel is always CONS and EXT and, conversely, if Q 0 is CONS and EXT defined by QAY , Q A AY , Q This simple fact often enables one to generalize results about simple unary quantifiers to determiner denotations. Notice also that a simple unary Q is monotone iff Q rel is right monotone (mono- tone in the verb argument). For example, the proportional quantifiers above are right monotone. Thus, NP semantics involves monadic quantifiers. But sentences can combine several NPs, together with transitive or ditransitive verbs denoting not sets but binary or ternary relations. Their truth conditions can then be given by means of polyadic quantifiers. Indeed, common sentential structures correspond to ways of lifting monadic quantifiers to polyadic ones. The most ubiquitous lift is iteration, as in (3) Most students criticized three teachers with a quantified subject and object (Q 1 X and Q transitive verb (R ). The truth conditions (of one of the two readings of (3)) are or, equivalently, the result of applying the lifted quantifier It(Q to the arguments X;Y and R : define (suppressing (R seemingly similar sentences sometimes correspond to other lifts than iteration. For example, researchers wrote thirty-two papers for the Handbook presumably means neither that each of the forty researchers wrote thirty-two papers, nor that each of the thirty-two papers was written by forty (co)authors. Instead, this is an instance of so-called cumulative quantification: each of the forty researchers authored at least one paper for the Handbook, and each of the thirty-two papers was authored by at least one researcher. Thus, we have the lift We now come to the lifts studied in this paper. The first was introduced by Barwise in [1] with examples such as Most boys in your class and most girls in my class have all dated each other Here the lift in question, on at least one plausible reading, is branching. 1.2 Definition. For right monotone type h1; 1i quantifiers (again suppressing the universe) We write Br k (Q) for arguments). The branching of k monotone type h1i quantifiers is defined similarly (suppress So (8) means that there is a set X containing more than half the boys in your class and a set Y containing more than half the girls in my class such that any pair in X \Theta Y is a 'dating' pair. A similar lift occurs in certain reciprocal sentences like At least two thirds of the boys in your class all like each other which can be taken as saying that there is a set containing at least two thirds of the boys in your class such that any two distinct boys in this set are in the 'like' relation: 1.3 Definition. Let Q be a right monotone type h1; 1i quantifier. (where I(Y )g). Similarly for a monotone simple unary Q. We call the values of this lift Ramsey quantifiers, extending slightly the common usage of this term (where Q is the quantifier 'there exists a least @ ff '). Our final example of a polyadic lift, resumption, amounts to using a monadic quantifier to quantify over k-tuples of individuals. This lift has found uses in computer science (cf. [4] and [16]) but can be given a linguistic motivation as well; cf. cases like Most neighbours like each other Most twins never separate 1.4 Definition. Let Q be of type h1; 1i. Define (Res k (Q))ARS () Q A k RS Similarly for simple unary quantifiers. The route to the lifts here went via linguistics, but they are natural in other contexts too. For example, to say that a graph contains so and so many vertices of degree 3 one iterates monadic quantifiers. But to say that a graph contains a clique of such and such size, one needs a Ramsey quantifier. Of course, one could also use second-order logic. But the Ramsey quantifier gives a better estimate of just how much expressive power one needs to add to first-order logic in this case. To know how powerful the lifts are one needs first of all to answer the basic logical definability question: When is (The question is also significant in a linguistic context. For example, non-definability may imply that a certain kind of grammar for the corresponding expressions does not exist.) Trivially, iterations and Boolean combinations thereof are so definable. We shall prove below that the other polyadic lifts - branching, Ramseyfication and resumption - are usually not definable in this way. Monotone quantifiers ?From now on, unless otherwise noted, we restrict attention to finite models. Then, a monotone simple unary quantifier can be identified with a function quantifier corresponding to f says of a set X that it has at least f(n) elements, where n is the cardinality of the universe (if f(n) ? n this is trivially false). More precisely, given f , define for A. Conversely, if Q is monotone, and least k s. t. 9A9X ' if such a k exists then Rather perspicuous definability criteria can be given for monotone simple unary quantifiers. A full characterization can be found in [18]; here we quote for illustration the following special cases: 2.1 Theorem. 1. ([11], [20]) The quantifier Q f is first order definable iff either f(n) or n \Gamma f(n) is eventually constant. 2. ([18]) Suppose that lim n!1 that either lim is definable in L !! (Q f ) iff at least one of eventually constant. In the second of these statements Q f is assumed to satisfy a strong unboundedness condition. For most of this paper, the following weaker notion is sufficient. 2.2 Definition. We say that the function f (or quantifier Q f ) is unbounded if Canonical examples of unbounded functions are n ], and [log(n)] (where [a] is the integer part of a). Typical bounded functions are (corresponding to 9), 2.3 Theorem. If f is bounded, then for any k - 2, Br k (Q f ) and Ram k (Q f ) are definable in L !! (Q f ). Proof. By hypothesis, there is a number m such that for every n, either 0, or n. We can use Q f to uniquely characterize each of these possibilities for f(n). To see this, note first that are the sentences respectively. Second, for p ? 0 we have Third, let / q be the conjunction of y y and :- . Then Moreover, for each one of these finitely many possibilities, the branching conditio is expressible in first-order logic. Indeed, if condition is equivalent to the sentence states that all components of x are distinct, and H is the set of all functions h pg. In the case the branching condition is equivalent to the sentence simply finally, if 1), then the branching condition is trivially true (false). Putting all this together we conclude that Br k (Q f )x equivalent to the sentence 1-p!m q!m The case of Ram k (Q f ) is similar. 2 The relativization Q rel f of Q f is the right monotone type h1; 1i quantifier (Q rel We say that f is bounded if f is bounded. Notice that Q f is definable in f ), that Br(Q f 1 is definable in L !! (Br(Q rel )), and that Ram k (Q f ) is definable in L !! (Ram k (Q rel f )). ?From Fact 1.1 we see that right monotone determiner denotations usually are of the form Q rel f . For example, the quantifier most is Q rel f for 1. And in most of these cases, relativization increases expressive power. In fact, the following result tells us precisely when this happens. 2.4 Theorem. ([11], [20]) The quantifier Q rel f is definable in L !! (Q f ) (or in eventually constant (i.e., iff Q f is first-order definable). Note that Q rel is definable in L(Q) iff L(Q) has the relativization property, i.e., for any L(Q)-sentence ', ' (P ) is equivalent to an L(Q)-sentence. We end this section by stating a relativized version of Theorem 2.3 (these theorems will be generalized to the branching of several quantifiers in section 3). 2.5 Theorem. If f is bounded, then Br k (Q rel f ) and Ram k (Q rel f are definable in Proof. Similar to the proof of Theorem 2.3. To formulate the branching condition Br k (Q rel over a universe A, we need to express statements of the (p ? 0, q - 0), and this can be done with the quantifier Q rel f . 2 3 Definability of branching In this section we first recall definitions and basic properties of some Ehrenfeucht- Fra-iss'e type games, and then apply these games to the definability of branching. The bijective Ehrenfeucht-Fra-iss'e game of length q , BEF q (A; B), has two play- ers, which we call Duplicator and Spoiler , respectively. In each round 1 of the game Duplicator chooses first a bijection f by choosing an element a i 2 A. These q pairs of moves of the players determine a relation Duplicator wins the game if p is a partial isomorphism A ! B, i.e., if p is an injective function such that for each relation R of A and the corresponding relation R 0 of B, and for all tuples (b (b Spoiler wins if this is not the case, or if Duplicator cannot make his moves, i.e., if there are no bijections As usual, what happens in some particular play of BEF q (A; B) is not important. It is rather the question "which one of the players has a winning strategy" that is relevant for us. Here we say that Duplicator (Spoiler) has a winning strategy if he has a systematic way of choosing the bijections f i (the elements a i ) such that using it he always wins the game, no matter how the other player moves. The following result is a direct consequence of Theorem 2.5 of [6]. For the sake of completeness we will explain here the basic idea behind its proof. 3.1 Proposition. If Duplicator has a winning strategy in the game BEF q (A; B), then A j q B. Proof. (idea) Consider a formula of L !! (Q 1 ) and a partial function a 7! b. If f is a bijection such that for every a 2 A and each 1 - j - m, A then the function f is an isomorphism between the defined structures (b; \Delta)) . Hence, A only if B no matter what the interpretation of the quantifier Q is. Using this observation it is easy to prove by induction on r that if f are bijections which Duplicator has chosen according to his winning strategy, and are the elements chosen by Spoiler, then A for every formula ' of L !! (Q 1 ) with qr(') - r . In particular, the equivalence A holds for every sentence ' of quantifier rank - q . 2 If the structures A and B are finite, then the converse of Proposition 3.1 is also true: A j q only if Duplicator has a winnning strategy in BEF q (A; B). Furthermore, this equivalence holds for infinite structures, too, if the logic L !! (Q 1 ) is replaced with the infinitary logic L1! (Q 1 ) (cf. [6]). However, the implication stated in Proposition 3.1 is all that is needed for proving undefinability results. 3.2 Corollary. Let K be a class of structures. If for every natural number q there exist structures A and B such that A 2 K , B 62 K and Duplicator has a winning strategy in BEF q (A; B), then K is not definable in L !! (Q 1 ). Below we will give a non-trivial application of BEF q (A; B) to finite models. But to get a feeling for the game, it may be helpful to to first see the use of Corollary 3.2 with a much simpler example, which involves infinite models. 3.3 Example. For each natural number m, let Em be an equivalence relation on equivalence classes, all of cardinality ! . Similarly, let an equivalence relation with ! equivalence classes of cardinality ! . We claim that To see this, consider an arbitrary partial isomorphism such that . For each i, let A i be the set of all elements of ! which are Em -equivalent with a i , and let B i be the set of those elements which are E n -equivalent with b i . Clearly all these sets are of cardinality ! . Furthermore, since r ! q - m;n, the sets are also of cardinality . Hence, there exists a bijection extending p which maps A i onto B i is a partial isomorphism In particular, Duplicator can keep choosing bijections in the game BEF q ((!; Em in such a way that the mapping f (a is always a partial isomorphism. Let K be the class of all structures (A; E) such that E is an equivalence relation on A with finitely many equivalence classes. Since (!; Em but it follows from Corollary 3.2 that K is not definable in L !! (Q 1 ). In a similar fashion we see that the class K 0 consisting of those equivalence relations E) that have an even number of equivalence classes is not definable in L !! (Q 1 ). Historical remark. The example is a variant of Keisler's counterexample to interpolation in L !! (Q 1 ). It was realized by Caicedo [3] and V-a-an-anen (cf. [12]) that this kind of example (though with infinitely many equivalence classes) works for all monadic quantifiers; a proof using the BEF game was given in [6]. 3.1 Branching of one quantifier In Section 2 we saw that branching a bounded quantifier does not have any effect on its expressive power. However, the following result shows that branching usually increases the expressive power even beyond the reach of any monadic quantifiers. 3.4 Theorem. Let Q be a monotone simple monadic quantifier. If Q is un- bounded, then Br 2 (Q) is not definable in L !! (Q 1 ). Putting together Theorems 2.3 and 3.4 we obtain a complete characterization of the definability of Br 2 (Q) in terms of Q for any monotone simple monadic quantifier Q. Using Theorem 2.5 (and the fact that Br 2 (Q f ) is definable in L !! (Br 2 (Q rel we get the same characterization for relativizations of such quantifiers. Moreover, 2 can be replaced by any k ? 1, since we have 3.5 Lemma. If Q is a monotone simple unary quantifier, or the relativization of such a quantifier, then Br k (Q) is definable in L !! (Br k+1 (Q)). Proof. When simply which is seen by considering separately the three cases (1) 1. The relativized case is similar. 2 Applying also Fact 1.1, we obtain the following characterization. 3.6 Corollary. Suppose that Q is either a monotone type h1i quantifier, or a right monotone type h1; 1i quantifier satisfying CONS and EXT. Then the following conditions are equivalent, for any k ? 1: (1) Br k (Q) is definable in L !! (Q). (2) Br k (Q) is definable in L !! (Q 1 ). (3) Q is bounded. In [9], Hella and Sandu proved, modifying a model construction by Fagin [5], that connectivity of (finite) graphs is not definable in terms of monadic quantifiers. Their proof implies a special case of Corollary 3.6: Br 2 (most) is not definable in Here we will use a simple trick to modify this proof so that it works for any unbounded quantifier Q f . 3.7 Definition. Let n be a natural number and E) a graph. We say that A is n-separable if there exist subsets C; D ' A such that (C \Theta D) " by K f the class of all graphs A which are f(jAj)-separable. Here all graphs are assumed to be undirected, but not necessarily irreflexive. Thus, a structure (A; E) is a graph if E is a symmetric binary relation on A, possibly containing self-loops (a; a). Observe now that a graph (A; E) is in the class K f if and only if its complement graph E) is in the quantifier Br 2 (Q f ). Hence, we have 3.8 Lemma. For any function f the class K f is definable in the logic We shall now define two families of graphs that are the heart of the proof of Theorem 3.4. 3.9 Definition. Let q , r and s be natural numbers, and let P , P 0 , R and S be mutually disjoint sets of cardinalities respectively. Assume further that 1. We define two graphs A q;r;s = (A; E) and B g. In the case simply by A q and B q . Thus, one can visualize the graphs A q;r;s and B q;r;s as follows. A q;r;s consists of a cycle (with self-loops) of length two 'boxes' R and S containing r and s elements, respectively. Each element in S is connected by an edge to all elements of A (including itself), whereas the elements in R are not adjacent to any other elements than those in S . The graph B q;r;s is similar, except that instead of the big cycle it contains two cycles of length 2 3.10 Lemma. For all natural numbers q , r and s, Duplicator has a winning strategy in the game BEF q Proof. We start by observing that the subsets R and S of the universe of A q;r;s and B q;r;s can be omitted in this proof without loss of generality: Duplicator has a winning strategy in BEF q only if he has one in BEF q Indeed, the mapping f 7! f [ id R [ id S for each bijection any winning strategy of Duplicator in BEF q to a winning strategy in BEF q On the other hand, if Duplicator wins the game between the larger structures A q;r;s and B q;r;s , then he wins the game between A q and B q simply by choosing the restrictions of his bijections to the set P . Hence, it suffices to show that Duplicator can always win the game BEF q This was already done in [9] (for cycles without self-loops), but for the sake of completeness we give here a new proof. For each k ! be the bijection defined by ae (Here we are stipulating that if t - similarly for e i .) Thus, g k arises by splitting the big cycle in A q into two halves and mapping each of these halves, in a uniform way, onto one of the small cycles in B q . We will show below that Duplicator wins the game BEF q choosing bijections of the form g k for suitable k . Before describing this winning strategy we need to introduce some auxiliary concepts. Let C be a subset of P and k and l integers ! 2t. We say that k and l are C -equivalent if g l (a) = g k (a) for every a 2 C . The splitting points of g k are . The distance of C from the splitting corresponding to g k is splitting pointg; denotes the usual distance between elements in a graph. Furthermore, for each a 2 P , we denote by h k (a) the splitting point which is closest to a. Note that if d k (C) ? 0, then the restriction of the bijection g k to the set C [ fag is a partial isomorphism A q ! B q for any a 2 P . We claim now that Duplicator can choose his bijections f in the game BEF q in such a way that, for each 1 - i - q , (1) there is k ! 2t such that f g-equivalent to f where a are the elements chosen by Spoiler. This constitutes the promised winning strategy for Duplicator, since condition (2) implies that the mapping is the restriction of f q to the set fa and this restriction is a partial isomorphism by condition (1) for . The claim is proved by induction on i: (i) We let f note that conditions (1) and (2) are trivially satisfied (ii) If a is the unique natural number ! 2t such that k+2 q\Gamma2 assume as induction hypothesis that conditions (1) and (2) hold for the bijection f . There are two possibilities: If h k (a i g, we choose f l for . Then we have d l (fa and since there are no elements of fa l or between c t+k and c t+l , g l is g-equivalent to g k . In the case h k (a i we let f l for conditions (1) and (2) are seen to hold by a similar argument.Proof of Theorem 3.4. Note first that in the graph B q;r;s there are no edges between the sets whence B q;r;s is (t+r)-separable. On the other hand, A q;r;s is not (t+r)-separable, since if C; D ' A are sets of cardinality t +r , then either there is an element a of S in C or D , or there are elements 1g. In the former case, there is an edge between a and every element of C and D , and in the latter case, there is an edge between c i and c j (here is where the self-loops are used). Now given Q f with f unbounded we can choose n such that t - f(n) - t. equals the size of the graphs A q;r;s and B q;r;s , and . By the observation above, B q;r;s is f(n)-separable, but A q;r;s is not. From Corollary 3.2 and Lemma 3.10 it follows that the class K f is not definable in L !! (Q 1 ). Since, by Lemma 3.8, K f is definable in L !! (Br 2 (Q f )), it follows that Br 2 (Q f ) is not definable in L !! (Q 1 3.2 Branching of several quantifiers We shall now extend the characterization in Corollary 3.6 to the branching of more than one quantifier. The first thing to observe is that if both f and g are unbounded, it does not necessarily follow that Br(Q f ; Q g ) is undefinable in L !! (Q f ; Q g ). For example, let f(n) be n=2 if n is even and n+1 otherwise, and let g(n) be (n+1)=2 if n is odd and n+1 otherwise. Then Br(Q trivially false of every relation, hence definable in L !! , although Br 2 (Q f ) and Br 2 (Q g ) are both undefinable in What we need for ) is that f and g are 'jointly unbounded' in the sense that the proof of Theorem 3.4 can be modified to show that Br(Q f ; Q g ) is not definable in L !! (Q 1 ). And in fact the converse also holds: if f; g are not jointly unbounded it can be shown that Br(Q f ; Q g ) is definable in L !! (Both of these claims will follow from from Theorem 3.12 below.) But these observations on the branching of two quantifiers do not generalize immediately to the branching of k quantifiers. First, the condition on f; g; h of not being jointly unbounded in the above sense turns out to be too weak to permit the conclusion that Br(Q f ; Q Lemma 3.5 does not generalize directly to the branching of several quantifiers. This is because on a model A where false, whereas if If we know that h(n) is never 0 or n definable in terms of as in the proof of Lemma 3.5, but not necessarily otherwise. It turns out, however, that with a more careful generalization of the notion of (un)boundedness to sequences of functions, we can obtain a necessary and sufficient condition for the definability of the branching of k quantifiers. 3.11 bounded if there exists a number t such that for all n, either holds for at most one i. Otherwise hf We also say that are (un)bounded under these circumstances. The relation to the previous concept of boundedness for one function is the (observe that if P holds of f i for at most one i, k ? 1, and all the f i are equal, then P does not hold of any f i ). Also, note that if each of f or even if all but one of them are bounded, then hf The following theorem generalizes Corollary 3.6. 3.12 Theorem. Let hQ be either a sequence of monotone type h1i quan- tifiers, or a sequence of right monotone type h1; 1i quantifiers satisfying CONS and 1). Then the following conditions are equivalent: definable in L !! (Q 1 ). Proof. (outline) Let hQ i. Claim: There are such that for every t there is n such that (i) for all l , f l (n) 6= 0 and f l (n) For, by the unboundedness of hf holds that for every t there is n and a pair such that (i) and (ii) hold. But there are only finitely many pairs so for at least one of them, (ii) holds for infinitely many t. From this the readily follows. Let, be as in the Claim. Note that now we are essentially back to the case mentioned above with two 'jointly unbounded' functions. The proof of Theorem 3.4 is modified as follows. In the two models A q;r;s and B q;r;s , split the into two subsets R 1 and R 2 with r 1 and r 2 elements, respectively. R 1 and need not be disjoint but their union is R . Expand the models to A q;r 1 ;r 2 ;s and by adjoining two unary predicate symbols P 1 and P 2 , which in A q;r 1 ;r 2 ;s are interpreted as P respectively, and in B q;r 1 ;r 2 ;s as and elements in R 1 and R 2 are only connected to elements in S , Duplicator still has a winning strategy in the game BEF q Consider the sentence Take any quantifier depth q and choose n such that (i) and (ii) of the Claim hold. Say So we take R R of size r 2 , and R 1 as a subset of R of size r 1 . Then n is the size of the models A q;r 1 ;r 2 ;s and B q;r 1 ;r 2 ;s and f i . It is then easy to see, using the fact that none of f 1 that the sentence above is true in B q;r 1 ;r 2 ;s but false in A q;r 1 ;r 2 ;s . Assume that hf is bounded and let t be the bound. For i be L !! (Q f i )-sentences such that (cf. the proof of Theorem 2.3) in a model with universe A, Let Finally, let \Phi s be a list of all conjunctions built up from the sentences ' i , - i , ' r which specify the value of f i (jAj) for all but one i, and also state that f i (jAj) is not jAj + 1 or 0 for any i. It follows from the assumption that in any model, either \Theta or \Xi or one of In models of \Theta the branching sentence is false, and in models of :\Theta - \Xi it is true. Moreover, we make the following Claim: For each \Phi j there is a sentence \Psi j in L !! (Q f 1 such that in models of \Phi j , ?From this it follows that the branching sentence is logically equivalent to the sentence (:\Theta - \Xi) - s Proof of the Claim: An example will suffice to give the idea. Suppose that and that \Phi j says that f 2 that none of f 1 (jAj), (written semi-formally) is 9 distinct y (We have and the conjunct " x 3 62 fz above.) For this sentence says that there are sets Y and Z g) such that jY Using the monotonicity of Q f 1 one sees that this is the same as saying that there are sets X;Y;Z such that jY which, by the monotonicity of Q f 2 and , is precisely what the branching sentence holds. This proves the Claim, and thereby the Theorem in the unrelativized case. The case with relativized quantifiers is similar (cf. the proof of Theorem 2.5). 2 4 Definability of Ramseyfication In this section we show that the Ramsey lift Ram k+1 (Q) of an unbounded quantifier Q is not definable in L !! (Q). So the situation is analogous to the case of the branching lift: the lift of Q is definable in L !! (Q) if and only if Q is bounded. However, in two important respects our results about the Ramsey lift are stronger. Namely, we prove that for unbounded Q the lift Ram k+1 (Q) is not definable even in Recall that L ! 1! is the fragment of the infinitary language L1! in which every formula contains finitely many different variables only, and Q k is the family of all k-ary generalized quantifiers. Thus we can prove the undefinability of Ram k+1 (Q) even with respect to infinite disjunctions and conjunctions and even with arbitrary k-ary generalized quantifiers. The main interest of L ! is in the fact that it contains various fixpoint extensions of L !! (Q k ) (see [7, 11]). Intuitively speaking, L ! by allowing recursive definitions. In Section 3 we introduced the game BEF q (A; B). The point of this was that if Duplicator has a winning strategy in BEF q (A; B), then A j q B. We shall now recall a modification BP l k (A; B) of BEF q (A; B) from [7] in order to get a criterion for A j L l 1! (Q k ) B. The k -bijective l -pebble Ehrenfeucht-Fra-iss'e game, BP l k (A; B) is defined as fol- lows: In each round q of the game Duplicator chooses first a bijection f and Spoiler responds by choosing sets C q ' A and D q ' A so that jD q and jC q [ D q j - l . Intuitively speaking, Duplicator claims that f q is an isomorphism and Spoiler tries to dispute this by pointing to the part C q [ D q of A where he thinks f q does not preserve structure. To check this out the referee defines (letting rule dictates that Spoiler has to play in such a way that C q ' dom(p wins the game if for all q the relation p q is a partial isomorphism A ! B. Spoiler wins if this is not the case, or if Duplicator cannot make his moves, i.e., if there are no bijections The difference between BP l is the following: In the latter game all moves of Spoiler remain in the game and Duplicator has to create bigger and bigger partial isomorphisms, whereas in the former game Spoiler can only up to l previously played elements so the task of Duplicator is easier but, on the other hand, Duplicator does not know how many moves the game has. The game BP l introduces the new feature that Spoiler can let the set D q of "new" elements contain up to k elements. 4.1 Proposition. A j L l 1! Duplicator has a winning strategy in the game BP l Proof. See the proof of Proposition 3.1 and the proof of Corollary 5.9 in [7]. 2 4.2 Corollary. Let Q be a quantifier. If for some sentence ' of L !! (Q) and for all natural numbers l there exist structures A and B such that A and Duplicator has a winning strategy in BP l then Q is not definable in We shall consider the lift Ram 2 (Q) first, because the construction behind our result on the general case of Ram k (Q) is different. 4.3 Theorem. If f is unbounded, then Ram 2 (Q f ) is not definable in L ! Proof. Let l - 1 be arbitrary. We show that Ram 2 (Q) is not definable in The proof is based on a modification of the proof of the main result of [2]. As in [2], let X be the colored graph which has the following vertices: ffl a for each subset S of ng of even cardinality. ffl The vertices a i and b i have the same color i and are called mates of each other, ffl The vertices m S are colored magenta, ffl If i 6= j , then a i and a j have different color, and the following edges: ffl a i and m S are joined by an edge if i 2 S , are joined by an edge if i 62 S . The relevant properties of these graphs are ([2]): ffl For any even number of pairs there is an automorphism of X n which swaps a i and b i for these i and leaves other a j and b j fixed. Every automorphism of X n is obtained in this way. 2. Let T be a copy of K n , i.e. the complete graph with n vertices. We define a new graph G as follows: For each vertex v of T, we replace v by a copy X(v) of X n\Gamma1 . These copies are called gadgets. We endow X(v) with an additional color which is different for different v . For each w 6= v in T we select one pair from X(v) and one pair fa i(w;v) ; b i(w;v) g from X(w), in such a way that all selected pairs in each gadget are distinct. Then we connect with an edge a i(v;w) to a i(w;v) and b i(v;w) to b i(w;v) . This ends the description of G. The graph H is defined similarly, except that for one edge (v T, we create a "twist" by connecting a i(v 0 ;w 0 ) to b i(w 0 ;v 0 ) and b i(v 0 ;w 0 ) to a i(w 0 ;v 0 ) . Note that because of the coloring, an automorphism of G or H induces an automorphism of each gadget, and vice versa. The crucial properties of G and H are: ffl For any even number of edges of T there is an automorphism of G (H) which twists f(v leaves other edges fixed. Every automorphism of G and H is obtained in this way. ffl For any edge (v; w) of T there is an isomorphism G ! H, where G is obtained from G by twisting (v; w) and leaving other edges fixed. ffl G and H are not isomorphic. Since f is unbounded, there is a natural number m so that d - be the disjoint union of G and two complete graphs K u and K v , where Likewise, let H 0 be the disjoint union of H, K u and K v . In both graphs the clique K u is colored beige and K v violet. Note that the number of vertices of both G 0 and H 0 is exactly m. The automorphisms of G 0 consist of an automorphism of G plus permutations of K u and K v . The same holds for H 0 . Thus all the automorphisms of G 0 and H 0 are known. The same is true of isomorphisms G 0 ! H 0 , where G 0 is obtained from G 0 by twisting an edge (v; w) of T and leaving other edges fixed. We shall now show that G 0 j L l 1! by describing a winning strategy of Duplicator in BP l The task of Duplicator is to choose bijections G Suppose Duplicator has played f q\Gamma1 , Spoiler has played C q\Gamma1 and D q\Gamma1 and the referee has defined which is indeed a partial We assume that Duplicator found f by choosing a vertex v q\Gamma1 from T, twisting an edge adjacent to it and letting f be the arising isomorphim onto H 0 . Naturally, v q\Gamma1 had to be chosen carefully. Now, Duplicator chooses v q 6= v q\Gamma1 so that X(v q ) does not meet C This is possible because n ? l + 1. Then he twists the edge (v q ; v q\Gamma1 ) and lets f q be the arising isomorphism onto H 0 such that p q\Gamma1 ' f q . Next Spoiler plays C q and D q and the referee defines p This is a partial isomorphism is a singleton and no edge with one end in dom(p twisted when f q was defined. This ends the description of the winning strategy of Duplicator in BP k ending the proof of G We shall next describe the sentence of L !! (Ram 2 (Q)) that separates these two graphs. The sentence stipulates the existence of a set which contains all beige elements and exactly one element from the magenta part of each gadget X Once we have this one fixed element x in the magenta part of a gadget, we can divide each pair (a into the element with an edge to x - call it "lower" - and the element without an edge to x - call it "upper" . The sentence we need says: There is a set X such that jXj - f(m) and any two distinct elements x and y from 1. x is magenta or beige, 2. If x and y are not beige, then they are in different gadgets, 3. If there is a connection from the gadget of x to the gadget of y , it connects "upper" elements to "upper" elements. To see that this sentence really separates G 0 and H 0 we first note that the sentence is true in G 0 , since we can let X consist of all beige elements plus the element m ; from all gadgets. Then On the other hand, suppose such a set X existed in H 0 . Since elements of X are either beige or in different gadgets, actually consists of exactly the beige elements and exactly one vertex from the magenta part of each gadget. This element has the form m S for some set S of even cardinality, so the a i for i 2 S are "lower". Now, in each gadget, flip the even number of pairs (a leave the others. This results in an automorphism h of H 0 such that, in each gadget, But with respect to m ; , all the a i are "upper". Since part 3 of the sentence is true (for h(x) and h(y)), it follows that H 0 must be isomorphic to G 0 , which is a contradiction. Thus the sentence is false in H 0 , and the proof is finished.4.4 Theorem. If f is unbounded, then Ram k+1 (Q f ) is not definable in L ! Proof. The case Theorem 4.3, so we assume k ? 1. Fix l - 1. We use the models constructed in [7]. By general results in [7] we have A j L l 1! (Q k ) B. By [7, Corollary 8.8] A and B can be separated by We shall make a little modification to the proof of [7, Corollary 8.8] in order to cover an arbitrary unbounded f . Let x be the size of the graph G. Note that A has a homogeneous subset of cardinality does not. With this in mind, and since f is unbounded, we choose an n with extend the model A with two disjoint sets U and V , where jU and f(n). The elements of U are colored beige and the elements of V violet. The same extension is applied to B. We get two new models A 0 and . We can separate these models with the sentence Ram k+1 (Q)x is the disjunction of /(x Corollary 8.8] and the formula " x 1 is beige".Combining Theorems 2.3 and 4.4 we obtain a complete characterization of the definability of Ramseyfication for all monotone simple monadic quantifiers. And as in the case of branching, the same characterization holds also for relativizations of such quantifiers (this follows from Theorem 2.5 and the fact that Ram k (Q f ) is definable in L !! (Ram k (Q rel f ))). 4.5 Corollary. Suppose that Q is either a monotone type h1i quantifier, or a right monotone type h1; 1i quantifier satisfying CONS and EXT. Then the following conditions are equivalent, for any k - 1: (1) Ram k+1 (Q) is definable in L !! (Q). (2) Ram k+1 (Q) is definable in L !! (Q k ). (3) Q is bounded. 5 Definability of resumption We do not have as strong results for Res k (Q) as for Br k (Q) and Ram k+1 (Q) above. In fact, it is proved in [11] that for type h1i Q, Res k (Q) is always definable in extends to resumptions of any monadic quantifiers). But we prove that if f is unbounded in a sense which is appropriate from the resumption point of view, then Res k (Q f ) is not definable in L ! and hence not in L !! (M), where M is the collection of all monotone type h1i quantifiers. For this result we need again a new version of an Ehrenfeucht-Fra-iss'e type game from [11]. Ig be a family of type h1i quantifiers, A and B two structures, and l a positive integer. The game MEF l (Q; A;B) has the following rules: There are l pairs of pebbles. During the game Spoiler may take a pebble and put it on an element of one of the models, and then Duplicator has to take the corresponding pebble and put it on an element of the other model. If all pebbles have been used already, Spoiler can remove one of his pebbles from one of the models and reposition it on another element of one of the model. In such a case Duplicator has to reposition the corresponding pebble on some element of the other model. These pebble moves are called elementary moves. There is another type of move that Spoiler can choose to make. This is called a quantifier move. In this move Spoiler takes a pebble, one of the structures A and B, say A, a quantifier Q j from the family Q, and a subset X of the universe A of A such that the structure (A; X) is in the quantifier Q j , and the set X is invariant under all automorphisms of A that fix all pebbled elements. Duplicator must respond by choosing a subset Y of the universe of the other structure (in this case Y must be a subset of the universe B of B) such that the structure (B; Y ) is in the quantifier Q j . Then Spoiler places the pebble on an element b 1 2 Y and Duplicator must respond by placing a pebble on an element a 1 2 X . In each round of the game Spoiler first makes an elementary move or a quantifier move and then Duplicator responds as described above. This is repeated indefinitely. After each round the referee checks if Spoiler has won already. For this, let a i and , be the elements of A and B respectively pebbled by the two players using the up to k pebbles available. If the mapping a to be a partial isomorphism between A and B, then Spoiler wins. Otherwise, the game goes on. If the game lasts for infinitely many moves without Spoiler winning, then Duplicator is declared the winner. 5.1 Proposition. ([11]) Let Ig be a family of monotone type quantifiers, A and B two finite structures, and l a positive integer. Then the following statements are equivalent: (3) Duplicator has a winning strategy for the game MEF l (Q; A;B). 5.2 Corollary. ([11]) Let Q be a quantifier and Q a family of monotone type h1i quantifiers. If for some sentence ' of L !! (Q) and for all natural numbers l there exist structures A and B such that A and Duplicator has a winning strategy in MEF l (Q; A;B), then Q is not definable in L ! 1! (Q). We shall now give some examples concerning the definability of Res k (Q f ) in The examples display characteristic features of resumption in this respect. 5.3 Proposition. Both boundedness of f and unboundedness of f are consistent with Res k (Q f ) being definable in L !! (Q f ), but neither condition is sufficient or necessary for this to hold. Proof. Suppose is a k 'th power and Then f is bounded and Q f is not first order (cf. Theorem 2.1), but Res k (Q f ) is. Hence Res k (Q f ) is trivially definable from Q f . Thus boundedness is consistent with Res k (Q f ) being definable in L !! even when Q f is not. To see that boundedness is not sufficient for Res k (Q f ) to be definable from Q f , suppose is the k'th power of a prime and Then f is bounded and Q f is not first order. Moreover, Res k (Q f ) is not definable from Q f , as the following argument shows: We can use Res k (Q f ) to define the property "the size of the universe is a prime". But we can use Corollary 5.2 to show that this property is not definable from l is a natural number. l be a prime such that wins the game Unboundedness of f is consistent with Res k (Q f ) being definable in L !! even when Q f is not. To see this, suppose is a k 'th power and otherwise. Then f is unbounded and Q f is not first order, but Res k (Q f ) is. Finally, suppose is the k'th power of a prime and otherwise. Then f is unbounded, Q f is not first order, and Res k (Q f ) is not definable from Q f , because "the size of the universe is a prime" is definable from Res k (Q f ), but not from Q f . 2 5.4 Definition. Let f be a function. The function f 0 is said to be determined by f if for all m;n 5.5 Proposition. Suppose f is bounded and f(n k ) is determined by f . Then Res k (Q f ) is definable in L !! (Q f ). Proof. Let m, - , ' p , / q (p ? 0; q - 0) be as in the proof of Theorem 2.3, and let ' sentence in L !! saying of a k-ary predicate R that the number of k-tuples it contains is at least i (jAj k \Gamma i). By assumption there are functions g and h such that f(n k Res k (Q f )x equivalent to the L !! (Q f )-sentence i!m i!m )The condition of Theorem 5.5 holds, for example, for the following divides n Thus Res k (Q fp ) is definable from Q fp . By a result of M. Mostowski ([17]), resump- tions of the stronger divisibility quantifiers divides the number of elements x satisfying OE(x): are likewise definable from the quantifiers themselves. is unbounded on X , if is bounded on X . The following theorem is the main result of this section: 5.7 Theorem. Suppose f is unbounded on fn k 2. Then Res k (Q f ) is not definable in L ! Proof. For every m there is an n so that m - f(n k m. This f(n k ) is either - [n k =2] or ? [n k =2]. One of these possibilities occurs infinitely often, and hence we have or We start by assuming (1) and indicate at the end of the proof how to handle the case that (2) holds. Suppose are functions and l is a natural number. We are going to prove that Res k (Q f ) is not definable in L l consist of intervals of the following form: We call a real a 2 [0; n] n-good, if a is not on any of the intervals in I n . A set if each of its elements is. A direct calculation reveals that if I is a subinterval of [0; n] of length - there is an n-good subinterval J of I of length x. Let C . By (1) there is an n such that -=n . s s In either case we have proved the claim. 2. There are an n-good x and a non-integer n-good y ? x on [0; n] so that x By Claim 1, there is an n-good subinterval I of [-] of length ? C 1 . On the interval [-] the function F derivative so the interval F 00 I has length ? C 1 . Hence there is a non-integer n-good I . Now also x is n-good and x proved. Let a be the integer part of x, and b the integer part of y . We define now two structures ng ag bg. 3. Duplicator wins the game MEF l (fQ The elementary moves are easy for Duplicator. Suppose then Spoiler chooses a pebble, the model A, the quantifier Q and an invariant subset X of A with furthermore that the pebbled elements in A are a that Duplicator has not lost the game yet. Let the correspondingly pebbled elements of B be b . For any set Z and d 2 f0; 1g let Z is invariant under automorphisms of A that fix the elements there are d 1 ; d 2 ; d 3 2 f0; 1g so that where Let and even then jjX Therefore we can make the following inference, making use of the n-goodness of x t. Hence jX contrary to assumption. This inference shows that jY j - g j (n). So Duplicator can play the set Y . Next Spoiler puts the pebble on some element y of Y . It is now easy for Duplicator to put his pebble on an element x of X so that the following condition are satisfied: only if To end the proof we have to exhibit a sentence of L !! (Res k (Q f )) which distinguishes A and B. Consider the sentence of L !! (Res k (Q f )). This sentence is true in A but false in B because a We made the assumption that (1) holds. Suppose now (2) holds. We proceed as above until the definition of n and -. This time we use (2) to find an n so that we can continue as above. When we come to the sentence ', we replace it with :Res k (Q f )x The theorem is proved. 2 A special case of the above theorem - that Res 2 (Q f ) is not definable in L !! (Q f ) for proved in [19]. 5.8 Theorem. The following conditions are equivalent for any monotone simple monadic quantifier Q and any k - 2: Res k (Q) is definable in L !! (M). (2) Q is bounded on fn k !g. Proof. Theorem 5.7 gives (1) ) (2). Assume then (2). So there is a number m such that for all n, either or sentence in L !! (Q g ) (L !! (Q h )) saying that i), and let \Phi i and \Psi i ) be as in the proof of Proposition 5.5. Then the sentence Res k (Q f )x equivalent to the L !! (Q )-sentence i!m i!m The field of definability of generalized quantifiers on finite structures is full of open problems. Especially the polyadic quantifiers as well as the non-monotone ones give rise to problems that no one knows how to approach. In this paper we have considered certain polyadic lifts of monotone quantifiers, obtaining the following results I. definable in L !! (Q f 1 is definable in L !! II. Ram k (Q f ) is definable in L !! (Q f ) iff f is bounded. III. Res k (Q f ) is definable in L !! (M) iff f is bounded on fn k !g. I and II were also proved for relativizations of monotone simple unary quantifiers. The following questions seem to be the most natural direction to continue the work of this paper: 1. Find necessary and sufficient conditions for Br(Q f 1 to be definable in 2. Find necessary and sufficient conditions for Res k (Q) to be definable in L !! (Q). This question makes sense for non-monadic and non-monotone quantifiers, too. 3. Find necessary and sufficient conditions for Res 2 (Q f ) to be definable in L !! (Q 1 ). (As we mentioned in the Introduction, K. Luosto [15] has proved that Res 2 (Q f ) is not definable in L !! 4. Find a suitable extension of III to relativizations of monotone simple unary quantifiers. --R On branching quantifiers in English Maximality and Interpolation in Abstract Logic Journal of Logic and Computation Definability hierarchies of generalized quantifiers Logical hierarchies in PTIME Partially ordered connectives and finite graphs Kolaitis and J. First order predicate logic with generalized quantifiers On extensions of elementary logic The hierarchy of monadic quantifiers Computable quantifiers and logics over finite struc- tures The logic of divisibility Quantifiers in formal and natural languages Relativization of quantifiers in finite models Quantifiers in natural language - a survey of some recent work --TR --CTR Jouko Vnnen, Unary Quantifiers on Finite Models, Journal of Logic, Language and Information, v.6 n.3, p.275-304, July 1997
resumption;ramsey quantifier;branching;finite model theory;generalized quantifier
595891
Reasoning about Information Change.
In this paper we introduce Dynamic Epistemic Logic, which is a logic for reasoning about information change in a multi-agent system. The information structures we use are based on non-well-founded sets, and can be conceived as bisimulation classes of Kripke models. On these structures, we define a notion of information change that is inspired by Update Semantics (Veltman, 1996). We give a sound and complete axiomatization of the resulting logic, and we discuss applications to the puzzle of the dirty children, and to knowledge programs.
Introduction Recently the notions of information and information change have gained a prominent place in several fields of scientific research, such as philosophy, the formal semantics of natural language and computer science. The present paper highlights two of such developments, namely the update semantics of Veltman (1996), and the analysis of communication in distributed systems, in particular the approach of Fagin et al. (1995). Our main goal here is to show that tools of the former may provide useful supplements to the approach of the latter. Update Semantics In his influential paper 'Defaults in Update Semantics,' Frank Veltman presents a dynamic view on meaning. Following the slogan "You know the meaning of a sentence if you know the change it brings about in the information state of anyone who accepts the news conveyed by it," the meaning of a sentence is associated with a function on information states. We will discuss Veltman's update semantics as a simple example to illustrate the main notions involved. The language of update semantics is a standard propositional modal language. Definition 2.1 (Language) Given a set of propositional variables P , the language LP of update semantics is the smallest set containing P such that if OE and / are in LP , then OE - /, :OE and 3OE are in LP . 2 As said, we will interpret sentences as functions on information states. The notion of information state used in update semantics is a very simple one. 1 Definition 2.2 (Information states) ffl A possible world w assigns to each propositional variable a truth-value; it is a function 1g. ffl An information state oe is a set of possible worlds. 2 Veltman introduces more complex notions of information state to model reasoning with default rules. Intuitively, if an agent is in information state oe, then oe contains all worlds that are compatible with the agent's information: for all the agent knows, each possible world in oe may picture reality correctly. For example, the information state consisting of the set of all possible worlds represents an information state of an agent that has no information about the world at all, and the empty set is an information state in which an agent has contradictory information. When oe is an information state, and OE a sentence, we will write oe[OE] for the result of updating oe with the sentence OE. Intuitively, oe[OE] is the state that results when the agent, being in state oe, gets the information expressed by OE. Definition 2.3 (Interpretation) ae oe if oe[OE] update of an information state oe with a sentence p results in a state containing all and only the worlds in oe in which p is true. The result of updating a state oe with a negated sentence is a state containing all worlds in oe that do not survive in oe updated with OE. Conjunction is defined as intersection. A sentence of the form 3OE gets an interpretation roughly corresponding to the intuitive meaning of 'It might be the case that OE.' An update of oe with 3OE either returns the same information state (when oe updated with OE does not result in the empty state, i.e. when OE is compatible with the information contained in oe), or it returns the inconsistent state (when OE is not compatible with the information contained in oe). This reflects the assumption that an agent in a state oe already knows what she considers possible and what not, which means that a sentence of the form 'It might be that.' can never provide new information; at most, it can be inconsistent with the information contained in oe. Definition 2.4 (Acceptance and validity) ffl A sentence OE is accepted in an information state oe iff oe k\Gamma\Gamma OE. ffl A sequence of sentences acceptable iff there is a oe such that ffl An argument OE updating any information state with the premises in the order they are given, result in a state in which the conclusion is accepted: A sentence is accepted in an information state if an update with OE does not change the information state. Intuitively, this happens only when the information that OE is already contained in oe. A sequence of sentences is acceptable when there is a state in which updating with the sentences in the order they are given does not result in the inconsistent empty state. An argument is valid iff updating an information state with the premises in the order they are given, results in a state in which the conclusion is accepted. One of the interesting features of update semantics is that the order in which sentences in a text occur is important. In general, it is not the case that correctly reflects the fact that changing the order of the sentences of a text will in general produce a different story, which need not even be coherent. Consider for example the following two examples: Someone is knocking at the door.It might be John.It is Mary. Someone is knocking at the door.It is Mary.It might be John. The first sequence of sentences is acceptable, while the second is not: once one knows it is Mary who is knocking at the door, it cannot be John anymore. This is reflected in update semantics: 3p; :p is acceptable, while :p; 3p is not. Another feature of update semantics is that updates always imply an increase of information, in the sense that an update of an information state oe always results in a state that is a subset of oe. i.e. oe[OE] ' oe. But this does not mean all sentences that are accepted in the first state are also accepted in the updated state. A typical case is a state oe containing two worlds w and v, with and This is a state in which the agent does not know whether p is true or not. In such a state, 3p is accepted. But as soon as the agent learns that p is not the case, p is not considered possible anymore: in the state oe[:p], 3p is not accepted. 3 Dynamic Epistemic Semantics Update semantics is a semantics that models the information change of a single agent. In this section, we develop a semantics for a language in which it is possible to express facts about the information and information change of several agents. 2 The language is the following: Definition 3.1 (Language of DES) Let A be a non-empty set of agents, and let P be a set of propositional atoms. The language L A P of DES is the smallest set such that P ' L, and if OE and / are in L A P , and a 2 A, then :OE; OE - /; 2 a OE and [OE] a / are in L A 2 For an approach that uses a similar language, and employs a notion of constructive update over partial Kripke models, see Jaspars (1994). Other logical constants, such as -, ! and 3 a , are defined in the standard way. We will refer to the part of the language that does not contain the [\Delta] a -operator as the 'classical fragment of the language'. This is just the language of classical multi-modal logic. The intended interpretation of 2 a OE is that agent a has the information that OE. The intended meaning of [OE] a / is that an update of a's information with OE results in a situation where / is true. There is an operator [OE] a for each agent a and each sentence OE in the language, which reflects the idea that any statement about the system of agents and their beliefs is something which the the agents may learn. This makes the agents effectively as 'intelligent' as us theoreticians, i.e. in principle any property of the system we are able to formulate in the object language may be known or learned by the agents. To give a semantics for this language, we first need to make a choice as to how to represent the information of the agents. We believe that a representation that is based on non-well-founded sets is the most elegant way to do this. 3 Definition 3.2 (Possibilities) Let A, a set of agents, and P , a set of propositional variables, be given. ffl A possibility w is a function that assigns to each propositional variable 1g, and to each agent a 2 A an information state w(a). ffl An information state oe is a set of possibilities. 2 Clearly, this definition is circular, since possibilities are defined in terms of information states, and information states are sets of possibilities. In the universe of non-well-founded sets of Aczel (1988), this circularity is harmless. 4 A possibility w characterizes which propositions are true and which are false by assigning to each atomic sentence a truth value, and it characterizes the information of each of the agents by assigning to each agent an information state. The information of an agent is represented, as it is in update semantics, as a set of possible ways the world might be, according to that agent. In this case, this is a set of possibilities. There is a close relation between these non-well-founded models and Kripke- structures; to be precise, there is a one-one-relation between possibilities and bisimulation classes of worlds in Kripke models, that preserves truth of the classical fragment of the language. Definition 3.3 Let (R a A decoration of M is a function that assigns to each w 2 W a function d(w) on such that 3 See Groeneveld (1995) for a discussion of the semantics of DES using Kripke models and the knowledge structures of Fagin and Halpern (Fagin et al. 1991). 4 The underlying set-theory is axiomatized by ZFC \Gamma (the Zermelo-Fraenkel axioms minus the axiom of foundation) plus Aczel's Anti-Foundation Axiom (AFA). assigns to each propositional variable the same truth-value as it has in w in the model. assigns to each agent a the set of functions associated with worlds reachable from w by R a . If d is a decoration of M , and w a world in M , we say that d(w) is the solution of w in M , and (M;w) is a picture of d(w). 2 Proposition 3.4 ffl Each Kripke model M has a unique decoration. This decoration assigns to each world in M a possibility. ffl Each possibility has a picture. ffl w in M and w 0 in M 0 have the same solution iff w and w 0 are bisimilar. 2 This means that possibilities can be seen as representing bisimulation classes of worlds in Kripke models. Moreover, it implies the Bisimulation Principle (3.5 below) which we will frequently use later. In the following definition, we use the notation w[B]v, for B a set of agent, to stand for the fact that w and v differ at most from each other in the information states they assign to agent in B. Proposition 3.5 (Bisimulation Principle) A bisimulation between possibilities is any relation B such that wBv iff w[A]v and for each a 2 A, if w there is a v 0 2 v(a) such that w 0 Bv 0 , and if v 0 2 v(a), then there is a w such that w 0 Bv 0 . If B is a bisimulation, then for all possibilities w; v: if wBv then Properties of possibilities One of the charms of Kripke semantics is the fact that properties of information such as positive introspection or consistency correspond to certain simple properties on frames, such as transitivity and seriality of the accessibility rela- tions. Here are some examples of constraints on possibilities that correspond to familiar frame constraints. Definition 3.6 Call a class of possibilities S closed iff it holds that if w 2 S 1. C, the class of consistent possibilities is the largest closed class such that 2. T , the class of truthful possibilities is the largest closed class such that 3. P , the class of positive introspective possibilities is the largest closed class such that w 2 P and v 2 w(a) imply v(a) ' w(a) 4. N , the class of negative introspective possibilities is the largest closed class such that w 2 N and v 2 w(a) imply w(a) ' v(a) 2 Of special interest is the class of fully introspective possibilities P " N (which is a closed class). Conscious updates In update semantics, if an agent updates her information with p she will discard all possible worlds in which p is false. But if her epistemic alternatives are not classical possible worlds, but possibilities as we have defined them, there will be no point in also preserving those options in which p is true but in which the agent does not have the information that p. Ideally, she will also accommodate for the fact that she has learned p, and after having learned that p, she will not only have the information that p, but on top of that have the information that she has the information that p. And so on. We will refer to such an update as a 'conscious' update: the agent who gets new information is conscious of the fact that she gets this new information. Note that this is not an 'eliminative' process. A conscious update is not one in which one simply discards possibilities to reach a state in which one has more information. For example, consider a situation in which an agent a does not know whether p, and knows that she does not know this. This will be modeled by a possibility w such that the set of possibilities w(a) will only contain possibilities in which a does not know whether p. Removing all non-p- possibilities from this information state leaves the agent with a set of possibilities in which p is true, but in which she does not know whether p. We use the idea of conscious update for for interpreting sentences of the form [OE] a /, which will be interpreted as 'after a consciously updates with OE, / is true.' To do this, we first have to give a formal definition of conscious update. We will define for each sentence OE and each a 2 A a function [[OE]] a on possibilities in such a way that applying this function to a possibility returns a new possibility that is the result of updating a's information state consciously with the information that OE. Remember that we use the notation w[a]v as an abbreviation for the statement that w and v differ at most in the information state they assign to a. Definition 3.7 (Conscious updates) a is that w 0 such that w[a]w 0 and w OEg. So a conscious update of a's information in a possibility w changes the possibility w in such a way that only a's information state is changed (i.e. the new possibility differs from the old one only in the information state that is assigned to a), and this is done in the following way: all possibilities in which OE is not true are eliminated from a's information state, and in all remaining possibilities, a's information state is consciously updated with OE. That this notion of update is well-defined needs some proof. We will rely on the Solution Lemma of Aczel (1988), which is standardly used in the set theory ZFC/AFA for establishing the existence of non-well-founded sets, to prove that that there is in fact a relation that conforms to definition 3.7. Then, we will give an argument using the bisimulation principle to show that this relation is the only relation conforming to the definition. To prove the existence of the update function, fix an actor a 2 A and some proposition p (i.e. p is a class of possibilities). For each possibility w, introduce an indeterminate xw , and consider the class of equations defined by the stipulations ag [ f(a; fx By the Solution Lemma this system has a unique solution, which in this case is a map - from indeterminates to possibilities. Then define cu(a; p), the conscious update with p for a by It is not hard to check that it holds that (w; v) 2 cu(a; p) iff w[a]v and which shows that cu(a; p) is the function we were looking for. To show that the update function is unique for each OE and a, assume that there are in fact two functions f and f 0 that conform to definition 3.7. Note that by the definition, both of these functions will be total on the class of all possibilities. We will show that it holds for each possibility w, that (w). For define a relation B as: We claim that B is a bisimulation, from which it follows by the bisimulation principle that so we need to show that for each w there is a v 0 2 v(a) such that w 0 Bv 0 , and vice versa. Take any w 0 2 w(a). Then, there must be a u 0 2 u(a) such that u 0 f 0 is a total function, there must be a v 0 2 v(a) such that f 0 this v 0 it holds that w 0 Bv 0 . The other direction is completely symmetric, which shows that B is a bisimulation. Definition 3.8 (Truth) Technically we have to conceive of definitions 3.7 and 3.8 as one simultaneous definition, since the definition of update uses the notion of truth and vice versa. This offers no problems, however, and we have only separated the two for clarity. All classical logical operators are interpreted classically: a conjunction is true just in case both conjuncts are, a negation is true iff the negated sentence is not true, 2 a OE is true just in case OE is true in each possibility in a's information state. New is the definition for [OE] a /: such a sentence is true in a possibility w exactly when / is true in the possibility that results from updating a's information state in w with OE. We define validity in the standard way, i.e. for each possibility w, and for each set of sentences \Gamma, \Gamma for each possibility w such that w Update Semantics It turns out that validity in update semantics can be expressed in DES by identifying a US-update with a conscious update in DES in a fully introspective possibility. The validity of an argument-updating with the premises results in an information state in which the conclusion is accepted-can then be expressed in DES by a sentence expressing that after an agent consciously updates with the premises, then she will accept the conclusion, i.e. she will have the information that the conclusion holds. Proposition 3.9 Let for each OE in the language of update semantics, OE 0 be just like OE, except that each occurrence of 3 is replaced by 3 a . Then: fully introspective w, w Group Updates Common knowledge is a concept that crops up in several places in the literature on distributed systems in computer science, in the literature on game theory in philosophy and economics and in the literature on pragmatics in linguistics. The concept is most easily explained as follows: a sentence OE is common knowledge between a group of agents B just in case each agent in B knows OE, each agent knows of each other agent that he knows OE, and so on. What we will do here is model the effect of a sentence becoming common knowledge between a certain group of agents, and add operators [OE] B for each A to express this in the object language. Such an operator may be useful, for example, to formalize an idea in the theory of discourse that the purpose of an assertion is to extend the common knowledge of speaker and hearer. Definition 3.10 (Conscious group update) B is that w 0 such that w[B]w 0 and w for each a 2 B. A group update with OE in a possibility w results in a new possibility in which only the information of the agents in the group has changed. For each agent in the group, the new information state consists of all the old possibilities in which OE is true updated with the information that OE becomes common knowledge between the agents in the group. This definition is structurally similar to the definition given for conscious updates above. In fact, it holds that for each w, w[[OE]] a = w[[OE]] fag . Also, the proof that definition 3.10 is in fact correct is entirely analogous to the coinduction argument we gave for definition 3.7, for which reason we won't repeat the argument. We can now extend the truth definition of DES with the following clause: Axiomatization In this section, we provide an axiomatization of the language of DES with group updates, and prove that it is sound and complete with respect to the semantics. Definition 3.11 (Conscious K) The system CK is defined by the following axioms and rules. Axioms is valid in classical propositional logic A3 is an atom. (independence) (Generalized Ramsey Axiom) a OE if a 62 B. (Privacy Axiom) Rules Nec2 If ' OE then ' 2 a OE Nec[\Delta] If ' OE then ' [/] B OE there is a derivation of OE from assumptions in \Gamma. 2 So, in addition to the rules and axioms of classical model logic, the deduction system consists axioms describing the interaction between the dynamic operators and the classical logical constants. Axiom 3 together with the rule Nec[\Delta] guarantee that the dynamic operators behave as normal modal operators. Axiom 4 expresses that updates are functional: if it is not the case that a certain sentence is true after an update with a certain sentence, then, since the update always gives a unique result, it must be the case that the negation of that sentence is true in the updated possibility. Axiom 5 expresses that the update of an information state has no effect on the 'real' world; the same propositional atoms will be true or false before and after an update. Axiom 6 expresses that if it is the case that after a group update with OE, some agent in the group knows that /, then that agent already knew that if OE were true, then after a group update with OE, / would be true, and vice versa. Axiom 7, finally, expresses that a group update has no effect on the information of agents outside of that group. Proposition 3.12 (Soundness) If \Gamma 'CK OE then \Gamma proof: A standard induction; by way of illustration, we show the correctness of axiom 6, and leave the remaining cases to the reader. We have the following equivalences, if a 2 B: Proposition 3.13 proof: We use a variation on the classical Henkin proof for completeness of modal logic, showing that for each consistent set of sentences there is a possibility in which these sentences are true. It is easy to show that each consistent set can be extended to a maximal consistent set (we will refer to this as 'Lindenbaum's Lemma'). Let, for each maximal consistent set \Sigma, w \Sigma be that possibility such that w \Sigma and for each agent b: w \Sigma consistent and if 2 b / 2 \Sigma, then \Gammag. 5 We prove the usual Truth Lemma, that is, for each sentence - it holds that - 2 \Sigma iff w \Sigma Completeness then follows by standard argumentation. The Truth Lemma is proven by an induction on the structure of -, in which all cases are standard, except the case where - is of the form 5 In the terminology of definition 3.3, this model is the solution of the standard canonical model for the minimal modal logic K. Alternatively, the existence of the possibilities w \Gamma is easily proven by defining the appropriate set of equations, and an appeal to the Solution Lemma. /. The proof for this case rests on the following idea. Just as membership in w \Sigma (a) depends on the formulae of the form 2 a OE in \Sigma, the B-update of w \Sigma with OE depends on the formulae of the form [OE] B / in \Sigma. This is reflected by the following operation on maximal consistent sets: \Gammag The functionality axiom A4 will ensure that will be a maximal consistent set whenever \Gamma is. The step in the Truth Lemma for formulae of the form [OE] B / then proceeds as follows: (by 3.14 and induction hypothesis on OE) (by induction hypothesis on /) (by definition of \Sigma ffl B OE) Which only leaves the next lemma. 2 Lemma 3.14 Let OE be fixed, and assume that for each maximal consistent \Sigma, proof: Define a relation R on possibilities by there exists a maximal consistent set \Sigma such that We will show that R is a bisimulation. The Bisimulation Principle 3.5 then implies that R actually is an identity relation, which proves the lemma. Let wRv, and suppose (the case that easy). We need to show three things: 1. w[A]v. (by the semantics) iff p 2 \Sigma (by the definition of 2. Next we must show that for each b 2 A, if w 0 2 w(b) then there is a We distinguish two cases: b 2 B, and b 62 B. Assume that b 62 B. It then follows by axiom 7 that which implies that w \Sigma (b). But by the definition of and the fact that b 62 B, we have w \Sigma actually v(b) in this case, which is certainly sufficient. For the other case, let b 2 B and take any w (b). Then there must be a \Sigma 0 that is maximal consistent such that w \Sigma 0 and w \Sigma (b). The latter implies that if 2 b - 2 \Sigma, then - 2 \Sigma 0 . We want to find a v 0 such that (I) v the set \Sigma . It is then immediate that (II) holds. To see that (I) holds, take any 2 b / 2 and hence, by axiom Because w \Sigma 0 by the main assumption of this lemma that which means that / 2 \Sigma was arbitrary, it follows that w \Sigma (b), which is what we wanted to prove. 3. Finally we must show that for each b 2 A, if v 0 2 v(b) then there is a If b 62 B, we can use the same argument as in case (2). So assume instead that b 2 B, and take any w (b). Consider the set \Sigmag. Below we show that \Delta is consistent (III). Then \Delta can be extended by Lindenbaum's Lemma to a maximal consistent set \Sigma 0 , for which it holds, by definition of \Delta that w \Sigma This implies that w \Sigma 0 it is not hard to see that by definition of \Delta and the functionality axiom it holds that It then follows that that To see (III), suppose to the contrary that \Delta is inconsistent. Then there must be That means that: (using A3 and A4) which means that by the fact that contradicting the fact that \Gamma is consistent.Some observations and remarks Here follows a short list of validities and non-validities. Proposition 3.15 1. (there is no difference between updating with updating with / - OE) 2. (updating first with OE and then with / is different from updating with / first, and after that with OE.) 3. (updates of the information of one group does not affect updates of the information of a disjoint group) 4. If ' OE (if two sentences have the same truth conditions, they are also equivalent as updates) It is possible to combine the notion of conscious update with stronger epistemic logics. This is unproblematic for the logic of introspection: if we add the introspection axioms of K45 to the axioms of CK, the resulting logic CK45 is sound and complete with respect to the class of introspective possibilities. For other well known epistemic logics such as KD45 or S5, updates will be partial functions, since conscious updates don't necessarily preserve the properties of consistency and correctness. We sketch some details for one special case, S5. We take K to be the class of truthful and introspective possibilities (positive and negative; see definition 3.6). The K-update a is then defined as the restriction of [[OE]] a to possibilities in K. This will make the updates partial functions. For example, for an atom p this will have the effect that an agent a can only learn p in a possibility w if w As for the effect on the axiomatics, we conjecture that the following is complete: add the axioms of S5 to CK; weaken the functionality axiom to :[OE] a [OE] a :/; and compensate the loss of the existential part of the functionality axiom by adding the axiom hOEi a ? We end this brief discussion by a comment on our choice of operators. We could have added operators of the form CB to the language, one for each (non- empty) set of agents B, to express the static concept of OE being common knowledge between the agents in B. Transferring the definition of Fagin et al. (1995) to our framework, its definition could be the following: an OE for each an The reason we have not added these operators is that we have not yet found an axiomatization for the language with these operators. We hope to correct this omission in the near future. 4 An application: Automated Dirty Children We have developed a language and a semantics for reasoning about information and information change of several agents, which in turn can reason about their own information and the information of other agents. We have provided an axiomatization for this semantics. In this section, we want to show that the logic developed above can be useful as a tool for analyzing problems concerning reasoning about information in a multi-agent setting. To show this, we consider a textbook case, the puzzle of the dirty children. This puzzle occurs under different guises in the literature: it is a variant of the puzzle of the cheating husbands (see for example Moses et al., 1986), the wise men puzzle (in e.g. McCarthy, 1990) and the the Conway-paradox (e.g. van Emde Boaz et al., 1980). The puzzle The description of the dirty children puzzle that we give here is adapted from Barwise (1981). There are n children playing together. During their play some of the children, say k of them, get mud on their foreheads. Each can see the mud on others but not on his own forehead. Along comes a father, who says, "At least one of you has mud on your head." He then asks the following question, over and over: "Can any of you prove that you have mud on your head?" Assuming that all the children are perceptive, intelligent, truthful, and that they answer simultaneously, what will happen? There is a "proof" that the first k \Gamma 1 times the father asks the question, the children will all say "no" but that the k-th time the children that are dirty will answer "yes." The proof is by induction on the number of dirty children k. For result is obvious: the dirty child sees that no one else is muddy, so he must be the muddy one. If there are two dirty children, say a and b, each answers "no" the first time, because of the mud on the other. But, when b says "no," a realizes he must be muddy, for otherwise b would have known the mud was on his head and answered "yes" the first time. Thus a answers "yes" the second time. b goes through the same reasoning. Now suppose there are three dirty children, a, b, c. Child a argues as follows. Assume I don't have mud on my head. Then, by the both b and c will answer "yes" the second time. When they don't, he realizes that the assumption was false, that he is muddy, and so will answer "yes" on the third question. Similarly for b and c. Formalization We will show how one can formalize the description of the puzzle, and the reasoning involved, in dynamic epistemic semantics. The result, that after answers to the father's question, the children that are dirty know that they are dirty, will then be a theorem in the logic. Let A be a set of children playing in the mud. Consider a language that contains a propositional atom p a for each a 2 A, which we will take to express that child a is dirty. We start by introducing some convenient abbreviations: ffl Each child can see the forehead of each of the other children. So, if a child is dirty, each of the other children knows that she is dirty. This can be expressed by the conjunction of all sentences of the form (p a ! for each a and b in A such that a 6= b. We abbreviate this conjunction by vision . ffl It is common knowledge between all children that each forehead can be seen by each of the other children, i.e. vision is common knowledge between all children. We can express this as CA vision. ffl Before asking the children whether they know if they are dirty or not, the father announces in face of all children that at least one of them has a dirty forehead. Let father be the sentence fp a j a 2 Ag, which expresses that at least one of the children is dirty. ffl After the father's announcement, all children answer the question 'Do you know whether you are dirty or not?' The children answer either `yes' or 'no'. Let no be the sentence f(:2 a p a -:2 a :p a ) j a 2 Ag, which is the sentence that expresses that none of the children knows that she is dirty (i.e. the information expressed by all children answering 'no' at the same time.) ffl Finally, we let for each B ' A, dirty(B) abbreviate V This sentence expresses that all and only the children in B have dirty foreheads. We can now express in DES that if exactly m children are dirty and it is commonly known between all children that they can see each other, then it holds that after a common update with the father's statement that at least one of the children is dirty, and commonly updating times with the fact that all children answer 'no', the resulting state is a situation in which all dirty children know that they are dirty. Formally expressed, this boils down to the following statement: Proposition 4.1 Let B be a set containing exactly m children (m - 1), and stand for a sequence of updates with OE. Then it holds for all a 2 B: dirty(B); vision; CA vision A 2 a p a proof: We will provide a syntactical proof of this statement. Although we do not have an axiomatization for the language containing the common knowledge operators CB , it is easy to see that that following axiom is sound, if b 2 B: We will make use of this axiom in the proof. The proof of the proposition is by induction on the number of dirty children m. Assume first that only one child, say a, is dirty. In classical modal logic, it holds that if a is the only dirty child, and a can see all other children, then a knows that if at least one child is dirty, it must be herself: We use axiom 5 to conclude that (omitting the subscript A from the update operators for legibility): whence, by axiom 6 dirty(fag); vision; CA vision ' [father]2 a p a For the induction step, let B be a set of m+1 children, and a; b 2 B. It holds by induction hypothesis and the necessitation rule that if B 0 has m elements, then Since CAvision ' 2 a (vision - CA vision) and dirty(B); vision ' 2 a (:p a ! dirty(B=fag)), it follows that from which it follows, using axioms 3 and 5, and the fact that ' no ! :2 b p b , that By axiom 6, then and using the lemma below, finally, Lemma 4.2 For each m: 2 a [no] m (no ! p a ) ' [no] m+1 2 a p a This is proven by induction on m. If a ) is equivalent by axiom 5 to 2 a (no ! [no]p a ), which is, by axiom 6, equivalent to [no]2 a p a . For the induction step, assume that: This implies that From which it follows by axiom 6 that [no]2 a [no] m (no ! p a ) whence, by induction hypothesis, that This completes the proof of proposition 4.1. 6 2 Discussion The puzzle of the dirty children and related puzzles have been discussed relatively extensively in the literature, and several formalizations have been given. Our analysis adds to earlier approaches in an essential way, we believe. First of all, we have rephrased the informal description of the puzzle in the object language of an independently motivated logic. Something like that has not been done before: all earlier formalizations of the puzzle that we know of consist of a more or less ad hoc model of the information and information change involved in the puzzle. That means that each variant of the puzzle that differs from the present one calls for a new analysis and the construction of a new model. The relatively straightforward way in which the puzzle can be formalized in DES suggests that similar problems may be formulated in the same way. Secondly, the fact that our formalization of the puzzle gives results similar to Barwise's semi-formal results, shows that the paradoxical flavor of the puzzle does not stem from a logical mistake. This suggests strongly that the discrepancy between the ideal situation described in the puzzle and a 'real life' situation should not be explained as a difference in principles of logic, but as a result of the complexity of the reasoning involved in the puzzle and the way it depends on the strong trust they should have in the each other's reasoning capabilities. Thirdly, the formalization given above makes the role that the father's announcement and the children's answers play quite explicit. For example, one of the 'paradoxical' aspects of the puzzle is that the father's statement seems superfluous at first sight if there two or more dirty children present. In such a situation, each of the children already knows that one of the children is dirty (since everyone can see a dirty child). The formal correlate of this fact is a theo- rem: father. The point of the father's statement lies in the assumption that his announcement makes it common knowledge that at least on child is dirty, which was not the case: p a -p b ; vision; CA vision 6' CAfather. This observation is not new, but our analysis adds to earlier ones in that it is now possible to formulate such facts in the object language. Another puzzling aspect of the puzzle that is highlighted in our analysis is the fact that the children keep on saying 'no' until `suddenly' some children 6 We have not proven that the children do not know that they are dirty before they have answered the question times. To show that, one needs an extra assumption that in the initial possibility, none of the children knows whether she is dirty of not, and that this fact is common knowledge. A proof can then be given along the lines of the proof given here. answer yes, suggests that each answer supplies new information, although, 'syn- tactically', the children say the same thing each time. This is directly reflected in our semantics: an update with no changes the possibility in a certain fixed way, resulting in a new possibility in which another update with no may change the possibility again. This is an example of the failure in DES of the following principle of Success: which states that after a group update with a sentence OE, each member in the group knows that OE. This is not a property of updates in general, and the example of no suggests that this is right. 7 5 Epistemic propositional dynamic logic In this section, we use the ideas of dynamic epistemic logic for developing a logic, which we call EPDL, for 'epistemic propositional dynamic logic'. Besides update actions this logic will also have send actions and test actions. The main thrust of this move is very much in the line of our discussion in the previous section: by incorporating more notions, such as send actions, that are crucial for understanding communication, into the logical object language, it becomes possible to formalize processes that otherwise would remain part of the meta-language The language of EPDL may be used to specify or describe the behavior of a group of communicating agents, in a very general sense. The phenomena modeled might be human agents speaking to each other, the behavior of processors in a distributed network, or the behavior of a knowledge base and a human agent querying it. The idea that extensions or variations of epistemic logic may be used to describe such kind of applications is not new. In computer science, the work of Fagin et al. (1995) and Shoham (1993) on agent oriented programming are prime examples. Another example is the work of McCarty (1990). See also de Rijke (1993). The basic idea here is to treat the update modalities of the previous sections as programs, and extend the language with certain program operators familiar from PDL (cf. Pratt (1976), Goldblatt (1987)). We consider a language in which there are three kinds of basic programs. Firstly, there are update programs of the form U(B; OE), that have the effect that OE becomes common knowledge in the group of agents B. Secondly, a program S(a; B; OE) will stand for the action of agent a sending the message OE to all agents in B. Thirdly, the local tests ?(a; OE) stand for the action of agent a testing whether she knows that OE. In addition 7 For this particular example, we even have the surprising fact that there are possibilities w such that w to these basic programs, we add the program operators of PDL: composition, union and iteration. Definition 5.1 (Epistemic Propositional Dynamic Logic) Given an atomic vocabulary P and a set of actors A, we define a set of assertions \Phi and a set of programs \Pi by simultaneous induction: where a We will interpret programs as relations on possibilities. Since we have included union in the language, which is interpreted as choice in our semantics, programs will be non-deterministic in general: running a particular program may lead to several different outcomes. That means that, in contrast with the updates described in the previous sections, not all programs will have functional interpretations. Definition 5.2 (Semantics of EPDL) An interpretation for EPDL is a function I that assigns to each triple consisting of an actor a, a set of actors B and a formula OE a binary relation between possibilities. Relative to such a interpretation we define the truth conditions of assertions and the interpretation of programs inductively by: truth conditions program interpretations w[[- The interpretation function I plays a role only in the clause for the send actions OE). The basic idea is that I(a; B; OE) describes the effects of the sending action. Of course, the real interest in sending a message is to exchange information. Thus the real burden is to relate send actions to epistemic effects. This can be achieved in EPDL by formulating extra constraints that relate send actions to update actions and to the information of the actors. Typically, these extra constraints will reflect certain properties of the communication channel, or certain pragmatic rules that the actors follow. For example, the axiom a sincerity condition for actor a (here h-iOE is the existential dual of actor a can only send the message OE if he has the information that OE. As an example of the application of EPDL to communicative situations we discuss a simple example which is called 'the bit-transmission problem' in Fagin et al. (1995, pp. 107ff. We consider two agents, a sender s and a receiver r . The sender has a certain piece of information (for example that the value of register x is either 0 or 1) which she wants to communicate to the receiver. We let the proposition p represent the information that x has value 1 (and :p that the value is 0). We assume the communication line may be faulty; for simplicity's sake we assume that messages either arrive immediately or are lost forever. Since s cannot be sure that her message has arrived, she will continue sending the message p until she has received an acknowledgment from the receiver that he has gotten her message. Another way of describing the behavior of s and r is as follows. s will send the message p to r until she knows that r knows that p. As soon as r knows that p, he will send a message 'I know that p' to s . Such descriptions that make use of concept such as 'knowledge' are descriptions that we can fairly straightforwardly translate into our language. For capturing the behavior of this system in EPDL we first of all need to express what the effect of sending a message is. Under the assumptions we have made, that messages either arrive immediately, or are irredeemably lost, the effect of sending a message OE to r can be described by the program 's information state is updated with OE, or nothing at all happens (the test ?(r ; ?) will always succeed). The corresponding interpretation function I interprets send actions as follows: where Id is the identity relation over possibilities. This interpretation of send actions corresponds to the syntactical characterization: The action of s sending r the value of the bit p is can be described by the program While the receiver can be described as performing the following program: (- r :p)). We can now formulate statements about such programs in the object language. For example, the statement "If the value of x is 1, then the receiver will eventually (that is, after repeating both programs again and again) know that the x is 1" may be represented by: 2 s p. In fact, this sentence turns out to be valid under the interpretation I given above. Properties of programs We may consider how EPDL programs behave with respect to the properties of situations introduced in definition 3.6. In particular, we may ask which properties are preserved under updating with certain programs. Definition 5.3 Let S be a class of possibilities, I an interpretation. A program - is persistent over a class S of possibilities under the interpretation I iff w 2 S and The following result claims that if I interprets all send actions as actions that preserve positive or negative introspection, then all programs will. Proposition 5.4 If each send action S(a; B; OE) is persistent over P (N , respectively) under I, then each program - is persistent over P (N , In general, programs are not persistent over the class of truthful situations, nor over the class of consistent situations. One example is the program - s above. In a situation where S knows that r does not know the value of the bit, the may result in a possibility in which r knows the value, but s still believes that r does not know it. Knowledge programs As observed by Fagin et al. (1995), one way of looking at a problem like the bit transmission problem is in terms of so called knowledge programs. That is, we can model the actors as executing a certain set of instructions of the form OE ) ff, which are read as "if OE then do ff", where OE is a formula of epistemic logic, and ff is some action. Formally, they define a knowledge program for an actor a as a set of these instructions g. These programs are interpreted indeterministically by requiring that a performs one of the actions ff i for which the test OE i succeeds. The models they consider consist of 'local states' of agents standing in a certain relation to each other. These 'local states' are meant to correspond in a relatively direct way to states that the agents may actually be in. Actions are interpreted as operations on such representations, while sentences of modal logic are interpreted in a possible worlds model that is derived from this representa- tion. In addition to this 'two-level' architecture, the model contains an explicit representation of time, and a simple logic of time is added to the language. It is clear that the in EPDL there are programs similar to these knowledge programs, i.e. programs that make an action conditional on the epistemic state of the actor. We will make a few remarks about how our framework compares with the approach adopted by Fagin et al. One of the most salient differences is that the former has an ontology that is much richer. Although this means that the model allows for distinctions that cannot be drawn in our model and that the behavior of a system can be described in much more detail than in our approach, it also implies, as the authors themselves note, that it is often unclear what part of the behavior of a system should be modeled by what part of the semantics. In the work of Fagin et al., information change on the level of Kripke structures is a notion that is derived from change in the underlying model. By contrast, we have given an explicit semantics for the notion of epistemic update on this level, thereby providing a semantics in which it is much more clear what is going on. Moreover, this means that we are not restricted to using only S5 models, which in the architecture of the system of Fagin et al. seems unavoid- able. This is interesting, because it makes it possible to describe situations in which agents are misinformed about either their environment or about the information of the other agents. 6 Conclusions In this paper, we have combined techniques from epistemic and dynamic logic to arrive at a logic for describing multi-agent information change. The key concept of dynamic semantics is that the meaning of an assertion is the way in which the assertion changes the information of the hearer. Thus a dynamic epistemic semantics consist in a explicit formal definition of the information change potential of a sentence. We used these ideas to arrive at the system of Dynamic Epistemic Semantics, which is semantics for a language describing information change in a multi-agent setting. This semantics proved useful for analyzing the Muddy Children paradox, and also for giving a semantics for knowledge programs, since it enabled us to model knowledge change by giving an explicit semantics to the triggers of the information change (the latter being the assertions made, or the messages sent). We feel that this is an important extension, since standard approaches to for example the Muddy Children (e.g. Fagin et al. generally use static epistemic logics like S5 to describe the situation before and after a certain epistemic event, leaving the transition between 'before' and 'after' to considerations in the meta-language. 8 In contrast, in dynamic epistemic logic, epistemic actions like updates are first class citizens of the object language of DES. For one thing, this opens the possibility of making artificial agents a bit more intelligent, by giving them an axiomatics for DEL as their tool for reasoning about knowledge change. Authors adress: Department of Philosophy University of Amsterdam Nieuwe Doelenstraat 15 The Netherlands gerbrand@illc.uva.nl groenev@illc.uva.nl --R The Journal of Philosophy A system of dynamic modal logic. Reasoning about Knowl- edge A modeltheoretic analysis of knowledge. Knowledge in Flux. Logics of Time and Computation. Logical Investigations into Dynamic Semantics. Calculi for Constructive Communication. On the difference between updating a knowledge base and revising it. Formalization of two puzzles involving knowedge. Formalizing Common Sense: Papers by John McCarthy. Cheating husbands and other stories: A case study of knowledge Semantical considerations on floyd-hoare logic Belief revision from the point of view of doxastic logic. Agent oriented programming. Peter van Emde Boas Defaults in update semantics. --TR --CTR Alexandru Baltag , Lawrence S. Moss , Slawomir Solecki, The logic of public announcements, common knowledge, and private suspicions, Proceedings of the 7th conference on Theoretical aspects of rationality and knowledge, July 22-24, 1998, Evanston, Illinois Fernando R. Velzquez-Quesada , Francisco Hernndez-Quiroz, Some semantics for a logical language for the game of dominoes, Proceedings of the 24th IASTED international conference on Artificial intelligence and applications, p.293-298, February 13-16, 2006, Innsbruck, Austria Hans P. Van Ditmarsch, Descriptions of Game Actions, Journal of Logic, Language and Information, v.11 n.3, p.349-365, Summer 2002 Jan-Willem Roorda , Wiebe van der Hoek , John-Jules Meyer, Iterated belief change in multi-agent systems, Proceedings of the first international joint conference on Autonomous agents and multiagent systems: part 2, July 15-19, 2002, Bologna, Italy Alexandru Baltag, Logics for insecure communication, Proceedings of the 8th conference on Theoretical aspects of rationality and knowledge, July 08-10, 2001, Siena, Italy John Cantwell, A Formal Model of Multi-Agent Belief-Interaction, Journal of Logic, Language and Information, v.14 n.4, p.397-422, October 2005 Chitta Baral , Yan Zhang, Knowledge updates: semantics and complexity issues, Artificial Intelligence, v.164 n.1-2, p.209-243, May 2005 H. P. van Ditmarsch , W. van der Hoek , B. P. Kooi, Concurrent dynamic epistemic logic for MAS, Proceedings of the second international joint conference on Autonomous agents and multiagent systems, July 14-18, 2003, Melbourne, Australia Robert Van Rooy, Quality and Quantity of Information Exchange, Journal of Logic, Language and Information, v.12 n.4, p.423-451, Fall Barteld P. Kooi, Probabilistic Dynamic Epistemic Logic, Journal of Logic, Language and Information, v.12 n.4, p.381-408, Fall Andreas Herzig, Modal Probability, Belief, and Actions, Fundamenta Informaticae, v.57 n.2-4, p.323-344, April Andreas Herzig, Modal probability, belief, and actions, Fundamenta Informaticae, v.57 n.2-4, p.323-344, October Churn-Jung Liau, A Logical Analysis of the Relationship between Commitment and Obligation, Journal of Logic, Language and Information, v.10 n.2, p.237-261, 2001 Paul Dekker, Meaning and Use of Indefinite Expressions, Journal of Logic, Language and Information, v.11 n.2, p.141-194, Spring 2002 J.-J. Ch. Meyer, Dynamic logic for reasoning about actions and agents, Logic-based artificial intelligence, Kluwer Academic Publishers, Norwell, MA, 2000 Churn-Jung Liau, Belief, information acquisition, and trust in multi-agent systems: a modal logic formulation, Artificial Intelligence, v.149 n.1, p.31-60, September Barbara Dunin-Keplicz , Rineke Verbrugge, Evolution of Collective Commitment during Teamwork, Fundamenta Informaticae, v.56 n.4, p.329-371, December
dynamic logic;epistemic logic;non-well-founded set theory;information exchange
595905
Contexts in Dynamic Predicate Logic.
In this paper we introduce a notion of context for Groenendijk & Stokhofs Dynamic Predicate Logic DPL. We use these contexts to give a characterization of the relations on assignments that can be generated by composition from tests and random resettings in the case that we are working over an infinite domain. These relations are precisely the ones expressible in DPL if we allow ourselves arbitrary tests as a starting point. We discuss some possible extensions of DPL and the way these extensions interact with our notion of context.
Introduction Dynamic Predicate Logic (DPL) was invented by Jeroen Groenendijk and Martin Stokhof (see [1], see also our section 2) as a specification language (or better: as a module for a specification language) of meanings for fragments of natural language. Most of the research concerning DPL has gone into integrating it with versions of Montague Grammar (see [5]) and into integrating it with Frank Veltmans Update Semantics (see [6]). DPL is a theory of testing and resetting of variables/registers. These are fundamental operations in computer science. Thus, apart from its use in Logical Semantics, DPL is a simple theory of these basic operations. 1 DPL is a natural variant of Predicate Logic. It mainly differs in the treatment of the scope of the existential quantifier. Certain basic truths about variables in Predicate Logic, however, fail in DPL (see section 2, see also [2] for similar observations on DPLE). The study of DPL and its kin makes the dependence of these truths on the specific choice of scoping mechanisms in standard Predicate Logic visible. In the light of the varied interest of DPL, it seems a good idea to make a closer study of its metamathematical properties. We focus on the closely related questions: ffl Which relations between assignments are definable in DPL (in a sense we will specify later)? ffl How does DPL treat its variables? To throw some light on both questions a good notion of context in DPL is indispensable. When studying classical DPL, which is based on total assignments, contexts appear as objets trouv'es. They are not part of the design of the language, but -as will become clear in the paper- can be viewed as the result of 'abstracting away' or `erasing' certain properties of the predicate logical language (both vocabulary and structure), thus yielding an underspecified language. Underspecification simply means here that the denotations or meanings of the contexts are properties of relations, rather than relations. 1 DPL is just one theory in a family of alternatives to predicate logic. In these alternatives 'resetting a register' is replaced by other related actions, like 'create a new register'. See section 7 for references. We can view what happens in the paper in a different light. Our deeper interest is in such grand questions as: ffl What should a general theory of information processing look like? ffl What is the nature of the variable? It seems to me that these questions are closely related and that ideas involving both Dynamics and Contexts should play a role in the answer. The study of DPL, here, is analogous to the study of a fruitfly in a laboratory situation. It allows us to focus on problems involving dynamics, variables, information processing. Still, these problems remain feasible. We end this introduction with a brief sketch of the paper. Section 2 is a straightforward introduction to predicate logic. It contains all the technical material the reader needs to know. For a discussion of the applications to discourse phenomena, however, the reader should consult [1]. Section 3 presents our theory of contexts in DPL. We study contexts as mathematical objects in their own right and establish their connections to language and semantics. Some materials concerning the information ordering on contexts are placed in an appendix. In section 4, we treat the Switching Propery. This property is characteristic for the DPL-definable relations. Section 5 contains the main result of the paper: a relation over an infinite domain is DPL-definable iff it 'has' a context and satisfies the Switching Property. In the next section, we touch on the subject of extending the DPL-language with new operations such as conjunction and disjunction. We consider the question whether such extensions support a good theory of contexts. We will produce two extensions that are complete for all relations that have a context. In other words: all such relations are definable in those extensions. Our last section 7 is devoted to the idea of making the context part of the semantics. 2 What is Dynamic Predicate Logic? We provide the basic definitions of DPL. Nothing in this section pretends to be original. We start by introducing some basic relational notions. Definition 2.1 Let X be any non-empty set. Rel(X) is the set of binary relations on X , i.e., X). Let R; S 2 Rel(X ). We define: 1. The composition R ffi S of R and S is defined by: x(R ffi S)y :, 9z xRzSy. Note that composition is in the order of application. 2. The dynamic implication (R!S) between R and S is defined by: x(R!S)y Our use of ! here overloads the symbol, since we also use it for implication in the objectlanguage. We write :(R) for (R ! ;) 3. id X is the identity relation. R is a condition or test if R ' id X . 4. Consider Y ' X . We define diag(Y g. 5. dom(R) := fx2X j 9y xRyg and cod(R) := fy2X j 9x xRyg. The notion of dynamic implication was first introduced by Hans Kamp in his pioneering paper [4]. Note dom(R). The relations in the range of diag are precisely the conditions. Writing (Y ! Z) := (XnY ) [Z, we have: Thus, diag is a homomorphic embedding of the structure h-X ; in the structure We will sometimes confuse, in the relational context, the set X with the relation diag(X). We need some further relational notions specifically concerned with relations between assignments. Definition 2.2 Let D be a non-empty domain and let Var be a set of variables. Let R2Rel(D Var ), We define: is the result of changing the values of f on the v i to d i . I VarnV . We write [v] for: [fvg]. ffl Y is an hV i-set if f 2 Y and fI V g ) finitely restricted if Y is a hIi-set for some finite I . A condition is finitely restricted iff it is the image of a finitely restricted set. If we want to make the dependence of I or [:] on Var or D visible, we add them as subscripts. We collect some simple facts concerning these notions. We have: Note that the classical meaning of the existential quantifier as a 'cylindrification' can be given as: We turn to the definition of DPL. Definition 2.3 A DPL-language L is a structure hPred ; Ar ; Var ; Coni, where Pred is a set of predicate symbols; Ar is a function from Pred to the natural numbers (including 0); Var is a, possibly empty, set of variables, Con is a, possibly empty, set of constants. Let Ref := Var [Con be the set of referents. We will use v; constants and referents. The set of L-formulas, For L , is the smallest set such that: are in For L for ffl If OE; / 2 For L , then so are OE:/ and (OE ! /) I feel that it is more faithful to the semantics to leave out the brackets in the formation rule for the dot officially, but nothing important hangs on this choice in this paper. We get an ambiguous syntax, but still unique meanings, since the operation of composition -the semantic counterpart of "."- is associative. An alternative notation for 9v, is [v :=?] (random reset). We use :(OE) and 8v(OE) as abbreviations of, respectively, (OE ! ?) and (9v ! OE). If x 2 Var and r 2 Ref and x and r are distinct, we write [x := r] for: Definition 2.4 A DPL-model M for a DPL-language L is a structure hD; Ii, where D is a non-empty set, the domain of M; I is a function which assigns to each predicate symbol P of Pred L an Ar(P)-ary relation on D and to each constant c an element of D. AssM , the set of assignments for M, is D Var . Consider r 2 Ref . We define: ae The interpretation function [:] M : For We write OE jM / for [OE] . We define validity in DPL by: As usual, OE appropriate for the given language. binary relation R is definable in a DPL-model M for a language L if there is an L-formula OE, which defines R, i.e., We will often suppress the subscript M, when the model is clear from the context. We could extend the DPL-language with function symbols by copying the way this is done in ordinary predicate logic. However, for the kind of result we are after such an extension is immaterial, since the usual trick to eliminate function symbols works also in DPL -with a small twist. E.g., P (f(g(x))) will be translated to: ::(9u:G(x; u):9v:F (u; v):P (v)). We remind the reader of Geach's Donkey Sentence: If a farmer owns a donkey, he beats it. This sentence can be translated into DPL in a compositional way as: One striking feature of DPL is that it is not 'structural': the values the predicate symbols may assume are not all the possible meaning objects provided by the semantics; we only allow tests. A second striking feature is the time symmetry of resetting and composition, which constrast strongly with our time asymetric intuition about, say, the meaning of P (x):9x:Q(x). The asymmetry of our intuition may be explained by the fact that we tend to think more in terms of successful resetting, just in terms of what the resetting relation is. Ordinary predicate logic can be interpreted in DPL as follows. We suppose that the predicate logical language has as connectives and quantifiers: ?, ?, -, !, 9x. We translate as follows: ffl (:) commutes with atomic formulas and with ! We find: [OE is the usual valuation function of Predicate Logic. Our translation is compositional. It shows that we may consider Predicate Logic as a subsystem of DPL. There is also a kind of inverse translation (:) ffi , which satisfies: [[OE involves renaming of variables and cannot be taken to show that DPL is a subsystem of Predicate Logic. 3 Contexts for Dynamic Predicate Logic In this section, we study the notion of context and its connections with relations and language. We placed some materials concerning the information ordering on contexts in an appendix, since, on the one hand, they are conceptually relevant and have a clear place in the total picture, but, on the other hand, they have no direct bearing upon the main results of the paper. 3.1 Introductory remarks To motivate our notion of contexts, we first give an intuitive discussion about substitution and kinds of variable occurrences in DPL. 2 In Predicate Logic variables may occur in a formula in two ways: freely and bound. The free variables admit (under certain conditions) substitution. The bound variables may be renamed salva significatione (ff-conversion). Let's write oe t x (OE) for: the result of substituting t for x in OE. In Predicate Logic we have, e.g., f I(c) What is the proper analogue of this fact for DPL? To simplify the discussion we will only treat a special case and refrain from giving official definitions. Consider the DPL-formula P (x):9x:Q(x):9x:R(x). We 1. f I(c) 2. f [P (x):9x:Q(x):9x:R(x)] M g I(c) Meditation upon (1) and (2) suggests, that, in DPL, we have to distinguish two kinds of substitution left substitution and right substitution and corresponding to these kinds two kinds of 'free occurrence': right free. We also speak of input occurrences and output occurrences. Following temporal intuitions -ignoring the essentially time-symmetric character of resetting and composition- we may also call the left free occurrences simply free and the right free occurrences actively bound. Now consider the following formula, say OE 0 , in which we have tagged occurrences of x with superscript numerals. We see that x 1 is a (left) free or input occurrence. Left substitution for x will cause it to be replaced. If we form T in the semantics the values assigned to x 0 and x 1 will be unified. If we form 9x:OE 0 , x 1 will be 'bound' or `initialized' by the new 'quantifier'. Symmetrically, x 7 is right free or actively bound. It will be in the scope of right substitution. If we form OE 0 :T the values of x 7 and x 0 will be unified, If we form OE 0 :9x, x 7 will be 'aborted'. Neither x 1 nor x 7 are open to ff-conversion salva significatione. x 3 is not accessible for substitution, nor is it ff-convertible: replacing x 2 , x 3 and x 4 by, say, y, will result in a formula that resets y, which OE 0 doesn't do. We call x 3 a garbage occurrence: it is something that 'exists', but is no longer `used'. 3 x 6 is also inaccessible for substitution, but in addition it can be ff-converted: replacing x 5 and x 6 by y does not change the meaning of OE 0 . We say that x 6 is clasically bound. Finally, we consider x 2 , x 4 and x 5 . These are 'occurrences' in a purely syntactical sense only: they do not represent 'files' carrying information, but just signal that incoming files labeled x should not be 'unified' with outgoing files labeled x. We say that these `occurrences' are blockers. x 5 is not a blocker in OE 0 as a whole. 4 2 The subject of kinds of variables and substitution would merit a far more extensive discussion. Regrettably, such a discussion is beyond the scope of the present paper. 3 The notion of garbage is studied in [8] and in [9] 4 Even if Vermeulens DPLE (see [7]) is quite close to DPL, the discussion of kinds of occurrences would be very different. Contexts, in our present set-up, signal the presence of input occurrences, of blockers and of output occurrences. 5 They are abstract -in comparison with formulas- in the sense that they contain no information about the number or the place of these occurrences. Contexts can be studied independently from their connection with the logical language. Contexts are familiar from Predicate Logic. There the context associated to a formula OE is simply the set F of free variables of OE. 6 A salient property of contexts in Predicate Logic is as follows. Suppose F is a context for OE, then Most of the work in this section will be devoted to proving the appropriate DPL-analogue of this property of Predicate Logic. 3.2 Contexts, considered by themselves In this subsection, we treat contexts as mathematical objects in their own right. The natural connection with DPL will surface in the subsequent subsections. Definition 3.1 A DPL-context is a triple hI ; B; Oi, where I , B and O are finite sets of variables and where equivalently, I [B. The set I is the input set, i.e., the set on which the incoming assignments are constrained. The set O is the output set, i.e. the set on which the outgoing assignments are constrained. Finally, the set B is the set of blocks. This is the set of variables for which the identity between input and output value is cut through. The 'block' is a barrier between past and future, breaking the link between input- and output-value. We write # for is defined or converges, and " for is undefined or diverges. Define: ffl id := h;; ;; ;i " is a partial operation on contexts, defined by: " otherwise We will use c, d, . as variables over contexts. We write I c for the first component of c, etcetera. The meanings of these objects, relations and operations will become apparent in subsection 3.3. Lemma 3.2 The operations ffl, ! and " are well defined. Proof To see that ffl is well defined, note that: 5 In fact, there are good reasons also to put witnesses of garbage into the contexts. We will do not this in the present paper, since it is not necessery for our results here. Moreover, adding garbage leads to considerable complication of the framework and it necessitates bringing in Category Theory. We refer the reader further to [9]. 6 The reader is referred to [3] for a category-theoretical framework appropriate for the study of contexts in Predicate Logic. It is trivial that ! is well defined. For the proof that " is well defined we refer the reader to the appendix. 2 Theorem 3.3 The contexts with id and ffl form a monoid. Moreover, - is a partial ordering. The proof of the theorem is easy. In the appendix we will show that " , if defined, is the infimum, w.r.t. Consider the monoid of contexts. It can be represented in an alternative way, as follows. The monoid A is the monoid on two generators a and b, given by the equations: a ffl a b. The tabel of the monoidal operation ffl is a follows. ffl e a b ab ba aba e e a b ab ba aba a a a ab ab aba aba ab ab aba ab ab aba aba ba ba ba b b ba ba aba aba aba ab ab aba aba The monoid of contexts is now given as the set of functions from Var to A that are on all, but finitely many arguments equal to e. We put: (f ffl g)(v) := f(v) ffl g(v). A triple hI ; B; Oi 'translates' to a function f with, e.g., translates to a triple abag, etcetera. It is easily seen that these 'translations' give us an isomorphism of monoids between the representations. A is in fact isomorphic to the monoid of contexts in the case that fxg. The alternative representation is possible by the fact that in our monoidal operation treats all variables 'independently'. It is not difficult to extend the structure on to get a function representation also for !, - and " . In this paper we will stick to the set representation, since this representation is closest to the relational notions we will need to formulate our theorem on contexts and relations -the theorem that tells us what the contexts do. The function representation, however, has two advantages. First, it is easier to use for doing computations 'in the head'. Secondly, its connection with the framework developed in [9] to study contexts, is more perspicuous. 3.3 Contexts and relations We turn to the connection between contexts and relations. 7 We show that this connection 'commutes' w.r.t. ffl=ffi, !=! and -='. We fix a non-empty domain D. Definition 3.4 Consider a relation R on D Var . R is an hI ; B; Oi-relation if We say that c is a context for R, if R is an c-relation. R is an IBO-relation if R is an hI ; B; Oi- relation for some hI ; B; Oi. We assign to each context c the property or 'meaning' fcg D , the set of all c-relations on D Var . (We will often suppress the subscript D.) The heuristic for this property is as follows. R is means that R is only concerned with the values of the incoming assignment on I ; R only cares about the values of the outgoing assingment on O; all this under the constraint that in going from input to output only values of variables in B are changed. Before proving some facts about the notion introduced above, we sample some immediate insights. 7 The semantics for contexts given here certainly does not exhaust all possible uses of contexts. E.g., the problem of explaining what it is to be a variable occurrence of a certain kind is left untouched. Undoubtedly, contexts will play a role in solving this problem. ffl R is a hI ; ;; Ii-relation precisely if R is an hIi-condition. We show that - on contexts describes the 'information ordering' on contexts. The idea is simply that c is more informative than d if fcg ' fdg. Theorem 3.5 1. c - d ) fcg ' fdg. 2. Suppose jDj - 2. Then: fcg ' fdg ) c - d. Proof We prove (1). Let c - d and R 2 fcg. We want to show: Trivially, R is contained in (I I 0 show I I 0 . It is immediate, that f 0 I I fRgIO g 0 . We show that f [B]g. Suppose since fRg, R ' [B] and v 62 B. Finally, . Since R is a c-relation, we may conclude that f 0 Rg 0 . We prove (2). We write X c for VarnX . Let jDj - 2 and fcg ' fdg. Since, [B] 2 fcg, we have fdg. Hence, using the facts of page 3, we find: [B [O Now it is immediate The arguments that I ' I 0 and that O ' O 0 are analogous to one another. We give the argument for the I-case. Suppose, for a reductio, that v 2 InI 0 . Let d and e be distinct elements of D. We for the test: is f(v)=d?. Consider the relation R := so, R 2 fdg. Consider f with In case v 2 B 0 , we have: f e I I 0 fRgIO 0 g. Since, as we have already shown, ]g. Hence, since R 2 fdg, we have f d Rg. Quod non. We turn to the case that v 62 B 0 . Since I 0 nB I I 0 fRgIO 0 g e we find f e . Thus, since R 2 fdg, we have f d v . Quod non. 2 Theorem 3.5 shows clearly that contexts stand in a many-many relation to relations. Thus, the question What is the context of R? has no definite answer. In the appendix we show that, but for one notable exception, every relation has a most informative context. The next lemma may be used in some cases to simplify the verification that a relation is hI ; B; Oi. Lemma 3.6 R is a Oi-relation Proof It is clearly sufficient to show that for any R ' [B]: I ?From left to right is immediate, since IO ' IO " B . For the converse, suppose f 0 I I fRgIO " B g 0 and We have to show: gIO g 0 . Consider v 2 O. In case v 2 B, we have I . Also since fRg, R ' [B] and v 62 B. Putting the identities together, we find The following lemma is quite useful in applications (see example 3.14). We write A := I [B. Lemma 3.7 Suppose R is an hI ; B; Oi-relation and f 0 I I fRg. Then there is a unique g 0 , such that g. This g 0 has the following property: for any set of variables J , if f 0 I J f , then g 0 I J g. As a consequence, we find that g 0 I I g and, hence, g 0 IA g. So, also, g 0 IO g. Proof Let R be an hI ; B; Oi-relation with f 0 I I fRg. Any g 0 with f 0 Rg 0 IB g, must satisfy: f 0 [B]g 0 IB g. So the only possible choice of such a We verify that g 0 , thus defined, satisfies . It is sufficient to show that gIO g 0 . This, in its turn, follows immediately from the property in the last part of the theorem. Consider any set J such that f 0 I J f . Suppose v 2 J . In case v 2 B, we have (v). In case Finally f 0 since v 62 B. Putting things together, we find Note that in the lemma O plays no significant role. Due to the 'forward looking' and time asymmetric nature, however, of the definitions of implication and validity in DPL, it is sufficient for most applications. An immediate consequence of the lemma is that if R is an hI ; B; Oi-relation, then dom(R) is an hIi-condition and (by symmetry) cod(R) is an hOi-condition. Theorem 3.8 Suppose R is a c-relation and S is a d-relation. Then R ffi S is a c ffl d-relation. Proof It is easy to see (R For the converse, suppose that f 0 [B 00 We have to show: f 0 (R ffi S)g 0 . For some h, we have fRhSg. We partition Var into three sets We show that f 0 Rh 0 Sg 0 . We first prove that f 0 Rh 0 . We check the conditions for applying the fact that R is an hI ; B; Oi-relation. 1. f 0 I I 00 f and, hence, since I ' I 00 , f 0 I I f . 2. fRh. 3. hIO[ I 0 h 0 and, hence, hIOh 0 . 4. We show that f 0 [B]h 0 . Consider a variable v not in B. We have to show f 0 (v). We can only run into trouble in case v is not in X 3 , i.e., if v is in BnB 0 or in O [ I 0 . The first possibility is excluded, by the fact that v is not in B. Suppose v is in O [ I 0 . Then: h 0 R is an hI ; B; Oi-relation, so R ' [B]. We may conclude that In case I 00 . In case v 2 I 0 , we find: v 2 I 0 nB ' I 00 . So in both cases: v 2 I 00 . Since f 0 I I 00 f , it follows that (v). Composing the identities, we find By (1)-(4), we may conclude that f 0 Rh 0 . Next, we check the conditions for applying the fact that S is an hI 1. We have hIO[ I 0 h 0 and, hence, h 0 I I 0 h. 2. hSg. 3. We have gIO 00 g 0 and, hence, since O 0 ' O 00 , gIO 0 g 0 . 4. We show that h 0 [B 0 ]g 0 . Consider a variable v not in B 0 . We have to show h 0 Inspecting the definition of h 0 , we see that our desired identity can only fail if either v 62 B or . We consider the case that v 62 B. We already showed that f 0 [B]h 0 . Moreover, we assumed f 0 [B 00 we find: h 0 (v). Next, we consider the case that v 2 B and v 2 O [ I 0 [B we choose v outside of B 0 , we need only consider the possibility that v 2 O [ I 0 . We have, by our early assumption, gIO 00 g 0 , where O we have v 2 I 0 nB (v). Putting the identities together, we find h 0 By (1)-(4), we have: h 0 Sg 0 . 2 Theorem 3.9 Suppose R is a c-relation and S is a d-relation. Then R ! S is a c!d-relation. Proof Suppose R is a c-relation and S is a d-relation. Let Trivially, (R ! S) ' [;]. Moreover, since id ' I I 00 , (R To prove the converse, suppose f 0 I I 00 f and f(R ! S)f . We have to show that f 0 (R ! S)f 0 . Suppose f 0 Rg 0 . By lemma 3.7, there is a g such that fRgI I 00 [B g 0 . It follows that g 0 I I 0 g. Since , we can find an h, such that gSh. Again applying lemma 3.7, we find an h 0 with g 0 Sh 0 .We close this subsection with a language-free soundness result. Definition 3.10 A relation on D Var is DPL-definable over D iff it can be generated by composition from resettings [v] and finitely restricted conditions over D Theorem 3.11 Every DPL-definable relation over D is an IBO-relation. The proof is an obvious induction on the way the relation is generated. 3.4 Contexts and language We turn to our discussion of how contexts are connected to formulas. Definition 3.12 We assign to every DPL-formula OE a context c OE . Define: We write I OE for I c OE , etcetera. Note that the definition correctly defines a function, by the associativity of ffl. We now prove the main theorem of this section. Theorem 3.13 For every formula OE, [OE] is a c OE -relation on D Var . Proof The proof is by induction on OE using theorems 3.8 and 3.3. The atomic cases are easy. 2 We obtain the following picture of the way contexts work: OE is mapped to c OE by abstracting both from part of the vocabulary and part of the structure. c OE is mapped to fc OE g D , a property of relations. Via a different route OE is mapped to the relation [OE] M . The two routes are connected by the theorem that Note that c OE is not always the -minimal context of [OE] M , as is witnessed by the fact that c is the identity on D Var , thus admitting the context h;; ;; ;i Example 3.14 We provide two examples of how theorem 3.13 in combination with lemma 3.7 can be used to verify a valid principle for DPL. We first prove: Suppose [-:OE]g. We have to produce an h with g[OE]h. Since f [-:OE]g, we can find a j with j[OE]g. By theorem 3.13: j[B OE ]g. I OE j. So, gI I OE j[OE]g. By theorem 3.13 and lemma 3.7 we can find an h with g[OE]h. As a second example we prove: Suppose [-]g. We have to produce an h with f [/]h. By our assumptions, there are i and j such that f [OE]i[/]j. Hence, f [B OE ]i and so fI I / i. Thus, fI I / i[/]j. We may conclude that there is an h with f [/]h. The examples demonstrate the role contexts must play in the formulation of schematic principles for DPL. 4 The Switching Property In section 3 we introduced contexts or IBO's as properties of relations and showed that every [OE] is an IBO-relations. A first conjecture for characterizing the DPL-definable relations would be that these are precisely the IBO-relations. We will see, however, that this conjecture is false. To characterize the DPL-definable relations we need one extra property: the Switching Property. In the present section we will prove that the DPL-definable relations do have the Switching Property (soundness). In section 5 will will show that every IBO-relation on an infinite domain that has the Switching Property is DPL-definable (completeness). Definition 4.1 A relation R on D Var has the switching property if it is either a condition or there are variables x and y (not necessarily distinct), such that cod(R). If the second case obtains, we call the variables x and y involved a pair of switching variables. There might be more than one pair of switching variables. There are various other ways to define the Switching Property, but, I submit, the one presented here is the most natural one. In the lemma below, we collect some helpful insights. Lemma 4.2 Suppose R is a relation on D Var 1. 2. Suppose R is C are conditions and where T is a relation. Then dom(R) 3. Suppose C is a condition. Then, C The easy proof is left to the reader. Theorem 4.3 Every DPL-definable relation over D has the switching property. Proof In case R is a condition, we are done. Suppose R is not a condition. As is easily seen, R must be of the form C y, some conditions C, C 0 , and some relation S. (In a formula OE defining R, x would correspond to the first existential quantifier occurring in OE, y to the last. Note that we allow x and y to be the same variable and even the first and last existential quantifier occurrence to be the same occurrence.) We have, using lemma 4.2: dom(R) dom(R) dom(R) dom(R) RThe results of section 3 and of this section combine to the obvious 'soundness'-result: Theorem 4.4 Every DPL-definable relation over D is an IBO-relation with the Switching Property. Example 4.5 We show how to use the Switching Property to prove that certain relations are not DPL-definable. Suppose jDj - 2. ffl Let R := [x := x;y . R is an IBO-relation, with context hfx; yg; fx; yg; fx; ygi. Suppose R has the Switching Property. R is evidently not a condition. Let v; w be a pair of switching variables. Let fRg, d, and d 6= e. Using the fact that the domain of R is the set of all assignments, we find: Hence, by the switching property: f e Rg. But R is obviously injective. So we have a contradiction Suppose our model M is the usual structure of the natural numbers. Let S := [x := x x . S is an IBO-relation with context hfxg; fxg; fxgi. S does not have the Switching Property since: S is not a condition; S has as domain the set of all assignments; S is injective. is an IBO-relation. Surprisingly, the best context we can find for it is hfx; yg; fx; yg; fx; ygi. 8 Suppose that T has the Switching Property. T is not a condition, so we can find switching variables v and w. By symmetry we may assume that v 6= x. We can find f and g with fSg and f(x) 6= g(x). Choose d with d 6= g(v). Using the fact that the domain of T is the set of all assignments, we find: f d By the Switching Property: f d g. But we have two distinct variables x and v such that f d g(v). This is clearly impossible. 8 We will discuss this fenomenon in more detail in subsection 6.3. 5 The DPL-definable relations on an infinite domain In sections 3 and 4 we have seen that the DPL-definable relations are IBO-relations with the Switching Property. Here we show the converse -for the case that the domain, D, is infinite. Theorem 5.1 Let D be an infinite set. Then the DPL-definable relations over D are precisely the IBO-relations with the switching property on D Var . Proof One direction is by our previous results. Let R be an hI ; B; Oi-relation with the switching property. In case R is a condition we are done. Suppose R is not a condition and let x, y be a pair of switching variables. By the switching property cod(y). Note that dom(R) is an hIi-condition and that cod(R) is an hOi-condition. Thus, it is sufficient to show that [x] DPL-definable. [x] is an hInfxg; B; Onfygi-relation, where x; y 2 B. After renaming, we see that it is sufficient to prove that any hI ; B; Oi-relation R, with x; y 2 B and x 62 I and y 62 O is DPL-definable. We will assume x 6= y. In case y, the proof is simpler. To increase readability, we will specify R in a DPL-language that we introduce along the way. Suppose g. Here the i k are supposed to be mutually distinct and similarly for the other sets. Since D is infinite there is a coding of finite sequences of elements of D in D. Par abus de langage, we will confuse this coding with our ordinary sequences of elements of D. Our language has an (m 1)-ary predicate symbol P , where: Remember that [y := x] is short for 9y:y = x. The formula OE is given by: Here stands for the obvious condition. Note that [OE] is an hI ; B; Oi-relation. We claim that R = [OE]. Suppose first that fRg. Take: x . Remember that x 62 g. We have: Clearly, f and g witness that hf(i 1 )ii is in I(P ). Next we set: y We find (using y 62 Since f [B]g, it is easy to see that For the converse, suppose f 0 [OE]g 0 . Let h 1 , be such that: We are going to apply the fact that R is an hI ; B; Oi-relation. By the fact that h 1 and by the definition of P , we can find f and g such that fRg, f(i 1 (x). Since x 62 , it follows that f 0 I I f . Collecting what we have, we see: fRg. By lemma 3.6 we need to check: Consider y 62 O, we find: h 3 Finally, v is not among the b Putting the itemized insights together, we may conclude: f 0 Rg 0 . 2 6 Extensions of the DPL-language We consider three extensions of the DPL-language. One with conjunction interpreted as intersection of relations, one with a new quantifier 99 and one with disjunction interpreted as union of relations. We will show that our contexts work for each of these extensions. The contexts provided for disjunction are not optimally informative and intuitively queer, however. We will give some hints on how we think this apparent defect should be repaired. We show that 99 is definable using - and that in the system with 99 all IBO-relations are definable. 6.1 Conjunction We study the effect of adding intersection of relations to the DPL-repertoire. One way of of thinking about R " S is as: reset simultaneously via R and via S, and compare the results. If they are equal, make the output of our new relation the shared output, otherwise abort. At the syntactical level, we reflect the new operation by extending the language of DPL by adding the clause: ffl If OE; / 2 L, then (OE-/) 2 L. We will call the new language: L(-). The semantic clause is: [OE-/] := [OE] " [/]. We define intersection of contexts as follows. Note that: I So - is a well-defined operation on contexts. We define: Note the difference between - and -. It is easy to see that - is precisely the infimum with respect to -. Theorem 6.1 Let R be a c-relation and let S be a d-relation. Then, R " S is a (c-d)-relation. Proof Suppose that the conditions of the theorem are fulfilled. Let . Suppose f(R " S)g. It follows that f ([B] " [B 0 ])g, (R " S) ' (I I 00 For the converse, suppose f 0 I I 00 f , We extend the definition of c OE to the new language by adding the clause c (OE-/) Theorem 6.1 immediately yields the next theorem. Theorem 6.2 [OE] is a c OE -relation for every OE2L(-). We consider an example. Let c g. Let / := (OE:9b 1 We have: Thus, in this example our conjunction on contexts, gives us the 'intuitive result', i.e., [/] is an I- condition. Let's say that the DPL(-relations over a given domain D are the relations on this domain generated by finitely restricted conditions and resettings using composition and intersection. The results of the present section show that the DPL(-relations are all IBO-relations. The results of the next section, will imply that, conversely, every IBO-relation is DPL(-). 6.2 A new existential quantifier We define: 99x(R) := ([x] In case R is an hI ; B; Oi-relation, we see that 99x(R) := ([x] ffi " [Bnfxg]. It follows -by the result of subsection 6.1- that 99(R) is an hInfxg; Bnfxg; Onfxgi- relation. We extend the language of DPL by adding the clause: ffl If OE 2 L and v 2 Var then 99v(OE) 2 L Note the overloading of notations. The new language will be L(99). The new semantical clause is the obvious: [99vOE] . 99v is definable in L(-) as follows. Suppose c then we can put (9x:OE:9x - 9b Theorem 6.3 For any non-empty domain D, the DPL(99)-definable relations are precisely the IBO- relations. Proof [Sketch] We have already seen that every DPL(99)-definable relation is IBO. For the converse, suppose that R is an g. Here the i k are supposed to be mutually distinct and similarly for the other sets. Take a DPL(99)- language with an (m p)-ary predicate symbol P , where: and f(o 1 Let be variables disjoint from I [B [O. Let OE be given by: Clearly, that [OE] is an hI ; B; Oi-relation. The verification that R = [OE] is along the lines of the proof of theorem 5.1. 2 Example 6.4 We show how to define the three relations of example 4.5. We do a bit more than the theorem promises, because we give explicit descriptions of the 'P '. can be defined by: Note that this gives us the expected context: hfx; yg; fx; yg; fx; ygi. can be defined (Strictly speaking we are working in a relational language, so suggestive notation for, say, S(v; x).) The context produced by the formula is as expected. ffl We can define [(9x-9y)] by: This gives us the context hfx; yg; fx; yg; fx; ygi. We will discuss this context in the next subsection Since 99 is definable using -, the 'expressive completeness' of 99 w.r.t. the IBO-relations implies the 'expressive completeness' of -. Finally, we can translate Predicate Logic into DPL(-), by changing the 9-clause of our earlier translation to: (9(OE)) := 99(OE ). Remarkably, the old and the new translation produce precisely the same relations at the semantical level. Operations like !, - and thereis are not themselves actions in the sense of our semantics. They are transformers of actions. Yet there is a tendency to understand 99x(OE) dynamically as a sequence of actions: reset x; execute OE; set x back to its original value. The problem with this way of viewing things is summarized with the question: where do we store the original value of x, so that we can restore it at the end? DPL-semantics does not supply the right kind of 'memory' to realize 99(OE) as a sequence of actions. We can do that (or, rather, something very much like it) in the richer semantics of Kees Vermeulen's DPLE (see [7]), where under a variablename we do not store just one value, but a stack of values. Here the original value of x is simply stored 'under' the new one. RELATIONS IN CONTEXT 17 6.3 Disjunction In this section we have a brief look at the problem of adding disjunction/union to DPL. Adding disjunction/union evokes problems that are definitely beyond the scope of the present paper. So we can only offer some tantalizing remarks. One way of of thinking about R[S is as: Choose between R and S, and reset via the relation chosen. At the syntactical level, we reflect the new operation by extending the language of DPL by adding the clause: ffl If OE; / 2 L, then (OE-/) 2 L. We will call the new language: L(-). The semantic clause is: [OE-/] := [OE] [ [/]. What could be a context for [x] [ [y]? Some experimentation shows that the best we can do is: This seems a wasteful way to represent the variable handling of this relation. Our intuition tells us that [x] [ [y] is a pure resetter and not something that 'constrains' w.r.t. x and y. The resetting part of our contexts is somehow too crude to represent 'choice' well. The example does not tell us that in any strict sense our present framework is wrong. It just suggests that, possibly, we could do better. We might try out richer notions of context. The most obvious proposal is to take as a context in the new sense a set of contexts in the old sense, where the set is given 'disjunctive reading'. 9 So, e.g., we would have: c (9x-9y):P (x;y) Note that e.g. the second occurrence of x in (9x-9y):P (x; y) seems to be ambiguous between free and actively bound. So what is an ambiguous occurrence and how do we handle it theoretically? We propose to addres this question elsewhere. 7 Relations in context In DPL meanings are relations. The contexts we studied appear as properties of these relations. We could give an alternative semantics for DPL by building the context into the meaning. Thus we take as meanings pairs hc; Ri, where R 2 fcg. Let's call such a pair a c-relation. We define: The new domain of meanings is, on the one hand, essentially richer than the old one, since the same relation falls under several contexts. On the other hand, we threw all non-IBO-relations away. We can 'lift' the notions intoduced in this paper to c-relations: ffl A c-condition is a c-relations of the form hhI ; ;; Ii; Ri. ffl If diag(dom(R))i. Similarly for cod. ffl A c-relation R has the Switching Property if it is either a c-condition or there are variables v and w such that: ffl A c-relation is DPL-definable (over a given domain D) if it can be generated using ffl from c-conditions and resettings 9 In fact, I think, this proposal is in the right direction, but still not quite right. I am convinced that the enriched semantics is better than the usual one, but I will not argue the case here. RELATIONS IN CONTEXT In a similar way we can upgrade - and 99. Inspection of the proofs in this paper shows that, in case D is infinite, the DPL-definable c-relations are precisely the ones with the Switching Property. Moreover, all c-relations over the given domain -infinite or not- are DPL(99)-definable. We consider an example. Remember that: Let a model with domain D be given. We assume that D has at least two elements. Let id := id D Var . We consider three c-relations with associate relation id. 1. 2. 3. (1) and (2) are c-conditions and, hence DPL-definable. In fact, they can be defined in the language by ?, respectively In contrast, (3) is not a condition. It is easy to see that (3) does not have the Switching Property, since the domain of its internal relation id consists of all assignments and id is injective. Hence (3) is not DPL-definable. Note that hh;; fxg; ;i; idi is not a c-relation at all. A further step in modifying our semantics is to make the assignments 'local'. The idea is that the context 'provides' the files/discourse referents/variables on which the variables are defined. Thus, our meaning objects would be of the form hhI ; B; Oi; Ri, where R would be a relation taking input assignments defined on I and yielding output assignments defined on O. This approach leads to a semantics very much like Vermeulen's Referent Systems (see [8]). One effect of this further modification is that it leads to a somewhat different view of contexts. In the local approach, contexts are the central 'engines' that manage the flow of the files in the interactions of meanings. This more dramatic view of contexts is elaborated in [9]. Concluding remarks In this paper we introduced a notion of context and specified its connections with relational semantics and language. We used these contexts to prove a characterization of the DPL-definable relations. Moreover, we illustrated the usefulness of contexts both in formulating and in verifying valid sequents of DPL. We illustrated the fact that 'understanding of what is going on' is not automatically preserved if we extend the DPL-language. E.g., adding disjunction leads to ambiguous occurrences of variables. This observation tells us that the study of extensions will provide us clues regarding the question: what is it to be a variable occurrence of a certain kind? So -apart from the concrete results- what general conclusions may we draw from the paper? A first one is, surely, that a study of the elementaria of DPL is both necessary and rewarding. Questions on the nature of variable occurrences, the proper notion of syntactic substitution, etcetera, appear in a new light. The fruitfulness of the study of DPL is independent of the question whether DPL is really the best choice as a medium for representing dynamic phenomena. One reason is that, in a sense, the relational semantics of DPL is very simple and that it is, therefore, easier to make progress. A second conclusion is that it is rewarding to engage in a study that stresses the differences between DPL and Predicate Logic. Much effort has gone into integrating DPL into the classical Montague framework. This project has unavoidably a conservative flavour. The result has been that the unfamiliarity, the strangeness of DPL has been underadvertised. Precisely mastering the strangeness provides us with better insight into the formerly familiar notions. My third conclusion is simply: contexts are essential in the study of DPL and its kin. We may want to vary the contexts, e.g., we may want to add 'garbage elements' or to ignore the O-component, but contexts per se are there to stay. Our third conclusion points to a larger programmatic point: the study of contexts and the way they are contexts of their contents should be one of the central endeavors of the study of Information Processing and Dynamics. REFERENCES 19 Acknowledgements I thank Kees Vermeulen for many enlightening conversations and for his careful reading of a version of this paper. I am grateful the students of various classes in Dynamic Logic both in Amsterdam and in Utrecht for providing a critical forum for some of my ideas on DPL and Dynamics. --R Dynamic predicate logic. Counting variables in a dynamic setting. Categorical Type Theory. A theory of truth and semantic representation. Anaphora and the logic of change. Defaults in update semantics. Sequence semantics for dynamic predicate logic. Merging without mystery Dynamic bracketing and discourse representation. --TR --CTR C. F. M. Vermeulen, Variables as Stacks, Journal of Logic, Language and Information, v.9 n.2, p.143-167, April 2000 C. F. M. Vermeulen, Text Structure and Proof Structure, Journal of Logic, Language and Information, v.9 n.3, p.273-311, July 2000
variable;expressability;predicate logic;resetting;dynamics;context;register;relation;definability
595906
Communication across Viewpoints.
A case situation arising in a normal interaction among people is the baseline for discussing properties of the theory of viewpoints. In particular we consider how to ensure agreement on the meaning of certain utterances by agents who have different perspectives on the situation, while maintaning other knowledge as private. We argue that communication should be modeled as adding facts to the common knowledge of agents. We introduce the principle of referent sharing in communications and argue that common knowledge resulting from communication should only use constants whose referent is manifest to the parties involved.
Introduction When several agents interact, their understanding of each other behaviour depends on some common knowledge they all share as well as on assumptions about each other private knowledge. The question arises then of how such common knowledge comes about. Communication between agents seems to involve two aspects: ensuring that they all understand the subjects of their sentences and becoming all aware that the information has been transferred. The first aspect requires that the "referents" in communications be shared. The second entails that communication not only involves transfer of knowledge from one agent to another but also extends their common knowledge. These two facts together imply that common knowledge resulting from communication should only use constants whose referent is "manifest" to the parties involved. We explore these issues in the framework of the theory of viewpoints [2, 4, 5]. This theory was conceived as a general and unified formalism for expressing several varieties of relativised beliefs contexts, situations, truth, partitioning a knowledge base in microtheories and so on. Each of these notions can be represented through viewpoints whose specific properties are captured by axioms, added to the basic theory in order to characterise those viewpoints [26]. The theory of viewpoints is a reflective first order logic which amalgamates object and metalanguage by using names for each term and statement of the language and which contains an axiomatisation of provability in the style of natural deduction. Reflection rules provide the link between object and metalanguage and lead to a non conservative but consistent extension of first order logic; these rules have been carefully formulated in order to avoid paradoxes arising from self referential sentences, which trickle in by diagonalisation [20]. Viewpoints denote sets of sentences which represent the assumptions of a theory. A statement of the form in( where vp is a viewpoint expression, is interpreted as "statement A is entailed by the assumptions denoted by vp in the current interpretation" 1 . The motivations and properties of the theory of viewpoints differ in several important respects from existing proposals and formal accounts of contexts [18, 13, 25, 7, 8]. In the formal system by Buva-c and Mason for example [8], the semantics of ist(c; p), to be read as "it is true in context c that p", is essentially (language restrictions apart) entailment: the semantics associates to a context a set of models and p is assessed to be true in such models. As a consequence a context will have complete information of what is true in other contexts: Moreover the context structure, at least in the quantified version [9], is flat, according to their own In the theory of viewpoints, since they are meant also for modeling belief contexts, the perspective of observation is important: the fact that something holds in a viewpoint does not imply that "the fact that it holds" is true in any other viewpoint. Therefore the property is not desired and is not valid. To achieve this, the semantics provided for in is not entailment but rather contextual entailment, so that the viewpoint vp 1 is resolved (interpreted) with respect to the context where it appears (i.e. vp 2 rather than in a global context. This means that the viewpoint expression vp 1 in the antecedent might denote a different set of assumptions from the set of assumptions denoted by the vp 1 in the consequent. Note however that: is valid, since contextual entailment ensures that viewpoint expressions at different levels of nesting are interpreted coherently: the entailment is in fact restricted to a subclass of all models, those "coherent" with the current context in the interpretation of the viewpoint expressions. As an additional consequence the rules for entering and exiting a context have to preserve the nesting of contexts, as argumented in [6] and as in the propositional version of the theory of contexts by Buva-c and Mason [7, 8] but unlike in their quantified version [9]. The expressivity we require calls for implicitly defined viewpoints and a syntactic treatment of in, so that self referential and mutually referential viewpoints can be defined. In many real applications 1 The notation 0 is an abbreviation for a term denoting sentence A. The naming device is not important in this paper; we will feel free to drop quotes inside in when there is no risk of ambiguities. contexts cannot be characterised explicitly by listing the complete set of assumptions. This is also an issue raised by McCarthy when he suggests that contexts are to be considered rich objects [18]. Therefore, in addition to viewpoints expressed by means of sentence names, we also allow for viewpoint constants and functions and rely on a partial characterisation by means of axioms stating what holds in a viewpoint. In particular one can express the rules for the derivation of facts of type vp). Examples of such assertions are lifting rules [18], which relate two separate viewpoints; for example it is possible to state that whenever a formula satisfying some condition holds in a viewpoint vp 1 then a related formula holds in viewpoint vp 2 As a special case, we can state that vp 1 subsumes vp 2 This allows for compact statements of problems and leaves to the logic machinery the burden of extracting statement present in viewpoints as necessary during proofs from the rules which charaterise them. Note that the possibility of quantification over sentences inside the in operator, excluded in modal logics, is essential to this aim; hence our choice of a syntactic treatment of the in operator. Another important characteristic of the theory of viewpoints is that it is a reflective theory, as opposed to a stratification of theories as in [27, 12], despite the troubles created by paradoxes. This allows expressing facts which simply cannot be stated without the use of self referential viewpoints. Here are some examples. "John believes that he has a false belief " [21] "Agent a believes that whatever he and agent b believe is true, while b does not believe so" "Agent a and agent b have common knowledge (or belief) that p" Note that to faithfully represent common knowledge it is not enough to state that both agents know p, but it is also necessary to express that they know that they know p, that they know that they know that they know The possibility of recursive fedinitions of viewpoints (for example through lifting axioms or in particular "auto-lifting" axioms such as CK-3) and of expressing self referential viewpoints leads to non finite viewpoints. 2 Syntax and proof theory The logic of viewpoints has the syntax of classical first order logic, extended with the statements is the name of a statement and vp can be: 1. a finite set of statement names: f 0 A 0; 0 A 0; :::; 0 A 0 2. a viewpoint term, consisting of a viewpoint function and a list of terms. A viewpoint function of no argument is a viewpoint constant. For each symbol s of the language (either variable, function, predicate or viewpoint), we denote by its name. Names for terms and literals are obtained through an appropriate naming schema as in [5]. In the following we use the convention that A, B are metavariables for statements, 0 A 0 , are quoted statements while vp; vp 1 are metavariables for viewpoints, i.e. expressions denoting sets of sentences. To simplify the notation, we will sometimes use vp; for sets of sentences or, in the case of finite viewpoints, for the sentence which is the conjunction of the sentences in the viewpoint. It should be clear from the context which is the intended meaning for such meta-variables. The proof theory for viewpoints can be conveniently presented in the style of natural deduction. 2.1 Inference rules for classical natural deduction As customary, the notation \Gamma ' A indicates the pending assumptions \Gamma in the proof of A in rules where some of the assumptions are discharged, like in the cases of implication introduction and negation introduction. When the pending assumptions are the same in the antecedent and consequent of a rule they are omitted. The rules for natural deduction are quite standard. For example: (- E) are the rules for conjunction introduction and elimination, respectively, and E) are the rules for implication introduction and elimination. The full set of classical rules used is presented in [6]. 2.2 Metalevel axioms The behaviour of in is characterised by the following axioms and inference rules, which allow classical reasoning to be performed inside any viewpoint, and at any level of nesting. The first axiom asserts that all the sentences which constitute a viewpoint hold in the viewpoint itself, while the second states monotonicity of viewpoints. An additional axiom establishes a principle which could be called positive introspection, if we choose an epistemic interpretation for in. (Positive introspection) 2.3 Metalevel inference rules The following are the rules of reflection and proof in context for the theory of viewpoints: they are more powerful than classical reflection and reification rules, but still safe from paradoxes as discussed in [4]. (Reflection) We have argued elsewhere for the usefulness of such a strong version of reflection [4], [5]. The metalevel should have the same inferencing capabilities of the object level. This could be provided by means of one meta-inference rule for each classical object level inference rule, as described in [5], or, more succintly, through the following rule of proof in context, together with a meta-rule for implication introduction, which could be regarded as a principle of decontextualisation [13]. Axiom1 and Axiom2 are also necessary to allow for classical reasoning in any viewpoint. (Proof in context) where the vp 1 in the antecedent is a finite set of sentences and the vp 1 in the consequent should be read as the conjunction of the formulae in vp 1 . The notation 'C stands for classically derivable or derivable without using reflection and positive introspection. (Decontextualisation) The converse rule of contextualisation is a derived rule: The rule of proof in context asserts that a classical proof can be performed in any viewpoint provided the premises of the proof are in the viewpoint. Notice that the consequent of the rule is still a classical derivation, therefore the rule can be applied repeatedly, to carry out a proof at any level of nesting within viewpoints. Therefore, if A 'C B then: A corollary of proof in context is the following: Theorem 1 in( logical theorem A of classical natural deduction and for any view-point vp. Proof in context is more powerful and a much more useful rule than classical reification when working with implicit viewpoints [6]. Classical reification is also a derived rule: vp 'C A (Reification) 2.4 Entering and exiting contexts A useful mechanism when performing proofs in context is the ability to switch contexts and perform natural deduction proofs within viewpoints. The safest way to interpret context switching in the framework of natural deduction proofs with pending assumptions and implicit contexts is simply to go one level deeper or shallower in nesting, or in other words unnesting and nesting. This means for instance that in order to prove a statement of the form one may pretend to move inside vp 1 , and perform a proof using available facts of the form ). If the formula A is itself of the form in( one will have to go one level deeper to prove B by using this time just facts of the form in( In [6] we introduced a box notation for natural deduction proofs and viewpoints, making visible the structure of dependencies and nested viewpoints in contextual reasoning. The notation is an extension of the box notation introduced by Kalish and Montague [16]. For example, the following schema corresponds to the inference rule of negation introduction and should be read as: "if assuming A a contradiction is reached, then :A is proved". A (assum.) The box notation is useful to visualise the scope of the assumptions made during a natural deduction proof. In performing a proof within a box one can use facts proved or assumed in the same box or in enclosing boxes. Facts cannot be exported from within a box to an enclosing or unrelated box. For proofs in context we introduce a different kind of box, with a double border, to suggest boundaries which are more difficult to traverse. The double box represents a viewpoint, i.e. a theory, whose assumptions, if known, are listed in the heading of the box. If the assumptions are not known the name of the viewpoint is shown. The only two rules for carrying facts in and out of a double box are the rules corresponding to unnesting and nesting. Importing a fact in a viewpoint: vp A Exporting a fact from a viewpoint: vp A (nesting) The only way to import a fact A in a double box vp is to have a statement in( in the environment immediately outside the box. Symmetrically one can obtain in( in the environment immediately outside a double box vp if A appears in a line immediately inside the double box (not inside a further single or double box within the double box). Note that to import a fact into a nested double box the rule of unnesting must be applied repeatedly. According to Axiom1, the assumptions of a viewpoint can also be used inside the viewpoint: The main problem in providing a semantics to a reflective theory is to decide which truth value to assign to statements of the form R , F ( 0 :R 0 ), which can be built by diagonalisation. Solutions proposed in the context of logics for truth, can be a source of inspiration. Kripke's account of truth [17] is based on Kleene's three valued logic. In his model the law of excluded middle holds, while the principle of bivalence does not: in( does not hold whenever A is a classical tautology, like B - :B. A counterexample is R - :R, where R is a paradoxical sentence. The solution by Perlis [21] also does not preserve the principle of bivalence. We adopt therefore a different solution based on the idea of reaching a stable truth value for a formula through an iterative process of revision as in Gupta-Herzberger semantic theory [14, 15]. We however retain the notion of truth and validity which Gupta-Herzberger leave undeterminate for some statements. An adequate semantics must avoid to sanction dangerous formulae. For example in consistent reflective theories, a T axiom like is often allowed, but in this case the reification rule (called necessitation in the context of modal logics) must be restricted so that is not derivable. Its derivation in fact would immediately lead to an inconsistent theory [20]. The same argument applies to the reified positive introspection. Another problem is accounting for positive introspection, which is a useful principle for our pur- poses, without also having to admit the slightly different, stronger which is not desirable and in fact cannot be proved in our theory. As a consequence, an interpretation of in as straightforward entailment or derivability is not satisfactory. Our answer to this problem is an interpretation of in as contextual entailment , i.e. entailment restricted to a subclass of all models, to ensure that viewpoints at different levels of nesting are interpreted coherently. We want to interpret in(t 1 in a certain model M, as entailment of t 1 from t 2 , i.e. that whenever holds, then t 1 holds. However M establishes an interpretation for viewpoint functions some of which may appear in t 1 ; we must ensure that the same interpretation is used in evaluating t 1 Consider for instance: When examining the validity of the consequent, we are already restricted to models where vp is a viewpoint which entails A. Therefore, when we consider whether vp entails in( must carry over this restriction, and in fact we conclude that in all such models A is entailed by vp. The interpretation of in as contextual entailment rests upon the notion of coherence of models. We say that two models are coherent with respect to a viewpoint expression vp when they agree on the interpretation given to the viewpoint constants and functions appearing in vp; formally: Coherence is a reflexive, symmetric and transitive relation between models. An interpretation structure M for the theory of viewpoints is a pair hD; Ii defined as follows: 1. The domain of interpretation D contains also the set of statements S of the language. 2. I associates an n-ary function f I : D n ! D, to each n-ary function symbol f of the language, except the syntax constructors, i.e. those functions that we use to name sentences and terms of the language, whose interpretation is fixed. 3. I associates to each n-ary predicate symbol p of the language, an n-ary predicate p I ' D n . 4. I associates to each n-ary viewpoint function vp, a function vp I : D n ! RecSubs(S); the recursive subsets of S. Terms are interpreted with respect to an assignment function g, which assigns elements of D to variables. The interpretation of terms is classical except for naming terms, whose interpretation is the term or statement they name. In particular: In short we could write, for any term t without variables: For each formula A, means that A is true in the interpretation M with assignment function g. The notation g[i=x] represents the assignment function identical to g except that x is bound to i. For the interpretation of statements we proceed by defining the notion of true at level n. For statements which do not contain variables quantified across in statements, the level corresponds to the level of nesting of in statements. A statement with a nesting level of n will receive its definitive truth value at level n; while paradoxical statements, to which no finite level of nesting can be assigned, will keep oscillating periodically between true and false. At level 0, all statements are considered false: not M;g A. For any n ? 0: M;g (:A) iff not M;g A M;g A and M;g A or M;g (8x:A) iff for all d in D, 9k: M;g[d=x] A The truth of in statements at level n is defined on the basis of the truth at level n \Gamma 1. The interpretation of in is contextual entailment , i.e. entailment restricted to a subclass of all models, to ensure that viewpoints at different levels of nesting are interpreted coherently. such that (N j t 2 M), not or In most cases, this process of revision stabilises in the sense that the truth value of statements settles to either true or false from a certain level onward. There are however paradoxical statements, like the counterpart of the liar statement, which do not stabilise at any finite level of revision but continue to oscillate between different truth values. To such statements false will be assigned as ultimate truth value. A statement is said to be stably true accordingly to whether or not the revision process stabilises: M;g A Stable truth is however too coarse, since all statements which do not stabilise are considered false: therefore for any such statement R both R and :R would be false. One further step is necessary to distinguish among the statements which do not stabilise. The notion of truth is classical except for in statements which relies on stable Validity is defined as usual: The notation vp j= A means that A is true in every model of vp. We can verify that our semantics has the required properties. Note. The law of excluded middle holds both at the object level: A - :A and at the metalevel: However this does not imply that in( which would be undesirable since it would sanction the completeness of any viewpoint. Consider the formula R, such that R , in( which corresponds to the Liar sentence. The paradox is avoided since the Liar sentence is false in every model: We summarise a few relevant properties of our semantics: 1. M A implies A but not vice versa. 2. in( 3. (in( 0 A 4. in( is not valid. 5. in( vp) is not valid. 6. in( 7. in( ) is not valid. 4 Sharing of referents in communication We will illustrate with an example several features of the theory of viewpoints and how to reason in it. The example is meant to illustrate several issues related to the modeling of communications among agents holding different viewpoints about the same state of affairs and in particular using different expressions to denote the objects of interest. One morning Beppe told Maria about the following dream. You and me are traveling by train, and you have both our tickets. I go to the toilet. After my return the ticket inspector passes and asks for unchecked tickets. I do nothing, since I know that you have my ticket. You do nothing since you have already shown both tickets to the inspector while I was away. The inspector does not remember seeing my ticket and therefore he asks me for it. At this point I ask you why you do not show the ticket to the inspector. We will attempt a formalisation of the reasoning of each agent involved in the situation described above at different times, providing a rational account for the agent's behaviour. In the reasoning by the inspector and by Maria, they both assume knowing certain information about each other. Each one should have obtained such information through the transaction that happened between them when Maria showed the tickets to the inspector. The transaction should be modeled in such a way that some information flows from one agent to the other and that both agents become aware not only of the transaction but also that the other is aware of the information conveyed in the transaction. The proper way to model the effect of a communication act is by extending the common knowledge of the participants. The effect is not only that some information is added to the knowledge of the recipients of the communication, but also that the recipients' awareness of such information enriches the viewpoint of the originator of the communication, and the recipients become aware of this and so on. Communication also raises a problem about terms and denotations: i.e. which constants should be used in the formalisation of the communicated knowledge. For instance, in the interaction between Maria and the inspector while Beppe is away, how can Maria refer to Beppe? In our solution we decided that she can't refer to Beppe directly, since she has no way to ensure that her reference to Beppe will be the same as the one of the inspector. The only things to which she can refer in her interaction with the inspector are what we call "manifest constants", i.e. constants to which she can point directly (for instance because they are objects in the scene), or indirectly through terms built upon other manifest constants (for instance Husband-of(M), or Person-sitting(there)). We could even use Person-named("Beppe"), since the name "Beppe" is part of the common vocabulary, however the referent of such name may not be the same in all viewpoints. Quine speaks about "observations terms", which are terms that are or can be taught by ostension, and whose application in each particular case can therefore be checked intersubjectively [24]. Learning, and in particular language learning, is predicated on observation terms and hence on ostension: as children learning the language, we get on to various simple terms and key phrases by direct association with appropriate experiences. When we have progressed a bit with this kind of learning, we learn further usages contextually [23]. We require that any statement expressing common knowledge between two agents, gained as a result of a communication, only uses manifest constants: we call this the principle of referent sharing in communicating knowledge. Therefore in the example, the transaction between Maria and the inspector will be expressed by means of the manifest constant ticket2 which refers to the actual ticket that Maria hands over to the inspector. If we were to use T icket(B), where B is the constant denoting Beppe, we would unduly transfer Maria's knowledge that Beppe's ticket has been checked to the viewpoint of the inspector. A different approach would be not to worry about terms used in the communication and rely on a theory which provides for different denotations of terms in different contexts, as for example the theories of context discussed by Buva-c, Guha and McCarthy. 5 Formalisation of Beppe's dream The following notation will be used in the formalisation of the example introduced above. B: Beppe Beppe's viewpoint at time t Maria's viewpoint at time t vp(C; t): the ticket inspector's viewpoint at time t knowledge of the set of agents denoted by x at time t; in the example we will use All for everybody, BM for Beppe and Maria and MC for Maria and the inspector; Has(x; shows ticket y to the inspector at time t; ticket x has been checked. Common knowledge of a set of agents plays a significant role in this example and we express it as one theory for lifting axioms provided for each involved agent to access it. A general formulation of common knowledge which will serve the purpose of the example, is the following. Axioms for common knowledge: to be used in conjunction with the following: Axioms (1) and (2) provide a proper account of common knowledge to a group of agents, allowing to derive the commonly known facts in any viewpoint, no matter how nested. In particular axiom (2) is used to achieve the appropriate level of nesting in CK, and axiom (1) to lift from the CK viewpoint at time t to any other relevant viewpoint at time t. For example assuming: (a) it is possible to derive which will appear in the proof as line (33), with the following steps: (c) in(in(Has(M; Ticket(B)); vp(M; t 3 (d) in(in(Has(M; Ticket(B)); vp(M; t 3 To be more concise, we will skip similar derivation steps from common knowledge from now on. The specific knowledge of the problem can be expressed as follows: Both Beppe and Maria know that Maria has both tickets, at any time; Corresponds to the common sense knowledge that only one person can hold a given object. Next we need axioms to describe Maria's act of showing the tickets to the inspector and its effect, i.e. the information conveyed in that act. They deserve some discussion. In formalising the information flow one must avoid that also unintended information be transferred between viewpoints. For example, one might express as follows the fact that once Maria has shown her tickets, the inspector checks them and thereafter they both know that they have been checked: Maria knows that she has shown both tickets at time t 0 Maria knows that the second ticket is Beppe's; It is common knowledge that after Maria has shown a ticket, at any later time she and the inspector know that the ticket has been checked. However this solution has a flaw: applying the law of substitutivity in Maria's viewpoint, Maria knows that Shows(M; ticket2; t 0 therefore she can con- clude, for instance, that in(Checked(Ticket(B)); vp(C; t 3 )). The effect is therefore that she unduly transfers her information about ticket2 being Beppe's ticket to the viewpoint of the inspector. The solution is then to separate the information flow from the conclusion that each agent is able to draw from the information gathered. First, we state that showing the ticket is common knowledge of both Maria and the inspector. This is the information that is actually transmitted. Maria and the inspector know that she has shown both tickets at time t 0 while the first ticket can be referred to as Maria's ticket the second one is simply ``a second ticket'' in the common knowledge viewpoint; Maria knows that the second ticket is Beppe's (while the inspector doesn't); The first statement uses ticket2 and cannot use T icket(B) since this is not a manifest constant and is therefore forbidden by the principle of referent sharing in a statement on common knowledge. And now we allow each agent to draw his own conclusion about which tickets have been checked: It is common knowledge of Maria and the inspector that after Maria has shown a ticket, the ticket has been checked. Using these statements, we can easily prove the following two lemmas that will be used in the formal account of the reasoning. The first lemma, corresponds to the fact that Maria knows that her ticket and Beppe's have been checked by the inspector. The second lemma, that will be used in Maria's reasoning, corresponds to the fact that she knows that the inspector is aware of having checked her ticket and a second one, referred to as ticket2. in particular, the following facts that Maria will use in her reasoning, can also be proved in viewpoint while it is not the case that the inspector, nor Maria reasoning about the inspector, can deduce: Finally we formalise the first request from the inspector (the request to show unseen tickets), which, at time t 1 , becomes part of common knowledge in the following form: Figure 1 summarises the axioms that have been used in the statement of the problem. We can now proceed to give a formal account of the reasoning of the different agents at the different times involved. This is the reasoning in Beppe's viewpoint, at time t 1 Figure 1: Problem statement Maria has my ticket; I do not have my ticket; Since I do not have a ticket the request from the inspector does not apply to me. I do not have to show my ticket. Therefore Beppe does nothing. Maria reasons as follows at the same time. The inspector has already checked our tickets; The request of the inspector does not apply to Beppe's ticket; I do not have to show Beppe's ticket; The request of the inspector does not apply to my ticket; I do not have to show my ticket. Therefore also Maria has good reasons for doing nothing. After this, at time t 2 , the common knowledge is enriched as follows for the fact that nobody shows any ticket: Moreover we have the fact that the inspector doesn't remember seeing Beppe's ticket, which is part of the statement of the problem; this however is not common knowledge but it is asserted outside any viewpoint. Reasoning about the inspector in the external viewpoint, where the problem has been stated, and including axioms (1)-(10) plus axiom (21). the inspector notices, as anybody, that nobody shows Beppe's ticket at time t 1 he reasons that either the ticket has been checked or nobody has Beppe's ticket. Assume the inspector believes that somebody has Beppe's the ticket; he should at least believe that Beppe's ticket has been checked; but he doesn't by (21), therefore: the inspector cannot deduce that somebody has Beppe's ticket. This same reasoning can be performed by the inspector himself, provided we postulate the addi- tional, and reasonable, assumption that he is aware of the fact that he doesn't remember. This can be made explicit in the statement of the problem or alternatively derived from (21) by negative introspection (assuming this is a valid principle in general). a realisation by the inspector of the fact that he cannot tell whether somebody has Beppe's ticket. As a consequence he asks Beppe to show his ticket. It is now common knowledge, at time t 3 , that the inspector does not know whether somebody has Beppe's ticket. Beppe now, at time t 3 The inspector does not know whether somebody has my ticket; Assume the inspector thinks that he has seen my ticket; He should also believe that somebody has my ticket, but he doesn't by (30); therefore he does not believe that my ticket has been checked; Maria knows that she did not show my ticket when asked; Maria knows that she has the ticket; Maria is aware of what the inspector said; Maria believes that my ticket has been checked. Therefore Beppe notices a difference between the inspector's viewpoint and Maria's viewpoint (lines 33 and 37); that's why he asks Maria why she doesn't show his ticket. Finally, solicited by the request of the inspector and, further, from Beppe himself, Maria reasons The inspector is not aware of the fact that he checked Beppe's ticket (similar to Beppe's reasoning above). but he certainly remembers having checked a second ticket; he doesn't realize that the second ticket is Beppe's one (otherwise there would be a contradiction). she decides that in order to resolve the question she must tell the inspector that the second ticket is Beppe's one. This goes a little beyond Beppe's dream, but it looks like a plausible reason of the fact that the inspector does not remember: Maria, showing both tickets Beppe being absent could not refer to Beppe as the owner of the second ticket. 6 Technical points about this example The reasoning performed by the agents involved in the situation described by the example is simple and intuitive, yet it appears that its formal account has some requisites in terms of expressivity of the logical representation language to make it an interesting case study. In the following we will discuss additional aspects of the proposed formalisation. 6.1 Viewpoint functions We have several agents (Maria, Beppe, the inspector) and several times corresponding to situations with different knowledge involved. We use vp(a; t), a viewpoint function denoting the viewpoint of agent a at time t; the example shows the importance of viewpoint functions with variables quantified in; this way it is possible to express, for example, that a certain fact holds from a certain time on, or, if necessary, that there is a time when a fact holds. 6.2 Propagation of ignorance The inspector does not remember having checked Beppe's ticket, and as a consequence he makes everybody aware of the fact that he does not know whether Beppe has a ticket, and as a further consequence Maria can realize that the inspector does not know that the second ticket he has seen is Beppe's ticket. Note that in the example we do not deal with deduction of ignorance from the fact somebody cannot deduce something, a property which is in general undecidable in a first order setting (see [1, 10] for examples of this approach). In other words ignorance in not "introduced" but only "propagated". All we do is to use a much more conservative and sound pattern of reasoning for propagation of ignorance, namely: if we have :in(A; vp) and we want to prove :in(B; vp), we can assume in(B; vp) and try to prove in(A; vp); if this succeeds, we have succeeded in proving :in(B; vp). In other words the following is a sound rule of inference: (Propagation of Ignorance) 6.3 Viewpoint consistency is an option In the example we do not deal with nonmonotonic aspects such as changes of mind, assumptions by default, resolving contradictions, and so on. In this problem once a piece of knowledge is acquired in a viewpoint it is also assumed to hold in successive viewpoints in time. Dealing with nonmonotonic aspects would require writing problem dependent axioms stating what changes and what doesn't from one viewpoint to the next in time. In this respect some of the ideas presented in connection with the step/active logics by Perlis and co-authors look promising [22, 19]. As an example we could exploit their approach to resolving contradictions, once they manifest explicitly, by failing to inherit to the next step contradicting assumptions. This can be done since consistency of viewpoints is an option: the axiom :in(false; vp) does not hold in general; a viewpoint can become inconsistent without affecting other viewpoints nor the global reasoning context. This means that in the presence of a contradictory viewpoint, without the consistency axiom, the inconsistency does not propagate outside and one can consistently reason about it and adopt the best strategy available to create a new consistent viewpoint. 7 Implementation of viewpoints The implementation of the Omega description logic [3] provided a viewpoints mechanism which emphasised a hierarchical organisation of viewpoints. Viewpoints, as the other descriptions of the Omega system, could be arranged in a lattice, where a viewpoint inherited from another by including all the sentences belonging to it. Thereafter a viewpoint would inherit all the logical consequences from its ancestor viewpoints. This proved useful for instance to create viewpoints describing a basic theory (e.g. natural deduction), from which more specific viewpoints could be created by adding new statements. A new implementation of the theory of viewpoints is in progress using the Coq Proof Assistant which supports the development of higher order logics [11]. The logical language used by Coq is a variety of type theory, called the Calculus of Inductive Constructions. The -calculus notation can be used as a uniform encoding for expressions, assertions and proofs. Type checking rules enforce well-formedness conditions. According to the Curry-Howard isomor- phism, assertions are represented as types and proofs are represented as terms whose type is the formula they prove. Proving a formula is therefore seen as exhibiting a term of a given type (proving constructively that the type is populated). For defining a new logic system one has to provide: 1. a definition of wff's: by defining a signature for terms and formulas; 2. a definition of the logical axioms: asserting the existence of constants, whose type is the formula corresponding to the axiom schema; 3. a definition of the inference rules: asserting the existence of constants, of functional type, mapping the premises into the conclusions; Using the higher order features of the language the naming device used for reifying sentences at the metalevel and for the formulation of the reflection rules is greatly simplified: sentences are represented as terms at any level. An important task will be to prove that the presentation of the logic (by means of a specific signature) is adequate in the sense that the encoding is a compositional bijection between the syntactic entities (terms, formulas, proofs) of the logical system and certain valid -terms in the signature. In particular we will have to prove that the higher order does not introduce any additional theorems. 8 Conclusion We have discussed properties of a theory of viewpoints which extends a classical natural deduction calculus with sentences of the type in(A; vp) and with axioms and inference rules for contextual reasoning. We argued that an appropriate formalisation of the effect of communication acts is to extend the common knowledge of the participants with statements which can only contain terms made out of manifest constants, so that referent sharing is ensured. --R Reasoning about Knowledge and Ignorance Metalanguage and reasoning across viewpoints A Description Oriented Logic for Building Knowledge Bases Reflections about reflection A formalisation of viewpoints Proofs in context Propositional Logic in Context Metamathematics of Contexts Quantificational Logic of Context Agent Reasoning with Belief Contexts: the Approach and a Case Study The Coq Proof Assistant Contexts: A Formalization and Some Applications Truth and paradox Notes on naive semantics Logic: techniques of formal reasoning Outline of a theory of truth Notes on Formalizing Context Context Shifts and Clashes in Dialogues: An Active Logic Perspective Syntactical treatment of modalities Languages with self-reference I: foundations Reasoning Situated in Time I: Basic Concepts On Mental Entities Linguistics and Philosophy Varieties of contexts Viewpoints subsume belief Prolegomena to a theory of mechanized formal reasoning --TR --CTR Richard W. Weyhrauch , Marco Cadoli , Carolyn L. Talcott, Using Abstract Resources to Control Reasoning, Journal of Logic, Language and Information, v.7 n.1, p.77-101, January 1998
metalevel theories;contexts;reflective theories;viewpoints;natural deduction
595925
Off-Line Parsability and the Well-Foundedness of Subsumption.
Typed feature structures are used extensively for the specification of linguistic information in many formalisms. The subsumption relation orders TFSs by their information content. We prove that subsumption of acyclic TFSs is well founded, whereas in the presence of cycles general TFS subsumption is not well founded. We show an application of this result for parsing, where the well-foundedness of subsumption is used to guarantee termination for grammars that are off-line parsable. We define a new version of off-line parsability that is less strict than the existing one; thus termination is guaranteed for parsing with a larger set of grammars.
Introduction Feature structures serve as a means for the specification of linguistic information in current linguistic formalisms such as LFG (Kaplan and Bresnan, 1982), HPSG (Pollard and Sag, 1994) or (some versions of) Categorial Grammar (Haddock, Klein, and Morill, 1987). This paper focuses on typed feature structures (TFSs) that are a generalization of their untyped counterparts. TFSs are related by subsumption (see (Carpenter, 1992b)) according to their Journal of Logic, Language and Information, accepted for publication information content. We show that the subsumption relation is well-founded for acyclic TFSs, but not for cyclic ones. We use this result to prove that parsing is terminating for grammars that are off-line parsable: this proposition is cited, but not proved, in (Shieber, 1992). We also suggest a less strict definition for off-line parsability that guarantees termination in the case of acyclic TFSs. This work has originated out of our interest in the theoretical aspects of parsing with grammars that are based on TFSs (see (Wintner and Francez, 1995b)). While the results presented here are basically theoretical, we have implemented a system for efficient processing of such grammars, based on abstract machine techniques; this work is presented in (Wintner and Francez, 1995a). The rest of this paper is organized as follows: section 2 outlines the theory of TFSs of (Carpenter, 1992b). In section 3 we discuss the well-foundedness of TFS subsumption. We sketch a theory of parsing in section 4 and discuss off-line parsability of TFS-based grammars in section 5. 2 Theory of feature structures This section summarizes some preliminary notions along the lines of (Carpenter, 1992b). While we use the terminology of typed feature structures, all the results are valid for untyped structures, that are a special case of TFSs. (Carpenter, 1992b) defines well-typed and totally well-typed feature structures that are subsets of the set of TFSs; for generality, we assume nothing about the well-typedness of TFSs below. However, the larger context of our work is done in a setup where features are assigned to types through an appropriateness specification, and hence we retain the term typed feature structures rather than sorted ones. For the following discussion we fix non-empty, finite, disjoint sets Types and Feats of types and features, respectively, and an infinite set Nodes of nodes, disjoint of Types and Feats, each member of which is assigned a type from Types through a fixed typing function Nodes ! Types. The set Nodes is 'rich' in the sense that for every t 2 Types, the set Below, the meta-variable T ranges over subsets of Types, t - over types, f - over features and q - over nodes. For a partial function F , 'F (x)#' means that F is defined for the value x. Whenever the result of an application of a partial function is used as an operand, it is meant that the function is defined for its arguments. IN denotes the set of natural numbers. A partial order v over Types is a type hierarchy if it is bounded complete, i.e., if every up-bounded subset T of Types has a (unique) least upper bound, tT . If t v t 0 , t is said to be more general than t 0 , which is more specific than t. t 0 is a subtype of t. is the most general tTypes is the most specific, inconsistent type. All occurrences of ? are identified. A feature structure (over the parameters Nodes, Types and Feats) is a directed, connected, labeled graph consisting of a finite, nonempty set of nodes Q ' Nodes, a root Q, and a partial function specifying the arcs, such that every node accessible from - q. A; B (with or without subscripts) range over feature structures and Q; - (with the same subscripts) over their constituents. 1 Let TFSs be the set of all typed feature structures (over the fixed parameters Nodes, Types and Feats). A path is a finite sequence of features, and the set Feats is the collection of paths. ffl is the empty path. -; ff (with or without subscripts) range over paths. The definition of ffi is extended to paths in the natural way: ffi(q; -). The paths of a feature structure A are -)#g. Note that for every TFS A, \Pi(A) 6= OE since ffl 2 \Pi(A) for every A. A feature structure q; ffi) is cyclic if there exist a non-empty path ff 2 Paths and a node q 2 Q such that ffi(q; q. It is acyclic otherwise. Let ATFSs be the set of all acyclic TFSs (over the fixed parameters). A feature structure q; ffi) is reentrant iff there exist two different paths Definition 2.1 (Subsumption) A exists a total function subsumption morphism) such that ffl for every q 2 ffl for every q 2 Q 1 and for every f such that Untyped feature structures can be modeled by TFSs: consider a particular type hierarchy in which the set of types is the set of atoms, plus the types complex and ?, ? subsumes every other type, and the rest of the types are incomparable. All features are appropriate for the type complex only, with ? as their appropriate value. Atomic nodes are labeled by an atom, non-atomic nodes - by complex and variables - by ?. The symbol 'v' is overloaded to denote subsumption (in addition to the subtype relation). The morphism h associates with every node in Q 1 a node in Q 2 with at least as specific a moreover, if an arc labeled f connects q with q 0 in A 1 , then such an arc connects h(q) with h(q 0 ) in A 2 . Two properties follow directly from the definition: If A v B then every path defined in A is defined in B, and if two paths are reentrant in A they are reentrant in B. If two feature structures subsume each other then they have exactly the same structure. The only thing that distinguishes between them is the identity of the nodes. This information is usually irrelevant, and thus an isomorphism is defined over TFSs as follows: A and B are alphabetic variants (denoted A - B) iff A v B and B v A. A strictly subsumes B If A strictly subsumes B then one of following cases must hold: either B contains paths that A doesn't; or there is a path in B that ends in a node with a type that is greater than its counterpart in A; or B contains 'more reentrancies': paths that lead to the same node in B lead to different nodes in A. Lemma 2.2 If A ! B (through the subsumption morphism h) then at least one of the following conditions holds: 1. There exists a path - 2 \Pi(B) n \Pi(A) 2. There exists a node q 2 QA such that '(q) ! '(h(q)), 3. There exist paths Well-foundedness In this section we discuss the well-foundedness of TFS subsumption. A partial order - on a set D is well-founded iff there does not exist an infinite decreasing sequence d 0 - d 1 - d of elements of D. We prove that subsumption of acyclic TFSs is well-founded, and show an example of general (cyclic) TFSs for which subsumption is not well-founded. While these results are not surprising, and in fact might be deduced from works such as, e.g., (Moshier and Rounds, 1987) or (Shieber, 1992), they were not, to the best of our knowledge, spelled out explicitly before. Lemma 3.1 A TFS A is acyclic iff \Pi(A) is finite. Proof: If A is cyclic, there exists a node q 2 Q and a non-empty path ff that ffi(q; Since q is accessible, let - be the path from the root to q: ffi( - q. The infinite set of paths contained in \Pi(A). If A is acyclic then for every non-empty path ff 2 Paths and every q 2 Q, ffi(q; ff) 6= q. Q is finite, and so is Feats, so the out-degree of every node is finite. Therefore the number of different paths leaving - q is bounded, and hence \Pi(A) is finite. 2 Definition 3.2 (Rank) Let r : Types ! IN be a total function such that r(t) ! r(t 0 ) if . For an acyclic TFS A, let IN be defined by \Delta(A). By lemma 3.1, rank is well defined for acyclic TFSs. \Delta(A) can be thought of as the 'number of reentrancies' in A: every node q 2 QA contributes to \Delta(A). For every acyclic TFS A, \Delta(A) - 0 (clearly Lemma 3.3 If A ! B and both are acyclic then Proof: Since A v B, \Pi(A) ' \Pi(B). Consider the two possible cases: ffl If by the definitions of \Theta and subsumption; by the definition of subsumption, since every reentrancy in A is a reentrancy in B; - By lemma 2.2, either case (2) holds), or \Delta(A) ! \Delta(B) (if case (3) holds). Hence ffl If \Pi(A) ae \Pi(B) then it might be the case that jQA j ! jQB j. But for every node q 2 QB that is not the image of any node in QA , there exists a path - such that Hence Theorem 3.4 Subsumption of acyclic TFSs is well-founded. Proof: For every acyclic TFS A, rank(B). If an infinite decreasing sequence of acyclic TFSs existed, rank would have mapped them to an infinite decreasing sequence in IN , which is a contradiction. Hence subsumption is well-founded for acyclic TFSs. 2 Theorem 3.5 Subsumption of TFSs is not well-founded. Proof: Consider the infinite sequence of TFSs A depicted graphically in figure 1. For every i - 0, A to see that consider the morphism h i that maps - q i+1 to - (i.e., the first i nodes of A i+1 are mapped to the first i nodes of A i , and the additional node of A i+1 is mapped to the last node of A i ). Clearly, for every i - 0, h i is a subsumption morphism. Hence, for every i - 0; A i w A i+1 . To show strictness, assume a subsumption morphism . By the third requirement of subsumption (h commuting with ffi), the first nodes in A i have to be mapped by h 0 to the first i nodes in A i+1 . However, if q is the 1-th node of A i , then ffi i (q) leads back to q, while ffi i+1 leads to the last node of A i+1 (the cyclic node), and hence h 0 does not commute with ffi, a contradiction. Hence, Thus, there exists a strictly decreasing infinite sequence of cyclic TFSs and therefore subsumption is not well-founded. 2 To conclude this section, note that specification, which is the inverse relation to subsump- tion, is not well-founded even when cyclic feature structures are ruled out. This fact can easily be seen by considering the sequence of feature structures f f f f f f f f f A3 Figure 1: An infinite decreasing sequence of TFSs of nodes, the first i of which are labeled t and the last - ?, and an f-arc leads from every node to its successor. Clearly, and the sequence is infinite. This is true whether or not appropriateness constraints are imposed on the feature structures involved. In the general case, then, given a feature structure A it might be possible to construct, starting from A, both an infinite decreasing sequence of TFSs (by expanding cycles) and an infinite increasing sequence (by adding paths). Parsing Parsing is the process of determining whether a given string belongs to the language defined by a given grammar, and assigning a structure to the permissible strings. A large variety of parsing algorithms exists for various classes of grammars (for a detailed treatment of the theory of parsing with grammars that are based on feature structures, refer to (Shieber, 1992; Sikkel, 1993; Wintner and Francez, 1995b)). We define below a simple algorithm for grammars that are based on TFSs, but it must be emphasized that the results presented in this paper are independent of the particular algorithm; they hold for a wide range of different algorithms. To be able to represent complex linguistic information, such as phrase structure, the notion of feature structures is usually extended. There are two different approaches for representing phrase structure in feature structures: by adding special, designated features to the FSs themselves; or by defining an extended notion of FSs. The first approach is employed by HPSG: special features, such as DTRS (daughters), encode trees in TFSs as lists. This makes it impossible to directly access a particular daughter. (Shieber, 1992) uses a variant of this approach, where a denumerable set of special features, namely 0; are added to encode the order of daughters in a tree. In a typed system such as ours, this method would necessitate the addition of special types as well; in general, no bound can be placed on the number of features and types necessary to state rules (see (Carpenter, 1992b, p. 194)). We adopt below the other approach: a new notion of multi-rooted feature structures, suggested by (Sikkel, 1993) and (Wintner and Francez, 1995b), is being used. Definition 4.1 (Multi-rooted structures) A multi-rooted feature structure (MRS) is a pair h - is a finite, directed, labeled graph consisting of a set Nodes of nodes and a partial function specifying the arcs, and - Q is an ordered (repetition-free) set of distinguished nodes in Q called roots. G is not necessarily connected, but the union of all the nodes reachable from all the roots in - Q is required to yield exactly Q. The length of a MRS is the number of its roots, j - Qj. - denotes the empty MRS (where is cyclic under the same conditions a TFS is. A MRS is reentrant if it contains a node that can be reached either from two different roots or through two different paths. Meta-variables oe; ae range over MRSs, and ffi; Q; - constituents. If Q; Gi is a MRS and - q i is a root in - naturally induces a feature structure A is the set of nodes reachable from - Thus oe can be viewed as an ordered sequence hA necessarily disjoint) feature structures. We use the two views of MRSs interchangeably. Not only can nodes be shared by more than one element of the sequence; paths that start in one root can reach a different root. In particular, cycles can involve more than one root. Still, it is possible to define sub-structures of MRSs by considering only the sub-graph that is accessible from a sub-sequence of the roots. Definition 4.2 (Sub-structure) The sub-structure of induced by the pair hi; ji and denoted oe i:::j , is hA use oe i for oe i:::i . Definition 4.3 (Subsumption of multi-rooted structures) A MRS Q; Gi subsumes a MRS oe (denoted by oe v oe 0 there exists a total ffl for every root - ffl for every q 2 Q, '(q) v '(h(q)) ffl for every q 2 Q and f 2 Feats, if ffi(q; f)# then h(ffi(q; Many of the properties of TFSs are easily adaptable to MRSs. Let q is the i-th root in - -)#g. Then it is easy to show that if oe v ae then every reentrancy in oe is a reentrancy in ae. Moreover, if oe ! ae (strictly) then at least one of the three conditions listed in lemma 2.2 holds. The well-foundedness result of the previous section are easily extended to MRSs as well. and the same rank function of TFSs can be used to show the well-foundedness of (acyclic) MRSs. The reverse direction is immediate: in the presence of cycles, duplicate the example of the previous section k times and an infinite decreasing sequence of MRSs of length k is obtained, for any k ? 0. For a detailed discussion of the properties of MRSs, refer to (Wintner and Francez, 1995b). Rules and grammars are defined over an additional parameter, a fixed, finite set Words of words (in addition to the parameters Nodes, Feats and Types). The lexicon associates with every word w a feature structure Cat(w), its category. 2 The categories are assumed to be disjoint. The input for the parser, therefore, is a sequence of (disjoint) TFSs rather than a string of words. Definition 4.4 (Pre-terminals) Let defined iff case it is the MRS hCat(w j ); Cat(w j+1 PTw (j; 2 For the sake of simplicity, we assume that lexical entries are not ambiguous. In the case of ambiguity, Cat(w) is a set of TFSs. While the definitions become more cumbersome, all the results still obtain. If a single word occurs more than once in the input (that is, w its category is copied (with remaned nodes) more than once in PT . Definition 4.5 (Grammars) A rule is an MRS of length greater than or equal to 1 with a designated (first) element, the head of the rule. The rest of the elements form the rule's body (which may be empty). A grammar G = (R; A s ) is a finite set of rules R and a start A s that is a TFS. Figure 2 depicts an example grammar (we use AVM notation for this rule; tags such as 1 denote reentrancy). The type hierarchy on which the grammar is based is omitted here. Initial phrase s Lexicon: John her loves6 6 6 6 6 6 6 4 word \Theta AGR :4 agr \Theta 3rd \Theta sg sem \Theta john word \Theta acc AGR :4 agr \Theta 3rd \Theta sg sem \Theta she word \Theta AGR :4 agr \Theta 3rd \Theta sg sem \Theta love phrase \Theta s sem sign \Theta nom sem sign \Theta phrase \Theta sem \Gamma!6 6 4 sign \Theta sign \Theta acc sem Figure 2: An example grammar In what follows we define the notion of derivation (or rewriting) with respect to TFS-based grammars. Informally, this relation (denoted ';') is defined over MRSs such that oe ; ae iff ae can be obtained from oe by successive application of grammar rules. The reader is referred to, e.g., (Sikkel, 1993; Shieber, Schabes, and Pereira, 1994; Wintner and Francez, 1995b) for a detailed formulation of this concept for a variety of formalisms. To define derivations we first define immediate derivation. Informally, two MRSs A and are related by immediate derivation if there exists some grammar rule ae that licenses the derivation. ae can license a derivation by some MRS R that it subsumes; the head of R must be identified with some element i in A, and the body of R must be identified with a sub-structure of B, starting from i. The parts of A prior to and following i remain intact in B. Note that R might carry reentrancies from A to B: if a path - 2 leaving the i-th element of A is reentrant with some path - 1 leaving the a-th element, and - 2 , starting from the head of R, is reentrant with - 3 in some element b in R, then - 1 and - 3 are reentrant in B, starting from the elements in B that correspond to a and b, respectively. Definition 4.6 A MRS immediately derives a MRS (denoted A ! B) iff there exist a rule ae 2 R of length n and a MRS R w ae, such that: ffl R's head is identified with some element i of A: R ffl R's body is identified with a sub-structure of B: R ffl The first are identical: A ffl The last k \Gamma i elements of A and B are identical: A The reflexive transitive closure of '!', denoted ' !', is defined as follows: A or if there exists A 0 such that A ! A 0 and A 0 A 00 . Definition 4.7 A MRS A derives a MRS (denoted A ; B) iff there exist MRSs A such that A v A 0 , Immediate derivation is based on the more traditional notion of substituting some symbol which constitutes the head of some rule with the body of the rule, leaving the context intact. However, as our rules are based on TFSs, the context of the "symbol" to be substituted might be affected by the substitution. To this end we require identity, and not only unifiability, of the contexts. MRSs related by derivations should be viewed as being "as specific as needed", i.e., containing all the information that is added by the rule that licenses the derivation. This is also the reason for the weaker conditions on the ';' relation: it allows an MRS A to derive an MRS B if there is a sequence of immediate derivations that starts with a sepcification of A and ends in a specification of B. Figure 3 depicts a derivation of the string "John loves her" with respect to the example grammar. The scope of reentrancy tags should be limited to one MRS, but we use the same tags across different MRSs to emphasize the flow of information during derivation.6 6 6 6 6 6 6 6 6 6 6 6 4 phrase \Theta s agr \Theta 3rd \Theta sg sem \Theta love \Theta john \Theta she (1) sign \Theta nom agr \Theta 3rd \Theta sg sem \Theta john sign \Theta sem \Theta love (2) word \Theta nom agr \Theta 3rd \Theta sg sem \Theta john word \Theta agr \Theta 3rd \Theta sg sem \Theta love word \Theta acc AGR :4 agr \Theta 3rd \Theta sg sem \Theta she John loves her Figure 3: A leftmost derivation Definition 4.8 (Language) The language of a grammar G is n)g. The derivation example of figure 3 shows that the sentence "John loves her" is in the language of the example grammar, since the derivation starts with a TFS that is more specific than the initial symbol and ends in a specification of the lexical entries of the sentences' words. Parsing is a computational process triggered by some input string of words of length n - 0. For the following discussion we fix a particular grammar particular input string w of length n. A state of the computation consists of a set of items. Definition 4.9 (Items) An item is a tuple [i; oe; j; k], where oe is an MRS joej. an item is active if k ! joej, otherwise it is complete. Items is the collection of all items. If [i; oe; j; k] is an item, oe 1:::k is said to span the input from position position j (the parsing invariant below motivates this term). oe and k can be seen as a representation of a dotted rule, or edge: during parsing all generated items are such that oe is (possibly more specific than) some grammar rule. k is a position in oe, indicating the location of the dot. The part of oe prior to the dot was already seen; the part following the dot is still expected. When the entire body of oe is seen, the edge becomes complete. A computation amounts to successively generating items; we assume that item generation is done through a finite set of deterministic operations that create an item on the basis of previously generated (zero or more) items. Also, if an item was generated on the basis of some existing items, those items are not used again by the same operation. This assumption is realized by an important class of parsing algorithms known as chart parsers. A computation is terminating if and when no new items can be generated. A computation is successful if, upon termination, a complete item that spans the entire input and contains the initial symbol was generated: the final state of the computation should contain the item [0; oe; n; 1], where A - A s and n is the input's length. Different algorithms assign different meanings to items, and generate them in various orders (see, e.g., (Shieber, Schabes, and Pereira, 1994; Sikkel, 1993)). To be as general as possible, we only assume that the following invariant holds: Parsing invariant In a computation triggered by w, if an item [i; oe; j; k] is generated then One immediate consequence of the invariant is that for all the items [i; oe; j; k] generated when parsing A parsing algorithm is required to be correct: Correctness A computation triggered by w is successful iff w 2 L(G). Although (Shieber, 1992) uses a different notation than (Wintner and Francez, 1995b), this property is proven by both. 5 Termination of parsing It is well-known (see, e.g., (Pereira and Warren, 1983; Johnson, 1988)) that unification-based grammar formalisms are Turing-equivalent, and therefore the parsing problem is undecidable in the general case. However, for grammars that satisfy a certain restriction, termination of the computation can be guaranteed. We make use of the well-foundedness of subsumption (section to prove that parsing is terminating for off-line parsable grammars. To assure efficient computation and avoid maintenance of redundant items, many parsing algorithms employ a mechanism called subsumption check (see, e.g., (Shieber, 1992; Sikkel, 1993)) to filter out certain generated items. Define a (partial) order over items: . The subsumption filter is realized by preserving an item x only if no item x 0 such that x 0 - x was generated previously. Thus, for all items that span the same substring, only the most general one is maintained. (Shieber, 1992; Wintner and Francez, 1995b) prove that by admitting the subsumption check, the correctness of the computation is preserved. Off-line parsability was introduced by (Kaplan and Bresnan, 1982) and was adopted by (Pereira and Warren, 1983), according to which "A grammar is off-line parsable if its context-free skeleton is not infinitely ambiguous". As (Johnson, 1988) points out, this restriction (which he defines in slightly different terms) "ensures that the number of constituent structures that have a given string as their yield is bounded by a computable function of the length of that string". The problem with this definition is demonstrated by (Haas, 1989): "Not every natural unification grammar has a context-free backbone". A context-free backbone is inherent in LFG, due to the separation of c-structure from f-structure and the explicit demand that the c-structure be context-free. However, this notion is not well-defined in HPSG, where phrase structure is encoded within feature structures (indeed, HPSG itself is not well-defined in the formal language sense). Such a backbone is certainly missing in Categorial Grammar, as there might be infinitely many categories. (Shieber, 1992) generalizes the concept of off-line parsability but doesn't prove that parsing with off-line parsable grammars is terminating. We use an adaptation of his definition below and provide a proof. Definition 5.1 (Finite-range decreasing functions) A total function F : D ! D, where D is a partially-ordered set, is finite-range decreasing (FRD) iff the range of F is finite and for every d 2 D;F (d) - d. Definition 5.2 (Strong off-line parsability) A grammar is strongly off-line parsable iff there exists an FRD-function F from MRSs to MRSs (partially ordered by subsumption) such that for every string w and different MRSs oe; ae such that oe ; ae, if oe ; PTw (i Strong off-line parsability guarantees that any particular sub-string of the input can only be spanned by a finite number of MRSs: if a grammar is strongly off-line parsable, there can not exist an infinite set S of MRSs, such that for some S. If such a set existed, F would have mapped its elements to the set fF This set is infinite since S is infinite and F doesn't map two different items to the same image, and thus the finite range assumption on F is contradicted. As (Shieber, 1992) points out, "there are non-off-line-parsable grammars for which termination holds". We use below a more general notion of this restriction: we require that F produce a different output on oe and ae only if they are incomparable with respect to subsump- tion. We thereby extend the class of grammars for which parsing is guaranteed to terminate (although there still remain decidable grammars for which even the weaker restriction doesn't hold). Definition 5.3 (Weak off-line parsability) A grammar G is weakly off-line parsable iff there exists an FRD-function F from MRSs to MRSs (partially ordered by subsumption) such that for every string w and different MRSs oe; ae such that oe ; ae, if oe ; PTw (i ae and ae 6v oe, then F (oe) 6= F (ae). Clearly, strong off-line parsability implies weak off-line parsability. However, as we show below, the inverse implication does not hold. We now prove that weakly off-line parsable grammars guarantee termination of parsing in the presence of acyclic MRSs. We prove that if these conditions hold, only a finite number of different items can be generated during a computation. The main idea is the following: if an infinite number of different items were generated, then an infinite number of different items must span the same sub-string of the input (since the input is fixed and finite). By the parsing invariant, this would mean that an infinite number of MRSs derive the same sub-string of the input. This, in turn, contradicts the weak off-line parsability constraint. Theorem 5.4 If G is weakly off-line parsable and MRSs are acyclic then every computation terminates. Proof: Fix a computation triggered by w of length n. By the consequence of the parsing invariant, the indices that determine the span of items are limited (0 - n), as are the dot positions (0 ! k - joej). It remains to show that for every selection of i, j and k, only a finite number of MRSs are generated. Let be a generated item. Suppose another item is generated where only the MRS is different: x x 0 will not be preserved because of the subsumption test. If ae v oe, x can be replaced by x 0 . There is only a finite number of such replacements, since subsumption is well-founded for acyclic MRSs. Now suppose oe 6v ae and ae 6v oe. By the parsing invariant, oe 1:::k ; PTw (i+1; and ae 1:::k ; PTw (i the range of F is finite, there are only finitely many items with equal span that are pairwise incomparable. Since only a finite number of items can be generated and the computation uses a finite number of operations, every computation ends within a finite number of steps.The above proof relies on the well-foundedness of subsumption, and indeed termination of parsing is not guaranteed by weak off-line parsability for grammars based on cyclic TFSs. Obviously, cycles can occur during unification even if the unificands are acyclic. However, it is possible (albeit costly, from a practical point of view) to spot them during parsing. Indeed, many implementations of logic programming languages, as well as of unification-based grammars (e.g., ALE (Carpenter, 1992a)) do not check for cycles. If cyclic TFSs are allowed, the more strict notion of strong off-line parsability is needed. Under the strong condition the above proof is applicable for the case of non-well-founded subsumption as well. To exemplify the difference between strong and weak off-line parsability, consider a grammar G that contains the following single rule:6 4 and the single lexical entry, w 1 , whose category is: This lexical entry can be derived by an infinite number of TFSs: It is easy to see that no FRD-function can distinguish (in pairs) among these TFSs, and hence the grammar is not strongly off-line parsable. The grammar is, however, weakly off-line parsable: since the TFSs that derive each lexical entry form a subsumption chain, the antecedent of the implication in the definition for weak off-line parsability never holds; even trivial functions such as the function that returns the empty TFS for every input are appropriate FRD-functions. Thus parsing is guaranteed to terminate with this grammar. It might be claimed that the example rule is not a part of any grammar for a natural language. It is unclear whether the distinction between weak and strong off-line parsability is relevant when "natural" grammars are concerned. Still, it is important when the formal, mathematical and computational properties of grammars are concerned. We believe that a better understanding of formal properties leads to a better understanding of "natural" grammars as well. Furthermore, what might be seem un-natural today can be common practice in the future. Acknowledgments This work is supported by a grant from the Israeli Ministry of Science: "Programming Languages Induced Computational Linguistics". The work of the second author was also partially supported by the Fund for the Promotion of Research in the Technion. We wish to thank the anonymous referees for their enlightening comments. --R The Logic of Typed Feature Structures. A parsing algorithm for unification grammar. Categorial Grammar Lexical functional grammar: A formal system for grammatical representation. A logic for partially specified data struc- tures Parsing as deduction. Parsing Schemata. An abstract machine for typed feature struc- tures Parsing with typed feature structures. --TR --CTR Efrat Jaeger , Nissim Francez , Shuly Wintner, Unification Grammars and Off-Line Parsability, Journal of Logic, Language and Information, v.14 n.2, p.199-234, March 2005 Shuly Wintner, Compositional semantics for linguistic formalisms, Proceedings of the 37th annual meeting of the Association for Computational Linguistics on Computational Linguistics, p.96-103, June 20-26, 1999, College Park, Maryland
parsing;computational linguistics;feature structures;unification
595932
Group Theory and Computational Linguistics.
There is currently much interest in bringing together the tradition of categorial grammar, and especially the Lambek calculus, with the recent paradigm of linear logic to which it has strong ties. One active research area is designing non-commutative versions of linear logic (Abrusci, 1995; Retor, 1993) which can be sensitive to word order while retaining the hypothetical reasoning capabilities of standard (commutative) linear logic (Dalrymple et al., 1995). Some connections between the Lambek calculus and computations in groups have long been known (van Benthem, 1986) but no serious attempt has been made to base a theory of linguistic processing solely on group structure. This paper presents such a model, and demonstrates the connection between linguistic processing and the classical algebraic notions of non-commutative free group, conjugacy, and group presentations. A grammar in this model, or G-grammar is a collection of lexical expressions which are products of logical forms, phonological forms, and inverses of those. Phrasal descriptions are obtained by forming products of lexical expressions and by cancelling contiguous elements which are inverses of each other. A G-grammar provides a symmetrical specification of the relation between a logical form and a phonological string that is neutral between parsing and generation modes. We show how the G-grammar can be oriented for each of the modes by reformulating the lexical expressions as rewriting rules adapted to parsing or generation, which then have strong decidability properties (inherent reversibility). We give examples showing the value of conjugacy for handling long-distance movement and quantifier scoping both in parsing and generation. The paper argues that by moving from the free monoid over a vocabulary V (standard in formal language theory) to the free group over V, deep affinities between linguistic phenomena and classical algebra come to the surface, and that the consequences of tapping the mathematical connections thus established can be considerable.
Introduction There is currently much interest in bringing together the tradition of categorial grammar, and especially the Lambek calculus [10], with the more recent paradigm of linear logic [8] to which it has strong ties. One active research area concerns the design of non-commutative versions of linear logic [1, 14] which can be sensitive to word order while retaining the hypothetical reasoning capabilities of standard (commutative) linear logic that make it so well-adapted to handling such phenomena as quantifier scoping [4]. Some connections between the Lambek calculus and group structure have long been known [16], and linear logic itself has some aspects strongly reminiscent of groups (the producer/consumer duality of a formula A with its linear negation A ? ), but no serious attempt has been made so far to base a theory of linguistic description solely on group structure. This paper presents such a model, G-grammars (for "group grammars"), and argues that: ffl The standard group-theoretic notion of conjugacy, which is central in G- grammars, is well-suited to a uniform description of commutative and non-commutative aspects of language. ffl The use of conjugacy provides an elegant approach to long-distance dependency and scoping phenomena, both in parsing and in generation. ffl G-grammars give a symmetrical account of the semantics-phonology rela- tion, from which it is possible to extract, via simple group calculations, rewriting systems with strong decidability properties computing this relation for the parsing and generation modes. The paper is organized as follows. In Section 2 we introduce a "group com- putation" model, using standard algebraic tools such as free groups, conjugacy and normal subsets. The main deviation from traditional mathematical practice is in the focus given to the notions of compatible preorder and normal submonoid, whereas those of compatible equivalence relation and normal subgroup are more usual in algebra. Section 3 applies this model to linguistic description, and presents a G-grammar for a fragment of English involving quantification and relative pronouns. Sections 4 and 5 are concerned with generation and parsing, which correspond to two ways of exploiting the relation of preorder associated with the G-grammar, one (generation) in which logical forms are iteratively rewritten as combinations of logical forms and phonological forms until only phonological forms are left, the other (parsing) in which phonological forms are rewritten as combinations of logical forms and inverses of those until, after cancellation of adjacent inverses, exactly one logical form is left. Section 6 introduces the concept of diagrams, which provide an intuitive and powerful geometrical representation for G-grammars. These diagrams originate in the work of Van Kampen [17], and have found applications in combinatorial group theory and in studies of decidable subclasses of the word problem for groups [11]. They have also been used to produce complete term rewriting specifications for certain classes of groups [2]. Section 7 discusses in detail the conditions under which G-grammars and the associated rewriting systems lead to equivalent definitions of the semantics-phonology relation. These results are used for comparing G-grammars, context-free grammars, DCG's and categorial grammars. The section also provides a short discussion of the advantages of G-grammars for describing in a uniform way commutative and non-commutative aspects of language. Section 8 applies certain group morphisms to show some strong computability properties of G-grammars, both for parsing and for generation. Computation monoid M is a set M together with a product M \Theta M !M , written ab, such that: ffl This product is associative; ffl There is an element 1 2 M (the neutral element) with a for all A group is a monoid in which every element a has an inverse a \Gamma1 such that a A preorder on a set is a reflexive and transitive relation on this set. When the relation is also symmetrical, that is, R(x; y) ) R(y; x), then the preorder is called an equivalence relation. When it is antisymmetrical, that is that is, y, it is called a partial order. A preorder R on a group G will be said to be compatible with the group product iff, whenever R(x; y) and R(x 2.1 Normal submonoids of a group. We consider a compatible preorder notated x ! y on a group G. The following properties, for any x; y 2 G, are immediate: G: Two elements in a group G are said to be conjugate if there exists y 2 G such that x . The fourth property above says that the set M of elements G such that x ! 1 is a set which contains along with an element all its conjugates, that is, a normal subset of G. As M is clearly a submonoid of G, it will be called a normal submonoid of G. Conversely, it is easy to show that with any normal submonoid M of G one can associate a preorder compatible with G. Indeed let's define x ! y as xy M . The relation ! is clearly reflexive and transitive, hence is a preorder. It is also compatible with G, for if x are in M ; hence x \Gamma1 is in M , implying that x 1 that is, that the preorder is compatible. Remark. In general M is not a subgroup of G. It is iff x ! y implies y ! x, that is, if the compatible preorder ! is an equivalence relation (and, therefore, a congruence) on G. When this is the case, M is a normal subgroup of G. This notion plays a pivotal role in classical algebra. Its generalization to submonoids of G is basic for the algebraic theory of computation presented here. If S is a subset of G, the intersection of all normal submonoids of G containing S (resp. of all subgroups of G containing S) is a normal submonoid of G (resp. a normal subgroup of G) and is called the normal submonoid closure NM(S) of S in G (resp. the normal subgroup closure NG(S) of S in G). 2.2 The free group over V. We now consider an arbitrary set V , called the vocabulary, and we form the so-called set of atoms on V , which is notated and is obtained by taking both elements v in V and the formal inverses v \Gamma1 of these elements. We now consider the set F (V ) consisting of the empty string, notated 1, and of strings of the form x 1 x 2 :::x n , where x i is an atom on V . It is assumed that such a string is reduced, that is, never contains two consecutive atoms which are inverse of each other: no substring vv \Gamma1 or v \Gamma1 v is allowed to appear in a reduced string. When ff and fi are two reduced strings, their concatenation fffi can be reduced by eliminating all substrings of the form vv \Gamma1 or v \Gamma1 v. It can be proven that the reduced string fl obtained in this way is independent of the order of such eliminations. In this way, a product on F (V ) is defined, and it is easily shown that F (V ) becomes a (non-commutative) group, called the free group 2.3 Group computation We will say that an ordered pair is a group computation structure if: 1. V is a set, called the vocabulary, or the set of generators 2. R is a subset of F (V ), called the lexicon, or the set of relators. 1 1 For readers familiar with group theory, this terminology will evoke the classical notion of group presentation through generators and relators. The main difference with our definition is that, in the classical case, the set of relators is taken to be symmetrical, that is, to contain contains r. When this additional assumption is made, our preorder becomes an equivalence relation. The submonoid closure NM(R) of R in F (V ) is called the result monoid of the group computation structure GCS. The elements of NM(R) will be called computation results, or simply results. If r is a relator, and if ff is an arbitrary element of F (V ), then ffrff \Gamma1 will be called a quasi-relator of the group computation structure. It is easily seen that the set RN of quasi-relators is equal to the normal subset closure of R in F (V ), and that NM(RN ) is equal to NM(R). A computation relative to GCS is a finite sequence quasi-relators. The product r 1 evidently a result, and is called the result of the computation c. It can be shown that the result monoid is entirely covered in this way: each result is the result of some computation. A computation can thus be seen as a "witness", or as a "proof", of the fact that a given element of F (V ) is a result of the computation structure. 2 For specific computation tasks, one focusses on results of a certain sort, for instance results which express a relationship of input-output, where input and output are assumed to belong to certain object types. For example, in computational linguistics, one is often interested in results which express a relationship between a fixed semantic input and a possible textual output (generation mode) or conversely in results which express a relationship between a fixed textual input and a possible semantic output (parsing mode). is a group computation structure, and if A is a given subset of F (V ), then we will call the pair GCSA = (GCS;A) a group computation structure with acceptors. We will say that A is the set of acceptors, or the public interface, of GCSA. A result of GCS which belongs to the public interface will be called a public result of GCSA. G-Grammars We will now show how the formal concepts introduced above can be applied to the problems of grammatical description and computation. We start by introducing a grammar, which we will call a G-Grammar (for "Group Grammar"), for a fragment of English (see Fig. 1). A G-grammar is a group computation structure with acceptors over a vocabulary log [V phon consisting of a set of logical forms V log and a disjoint set of phonological elements (in the example, words) V phon . Examples of phonological elements are john, saw, every, examples of logical forms j, s(j,l), these logical forms can be glossed respectively as "john", "john saw louise" and "for every man x, for some woman y, x saw y". The grammar lexicon, or set of relators, R is given as a list of "lexical schemes". An example is given in Fig. 1. Each line is a lexical scheme and represents a set of relators in F (V ). The first line is a ground scheme, which corresponds to the single relator j john \Gamma1 , and so are the next four lines. The sixth line is a non-ground scheme, which corresponds to an infinite set of relators, 2 The analogy with the view in constructive logics is clear. There what we call a result is called a formula or a type, and what we call a computation is called a proof. l louise man A A Figure 1: A G-grammar for a fragment of English obtained by instanciating the term meta-variable A (notated in uppercase) to a logical form. So are the remaining lines. We use Greek letters for expression meta-variables such as ff, which can be replaced by an arbitrary expression of F (V ); thus, whereas the term meta-variables A, B, ., range over logical forms, the expression meta-variables ff, fi, ., range over products of logical forms and phonological elements (or their inverses) in F (V ). 3 The notation P[x] is employed to express the fact that a logical form containing an argument identifier x is equal to the application of the abstraction P to x. The identifier meta-variable X in P[X] ranges over such identifiers (x, y, z, .), which are notated in lower-case italics (and are always ground). The meta-variable P ranges over logical form abstractions missing one argument (for instance z.s(j,z)). When matching meta-variables in logical forms, we will allow limited use of higher-order unification. For instance, one can match P[X] to s(j,x) by taking The vocabulary and the set of relators that we have just specified define a group computation structure We will now describe a set of acceptors A for this computation structure. We take A to be the set of elements of F (V ) which are products of the following form: where S is a logical form (S stands for "semantics"), and where each W i is a phonological element (W stands for "word"). The expression above is a way of encoding the ordered pair consisting of the logical form S and the phonological string W 1 (that is, the inverse of the product Wn A public result SWn in the group computation structure 3 Expression meta-variables are employed in the grammar for forming the set of conjugates ff exp ff \Gamma1 of certain expressions exp (in our example, exp is ev(N,X,P[X]) P[X] \Gamma1 , Conjugacy allows the enclosed material exp to move as a block in expressions of F (V ), see sections 4 and 5. l * louise man w * woman ran t(N) * the N tt(N,X,P[X]) * N that ff Figure 2: Generation-oriented rules with acceptors ((V; R); the G-grammar -will be interpreted as meaning that the logical form S can be expressed as the phonological string W 1 Let us give an example of a public result relative to the grammar of Fig. 1. We consider the relators (instanciations of relator schemes): and the quasi-relators: Then we have: which means that s(j,l) louise is the result of a computation This result is obviously a public one, which means that the logical form s(j,l) can be verbalized as the phonological string john saw louise. Generation Applying directly, as we have just done, the definition of a group computation structure in order to obtain public results can be somewhat unintuitive. It is often easier to use the preorder ! . If, for a; b; c 2 F (V ), abc is a relator, then abc ! 1, and therefore b ! a . Taking this remark into account, it is possible to write the relators of our G-grammar as the "rewriting rules" of Fig. 2; we use the notation * instead of ! to distinguish these rules from the parsing rules which will be introduced in the next section. The rules of Fig. 2 have a systematic structure. The left-hand side of each rule consists of a single logical form, taken from the corresponding relator in the G-grammar; the right-hand side is obtained by "moving" all the remaining elements in the relator to the right of the arrow. Because the rules of Fig. 2 privilege the rewriting of a logical form into an expression of F (V ), they are called generation-oriented rules associated with the G-grammar. Using these rules, and the fact that the preorder * is compatible with the product of F (V ), the fact that s(j,l) louise can be obtained in a simpler way than previously. We have: l * louise by the seventh, first and second rules (properly instanciated), and therefore, by transitivity and compatibility of the preorder: * john saw l * john saw louise which proves that s(j,l) * john saw louise, which is equivalent to saying that s(j,l) louise result. Some other generation examples are given in Fig. 3. The first example is straightforward and works similarly to the one we have just seen: from the logical form i(s(j,l),p) one can derive the phonological string john saw louise in paris. 4.1 Long-distance movement and quantifiers The second and third examples are parallel to each other and show the derivation of the same string every man saw some woman from two different logical forms. The penultimate and last steps of each example are the most interesting. In the penultimate step of the second example, fi is instanciated to saw has the effect of "moving" as a whole the expression some woman y \Gamma1 to the position just before y, and therefore to allow for the cancellation of y \Gamma1 and y. The net effect is thus to "replace" the identifier y by the string some woman; in the last step ff is instanciated to the neutral element 1, which has the effect of replacing x by every man. In the penultimate step of the third example, ff is instanciated to the neutral element, which has the effect of replacing x by every man; then fi is instanciated to saw man which has the effect of replacing y by some woman. Remark. In all cases in which an expression similar to ff a (with the a i arbitrary vocabulary elements), it is easily seen that, by giving * s(j,l) in p * john saw l in p * john saw louise in p * john saw louise in paris woman (by taking * every man saw some woman (by taking ff \Gamma1 every man x man saw y (by taking * every man saw some woman (by taking man * t(tt(m,x,s(l,x))) ran * the tt(m,x,s(l,x)) ran * the m that ff * the man that ff * the man that ff * the man that ff * the man that louise saw ran (by taking Figure 3: Generation examples ff an appropriate value in F (V ), the a move arbitrarily to the left or to the right, but only together in solidarity; they can also freely permute cycli- cally, that is, by giving an appropriate value to ff, the expression ff a ff can take on the value a k a (other permutations are in general not possible). The values given to the ff, fi, etc., in the examples of this paper can be understood intuitively in terms of these two properties. We see that, by this mechanism of concerted movement, quantified noun phrases can move to whatever place is assigned to them after the expansion of their "scope" predicate, a place which was unpredictable at the time of the expansion of the quantified logical form. The identifiers act as "target markers" for the quantified noun phrase: the only way to "get rid" of an identifier x is by moving x \Gamma1 , and therefore with it the corresponding quantified noun phrase, to a place where it can cancel with x . The fourth example exploits a similar mechanism for handling relative clauses. At the time the relative pronoun is produced, an identifier inverse x \Gamma1 is also produced which has the capability of moving to whatever position is finally assigned to the relative verb's argument x . 4.2 Word movement and group morphisms The derivations of Fig. 3 show possible rewritings of the four given logical forms into phonological strings. It is natural to ask whether these rewritings are the only ones possible. That is in fact the case, but for now we will confine ourselves to showing that the expression: can only be rewritten into the phonological string: every man saw some woman ; whatever values one may choose for the expression meta-variables ff and fi. 4 We start by considering a variant of the original vocabulary V , namely the vocabulary are new letters not appearing in . We then consider the application OE reduced expression g of F (V ) into the expression of F (V 0 ) obtained by replacing each x by a \Gamma1 every man (resp. man \Gamma1 every \Gamma1 a), and similarly by replacing each y (resp. y \Gamma1 ) in g by b \Gamma1 some woman (resp. woman The application OE is clearly a group morphism from F (V ) to F (V 0 ); in fact it is an isomorphism, where OE \Gamma1 maps a (resp. b) into every man x some woman y We then consider the vocabulary V and the application obtained by mapping, in reduced expressions of F (V 0 ), 4 Of course, for certain values of ff; fi, the first expression does not rewrite into a phonological string at all. OE oe Figure 4: Group morphisms between free groups over different vocabularies. The group F (V ) is isomorphic to F (V 0 ), which in turns projects onto F (V 00 ). The "phonological" subgroup F (V phon ) is kept invariant in the three morphisms. each a (resp. a 1. This is again clearly a group morphism, and we can consider the composition of morphisms Fig. 4). This morphism oe has the following properties: ffl It maps all phonological strings into themselves. ffl It maps x (resp. y) into every man (resp. some woman). ffl It maps into every man saw b some woman ); that is, into every man saw some woman : Because phonological strings are kept invariant in the mapping oe, whenever one gives values to ff; fi such that fl is a phonological string, then, for these values, but we have just seen that, for any values of ff; fi one has man saw some woman . This establishes the initial claim. This simple proof illustrates clearly the power of group operations. By changing the presentation of the group F (V ) into an isomorphic one in which x; y have been reexpressed in terms of a; b and of phonological elements, one obtains effects which can be interpreted as constituent movements which furthermore are mandatory if the output is constrained to be a string containing only phonological elements. Parsing To the compatible preorder ! on F (V ) there corresponds a "reverse" compatible preorder defined as a equivalently, a . The normal submonoid M 0 in F (V ) associated with + is the inverse monoid of the normal submonoid M associated with ! , that is, M 0 contains a iff M contains a It is then clear that one can present the relations: A etc. in the equivalent way: ran some etc. louise man woman ran in the every some that Figure 5: Parsing-oriented rules Suppose now that we move to the right of the + arrow all elements appearing on the left of it, but for the single phonological element of each relator. We obtain the rules of Fig. 5, which we call the "parsing-oriented" rules associated with the G-grammar. By the same reasoning as in the generation case, it is easy to show that any derivation using these rules and leading to the relation PS +LF , where PS is a phonological string and LF a logical form, corresponds to a public result in the G-grammar. A few parsing examples are given in Fig. 6; they are the converses of the generation examples given earlier. In the first example, we first rewrite each of the phonological elements into the expression appearing on the right-hand side of the rules (and where the meta-variables have been renamed in the standard way to avoid name clashes). The rewriting has taken place in parallel, which is of course permitted (we could have obtained the same result by rewriting the words one by one). We then perform certain unifications: A is unified with j, C with p; then B is unified to l. Finally 5 Another possibility at this point would be to unify l with E rather than with B. This would lead to the construction of the logical form i(l,p), and, after unification of E with john saw louise in paris every man saw some woman and then either: the man that louise saw ran Figure Parsing examples E is unified with s(j,l), and we obtain the logical form i(s(j,l),p). In this last step, it might seem feasible to unify E to i(E,p) instead, but that is in fact forbidden for it would mean that the logical form i(E,p) is not a finite tree, as we do require. This condition prevents "self-cancellation" of a logical form with a logical form that it strictly contains. 5.1 Quantifier scoping In the second example, we start by unifying m with N and w with M; then we next to s(A,B) by taking again we "move" next to s(A,B) by taking unified with A and y with B. This leads to the expression: where we now have a choice. We can either unify s(x,y) with Q[y], or with P[x]. In the first case, we continue by now unifying P[x] with sm(w,y,s(x,y)), leading to the output ev(m,x,sm(w,y,s(x,y))). In the second case, we continue by now unifying Q[y] with ev(m,x,s(x,y)), leading to the output sm(w,y,ev(m,x,s(x,y)). The two possible quantifier scopings for the input string are thus obtained, each corresponding to a certain order of performing the unifications. In the last example, the most interesting step is the one (third step) in which ff is instanciated to s(l,B) \Gamma1 , which has the effect of "moving" x close to the "missing" argument B of "louise saw", to cancel it by unification with B and consequently to fill the second argument position in the logical form headed by s. After this step, P[x] is ready to be unified with s(l,x), finally leading to the expected logical form output for the sentence. 6 Diagrams 6.1 Definition Definition. A diagram over the vocabulary V is a finite graph which is (1) planar, that is, embedded in the plane in such a way that the edges do not cross; (2) connected; (3) directed, that is, the edges carry an orientation; (4) labelled, that is, each edge carries a label taken in V . A diagram separates the plane in n+1 connected open sets: the exterior (set of points that can be connected to a point at infinity without crossing an edge), that logical form, would conduct to the output s(j,i(l,p)). If one wants to prevent this output, several approaches are possible. The first one consists in typing the logical form with syntactic categories. The second one is to have some notion of logical-form well-formedness (or perhaps interpretability) disallowing the logical forms i(l,p) [louise in paris] or i(t(w),p) [(the woman) in paris], although it might allow the form t(i(w,p)) [the (woman in paris)]. 6 We have assumed that the meta-variables corresponding to identifiers in P and Q have been instanciated to arbitrary, but different, values x and y. We leave a discussion of this point to a future paper. and n open internal regions each consisting of points which can be connected without crossing an edge, but which are separated from the exterior. An internal region will be called a cell. An example of a diagram over the vocabulary fa; b; cg is given in Fig. 7.000000001111111111 c a c c O c a c a a Figure 7: A diagram. The boundary of a cell is the set of edges which constitute its topological boundary. The boundary of a diagram is the set of edges which are such that all their points are connected to the exterior. If one choses an arbitrary vertex (such as O in the figure) on the boundary of a diagram, and if one moves on the boundary in a conventional clockwise fashion (an orientation that we adopt throughout in the sequel), then one collects a list of edges which are either directed in the same way as the movement, or contrary to it. By producing a sequence of atoms, positive in the first case, negative in the second one, one can then construct a word of F (V ); this word is said to be a boundary word of the diagram. 6.2 Reduced diagrams We will say that a diagram is reduced if there does not exist a pair of edges with a common vertex O, with the same label, oriented oppositely relative to O (that is, both edges point towards O or both point from O), and such that at least one of the two "angles" formed by the two edges is "free", that is, does not "contain" another diagram edge (see Fig. 8). 6.3 Cyclically reduced words word w on V [ V \Gamma1 is said to be cyclically reduced iff every cyclic permutation of it is reduced. It is easy to see that (1) a reduced word is cyclically reduced iff it is not of the form aw 0 a \Gamma1 with a an atom (positive or negative), (2) the conjugate class of any word contains cyclically reduced words, which are cyclic permutations of each other. a c c O c a a Figure 8: A reduced diagram. The diagram of the previous figure was not reduced because of the two c edges outgoing from vertex O. 6.4 Relator cells Consider a group computation structure loss of gener- ality, it can be assumed that the relators in R are cyclically reduced, because the result monoid is invariant when one considers a new set of relators consisting of conjugates of the original ones. From now on, unless especially stated otherwise, this assumption will be made for all relators considered. Take any such cyclically reduced relator construct a labelled cell in the following way: take a circle and divide it in n arcs; label the clockwise-ith arc x i and orient it clockwise if e anti-clockwise otherwise. The labelled cell thus obtained is call the relator cell associated with r. Rather than presenting the GCS through a set of relator words as we have done before, it is now possible to present it through a set of relator cells; if one gives such a set, a standard presentation of the GCS can be derived by taking an arbitrary origin on each cell and "reading" the relator word clockwise from this origin; the origin chosen does not matter: any other origin leads to a conjugate relator, and this does not affect the notion of result. 6.5 Fundamental theorem of combinatorial group theory We are now able to state what J. Rotman calls "the fundamental theorem of combinatorial group theory" [15]. We give the theorem in a slightly extended form, adapted to the case of the normal sub-monoid closure; the standard case of normal subgroup closure follows immediately by taking a set or relators containing along with r. Theorem 1 Let be a group computation structure such that all relators r 2 R are cyclically reduced. If w is a cyclically reduced word in F (V ), then w 2 NM(R) if and only if there exists a reduced diagram having boundary word w and whose regions are relator cells associated with the elements of R. O Figure 9: Star diagram. The proof is not provided; it can easily be recovered from the property demonstrated in [11] (chapter 5, section 1). The proof involves the following remark. If one considers a product arbitrary elements of F (V ), this product can be read as the boundary word of the "star" diagram represented in Fig. 9, starting at O and progressing clockwise. This star diagram is in general not in reduced form, but it can be reduced by a stepwise process of equating edges which do not respect the definition of a reduced diagram. Example. Let's consider a GCS with vocabulary set of (cyclically reduced) relators The cyclically reduced word c is an element of NM(R), for it can be obtained by forming the product If we form the star diagram for this product, we obtain the first diagram shown in Fig. 10. This diagram is not reduced, for instance the two straight edges labelled c are offending the reduction condition. If one "stitches" these two edges together, one obtains the second diagram in the figure. This stitching corresponds to a one-step reduction of the boundary word of the first diagram, into the boundary word of the second By continuing in this way, one obtains the fifth diagram of the figure, which is reduced, and whose boundary is the wanted 6.6 Linguistic examples The previous considerations can be applied to the linguistic examples given in the body of the article. Here we only consider examples which do not involve a c c O c a a a c c c c a O a a c c c a a c a c c O a c c a c c O a a c a c c O c a c a a Figure 10: Transformation of a diagram into reduced form (adapted from [11]). long-distance dependencies; the more complex examples will be treated after we have introduced multi-relators in 7.6. l p saw in paris louise john (1) Figure 11: Cells associated with some grammar relators. saw john louise in paris l Figure 12: A diagram establishing the relationship between a logical form and a phonological string. Let's consider the relator schemes, from Fig. 1: l louise A In Fig. 11, cells (which we have numbered from 1 to 5) associated with some instanciations of these schemes are shown. In Fig. 12 we construct a reduced diagram whose cells are the relator cells of Fig. 11. The boundary word of this diagram is the expression i(s(j,l),p) paris john \Gamma1 . This proves that, in the G-grammar, the logical form i(s(j,l),p) is associated with the phonological string john saw louise in paris. Note (1) the analogy between reading the diagram from top to bottom (resp. bottom to top) and a generation (resp. parsing) process. 7 7 G-grammars and rewriting In the discussion of parsing and generation, we saw how a derivation according to the rewriting rules of Figs. 2 and 5 is always "sound" with respect to the group computation structure. We did not consider the opposite question, namely whether it is "complete" with respect to it: can any public result relative to the GCS be obtained by such rewritings? The answer to this question will be given by theorems demonstrated below (theorems 4 and 7), which roughly state that such a rewriting system is complete relative to the GCS if the system does not contain "ground cycles", that is situations where a ground term T can derive This condition is true of both the rewriting systems of Figs 2 and 5. For instance, in the generation case, it can be checked that any ground logical form that appears on the right-hand side of a rule of Fig. 2 is strictly smaller than the ground logical form on the left-hand side, therefore precluding ground cycles. This condition is related to the decidability of generation and parsing, and we will prove in 8 that our G-grammar is finitely enumerable both in parsing and in generation, that is, in the terminology of [6], that it is inherently reversible. This property is difficult to guarantee in formalisms relying on empty categories for long-distance dependencies, a problem which is avoided by the use group structure for the same purpose. 7.1 Oriented relators and rewriting rules Suppose an is a relator. This relator is said to be oriented at index if a number i with 1 i n has been chosen. A QCF rule (QCF stands for "quasi-context-free") on F (V ) is a pair, notated a 7! ff, with a is a group computation structure, and if r is a relator in R such that r = a an in F (V ), then the rule a i 7! a an is said to be the rule associated with the relator r oriented at index i. 7 There is also an analogy between a bottom-up reading and a chart parsing approach in the case of this example. In more complex examples (for instance multi-word expression parsing) a better analogy would be with Colmerauer's Q-systems [3]. However, both analogies break down when considering examples involving long-distance dependencies. If for each element in a set of relators R, an orientation is chosen, one says that an orientation has been given to the GCS. By associating a QCF rule to each relator, one obtains the QCF system associated with the orientation. A derivation relative to a QCF system is a finite sequence of elements (that is, of strings formed over elements of V or their inverses) such that, for each i, there exists ff; fi 1. and a 7! b is a rule of the rewriting system [replace- ment step], 2. or s i If for two strings s and t in (V one can find a derivation are identical, then we will say that s derives into t and we will write s 7! t (using the same notation as for rewrite rules; the context will make clear which case is intended). The following proposition, stating that rewriting implies preorder, is straightforward. Theorem 2 If 7! is the derivation relation associated with an oriented GCS, then, for any s; t 2 Proof. Immediate consequence of the fact that if r = a an is a relator, then and therefore an ut 7.2 Anteriority Let's consider the QCF system associated with an oriented GCS. If a and b are elements there exists a rule in the system of the form: a 7! ffbfi; then we will say that a is immediately anterior to b relative to the system (or, equivalently, relative to the oriented GCS). If there exists a finite sequence b such that c i is immediately anterior to c i+1 then we will say that a is anterior to b, which will be noted: a b: We will say that the anteriority relation is acyclic iff it is never the case that a a. We will say that it is noetherian iff any "descending" chain a 1 a noetherian it is a fortiori acyclic. If a rule in the QCF system has left-hand side a, then a will be called the mother of the rule (or, equivalently, of the oriented relator corresponding to the rule). The set M of all rule mothers is a subset of the set of atoms called the set of potential mothers associated with the QCF system. We will note M \Gamma1 the set of atoms which are inverses of potential mothers. 7.3 Rewriting theorem (basic case) We are now ready to state a theorem which provides conditions under which the reciprocal of theorem 2 holds. We will only sketch proofs. We first need a lemma. Lemma 3 Suppose that \Delta is a diagram relative to an oriented acyclic GCS, and also that the set of potential mothers does not simultaneously contain an element and its inverse. Then there is a relator cell in this diagram such that its mother m is on the boundary of the diagram. Proof. Take any cell \Gamma in the diagram, and consider the mother m in this cell. If m is on the boundary of the diagram, we are done. Otherwise, m is on an edge separating \Gamma from another cell \Gamma 0 , and therefore m \Gamma1 is on the boundary of \Gamma 0 (as a consequence of the clockwise orientations of the cells). Because the condition about potential mothers, m \Gamma1 cannot be the mother of \Gamma 0 , and therefore the mother m 0 of \Gamma 0 is anterior to m. If m 0 is on the diagram boundary, we are done, otherwise we repeat the operation. Because the anteriority relation is acyclic, the iteration must stop at some point, which proves the lemma. ut We can now state the theorem. Theorem 4 If ' 7! ' is the derivation relation associated with an oriented GCS where (i) anteriority is acyclic, and (2) the set of potential mothers does not contain simultaneously an element and its inverse, then, for any s 2 and for any t 2 ((V Proof. Suppose that s ! t, then there exists a diagram \Delta having k cells whose boundary is st \Gamma1 . We will prove the result by induction on k. In the base case the result is immediate, because in a diagram with 0 cells of boundary st one necessarily has t. Let's now address the case of k ? 0. From the lemma, there is a cell \Gamma in this diagram whose mother m lies on the boundary of \Delta. Because of the condition on t, this m cannot lie in the t portion of the boundary, and so it must lie in the s portion (see Fig. 13). The boundary of \Gamma can be written in the form mw \Gamma1 for some word w 2 therefore s can be expressed in the form and s 2 subparts of s (see Fig. 14). But now one has: by definition of the rewriting relation. Furthermore the word is on the boundary of a diagram in cells, obtained from \Delta by removing the we will not explain here in detail what it means to "remove a cell", and will not justify formally why the remaining construct is still a diagram, but the way the operation works may be understood on the basis of an example (see Fig. 15). We now have: and, by the induction hypothesis, one has and therefore ut A s G Figure 13: The diagram \Delta in k cells, whose boundary is st \Gamma1 . The word s can be read clockwise from vertex A to vertex D, the word t counter-clockwise from A to D. There exists a cell \Gamma whose mother m lies on the s portion of the boundary. 7.4 G-grammars compared to CFG's, DCG's and Categorial Grammars Rather than starting from a G-grammar and associating rewriting systems to it, one can start from a conventional grammar represented as a rewriting system A Figure 14: Removing the cell \Gamma by rewriting m into w. and associate to it a G-grammar. An obvious question is then whether the input-output relationship as defined in the two models agree. Let's for instance consider the context-free grammar presented in Fig. 16, (a). The rewriting system (a) is a QCF system associated with an orientation of the G-grammar given in (a'), over the vocabulary U is the set of nonterminals fs; np; vpg and W is the set of terminals john ; walked; often. By Theorem 2, if we can derive a string of words wn from s, then wn relative to the G-grammar. However, the system (a) is not acyclic for anteriority, because of its fourth rule. The translation of this rule is the relator vp vp often \Gamma1 which is equal to often \Gamma1 . This means that often often. This in turn, because of the compatibility of the preorder, implies that, whenever one has then one has s ! ff often fi, for any ff; fi 2 F (V ). In particular one has string that cannot be derived from grammar (a). More generally, it can be shown that, relative to the GCS, the word 'often' can be added freely in any string generated from s! One way of making the grammar (a) acyclic is to upgrade it to a Definite Clause Grammar [13] such as (b). With each nonterminal is associated a term representing the syntactic structure of the string spanned by this nontermi- nal. The crucial point here is that the structure of the mother nonterminal (for instance vp(.(often,VP))) strictly contains the structures of the daughter non-terminals (for instance vp(VP)). It is then easy to show that the system is acyclic. Then Theorem 4 holds, and the notions of derivations relative to (b) and (b') coincide. In the example, the relator vp(:(often; VP))vp(VP) does not collapse to often \Gamma1 anymore, for vp(:(often; VP)) can never cancel out. From theorem 4, we now see that (b) and (b') are equivalent. Another way to obtain an acyclic grammar is shown in (c). Here we have an infinite set of rules indexed by integers corresponding to string length. The length of a mother nonterminal is strictly larger than the length of any daughter, saw john louise in paris l saw john louise in paris l Figure 15: A cell-removal step. If one assumes that s(j,l) in p 7! john saw louise in paris, then one has i(s(j,l),p) 7! s(j,l) in p 7! john saw louise in paris. and this again ensures acyclicity. 7.4.1 G-grammars and categorial grammars The context-free grammar (a) and its associated G-grammar (a') can be used to illustrate a crucial difference between G-grammars and categorial grammars. Whereas, in categorial grammar, the expression vp=vp "expects" a vp on the right - and nothing else - and then "returns" a vp, when working with groups, an expression such as vp vp \Gamma1 is formally undistinguishable from 1. It is this crucial property that permits G-grammars to profit from many standard mathematical tools. (a) s 7! np vp (a') s vp vp 7! walked vp walked \Gamma1 vp 7! often vp vp vp (b) vp(walked) 7! walked vp(walked) walked \Gamma1 (c) s(x+y) 7! np(x) vp(y) (c') s(x+y) vp(x) walked walked Figure G-grammars. 7.5 Mixing commutative and non-commutative phenom- ena, logic programs We have already seen examples where commutative and non-commutative aspects are both present in a lexical entry. Thus, the presence of ff; ff \Gamma1 in the entry for 'every' in Fig. 2 allows the expression every N X to move as a block to the position where the argument X of the verb is eventually found; in this movement, it is however impossible for every and N to exchange their relative positions, and it can be shown that, for the input logical forms of Fig. 3, only the four phonological strings listed can be obtained. Let us briefly indicate how this commutative/non-commutative partnership could be used for error correction purposes. Suppose that a relator of the Report is added to the grammar, where Error is some erroneous input (for instance it could be the word 'principle' improperly used in a situation where 'principal' is needed), Repair is what the input should have been (in our example, the word 'principal'), and Report is a report which tells us how the error was corrected. Then, the expression Error \Gamma1 Repair can move in block to the spot where the error occurs, replacing the erroneous word by the correction and allowing normal processing to continue. The report can then be used for warning or evaluation purposes. Fully commutative structures and logic programs Suppose that one wants to have a group computation structure which is completely commutative, that is, one in which elements can move freely relative to each other. This property could be stipulated by introducing a notion of "commutative group compu- tation" using the free commutative group FC(V ) rather than F (V ). Another possibility, which illustrates the flexibility of the group computation approach, is just to add a relator scheme fffiff to R, where ff and fi are expression meta-variables. This expression is called a commutator of ff and fi because when multiplied by fiff it yields fffi. This single relator scheme permits to permute elements in any expression, and has the same effect as using FC(V ). An example where commutative structures are useful is the case of logic programs. It can be shown that if one encodes a clause of a logic program in a commutative structure, and defines public results to consist of a single ground predicate P , then public results in the group computation structure coincide with consequences of the program. 8 7.6 Multi-relators Theorem 4 already gives interesting equivalence for several types of GCS's, such as the ones associated with acyclic DCGs, or for a G-grammar which does not make use of expression meta-variables, such as the G-grammar obtained by eliminating the entries for every, some and that from the G-grammar of Fig. 1. However, it is not sufficient for the long-dependency situations such as these where expression meta-variables do appear; in these cases, the conditions of Theorem 4 are not met, because by giving a well-chosen value to the expression meta-variable, the condition of anteriority can be easily violated. We need a stronger version of the theorem which is able to take into account the situation where expression metavariables appear in pairs ff; ff \Gamma1 . In order to state this extended version of Theorem 4, we need to introduce some terminology. Consider a finite list wn of elements of (V wn ) be the set: The following lemma is straightforward. Lemma 5 . The set MC(w wn ) has the following properties: ffl It is a normal subset of F (V ) (that is, is self-conjugate); ffl It is invariant up to arbitrary permutations of the list ffl It is invariant up to cyclic permutations of any of the words w be the set: wn ) is the set of conjugates of elements of B in F (V ). 8 This result, in contrast to the corresponding result concerning DCG's (see below) does not depend on a condition of acyclicity. This difference is related to the fact that, in the case of logic programs, a succesful sequence of rewritings results in the empty string, whereas in the case of DCGs, it results into a string of words. The acyclicity condition for DCGs ensures that this string of words contains all words that could be produced using the GCS. For logic programs, this property is irrelevant. l louise man ran Figure 17: Multi-relator version of our example G-grammar. We have made use of the transformations permitted by Lemma 5. Because of the second of these properties the set MC(w wn ) is defined as soon as the multiset (unordered list) hw wn i, which we will call a multi- relator, is given. We will call MC(w wn ) the set of multi-conjugates of the multi-relator hw wn i. If we are given a (finite or infinite) collection MR of multi-relators, we can consider the group computation structure obtained by taking as set R of relators the set We will call the GCS thus obtained the group computation structure with multi-relators When presenting a group computation structure with multi-relators, we will sometimes use the notation: wn for listing the multi-relators. With such a notation, the presentation of our example G-grammar of Fig. 1 becomes the presentation shown in Fig. 17. The presentation of the grammar using multi-relators is more symmetrical than that of Fig. 1. Remark that expression meta-variables are not used anymore 7.6.1 Multi-relators and diagrams We will now state a theorem which is an extension of the the fundamental theorem of combinatorial group theory for the case of GCS's with multi-relators. We first need the notion of multi-cell associated with a multi-relator. We first remark that we can always assume that a multi-relator wn i is such that each w i is cyclically reduced, because taking the cyclically reduced conjugate of w i does not change the notion of multi-conjugate. We will assume this is always the case in the sequel. wn i is a multi-relator, and if \Gamma i is the cell associated with w i in the manner of 6.4, then we will call the multiset of cells ng the multi-cell associated with this multi-relator. Consider a finite multiset of multi-relators and take the multi-set obtained by forming the multiset U k \Theta k of the multi-cells associated with these multi-relators. A diagram whose cells are exactly (that is, taking account of the cell counts) those of the multiset\Omega will be said to be a diagram relative to the GCS with multi-relators under consideration. Theorem 6 . Let be a group computation structure with multi-relators. If w is a cyclically reduced word in F (V ), then w is in the result monoid of GCS-MR iff there exists a reduced diagram relative to GCS-MR having boundary w. The proof of this theorem is a straightforward extension of the proof of Theorem 1 and is not provided. 7.6.2 Linguistic examples These results can now be applied to a diagrammatic proof of the fact that, in our example G-grammar, one has (see 4.1): every man saw some woman: In Fig. 18 we show multi-cells corresponding to the G-grammar entries relevant for this example (see Fig. 17), and in Fig. 19 a diagrammatic proof of the example; the multi-cells are lettered (1) to (5) for easy comparison between the figures. 7.6.3 Multi-relators and QCF systems One can extend the notion of orientation of 7 to multi-relators in the following way. We will say that a multi-relator hw wn i is oriented at index if one of the words w i has been chosen and if, writing w , with an index j with 1 j m i has also been chosen. The QCF rule scheme associated with the oriented multi-relator is then the rewriting scheme: a ij 7! a i(j \Gamma1) Y where the ff k are expression metavariables (which can therefore take arbitrary values in F (V )). One can extend the notion of anteriority to oriented multi-relators. We simply define the atom a ij to be anterior to each of the atoms a kl \Gamma1 appearing in the man (1) saw x y every m woman (2) y some w Figure 18: Multi-cells for some multi-relators in the grammar. Dotted lines indicate that two cells are part of the same multi-cell: they will appear together in a diagram or not at all. right-hand side of the rewriting rule, that is to say, all inverses of atoms appearing in any of the w k , apart of course from a ij \Gamma1 itself, and not considering the atoms appearing in any of the ff k ; ff k \Gamma1 (it follows that the anteriority relation is defined between atoms or inverses of atoms that appear in hw wn i). The notion of acyclic and noetherian QCF system is defined just as before on the basis of this anteriority relation. The following theorem is a straightforward extension of Theorem 2. Theorem 7 If 7! is the derivation relation associated with an oriented multi- relator GCS, then, for any s; t 2 saw every x man woman y some (1) (2) Figure 19: A diagram using multi-cells establishing a relationship between the sentence every man saw some woman and its logical form ev(m,x,sm(w,y,s(x,y))). 7.6.4 Rewriting theorem (multi-relator case) We now state a theorem which extends Theorem 4 to the case of multi-relators GCS's. The proof is only broadly outlined. Theorem 8 If ' 7! ' is the derivation relation associated with an oriented multi- relator GCS where (i) anteriority is acyclic, and (2) the set of potential mothers does not contain simultaneously an element and its inverse, then, for any s 2 Proof. Rough sketch. The proof is similar to that of theorem 4. The difference is that we must perform multi-cell removal rather than simple cell-removal, and that this may involve more complex "topological surgery". An example may serve to illustrate the general method. Let's consider a GCS defined by the three multi-relators: hfd these multi-relators are oriented by distinguishing f in the first, d in the sec- ond, and a in the third. Then the corresponding rewriting system has the rules: is an expression meta-variable. Anteriority is clearly acyclic. Let us now consider the diagram \Delta at the top-left of Fig. 20. This diagram contains three multi-cells corresponding to the three grammar multi-relators, one of which, \Gamma, is itself formed of two cells, and corresponds to the third multi-relator. The mother in each multi-cell is underlined. By reading the boundary of the diagram, we see that a ! hm. We want to show how multi-cell removal can be used to obtain a 7! hm. First, by a reasoning similar to the proof of Lemma 3, we know that some mother must appear on the boundary. In the example, this is a. We want to remove the corresponding multi-cell \Gamma from the diagram without destroying the topological property of being a diagram, and, in particular, without destroying simple-connectivity. In order to do that, we have to find a "subdiagram" (that, is, a topological constructs which is itself a diagram) \Delta 0 of \Delta consisting only of the cells appearing in the multi-cell under consideration (simple reasoning shows this to always be possible). In our example this subdiagram is constituted by the two cells labelled \Gamma linked by the edge labelled d. The surgery now consists in "forming a tube" wherever a "thin" edge, that is, an edge which is not on the surface of a subdiagram cell, such as the d edge in the example, and in deleting from the diagram both the subdiagram cells and the "tubes" thus formed. What remains is a well-formed diagram - which may well have thin edges itself - in which the removed tubes leave behind couple of parallel edges carrying the same label (second diagram in our figure, consisting of two cells). This multi-cell removal corresponds to a rewriting of the multi-cell mother into a right-hand side, where expression multi-variables have been given proper values (in our example ff takes the value c \Gamma1 d) allowing for the construction of the subdiagram. The operation is repeated until all multi-cells have been deleted, in the same way as in the previous proof. In the example given in illustration, after the first multi-cell, having its mother a on the diagram boundary, has been removed, one obtains the second diagram; at this point a has been rewritten (using the rewriting rule for a) into a 7! bc c c. The new diagram contains two multi-cells, each consisting of a single cell. Again, there exists a mother on the diagram bound- ary, this time f . The corresponding cell is removed, and at this point a has been rewritten (using the rewriting rule for f) into bdeg The last cell, with mother d is finally removed, leaving the last diagram (with only edges, but no cell). At this point a has been finally rewritten (using the rewriting rule for d) into Application to G-grammars When the G-grammar presented as a multi- relator GCS in Fig. 17 is oriented by privileging as a left-hand side the "largest" logical-form atom, its associated QCF system becomes the system of Fig. 2. It can be checked that this "generation" system is acyclic and even noetherian. The theorems 7 and 8 then say that generation by using the rewriting system is equivalent to generation by using the G-grammar. A similar result can be shown for the "parsing-oriented" rewriting system of Fig. 5. f d c e d f a d c e G G e Figure 20: Multi-cell removal and rewriting. Computability We will not attempt here a systematic treatment of the computability properties of G-grammars, but will limit ourselves to prove certain strong decidability properties of our example G-grammar. Although a general synthesis is not proposed here, the methods show clearly the power of group morphisms to study computability in group computation structures. If we consider a grammar to be the specification of a recursively enumerable list L of pairs (sem; phon), then, using the terminology of [6], we will say that generation (resp. parsing) is finitely enumerable if the subset of L consisting of the pairs (sem; phon) with sem known (resp. with phon known) is finite and computable in finite time. If parsing and generation are both finitely enumer- able, the grammar is said to be inherently reversible. We will sketch a proof of the fact that the G-grammar of Fig. 17 is inherently reversible. 9 9 In [6], several types of computability of generation (resp. parsing) are defined and it is shown there that, unless certain specific conditions are met, whatever the type of computability considered, the computability of parsing and the computability of generation bear in general no relation to each other. 8.1 Bounding of diagram complexity in function of the input Let's first consider a group morphism f from F (V ) to the additive group of reals (R;+). Thus f is a real-valued function on F (V ) which is such that for f(a)+f(b). Remark that because f takes its value in a commutative group, f is invariant by permutations (that is, We can therefore define the value of f on a multi-relator be its value on any of its multi-conjugates, that is, to be f(r 1 If a result s is obtained by taking the product of multi-conjugates of the multi- relators can be computed simply by taking the sum of all the multi-relator components r ij , where mr i. Let's now suppose that there exists a strictly positive real ! such that, for any multi-relator mr in the group computation structure, one has: If such is the case, we will then say that f is a relator-lower-bounded or rlb morphism. Similarly, we will say that f is a relator-upper-bounded or rub morphism iff there exists a strictly positive real ! such that, for any multi-relator mr in the group computation structure, one has: The following property is immediate. Lemma 9 If f is an rlb morphism from F (V ) to R, and if s is a result in the GCS, then (1) f(s) 0, and (2) if a computation of s involves k multi-relators, then k f(s)=!. Similary, if f is an rub morphism, then, if a computation of s involves k multi-relators, then f(s) k!. If f is rlb, and if s an expression in F (V ) for which the value of f is known, then we see that in order to test whether s is a result, we only need to test diagrams with an a priori bounded number of multi-cells. Now consider the following two real-valued functions on F (V ffl For e 2 F (V ), e reduced, f(e) is defined as the difference between the number of phonological elements which appear in e with a negative polarity with the number of such elements which appear with a positive polarity. For instance the value of f on john \Gamma1 s(j; m) in the park ffl Suppose that the size of a term T is defined recursively by the following equations: (1) if T is an argument identifier (such as x, y, .) then is of arity 0, then of arity n, of the form being the sum of the sizes of the terms of positive polarity minus the sum of the sizes of the terms of negative polarity. For instance, the value of g on the expression It is immediate that f and g are morphisms. It is also a simple observation that, for any multirelator mr of Fig. 17, the value of both f and g on mr is uniformly lower-bounded by and g are uniformly equal to 1 in our example). For instance if e is any expression of the form 1. It is also clear that both f and g are uniformly upper-bounded Let's now consider parsing. In this case, we are looking at all results of the form sem phon \Gamma1 , with phon known, and therefore with f known (it is just the length of the phonological string). We therefore have only to consider diagrams with fewer than length(phon) multicells and having a boundary of the form sem phon \Gamma1 . If we instead consider generation, we are looking at all results of the form sem phon \Gamma1 , with sem known, and therefore with g known. We therefore have only to consider diagrams with fewer than size(sem) multicells and having a boundary of the form sem phon \Gamma1 . 8.2 Restriction of the set of multi-cells to consider We are now very close to results of finite enumerability for parsing and genera- tion. We know a bound for the total number of cells in any relevant diagram. Still, it could be conceivable that the repertory of relevant cells is not finite. We will now show that this cannot be the case. Let's introduce the notion of subharmonic function on a GCS. 10 Let h be a function from V [ V \Gamma1 to R. We will say that h is subharmonic relative to the GCS if, for any diagram \Delta relative to the CGS, h can only reach its maximum on the boundary of \Delta (therefore not on an "internal" edge in the diagram). Let's define h as the function which takes a phonological element, whatever its polarity, to 0, and which takes a logical form T , as well as its inverse T \Gamma1 , to From the discussion of acyclicity of our G-grammar, and from Lemma 3, we see that h is subharmonic for the G-grammar. This means that any multi-cell appearing in a diagram which is a proof of sem phon \Gamma1 has all its semantics (of whatever polarity) bounded by the size of sem. It is easy to check that there are only finitely many multi-cells according to the G-grammar for which the largest semantic element has a bounded size. 11 10 The terminology is borrowed from analysis, where subharmonic functions are functions which have a property of reaching their maximum only on the boundary of disks. 11 We are finessing the case of the identifier meta-variable X , whose domain of instanciation has not been described precisely. We will assume here that its domain of instanciation is constituted by a finite, fixed, set of identifiers x; This can be assumed both for generation and for parsing on the basis of a simple argument. A more satisfying treatment is left to future research. So, if we are in generation mode, and know sem, we have only a fixed, finite, repertory of multi-cells to consider for diagrams pretending to the status of proofs of a result of the form sem phon \Gamma1 . Because of the fact noted above that these cells can appear conjointly only a finite number of times, we have now proven the finite enumerability of generation. For parsing, we only know phon. But it is immediate from the fact that f is rlb and g rub that sem can be bounded from the fact that phon is known. We conclude that only a finite repertory of cells have to be considered for diagrams pretending to the status of proof of results of the form sem phon \Gamma1 . By the same reasoning as previously, this proves the finite enumerability of parsing. Acknowledgements Thanks to Sylvain Pogodalla, Christian Retor'e and to the anonymous reviewers for their remarks and comments. --R 'Les syst'emes-Q ou un formalisme pour analyser et synth'etiser des phrases sur ordinateur' 'Transformations de grammaires logiques et r'eversibilit'e en Traduction Automatique'. Presentations of Groups. Combinatorial Group Theory. Lambek grammars are context free. 'R'eseaux et s'equents ordonn'es'. An Introduction to the Theory of Groups. Essays in Logical Semantics. --TR
parsing;generation;group theory;diagrams
595959
On Automating Diagrammatic Proofs of Arithmetic Arguments.
Theorems in automated theorem proving are usually proved by formal logical proofs. However, there is a subset of problems which humans can prove by the use of geometric operations on diagrams, so called diagrammatic proofs. Insight is often more clearly perceived in these proofs than in the corresponding algebraic proofs; they capture an intuitive notion of truthfulness that humans find easy to see and understand. We are investigating and automating such diagrammatic reasoning about mathematical theorems. Concrete, rather than general diagrams are used to prove particular concrete instances of the universally quantified theorem. The diagrammatic proof is captured by the use of geometric operations on the diagram. These operations are the inference steps of the proof. An abstracted schematic proof of the universally quantified theorem is induced from these proof instances. The constructive -rule provides the mathematical basis for this step from schematic proofs to theoremhood. In this way we avoid the difficulty of treating a general case in a diagram. One method of confirming that the abstraction of the schematic proof from the proof instances is sound is proving the correctness of schematic proofs in the meta-theory of diagrams. These ideas have been implemented in the system, called Diamond, which is presented here.
Introduction It requires only basic secondary school knowledge of mathematics to realise that the diagram above is a proof of a theorem about the sum of odd naturals. It is an interesting property of diagrams that allows us to "see" and understand so much just by looking at a simple diagram. Not only do Portions of this work were published in the Proceedings of the International Joint Conference on Artificial Intelligence in (Jamnik et al, 1997). Mateja Jamnik et al. we know what theorem the diagram represents, but we also understand the proof of the theorem represented by the diagram and believe it is correct. Is it possible to simulate and formalise this sort of diagrammatic reasoning on machines? Or is it a kind of intuitive reasoning particular to humans that mere machines are incapable of? Roger Penrose claims that it is not possible to automate such diagrammatic proofs. 1 We are taking his position as a challenge and are trying to capture the kind of diagrammatic reasoning that Penrose is talking about so that we will be able to emulate it on a computer. The importance of diagrams in many domains of reasoning has been extensively discussed by Larkin and Simon (Larkin & Simon, 1987), who claim that "a diagram is (sometimes) worth ten thousand words". The advantage of a diagram is that it concisely stores information, explicitly represents the relations among the elements of the diagram, and it supports a lot of perceptual inferences that are very easy for humans. Diagrams have been extensively used in the history of mathematics to aid informal mathematical reasoning. The use of diagrams in explanations of theorems and proofs of geometry dates back to Ancient Greece, and the time of Aristotle and Euclid. Thus it is surprising perhaps that more recently, starting with the invention of formal axiomatic logic in the sense of Frege, Russell and Hilbert, diagrams have been denied a formal role in theorem proving. It is generally thought by logicians that diagrams have no accepted syntax nor semantic theory in a logical formalism which would make them rigorous enough to be used in formal proofs. Only very recently, in the last two decades, have there been efforts to fill this gap and investigate whether and how diagrams can be used in formal proofs. For instance, the investigation of Pierce's existential graphs in (Sowa 84), the work on GROVER in (Barker-Plummer & Bailin, 1997), the work on Hyperproof in (Bar- wise & Etchemendy, 1991), the introduction of computational models for interpreting Euler's circles in (Stenning & Oberlander 95), the analysis of the use of Venn diagrams as a formal system in (Shin 95), and the formalisation of a logical theory of Venn diagrams in (Hammer 95). Our work contributes in some sense to the effort in the research from the formal perspective on the use of diagrams, especially that of automated reasoning systems which use diagrams in the reasoning pro- cess. Our aim is to formalise diagrammatic reasoning and to show that diagrams can be used for proofs. In this paper we show how diagrams can be used for proofs in a formal system. We look into how theorems Roger Penrose presented his position in the lecture at International Centre for Mathematical Sciences in Edinburgh, in celebration of the 50th anniversary of UNESCO on 8 November, 1995. On Automating Diagrammatic Proofs of Arithmetic Arguments 3 of mathematics can be expressed as diagrams for some concrete values, i.e., ground instantiations of a theorem. The initial diagram is manipulated using some geometric operations. The sequence of geometric operations on a diagram represents the inference steps of a diagrammatic proof. Such a concrete proof instance is called an example proof. The set of all available operations defines the proof search space. A general pattern is extracted from these proof instances, and is captured in a recursive program. This recursive program constitutes a general diagrammatic proof for the universally quantified theorem. An existing technique in logic, namely the constructive !-rule, justifies the step from schematic proofs to theoremhood. We also aim to investigate the relation between formal algebraic proofs and more "informal" diagrammatic proofs. Usually, theorems are formally proved with the use of inference steps which often do not convey an intuitive notion of truthfulness to humans. The inference steps of a formal symbolic (as opposed to diagrammatic) proof are statements that follow the rules of some logic. The reason we trust that they are correct is that the logic has been previously proved to be sound. Following and applying the rules of such a logic guarantees that there is no mistake in the proof. We want to have such a guarantee in our proof system, and moreover, to gain an insight into the proof. Ultimately, the entire process of diagrammatically proving theorems will illuminate the issues of formality, rigour, truthfulness and power of diagrammatic proofs. We implemented a diagrammatic proof system called Diamond, which automates such diagrammatic reasoning and applies it to problem solving in mathematics. The user interactively constructs example proofs by choosing an initial diagram which represents the theorem, and then applies diagrammatic operations to build a proof. Diamond then automatically extracts a general pattern from these instances, and captures it in a recursive program. First, we list some of the theorems and their diagrammatic proofs. These help us define our problem domain. Second, we present Dia- mond's architecture, some operations required, the abstraction mechanism employed, and indicate how to verify the abstracted proof. Next, we report on some of our results and discuss future work. Then, we discuss some of the related diagrammatic reasoning systems. Finally, we conclude by summarising the main points of this paper. 2. 'Diagrammatic' Theorems We are interested in mathematical theorems that admit diagrammatic proofs. In order to clarify what we mean by diagrammatic proofs 4 Mateja Jamnik et al. we first list some example theorems. Then, we introduce a taxonomy for categorising these examples in order to be able to characterise the domain of problems under consideration. 2.1. Examples 2.1.1. Pythagoras' Theorem Pythagoras' Theorem states that the square of the hypotenuse of a right angle triangle equals the sum of the squares of its other two sides. Here is one of the many different diagrammatic proofs of this theorem, taken from (Nelsen, 1993, page 3): a c a b a a a The proof consists of first taking any right angle triangle, completing a bigger square by joining to it identical triangles and squares along its sides, and then rearranging the triangles in a bigger square. 2.1.2. Sum of Odd Naturals This example is also taken from (Nelsen, 1993, page 71). The theorem about the sum of odd naturals states the following: Note the use of parameter n. If we take a square we can cut it into as many ells (which are made up of two adjacent sides of the square) as the magnitude of the side of the square. Note that one ell is made out of two sides, i.e., 2n, but the shared vertex has been counted twice. Therefore, one ell has a magnitude of (2n \Gamma 1), where n is the magnitude of the square. 2.1.3. Geometric Sum This example is also taken from (Nelsen, 1993, page 118). A theorem about a geometric sum of 1 states the following: On Automating Diagrammatic Proofs of Arithmetic Arguments 52 1 _Note the use of ellipsis in the diagram. Take a square of unit magnitude. Cut it down the middle. Now, cut one half of the previously cut square into halves again. This will create two identical squares making up a half of the original square. Take one of these two squares and continue doing this procedure indefinitely. 2.2. Classification From the analysis of the examples that we presented above, and many others (see (Jamnik, 1998) and (Nelsen, 1993)), three categories of proofs can be distinguished: Category 1: Non-inductive theorems. Usually, there is only one representative diagram for all instances of the theorem. There is no need for induction to prove the general case: proofs are not schem- atic. Simple geometric manipulations of a diagram prove the individual case. Abstraction is required to show that this proof will hold for all a; b. Theorems are of continuous space. Example the- orem: Pythagoras' theorem. Category 2: Inductive theorems with a parameter. A diagram is a representative of a particular instance of a theorem. Proofs are schem- atic: they require induction for the general diagram of magnitude n (a concrete diagram cannot be drawn for this instance). An alternative method can sometimes be used to capture the generality of the proof. Theorems are of discrete space. Example theorem: sum of odd naturals. Category 3: Theorems whose proofs are inherently inductive: for each individual concrete case of the diagram they need an inductive step to prove the theorem. Every particular instance of a theorem, when represented as a diagram requires the use of abstractions to represent infinity. Theorems are of continuous space. Example theorem: geometric sum. 6 Mateja Jamnik et al. 2.3. Problem Domain We choose mathematics as our domain for theorems since it allows us to make formal statements about the reasoning, proof search, induction, generalisations, abstractions and such issues. Having introduced the examples and their categorisation, which is by no means exhaustive, we are now able to further restrict our domain of mathematical theorems. First, we narrow down the domain to a subset of theorems that can be represented as diagrams without the need for abstraction (e.g., the use of ellipsis, as in the above example theorem for geometric sum). Conducting proofs and using abstractions in diagrams is problemat- ic, since it is very difficult to keep track of these abstractions while manipulating the diagram during the proof procedure. Second, we consider diagrammatic proofs that require induction to prove the general case (i.e., Category 2 above). Namely, diagrams can be drawn only for concrete situations and objects. We cannot draw, for example, an n \Theta n square without some abstraction device, e.g., ellipsis. Our challenge is to find a mechanism for extracting a general proof that does not require using abstractions in diagrams. 2 The generality of the proof will be captured in an alternative way (by using the constructive !-rule - see x2.4). Third, to date we consider theorems of natural number arithmetic only. Diamond is designed to prove examples of Category 2, where diagrams represent natural numbers. A more formal definition of diagrams is given in x3.4.1. We may extend diagrammatic theorem proving for examples of Category 1 as well. One of the possibilities for future work is to consider a need for a more formal problem domain definition. 2.4. Constructive !-Rule As mentioned above we use the constructive !-rule to prove theorems of Category 2. Siani Baker in (Baker et al, 1992) did some work on the constructive !-rule and schematic proofs for theorems of arithmetic. Here, we explain the idea behind constructive !-rule and schematic proofs and how they can be applied to diagrammatic proofs. 2.4.1. Schematic Proof Schematic proofs use the constructive !-rule which is an alternative to induction. The constructive !-rule allows inference of the sentence 8xP (x) from an infinite sequence P (n) of sentences. 2 Note that (Barker-Plummer & Bailin, 1997) formalises the use of abstractions, however in the domain of well founded relations. On Automating Diagrammatic Proofs of Arithmetic Arguments 7 where "if each P (n) can be proved in a uniform way (from parameter n), then conclude 8nP (n)." The criterion for uniformity of the procedure of proof using the constructive !-rule is taken to be the provision of a general schematic proof, namely the proof of P (n) in terms of n, where some rules R are applied some function of n (i.e., f R (n)) times (a rule can also be applied a constant number of times). Let the proof of P (n) be captured using a recursive function proof(n). Now, proof(n) is schematic in n, since we applied some rule R n times. The following procedure summarises the essence of using the constructive !-rule in schematic proofs: 1. Prove a few special cases (e.g. P (2), P (16), . 2. Abstract (guess) proof(n) (e.g. from proof(2), proof(16), . 3. Prove that proof(n) proves P (n) by meta-induction on n. The general pattern is extracted (guessed) from the individual proof instances by (learning type) inductive inference. By meta mathematical induction we mean that we introduce system Meta such that for all n: where ":" stands for "is a proof of ". Baker used PA ! (i.e., Peano arithmetic with !-rule) for the system Meta (Baker et al, 1992). The meta inductive rule is defined as follows: This essentially says that by using the rules on P (s(n)) we can reduce it to P (n). For more information, see (Baker et al, 1992). 2.4.2. Diagrams and Schematic Proofs We claim that we can extend Baker's work on schematic proofs to our diagrammatic proofs so that the generality of the diagrammatic proof is embedded in the schematic proof. Thus, we eliminate the need for abstractions in diagrams, and can extract a general schematic proof from manipulations on concrete diagrams. The diagrammatic schematic proof starts with a few particular concrete cases of the theorem represented by the diagram. The diagrammatic procedures (i.e., operations) on the diagram are performed next, 8 Mateja Jamnik et al. capturing the inference steps of the diagrammatic proof. This step corresponds to the first step of the schematic proof procedure given in the previous section x2.4. The second step is to abstract the operations involved in the schematic proof for n. Note that the generality is represented as a recursive program which specifies a sequence of diagrammatic procedures (oper- ations) that are used on a diagram, and not as a general representation of a diagram. More precisely, the basic idea is to consider proofs for which can be reduced to proofs for n (or conversely, such proofs for n which can be extended to proofs for by adding to them some additional sequence of operations). The difference between the proof for (n and the proof for n, i.e., the additional sequence of operations in the proof for (n respect to the proof for n is referred to as the step case of the abstracted schematic proof. The last step in the schematic proof procedure is to prove by meta- induction that the abstracted diagrammatic schematic proof is indeed correct. One way of proving the correctness of schematic proofs is to create a theory of diagrams that models the processes in a diagrammatic reasoning system and prove correctness there. A formal definition of a diagrammatic proof of an arithmetic statement, and the correctness of this diagrammatic proof will be discussed in x3.4. 2.4.3. Schematic Diagrammatic Proof for the Sum of Odd Naturals Now we can structure the diagrammatic proofs in a more formal way. Here we list the proof for the theorem about the sum of odd naturals as a sequence of steps that need to be performed on the diagram: 1. Cut a square into n ells, where an ell consists of 2 adjacent sides of the square. 2. For each ell, continue splitting from an ell pairs of dots at the end of two adjacent sides of the ell until only 1 dot is left (note that for each ell of magnitude n, we will have of dots plus another dot which is a vertex of the two adjacent sides, i.e., 1). Identifying the operations (i.e., geometric manipulations) that were required to prove the theorem will help us define a large repertoire of such operations which will be used in the diagrammatic proofs. The generality of the proof is captured by the use of the constructive !- rule, by which we take a few special cases of the diagram (say squares of magnitudes 15 and 16), and find the general pattern of the proof that will hold for each case (e.g., the schematic proof given above). On Automating Diagrammatic Proofs of Arithmetic Arguments 9 3. Diamond System The diagrammatic proof system Diamond is an embodiment of some of the ideas presented in this paper. Diamond stands for Diagrammatic Reasoning and Deduction. Clearly, an important issue in the development of Diamond is the internal representation of diagrams and operations on them. It was George P'olya who was first to advise us on the importance of knowledge representation (see (P'olya, 1945) and (P'olya, 1965)). Simon argued P'olya's point further in (Simon 96) by stating that solving a problem means representing it so that the solution becomes trivial, or at least transparent. In automated reasoning it is difficult to see how to use this advice, since there is normally only one representation scheme for the problem which is available to the system. In Diamond we choose a representation which we hope captures the intuitiveness, rigour and simplicity of human reasoning with diagrams. We aim to represent diagrams in a way which enables a theorem prover to prove theorems using diagram. In Diamond we use a mixture of Cartesian and topological rep- resentations. Diamond uses a primitive notion of a diagram, a dot. All other elementary and derived diagrams (e.g., rows, columns, ells, frames, squares, triangles, rectangles, etc.) are composed in various ways out of dots. The advice of P'olya about alternative representations can readily be used in Diamond. Namely, diagrams can be represented in a variety of different ways. For instance, a square is represented as: a sequence of rows; a sequence of columns; a concentric sequence of cir- cumferences, each of which is called a frame; a nested sequence of ells; a sequence of four similar squares; a matrix; a sequence of diagonals. The choice of the representation that Diamond uses is important. Most of the proofs that Diamond proves require some kind of recursive decomposition of a diagram. Each alternative representation makes available a different form of recursive decomposition. For more information on the choice of internal representation for diagrams in Diamond see (Jamnik, 1998). The architecture of Diamond consists of two parts. The diagrammatic component forms and processes the diagram. It is the interface between Diamond and the user. The inference engine deals with the diagrammatic inference steps. It processes the operations on the dia- gram. An important submodule is the abstraction mechanism which is used to extract general schematic proofs from example proofs. The rest of this section presents the operations used to construct proofs, the structure of proofs and the abstraction mechanism used in Diamond. Mateja Jamnik et al. 3.1. Geometric Operations Geometric operations (also referred to as manipulations or procedures) capture the inference steps of the proof. Thus, a sufficiently large number of such operations which are then available to the user in the search for the proof, needs to be identified and formalised. Since we are not generating, i.e., discovering diagrammatic proofs, but rather we are trying to understand them, we can expect from the user to input these operations. To date, a small number of such operations has been implemented and is available to the user. Diamond is targeted to prove theorems of discrete arithmetic. Diagrams are a way of representing natural numbers. The interest lies in the effect on the numbers that diagrams represent after an operation has been applied on the diagrams. Thus, the operations join and split diagrams apart in various ways. Some operations are just simple ones (e.g., split a row from a square), and some are more complicated ones (e.g. decompose a square into a sequence of rows). Hence, Diamond distinguishes between two types of operations, atomic and composite: 3 Atomic operations: are basic one-step operations that can be combined into more complex operations. Examples of such operations are: rotate, translate, cut, split, join, remove, insert a segment,. To date, there are fourteen atomic operations implemented in Diamond Composite operations: are more complex, typically recursive opera- tions, composed from simple atomic ones. One can think of them as tactics in automated reasoning. Composite operations are defined in terms of decomposition of different recursive representations of diagrams. Depending on the theorem at hand, the diagram is viewed using a particular representation, which enables one to use a particular recursive composite operation. Ideally, the internal representation of the diagram is pertinent to the composite operation that is being carried out on it. Such a representation would render an operation very easy to apply. It would be just a simple decomposition of the representation of a diagram. Examples of such operations are: recursive decomposition of a square into rows, or columns, or ells, or frames,. In the example of the theorem for the sum of odd naturals the proof consists of the following operations: lcut and split ends. 3 A complete list of operations can be found in (Jamnik, 1998), however, a more formal definition of operations as part of the diagrammatic theory can be found in On Automating Diagrammatic Proofs of Arithmetic Arguments 11 3.2. Constructing a Proof Diamond's example proof consists of a sequence of applications of geometric operations on a diagram. The abstraction is then carried out automatically, if any such abstraction exists for the two example proofs given. 4 Diamond expects the example proofs to be formulated in a particular way where the order of operations in the user's formulation of the example proofs is crucial. Both example proofs are expected to be given with the same order of operations, but with some extra operations in the case of the proof of (n + 1) for some particular n. Consider the example for the sum of odd naturals. The step cases for proofs for look as follows: The aim is to recognise automatically the structure of the proof from a linear sequence of applications of operations, so that the example proofs for n and n+ 1 can be reformulated in the general case into the where for each n, A(n) is a step case consisting of a sequence of applications of some operations and B is a base case for Alternatively, we seek this recursive reformulation: Note that proof(0) is often an empty list of operations, because often no diagram is defined for a diagram which consists of no dots. 4 If the proof contains a case split for say, even and odd integers, and the two example proofs given are for two different cases, then Diamond cannot abstract from them. However, Diamond recognises that the example proofs were given for different cases, and requests the user to supply another example proof for each case, in order for it to be able to abstract. This will be further explained in x 3.3.2. 12 Mateja Jamnik et al. A further issue that we are investigating currently is to relax the requirement for a particular ordering of operations in formulating example proofs. Sets with partial ordering could be used as an alternative. 3.3. Abstraction Given some example proofs Diamond needs to abstract from them, so that the final diagrammatic proof is not only for the cases of specific n's, but holds for all n. Such a schematic proof is a general program which specifies the applications of some operations, where the number of application of each operation is dependent on n or is a constant. We distinguish between two types of example proofs: destructor, i.e., the example proofs which are formulated so that the base case operations are performed last (in a sense, the initial diagram is "destructed" by the application of operations down to a trivial diagram, forming the proof along in this way); and constructor, i.e., the base case operations are performed first followed by the step case operations. In Diamond we arbitrarily choose to use destructor schematic proofs. A proof that has the same structure for all n, i.e., one recursive function defines a complete proof for all n, is called a 1-homogeneous proof. Proofs can be c-homogeneous; then there are c cases of the proof. For instance, when there are two cases of a proof, one for odd and one for even natural numbers, then there need to be two recursive functions defining each case of the proof - the proof is 2-homogeneous. We say that if all concrete instances of the proof (for instances of numbers that "equal modulo c") have the same structure and can be abstracted, then the proof is c-homogeneous. If there are c cases, then there are c different abstracted proofs, one for each case. We seek the smallest complete recursive definition of a proof, i.e., c potentially different schematic proofs, if there are c cases. The following theorem and corollary will help us define what we mean by the smallest complete proof: THEOREM 1. If a proof is c-homogeneous, then it is also (kc)-homo- geneous for every natural number k ? 0. The immediate consequence of Theorem 1 is: COROLLARY 1. If a proof is not c-homogeneous, then it is also not f-homogeneous for every factor f of c. In a c-homogeneous proof we will denote by B r a base case for a branch of numbers which give remainder r when divided by c. B r is actually a proof for the smallest natural number that gives remainder r when divided by c. A schematic proof is defined to be the smallest complete proof if there is no other f-homogeneous proof obtainable from a c-homogeneous On Automating Diagrammatic Proofs of Arithmetic Arguments 13 proof for any factor f of c, and all f schematic proofs for f cases are defined. The general representation of a destructor proof is formalised as number of cases and r ! c, and 1. Then the recursive definition of a general proof is: where A r is a step case and B r is a base case for a class of proofs where denotes concatenation of sequences of operations: "do operations of A(n+ 1), then proof(n)". The formalisation of abstracted proof for constructor proofs is symmetric to the one given above. 3.3.1. Abstracting For All Linear Functions As mentioned above, we aim to recognise the particular recursive structure of the given example proofs. More precisely, we want to extract the step case A and the base case B of the proof and then abstract them for all n. The general methodology employed for doing this can be demonstrated as: The first step of the abstraction algorithm is to extract the difference between the two given example proofs for in the hope that this, when abstracted, will be the step case A of the proof. This is done by commutative and associative matching 5 which detects and returns the difference between the two example proofs. Now we have a concrete step case of the proof. This difference consists of a few operations op k each applied x k;n 1 times for some natural k. To make a step case general, we need to find the dependency function between every x k;n 1 and n 1 . This demands identifying a function of n 1 , which would give a specific x k;n 1 for some k and assumes that the dependency is linear: an + b. This is a heuristically adequate choice. Thus, let us write for each op k a linear equation an 1 , where n 1 and x k;n 1 are known. The subsequent stage of the abstraction is to extract the next step case from the rest of the example proof for the corresponding new n 5 Using commutative and associative matching reduces the sensitivity to the order of proof steps (Jamnik, 1998). 14 Mateja Jamnik et al. (i.e., n 2 ). If successful, continue extracting step cases for the corresponding n's from the rest of the proof until only the base case is left. Since we are dealing with inductive proofs, it is expected that every step case of a proof will have the same structure, i.e., will consist of the same sequence of application of operations, but a different number of times. Thus, we could in the same way as above for every operation op k write a linear equation an However, the number x k;n2 of applications of a particular operation op k in the next step case is not known. A possible value of x k;n 2 is acquired by counting the number x 0 of times every operation op k of the initial step case occurs in the rest of the proof. The actual value of the number of occurrences of each operation could be any number from 0 to x 0 . Thus, we do branching for all such values and thus we have: an an are known, so the equations can be solved for a and b, and x k;n 2 takes values from 0 to x 0 . This results in several possible potential abstractions of the step case. The aim is to eliminate those that are impossible. After checking if step cases for all n down to the base case are structurally consistent one hopes to be left with at least one possible abstraction of the example proofs. The step case is rejected when the sequence of operations in the subsequent step cases is impossible, i.e., the functions were wrong. This normally occurs when the dependency function gives a negative number of applications of a particular operation, when the calculated sequence is not identical to the rest of the example proof, or when there is no integer solution to our equations. Usually, there will be only one possible abstraction of the two given example proofs. The example proof for the sum of odd naturals is abstracted into the following step case and base case: diagonal ends,n-1)] where the function in parentheses indicates the number of times that the operations are applied for each particular n. 3.3.2. f-Homogeneous Proof Assume two example proofs for the sum of odd naturals (the example proof would consist of making n lcuts, and then showing that each ell consists of an odd number of dots). If the user supplies two example proofs for values of n and n + 1, for some concrete n, then there is no On Automating Diagrammatic Proofs of Arithmetic Arguments 15 problem, so Diamond will abstract normally and determine that the proof is 1-homogeneous. However, should the user supply proofs for concrete n, the first stage of abstraction would determine that the step case consists of two lcuts. However, a complete recursive function for abstraction requires a step case to consist of one lcut only. Diamond checks this by trying to split the step case into a further f structurally the same sequences of operations, for all factors f of c in order to obtain an f-homogeneous proof. If the method fails, then there is no such f-homogeneous further abstraction of the step case A(n). If the method succeeds, and Diamond finds a new abstraction of the step case, call this A 0 (n), then it also needs to find a new base case r 0 where the previous r for c was such that and the new r 0 is now such that 3.4. Correctness of the Schematic Proof The last stage of extracting a diagrammatic proof is to check that the guessed general schematic proof is indeed correct. To prove that the schematic proof is correct we need to show in some meta-theory that proof(n) uniformly proves P (n) for all n, i.e., it gives a proof tree with P (n) at its root, and axioms at its leaves. This requires reasoning about proofs, i.e., meta-level reasoning. A meta-level proof using general diagrams would be an obvious method for verifying our schematic proof. However, such a meta-level proof would reintroduce the need for abstractions (e.g. ellipsis) of diagrams, which we are trying to avoid. One way of overcoming this problem is to define diagrams and operations in a theory of diagrams where we can express abstract diagrams symbolically rather than diagrammatically. In this theory we can verify schematic proofs by defining the notion of applicability of a posited proof. Given that a particular theorem is expressed as an equality, its schematic proof is correct if applying the operations specified in the schematic proof of the diagrammatic representation of the left hand side of the theorem results in the diagrammatic representation of the right hand side of the theorem. There are two conditions that need to be satisfied. The first condition is that there is an appropriate diagrammatic representation available for the conversion of the theorem into its diagrammatic representation. The second condition is that the operations of the schematic proof are defined on those diagrams. Before we can state the definition of the correctness property of schematic proofs, we need to formalise the machinery which will enable us to model the processes of a diagrammatic proof. Therefore, we need to formally define diagrams, operations on them, and the applicability of operations of a schematic proof. Mateja Jamnik et al. 3.4.1. Diagrams Diagrams in the theory are defined to be of object type. Some examples of the different kinds of object names in the theory are: row, column, ell, frame, square, rectangle, and triangle. Diagrams of the theory model natural numbers. Diamond's primitive notion of a concrete diagram, a dot, is represented in the theory as the natural number 1. Objects are introduced via a constructor func- tion, diagram, which takes the name of the type of a diagram and the list of parameters of its magnitude. Thus, the type of constructor function diagram is name \Theta pnat list ! object. So, for instance, a square of magnitude 4 is expressed in the theory as diagram(square,[4]). All elementary and derived concrete diagrams are expressed using a primitive object dot, hence in the theory they can be expressed using a constructor function, the object name and some parameter representing a natural number for the magnitude of the diagram. Constant ; denotes a null diagram, or in other words an empty diagram. We define that any diagram that is of 0 magnitude is an empty diagram (note that a 2 b denotes that a natural number a is an element of a list b; thus the type of an infix 2 is: pnat \Theta pnat list ! Note also, that all triangles are equilateral. 6 here are some examples of diagrams: diagram(row,n), diagram(column,n), diagram(square,n) and diagram(ell,n). 3.4.2. Operators This section gives the operators available in the theory. First, we write diagrammatic equality using d which denotes that two lists of diagrams are identical. here is the definition of d where the function count can be defined by: count(d; d :: d 6 It is hard to represent discrete triangles that are of any magnitude, i.e., the sides are of different and any magnitudes. Triangles are represented in a discrete space. Hence, they appear to be right-angle triangles, despite the fact the all the sides of any triangle are of equal discrete magnitude. Were we to extend Diamond to prove theorems of real arithmetic (see x4), then there would be a need for continuous space, and therefore a scope for triangles of any magnitude. On Automating Diagrammatic Proofs of Arithmetic Arguments 17 Diagrammatic equality d is a larger relation than an arithmetic equality =, because it has all the properties of =, i.e., reflexivity, sym- metry, transitivity and substitution properties, plus an additional one - the order of elements in a list does not matter. Therefore, two lists of diagrams, X and Y, are diagrammatically equal, X d even if the orders in which the diagrams are listed in both lists differ. 7 We now define some operators that introduce the existence of several diagrams (note that the data type pnat stands for non-negative natural number of Peano arithmetic): @ is append on lists,; :: and nil are list constructors (concatenation of elements onto a list, and an empty list); \Omega is an infix operator which introduces a combination of a number of identical lists of diagrams; U denotes a collection of diagrams of increasing magnitudes which are all of the same kind - it is analogous to for summation of integers. here is the recursive definition of for all a b: 8 a d a d diagram(name; f(i))@ Note that f is some function which generates a list of natural numbers for a given number i. This list denotes the parameters of a magnitude of a diagram. 3.4.3. Operations Diagrammatic operations are represented via a function op : opname \Theta object list ! object list. We give here a definition of one operation only, but there are many more operations defined in the theory - see (Jamnik, 1998). op(lcut; diagram(square; n):: D) d diagram(ell; n)]@D (4) 3.4.4. One Apply and Apply here we define what it means to apply an operation on a diagram several times. We use a function apply and function one apply. Let: 7 Note that our definition of diagrammatic equality of lists is equivalent to bag equality. The order of the elements in a bag does not matter. 8 Note that to simplify the notation we write D(i) instead of U (a; b; i:D(i)). Mateja Jamnik et al. one apply(0; opnm; D) d one apply(n apply([ ]; D) d 3.4.5. Equations here we give a theorem which will be needed. (9) is provable from (7) and (8). Its proof is not given here, but can be found in (Jamnik, 1998). apply(ops; D :: D s Let dmap denote a relation between a particular class of statements of arithmetic and their equivalent diagrammatic expressions in the theory of diagrams. The equivalence is defined to be over the size of the diagram. The size of a diagram is defined to be the number of counters (dots) in the diagram, i.e., the natural number that the diagram rep- resents. dmap takes two arguments, an arithmetic expression and a list of diagrams which could collectively represent this expression. Hence, the type of the relation dmap is pnat \Theta object list. here are some general conversions: We have now formalised enough machinery to be able to define the correctness property of a schematic proof. DEFINITION 1 (Correctness of Schematic Proofs). proof is a correct schematic proof of a particular conjecture 8n R(n) if for all n there exist two lists of diagrams D and E such that dmap(L(n); D) and dmap(R(n); E), and apply (proof (n); D) d On Automating Diagrammatic Proofs of Arithmetic Arguments 19 It is possible to prove the property in Definition 1 only if L(n), R(n) and proof are known, i.e., for a specific case of a conjecture and a schematic proof. Knowing L(n) and R(n) allows us to infer some conversion relations which specify two lists of diagrams D and E. This satisfies the first part of Definition 1. In the next section we prove the correctness of a schematic proof for a particular conjecture at hand. 3.4.7. Proof of Correctness of Schematic Proofs for an Example here we prove the property given in Definition 1 for an example of a schematic proof of a theorem about the sum of odd naturals. The theorem is stated as 1). The schematic proof of this theorem is given as: 9 The proof of correctness of a schematic proof for this particular example requires induction on n. The base case for since by (10) no operations are applied to an empty diagram list which results in [ ]. We consider a step case of induction. Step case: Hypothesis: for n Using hence let Using (13) and (12) notice dmap( diagram(ell; [i])), hence diagram(ell; [i]). diagram(ell; [i]) Conclusion: for Similarly to the hypothesis, D and E are converted for n + 1. diagram(ell; [i]) diagram(ell; [i]) 9 For the brevity of presentation we take a simpler version of the schematic proof which does not include the operation split ends. 20 Mateja Jamnik et al. apply(proof(n); one apply(1; lcut; diagram(ell; [i]) diagram(ell; [i]) diagram(ell; [i]) diagram(ell; [i]) (RHS of hypothesis) diagram(ell; diagram(ell; [i]) diagram(ell; [i]) d diagram(ell; [i]) 3.5. Arithmetic Conjecture and Diagrammatic Proof Definition 1 makes no claims about the link between a schematic proof and the theoremhood of a conjecture 8n We still need to consider the possibility of a correct schematic proof of a false conjecture. To establish that the conjecture is true when proved by a schematic proof, an explicit algebraic link between them needs to be defined. We establish this link via the size of diagrams. We first define the size of a diagram, and later, in Theorem 2, we state the theorem about the algebraic correctness of a schematic proof for a given conjecture. Let us denote the size of the diagram D by j D j. here is a definition for the size of a diagram: DEFINITION 2 (Size of Diagrams). The size of a list of diagrams is equal to the value of the arithmetic expression that it represents: if dmap(e; D) then j D Note that the type of j j is: object list ! pnat. Using the property of size defined in Definition 2 on formulae from (10) to (13), we have the On Automating Diagrammatic Proofs of Arithmetic Arguments 21 Apart from being diagrammatically correct, we want every schematic proof to be algebraically correct as well. A schematic proof is algebraically correct if the sizes of the diagrams representing both sides of the proposition after the operations of the schematic proof have been applied are the same. Theorem 2 states the property of algebraic correctness for any schematic proof. of Schematic Proofs). For all instances of a schematic proof P and for all pairs of lists of diagrams D and E, a schematic proof P is algebraically correct if and only if apply The proof of Theorem 2 is straightforward by appealing to the properties of diagram size invariance under applications of multiple opera- tions. The lemmas about these properties and the proof of Theorem 2 are not given here, but can be found in (Jamnik, 1998). There is one last theorem needed in the formalisation of diagrammatic theory which will allow us to prove theorems of arithmetic using diagrammatic proofs. We state in Theorem 3 the property about the diagrammatic provability of arithmetic arguments. THEOREM 3 (Diagrammatic Provability of Arithmetic Conjecture). A conjecture 8n diagrammatically provable if and only if for all n there exist two lists of diagrams D and E such that dmap(L(n); D) and dmap(R(n); E), and The proof of Theorem 3 is trivial by the definition of size of a list of diagrams given in Definition 2. 3.5.1. Diagrammatic Provability for an Example We consider now an example of an arithmetic conjecture and prove it diagrammatically using a schematic proof that Diamond extracts. Let the arithmetic conjecture be 22 Mateja Jamnik et al. and the schematic proof proof that Diamond extracted be as defined in (14) and (15). here are the reasoning steps of the proof: 1. Appealing to Theorem 3 we can discharge the conjecture by: using (11) notice dmap(n 2 ; [diagram(square; [n])]), hence let using (13) and (12) notice dmap( diagram(ell; [i])), hence diagram(ell; [i]), and proving for all n diagram(ell; [i]) 2. Appealing to Theorem 2 and proof(n) that Diamond extracted, we can discharge the expression in (20) by proving for all n apply (proof(n); [diagram(square; [n])]) d diagram(ell; [i]) (21) 3. Finally, notice that we already proved (21) in x3.4.7. 4. Results and Further Work Diamond is implemented in Standard ML of New Jersey, Version 109. 10 The code is available upon request to the first author. The entire process of interactive construction of proofs, automatic abstraction from example proofs, and automatic verification of schematic proofs in the theory of diagrams have been implemented in Dia- mond. In the evaluation of Diamond we distinguished between a development and a test set of theorems which we proved using Diamond. The development set of theorems included three theorems: the sum of odd naturals, i.e., the sum of all naturals, i.e., i\Gamma0 i, and an odd triangular sum, i.e., T ri 3T ri n , where T ri i is an i-th triangular number (Nelsen, 1993). The Standard ML of New Jersey (SML/NJ) is a compiler and programming environment for the Standard ML programming language. SML/NJ is publicly available via the internet on the following site: http://cm.bell-labs.com/cm/cs/what/smlnj/- index.html On Automating Diagrammatic Proofs of Arithmetic Arguments 23 test set included twenty six theorems. Some proofs of these theorems are reported more elaborately in (Jamnik, 1998). All of these theorems contribute to the significant range and depth of theorems proved using Diamond. For more information, the reader is referred to (Jamnik, 1998). We want to relax the restriction currently imposed on the formulation of example proofs. Our abstraction mechanism can deal with a linear sequence of operations. This sequence is in fact a linearisation of some partially ordered sequence of operation. We want an abstraction mechanism which would be sensitive to partially ordered sequences of operations. There is also a possibility of allowing non-linear dependency functions in general schematic proofs: e.g. exponential or polynomial function Some recognition and generalisation of diagrams using abstractions could be an interesting issue to consider. This requires some formalisation of abstractions (e.g. ellipsis) in diagrams. There is a possibility to extend Diamond's problem domain from natural number arithmetic to geometry or even further to a different field such as hardware verification. Extending our problem domain to geometry would enable us to prove theorems of Category 1 which are usually geometric theorems of continuous space. These do not require induction, hence there would be no need for Diamond's abstraction mechanism. The generality is embedded in the use of continuous space and diagrams of general magnitude. The existing operations and the formalisation of schematic proofs can be used. Additional operations for moving diagrams in various directions would need to be implemented. For more information on the possible extension of Diamond to other problem domains, see (Jamnik, 1998). Diamond is an interactive proof checker. A long term goal is to design an automated theorem prover capable of discovering diagrammatic proofs. 5. Related Work Several diagrammatic systems such as the Geometry Machine (Gel- ernter, 1963), Diagram Configuration model (Koedinger & Anderson, 1990), GROVER (Barker-Plummer & Bailin, 1997), and Hyperproof (Barwise & Etchemendy, 1991) have been implemented in the past and are of relevance to our system. Additional information about issues in reasoning with diagrams can be found in (Chandrasekaran et al, 1995) which is a good reference for demonstrating how extensive and important this field is. Mateja Jamnik et al. One of the first systems to use a diagram in proving theorems was Gelernter's Geometry Machine (Gelernter, 1963). The diagram in the geometry machine has two roles. Its negative role is to reject hypotheses (subgoals) that are not true in the diagram. In this way the search space is pruned. The positive role of the diagram is to shorten the inference paths by assuming various facts that are obvious in the diagram as true. (Koedinger & Anderson, 1990) implemented a geometry problem solver called the Diagram Configuration (DC) model. The key feature of the system is that it organises its data in perceptual chunks, called diagram configurations. These are analogical to key features of diagrams that humans recognise when they inspect a diagram. Therefore, during the process of generating a solution path, DC infers the key steps first, and ignores along the way the less important features of the diagram, i.e., the less important inference steps. "&"/GROVER, developed by (Barker-Plummer & Bailin, 1997) is an automated reasoning system which uses information from a diagram to guide proof search. Its problem domain are theorems of well founded relations. It consists of the "&" automated theorem prover, based on the sequent calculus for Zermelo set theory, and GROVER, which is the diagram interpreting component of the system. It passes information extracted from the diagram and translated into logical formulae in the language of "&" to the "&" theorem prover. These formulae are then used as additional hypotheses to the main proof of the conjecture. GROVER considers only subgoals that are known to be true in the diagram. The main common feature of these three systems is their use of dia- grams. The diagram is used to model algebraic statements, and the system uses these models for heuristic guidance while searching for an algebraic proof. Thus, the basic underlying reasoning process is non- diagrammatic. In contrast, proofs in Diamond are explicitly constructed by operations on diagrams, thus the inference steps of the proof are entirely diagrammatic. Perhaps a more closely related system to Diamond is Hyperproof (Barwise & Etchemendy, 1991). Hyperproof reasons about the blocks world. It is an educational tool to teach principles of logic reasoning and proof construction. Hyperproof is an interactive tool for proof checking, as opposed to an automated theorem prover. It is a heterogeneous logic system, because it models inferencing between different kinds of rep- resentation. Unlike traditional systems for first-order logic, which use sentential representation, Hyperproof uses sentential and diagrammatic (graphical) representation. It uses a diagram for a concise representation of a complex system aiming to aid human reasoning. The user On Automating Diagrammatic Proofs of Arithmetic Arguments 25 can take advantage either of conventional sentential inference rules or diagrammatic inference rules. It differs from Diamond in that Hyper- proof's diagrammatic inference rules deduce from a diagram to a sentential formula or vice versa. It does not have diagrammatic inference rules between two diagrams, as it is the case in Diamond. Moreover, in Hyperproof the sentential inference rules (as well as diagrammatic description of a situation) are essential to construct a proof. Closer to our work is work done by Siani Baker (Baker et al, 1992) described in x2.4 on constructive !-rule, whereby she exploits the uniform structure of inductive proofs to abstract from example proofs. The main difference between Baker's work and ours is the problem domain. Baker implemented the use of constructive !-rule for proving theorems. Our domain, on the other hand, is diagrammatic theorems. Furthermore, Baker's motivation was to use schematic proofs for theorems that require a cut rule in the inductive proof, otherwise the proof cannot be carried out automatically. Schematic proofs avoid the need for a cut rule. On the other hand, we use the constructive !-rule in order to justify the automatic provision of general arguments about theorems and their proofs from particular instances. 6. Conclusion One of the aims in the research reported here is to see whether it is possible to automate the use of diagrams in formal proofs. The hope is that automating the 'informal' diagrammatic reasoning of humans will shed light on the issues of formality, informality, rigour and 'intuit- ive' understanding of the correctness of diagrammatic proofs. We have made good progress in exploring this important and difficult area. In particular, we have an explicit handle on abstraction. We showed that diagrams can be used for formal proofs. We presented, as an example, a diagrammatic reasoning system, Diamond, which supports interactive construction of diagrammatic proofs. Diamond applies diagrammatic reasoning to problem solving in mathematics. The user proves concrete examples of a theorem, and the system automatically abstracts these instances to give a general schematic proof which we hope holds for all n. In Diamond we have the logical machinery (meta-theory, constructive !-rule) to subsequently justify that the schematic proof does indeed prove the original theorem. Acknowledgements The research reported in this paper was supported by an Artificial Intelligence Department Studentship, from the University of Edinburgh, 26 Mateja Jamnik et al. and a Slovenian Scientific Foundation supplementary studentship for the first author, and by EPSRC grant GR/L/11724 for the other two authors. --R On the use of the constructive omega rule within automated deduction. Visual information and valid reasoning. In Zimmerman Realization of a geometry theorem-proving machine Logic and visual information. Automation of diagrammatic reasoning. In Pollack Also available from Edinburgh as DAI Research Paper No. Automating diagrammatic proofs of arithmetic arguments. Forthcoming PhD thesis. "Diagrammatic Reasoning: Cognitive and Computational Perspectives" "Diagrammatic Reasoning: Cognitive and Computational Perspectives" Proofs Without Words: Exercises in Visual Thinking. How to solve it. Mathematical discovery. The Logical Status of Diagrams. The Sciences of the Artificial. Conceptual Structures: Information Processing in Mind and Machine. A cognitive theory of graphical and linguistic reasoning: Logic and implementation. --TR --CTR Oliver Lemon , Maarten De Rijke , Atsushi Shimojima, Editorial: Efficacy of Diagrammatic Reasoning, Journal of Logic, Language and Information, v.8 n.3, p.265-271, July 1999
automated reasoning;theorem proving;diagrammatic reasoning
595967
A New Criterion for Comparing Fuzzy Logics for Uncertain Reasoning.
A new criterion is introduced for judging the suitability of various fuzzy logics for practical uncertain reasoning in a probabilistic world and the relationship of this criterion to several established criteria, and its consequences for truth functional belief, are investigated.
Introduction It is a rather widespread assumption in uncertain reasoning, and one that we shall make for the purpose of this paper, that a piece of uncertain knowledge can be adequately captured by attaching a real number (signifying the degree of uncertainty) on some scale to some unequivocal statement or conditional, and that an intelligent agent's knowledge base consists of a large, but nevertheless nite, set K of such expressions. Whether or not this is the correct picture for animate intelligent agents such as ourselves is, perhaps, questionable, but it is certainly the case that many expert systems (which one might feel should be included under the vague title of 'intelligent agent') have, by design a knowledge base which is precisely of this form. Supported by an EPSRC Advanced Course Studentship, ref. 94416613 y Supported by EPSRC Research Assistantship, ref. GR/H/43045 Commonly the scale on which the uncertainty is measured runs between zero and one where the two extremes represent, respectively, 'certainly not' and 'cer- tainly' and intermediate values represent a qualied degree of condence, or be- lief, or truth, depending on where they stand in relation to the categorical end- points. So, for example, if an expert meterologist expressed in natural language his 'knowledge' that Rain tomorrow is doubtful this might very well be represented in this form, with the expert's blessing, as something like Bel(Rain where the gure 0.2, the 'belief', or `truth', value the expert gives to 'Rain to- morrow', corresponds, in the expert's mind, to 'doubtful'. A knowledge base K of this form where degrees of uncertainty are represented by truth, or belief, values lying between the classical values one (i.e. true) and zero (i.e. false) is, like a library, of little practical use without some method of manipulating, or processing, the knowledge in order to infer truth values for other statements as required. Whilst a multitude of methods for doing this in practical expert systems have been tried, one commonly recurring assumption, or device, (which dates back as far as MYCIN) is that the connectives 'and', `or', and 'not' are, or at least may be treated as being, truth functional. To be precise, let be a language for the propositional calculus, so the p i are propositional variables, and let SL be the set of sentences formed from L using the connectives ^; _; : as usual. For 2 SL let Bel() 2 [0; 1] be the 'truth value' (whatever that means) of . So, formally Bel is a function from SL into the interval [0; 1] and we are in the situation where our knowledge base consists of the values of Bel on some 1 ; :::; m , or more generally certain relations between these values, and we want to use this knowledge to infer the value of Bel on some other sentence . The assumption of truth functionality of the function Bel is that there are xed functions such that for all Three particularly popular choices here are _ _ _ often refered to, especially the rst of these, as Fuzzy Logic (or Logics). [We hasten to add at this point that the term Fuzzy Logic is also used to describe the logic, or logics, of vagueness. This paper has nothing to do with vagueness, it is to do with the assumption of truth functionality in uncertain reasoning for infering degrees of belief. For a discussion of the dierence between these two qualities see, for example, [Kruse 1994] or [Dubois 1988]] The use of fuzzy logic in reasoning with uncertainty raises two questions (at least). One is why it is justied to assume truth functionality in the rst place, and, related to this, what such truth values could mean in this case. The second is, having decided in favour of truth functionality in a particular situation, what choice of F should be made? On the former we shall have something to say later, but turning rst to the second question { what choice should be made of F { there are, to our knowledge, two main approaches to this question in the literature. The rst is to argue that F should satisfy certain desirable properties, properties that we might feel the connectives 'and', `or', 'not' exhibit in everyday usage. Taking the case of negation rst we might feel that F : should satisfy: That is, (N1) is saying that the negation of a certainly false statement should be certainly true and the negation of a certainly true statement certainly false. (N2) says that increasing one's belief in should decrease one's belief in :. Finally (N3) says that two negations cancel each other out. In this case we have the following rather satisfying result due to Trillas, [Trillas 1979], which justies the choice of F : in F 3 , above. Theorem 1 (N1-3) hold if and only if the rst order structure h[0; isomorphic to h[0; 1]; 1 x; <i, i.e. Proceeding similarly in the case of F ^ the following have been proposed as desirable properties. increasing (not necessarily strictly); That is, (C1) is saying that the conjunction of a certainly false statement with a certainly true statement, in either order, is certainly false, whilst the conjunction of two certainly true statements is still certainly true. (C2) says that the truth value of ^ should not decrease if the truth value of one of the conjuncts increases (whilst the other remains unaltered). (C3) says, in essence, that microscopic changes in the truth values of and, or, should not produce a macroscopic change in the truth value of the conjuction ^. Finally (C4) might be justied by arguing that in everday language we in no way insert 'parentheses' when saying a multiple conjunction, suggesting that in pratice its truth value is entirely independent of any way of 'bracketing' it. In this case we have the following theorem which, appearing as it has in various forms and fragments in the literature (the earliest known to us is [Ling 1965]) probably now simply deserves to be treated as folklore (see [Paris 1994a] for a proof). Theorem 2 Let F ^ satisfy (C1-4). Let 0 a < b b: Then on [a; 1] [0; a] and on [0; a] [a; isomorphic to F 2 (on [0; or to F 3 (on [0; or In other words this theorem says that (C1-4) force F ^ to look like a chimera of F 1 and copies of F 2 and F 3 . Of course one might question why we should stop at (C1-4), why not add in further desirable properties of conjuction? In fact as Theorem 1 shows, one of the most obvious further properties, that F ^ is commutative, already follows from (C1-4) whilst adding the only remaining clearly missing desirable property, idempotence (i.e. F ^ be F 1 min. If we were to combine both the desirable properties of negation and conjunction, then in addition to (N1-3)+(C1-4)+idempotence certain other natural laws seem to emerge, for example but since there is, by the theorem, already only one choice for F ^ we no longer have any latitude in this matter and natural laws such as the one cited may indeed have to be renounced. [It is interesting to note here that whilst models of (N1-3)+(C1-4) may separately have a negation isomorphic to 1 x and a conjuction isomorphic to a chimera of F 1 it is only in the case where there is actually no contribution to this mixture from F 2 that we can ensure that both isomorphisms coincide. (see [Paris 1994a] for a proof)]. Exactly as for conjunction we can list some desirable properties of F _ as (D2) F _ is increasing (not necessarily strictly); F _ is continuous; F _ is associative i:e: F _ (x; F _ (y; and again prove a version of Theorem 1 with the subcript ^ replaced everywhere by _ (and min by max). Once the decision has been made to assume truth functionality these considerations clearly distinguish the fuzzy logics F 3 (and especially F leading candidates in the context of processing uncertain knowledge. A second argument, this time for one particular choice for F ^ , has been given by Krienovich et al in [Kreinovich 1988], [Kreinovich 1990]. The argument runs as follows. Consider an expert giving belief values, or truth values if you prefer, to certain statements today and then repeating the exercise tomorrow. Unless he, or she, was prewarned that the exercise would be repeated it seems rather unlikely that they would give exactly the same gures again, even if they had in the meantime received no new information justifying a revision of their beliefs. However it does seem much more reasonable to suppose that the ordering of the values given to various statements would remain the same. [In a way we are saying here that, apart from zero and one the actual gures given don't mean very much, it is their relative sizes that matter.] Now if the expert was interested simply in these statements it would not matter whether he or she was using today's or tomorrow's values. However if the expert was combining, or these statements truth functionally according to some F then it would, or could, make a dierence if (which we assume are xed) did not themselves also preserve this relative ordering. Thinking of the drift in the expert's stated belief values between today and tomorrow as an arbitrary rescaling, that is order preserving one to one onto map g from [0; 1] to [0; 1] this preservation of relative ordering amongst sentences behoves, in the case of conjunction, that As Kreinovich et al show this requirement, for arbitrary rescalings, together with (C1) and (C3) forces F ^ to be equal to F 1 i.e. to min: A similar argument for disjunction forces F _ to be equal to F 1 _ , i.e. to max: Following this line of reasoning then leads to distinguishing F 1 _ as particularly appropriate choices _ in this context. [It should be added that in the case of negation and assuming that F : satises (N1), (N2), the above argument yields a choice for F : between the two functions 1 -(x); -(1 x), where -(x) takes value 1 for and value 0 for 0 x < 1. However, both of these functions fail to satisfy (N3).] Having described these arguments in favour of certain choices of F under the assumption of truth functionality we now turn to examining this latter assumption. The most immediate criticism of this assumption is that, in requiring no account is taken of any relationship between and . Thus, to take an extreme example, if (and surely this could easily happen) then we should be led to Bel(p _ whereas the average thinking man or woman might feel that the left hand side expressions should be 0, 1 respectively and the two right hand side expressions equal to Bel(p). Their problem here in accepting such conclusions, and the rules from which they are derived, is that as they stand there appears to be no semantics for belief which could possibly support them and still make sense within the context. Indeed the lack, up to this time, of an accepted, or even acceptable, semantics for truth functional belief, despite its frequent and continuing presence in expert systems is, to say the least, 'unfortunate'. [We hasten to repeat that we are talking here of belief not vagueness.] On the other hand there is an alternative theory which does have a supporting, and widely accepted semantics, namely probability theory. In our context this amounts to assuming not that the function Bel but that Bel is a probability function, that is it satises This interpretation of belief, or truth, values certainly can be supported in a wide range of contexts (for example in a medical expert system) with a suitable semantics, namely by identifying these values with frequencies. In addition there is the rather forceful Dutch Book argument for a rational agent's belief values being probabilities which results from identifying belief with willingness to bet. (See, for example [Paris 1994a].) Moreover, not only are there convincing arguments for belief being identied with probability, but there is a widespread trust (at least amongst statisticians) that uncertainties in the real world are themselves explainable as probabilities. That being the case it might seem reasonable to propose that, ideally, not only should the intelligent agent's belief values be probabilities (in the sense that they satisfy (P1-2)), but moreover that, ideally, they correspond to actual probabilities in the real world. If we were to accept this thesis (as we henceforth do in this paper), that ideally an intelligent agent's belief, or truth, values should be probabilities, and furthermore correspond to real world probabilities (without wishing to enter the mineeld of what exactly that means at this point), then it would seem that the problem of infering the truth values of other statements from K would move squarely into the realm of statistical inference. However, despite the rather general sympathy for this thesis, there is, of course, a major problem asssociated with attempting to go down this path, namely maintaining computational tractability (even more so it would seem for intelligent agents like ourselves with apparently very limited computational abil- ities). The key computational problem associated with working with probability functions is that if Bel is a probability function dened on SL (where as usual) then in order to specify Bel we need, in general, to know the value of Bel on all except one of the 2 k atoms of L, i.e. sentences of L of the i where the i 2 f0; 1g and p i i is dened to be p i if and to be 0: This su-ces because, by the disjunctive normal form theorem, for each 2 SL there is a subset S of the set At L (= of atoms of L such that so by (P2) [Only are needed to specify Bel since for a tautology S by (P1) in practical expert systems k would typically be at least 40 it is clearly unreasonable, in general, to be able to specify Bel, nor indeed would one expect that the 'knowledge' provided by the experts would by itself be su-cient to determine Bel given only that Bel is a probability function. On the other hand it is clear from (F1-3) that for truth functional Bel Bel(), for 2 SL, is determined simply by the k values Bel(p 1 ); :::; Bel(p k ), which, in general, will be directly available from the expert providing the knowledge base. So, it would seem that if we accept the thesis, as we do in this paper, that ideally an intelligent agent's belief, or truth, values should be probabilities, and furthermore correspond to real world probabilities, then we have a choice between moving away from probabilities, and getting the wrong answers, or sticking with probabilities and not, in general, being able to come up with any answers at all! [Of course this is an exaggeratedly pessimistic picture. There are, under suitable assumptions and situatuations, honestly statistical expert systems which have even be mooted as possible models for human uncertain reasoning (see for example [Pearl 1988]). Overall, however, their current applicability is limited.] Given this choice between no answer and the wrong (or, more accurately, an unjustied) answer one would surely feel that the latter was preferable (after all if an express train is thundering down the track towards you jumping either way has got to be better than doing nothing). And, having made that choice, of going for an answer which, necessarily, will sometimes be less than perfect, the next step is clear, within the parameters left open to you minimise the error. A new criterion There are many ways, of course, of measuring the error resulting from using a truth functional belief function to approximate a probability function. However a natural rst attempt might be to take the sum of the squared dierences where Bel() in this expression is determined, truth functionally, from F and the Bel(p i and w is this probability function on SL representing the 'state of nature'. Unfortunately this rst attempt fails to take into account the practical considerations appropriate to the situation we have in mind (i.e. of an intelligent agent reasoning in the real world). Worse still, except in very special circumstances, this sum will not even be dened! This is because we are summing over all the innitely many 2 SL and there is clearly no reason why (Bel() w()) 2 should tend to zero as the length of increases, indeed quite the opposite, we would expect that as the length of increased Bel() and w() would become ever more unrelated so that any closeness in their values would be simply by chance rather than by design. But clearly, from the agent's point of view, long sentences are very largely irrelevant because in the practical business of reasoning in the real world the agent is only going to encounter very short sentences involving rather few connectives. This consideration leads us to take the sum in the error not over all of SL but instead over the set ~ n of sentences formed from f using exactly n binary connectives chosen from f^; _g. We are thinking here, of course, of n being small, indeed in many of our results we shall limit ourselves to in which case the sentences in ~ are just single conjunctions or disjunctions of literals. On the other hand we would expect k to be very large, re ecting the very large number of (variable) features the agent would expect to encounter in any real world situation involving uncertainty, and, apart from this, to be essentially unknown to him. Indeed almost all our results will be given as assymptotic results as k tends to innity. The reason for choosing this particular class ~ n to work with is to some extent pragmatic, just as taking the square in the error term it helps make the mathematics more tractable (as [Paris 1994b] demonstrates). In practice one would hope that minor variations in the class of sentences whose probabilities we are trying to approximate truth functionally would have a rather marginal eect, a hope which is at least reasonably consistent with the results we shall shortly be presenting. Returning again to the appropriateness in this context of the error term E (now with ~ n in place of SL) we could suppose that the intelligent agent we have in mind would, like ourselves, not just hold beliefs in a few specialised areas where there was some hope of knowing the underlying probability distribution, but in a whole multitude of real situations, where, with maybe a small number of exceptions, the agent would know little or nothing about the underlying probabilities beyond (possibly) for those basic events which correspond to the propositional variables in our formalisation. [In fact even most expert systems, which one might have classed as 'very specialised', in practice actually approximate to this situation.] These considerations now lead us to make two further renements to (1). Firstly we shall assume that Bel(p i ) is simply a xed, increasing, function, That is we shall assume that the agent's belief in a basic event (such as we have denoted by a propositional variable) is purely determined by the probability of that event and that the more probable an event is the higher the belief the agent attaches to it. [By our earlier remarks on the ineable nature of k it seems unreasonable that Q n should also depend on this number.] Secondly, in view of the agent's large scale ignorance of the probability function w whose values s/he is endeavouring to approximate, it would seem perverse of the agent to treat all probabilty functions s/he encountered in a uniform manner as regards the choice of F Q. For that reason rather than take the error as in (1) for a single probability function w we shall take a weighted average of the error over all probability functions w on SL, weighted according to the likelihood, G L (w), that the agent will encounter that probability function. Precisely what we mean by this last italicised phrase is not a question we wish to discuss at length in this paper. It is, in another guise, the problem of picking a prior in Bayesian reasoning, and as such has already generated a voluminous literature without, apparently, any great concensus emerging. The notion might seem to make sense if we considered, say, the(?) probability distribution over all diseases, symptoms, signs etc. in medicine and then considered how the set of all restrictions (i.e. marginalizations) of this to a sublanguage with 3 propositional variables, so given by an eight-tuple of the values of the atoms, were scattered around in But rather than indulge in any further philosophical discussion we shall henceforth just talk of 'likelihood of being encountered', or as we shall call it 'nat- uralness', of a probability function and leave the reader to interpret it as s/he wishes. With these renements in place our new criterion can now be formalised in terms of minimising Z where the intention is that n is small, k is very large, Bel(p i xed increasing function Q n and G L (w) is some measure of the 'naturalness' of w, or of the 'likelihood of w being encountered', in the various real world situations that the agent might inhabit. We make explicit two assumptions about G L which, we believe, are quite reasonable properties to require of a measure of naturalness of probability functions w on SL. Firstly, we assume that G L satises weak renaming, that is, whenever is a permutation of At L resulting from transposing p , or from transposing probability functions w on SL, where w is the probability function satisfying Note that this simply amounts to saying it is irrelevant which propositional variable we use to denote a particular proposition, and whether or not it is used to denote that proposition or its negation. The property of weak renaming holds for all specic choices of G L that we consider in this paper. Secondly, we shall assume that for for every 1 i < j k the integral Z G L (w)dw is a non-zero, integrable function of ha; bi 2 [0; 1] 2 . Beyond this we shall not struggle to say exactly what well-behavedness conditions G L is to have, nor exactly how the integral is to be understood. Instead we shall take it that all our results carry with them the implicit assumption that all integrals appearing in the proofs are suitably well dened. In the particular examples that we cite it will be clear that this is the case. Returning now to the criterion of minimising the error E k n in (2), once we have xed n and G L we have various choices as to which of Q choose to x and which we will allow to vary in attempting to minimise E k n . In the next two sections we consider two applications of this criterion. [See also [Paris 1994c], [Bennett 1995] for earlier and related results along these lines.] Belief values in fuzzy logic In the introduction to this paper we provided arguments which, assuming belief was truth functional, rather favoured the choice of F 1 (= min) for F ^ . Indeed, in today's practical expert systems this choice is surely the most popular and is often refered to as 'fuzzy logic' (although this is also used as a generic term for any truth functional belief or logic of vagueness). Given its widespread use and popularity it is then natural to ask what function Q n , for a particular choice of G L when we to be F 1 (i.e. min; max; 1 x) respectively. One choice of G L for which Q n can be calculated explicitly is when G L is the completely independent distribution. That is, if w treats the p i 2 L as completely 1 Note that this denition is similar but does not correspond exactly to the denition of Weak Renaming for inference processes dened in [Paris 1994a] (where only the latter type of transpose is allowed). independent i:e: w then G L otherwise. As a measure of naturalness this distribution could be 'justied' on the grounds that in the wider world most events, or propositions, are independent. Of course this is a gross simplication. In particular in the narrow areas of expertise such as those covered by most expert systems we would certainly not expect, or want, all the features to be independent. (Having said that however, even in such situations it may not be so unreasonable to suppose that most small subsets of L will, in practice, be completely independent.) In this case we can show that (in the limit as k tends to innity) the Q n exist and are increasing polynomials of degree 2n x. [Proofs of the results stated in this and the following two sections are given in the appendix.] In particular we can derive a recursive formula for the Q n which for yields A surprising property of the Q 0 Q 15 that we have calculated explicitly is how little they vary, and how rapidly they appear to be converging. For example the maximum dierence between Q 0 and Q 1 is 0.093 whilst Q 7 and never dier by more than 0.005. In consequence the actual reduction in the error resulting from using these Q n is not staggering, an explicit calculation showing that the improvement in square error per sentence from ~ n as k tends to innity in using Q 1 rather than Q 0 (i.e. the identity function) is around 4% whilst the average square error per sentence resulting from using Bel (with F 1 to approximate w is around 3%. It is interesting to note that in a sense this gives a possible meaning, or se- mantics, to belief values in F 1 (and for truth functional belief in general), namely they are scaled probabilities, scaled so as to minimise the error. Notice also what has happened here to our earlier objection to truth functional belief that, for example, forces us to have If we explain belief values in this way then such identities are no longer a problem, all we need to say is that the approximations are rather poor in these cases. Indeed since we are arguing for an agent's beliefs being truth functional on purely practical, or pragmatic, grounds one might suppose that in simple cases like these an intelligent agent would adopt an alternative, better, approximation strategy, for example by simply setting In fact as k tends to innity sentences containing a repeated propositional variable account for an ever diminishing proportion of the sentences in ~ so that in the limit they are essentially neglected. It might be argued that this would not accurately re ect the actual experience of the agent and that on the contrary the agent would be disproportionately likely to meet such sentences. As with the simple case mentioned above however, there would be nothing to stop an intelligent agent in practice from adopting an alternative strategy in such cases if s/he saw some value in doing so. For small n this need, as in the example above, causes no problem because an intelligent agent might reasonably be expected to recognise the necessary immediate simplication. For larger however the problem with an agent attempting any sort of simplifying procedure to a sentence prior to approximating its probability is that, except where this was immediately recognisable, s/he would (apparently) risk a computationally very expensive diversion with no guarantee of being any better of at the end of it (see [Hopcroft 1979]). In the above example we have taken the rather drastic step of identifying 'nat- uralness' with `independence'. Whilst this may in some situations be a not unreasonable assumption, it is patently rather wide of the mark in many real world situations we could expect our intelligent agent to encounter, and these are completely disregarded by our function G L in this case. So are there not alternatives which might be prefered to the completely independent distribution? Unfortunately the problem of picking a prior in situations such as these of, essentially, complete ignorance, is one on which we continue to seek guidance from the statistics litera- ture. [ See [Good 1965], [Jerey 1948], [Paris 1994b], [Paris 1991] for approaches to the question in the univariate case and [Lawry 1994a], [Lawry 1994b] in the multivariate case.] The reader might wonder at this point why we do not simply assume that all probability functions on SL are equally natural, equivalently, equally likely to be encountered by our agent. In other words why not simply take G L to be the Uniform Distribution (i.e. G L functions w on SL)? Unfortunately such a distribution G L depends on L, or equivalently k, in a way which is perhaps unexpected, and certainly undesirable, in our context. To see the problem let us denote the language fp by Lm (so A k+1 n . Then for xed (which, notice, do not depend on the expected individual square errors Z or even the expected probabilities of , Z w()G Lm (w)dw; where G Lm (w) is identically 1, may dier according to whether 1. However from the point of view of our intelligent agent this seems counter intuitive because, as we have already said, the agent would not know k in general, nor, surely, would s/he consider its exact value at all relevant, beyond that it is very large. (Indeed it is questionable whether it even makes sense from the agent's standpoint to talk of 'the k'.) The underlying problem here is that if we take for each nite language G L to be the uniform distribution on the probability functions on SL then the family does not satisfy marginality, that is it does not satisfy that for L 0 L and w 0 a probability function on SL 0 , (w Z G L (w)dw: If we wish to keep G L k as the uniform distribution and have marginality then we must be prepared to use non-uniform distributions G L for other languages L. One way that this can be achieved is to take the G L to be, up to normalizing constants, the Dirichlet Priors, that is, G Y for some xed > 0, where and the i run through the atoms of SL. Taking (and indeed any G L with equal to the uniform distribution but forces the G L to be non-uniform distributions whenever jLj 6= k. Notice that these Dirichlet Priors do satisfy weak renaming. Unfortunately, when does not fully satisfy the conditions we earlier imposed on G L . As we shall see later (Theorem this family of priors has another unpleasant failing that would seem to make a second serious dent in their credibility as contenders for measures of naturalness (the rst dent, of course, being that we were led to them in the rst place by the intuition that all probability functions should be equally natural, whilst for all except possibly one value of k they fail to achieve this). We now see that in our context the problem of picking a prior actually involves picking a marginalising family of priors, or distributions. Fortunately the completely independent distributions do marginalise, which is the reason this problem of marginalisation did not appear earlier. Optimal choices for As a second application of our criterion we consider optimum choices of (which we shall denote by F opt _ ) to n in the limit as k !1 single conjunctions or disjunctions of literals) for a xed marginalizing family of G L . (We may assume that Q n is the identity, i.e. Bel(p i in this case since the Q n can be simply absorbed into the Throughout we shall assume that the G L satisfy marginality and weak renaming Having settled on some G L a rather vexing problem arises with specifying any particular F opt _ in that changing their values on a small number (or set) of points will not alter the value of the error E k 1 . To be more precise if F : agrees with F opt on a 1 -measure 1 set and F agree with F opt on a 2 -measure 1 set, where Z Z then replacing F opt _ by F in the error expression will have no eect on its value. Having noted this problem we will however henceforth talk about the F opt _ , it being acknowledged here that we mean almost everywhere (a.e.), that is to within a suitable -measure 0 set. With the above assumptions in place Theorem 3 Let 1. Then in the limit as k !1 F opt F opt R R Dx;y G L2 (w)dw F opt _ R R Dx;y G L 2 (w)dw where yg. (Notice that by an earlier assumption these denominators are non-zero.) For proofs, see the Appendix (where (31) gives the limiting value of the resulting mean square error). The result that F opt completely in agreement with the conclusion of Theorem 1 and seems to close the book on this question. As we shall see later the situation with F ^ and F _ is not quite so clear. Applying Theorem 3 in the case where G L is the completely independent distribution we obtain, just as would have been expected, that F opt _ In other words F 2 is the optimal choice in this case. If instead we take our G L to be the Dirichlet Priors with the uniform distribution on the probability functions on disregarding the ill-behaviour of the family members on SL k for k > 2 we obtain that F opt F opt _ _In other words the F opt _ in this case are the averages of their counterparts in F 1 and F 3 . If we now take instead the Dirichlet family with i.e. the uniform distribution on probability functions on SL 1 (extending F opt by continuity to the boundary of [0; comes out to be the value of an elliptic integral. Much worse however is that F opt arguments in (0; 1) 2 , not even increasing! Theorem 4 For the Dirichlet distribution with < 4, F opt fails to be monotone non-decreasing. Turning this observation on its head however, we certainly might be justied in adding the requirement that F opt ^ be increasing (and more generally the property of inferential monotonicity given in [Paris 1994c]) to the list of desiderata of a family G L of measures of naturalness. Returning for a moment to our earlier disappointment that we cannot have the uniform distribution for more than one value of k unless we are willing to marginality, it might at least be some comfort to observe that if we take a Dirichlet family in which the uniform distribution appears for some very large k (so is also very large) then the exact value of this k makes very little dierence to the G Lm for m small. In a sense then it could be argued that an agent who adopted some such G Lm for m small could 'accommodate' the thought that for a wide range of large k the corresponding G L k was the uniform distribution. In such a case we might further argue that (assuming it exists) the agent should take for his F ^ the limit of the F opt for the Dirichlet families as k !1. In fact by results in [Paris 1992] this limit of these functions does exist and on argument hx; yi gives the value of the middle root of the cubic d dz We now turn to stating some properties of the F opt Theorem 5 For 0 x; y 1 and under the assumptions of Theorem 3 F opt _ F opt (v) Notice that by using (iv) of this theorem we can obtain analogous results for the function F opt _ . As an immediate corollary of this result we have. Corollary 6 For 0 x 1 and under the assumptions of Theorem 3 F opt _ (0; F opt F opt do not satisfy (ii) of Theorem 5 we also obtain Corollary 7 Under the assumptions of Theorem 3, F 1 are never optimal (for any such G L , with On the other hand Theorem 5 elevates F 2 2 ) to a special status (see Appendix for the proof). Corollary 8 F 2 (i.e. multiplication) is the only F opt which is associative and continuous. Our next result shows that in a sense the converse to Theorem 5 holds. Theorem 9 If F ^ satises (i),(ii) of Theorem 5 then there is a marginalising family of priors G L satisfying weak renaming for which F ^ is F opt It should be noted here however that the G L constructed in the proof of this theorem are rather contrived and certainly make no pretentions to being 'measures of naturalness' ! In view of Theorem 9 it is interesting to compare the desirable properties (C1-4) of F ^ on which Theorem 2 rests and those properties of F opt emerging in Theorem 5. By Corollary 6 F opt although not necessarily (C2) or (C4) (although the failure of (C2) should, as we have already said, be viewed more as a question mark against the G L ). In the other direction we have already seen that (ii) of Theorem 5 does not follow from (C1-4) because it is not satised by either F 1 . On the other hand by Theorem 2 commutativity does follow from (C1-4) since it holds for each of F 1 Conclusion In this paper we have argued that in a probabilistic world an intelligent agent, with scant access to these probabilities and largely lacking the computational power required to draw statistical, or probabilistic, conclusions, may be justied in resorting to truth functional belief in order to function at all, and that in such a case the agent should adopt those which satisfy the criterion of minimising the resulting expected error. In order to apply this criterion it is necessary to rst x on a 'measure of the naturalness' of a probability distribution, that being identied with the likelihood that the agent would encounter this probability distribution. In this paper we have, without overwhelming enthusiasm, suggested two possible choices for this measure, the completely independent distribution and the Dirichlet families of priors. Other possibilities exist in the literature (see for example [Solomono 1978], [Li 1993], [Fine 1973], [Paris 1994b]), but at this level of generality the choice of a prior in a situation of, essentially, complete ignorance remains problematic. It is interesting to observe that in practical applications of statistical inference su-cient is usually known (or assumed ) about the situation to severely restrict the possible probability distributions which might be encountered and as a result the choice of prior is usually much less contentious. [It should also be pointed out that the customary conditioning on the statistical data tends to reduce the in uence of the prior so that the particular prior chosen is far less of an issue than in our case where we are not considering the possibility that the agent might somehow update his/her prior, or measure of naturalness, in the light of observations about probabilities in his/her world of experience.] our criterion favours the choice of 1 x for F : , a result very much in line with common practice and Trillas' earlier justication based on desirable properties, see [Trillas 1979]. When it comes to F ^ and F _ however our criterion as the only associative, continuous optimal F whilst F favoured in terms of desirable properties is never optimal under these conditions. In view of the fact that the case surely the most frequently encountered situation (apart from which is irrelevant here) this seems, all other factors being equal, a strong recommendation to use F rather than the ubiquitous F has to be resorted to. It would be pleasing, of course, to extend these results to n > 1, especially if Theorems 3 and 5 still held, at least for a wide class of measures G L . Unfortunately that appears, in general, to be a rather messy problem. A second generalisation of these results which we have not considered is the inclusion of implication amongst our connectives. Clearly there is little further interest in doing this if we are to view implication, ! , materially, as the equivalent of : _ . However, viewing implications as conditionals (and so as the conditional probability of given ) would appear to open up an interesting new area for further investigation. We close by mentioning that there is a rather natural generalisation (as it turns out) of our criterion which, instead of attempting to minimise the expected error incurred by using truth functional belief in place of the actual probabil- ity, attempts instead to minimise the average error resulting from using truth functional belief to determine truth values, on the basis of some knowledge base, rather than by using a particular probabilistic inference process, such as the Maximum Entropy Inference Process. Some preliminary results in this direction may be found in [Bennett 1995]. Appendix Calculation of the Q n for F 1 and the independent distribu- tion We seek an (increasing) function Q n (x) which is the limit as k ! 1 of the increasing n minimising the expected error Z -2f0;1g completely independent probability function and otherwise, and, in the second integral, w is the completely independent probability function which gives p i value x i , the assertion X is true and 0 otherwise, and ' p - i is true if (equivalently value Q k n is increasing) for k. Note that appearing in because we work with F 1 . In fact it may hold for more than one i with some and ~x, and thus we may be counting some terms more than once. However, the eect of this can be ignored since the set of the w for which this can happen has measure 0. Since ~ n is invariant under permuting the it is clearly su-cient here to minimise For let neg() be the result of transposing ^=_, p i =:p i for Then neg () :, ~ n is invariant under neg and if ' p - 1 . It follows then that since we obtain the same error if we use the (increasing) function 1 Q k place of 2 a with equality just if y, we see that the error for2 at most that for Q k (x). Since we are interested in the minimum error we may assume n (x) is already of this form, in which case it is easy to check that and, using (5), that the expression in (4) is twice where we can replace w(p 1 ) by x 1 . However since we are interested in the case where k ! 1, rather than minimising we shall instead rst attempt to minimise n is the set of those 2 ~ n with no repeated variables, and return later to the problem of minimising (6). The function (not necessarily increasing) to minimise (7) (up to a set of measure zero) is given by To see this notice that if Q 0 other function disagreeing with Q on a set of non-zero measure then substituting Q 0 subtracting (7) from it yields after some calculation which is strictly positive. Of course we would now like to show that Q n dened in (8) is increasing. Rather than do that at this point however we shall set about giving a specic recurrence relation dening the Q in fact. Once this is done it will be clear that Q n increasing. In order to simplify what follows let C n be those in B k n in which the propositional variables are p and, with the possible exception of p 1 , this is the order in which they appear when reading from left to right. Clearly appears in Let So, by marginality of the completely independent prior, Q Then, with the obvious abbreviations, for n > 0, m=n j+2 is the set of those in B k in which the variables p n j+2 , p n appear, and in that order. Taking the sum of the rst of these expressions within the brackets we obtain, for a xed 1 j n,X m=n j+2 Using the fact that for such we see that this is equal xn j+3;:::;x n+1 =0 is the set of sentences in C j 1 in which p n j+2 with the possible exception of p n j+2 , in that order when reading from left to right, and this expression is now seen to equal Considering now the sum of the second of the expressions within the brackets in and noting that equals where is the result of transposing p m and :pm throughout , we obtain similarly to (11) By similar arguments the third and fourth terms in (10) each yield Summing these contributions yields whilst directly from the denition, P 0 Repeating a similar argument with H n (x) gives From this it follows that H n (x) is actually independent of x, so and the solution to this is [see [Paris 1994b], lemma 2.2]. By induction on n we can now show that P n is a polynomial of degree at most 2n + 1. Consequently and dividing (14) by H n (x) and simplifying gives d n;j where Again by induction on n we can show that Q n (x) is increasing. It now follows from earlier remarks that Q n (x) is the increasing function minimizing (7). Turning back now to the original expression (6) suppose that Q k n is increasing and gives a value to (6) which cannot be improved by any other increasing function to within more than 1. Recall that we may assume that Q k Then, as in (9), for Q n as in (8) the expected error for Z when calculated using Q k n exceeds that when using Now by a straightforward induction on n for any x distinct [i.e. a measure one set of w] the value of is the same [since if exactly one of p(^) etc. Furthermore since for we see that X Hence the excess error in (18) is By the assumed 'near minimality' of Q k n the error (3) associated with using Q k cannot exceed that by using Q n by more than 1, so certainly, using (20), Z where Bel is calculated using Q n . Since (Bel this yields so 1: (21) n is increasing and Q n continuous if 0 < Q k n (a) Q n (a) for some a depending only on Q n (and similarly if Q n (a) Q k n (a)). It therefore follows from (21) that on (0,1) the Q k tend pointwise to Q n (x) as required. Concerning the improvement in the average error (per sentence in B k resulting from using the Q n (x) rather than simply the identity function, by where D n is the set of n in which, reading from left to right, the propositional variables appear in that order. Since for 2 D n and measure one set of ~x exactly one of ' so (21) simplies to Z 1 which can now be readily calculated for small n. As earlier this is also the limiting error with ~ n in place of B k n as k !1. As to the average error in using Q n , substituting (8) into (7) (which is clearly times the average error for simplifying yields Dividing by jB k using the value for H n previously shows the average error over n to be d~x. Then by splitting up the cases where 2 D n is a conjunction/disjunction we obtain that for n > 0 and d n;j as in (17), d n;j whilst Since we know Q enables us to calculate the average error over (and hence over ~ n in the limit as k !1), at least for small n.Proof of theorem 3 which give a value to Z which becomes arbitrarily close to the inmum (over all of values of this expression as k !1. [Here F () and the identity if As usual we shall start o by looking for to minimise the expression in but with i 6= j. In this case, by the renaming and marginalising properties of G, it su-ces to nd Z Now for w a probability function on L 2 and 1 ; 2 2 f0; 1g let be the permutation of At L2 such that for - renaming G L 2 whilst It follows then that Z Z (w(p 1)); F (w(p Z F since Z (w(p 1)); F (w(p Now let F (x; F so F (w(p 1)); F (w(p 2))): and, trivially, Hence replacing F in the expression in (24) for using the inequalityX we see that E(F It follows then that in our search for to minimise (23) we cannot do better than take shall shortly show that, in fact, no other F : can do this well, but for the present we shall only assume that F : our attention to F ^ . then by (24) and this choice of F : we seek F ^ to minimise Z A minimising function here (which we shall already denote by F opt F opt R R Dx;y G L 2 (w)dw where To see this suppose F ^ was any other possible choice (with F : Z Z Dx;y Z Dx;y Z Dx;y G L2 (w)dwdxdy Z Dx;y so if E(1 must be zero and F must agree on a 2 -measure 1 set, as required. An exactly similar argument applies for F opt _ . Returning now to F opt us suppose that (so by earlier results equality must hold here). Now since equality holds in (25) just if x that we must have the F ^ all equal (on a 2 measure 1 set of < w(p 1 the average of would give E(1 a.e. and, since E(1 E(1 (notice by weak renaming that theorem 4(ii) (whose proof does not depend on this theorem) F opt and adding (29) and (30) gives F : We have now shown that F opt _ as dened in the theorem when the are further restricted to be distinct. To remove this last assumption suppose that better, choice for arbitrarily large k. Then, by (28) and the discussion leading up to it, Z Dx;y Hence Z Dx;y O so since the right hand side can be made arbitrarily small (and by the assumptions on G L (w)dw) we must have that F Notice that from this discussion the expected square error incurred by approximating w() for 2 ~ A kby Bel() with tends to Z as k !1, where F opt ^ is given by (27). Proof of theorem 4 Let 0 a < b < 1and < 4. We shall show that for the Dirichlet distribution with this value of , F opt aby Corollary 6). 1. Then from the denition of F opt ^ in this case, F opt Z a y a Z aay a 1 a dy > 0 (by the change of variable y := y a) Z a0 y a 1 a y a a a 1 a dy > 0 (by splitting Z aainto Z a0 aand making the change of variable y := y in the latter) Z a0 y ay where a 1 a 1 a Now since A 1 (y) A 2 (y). Hence, since < 0, and the result follows. 2 Proof of theorem 5 (i) follows easily using the weak renaming property of G L . To show (ii) notice that F opt R R Dx;y G L (w)dw ; where, as usual, D R R Dx;y G L (w)dw R R Dx;y G L (w)dw Let be the permutation of At L 2 such that for . Then, since G L satises renaming Z Dx;y Z Z Similarly Z Dx;y G Z G L (w)dw (34) and (ii) follows from (32), (33), (34). (iii) follows from Z Dx;y Z Dx;y Z Dx;y G L (w)dw: To show the rst part of (iv) (the second part is proved similarly) notice that Z Dx;y Z Dx;y Z Dx;y Z Dx;y G L (w)dw Z by using the trick by which we derived (33). Again, as in (34), Z Dx;y G Z G L (w)dw and dividing both sides of (35) by this gives the required identity. (v) follows easily from the denition of F opt _ by noting that w(p 1 (or by straightforward manipulation using Proof of corollary 8 Suppose that F opt associative and continuous. We rst show by induction on m that F opt For this is true by Corollary 6. Assume it is true for m. Then if by Corollary 6 and the associativity of F opt F opt F opt by the inductive hypothesis, whilst F opt ; by the case i < 2 m above; ; as required: Hence, since F opt is continuous and agrees with multiplication on a dense set it must agree with multiplication everywhere. 2 Proof of Theorem 9 First dene for i;j;k to be the probability function on SL k such that for 2 At L k , Notice that properties (i)(ii) of Theorem 5 ensure that hence that w x;y i;j;k is well dened, and that w x;y Dene the probability functions w x i;k , for 1 i k, 0 x 1, and w k such that for 2 At L k on the probability functions on SL k by Now dene the prior G L k by It is straightforward, albeit rather tedious, to check that the G L k marginalise, renaming and give F opt --R Master's Thesis An Introduction to Possibilistic and Fuzzy Logics Theories of Probability The Estimation of Probabilities Introduction to automata theory On the foundations of fuzzy formalism: Explaining formulas for union Foundation of Fuzzy Systems Doctoral Thesis Representation of Associative Functions The uncertain reasoner's companion - a mathematical perspective Probabilistic reasoning in intelligent systems: Networks of plaussible inference Sobre functiones de negacion en la teoria de con- junctos difusos --TR --CTR Didier Dubois , Henri Prade, Possibility Theory, Probability Theory and Multiple-Valued Logics: A Clarification, Annals of Mathematics and Artificial Intelligence, v.32 n.1-4, p.35-66, August 2001
uncertain reasoning;truth functional belief;fuzzy logic;probability logic
595997
Beyond the Turing Test.
The main factor of intelligence is defined as the ability to comprehend, formalising this ability with the help of new constructs based on descriptional complexity. The result is a comprehension test, or C-test, which is exclusively defined in computational terms. Due to its absolute and non-anthropomorphic character, it is equally applicable to both humans and non-humans. Moreover, it correlates with classical psychometric tests, thus establishing the first firm connection between information theoretical notions and traditional IQ tests. The Turing Test is compared with the C-test and the combination of the two is questioned. In consequence, the idea of using the Turing Test as a practical test of intelligence should be surpassed, and substituted by computational and factorial tests of different cognitive abilities, a much more useful approach for artificial intelligence progress and for many other intriguing questions that present themselves beyond the Turing Test.
Introduction iThe original question, 'Can machines think?' I believe to be too meaningless to deserve discussionj (Turing 1950). Turing's intention was to leave behind that question and motivate the endeavour for making machines intelligent. Fifty years of philosophical discussions have shown that his goal has not been fullled, mainly because the Turing Test (TT) has been usually misunderstood as a test for intelligence and not simply as an imaginary test of humanity (Fostel 1993), a philosophical exercise contrived to stop the sterile debate that the advent of the rst computers were generating. Turing cannot be reproached for these misunderstandings. The TT iwas never intended as a test for intelligence, at least not by Turing himself, and it should not be criticized for failing something that it was not supposed to doj (Larsson 1993). In fact, there is no better way to state that intelligence can be reasonably judged by observation than Turing's imitation game, and that there is no reason to deny the intelligence of non-human beings. Nonetheless, the overuse of the TT as a referent or even a test of AI progress has had very negative consequences. AI has striven to imitate human's behaviour in many tasks, under the slogan iArticial intelligence is that thing that if made by humans would require intelligencej, which has promoted the view that ihuman intelligence subsumes machine intelligencej (Bradford and Wollowski 1995). Finally, the most negative result of the c Publishers. Printed in the Netherlands. Jose Hernandez-Orallo misinterpretation of the TT jointly with his celebrity is that there has not been the necessary eoeort for designing new alternative intelligence tests. The TT has even eclipsed such well-reputed proposals as Simon's early works on the relation between IQ tests and AI (Simon and Kotovsky 1963), on some heuristic approaches to solve analogy problems from IQ tests (Evans 1963), Chaitin's suggestion idevelop formal denitions of intelligence and measures of its various componentsj (Chaitin 1982). Even Johnson's call iNeeded: A New Test of Intelligencej (Johnson 1992) has been responded by formalisations of the TT (Bradford and Wollowski 1995), once again. From a practical point of view, it is vain to revive the little informative question iCan Machines Think?j (Epstein 1992) which has generated many words but few clues about how to progress in AI or how to understand intelligence. As any other discipline, AI requires an eoeective measure of its major issue, a gradual and detailed measure of intelligence. Concretely, a scientic measure of intelligence should comply with some requirements: intelligence is not an absolute attribute. From Darwin's imental continuityj to infant psychology, there is an unquestionable certainty that intelligence is a gradual aptitude. Any gradation of the TT as a function of the time of the test or the score of the judges shows the inappropriateness of the TT to measure intelligence in a gradual way, i.e., to give a continuous value of intelligence. The reason is obvious: the TT is a test of humanity (Fostel 1993), and the idea of being more or less human makes no sense. intelligence is multi-dimensional. It is quite unbelievable that there is a concrete ability that would be optimal for every context or world. Nonethe- less, it is conceivable that a concrete ability would be almost optimal for most contexts, and intelligence has sometimes been dened as isecond-best in everythingj. This may justify that a wide context as everyday life, which includes many other contexts, can distinguish a special kind of ability, the g factor, the primitive concept found in everyday life. Maybe the major problem of AI is that its reference of intelligent behaviour is human intelligence. Recently, AI researchers have paid attention to other 'intelligences': ants, rats, etc. in order to scale up the problem. However, the reference or the goal is always human intelligence. Nowadays the reason is not only pride and anthropocentrism but necessity, because ithere is not yet a solid denition of intelligence that doesn't depend on relating it to human intelligencej (McCarthy 1998). ffl Computationally based : According to the Church-Turing thesis, there is no reason to think that intelligent systems cannot be implemented in current computers. The only question is whether they have the necessary speed and memory for it. I share the opinion that icomputers of years ago were fast enough if only we knew how to program themj (McCarthy 1998), and the AI problem is just to discover what makes a program intelligent, or, in other words, iWhat kind of Information Processing is Intelligence?j (Chan- drasekaran 1990). Consequently, it is extremely signicant to be able to state the problem computationally, namely, to give the specication of the problem Beyond the Turing Test in computational terms, in order to solve the problem with AI means, which are exclusively machines and programs. intelligence is not an ethereal ability. It cannot be dened as iintelligence is what is measured by intelligence testsj. Intelligence must be expressed from its original meaning, the ability to comprehend, and this ability is what should be measured. Psychometrics has developed measurements of intelligence according to the rst two requirements. Since Spearman founded the eld (Spearman 1904), psychometrics has been more and more characterised by the scientic method: systematic experimentation and statistical rigour. iDespite the many short-comings of an IQ score, no other measure has been found to be related to so many other behaviors of theoretical or practical signicancej (Zigler and Seitz 1982). However, psychometrics has neglected, or failed, to incorporate the last three requirements, which, in fact, are highly related. Psychometrics is anthropomorphic by denition since it is the science of measuring human intelligence. Although there have been adaptations and essays with chimpanzees, dolphins and other animals, the reference is always the Homo Sapiens Sapiens. The frequently demanded theoretical foundation of psychometrics depends on the change of the point of reference, closely connected to the last three requirements. On the contrary, Computational Learning Theory is non-anthropomor- phic and computational. The question iWhat is to learn?j has been assimilated to the formal notion of identication in the limit (Gold 1967). After some discouraging results on the learnability of very simple languages, the complexity of learning has been studied for other paradigms, mainly PAC learning (Valiant 1984) and Query Learning (Angluin 1988). However, these theoretical results have not been used to develop rigorous measurements of learning ability. Contrarily, some contests and comparisons have been held for practical systems, but as collections of arbitrary examples extracted from the literature, without many justications of the theoretical complexity of each of them. The reason is that it is diOEcult to establish the complexity of an instance, when the results of computational learning theory apply to classes of concepts, and most results are asymptotical. Actually, the paradigm of identication in the limit is not applicable for nite instances, because they can be identied by themselves. Moreover, the philosophical problem of any measurement of learning ability is the same as the philosophical problem of prediction: given any nite set of examples, there are innite many concepts which are consistent with them and diverge in their predictions. This is exactly the 'subjectivity objection' of IQ tests: there may be controversy about the correct answer. In the following we adopt an information-theoretic approach to solve these problems in order to give a denition and a measure of intelligence according to the ve prerequisites mentioned above, exclusively based on 4 Jose Hernandez-Orallo concepts derived from the notion of Turing machine. Despite the title of this paper, the result is a veritable tribute to Alan Turing on the 50th anniversary of his celebrated paper. The paper is organised as follows. The following section introduces the necessary denitions and tools for the rest of the paper, as well as some technical diOEculties which are solved in the subsequent sections. In par- ticular, section 3 formalises the initially vague notion of comprehension in information-theoretical terms and relates it with the ontology that only intelligent systems are able to construct. Once settled on a computational setting for comprehension, Section 4 deals with its measurement by solving the 'sub- jectivity objection' under the notion of unquestionability, and by ordering the diOEculty of instances. This allows the construction of a comprehension test (C-test). Section 5 presents the results of applying it to humans and compares it with psychometrical tests. The applicability to AI is discussed. Section 6 studies the measurement of other factors (knowledge applicability, contextualisation, knowledge construction) under the same conditions that the C-test has been devised with. After the previous results and auspices, the TT is re-examined in section 7 and reduced to its original philosophical and even metaphorical character. Compared with the C-Test, the signicance of the TT is recognised, as well as the acute deciencies of its misinterpretation and incarnations, like the Loebner Prize. The nal section concludes with the proposal of a radical change of paradigm; a science of intelligence that would make it feasible to answer many new and fascinating questions. 2. Preliminaries and Technical Problems We choose any nite alphabet \Sigma composed of symbols (if not specied, f0; 1g). A string or object is any element from \Sigma , being \Delta the composition operator, usually omitted or represented by ha; b. The empty string or empty object is denoted by ". The term l(x) denotes the length or size of x in bits and log n will always denote the binary logarithm of n. For every string y, we denote with y n::m , being n m, the symbols from position n to position m. For every natural number n, we denote y 0::m , y n::l(y) , and y k::k+1 , respectively. A string x is a substring of y ioe there exist two strings z, w such that what is equivalent, string x is a prex of y ioe exists a string z such that what is equivalent, being Given any string x, we denote by x 0::l(x)\Gammad the prex of x with length d, i.e. the string x without its last d elements. Under the Church-Turing thesis, any description of any object of reality can be converted into a description in a universal computational machine (e.g. a universal Turing machine), so our idea of 'object' is, in this sense, Beyond the Turing Test universal. Given a universal descriptional machine OE, and a program p, we denote by OE(p) the result of executing p on OE. Two descriptions p and p 0 are extensionally equivalent (denoted It is easy to see that any description has innite equivalent descriptions. Denition 1. A k-projectible description for a string x is a program p on a descriptional mechanism OE such that 9y 2 \Sigma y, and 9w known as the prediction of p. The compression ratio of a program p is given by: CR OE The compression ratio of an innite projectible description is always innite. The relative compression ratio of a projectible description p for a nite string x is dened as CR OE The complexity of an object can be measured in many ways, one of them being its degree of randomness (Kolmogorov 1965), which turns out to be equal to the shortest description of it. Descriptional Complexity, Algorithmic Complexity or Kolmogorov Complexity was independently introduced by Solomonooe, Kolmogorov and Chaitin to formalise this idea, and it has been gradually recognised as a key issue in statistics, computer science, AI, epistemology and cognitive science (see e.g. Li and Vit#nyi 1997). Denition 2. The Kolmogorov Complexity of an object x given y on a descriptional mechanism (or bias) fi is dened as: where p denotes any iprex-freej fi-program, and OE fi (hp; yi) denotes the result of executing p using input y. The complexity of an object x is denoted by K fi It can be seen elsewhere (e.g. Li and Vit#nyi 1997) that Kolmogorov Complexity is an absolute and objective criterion of complexity, and it is independent (up to a constant term) of the descriptional mechanism fi. In other words, there is an invariance theorem that states that any universal machine can emulate another. For this reason many properties are proven just in an asymptotic way and fi is usually omitted. Throughout the paper we will use the relation as the asymptotical extension of !, namely a there exists an independent positive constant k such that a k. We will denote x as the rst (in lexicographical order) program for x such that K(x) also represents the idea of simplicity, deeply involved in the philosophy of inductive reasoning and learning theory. It is not surprising that learning and recognition were soon re-understood under this context. Solomonooe 1 It is obvious to see that 8x 2 \Sigma there is always a program of constant size of the form iprint the inputj. It is also easy to see that 8x 2 \Sigma because there is always a program of size less than l(x) plus a constant value of the form iprint xj. In the case that K(x) l(x) (i.e. l(x ) l(x)) we say that x is random. It is obvious that K(xjx since K(\Deltaj\Delta) is not computable, it is shown elsewhere (e.g. Li and Vit\Deltaanyi 1997) that the other way is just K(x 6 Jose Hernandez-Orallo proposed the view of unsupervised learning as information compression (Solo- monooe 1964) and Watanabe considered ipattern recognition as information compressionj (Watanabe 1972). The classical Occam's razor of the scien- tic method: igiven two alternative explanations, choose the simplest onej was formalised by Rissanen in 1978 under the name iMinimum Description nally re-formulated in its current one part code (Rissanen 1996). According to the MDL principle, given any sequence x, the optimal model in OE for it is x . If x is projectible, i.e. it allows the prediction of the subsequent symbols of the sequence x, then OE(x ) n+1 will be the most plausible prediction according to Occam's razor. In order to ensure that x is projectible we need to introduce a projectible variant of K: Denition 3. The k-Projectible Kolmogorov Complexity of an object x given y on a descriptional mechanism (or bias) fi is dened as: K 0 any iprex-freej fi-program. The literature has used Kolmogorov Complexity and not its projectible variant for prediction due to the following theorem: Theorem 1. For every string x, K 0 Proof. Every non-projectible program p can be transformed into a projectible program us denote by c the length of this extra coding of iand then print 1 foreverj. Hence, there exists a constant This can be extended to the denitions of K 0 and K, thus the theorem is proven. The contrary relationship does not hold. Consider the string nj. The projectible program iprint the natural numbers, orderedj has constant size, say l(p On the contrary, the non-projectible program iprint the rst n natural numbers, orderedj is, in the general case, not smaller than c Thus, the ideal MDL principle is represented by the rst (in lexicographical projectible description for x, denoted by x + , such that From here, a compression/prediction test based on Chaitin's proposal (Chaitin 1982) seems to be easily applicable. However, there are many technical reasons that explain that such an intriguing proposal has not been addressed yet 1. K(x) is not computable, so x + cannot be eoeectively computed. If a compression test is constructed, how do we know whether the subject's answer is a hit? 2. There are dioeerent equally alternative plausible descriptions: x + is just the rst one in lexicographic order of all the shortest descriptions. At least to the author's knowledge and as Chaitin himself has recognised (Chaitin 1998, personal communication). Beyond the Turing Test 3. Despite the invariance theorem that states that x + depends on OE only up to a constant, this constant is relevant if l(x) is small, and there is no reason to prefer one descriptional system over another. 4. The test intends to measure the ability of compression, but this does not match exactly 3 with the ability of comprehension, i.e., intelligence. The rst problem can be solved by incorporating time into the denition of K. The most appropriate way 4 to weigh space and time of a program, the log Cost fi (p x ), was introduced by Levin in the seventies (see e.g. Levin 1973). Then the next variant comes directly: Denition 4. The Levin's Length-Time Complexity of an object x given y on a descriptional mechanism fi: Kt where LT fi log Cost fi (hp; yi) This is a very practical alternative of Kolmogorov Complexity, because as well as avoiding intractable descriptions, it is computable. Moreover, it accounts better for the idea of simplicity, and Occam's razor should be better formalised under this variant. To extend LT-Complexity to projectible descriptions, we must measure Cost fi (p) in an asymptotical way. Consider a machine OE such that the output tape cannot be rectied. Cost fi (p)[::n] is dened as the time or machine steps such that the rst n symbols of the denite output are placed at the beginning of the output tape and Cost fi From here we only need LT fi (p x log Cost fi (p x )[n::m] and log Cost fi (p x )[::n] for the following denition: Denition 5. The k-Projectible Length-Time Complexity of an object x given y on a descriptional mechanism fi is dened as: Kt 0 This denition will serve as a start point to face the other three unsolved problems (2,3,4). In fact, they require rst to distinguish what is to comprehend (problem 4), which is addressed in the following section, and later on to solve how to measure it (problems 2 and 3). 3. Formalising Comprehension To comprehend is to understand either the inner mechanism or the plausible model of some evidence. In some way, comprehension is stricter than learning in terms of justication, because comprehension usually entails 3 iI just see how Kolmogorov Complexity and Intelligence could be well related but I don't think it would be `exactly' so.j (Hofstadter 1997, personal communication). 4 Intuitively, every algorithm must invest some eoeort either in time or demand- ing/essaying new information, in a relation which approximates the function LT . 8 Jose Hernandez-Orallo that the subject is able to explain the concept to others. In the case of innite concepts, this explanation is only possible if the subject has a nite description of the concept. Consequently, comprehension can be understood in terms of identication. However, if a concept is nite, like most everyday concepts, both notions diverge signicantly. A nite concept can be easily identied by the extensional description of the concept, which has no insight and surely has not identied any mechanism or pattern from it, if the evidence ever had one. This question is old in logic, where comprehension means the connotation of a term, opposed to its denotation or extension. Hence, an extensional description (by enumeration) has no connotation and consequently entails no comprehension at all. On the contrary, an intensional description (by comprehension) may have not discovered the right meaning or real mechanism of the evidence, but still has a chance of discovering it. There is a fundamental feature that determines this dioeerence: iThe Dened Thing CANNOT Appear in the Denitionj, which is known as Comprehension Requirement. It is also one of the four laws of denition, according to methodology (Bochenski 1965). It is frequently used as a criterion by teachers when asking their pupils whether they have comprehended a concept. In fact, the pioneer of the psychometric approach, Binet, designed his rst tests to avoid this irote learningj. Traditional use in mathematics also distinguishes informally an extensional denition from an intensional denition (or by comprehension). For innite sets, frequent in mathematics, every denition must be intensional (or by comprehension). Nonetheless, for nite sets there is still no formal dioeerence between an intensional description and an extensional one. At rst sight, Kolmogorov or Descriptional Complexity seems suOEcient to distinguish extensional descriptions from intensional ones. However, the MDL principle, which chooses the shortest description for a given concept x, does not ensure that the description is intensional. In the vast majority of cases, the data is not compressible, and the MDL principle will give the data itself, being the most extensional description, which does not give any hint about the comprehension of that data. Even in the rare cases where the data is compressible, a short description does not ensure that all the data is described intensionally; there could be a part that could be highly compressed and another part that could be quoted as an exception. The question is then more conspicuous: is there any way to distinguish pattern from exceptions, program from data? Koppel introduced the notion of sophistication with the goal of distinguishing the structural part of an object (Koppel 1988) from its data (or non-compressible part of it). Logical Depth, as dened by Bennett (Bennett 1988), is shown to be equivalent to sophistication up to a constant (Kop- pel 1987). Both, however, can 'disguise' a general eoeective interpreter as ctitious pattern and leave a great amount of real pattern as data. Beyond the Turing Test 9 It is then required a dioeerent approach to distinguish whether a description has exceptions (partially or totally extensional) or is composed exclusively of pattern (it is all structure or totally intensional). The idea is to compare the part which is used for all the data to the limit (the structure), with the part which is only used in some portion of the data (the exception). Denition 6. A description p 0 is (n; k)-equivalent in the limit to a description p ioe 9n 2 IN; n ? 0 and 9k 2 ZZ such that OE(p 0 Informally, two descriptions are equivalent in the limit if there is a point from which their predictions always match. If both descriptions are k-projectible with k nite they are always equivalent in the limit. If only one of both is 1- projectible then they cannot be equivalent in the limit. Hence, the denition applies when both descriptions are 1-projectible descriptions. Denition 7. A description p is a Fully Projectible Description of x given y ioe hp; yi is an 1-projectible description of x and :9p 0 s.t. hp hp; yi but (n; k)-equivalent in the limit to hp; yi and LT (p The rst condition that p 0 is not extensionally equivalent to p (p 0 6j p) is to avoid that given two or more equivalent descriptions, only the shortest one would be projectible. The second condition measures that this p 0 is simpler than p. Note that LT (and only applied to the rst chunk of length l(x) begin to be equivalent) is used instead of l. Example 1. Given the evidence i3, 12, 21, 30, 102, 111, 120j, we can consider several projectible descriptions. For instance, and 1 foreverj is not fully projectible because there exists a shorter description i1 foreverj which is equivalent in the limit. In the same way, D number 3. The following three numbers are obtained by adding 9 to the preceding one. Continue with number 102. The following numbers are obtained by adding 9 to the preceding onej is not fully projectible because there exists a shorter description iStart with number 3. The following numbers are obtained by adding 9 to the preceding onej which is equivalent in the limit. On the contrary, the description D 3 inumbers whose digits in decimal representation amounts to 3 orderedj is fully projectible. Similarly, the description D everj is fully projectible. Finally, the following description is also fully projectible values of a polynomial polynomial such that D 4 and D 5 may seem counterintuitive but it should be realised that a fully projectible description just formalises the idea of explanation (and not yet the comprehension describes the evidence, it accounts for all of it (there are no exceptions because it is fully projectible) and it can be related (explained) to others (because of the use of LT , descriptions which are extremely time consuming are avoided). Hence, D 4 , whether we like it or not, is an explanation for the evidence. For the moment, we can dene a new variant of descriptional complexity: Jose Hernandez-Orallo Denition 8. The Explanatory Complexity of an object x given y on a descriptional mechanism fi is dened as: fully projectible g The string y, which we have supposed empty in the previous example, represents the context or previous knowledge where the explanation must be applied. In the same way it is done with K and the MDL principle, we can denote with SED(xjy) the Shortest (in LT terms) Explanatory Description for x given y, i.e. the rst shorstest fully projectible (in lexicographic order) description for x given y. Logically, However, we still have that for most strings, SED(x) will be just the rote description irepeat x foreverj which does not follow the comprehension requirement. A rst idea to avoid this phenomenon is to force the description to be shorter than the data and to say that the data has no comprehensive explanation if this is not the case 5 . However, most of everyday data is not compressible and it is still comprehended. Another approach is the idea of reinforcement or cross-validation (Her- n#ndez-Orallo 1999a). For instance, if we remove the last element of the previous series, i.e. i3, 12, 21, 30, 102, 111j, it is not much expectable that 4 and D 0 5 would be produced but D 0 3 can still be generated. In general, Denition 9. Stability. A string x is m-stable on the right in the descriptional system In other words, a string x is m-stable on the right if taking m elements from the right, it still has the same best explanation. These m elements, if given a posteriori, are considered reinforcement or conrmation, and, if given a priori, are considered redundancy or hints to help to nd the explanation. Consequently, although rote learning can be trickily used to make an extensional description fully projectible, stability (like reinforcement or cross- validation) is a methodological criterion to avoid this phenomenon. This nally gives suOEcient characterisation to state that a description entails that the learner who has generated it has comprehended the data. There is still another reason to support the previous notion of comprehension as an ontological principle. Why must we avoid rote learning? Why must we anticipate? Why do children nd more complex patterns? (Marcus et al. 1999) Why are we genetically programmed to open any black box we are presented? This search for more informative hypotheses instead of the easiest ones may lead to fantasy, but this is not dangerous provided that the system can interact with the world in order to refute some of them. This informativeness or investment in the hypotheses was advocated by Popper for the scientic method, and as we have seen, it is equally applicable 5 A dioeerent approach is the notion of exception, studied and formalised in (Hern\Deltaandez- Orallo and Minaya-Collado 1998) and (Hern\Deltaandez-Orallo and Garc\Deltaa-Varea 1998). Beyond the Turing Test for cognition. Even if we make the very strong assumption of Occam's razor, i.e., things in nature are not complex unnecessarily, the previous rationale is justied by the fact that, just as every incompressible string has compressible substrings, most compressible strings have incompressible substrings, because the shorter the less worthy that is to compress. If the evidence is presented incrementally, it is better to invest in more informative or general hypotheses instead of nding the optimal one for each chunk, which will nally turn out not to be part of the whole description of the whole evidence. This rationale leads to the next theorem: Theorem 2. For every descriptional mechanism fi, there exists a constant c which depends exclusively on fi such that for every string x of length n with partition then SED(y) is not equivalent in the limit with s. Proof. Consider any string x and any prex y such that It has a fully projectible description p iprint y for everj with l(p y being the space which is required for coding iprint . for everj. Since the computational cost of p y is linear, say k is suOEcient to choose c c 0 to ensure that the description p y is shorter than s, and LT fi (p y log log l(x). Moreover, p y and s cannot be equivalent in the limit because s is fully projectible and, by denition, there does not exist a description with less LT equivalent in the limit. It is clear that the idea of stability or cross-validation is supported by the previous theorem. In fact, it is an innate aesthetic preference in the explanations that human beings generate. Why is it more pleasant the answer 23 to the series i3,7,11,15,19, .j than the answer 3? In Hofstadter words, iit would be nice if we could dene intelligence in some other way than ithat which gets the same meaning out of a sequence of symbols as we dojj (Hofstadter 1979). Theorem 2 simply states why we do in that way. As a result of this section, stable objects give SED descriptions where comprehension has taken place, i.e., comprehensive descriptions. The following section is devoted to ensure that the descriptions would give the same meaning out of a sequence, and how to measure their complexity. 4. Testing Comprehesion Ability Theoretically, there are two ways to know whether a system's operation is compliant with some requirements: by inspecting its code (or program) or by testing its behaviour. In general, for complex systems, as it has been nally recognised in software engineering, verication must be experimental, by means of sets of tests. It is an open and hard problem to devise a complete specication of intelligence, mainly because it depends on a consensus on the abilities that an intelligent system must have. However, it is possible to Jose Hernandez-Orallo distinguish some abilities that are fundamental for intelligence. A verication of intelligence behaviour should begin with these fundamental traits, and gradually add more diverse test cases in order to make the test set more ro- bust. Comprehending is the most important trait of intelligence, and we have formalised it in a computational framework. This allows the construction of exercises for a test which are not selected experimentally but theoretically, so, nally, we know what is to be measured, quite unlike psychometrics. However, if we intend to measure comprehensibility there are still two questions to solve. First, we must design unquestionable exercises, in order to avoid the 'subjectivity objection' of IQ tests. Secondly, we require an absolute referent of comprehension diOEculty in order to give a non-Boolean score independent to the mean ability of the subjects or society who have made the test before. Psychometrics has striven to show that it is not absurd to talk about the 'correct' solution. Its rationale is that if the great majority matches with some solution is because there are not alternative solutions of similar com- plexity, and, consequently, it is the most plausible. However, this assertion is made from a very subjective and informal point of view. Let us make formal and objective this idea. At rst sight it seems that stability avoids this but, if we restrict to stable descriptions, we can still modify any explanation p with the addendum iExecute p but print a '1' every hundred symbols that are printedj which would be comprehensive for the data but would dioeer from p in the limit, and would be only a little longer. For this reason, we must introduce the notion of plausibility: Denition 10. Plausibility. A fully projectible description p for a string x is (c; m)-plausible on the right in the descriptional system fi ioe 8d; 0 d Intuitively, a description is (c; m)-plausible if it is at most c bits longer (in LT terms) then the best explanation for x and this holds even if we remove up to m elements from the right of x. From here, we can face unquestionability in the following way: Denition 11. Unquestionability. A fully projectible description p for x is (c; m)-unquestionable in the descriptional system fi ioe it is (c; m)-plausible and there does not exist another (c; m)-plausible description p 0 for x. This is a more restrictive condition as c and m are greater. In order to still obtain some unquestionable descriptions we must make the strings larger. However, as we will see later, if c and m are tuned conveniently for a concrete descriptional mechanism, the tests can still be composed of short strings x such that their SED fi (x) is (c; m)-unquestionable. 6 6 This restriction to unquestionable descriptions not only preserves the goal of the test but even strengthens it, in ontological terms. It has been frequently argued in philosophy of Beyond the Turing Test Once we are able to obtain strings whose SED fi is (c; m)-unquestionable, we should ascertain the diOEculty of each problem, in order to be able to give a test set of exercises of dioeerent comprehensibility. The idea is to relate this diOEculty with explanatory complexity (Et) and the explicitness of the description wrt. the data: Denition 12. A string x is k-incomprehensible given y, denoted by incomp(xjy), in a descriptional system fi ioe k is the least positive integer number such that: Et fi (xjy) \Delta G(SED(xjy)jhx; yi) k \Delta log l(x). The use of the factor logl(x) is to compensate the fact that x must be printed and, therefore, for all x Et(x) log l(x). Consequently, for all x; k 1. As an example, consider a string x of length 256, with 50. The comprehensibility of x is k= 7. The function G corrects the degree of explicitness of the description wrt. the data and it is dened as follows (Hern#ndez-Orallo 1999b): Denition 13. The information gain of an object x wrt. an object y in a descriptional system fi is given by: G fi Denition 12 nally measures the real diOEculty of nding SED(x) from x, because descriptions of the form irepeat x for everj which have Et high (to quote x) are corrected by G, but the length of x is still important. Now we are prepared to construct a generic test of the ability of comprehension by generating a series of strings of gradual comprehensibility. However, as we have said, it is important that the answer is unquestionable, because if not, the answer would be an arbitrary choice from the examiner. A way is to provide redundant information to make the answer unquestionable, with some limitations, obviously, because if not, the problems would be much too long. For instance, given the series ia, c, c, a, c, c, c, a, c, c, c, c, a, .j it seems logical to expect that it would follow ic, c, c, c, c, a, c, .j, so it is redundant to present more than the necessary symbols. The measurement that is to be presented below requires the collaboration of the subject, which must employ all its resources to perform the test. It is not necessary that the subject understands the aim of the test but at least it should be programmed to do it. We can nally obtain the degree of intelligence of a given system as the value which results from applying the following test: Denition 14. C-Test. Let us select a descriptional system fi suOEciently expressive and impartial, composed of an alphabet of symbols\Omega fi and a set of science and induction that the plausibility and unquestionability of a theory or explanation not only depends on the intrinsic characteristics of the explanation but also on the ability of nding alternative explanations. In this sense we can see intelligence as the most important means to augment the plausibility and condence of explanations, and, consequently, the ontology of an 'intelligent' system. 14 Jose Hernandez-Orallo operations \Theta fi to manipulate these symbols, and their corresponding cost (or length). We provide (or programme) to S the alphabet, operations and cost. Depending on the expected intelligence of a system we select a suOEciently wide range 1::K of diOEculty. For each choose randomly p sequences x k;p , being k-incomprehensible, c-plausible, (c,m)-unquestionable and d-stable with d r, r being the number of redundant symbols (or hints) of each exercise. The questions are the K \Delta p sequences without their elements \Gamma(d+r) We give them to S and we ask for the following element according to the best explanation that is able to construct with\Omega fi and \Theta fi . We leave S a xed time t and we record its answers: guess(S; x k \Gammad+r+1 ). The result of this test of comprehensibility or (C-test) is measured as: \Gammad+r+1 the function hit is usually measured as hit(a; (negative values can be used to penalise errors). The value e is simply for weighting the diOEcult questions means that all have the same value). In an informal way, ithe test measures the ability of nding the best explanation for sequences of increasing comprehensibility in a xed timej. 7 5. Measurement of Pretended Intelligent Systems The preceding test is applicable to any system whose degree of intelligence is questioned. Selecting appropriately the descriptional system and the rest of parameters of the test, it can be used for humans, animals, computers, extraterrestrial beings and any collection of the preceding working jointly. Although Denition 14 evaluates a single ability, there are still many ways to realise a specic test. In (Hern#ndez-Orallo and Minaya-Collado 1998) the test was implemented by using an abstract machine quite similar to a state machine. From here, a variety of strings of dioeerent comprehensibility in that machine were generated. Although the set of k-potent numbers of length at most n can be computed in polynomial time in n (see a proof in (Li and Vit#nyi 1997)), the cost of O(n k ), forces to use some heuristics for this. In the same way, G was approximated. Finally, a sieve was applied for obtaining only c-plausible, (c,m)-unquestionable and d-stable sequences. The same work presents the results of applying the test to subjects from species Homo Sapiens Sapiens aged between 14 and jointly 7 One relevant feature of the test is that, although the subject is supposed to be a particular universal descriptional system OE s with a particular background knowledge (life experience) Bs , it is given a descriptional system fi over it, which highly minimises the inAEuence of the dioeerence between the computations performed by OE s and other subject i.e. the dioeerence between Ets(xjhBs ; fii) and Et t (xjhB t ; fii). This makes it possible for the notions of plausibility and unquestionability to be similar for both subjects. Beyond the Turing Test with a classic test of intelligence, the European IQ Test. The correlation between both tests was 0.77. This value only justies a further more exhaustive study over greater groups and several variations derived from Denition 14. Another remarkable experimental result shown in Fig. 1 is that the relation between hit ratio and k-uncomprehensibility is straight, which suggests that comprehensibility really estimates the diOEculty of each string. Figure Hit s @ l Logically, it is not expectable (for the moment) that contrasted and widely used IQ tests would be substituted by these C-tests. Nonetheless, this could be a startpoint towards a theoretical foundation of psychometrics free from the Homo Sapiens as a reference. However, it is not human intelligence but non-human what is urgent to be measured. A formal declaration of what is expected from an intelligent system should allow two important things: to derive more intelligent systems from a more concrete specication and, secondly, to evaluate them. Denition 14 provides a rst step for both things, a detailed scale for measuring the progress (in one intelligence factor) of generic systems in AI. As any other eld of science, a great advance in a discipline happens when one of its topics can be measured in an eoeective and justied way. AI, as a science, requires measurements of intelligence and its dioeerent factors. Modern AI systems are much more functional than systems from the sixties or the seventies. They solve problems in an automated way that before required human intervention. However, these complex problems are solved because a methodical solution is found by the system's designers, not because current systems are more intelligent than preceding ones. We share the opinion that iit is time to begin to distinguish between general, intelligent programs and the special performance systemsj (Nilsson 1995). This initial aim of being more general is nowadays still represented by two subelds of AI: automated reasoning and machine learning. Automated theorem provers are able to solve complex problems from dioeerent elds of mathematics. The great advance of the last two decades is mainly caused by the existence of sets of problems to compare dioeerent systems. Even these sets have evolved and grown to huge and complete libraries of theorem proving problems, like TPTP (Suttner and Sutclioee 1996). Machine learning Jose Hernandez-Orallo is also taking a more experimental character and dioeerent systems (from dioeerent paradigms) are evaluated according to classical problems in the literature. However, as we said in the introduction, there is no theoretical (nor empirical) measurement about the complexity of the problems which compose these test sets. This complexity could be obtained by adapting comprehensibility to several representational languages. 6. Factorisation During this century, psychometrics have striven for dioeerentiating between background knowledge (either evolutionary-acquired or life-acquired) and 'liquid intelligence' (or individual adaptability). Accordingly, exercises from IQ tests are strictly selected to avoid the inAEuence of background knowledge to be foolproof to 'idiots savants'. However, there are many knowledge- independent abilities (or factors) to measure. Some factors usually found in psychological tests are 'verbal ability', `visual ability', 'calculation / deductive ability', etc. The C-test measures one factor, which could empirically be identied with the g factor or liquid intelligence. There are more independent factors which could be measured by using extensions of the framework presented in the previous section. For instance, knowledge applicability, contextualisation and knowledge construction ability can be measured in the following way: Knowledge Applicability (or 'crystallized intelligence'): we provide a background knowledge B and we give a set of sequences x i such that incomp(x i are unquestionable with or without B. We can compare the dioeerence of performance between cases with B and without B. This test would actually measure the application of the background knowledge depending on two parameters: the complexity of B (i.e. Kt(B)) and the necessity or usefulness of B, measured by u. Contextualisation: it is measured in a similar way as knowledge applicability but providing dioeerent contexts dioeerent sequences x i;t such that incomp(x i;t jB t This multiplicity of background knowledge (a new parameter T ) dioeerentiates this factor from the previous one. Analogy tests generally resemble this type of exercises, as it was shown in (Hern#ndez-Orallo and Minaya-Collado 1998) where a denition (and measurement) of analogy was discussed. Knowledge Construction: we provide a set of sequences x i such that exists a common knowledge or context B and a constant u such that for incomp(x i jB) signicant increase of performance must take place between the rst sequence and the later sequences. The parameters are the same as the rst case, the complexity of B and the constant u. This learning from precedents has also been studied in AI (see e.g. Winston 1982). Knowledge Applicability may also be correlated with deductive abilities and these may also correlate with the idea of congruence or coherence, since Beyond the Turing Test it can be measured as constraint satisfaction (Thagard 1989). Other factors are more related with intentionality than intensionality and general intelligence. These are reactivity, pro-activity, interactivity and the recently elsewhere vindicated emotional abilities, that can be measured adopting notions from Query Learning paradigms (Angluin 1988), possibly formalised using interactive Turing machines. However, not every factor is meaningful for intelligence. Factors like iplay- ing chess wellj are much too specic to be robust to background knowledge. Other factors will result in being highly correlated (experimentally or the- oretically) to other more distinct factors. The inAEuence of the descriptional mechanism should also be studied for each factor. In the end, the matter at issue is to rene and extend all the previous ideas in order to make dioeerent and founded tests of intelligence, knowing exactly what is measured. This is an urgent and fascinating task for AI. 7. The C-test and The Turing Test The imitation game was conceived by Turing to dissipate the doubts about possibly non-human intelligent beings. He left no place for human's ex- clusivity: intelligence can be evaluated by an exclusively behavioural test. Unfortunately, instead of recognising this his most important contribution, the test is still understood as 'a goal' in AI. Nonetheless, this view has been responded by many authors, which criticise that the TT does provide little information about what intelligence is; it is just a test of humanity (Fostel 1993), that, in fact, if applied to human beings, yields many paradoxes. The result of applying it to ourselves is a recursive trap which is unable to answer the question of how intelligent the Homo Sapiens is. There have been unsuccessful attempts to correct the two main problems of the Turing Test for measuring intelligence: its informal character and its anthropocentrism. 8 There is still a third problem, which is the necessity of several intelligent 'judges' and a `referent' to make the test. The self-reference question arises again: Who is the rst intelligent being to start the game? These and other problems are incarnated in the Loebner Prize, which usually awards the participant who has devised the system more able to cheat the judge, because ihumans are surprisingly bad at distinguishing humans from computersj (Johnson 1992). Furthermore, there is no way of knowing who is cheating, the system or its designer. However, if fairly played, the imitation game is a hard examination for any pretended intelligent system. It is extremely diOEcult to behave like an In some cases, this has led to disparate proposals, as the so-called formalisation of the TT (Bradford and Wollowski 1995), sustained from the assumption that we are able to solve NP-complete problems in polynomial time. As (McCarthy 1998) claries: ihumans often solve problems in NP-complete domains in times much shorter than is guaranteed by the general algorithms, but can't solve them quickly in generalj. Jose Hernandez-Orallo average human being of this epoch (it is even diOEcult for some average human beings). For a non-human-contextualised being, it would be required to comprehend the complex behaviour of human beings of these times, their evolution-acquired traits, their language, their culture, their limitations, etc. It is much easier then to try to cheat the judges. In fact, the judges iare especially fooled into reading structure into chaos, reading meaning into non- sense.) Sensitivity to subtle patterns in our environment is extremely important to our ability to perceive, learn and communicatej (Shieber 1994). Curiously, it is precisely this 'lack' of the judges what the C-tests measure. However, the C-tests, as they have been presented, are necessary (at least to obtain a minimum value of I(S)) but not suOEcient (other important factors should be measured as well). It has been already suggested that both kind of tests (TT and factorial) could be combined in order to give a more accurate test of intelligence: iit is this posing of puzzles in arbitrary domains that is the hardest part of the Turing Test, and a part that no program has yet passedj (Shapiro 1992). This idea, however, would ultimately turn the TT into a lightweight and less rigorous version of a factorial C-Test. In our opinion, the TT should be celebrated as an extremely valuable philosophical exercise about the behavioural character of intelligence. How- ever, in practice, it should be substituted by progressively more accurate computational tests of dioeerent cognitive abilities. 8. Conclusions Turing devised a way for identifying intelligent beings from non-intelligent ones without solving the problem of what intelligence is. In fact, an imitation game is the only way to make sense from such an apparent paradox. However, the approach has shown numerous limitations and troubles which make it useless for AI in practice. With current theoretical and technical tools of computer science, it is diOEcult to develop non-human intelligence without a computational formalisation of the problem we are trying to solve. It is high time to address the fundamental problem: what intelligence is. This paper presents an important step in this line. A formalisation of one of the main factors of intelligence, the g factor or liquid intelligence is dened computationally. This denition has been used to develop an intelligence test, very dioeerent from the TT and in compliance with classical IQ tests. Like the latter it distinguishes acquired knowledge from liquid intelligence. More importantly, the C-Test, unlike the TT and IQ tests, is not anthropomorphic. The factor is dened as the ability of nding comprehensive explanations, and thus is meaningful. This makes it philosophically acceptable: intelligence is what allows us to comprehend the world. Sooner or later we will need to face the fact that computers will be closer and closer to human intelligence. Once arrived to this hallmark of Beyond the Turing Test 19 AI, the misunderstood goal of the TT, it will be indispensable to have an objective measure of intelligence, in order to solve the incipient technical and ethical problems that could be derived from here. The paradigm presented in this paper, since it is theoretically justied, allows the projection of the measurement of intelligence beyond human intelligence. There are many more interesting questions beyond the TT. How many independent computational factors does human intelligence have? How intelligent does the Homo Sapiens result in the end? Which main factors make a chimpanzee signicantly dioeerent from us? How intelligent might machines be with the current computational power? Psychometrics, Anthropology, Zoology and AI have only partially dealt with some of these problems. All these questions require new theoretical tools for a radical change of paradigm: a science of intelligence grounded in theoretical computer science and information theory. To construct this science, Turing's call is now more compelling then ever: iWe can only see a short distance ahead, but we can see plenty there that needs to be donej. Acknowledgements I am much obliged to the main inspirers of this work, G. Chaitin and D. Hofstadter, for their encouraging comments, especially during 1997, when the main ideas were taking shape. Since then, the work has matured with the help of numerous collaborations and suggestions: R. Beneyto and J.M. Lorente from the Dep. of Logic and Phil. of Science of the U. of Valencia, P. Thagard from the Dep. of Phil. of the U. of Waterloo, O. Pellicer from the Dep. of Psychobiology of the U. of Valencia and N.T. Crook from the Oxford Brookes U. Finally, I am especially grateful for the comments from K. Araque, R. Barreiro, E. Fueyo, I. Garc#a, E. Hern#ndez, N. Minaya and I. Soto on several drafts of this paper. --R Machine Learning The Economist Logical depth and physical complexity. The Methods of Contemporary Thought. A Formalization of the Turing Test (The Turing Test as an Interactive Proof System). What kind of Information Processing is Intelligence? Programs in the Search for Intelligent Machines: The Mistaken Foundations of AI. Can Machines Think? A Heuristic Program to Solve Geometric Analogy Problems. The Structure and Measurement of Intelligence. Language Identication in the Limit. The inference to the best explanation. The Turing Test Is Not a Trick: Turing Indistinguishability Is A Scientic Criterion. The universal Turing machine: a half-century survey of Intelligent Systems (EIS'98) A New Test of Intelligence. In (Herken The Turing Test Misunderstood. Universal search problems. Machines and Thought. Knowns and Unknowns. Eye on the Prize. Fisher information and stochastic complexity. What is AI The Turing Test and The Economist. Lessons from a Restricted Turing Test. Human acquisition of concepts for sequential patterns. A formal theory of inductive inference. 'General Intelligence' objectively determined and measured. Beyond Information. The TPTP Problem Library. On computable numbers with an application to the Entscheidungsproblem. Computing Machinery and Intelligence. A theory of the learnable. Pattern Recognition as Information Compression. Thinking Machines: Can There Be? Learning New Principles from Precedents and Exercises. Thinking Machines: Can There Be? --TR --CTR Shane Legg , Marcus Hutter, Universal Intelligence: A Definition of Machine Intelligence, Minds and Machines, v.17 n.4, p.391-444, December 2007
comprehension;measurement of intelligence;descriptional complexity;AI's anthropomorphism;inductive Inference;turing test;psychometrics
596040
The Constraint Language for Lambda Structures.
This paper presents the Constraint Language for Lambda Structures (CLLS), a first-order language for semantic underspecification that conservatively extends dominance constraints. It is interpreted over lambda structures, tree-like structures that encode -terms. Based on CLLS, we present an underspecified, uniform analysis of scope, ellipsis, anaphora, and their interactions. CLLS solves a variable capturing problem that is omnipresent in scope underspecification and can be processed efficiently.
Introduction Underspecication is a recent approach to the issue of combinatorial explosion caused by ambiguity. Its key idea is to derive a single, compact description of all readings instead of the (exponential number of) readings themselves. For as long as possible, later processing steps operate on these descriptions; readings are only enumerated by need. One type of ambiguity for which underspecication has been investigated particularly well is scope ambiguity. Common to many recent scope underspecication formalisms (Reyle, 1993; Muskens, 1995; Bos, 1996; Niehren et al., 1997b; Egg et al., 1998) is that they provide the 'semantic material' of which the description of the readings is built, plus structural relations that constrain possible combinations of this mate- rial. Dominance constraints, a tree logic which can express dominance of nodes in a tree, are a very natural way for doing this. Muskens (1995) was the rst to do this explicitly, but Reyle (1993) and Bos (1996) can be subsumed under this kind of approach as well. In this article, we show how to integrate a treatment of other linguistic phenomena { in particular, anaphora and VP ellipsis { with a scope underspecication formalism based on dominance constraints. The interaction of these three phenomena has been considered before in Crouch (1995) and Shieber et al. (1996). In our approach, it is modelled by a conservative extension of the language of dominance constraints c 2000 Kluwer Academic Publishers. Printed in the Netherlands. Markus Egg, Alexander Koller, Joachim Niehren by additional atomic formulas, viz., anaphoric binding constraints and parallelism constraints. While the former are based on an idea by Kehler (1993), the latter are formally equivalent to context unica- tion as used in Niehren et al. (1997b), which was shown in Niehren and Koller (1998). In addition, -binding constraints are introduced in order to solve a capturing problem that appears in all dominance-based underspecication formalisms. The combination of all of these constraints, called CLLS (Constraint Language for Lambda Structures), correctly describes many challenging cases from the literature that exhibit interaction of scope, anaphora, and ellipsis. In addition, its being an extension of dominance constraints makes for some appealing dierences to existing formalisms with comparable coverage: Unlike Shieber et al. (1996), the CLLS analysis is completely order-independent; and in contrast to the QLF- based analysis in Crouch (1995), it achieves a clean separation between underspecied descriptions and the described readings. The paper is structured as follows. In Section 2, we introduce the phenomena under consideration and present the intuitions behind our linguistic analysis without going into detail. Section 3 presents the formal denitions of lambda structures and CLLS and lls in the formal details that were omitted in Section 2. In Section 4, we extend the analysis to more involved cases from the literature. Section 5 denes a syntax/semantics interface that derives CLLS constraints on the basis of a simple phrase structure syntax. In Section 6, we give a brief overview over results on formal and computational aspects of CLLS. Finally, Section 7 concludes and outlines directions for future work. 2. Elements of CLLS This section introduces the three main phenomena considered in the paper { scope, ellipsis, and anaphora { and the major concepts used in their CLLS analyses. After discussing -structures and their correspondence to -terms, we go through the phenomena and present the intuitions underlying our analysis. The section does not even attempt to be formally precise; the formal details are put section 3. 2.1. Lambda structures and lambda terms The idea behind -structures is to represent a -term by a tree-like graph; a node in a -structure corresponds to an occurrence of a sub-term in the -term it represents. Consider e.g. the -term in (1), which happens to represent one of the readings of (3) below: The Constraint Language for Lambda Structures 3 (1) (a workshop)(x (every linguist)(y (attend We draw the corresponding -structure as the graph in (2). An occurrence of a -abstraction corresponds to a node labelled by lam, an occurrence of an application, to a node labelled by @, and an occurrence of a bound variable, to a node labelled by var. Variable binding is represented in a lambda structure by dashed arrows from the variables to their binders. For instance, if we identify each node by its address, i.e. the word of natural numbers that indicates the path from the root to the node, the unique occurrence of x in (1) corresponds to the node with the address 212 in (2); it binds the variable represented by the node with the address 21212. @ a workshop lam @ @ every linguist lam 212 @ @ visit var var 21212 Note that variable names in a -term (such as x and y in (1)) are not re ected by the node labels of the corresponding -structure. Con- sequently, a -structure represents a closed -term uniquely modulo consistent renaming of bound variables (-equality). Via the detour through -structures, we can use tree descriptions in semantic underspecication: We simply talk about -structures in the same way one usually talks about trees. The explicit binding relation is important in that context because variable names may not be sucient to ensure proper binding in a radically underspecied framework; we will come back to this in a minute. 2.2. Scope The rst class of phenomena we are concerned with is scope ambiguity, as in the following sentence. (3) Every linguist attends a workshop. This sentence has two readings, which dier in a permutation of quantiers only: Either the sentence means that there is one specic workshop which every linguist attends, or it means that every linguist attends a workshop, but not necessarily the same as the others. The rst reading is represented by the -term in (1), and the second by (4): 4 Markus Egg, Alexander Koller, Joachim Niehren (every linguist)(y (a workshop)(x (attend Closer inspection reveals that (1) and (4), and hence their corresponding lambda structures, are closely related. The lambda structures are composed from the same tree-like subgraphs, corresponding to the term fragments (every linguist)(x(:)), (a workshop)(y(:)), and ((attend x) y). But these fragments are composed in dierent order. We can describe both lambda structures at once by specifying the fragments and their relationships. Such a description is given in (5). a workshop lam X 3 @ attend var every linguist lam X 7 Intuitively, description (5) is satised by all -structures into which the graph in (5) can be embedded in such a way that no labelled nodes of the graph overlap. Dotted edges in (5) signify dominance: Of the two nodes they connect, the upper one must be above the lower one in the lambda structure. The dashed arrows, for -binding, act like elastic bands, which can be stretched indenitely without breaking. The description leaves the relative ordering between the two quan- tier fragments unspecied. But since both fragments dominate the nuclear scope and, just like trees, -structures cannot branch in the bottom-up direction, one of the two quantier fragments must dominate the other. This situation is very common in scope underspecication. Note that variable binders must always dominate their bound variables; so the -binding relations, logically entail the dominance relations, X 4 X 9 and X 8 X 9 . Inserting these dominance edges in (5) only enhances readability. In Section 3 we will give pictures like (5) a formal meaning as constraints over -structures, i.e. logical formulas that are interpreted over the class of -structures. Then nodes of the graph correspond to variables denoting nodes of a -structure, whereas labels, edges, and arrows correspond to various types of atomic constraints. In addition, such a graphical description imposes implicit inequality constraints between labelled nodes, which excludes proper overlap of fragments. We have said that a description like (5) is satised by all lambda structures into which they can be embedded. In particular, the lambda The Constraint Language for Lambda Structures 5 structure may contain additional fragments that are not mentioned in the description. This is a desired feature because we want to compute with limited information that can in principle be augmented at any later point; the fact that the information is complete must be expressed independently. As long as one is only concerned with scope, one can impose a general constructiveness restriction: each node in a lambda structure that we describe must be denoted by a variable in the de- scription. But constructiveness would be too strict for our treatment of ellipses below and also for an underspecied account of reinterpretation (Koller et al., 1999; Egg, 1999). 2.3. Ellipsis The second phenomenon we treat is VP ellipsis. VP ellipsis interacts with scope ambiguities, thus the treatment of both phenomena must be integrated. In the framework presented in this article, this integration simply amounts to adding further constraints on lambda structures. For ellipses, we introduce parallelism constraints, which permit to express structural equality between fragments of -structures. As an example, consider the elliptical sentence (6): Every man sleeps, and so does Mary. The meaning of the target sentence so does Mary is exactly like the meaning of the source sentence every man sleeps, except that the source parallel element every man is replaced by the target parallel element Mary. This can be modelled in two steps. First, we describe the -structure for the source sentence, as in (7), and the -structure for the target parallel element, as in (8). every man lam sleep var Mary Y 2 lam var In any lambda structure described by these constraints, the semantics for the source sentence will be below X 1 , and the semantics for the target sentence will be below Y 1 . The -structure for the entire sentence will be the conjunction of source and target sentence, like this: The second step is to impose the parallelism constraint (10), which says that the lambda structures for the source and target sentence (i.e. 6 Markus Egg, Alexander Koller, Joachim Niehren below X 0 and Y 0 ) are equal, with the exception of the sub-lambda structures representing the parallel elements (below X 2 and Y 2 Indeed, the intended semantic representation of (6), shown in (11), simultaneously solves all constraints in (7), (8), (9), and (10). @ @ every man lam @ sleep var @ Mary lam @ sleep var This shows that ellipsis resolution can be reduced to solving constraints over -structures. Since proper names and quantier NPs can be parallel elements in an ellipsis (as in (6)), our analysis requires type-raising of proper names; the constant Mary in (8) corresponds to a -term of type hhe; ti; ti. Where there is no ellipsis, proper names are freely analyzed as constants of type e, as this simplies the presentation. To distinguish the two analyses of proper names, we will capitalize the rst letter of label used in the former; that is, the label for the type-raised analysis of proper nouns is Mary and the one for the non-type-raised analysis, mary. 2.4. Anaphora Anaphora are the nal class of phenomena we consider. In order to represent anaphora, we extend our notion of -structures. Anaphora are represented as nodes labelled by ana which are linked to their antecedents by anaphoric links. We draw anaphoric links as dashed arrows using straight lines. For instance, a representation of (12) is shown in (13). For simplicity, we use a constant mother of here; we will analyze the possessive dierently later. liked his j mother. @ say @ @ like @ mother of ana X 3 In the example, the two anaphoric links form a link chain. Link chains go back to Kehler (1993, 1995); they have been worked out in The Constraint Language for Lambda Structures 7 this setting by Xu (1998). This representation of anaphoric reference can cope with a series of problematic cases from the literature, such as strict/sloppy ambiguities, the many-pronouns puzzle (of which (12) is the source clause), and cascaded ellipses. 2.5. The Capturing Problem A major dierence between CLLS and other underspecication formalisms is that variable binding is modelled by explicit -binding arrows and not by variable names. This looks like an unnecessary deviation from standard procedure, but it turns out that variable names are inconvenient in the context of underspecication and become completely unusable once we add parallelism. Let us assume that variable binding is modelled by using variable names; for instance, that we use new labels lam x and var x for each object-level variable x. Now imagine that a binder lam x appears twice in a description, and that there is a scope ambiguity between the two occurrences. In this case, it would be unclear which the intended binder of a node with label var x is; in each solution, the lower binder wins. The obvious way to avoid this problem is to name all variables apart when generating a description. But this restriction cannot be maintained once parallelism comes into play. This is best illustrated by corrections, as in the following example. Hans knows every student. No, Manfred knows every student. (15) is an underspecied description of the meaning of (14): correct @ @ every student lam x X 1 @ @ know var x X 3 hans X 2 @ @ every student lam x Y 1 @ @ know var x Y 3 manfred Y 2 In the analysis, we used the same variable name x in both binders. This is necessary since in solving the parallelism constraint, we must make X 1 correspond to Y 1 . If these two nodes carried dierent variable names, we would have to soften the meaning of parallelism constraints in that the labels of corresponding nodes need not be identical, but only uniable in some sense; but this raises very dicult problems of 8 Markus Egg, Alexander Koller, Joachim Niehren bookkeeping which variables have been identied with which. The only feasible option is using the same variable name twice. There are other problems, too. We have seen in Section 2.3 that we cannot even maintain the constructiveness restriction on satisfying lambda structures; parallelism may force us to copy material. But if we do not require constructiveness, we can get (16) as a solution for correct @ every student lam x X 1 @ @ every student lam x Y 0@ @ know var x X 3 hans X 2 @ every student lam x X 0@ @ every student lam x Y 1 @ @ know var x Y 3 manfred Y 2 More precisely, the problem is that parallelism (not only for correc- tions) may force us to copy binders, and if we cannot always rename variables apart, we must enforce that copying variables does not break the intended binding relations (as the one between X 1 and X 3 in (16)). The problems with variable binding are real problems in the presence of parallelism only; they can be avoided otherwise by naming variables apart and requiring constructiveness, both of which is still slightly un- pleasant. Modelling variable binding with an explicit binding relation between nodes oers a clean solution to all of these capturing problems: how much and what kind of material is added between a variable and its binder, the binding relation can never be broken. 3. Syntax and Semantics of CLLS Section 2 presented the general intuitions behind lambda structures. We showed how to describe -structures and how to apply their descriptions in an underspecied analysis of scope, ellipsis, and anaphora. Now we give a rigorous denition of the constraint language for lambda structure (CLLS), which we will use throughout the paper. We will start with the semantic level by rst presenting tree structures, then move over to lambda structures as an extension of tree structures. Once the semantics is xed, a denition of a syntax to talk about it is simple. The Constraint Language for Lambda Structures 9 3.1. Tree Structures A tree corresponds to a unique tree structure in the same way that a lambda structure relates to a unique -term modulo -equivalence. Lambda structures are extended tree structures because -terms are extended trees (trees plus variable binding). Let be a set of function symbols, f; g; a; b. Each function f symbol comes with a xed arity. We write f k for a function symbol with arity k 0. A (nite constructor) tree is simply a ground term (without variables) built from a set of function symbols. For instance, f(g(a; a)) is a tree built from function symbols f Taking IN to be the set of natural numbers starting at 1, we can dene a path as a word over IN . We identify each node of a tree with the path from the root to this node; the empty word, , identies the root. Paths are named with lowercase Greek letters, such as Concatenation of words and is written as , and a word is a prex of i there is a word 1 such that A tree structure is a rst-order model structure which characterizes a tree. The domain of a tree structure contains all nodes of the corresponding tree, and its relations specify labelling and childhood. E.g., the domain of the tree structure for f(g(a; a)) is f; 12g. The fact that the root is labelled with f and has a single child 1 is expressed by a relation we write as :f(1); the children and labels of the remaining nodes are specied by the relationships 1:g(11; 12), 11:a, and 12:a. More generally, a tree domain is a nite nonempty set of nodes which is prexed-closed (that is, 2 implies 2 ) and closed under left siblings (i.e. if i 2 then j 2 for all 1 j < i), and a tree structure is dened as follows: DEFINITION 3.1. A tree structure over consists of a tree domain IN and a relation :f n+1 for each n-ary function symbol , such that for all nodes 2 and the maximal n 2 IN with there is a unique symbol f n 2 such that Given nodes This relation expresses that 0 is a node of the tree structure with the label f and the children , in this order. Note that two nodes may be labelled by the same constant without being equal. The condition in the denition means that the number of children of a node is specied by the arity of its label. Markus Egg, Alexander Koller, Joachim Niehren 3.2. Lambda Structures A -structure extends a tree structure with two additional relations for -binding and anaphoric links. For -structures, we assume: DEFINITION 3.2. A -structure over consists of a tree structure over with tree domain and two additional binary relations and ante on which satisfy the following properties. 1. is a partial function that maps every node 2 with label var to a prex of which is labelled with lam. 2. ante is a partial function that maps every node with label ana to another node. The lambda structures we are interested in here have additional properties, which we have not included in the denition because they are not essential for what follows. For instance, there are no cyclic anaphoric link chains, and if a link chain ends at a node with label var, then each node in the chain (all of which are labelled by ana) must be below the binder of that var-labelled node. We draw lambda structures as tree-like graphs. E.g., the binding and anaphoric linking functions of the lambda structure (17) are specied by 1. Note that we annotate only those nodes with addresses in the picture that we want to refer to. @ @ like ana 2112 var 212 Lambda structures and -terms can almost be mapped to each other one-to-one. The two exceptions are that lambda structures do not mention variable names and that lambda structures contain anaphoric links. We do not elaborate on the further interpretation of anaphoric links roughly, they express that the interpretation of the last element of the link chain determines the interpretation of all other elements of the chain. The Constraint Language for Lambda Structures 11 3.3. Dominance and Parallelism Finally, we conservatively extend lambda structures by two additional relations between nodes: dominance and parallelism. These relations were not included in the original denition as they can be dened from the relations specied there; but they are essential for the application. Let be a -structure and ; two of its nodes. We say that dominates ( ) if it lies above it, i.e. if is a prex of . The dominance relation is a partial order on the domain of , i.e. re exive, transitive, and antisymmetric. For the denition of parallelism, we rst dene for any two nodes in the context of below in as the set of nodes of which are dominated by , but not by . Intuitively, the context of below is the set of nodes 'between' and . Taking the labelling relations of into account, a context looks almost like a tree domain starting at , except that one node (the mother of ) has one too few children. Based on this intuition, we call the root of the context and its hole. Intuitively, the parallelism holds for a -structure if imposes parallel structures on the contexts between 1 and 1 and between 2 and 2 . As it must also specify what it means for binding and anaphoric linking to be parallel, however, the following denition of the relation is rather involved. DEFINITION 3.3. Let be a -structure over and 1 nodes of its domain . Parallelism holds in i there are subsets 1 ; 2 and a bijective correspondence function c : such that P1{P6 are satised. P1 For 2g, the set i is the context i below i in : and not i g P2 imposes parallel tree relations on 1 and In particular, this enforces that c( 1 (that is, the holes correspond to each other) and that for all 2 IN , 1 2 1 [f 1 g ) (that is, corresponding nodes can be reached over the same path from the roots of their respective contexts). P3 Binding within 1 and 2 is parallel: Markus Egg, Alexander Koller, Joachim Niehren P4 For i 2 f1; 2g, no node below i is bound by a node inside P5 If binds a variable node within 1 properly above 1 or the corresponding variable node in 2 properly above 2 , then the two binder nodes are binding equivalent ('/ proper dominance). For the time being, we assume that binding equivalence is simply the node identity. We will modify this denition in Section 4.5. P6 If is an anaphoric node in 1 , then the corresponding anaphoric node in 2 must be linked either to (strict) or, if it exists, to the correspondence of ante() (sloppy). P1 species the parallel contexts and the correspondence function between them. P2 requires parallelism of tree structure. Conditions P3{P5 are concerned with the interaction of parallelism and variable binding. Condition P6 deals with the interaction of parallelism and anaphora, i.e. with strict/sloppy ambiguities. The parallelism relation is almost symmetric; the only exception is that the strict variant of the rule P6 only allows to link the anaphoric node in 2 to the corresponding node in 1 , but not vice versa. 3.4. The Constraint Language CLLS Now that -structures and their relations are dened, it is straight-forward to x a constraint language to go with them. We assume an innite set of node variables, ranged over by X;X pick relation symbols for all relations dened so far. To keep the presentation simple, we use the same symbols as for the relations themselves. The constraint language over lambda structures (CLLS) is given by the following abstract syntax: The Constraint Language for Lambda Structures 13 The semantics of CLLS is dened by interpretation of constraints over the class of lambda structures in the usual Tarski style. A constraint ' of CLLS can be evaluated to a truth value if a -structure and a variable assignment into the domain of are given. A solution of ' is a pair (; ) which satises '. 3.5. Constraint Graphs CLLS constraints as just dened can become unreadable rather quickly. Throughout the paper, we will use constraint graphs instead. Constraint graphs are an alternative, graphical syntax for CLLS; a constraint graph corresponding to a constraint ' contains one node for each variable in ' and labels and edges for the atomic constraints that appear in '. A simple example is given in (18). lam X var Z represents For every dominance constraint X Y , we draw a dotted line from X to Y . For every labelling constraint X:f(X the graph node for X with f and draw solid lines to the nodes for . For every -binding constraint (X)=Y , we draw a curved dashed arrow from X to Y ; and for every anaphoric binding constraint ante(X)=Y , we draw a dashed arrow from X to Y that only uses straight lines. Parallelism constraints cannot be represented easily in constraint graphs; we will simply write them next to the graph. Finally, a constraint graph represents inequality constraints X 1 6=X 2 for each two labelled nodes in the graph. This assumption nicely prevents 'overlaps' between the `fragments' of the constraint graph. Now we can interpret the intuitive diagrams from Section 2 as CLLS constraint graphs. This gives them a formal meaning as CLLS con- straints, which can be satised by lambda structures. These lambda structures, in turn, represent -terms which stand for the actual meaning(s) of the sentence. 4. Interaction of Quantiers, Anaphora, and Ellipsis Now that we know how to deal with simple examples, we turn to some well-known, more complex challenges in the area of quantiers, anaphora, and ellipses, in particular to involving their interactions. We look into quantier parallelism, strict/sloppy ambiguities, nested 14 Markus Egg, Alexander Koller, Joachim Niehren ellipses, a sentence involving interaction of all three, and antecedent- contained deletion (ACD). There are other interesting things that the analysis can do; e.g., to deal with gapping, parallelism could be extended to allow two exception positions rather than one: Niehren et al. (1997a) showed that this can be done in a dierent setting. Throughout, we assume a xed signature: We follow the convention that proper nouns are always analyzed as constants of type e, except as contrasting elements in ellipses where the other contrasting element is a quantier. We indicate the rst analysis by labelling the corresponding nodes in lowercase letters (john), the second one, by using labels with a capitalized rst letter (John). Fur- thermore, we draw dominance edges wherever dominance is entailed by -binding constraints. This is logically redundant, but may improve readability. 4.1. Quantifier Parallelism The rst phenomenon we consider is quantier parallelism, an interaction of scope and ellipsis. By way of example, we continue the sentence (3), repeated here as (19), with the elliptical sentence (20). Every linguist attends a workshop. Every computer scientist does, too. This pair of sentences, modelled after an example from Hirschbuhler (1982), has three readings: Either there is a single workshop which both all linguists and all computer scientist attend; or there is one workshop for all linguists and one for all computer scientists (but not necessarily the same one); or every linguist and computer scientist can attend a dierent workshop. This is surprising; if the target sentence means 'Every computer scientist attends a workshop', each sentence is two-way ambiguous, for a total of at least four readings. In CLLS, parallelism enforces that the quantier scopes in both sentences must be analogous; the ambiguities cannot be resolved inde- pendently. This is re ected by the following constraint (abbreviating workshop to ws, linguist to ling, and computer scientist to cs): The Constraint Language for Lambda Structures 15 @ a ws lam s every ling lam every cs @ @ attend var var s X s is the root of the semantics of the source sentence; X t is the root of the target sentence. The contrasting elements, every linguist and every computer scientist, are below the variables X 0 s tively. There is a dominance constraint between X s and X 2 , but not between X s and X 1 , since, like most quantiers, universal quantiers must obey a 'scope island constraint' that prevents them from escaping from a sentence. This restriction does not hold for indenites, such as a workshop. There are three places where the fragment below X 1 can go in a lambda structure satisfying the constraint: below the fragment starting at above the conjunction. In each case, the parallelism constraint enforces that whatever is below X s is also below replacing what is below X 0 s with what is below X 0 . This produces exactly the three readings of the sentence. Note that in the third reading, the indenite takes scope over the entire conjunction; it is not part of the parallel contexts. A particular consequence of the parallelism constraint enforcing structural parallelism of the sub-lambda structures below X s and X t is that the relative scopes of the quantiers must be identical in both the source and the target sentence (if the indenite is below X s , of course). 4.2. Strict/sloppy Ambiguities Strict/sloppy ambiguities are ambiguities of anaphoric reference that occur in the context of VP ellipses. They are an important benchmark for analyses of ellipsis. The prototypical example looks as follows: likes his mother, and Bill does too. This sentence has two readings: Either Bill likes John's mother (this is the strict reading), or Bill likes his own mother (the sloppy reading). We describe the meaning of (22) as in (23), using parallelism and anaphoric linking constraints. We have hidden the sub-lambda struc- Markus Egg, Alexander Koller, Joachim Niehren ture for his mother and only drawn an empty triangle, as the only node we are interested in is the anaphor. ana X a lam @ @ like var s bill X 0 Except for the fact that the source clause contains an anaphor, this constraint is totally analogous to the simple ellipses we considered above. The parallelism constraint rst enforces that the tree part of the lambda structure below X t is the same as the tree part of the one below X s , except for the contrasting elements, as follows: a lam @ @ like var bill X 0 (24) is not a complete lambda structure because the anaphor at X 0 a does not have an antecedent. According to P6 of Denition 3.3, there are two legal antecedents that this anaphor can be linked to. It can either be linked to the anaphor in the source context, i.e. to the node denoted by X a . In this case, the link chain starting at X 0 a ends at X s , which is labelled with john; so this is the strict reading. Or it can be linked to the node corresponding to the antecedent of X a . As X a is linked to the hole of the source context, X 0 a must be linked to the hole of the target context; this is X 0 , which produces the sloppy reading. 4.3. Nested Ellipses This analysis of strict/sloppy ambiguity carries over to more sophisticated cases like the Gawron and Peters (1990) example revised his paper before the teacher did, and so did Bill. This sentence comprises nested ellipsis: the source clause John revised his paper before the teacher did of the ellipsis is elliptical itself; ellipsis resolution is made even more interesting by the presence of an anaphor. We follow Dalrymple et al. (1991) in assuming ve readings (listed in (27)) for (25). These ve readings are described by the constraint in (26). The Constraint Language for Lambda Structures 17 @ ana X a lam @ s s lam before the teacher Bill X 0 @ @ revise var var s s For readability, we again compressed the subconstraint describing his paper into an empty triangle; the only visible (and interesting) node is X a , the anaphor. The constraint contains two parallelism constraints, one for each ellipsis. Note that the source sentence seems to contain a scope ambiguity between the three fragments corresponding to his pa- per, John, and before. However, the constraint entails that before takes scope over John because Y s =Y 0 s . The remaining scope ambiguity between his paper and the other two fragments is spurious; we chose to leave it in the constraint to be closer to what the syntax/semantics interface (Section 5) would produce. To understand how the ve (constructive) solutions come about, let us try to 'solve' the constraint by mimicking rst resolution of the inner and then of the outer ellipsis. An actual solution algorithm might behave dierently; in particular, it could interleave resolution steps for the inner and outer parallelism constraint and other constraints such as the dominance constraints for a scope ambiguity, as the formalism is totally declarative. We call the correspondence function on the parallel contexts that is induced by the inner parallelism constraint c i , and that of the outer one, c Resolution of the parallelism constraint for the inner ellipsis will make the description of the sentence the teacher revised his paper more explicit. Let us say that the node denoted by the variable for the original anaphor is a ; then its copy via this parallelism constraint is c i ( a ). There are two options where c i ( a ) can be linked: either to a (strict { this produces the rst three readings in (27)) or to 0 , the denotation of Y 0 (sloppy { this produces the other two readings). If we now resolve the parallelism constraint for the outer ellipsis, we produce copies of a and of c i ( a )). The node a ) can be linked either to a (strict) or to 0 , the denotation of X 0 Markus Egg, Alexander Koller, Joachim Niehren (sloppy). Likewise, c can be linked either to c i ( a ) (strict), or it can be linked to the respective value of c So there are ve possible combinations of linking relations: (27) 1. ante(c i (a))=a , ante(co(a))=a , ante(co(c i (a)))=c i (a) and John revised John's paper before the teacher revised John's paper, Bill revised John's paper before the teacher revised John's paper 2. ante(c i (a))=a , ante(co(a))= 0 John revised John's paper before the teacher revised John's paper, Bill revised Bill's paper before the teacher revised John's paper 3. ante(c i (a))=a , ante(co(a))= 0 John revised John's paper before the teacher revised John's paper, Bill revised Bill's paper before the teacher revised Bill's paper 4. ante(c i (a))= 0 John revised John's paper before the teacher revised the teacher's paper, Bill revised John's paper before the teacher revised the teacher's paper 5. ante(c i (a))= 0 John revised John's paper before the teacher revised the teacher's paper, Bill revised Bill's paper before the teacher revised the teacher's paper A reading where a , c i ( a ), and c are all linked to John and is linked to Bill, which is particularly dicult to exclude in other approaches, is impossible here, as c must be linked either to c i ( a ) or to c 4.4. A Complex Interaction Now we consider an example involving quantication, anaphora, and ellipsis where all three phenomena interact. Mary read a book she liked before Sue did. This sentence, modelled after a Gawron and Peters (1990) example, has three readings. In the rst, the indenite NP a book she liked takes wide scope over both clauses (a particular book liked by Mary is read by both Mary and Sue). The second and third arise from the strict/sloppy ambiguity that occurs if the operator before outscopes the indenite: either both read a book that Mary liked, or each read a book she liked herself. A constraint describing these three readings is shown in (29). The Constraint Language for Lambda Structures 19 lam before X 2 @ @ read var mary X 0 s sue X 0 For the sake of readability, we once again simplify the semantics of a book she liked to a mere empty triangle with the anaphoric node X 3 . The rst reading, with the NP taking wide scope, results when the relative scope of X 1 and X 2 is resolved such that X 1 dominates X 2 . In this case, the anaphoric node is not part of the parallel subdomain, which means that there is no strict/sloppy ambiguity. A schematic representation of this solution (in which we have compressed even more sub-lambda structures into empty triangles) is shown in (30). before X 2 mary X 0 s sue X 0 The other possible scoping is for X 2 to dominate X 1 . The two solutions this gives rise to are displayed in (31) and (32). Now the interpretation of the indenite, below X 1 , is dominated by X s and is copied by resolving the parallelism constraint. In particular, the anaphoric node, denoted by X 3 , is copied; the strict/sloppy ambiguity of its copy is dealt with in the now familiar way. mary X 0 s ana sue X 0 mary X 0 s ana sue X 0 Markus Egg, Alexander Koller, Joachim Niehren 4.5. Antecedent-Contained Ellipsis We conclude this section with a notoriously dicult case of ellipsis, antecedent-contained deletion (ACD), as in (33). John greeted every person that Max did. The problem here is is that the ellipsis is contained in the VP that it refers to (its 'antecedent'). A naive approach to ellipsis that simply copies phonetic material from the source to the target sentence runs into an innite loop as there is always an elliptic sub-sentence left. Generative syntax solves this problem by raising the universal quan- tier out of the sentence in LF (Sag, 1980; Fiengo and May, 1994); this has the eect that the ellipsis is no longer contained in the copied material. Dalrymple et al. (1991) and Crouch (1995) avoid a similar kind of recursion by an 'occurs check' of higher-order unication. In CLLS, the meaning of (33) can be described as follows. @ every lam X 4 @ person var @ lam X 2 @ var lam X 1 @ @ greet var X 9 Like in the other analyses, the universal quantier is not contained in the source context of the parallelism constraint. 1 As already discussed in Section 4.3, the constraint also entails that the universal quantier must take scope over John. There is one subtle problem, however. Condition P5 in Denition 3.3 requires that if is labelled with var, the binders of and c() must be binding equivalent, which we have taken to mean 'equal' so far. However, ACD forces one to soften this restriction: The var-node 3 is bound by X 2 , but X 2 cannot bind the copy of X 3 because it does not dominate it. Intuitively, it should be permitted for the copy (X 9 ) to be bound by X 1 because if it is, a sequence of -reductions will map 1 If the universal quantier was contained in the context, the only solutions of the constraint would be innite lambda structures. This means that it would be unsatisable according to Section 3, as we are interpreting CLLS over nite lambda structures. This is reminiscent of the occurs check, as the function of an occurs check is to avoid innite solutions to unication problems. The Constraint Language for Lambda Structures 21 its copy to the same variable. The separation of the two binders is merely an artifact of the type-raised analysis of every. However, checking for -equivalence while processing the constraint is unfeasible: It is computationally expensive, and it is unclear if it can be done at all before all readings have been enumerated. Instead, we use the following revised denition of binding equivalence, which is easier to check and works for the examples: DEFINITION 4.1. Let be a lambda structure. Binding equivalence in is the smallest equivalence relation on nodes in the domain of such that for all 1. if 1 :lam( 6 where Q is a quantier label, i.e. a member of 2. if 3 The rst branch models -equivalence; the precondition is shown in (35a). The second branch models -equivalence; its precondition is shown in (35b). If we modify our denition of the parallelism relation in this way, the constraint in (34) has exactly the correct solution. @ 4 lam 2 (b) lam 1 @ 3 lam 2 var 4 Note, by way of conclusion, that our analysis can also account for the dierence between (36) and (37) (Sag, 1980), the rst of which lacks one of the two readings of the second one (where the matrix verb want outscopes the universal quantier). wants Bill to read everything that Max does. wants Bill to read everything Max wants him to read. This blocking is achieved in the CLLS analysis because, as discussed above, John must be in the scope of the universal quantier (it is the contrasting element). As every NP argument outscopes the verb of which it is a syntactic argument, John must outscope want ; so the relative scope of the universal quantier and want is xed. If the elliptic material of (36) is spelt out like in (37), there is no parallelism constraint to enforce wide scope of the universal quantier over John. 22 Markus Egg, Alexander Koller, Joachim Niehren 5. The Syntax-Semantics Interface This section denes a grammar with a syntax/semantics interface which allows the derivation of the CLLS constraints for the example sentences discussed in this paper. As we have used a mixture of type-raised and rst-order representations of proper names in the examples to improve readability, we will dene an interface for each version. We will rst dene the grammar and then the rules for semantic construction, starting with the type-raised version and then showing what needs to be changed to get the rst-order representations. Finally, we will go through an example to show how the interface works. 5.1. Grammar The grammar is dened by the phrase structure rules in Fig. 1. Figure 1. The grammar Most labels for syntactic categories are self-explanatory, perhaps except for the following: PN is a category that comprises proper nouns and anaphora, Conj refers to the category of conjunctions like before, Coord, to the one of sentence coordinators like and, and RP and RCl, to the categories of relative pronoun and relative clause, respectively. Beyond transitive and intransitive verbs, we also allow 'AcI verbs' AV (such as expect) and sentence complement verbs SV (such as say). Conj and Coord are distinguished because they percolate scope domains in a dierent fashion. The lexicon is dened by a relation Lex, which relates words W and lexical categories 2 fDet; N; IV; TV; SV; productions expand lexical categories to words of this category. We make two types of simplifying assumptions about the grammar. First, we disregard all aspects of syntactic analysis that are inconvenient for a pure phrase structure grammar, such as agreement. Any The Constraint Language for Lambda Structures 23 serious NLP system would employ some type of unication grammar, in which these aspects could be taken care of easily. As a matter of fact, we have implemented an HPSG grammar that does this and produces essentially the same constraints as the one presented here. Second, we assume that the syntax provides coindexation that relates anaphoric pronouns to their antecedents and relative pronouns to their corresponding traces, and that the parallelism constraints that model VP ellipsis are generated by an independent source. Determining elliptic elements is, to our knowledge, an open problem and therefore beyond the scope of this paper. 5.2. Semantic construction The overall strategy of our syntax-semantics interface is to factor out as much of the constraint construction as possible into the interface rules. This allows us to keep the entries in the semantic lexicon extremely simple: most of them introduce just one labelling constraint. We generate a constraint for each node in the syntax tree; the constraint for the entire tree will then be simply the conjunction of these subconstraints. Each node 2 N in the syntax tree is associated with two variables, (the local scope domain of ) and X r (the root of the subconstraint for denotes the root of the 'semantic contribution' of the syntactic constituent below . X s is referred to in scope island constraints; for each quantier NP , except for indenites, a constraint keeps the quantier within its scope domain. The values of are set in the construction rules for embedded sentences; otherwise, daughters inherit the scope domains of their mothers. Furthermore, we associate with each NP index i that is used for coindexation in the syntax tree a variable X i . We will use these variables to model anaphoric binding and relative clauses. The (semantic) variables associated with syntactic nodes are related by the rules in Fig. 2, one for each syntactic rule from Fig. 1. Each construction rule is applied to syntactic nodes that have been expanded by a certain syntax rule. These rules are specied on the left hand sides of the arrows; [ :P Q R] means that node in the syntax tree is labelled with P , and its two daughter nodes 0 and 00 are labelled with Q and R, respectively. The complete constraint which the interface generates is the conjunction of all of these local constraints, plus the dominances X s for scope islands, plus constraints for the lexical entries. Most of the latter constraints are simple labelling constraints; for instance, the constraint we add for the intransitive verb sleep would be X r :sleep. Markus Egg, Alexander Koller, Joachim Niehren @ 000 000 @ lam @ @ lam lam X r @ var @ var lam X r 000 Figure 2. The syntax/semantics interface The Constraint Language for Lambda Structures 25 There are some exceptions to this general rule. First, the elliptic does (too) does not add a labelling constraint; its semantics is determined via a parallelism constraint. Furthermore, whenever coindexation signies a relation between an anaphor 0 and its antecedent , we add the constraint X r when we process and the constraint when we process 0 . Similarly, the constraint for a relative pronoun with index i at is X r and the constraint for the corresponding trace (say, at 0 ) is X r together with rule (b11), enforces correct binding of the trace. Finally, the constraints for possessive pronouns, such as his, are as in (38). Intuitively, we want to analyze the NP his i book just like the NP the book of him i ; the diagram in (38) hides most details of this construction for readability, but notice that there is an anaphoric linking constraint from the anaphoric node to the antecedent X i , as specied by the syntactic coindexation. ana The interface as presented so far will always produce type-raised representations of all NPs. A version of the interface which produces rst-order representations of proper names (i.e. as constants of type e) can be obtained by making exactly two changes to the interface. For one, the rule (b7) must be replaced by a rule that looks like (b9); second, the lexicon must be changed to label the nodes with the constants of type e (i.e., mary instead of Mary). 5.3. An Example To see how semantic construction works, consider example (39). person that Bill greeted The grammar assigns to (39) the parse tree in Fig. 3. From this parse tree, application of the synsem interface rules returns the constraint in Fig. 4. First, rule b7 builds the fragment below X 15 plus the node (with its -link to X 17 ). Next, rules b3 and b1 build up the subtree below X 9 from the semantics of the verb greet and its two NP arguments (without the label of X 12 ; recall that the corresponding trace contributes only an unlabelled node). Then rule b11 constructs the subtree below X 7 and introduces the -link for X 12 ; the label of X 12 comes from the semantics of the relative pronoun. The root node of the 26 Markus Egg, Alexander Koller, Joachim Niehren person RCl that i PN Bill greeted Figure 3. person that Bill greeted lam X 1 @ person X 4 var @ lam X 7 @ greet var X 12 var Bill lam X 17 Figure 4. Semantic representation of example (39). constraint for the relative clause is X 7 . Finally, the subtree below X 1 (the root node of the semantics of (39)) is built according to b10. The semantic contribution of the noun person is node X 4 . 6. Ecient Disambiguation A good underspecication formalism should not only allow to write down appropriate underspecied descriptions, but also to enumerate their readings eciently. The most optimistic case is that they can be enumerated in time polynomial in the total number of readings. In a framework that allows the analysis of ellipsis and anaphora in addition to scope, it is unlikely that such a low runtime can be obtained. A prime example is the Dalrymple et al. (1991) analysis of ellipsis: It uses higher-order unication, which is known to be undecidable. Similarly, the problem of disambiguating arbitrary CLLS descriptions (that is, of enumerating all most general solved forms of a constraint) is very complex. It was shown in Niehren and Koller (1998) that CLLS is equivalent to context unication (Schmidt-Schau and Schulz, 1998; Niehren et al., 1997a). There are semi-decision procedures The Constraint Language for Lambda Structures 27 for context unication, but its decidability is a prominent open problem in theoretical computer science (RTA, 1998). For the sublanguage of dominance constraints (which are sucient for scope underspecication alone), the situation is better: It was shown in Koller et al. (1998) that the satisability problem of dominance constraints is decidable, but NP-complete. This means that the most efcient solution procedure may take (deterministic) exponential time for enumerating all readings, even if the number of readings is polynomial. Nevertheless, there are solvers for dominance constraints that perform very eciently for all constraints from scope underspecication. These solvers are based on constraint programming with nite sets (Koller and Niehren, 1999); the general strategy is iterating deterministic propagation steps and non-deterministic choices. It turns out that despite the general NP-completeness, all choices made in the disambiguation of a constraint from scope underspecication are necessary to distinguish dierent readings; the search never fails. Recent research in this area focuses on obtaining formal complexity results that model this experience and on integrating a semi-decision procedure for the rest of CLLS with a dominance constraint solver. A demo system that derives dominance constraints describing a scope ambiguity and solves them eciently is available on the WWW (CHORUS Project, 1999). 7. Conclusion and Outlook In this paper, we have presented the Constraint Language for Lambda Structures (CLLS), a rst-order language for semantic underspecica- tion. CLLS allows the representation of scope ambiguities, anaphora, and ellipses in simple underspecied structures that are transparent and suitable for processing. It is a language of tree descriptions that conservatively extends the language of dominance constraints. We have shown that CLLS correctly represents many notorious problems from the literature involving scope, anaphora, ellipses, and their interactions. An advantage over earlier approaches to this area is that CLLS analyses are completely order-independent, and the formal foundations are very transparent. It also provides a clean solution to a problem with variable binding in the context of scope underspecication that becomes most obvious in the presence of ellipses. Although the linguistic coverage of CLLS includes various challenging problems, it still needs to be extended. The most obvious area for improvement is a more complete theory of ellipsis. Another interesting extension is to build dynamic accessibility relations into the denition of anaphoric links and make the eects of anaphora on quantier scope 28 Markus Egg, Alexander Koller, Joachim Niehren explicit by reasoning on underspecied structures (Koller and Niehren, 2000). Finally, CLLS can be used to model reinterpretation (meaning shift) of aspect and NPs in an underspecied way (Striegnitz, 1999; Egg, 1999; Koller et al., 1999). On the other hand, there are various more formal aspects of CLLS that can be pursued in the future. One essential problem for any under- specication formalism is to raise operations on the object level, such as deduction, to the level of underspecied descriptions. Although there are some rst results (Jaspars and Koller, 1999), the entire eld is still in its infancy. Furthermore, it is necessary to nd tractable algorithms for processing parallelism constraints (Erk and Niehren, 2000). Acknowledgements This work was supported by the SFB 378 (project chorus) at the Universitat des Saarlandes. The article is an updated and revised version of Egg et al. (1998) and Egg et al. (1999); we are indebted to Peter Ruhrberg and Feiyu Xu for their contributions to these papers. The authors wish to thank Manfred Pinkal, Gert Smolka, all members of the SFB 378 projects chorus, lisa, negra, and nep, and the commentators and participants at the Bad Teinach workshop on underspecication. --R CHORUS Project: de/cl/projects/chorus/demo. Arbeitspapiere der FG Sprachwissenschaft de/cl/projects/chorus/papers/egg99. Anaphora and Quanti Pustejovsky and P. Deletion and Logical Form. --TR --CTR Manuel Bodirsky , Katrin Erk , Joachim Niehren , Alexander Koller, Underspecified beta reduction, Proceedings of the 39th Annual Meeting on Association for Computational Linguistics, p.74-81, July 06-11, 2001, Toulouse, France Katrin Erk , Joachim Niehren, Well-nested parallelism constraints for ellipsis resolution, Proceedings of the tenth conference on European chapter of the Association for Computational Linguistics, April 12-17, 2003, Budapest, Hungary Alexander Koller , Stefan Thater, An improved redundancy elimination algorithm for underspecified representations, Proceedings of the 21st International Conference on Computational Linguistics and the 44th annual meeting of the ACL, p.409-416, July 17-18, 2006, Sydney, Australia Michael Schiehlen, Ellipsis resolution with underspecified scope, Proceedings of the 40th Annual Meeting on Association for Computational Linguistics, July 07-12, 2002, Philadelphia, Pennsylvania Katrin Erk , Joachim Niehren, Dominance constraints in stratified context unification, Information Processing Letters, v.101 n.4, p.141-147, February, 2007 Manuel Bodirsky , Denys Duchier , Joachim Niehren , Sebastian Miele, A new algorithm for normal dominance constraints, Proceedings of the fifteenth annual ACM-SIAM symposium on Discrete algorithms, January 11-14, 2004, New Orleans, Louisiana Alexander Koller , Joachim Niehren , Stefan Thater, Bridging the gap between underspecification formalisms: hole semantics as dominance constraints, Proceedings of the tenth conference on European chapter of the Association for Computational Linguistics, April 12-17, 2003, Budapest, Hungary Ralph Debusmann , Denys Duchier , Alexander Koller , Marco Kuhlmann , Gert Smolka , Stefan Thater, A relational syntax-semantics interface based on dependency grammar, Proceedings of the 20th international conference on Computational Linguistics, p.176-es, August 23-27, 2004, Geneva, Switzerland Alexander Koller , Stefan Thater, Efficient solving and exploration of scope ambiguities, Proceedings of the ACL 2005 on Interactive poster and demonstration sessions, p.9-12, June 25-30, 2005, Ann Arbor, Michigan Joachim Niehren , Stefan Thater, Bridging the gap between underspecification formalisms: minimal recursion semantics as dominance constraints, Proceedings of the 41st Annual Meeting on Association for Computational Linguistics, p.367-374, July 07-12, 2003, Sapporo, Japan Ruth Fuchss , Alexander Koller , Joachim Niehren , Stefan Thater, Minimal recursion semantics as dominance constraints: translation, evaluation, and analysis, Proceedings of the 42nd Annual Meeting on Association for Computational Linguistics, p.247-es, July 21-26, 2004, Barcelona, Spain Ernst Althaus , Denys Duchier , Alexander Koller , Kurt Mehlhorn , Joachim Niehren , Sven Thiel, An efficient graph algorithm for dominance constraints, Journal of Algorithms, v.48 n.1, p.194-219, August Denys Duchier, Dominance constraints with Boolean connectives: a model-eliminative treatment, Theoretical Computer Science, v.293 n.2, p.321-343, 6 February Manuel Bodirsky , Martin Kutz, Determining the consistency of partial tree descriptions, Artificial Intelligence, v.171 n.2-3, p.185-196, February, 2007
underspecification;ellipsis;anaphora;tree descriptions;scope ambiguity
596041
Towards the Use of Automated Reasoning in Discourse Disambiguation.
In this paper, we claim that the disambiguation of referring expressions in discourse can be formulated in terms automated reasoners can address. Specifically, we show that consistency, informativity and minimality are criteria which (i) can be implemented using automated reasoning tools and (ii) can be used to disambiguate noun-noun compounds, metonymy and definite descriptions.
Introduction It has long been thought that much of the burden of Natural Language understanding and generation in automated systems could be carried by inference. The most obvious early example of this is Winograd's SHRDLU system (1973), implemented in PLANNER (Hewitt, 1969), the earliest logic programming language. SHRDLU was a simulated, language-understanding robot. It interpreted Natural Language input in terms of decomposable procedures for achieving a state of the world in which the input was true or for assessing whether it was already true, with the resulting sub-goal structure (a plan or proof) then available for answering questions about the system's actions. Thus SHRDLU could understand and respond (either in language or simulated action) to input utterances by proving that an utterance was meaningful, with the proof providing a basis for its response. While a full survey of inference in Natural Language understanding is beyond the scope of this paper, besides SHRDLU it is worth mentioning Charniak's PhD thesis (1972) on children's story understanding, which assumes that stories are \immediately translated into an internal representation which is convenient for doing deduction" (Charniak, 1986), p.337. Charniak's system used rules triggered by evidence in a story, to make connections between story elements that are not expressed explicitly and thereby to answer questions about the story. In the 1980's, work on task-related dialogues by Allen and Perrault (1982), Litman and Allen (1990) and others, showed how systematic c 2001 Kluwer Academic Publishers. Printed in the Netherlands. Claire Gardent and Bonnie Webber inference on a hearer's part about a speaker's beliefs and goals could be used in understanding questions (often fragmentary indirect speech acts) and in responding to them cooperatively. Appelt's PhD thesis (1985) was the most detailed study of its time of the role inference could play in Natural Language generation, and the framework he developed was able to produce utterances that could satisfy a hearer's domain needs and his/her understanding abilities. However, for a long time, formal reasoners were too slow and the amount of knowledge they needed was too vast, to support their use in anything other than a toy system or PhD thesis. This situation is now changing: in Natural Language generation, fast automated reasoners are being used in integrated systems for micro-planning (Stone, 1998; Stone, 2000; Stone and Doran, 1997; Stone and Webber, 1998; Stone et al., 2001; Gardent and Striegnitz, 2001), while in Natural Language understanding { the concern of this paper { fast automated reasoners are being employed over an ever-expanding range of problems. In this paper, we focus on the role inference can play in disam- biguation, provided that disambiguation problems can be formulated in terms that automated reasoning tools can address. Here we look only at reasoning in rst-order logic. There are many types of ambiguities in natural language. Word sense ambiguity arises from the fact that a word can have dierent meanings. This has been addressed by e.g. (Charniak, 1986; Lehnert, 1986; Wilks, 1986) and (Hobbs et al., 1993) who showed how to identify a word meaning in a given context using script-based and abductive inference respectively. Structural ambiguity involves the possibility of multiple syntactic analyses of a string. But since each syntactic analysis may have a dierent semantic analysis, structural ambiguity often correlates with semantic ambiguity as well, as for instance in prepositional phrase and relative clause attachment (Crain and Steedman, 1985). Resolution ambiguity covers the many possible ways in which semantically underspecied elements such as anaphors or ellipses can be interpreted in a given context, while projection ambiguity refers to the various ways in which a presupposition can be integrated into the overall meaning of a text. The role that inference can play in resolving projection ambiguity has been demonstrated by (Blackburn and Bos, 1999; Blackburn et al., 1999) who showed that existing rst-order theorem provers and model builders can be used to weed out some of the impossible readings. Scope ambiguity refers to the fact that a quantier or a focus-sensitive operator such as \only" can include more or less of its matrix utterance within its scope. Again (Blackburn and Bos, have shown that rst-order automated reasoning tools can be used to rule out some of the illegitimate scopings namely, those that are incompatible with world knowledge. Finally, speech act ambiguity refers to the alternative possible dialog functions that an utterance can have. As mentioned earlier, plan-based inference techniques can help deal with this problem (Allen and Perrault, 1982; Chu-Carroll and Carberry, 1994; Litman and Allen, 1990; Lambert and Carberry, 1999), and statistical dialogue modelling may prove a help here as well (Stolcke et al., 2000). In this paper, we concentrate on how inference can be used to resolve ambiguities involving referring expressions in discourse: specically, we show that the method initiated in (Blackburn and Bos, 1999) for dealing with projection ambiguity can be successfully used for resolving underspecied entities and relations, by dispreferring resolutions that lead to inconsistent or uninformative readings. Additionally, we argue that model minimality is another factor that can be checked using automated reasoning tools and can serve ambiguity resolution: given several competing interpretations, these interpretations which are satised by the smallest model (smallest domain, smallest interpretation function) are often the preferred interpretations. The underlying intuition is that minimal models yield interpretations that are more \coherent" than non-minimal ones. The paper is structured as follows. We start in Section 2 by arguing for a certain type of interaction between NLP and automated reasoning and show how checks for consistency, informativity and minimality can be implemented using automated reasoning techniques. In Section 3, we then consider the interpretation of noun-noun compounds, metonymy and denite descriptions and show that their disambiguation through contextual reasoning can be done through checks for consistency, informativity and/or minimality. We conclude in Section 4 with pointers for further research. 2. NLP and Automated Reasoning There are essentially two ways to integrate automated reasoning into NL understanding: one can use it to resolve underspecied interpretations or one can use it to lter out unwanted interpretations resolved by other means. In the rst approach, which we will refer to as a constructive ap- proach, an underspecied semantic representation i.e., a formula containing rst-order and/or higher-order variables, has those variables resolved (i.e., instantiated) in the course of being proven with respect to the knowledge base. Thus a more specic representation is constructed from one that is underspecied. This approach is used in (Hobbs et al., 4 Claire Gardent and Bonnie Webber 1993; Gardent and Konrad, 2000a) and to a lesser degree in (Lascarides and Asher, 1993). Thus (Hobbs et al., 1993) uses abduction to prove an unresolved semantic interpretation. When this requires making assumptions about the values of variables that make the formula true, these assumptions become part of the interpretation, thereby specifying it further. Similarly, (Lascarides and Asher, 1993) applies a process of default inference to an underspecied discourse representation to compute its semantic interpretation. Defeasible content added by default also becomes part of the interpretation, again specifying it further. As others have remarked, the use of abduction for NL understanding has much in common with model-building, as in (Gardent and Konrad, 2000a; Baumgartner and Kuhn, 1999; Ramsay and Seville, 1999). model builder is a program that takes a set of logical formulae and returns some of the models that satisfy .) Where model-building is used for NL interpretation, the interpretation of a formula is some (carefully selected) model satisfying . As with abduction, a model may make information explicit that is only implicit in the input formula i.e. the semantic representation of the input discourse (Gardent and Konrad, 2000a). The merit of using reasoning to resolve underspecied interpretations is that it integrates knowledge-based reasoning directly into the interpretation process. In the case of model building, this has the advantage of providing a consistency check \for free", since the only models that will be built are both logically consistent and consistent with world knowledge. If model building and/or abduction is used to produce a set of possible interpretations for some input formula, members of the set can be compared and ranked according to any desired metrics. This provides a direct handle on the ranking of competing interpretations for a given discourse. The downside of using reasoning in this way is that ambiguity and underspecication are resolved purely on the basis of semantic content. This is in contrast with van der Sandt, who (in the case of presupposition projection) makes essential use of the structure of the semantic interpretation (van der Sandt, 1992). One can, of course, take this (and anything else) into account, but it requires explicit encoding in terms of additional predications. An approach that uses weighted abduction to resolve interpretations also suers from the problem of motivating the abductive cost of each type of assumption and verifying that it produces the intended result whenever used. The second way of integrating automated reasoning into NL understanding is to use it to complement separate specialised algorithms for resolving underspecication and ambiguity. In particular, when more than one possible reading is found for the input discourse, automated reasoning (in the form of theorem proving and/or model building) can be used to lter out unwanted ones. For instance, (Blackburn et al., 1999) compute all possible presupposition projections for an input discourse using van der Sandt's projection algorithm (van der Sandt, 1992), and then use rst-order theorem proving and model building to apply van der Sandt's pragmatic constraints on presupposition projection to eliminate implausible readings. The drawback of such a ltering approach is that it requires all ambiguities (including structural ambiguities, resolution ambiguities, projection ambiguities, etc.) to be identied and the resulting possible analyses enumerated. Dependencies between expressions and between types of ambiguities can make the process of computing all and only the possible analyses of a text, quite complex { a complexity that the ltering methods presented here do not address. Moreover, a complete enumeration seems far away from \how people process discourse", although it does take advantage of what systems can do well While a complete enumeration would be ine-cient if possible analyses were simply ltered sequentially, ignoring commonalities that would allow subsets to be ltered simulataneously, the input to ltering need not be the entire set of possible interpretations of a (small) discourse. If such interpretations could be generated in order of minimality (just as in breadth-rst search, solutions are generated in shortest-rst order, and in A* search, in best-rst order with respect to a given heuristic), then consistency and informativity could be applied locally to either accept the current interpretation, or reject it in favor of the one with the next most minimal model that satises these tests. (This is what is assumed in (Crain and Steedman, 1985).) Additional non-logical tests (not described here) could then be applied to choose between readings that were equally minimal or incomparable with respect to minimality. In either case, we look to take advantage of the resources provided by automated reasoners and tune them (through discourse-designed test suites) to better satisfy our needs. On the other hand, the value of a ltering approach is that it supports a precise specication of what is relevant to each decision and how it is used. For this reason, we adopt here a ltering approach, starting with an enumeration of readings expressed in a many-sorted rst-order logic. We then show that automated reasoning can use the criteria of consistency, informativity and minimality to identify and lter out unwanted readings. We shall see that in many cases, these three criteria su-ce to eliminate linguistically invalid readings and rank the remaining acceptable ones. First, however, we describe these criteria and show how they can be tested using techniques from automated reasoning. 6 Claire Gardent and Bonnie Webber 2.1. Consistency It is well-known that rst-order consistency and validity problems are undecidable: Given a formula , there is no guaranteed method of computing whether or not is consistent. However (Blackburn et al., 1999) shows that state-of-the-art, sophisticated rst-order theorem provers and model builders can be of practical use in checking consistency for at least some classes of linguistic problems. The idea is to use theorem provers and model builders in parallel. Specically, suppose we want to check the consistency of . We can simultaneously ask the theorem prover to prove : and the model builder to construct a model that satises . If the theorem prover succeeds, : is valid, hence is inconsistent. (If the theorem prover returns no result, either : is valid but the theorem prover ran of resources, or : is invalid.) On the other hand, if the model builder succeeds, then is satisable and hence consistent. Although there always remains the possibility either that both theorem prover and model builder run out resources, or that neither program terminates on some given input (even given unlimited resources), for at least the presupposition projection problem studied in BBKN, state-of-the-art automated reasoning systems return an answer almost all of the time and do so reasonably e-ciently. The question remains however, of whether the same will hold true of all inference tasks in discourse understanding. We can only tell this by clearly specifying the tasks. Developers can then try to tune their systems to perform e-ciently and eectively on these tasks. The aim of the present paper is to provide pointers to further language-related tasks. 2.2. Informativity BBKN show that informativity can be treated as a validity problem. Suppose NEW is the formula we want to test for informativity. Then it is informative with respect to a discourse context OLD and general world knowledge KB just in case NEW is not entailed by OLD ^ KB. Equivalently, NEW is informative with respect to OLD and KB just in case (OLD NEW is not valid. 2.3. Minimality Minimality, in some form, has continually been invoked as an important factor in choosing among competing readings of a text, including (Crain and Steedman, 1985; Johnson-Laird, 1983) inter alia. Intuitively, minimality resembles the early AI planning heuristic (Sacerdoti, 1977): \Use existing objects", but extended to include not just individuals but all semantic entities e.g., properties, relations and propositions, and taking \existing" to mean \discourse old" or \discourse salient". Minimality thus serves discourse coherence by favouring readings which continue to communicate about (or in terms of) the same things over those that introduce new ones. Although minimality is not as straightforward a concept as consistency and informativity, nevertheless one can begin to identify the easy cases and specify precisely how to capture various aspects of minimality in a framework of automated reasoning { in particular, in a framework of model building. Within such a framework, one can also begin to lay down guidelines for constructing models in less straightforward cases, to produce minimal models that accord with the speaker's intended meaning. As noted earlier, given a set of formulae , a model builder returns some of the models satisfying . A model is a mathematical structure that describes how the symbols of a logical theory are interpreted. Specically, a model for a rst-order language L is a pair hI; Di with D a non-empty set of entities (the domain of individuals) and I an interpretation function which maps relation symbols in L to relations of appropriate arity in D and constant symbols in L to elements of D. Here we identify these models with sets of positive assumptions that unambiguously dene the interpretation of the relation symbols and x the interpretation of terms to rst-order entities that carry a unique name. These are known in the literature as Herbrand models. For example, given sentence (1a) and its logical form (1b), the set (1c) is one of the Herbrand models satisfying (1b). In this model, there is an individual c 1 who is a friend of Jon and who Jon interred (i.e., buried). (1) a. Jon interred his friend. b. The model M 1 denes an interpretation of the predicates friend and inter over the universe of discourse g. It can also be taken as a valid interpretation of (1a). There are, however, innitely many other Herbrand models for (1b) which do not adequately capture the meaning of (1a) e.g., The model M 2 declares Jon to be his own interrer, which is not the intended or even a possible meaning of (1a) 1 . In the case of M 3 , the 1 The object of \inter" is, by denition, dead or at least taken to be so. 8 Claire Gardent and Bonnie Webber model contains an extraneous assertion namely, that Jon is married. As this is neither expressed nor entailed by the input discourse (1a), the model fails to appropriately capture its meaning: while it is possible to assume married(jon), it is not necessary for explaining the meaning or truth of (1a). In short, just as in the story The Three Bears, while some satisfying models are too small some seem \just right" to capture the preferred meanings that a hearer associates with a discourse. These are the ones we are interested in for Natural Language applications. To eliminate models that are \too big", we restrict the set of linguistically interesting models to those models that are in some sense minimal. The intuition is that the linguistically interesting models are often those that show the truth of the input theory by making the fewest possible assumptions and by using the fewest \objects" (i.e., individu- als, properties, relations, propositions). In what follows, we make use of a form of minimality introduced by (Gardent and Konrad, 2000a) for interpreting denite descriptions namely, local minimality. Since this notion incorporates both domain minimality and subset minimality, we rst explain these two concepts. Domain minimality (Hintikka, 1988; Lorenz, 1994) selects from among the many models satisfying the input specication , those models that have the smallest universe (i.e. domain of individuals). Domain Minimality: Let be a set of rst-order formulas and S be the set of Herbrand models of . Then a model hI; Di 2 S is domain minimal i there is no other model hI the cardinality of D 0 is smaller than the cardinality of D i.e. jD 0 j < jDj. By contrast, subset minimality (Bry and Torge, 1996; Schutz, 1999) minimizes facts rather than individuals 2 . Subset Minimality: Let be a set of rst-order formulas and D be the set of Herbrand models of . Then a model hI; Di 2 D is subset minimal i there is no other model hI such that I 0 I. Finally, local minimality requires models that satisfy some input within a minimal domain D of individuals and are subset-minimal with respect to all other domain minimal models. To spell this out explicitly: be a set of rst-order formulas and D be the set of Herbrand models of that use some nite domain D whose size is minimal. Then a model hI; Di 2 D is local minimal i there is no other model hI such that I 0 I . Subset minimality is also used in Circumscription (McCarthy, 1980), to dene a consequence relation j=c such that j=c i is true in all subset-minimal models of . In this paper, we do not use j=c , using subset minimality only as a way to help determine a preferred interpretation. Local minimal models are the simplest in the sense of Occam's Razor and often the best explanation for the truth of an observation. By preferring subset minimal models, for example, local minimality correctly rules out M 3 as not yielding a correct interpretation of example (1a). Eliminating models that are \too small" is less straightforward: the problem is that a locally minimal model satisfying the input semantic representation may not capture the preferred reading of the corresponding discourse. In such cases, model theoretic semantics interacts with other linguistic and non-linguistic knowledge sources (i.e., syntax, prosody, pragmatics and world knowledge) to determine the preferred reading, and it is this interaction that needs to be captured. One simple instance of this interaction is illustrated by the discourse (1a) and its locally minimal model M 2 , which does not adequately capture its meaning. If the input specication were augmented with appropriate world knowledge { namely, that one cannot inter oneself { then M 2 would never be constructed because the model builder can only build models that are both logically consistent and consistent with world knowledge. Thus M 1 would then be the locally minimal model and hence the preferred reading for (1b). More complex cases arise when factors other than world knowledge contribute to ruling out locally minimal models that a hearer would take to be \too small". For example, consider a. Jon likes his friend. b. d. Both M 1 and M 2 are satisfying models for (4b). But the locally mini- mal, hence preferred, model for (4b) is M 2 which states that Jon likes himself. Although this is a possible meaning for (4b) { i.e., if Jon's only friend is himself { mutatis mutandis, M 2 is not the preferred interpretation for (4a). Here, pragmatics comes into play, as models such as M 2 can often be eliminated by taking account of the Gricean maxim of manner: Be perspicuous (Grice, 1975). Specically, if the speaker intended two NPs to refer to the same object, she would have used a lexico-syntactic construction that customarily implies this (i.e., Jon likes himself ). So as long as appropriate knowledge from prosody, pragmatics, world knowledge, etc. can be identied and included in the input specication along with the semantic representation of the input discourse, model building can automatically rule out any model that is inconsistent with the overall input specication. On the other hand, speakers often violate the Gricean maxims for particular eects. For instance, in Gardent and Bonnie Webber (5) Jon admires his favorite person. the speaker may intend for the hearer to recognise that Jon's favorite person is Jon. In short, minimality, even when combined with inconsistency, does not produce a perfect test, ruling out all and only non-preferred read- ings. But if we can characterise the situations in which it is a good test (i.e., its false positive rate and false negative rate are low), it would be worth applying in those cases. Finally, note that minimality cannot, by itself, be used as a basis for deciding between competing possible readings of anaphoric pronouns. Consider Jon insisted that Fred drive [but, because] he was too tired. Models in which he co-refers with Jon and ones in which he co-refers with Fred are equally minimal, but one seems preferred in the case of \but" (i.e., Fred), while the other seems preferred in the case of \because" (i.e., Jon). So the preferred interpretation must be decided by other means. The approximation of discourse coherence given us by local minimal models is both less than and dierent from that used by (Lascarides and Copestake, 1999; Asher and Lascarides, 1998) to resolve noun-noun compounds, metonymy and denite descriptions. That treatment is grounded in discourse coherence associated with discourse relations (causality, narrative, background, concession etc.) that are taken to hold between eventualities. Establishing what relation holds in a particular case involves both logical deduction and defeasible reasoning. In contrast, the minimality-coherence discussed here, while it does not address the full discourse interpretation problem, involves only a simple bias that can be checked using rst-order model building. 3. Case Studies In the preceding section, we have proposed using automated reasoning to lter out possible but unwanted (or non-preferred) readings of a discourse, leaving either a single intended reading or a set of readings among which a most preferred one can be chosen using non-logical means (e.g., prior or conditional probabilities). We have suggested three criteria that support this selection process { consistency, informativity and minimality { and described how they can be assessed through automated reasoning. Here we show how these three criteria can be used, individually and jointly, in the discourse interpretation of three dierent phenomena that are well-known for ambiguity: noun-noun compounds, metonymy, and denite noun phrases. In each case, we start with the set of possible semantic representations of a text. Tests for consistency and informativity may then rule some of them out. Local minimality is then applied to the set of locally minimal models satisfying the remaining semantic representations. In some cases, the set may contain models that do not stand in a subset relation to each other, as when one model contains specic individuals and another, generic ones. Such models are incomparable with respect to local minimality. Where those models all embody legitimate and equally preferred readings, ranking is neither required nor desired. However, problems do arise where either incomparable models or comparable models of equal minimality embody readings with dierent degree of acceptability. In some cases, one may be able to appeal to other types of minimality which draw on relations other than sub- set. For instance, (Gardent and Konrad, 2000b) introduce conservative minimality, a conservative extension of local minimality, to model the interpretation of reciprocals. In other cases, factors other than minimality will probably have to be brought into play to choose between possible models, such as preferences for certain sources (e.g. world knowledge versus discourse context) or types (e.g. anaphoric link versus accommodation) of resolution, or notions of domain-specic salience or likelihood. Our point is to show that e-cient automated reasoning can quickly dispose of what is unwanted or non-preferred on formal grounds, not that it solves the problem of discourse interpretation in its entirety. 3.1. Noun-Noun Compounds Noun-noun compounds involve implicit relations between the denotations of their component nouns. Because world knowledge admits so many dierent relations { (Russell and Norvig, 1995) notes Compound Relation basketball shoes used-for baby shoes used-by alligator shoes made-of designer shoes made-by brake shoes part-of (to which can be added \toe shoes" (used-for-standing-on) and \tap shoes (containing), etc.) { noun-noun compounds have long held a fascination for NLP research (Finin, 1980; Hobbs et al., 1993; Levi, 1979). Recently, researchers such as (Lascarides and Copestake, 1999) have begun to consider noun-noun compounds in the context of discourse Gardent and Bonnie Webber understanding, where new relations introduced by the discourse may also be relevant to nding a referent for the noun phrase (NP) in which the compound occurs. For example, in the isolated utterance (7) The California student produced an excellent report. the compound California student involves some relation between California and student that can be attributed to the referent of the NP \the California student". Here the most plausible relation holding between a student (i.e., a person) x and a state y is known from world knowledge to be x being from y. Thus, \the California student" most plausibly refers to the student who is from California (whoever that is). However, (Hobbs et al., 1993; Lascarides and Copestake, 1999) show that in discourse, other relations are not only possible but preferred: Each student was assigned a state to study. The California student produced an excellent report. While it is still possible that the intended relation between \student" and \California" is that of being from, the most plausible relationship between them that could be attributed to the intended referent of the NP is student assigned to study California. This relation is one made available by the discourse context, overriding possibilities made available by general world or lexical knowledge. In what follows, we assume that the possible interpretations of a compound have already been identied by a separate process that has access to the discourse context, lexical knowledge, and general world knowledge. Checks for consistency, informativity and minimality can then be used to determine which interpretation is preferred. 3.1.1. Minimality Minimality can be used to determine the preferred interpretation of Example (8), repeated below. 3 Each student was assigned a state to study. The California student produced an excellent report. As noted earlier, we assume that possible readings of each utterance have already been identied { here, two dierent ways of resolving the relationship between \student" and \California" { the rst (9a) arising from the discourse context, and the second (9b), from the general world knowledge: a. The student assigned to study California produced an excellent report b. The student from California produced an excellent report. 3 In Example (7), the discourse context is empty, so the only relations available for resolving the compounds are those provided by lexical and world knowledge. To see how the preferred reading (9a) can be chosen on the basis of minimality from among those oered for ltering, consider the following models for Example (8). Because M b1 and M b2 are both models for the same interpretation of (8), M b2 would be eliminated from consideration by being less domain minimal than M b1 , even though M b1 contains information not entailed by reading (9b) { namely, that the student from California who produced an excellent report was assigned to do that report on California. Thus, as we saw before with Examples 1 and 4, minimality sometimes delivers interpretations that are overly specic. Applying local minimality to the remaining models M a and M b1 , the smallest model for (8) is M a . By taking the smallest (and hence, the simplest) model explaining the truth of a discourse as yielding the preferred reading { i.e., the one believed to be intended by the speaker { we can say that by minimality, (9a) is the preferred reading. 3.1.2. Consistency Consider the following variant of (8): After each student was assigned a state to study, the California student, who had been assigned Ohio, started work immediately. As with (8), there are at least two possible ways of resolving the relation between \California" and \student". These lead to the following meanings for the resolved discourse: a. The student from California, who had been assigned to study Ohio, started work immediately. b. The student assigned to study California, who had been assigned to study Ohio, started work immediately. Assuming that a \quantity implicature" (Grice, 1975) leads us to interpret \a state" in Example 11 as \exactly one state", reading (12b) yields a contradiction. This is something that can be detected by a consistency check, which would thereby rule it out. This leaves (12a) as the preferred reading for (11). Now consider another variant of Example 8: Each student was assigned a state to study. The California student agreed to work on Ohio. 14 Claire Gardent and Bonnie Webber Again, there are two possible ways of resolving the relation between \California" and \student", leading to the following further specica- tions of the second sentence: a. The student from California agreed to work on Ohio. b. The student assigned to study California agreed to work on Ohio. These interpretations are neither inconsistent nor uninformative. So the same reasoning as in (8) would lead one to expect a preference for (14b) by local minimality. Yet it is (14a) that seems preferred. One possible and detectable explanation for this is that under interpretation (14b), a presupposition triggered by the discourse becomes inconsistent. Specically, if x agreed to P is taken to presuppose x was asked to P 4 , then the presuppositions of the two readings in (14) are: a. The student from California was asked to work on Ohio. b. The student assigned to study California was asked to work on Ohio. If \assigned to study California" entails \asked to work on California", then presupposition (15b) is inconsistent. As (van der Sandt, 1992) shows, inconsistent presuppositions result in discourses that are pragmatically odd. 5 Assuming that consistency is applied to the presuppositions of a discourse as well as to the discourse itself (using (Blackburn et al., 1999) or perhaps the more e-cient approach described in (Monz, 1999)), (14a) will correctly be found to be the preferred reading for (13). 3.1.3. Informativity Informativity can also be used to determine the preferred interpretation of a noun-noun compound. Consider this next variant on Example 8: Each student was assigned a state to study. The California student was assigned California. a. The student from California was assigned California. b. The student assigned to study California was assigned California Although interpretation (16b) has a smaller locally minimal model than interpretation (16a), it is not preferred as it results in a non{informative utterance. As with consistency, however, informativity must also be applied to the presuppositions of each interpretation, as in this nal variant of Example 8: 4 This is because x didn't agree to P also seems to presuppose x was asked to P. 5 Presupposition (15b) is not inconsistent if one assumes dierent agents for \assign" and \ask" (i.e., a dierent person assigning the student to study California from the one asking her to work on Ohio). However, such a model would not be constructed since it is not locally minimal. Each student was assigned a state to study. The California student agreed to work on California. If the factive x agreed to P is taken to presuppose x was asked to P, then the presuppositions of the two interpretations of (17) are: a. The student from California was asked to work on California. b. The student assigned to study California was asked to work on California. The non-informative presupposition in (18b) thereby makes its corresponding reading dispreferred. 3.2. Metonymy The second case we will look at involves metonymy { i.e., using a word or phrase for another that it suggests { for example, using the container for the thing contained or vice versa, as in (19) The (wine, salt, meat) is on the table. used for (20) The (wine bottle, salt celler, platter of meat) is on the table. Some verbs such as \start", \nish", \enjoy", \loathe", etc. that have straightforward compositional readings when applied to events or activities as in nished preparing the peas. are metonymic when applied to non-activities, as in a. Jane nished the peas. b. Sarah enjoyed California. When the context is empty, as in (22), the preferred readings of \nish the peas" and \enjoy California" are those given by lexical knowledge i.e., nish preparing (or eating) the peas and enjoy visiting California. Given an appropriate context however, as in a. Jane was replenishing the shelves in the canned vegetable aisle. After she nished the peas, she moved on to the artichokes. b. Sarah read books about all the Western states. California she enjoyed, but not the others. the preferred readings are those that result in a minimal interpretation i.e., replenish the shelves of canned peas and read about California. We focus here on (23b), as \nish the peas" has a similar expla- nation. To see how minimality yields the preferred reading for this example, consider its models under the alternative reconstructions of the metonymy. (For simplicity, we give the models for \Sarah read about all the Western states. California she enjoyed." rather than for the whole discourse): Claire Gardent and Bonnie Webber Compared with M b , M a has fewer assertions and is therefore locally minimal. It also correctly re ects the preferred interpretation for (23b), the interpretation in which Sarah enjoyed reading about California rather than visiting it. minimality predicts that readings that \use existing objects" are preferred over those that do not and therefore that interpreting metonymy with respect to relations introduced by the previous discourse will be preferred over interpreting it with respect to relations made available by world knowledge. In Section 2.3, we noted that minimality can make wrong predictions in certain discourse situations and that it would be useful to be able to avoid the expense of applying a local minimality lter there. Cases of metonymy allow us to start identifying those situations. Consider the following examples: a. When Sarah read books about the American West, she really enjoyed California. b. After Sarah read books about the American West, she really enjoyed California. c. Before Sarah read books about the American West, she really enjoyed California. Example (25a) shows a preference for a discourse interpretation of metonymy \reading about California", while (25b) shows a marked dispreference, favoring the lexical interpretation { \visiting California". (25c) is similar, although the eect seems less strong. Since the only dierence in these examples is in the choice of con- nective, this must be the source of the dierence in preferred readings. What is known about \when" (Moens and Steedman, 1988) is that it signals a causal or contingent relation between the eventuality described in the \when" clause e when and that described in the matrix clause e matrix . In explaining this, Moens and Steedman invoke the idea of an event nucleus consisting of a preparatory process (prep(e)), followed by an instantaneous culmination (cul(e)), and nally a consequent state (conseq(e)). In these terms, a contingent relation holds between e when and e matrix when e matrix is part of prep(e when ) (as in Example 25a) or when it is part of conseq(e when ), as in When Jon spilled wine over Harold, he apologised profusely. On the other hand, \before" and \after" only convey a temporal relation between the two eventualities. 6 Examples such as (25) therefore suggest that local minimality is not an appropriate lter for discourse interpretations when temporally distinct eventualities are being described. This does not mean that the locally minimal interpretation may not, at times, be correct in such situations { just that minimality ltering would be the wrong basis on which to draw such a conclusion, since it will often lead to incorrect results. Notice that this predicts that one should also not use minimality ltering in \when" clauses where e matrix is a separate eventuality that is a contingent consequence of e when , as in When Sarah read all the library's books about the American West, she then really enjoyed California. As predicted, minimality ltering gives an incorrect result. The approach to metonymy presented here is complementary to that presented in (Markert and Hahn, 1997; Hahn and Markert, 1999). Markert and Hahn take the position that it is not type clashes that signal to a reader that a phrase should be interpreted metonymically, but rather a preference for interpreting denite noun phrases anaphorically (i.e., as referring to a previously mentioned entity) or bridging (i.e., as systematically related to such an entity). They address in detail the complex reasoning involved in recovering relevant possible interpretations of a metonymic phrase, as well as choosing the most likely. Their self-imposed constraint of incrementality, however, requires that they make this decision immediately before processing any further material in the utterance. Thus their approach does not allow them to take full advantage of tests provided by consistency, informativity and minimality. 3.3. Definite Descriptions In our nal case study, we show that minimality, consistency and informativity can also be used in resolving denite descriptions. Following (Strawson, 1950), singular denite descriptions (i.e. NPs of the form the N ) are usually taken to presuppose the existence of some 6 Such a relationship does not by itself appear su-cient for coherent discourse, and (25b) also implies positive causality (that the knowledge acquired through reading enhanced the experience), while if anything (25c) implies negative causality (that the knowledge acquired has now eliminated or diminished that enjoyment). But causality is not an intrinsic feature of either \before" or \after", but rather something deriving from the reader's beliefs about the eventualities themselves and their possible relations. Claire Gardent and Bonnie Webber individual uniquely satisfying the given description. In what follows, we will assume van der Sandt's anaphoric account of presuppositions (van der Sandt, 1992) { that they are anaphors with an unusually rich descriptive content. Like anaphors, presuppositions can be resolved to some antecedent made available by the context. (Standard terminology calls this cancelling). Unlike pronominal anaphors, however, it is possible for a presupposition to \repair" the context by creating its own antecedent, which is retrospectively added to the context in a process called accommodation (Lewis, 1979). The following example should clarify. Suppose a story were to start o with one of the following: a. The woman is single. b. Terry loves Sally. The woman is single. In (28a), the denite description the woman presupposes the existence of a uniquely identiable woman. Since none is made available by the null context, accommodation takes place and a woman is simply added to the context. By contrast, in (28b) the woman can be resolved to Sally (their referents are identied), and therefore no additional woman needs to be accommodated. The way a presupposition is resolved (whether and how it is resolved or accommodated) is pragmatically constrained. In particular, the resulting discourse must be both consistent and informative. For instance in: loves Sally. The woman is single. the reading in which Mia is the single woman is ruled out by the consistency constraint: a woman cannot simultaneously have a husband and be single. Similarly, informativity can rule out readings that are possible a priori. Thus in: loves Sally. The woman is married. although \the woman" could refer to Mia, this reading is dispreferred as it would result in the uninformative discourse: Mia has a husband and Mia is married. The role of both informativity and consistency in constraining the resolution possibilities of denite description presuppositions has been discussed at length in (van der Sandt, 1992), and given a computational treatment using automated inference systems in (Blackburn et al., 1999; Monz, 1999). We will therefore concentrate here on cases where minimality plays a decisive role in resolving denite descriptions. As discussed at length in (Clark and Marshall, 1981), context can make entities available for denite reference in several dierent ways, including explicit mention (which (Clark and Marshall, 1981) call \lin- guistic co-presence", and which we will refer to here as \anaphoric linking") and associations grounded in world knowledge (which (Clark and Marshall, 1981) call \indirect co-presence", which we will follow common practice and refer to as \bridging"). As (Gardent and Konrad, 2000a) shows, local minimality nicely captures the various ways in which a denite description can be resolved: to some previously mentioned entity (31a, anaphoric linking), to some associated entity (31b, bridging) or to some hearer-new entity (31c, accommodation). a. A woman sleeps. The woman dreams. (Anaphoric linking) b. A car stopped. The engine had broken down. (Bridging) c. The woman sleeps. (Accommodation) The preference for anaphoric linking over other possibilities can be attributed to the corresponding models having a smaller universe (do- main minimality): the entity denoted by the denite description is the same as that denoted by its antecedent. By contrast, both bridging and accommodation require the denite description to denote an entity that is distinct from all other entities explicitly introduced by the previous discourse. The preference for bridging over accommodation can be attributed to the corresponding models having fewer assertions (subset minimality). For instance, with (31b), its bridging reading (32a) is satised by model M a , whereas its accommodation reading (32b) requires something like model M b . 7 Of the two models, M a is locally minimal and therefore preferred. stopped. The engine [of that car] had broken down. b. A car stopped. The engine [unique but previously unmentioned] broke down. (Accommodation) broke down(c 3 )g There are cases of bridging however, where the entity denoted by the denite description is not itself entailed by the previous context. For instance, in: Jon entered the room. The chandelier was beautiful. the rst sentence does not entail the existence of a chandelier. Neverthe- less, \the chandelier" is interpreted as one hanging from the ceiling of 7 We assume the world knowledge that cars have engine, so that every model \containing" a car also contains its engine. Claire Gardent and Bonnie Webber the room John entered. Again this preference for bridging is captured by local minimality, as follows: if we assume that rooms have lights and that chandeliers are lights, the following models could both be constructed for contains one more entity than M a , M a is locally minimal and indicates the preferred reading. So far we have looked at simple examples with minimal contextual constraints. Constraints on the resolution of denite descriptions can arise from more complex contexts however { in particular, from relations between eventualities. Consider the following: John rode into the oasis at 3pm. The camels were under the palms, drinking water and resting. John rode into the oasis at 3pm. The camels are now under the palms, drinking water and resting. Although a bridging reading for \the camels" is preferred in both cases, the readings dier in what it is that \the camels" are associated with. In (35), \the camels" must belong to the oasis, while in (36), they could either belong to the oasis or have been John's means of transportation for getting there. As in example (25), the dierence between the two examples can be attributed to an interaction between the temporal relationships between the eventualities and consistency. Informally, the reasoning goes as follows. In both cases, the rst clause denotes an event (e 1 ), and the second a state (s 2 ). In (35), since both clauses are in the past tense, we assume following (Kamp, 1979; Hinrichs, 1986), that s 2 overlaps with e 1 . Because the eventuality e 1 denoted by the rst clause is an accomplishment, we assume following (Moens and Steedman, 1988), that it comprises a full event nucleus consisting of a preparatory process prep(e 1 ), culmination point cul(e 1 ), and consequent state conseq(e 1 ). Now e 1 (the arrival eventuality) involves a transition from John not being at the oasis in prep(e 1 ) to John being there in conseq(e 1 ). Since during the process of traveling, a person is collocated with their mode of transportation, it follows that e 1 involves a transition from John's means of transportation not being at the oasis during prep(e 1 ) to its being there during conseq(e 1 ). Turning to s 2 in (35), the claim in (Kamp, 1979; Hinrichs, 1986) that it must overlap e 1 means that there are only three ways this can happen: (1) s 2 starts during prep(e 1 ), or (2) it starts immediately at hence one of the intrinsic consequences of e 1 , or (3) it starts during conseq(e 1 ), with no commitment as to whether it is an intrinsic consequence or not. Here is where consistency and minimality interact with temporal reasoning: In the case where s 2 overlaps with prep(e 1 ), identifying the camels with John's means of transportation triggers an inconsistency, since it requires temporally overlapping states to have contradictory properties { John's camels not being at the oasis and John's camels being there. Hence the camels cannot be taken to be John's means of transportation and thus must be ones associated with the oasis. 8 In (36), on the other hand, the second clause is in the present tense, so there is no assumption that its s 2 overlaps temporally with e 1 , there is no such inconsistency. This allows \the camels" to be identied with either the oasis or John's means of transportation. Because the models are equally minimal by the arguments given so far, minimality can not be used to choose between them. Still, it is clear that (36) shows a preference for the \means of transportation" reading, and we leave for future work whether rst-order automated reasoning systems can be of help in nding it. 4. Conclusion We have shown how one can come to the preferred discourse interpretation of ambiguous sentences wholly or in part, by applying the criteria of consistency, informativity and minimality to the possible interpretations and their presuppositions. These were sentences with resolution ambiguities associated with the linguistic phenomena of noun-noun compounding, metonymy, and deniteness. As these criteria of con- sistency, informativity and minimality can be assessed by rst-order automated reasoning systems, we have in this way given further evidence for BBKN's observation in (Blackburn et al., 1999) that rst- order automated reasoning can play a role in automated discourse understanding. For this theoretical observation to be of practical interest however, automated reasoners must perform e-ciently on the type of reasoning tasks raised by the natural language data. Although Johan Bos's DORIS (http://www.coli.uni-sb.de/~bos) system already provides an Detecting inconsistencies based on temporal reasoning has not yet been addressed by the current generation of model builders. This may be one of the gap that test suites geared towards the problems of discourse understanding will cause to be addressed. 22 Claire Gardent and Bonnie Webber architecture in which to test this hypothesis, it only deals with a limited set of data namely, projection ambiguity and resolution ambiguity. It remains an open question how automated reasoners would scale to realistic grammars, longer discourses and other types of natural language related inference tasks such as those mentioned in the introduction. When optimising automated reasoning systems for mathematics, test suites of mathematics-related inference problems have allowed competing approaches to be systematically trained and compared, pushing them to become more sophisticated and e-cient at solving mathematics related problems. Similarly, we believe that having test suites for discourse understanding will allow the tuning of automated inference systems to excel at strategies that support e-cient inference in understanding Natural Language discourse. Acknowledgements We are grateful to Gann Bierner, Patrick Blackburn, Karsten Konrad, Katja Markert, Natalia Nygren, Erik Schwarz and Mark Steedman for helpful comments and discussion on material presented here. This work was partially supported by the Project C2 (LISA) in SFB{378, grant by the Deutsche Forschungsgemeinschaft to the University of Saarbrucken. --R Planning English Sentences. Cambridge University Press To appear in Journal of Language and Computation. Mental Models: Towards a cognitive science of language Morgan Kaufmann Publishers th International Joint Conference on Arti The Syntax and Semantics of Complex Nominals. Computational Linguistics Englewood Cli A Structure for Plans and Behavior. Submitted to Computational Linguistics. Journal of Semantics Morgan Kaufmann Publishers Schank and K. --TR --CTR Jochen L. Leidner , Gail Sinclair , Bonnie Webber, Grounding spatial named entities for information extraction and question answering, Proceedings of the HLT-NAACL workshop on Analysis of geographic references, p.31-38, May 31,
automated reasoning;discourse;semantics
596141
On the Complexity of Semidefinite Programs.
We show that the feasibility of a system of m linear inequalities over the cone of symmetric positive semidefinite matrices of order n can be tested in mn^O(min{m,n^2}) arithmetic operations with ln^O(min{m,n^2})-bit numbers, where l is the maximum binary size of the input coefficients. We also show that any feasible system of dimension (m,n) has a solution X such that log||X|| ln^O(min{m,n^2}).
Introduction This paper is concerned with the general semidefinite feasibility problem Given integral n \Theta n symmetric matrices A whether there exists a real n \Theta n symmetric matrix X such that where A ffl denotes the standard inner product on the space of real symmetric matrices and the notation (\Delta) - 0 indicates that (\Delta) is a symmetric positive semidefinite matrix We also consider the following (polynomially equivalent) problem Given integral n \Theta n symmetric matrices Q determine whether there are real numbers x xm such that The complexity status of problems (F) and (G) is a fundamental open issue in the theory of semidefinite programming. For the standard bit model of computation, it is known [7] that either these problems belong to the complexity class NP " coNP , or they are not in . For the real number model of computation problems (F) and (G) are known to be in NP " coNP [7], but the question of whether they can be solved in polynomial time remains open. We assume throughout the paper, that n - 2, because for the problems are trivial. The results of this paper are as follows. In Section 3, we obtain upper bounds on the norm of feasible solutions of (1) and (2). Specifically, we show that: (i) Any feasible system (1) has a solution in the Euclidean ball where log l is the maximum bitlength of the input coefficients. Moreover, the same bound applies to (2): any feasible system (2) has a solution x such that log If the solution sets of (1) or (2) are bounded, then the above bounds hold for any solution. In addition, we give examples of feasible systems (1) and (2) all of whose solutions have Euclidean norm at least R, where log In Section 4, we state lower bounds on the discrepancy of infeasible systems (1) and (2): (ii) If (1) is infeasible, then for any symmetric positive semidefinite matrix X 2 B, log . The corresponding result for an infeasible system (2) is that for any x that satisfies the upper bound of (i), the minimum eigenvalue - n of Q negative and \Gamma log(\Gamma- n We also give examples of infeasible systems (1) and (2) for which the quantities log max i=1;:::;m do not exceed \Gammal2 minfm;ng=2 . We prove the bounds of (i) and (ii) by using some results of Renegar [9] on decision methods for the first order theory of the reals and an analogue of the fundamental theorem of linear inequalities for positive semidefinite matrices. These auxiliary results are briefly reviewed in Section 2. In Section 5, we discuss the complexity of problems (F) and (G) . Due to (i) and (ii), solving (F) with the ellipsoid method requires lmn O(minfm;n 2 g) arithmetic operations with We use the decision method of Renegar [9] along with the derandomized version [2] of Clarkson's algorithm [3], [1] to improve this result as follows: (iii) Problem (F) can be solved in mn O(minfm;n 2 g) arithmetic operations over ln O(minfm;n 2 g) - bit numbers. In particular, (F) can be solved in strongly polynomial time for any fixed number of variables or constraints. Note also that for const, the required number of arithmetic operations grows linearly with m. In Section 5 we also argue that: Problem (G) can be solved in O(mn 4 operations over This extends the earlier result of Ramana [6] that for any fixed m the strict version problem (G) can be solved in polynomial time. Note that in the bit model of computation, each arithmetic operation with bit numbers can be replaced by n O(minfm;n 2 g) operations with l-bit numbers. For this reason, the bounds on the operations stated in (iii) and (iv) also apply to l-bit numbers. Finally, in Section 6 we briefly discuss some extensions of (iii) and (iv) to more general semidefinite optimization problems. Preliminaries In this section we introduce some notation and record a few auxiliary propositions, which are used in Sections 3 and 5. 2.1 Notation denotes the space of symmetric n \Theta n real matrices. Any matrix in S n can thus be viewed as a vector in IR 1n(n+1) . For a positive number R, we denote by CR the compact set is the cone of symmetric positive semidefinite matrices. - are the (real) eigenvalues of X 2 A formula (in the first-order theory of the reals) is an expression of the form where: are the free variables; ffl each Q is one of the quantifiers 9 or 8; of the form where 4 i is one of the "standard relations" ?; !; -; =; 6=, and the g i 's are real polynomials of degree at most d - 2: Note that the above formula is in prenex form: all quantifiers in (SF) occur in front. are called sentences. We say that y 2 IR k is a solution of (SF) if the sentence obtained by substituting y into (SF) is true. 2.2 Auxiliary Propositions Proposition 2.1 (Renegar [8]) If a formula (SF ) has only integer coefficients, each of bit length at most B; then every connected component of the set of its solutions intersects the ball fy 2 IR k j kyk - Rg; where R satisfies logR - B(rd) 2 O(!) k A quantifier elimination method for the first-order theory of the reals constructs for any input formula (SF) an equivalent quantifier free formula. Proposition 2.2 (Renegar [9]) There is an algorithm which, given a formula (SF ), finds an equivalent quantifier free formula of the form I where: 1 In this paper, all formulae will be explicitly written using the standard connectives -; and =) : ffl the degree of h ij (y) - (rd) 2 O(!) Q is one of the standard relations ?; !; -; =; 6=. The algorithm requires (rd) 2 O(!) k operations and (rd) O(k+ evaluations of the input formula. If the coefficients of the atomic polynomials are integers of bit length at most B; then the algorithm works with numbers of binary length This bound also holds for the binary length of the coefficients of the polynomials h ij . The following special case of the above result deals with the decision problem for the first-order theory of the reals: determine whether a sentence (SF) is true or false. Proposition 2.3 (Renegar [9]) There is an algorithm for the decision problem of the first-order theory of the reals that requires operations and (rd) O( evaluations of the input formula. When restricted to sentences involving only polynomials with integer coefficients of bit length at most B, the procedure works with numbers of binary length B(rd) 2 O(!) Q The inequality below is a well-known bound on nonzero roots of univariate polynomials (see, e.g., [5]). Proposition 2.4 Let a d be a univariate polynomial with integer coefficients, and let ff be a nonzero root of p(x). Then jffj is the height of p(x). We shall also need the following variant of the Fundamental Theorem of Linear Inequalities (see, e.g., [11]). Proposition 2.5 Consider a system of linear inequalities a T where a are non-zero vectors, b i 2 IR; let K be a convex set in Mg is not empty, then there exists a subset I ' M such that: (a) jIj - minfm;ng; Finally, for any R - 0 we have: To show the first of these identities, observe that where the last equality follows from the fact that A \Gamma - n (A)I n is a symmetric positive semidefinite matrix whose minimum eigenvalue is zero. To see the second identity, note that means that the minimum on the l.h.s. of (4) is negative and hence it is attained at a matrix X such that a consequence of (3). Upper Bounds on Feasible Solutions Theorem 3.1 (i) Any feasible system (1) has a solution X such that kXk - R, where log (ii) Moreover, if the feasible set of (1) is bounded, then the above bound holds for any solution of (1). Proof: Suppose that system (1) is feasible, and let (Recall that C is the cone of symmetric positive semidefinite matrices of order n.) From von Neumann's saddlepoint theorem (see, e.g., [10]) and (3), it follows, that for any R - 0, Consider the formula which can be written in the standard form (SF) as follows: (R - 0)g: It is easy to see that for any R 2 IR; the following statements are equivalent: ffl (1) has a feasible solution By our original assumption, (1) is feasible, and hence there is a nonnegative R that satisfies \Phi(R). Next, \Phi(R) is a standard formula (SF) of degree at most quantifiers. Furthermore, \Phi(R) consists of polynomial inequalities in m+ Since det( contains n! products of linear forms in m+1 variables with integer coefficients of height at most 2 l , each coefficient in \Phi(R) has binary length at most 1). Now from Proposition 2.1 it follows that \Phi(R) can be satisfied by a positive number R such that log By Proposition 2.5, there is a set I ' M of size at most n(n + 1)=2 such that the system is feasible, and any of its solutions solves the original system (1). For this rea- son, we can obtain a better bound on R by replacing m with minfm;n 2 g. Since of the theorem follows. To show part (ii), consider the formula \Phi R) g: Note that \Phi 0 (R) can be written in prenex form as (R It is easy to see that \Phi 0 (R) is satisfied if and only if Hence, we can apply Proposition 2.1 to \Phi 0 (R) to conclude that, similarly to (5), log It remains to show that m can be replaced by minfm;n 2 g. To this end, note that if the solution set of (1) is bounded, then there exists a system A with at most n(n + 1)=2 inequalities whose solution set is still bounded. This is because the solution set of (1) is bounded if and only if the recessive cone of (1) is trivial, i.e., is the halfspace fX 2 S n j A i ffl X - 0g. is equivalent to the emptiness of the intersection of the . By Helly's theorem (see, e.g., [10]) there exists a system of at most 1 whose intersection is still empty. Since any such system must contain\Omega 1 , the claim follows. Remark 3.2 The bounds of Theorems 3.1 apply to any mixed system of strict and/or non-strict inequalities where L is a linear subspace in IR n , and the constraint X -L 0 means that X is positive definite on L. This fact follows from the observation that for any t 2 (0; 1) and any solutions systems (7) and (1), respectively, the convex combination satisfies (7). Theorem 3.3 (i) Any feasible system (2) has a solution x such that log (ii) Moreover, if the feasible set of (2) is bounded, then the above estimate holds for any solution of (2). Proof: We can assume without loss of generality that the input matrices are linearly independent, and hence m - n(n 1)=2. Suppose that (2) is feasible and consider the formula or equivalently, \Psi(R) is a standard formula with quantifiers in n variables, respectively, which has polynomial inequalities of degree at most d = n, whose integer coefficients have at most Hence \Psi(R) can be satisfied by a number R such that log which implies part (i) of the theorem. To show part (ii), apply the above arguments to the modified formula Clearly, the bounds of Theorem 3.3 also hold for any strict and/or mixed system (2). We close this section with examples of ill-posed feasible systems (1) and (2). Example 3.4 Let n be an even number. Consider n \Theta n symmetric positive semidefinite matrices X satisfying the system of linear equations: It is easy to check that this instance of (1) is feasible and log X n;n - l2 n=2 for any of its solutions. A similar example for problem (2) is given below. Example 3.5 Given two matrices A 2 S n 1 , denote by their direct sum. For and x is a feasible instance of (2), any solution of which satisfies log xm - l2 m . Lower Bounds on the Discrepancy l) be the bound of Theorem 3.1, and let Rg: The discrepancy of (1) is the optimal value of the convex programming problem: Note that because of the compactness of CR , the minimum in (8) is always attained, and only if system (1) is feasible. Remark 4.1 There exist infeasible systems (1) such that inff ' j A For instance, this is true for the system of linear inequalities X 11 - is a symmetric positive semidefinite matrix of order 2. Theorem 4.2 If (1) is infeasible, then \Gamma log ' Although Theorem 4.2 can be proved analogously to Theorem 3.1, it is convenient to postpone its proof until Section 5. Now we consider systems (2). Let l) be the bound of Theorem 3.3, and let Rg be the m-dimensional ball of radius R centered at the origin. The discrepancy of (2) is the optimal value of the concave program: Clearly, (2) is feasible if and only if - - 0: Theorem 4.3 If (2) is infeasible, then \Gamma log(\Gamma- The proof of this theorem is also postponed until Section 5. We close this section with examples of infeasible systems (1) and (2) whose discrepancies are doubly exponentially small. Example 4.4 Let n be even, and consider n \Theta n symmetric positive semidefinite matrices X satisfying the equations: It is easy to check that this instance of (1) is infeasible and \Gamma log ' - l2 Example 4.5 For and is an infeasible instance of (2), and it can be verified that \Gamma log(\Gamma- 5 Complexity Bounds By Theorems 3.1 and 4.2, the feasibility of (1) can be determined by computing the optimal value ' of program (8) to an absolute accuracy of ffl, where This convex programming problem can be solved in O(n 4 log(2 l nR=ffl)) iterations of the ellipsoid method (see, e.g., [4]), where each iteration requires O(n 2 (m+n)) arithmetic operations over log(2 l nR=ffl)-bit numbers. Hence, we obtain an upper bound of lmn O(minfm;n 2 g) operations with testing the feasibility of (1). This result can be improved as follows. Theorem 5.1 The feasibility of (1) can be tested in mn O(minfm;n 2 g) arithmetic operations over Proof: We start with a weaker result. Lemma 5.2 The feasibility of (1) can be tested in (mn) O(minfm;n 2 g) arithmetic operations over l(mn) O(minfm;n 2 g) -bit numbers. Proof: The sentence states that (1) is feasible. Since the characteristic polynomial det(X \Gamma height 1, from Proposition 2.3 it follows that the validity of the above sentence can be determined in (mn) O(n 2 ) operations over To finish the proof of the lemma, it remains to show that the feasibility of (1) can also be decided in (mn) O(m) operations with l(mn) O(m) -bit numbers. Consider the sentence where \Phi(R) is the formula defined in the proof of Theorem 3.1. This sentence also states that (1) is feasible. Observe that (11) consists of polynomial inequalities of degree n in O(m) variables and has integer coefficients of binary size at most Since det( can be evaluated in poly(n; m) operations (or because all of its coefficients can be computed in n O(m) operations), the lemma follows from Proposition 2.3. We continue with the proof of Theorem 5.1. If m is bounded by a polynomial in n, the theorem follows from Lemma 5.2. We next show that for large m, determining the feasibility of (1) via Clarkson's algorithm [3] requires an expected mn O(minfm;n 2 g) operations over ln O(minfm;n 2 g) -bit numbers. Given a set I ' is the bound of Theorem 3.1 for the entire system (1). With this notation, we have ' by X(I) the (unique) least norm solution of the system A '(I)g be the set of constraints violated by X(I). A set I is called a basis, if V (J) 6= V (I) for any proper subset J ae I. A basis J is a basis for I, if J ' I and V for M is called optimal. In particular, if S is an optimal basis, then consequently, From Helly's theorem it follows that D : Given an optimal basis S, we can apply Lemma 5.2 to A and determine the feasibility of the original system (1) in n O(minfm;n 2 g) operations over finds an optimal basis by performing expected log m log m) - mpoly(n) violation tests. Each of these checks whether j 2 V (I) for a sample set I of cardinality log D) and an index Note that the inclusion can be written as the sentence where S I (X; ') is the quantifier g. Since the positive semidefiniteness of X can be expressed by the formula follows that V I;j is equivalent to: I (X; Each violation test can thus be represented by a sentence in prenex form with polynomial inequalities of degree In fact, by using the arguments of Section 4 in [3], one can verify that the above bounds on the number of violation tests and the size of sample sets are valid for computing an optimal basis for any mapping that satisfies the following two conditions: (i) V (I) ' M n I and (ii) any Note also that the coefficients of these polynomial inequalities are integers of binary length B - maxfl; log . Now from Proposition 2.3 it follows that each violation test can be accomplished in n O(minfm;n 2 g) operations over numbers. But the expected number of violation tests is bounded by mpoly(n). Hence we conclude that for all n and m, testing the feasibility of (1) requires expected mn O(minfm;n 2 g) operations over ln O(minfm;n 2 g) - bit numbers. Chazelle and Matousek [2] derandomized Clarkson's algorithm for a wide subclass of LP-type problems, which includes linear programming and the problem of computing the minimum volume circumscribed ellipsoid for a given m-point set in IR n . The analysis of their algorithm is based on an additional assumption which we state here in the following stronger form: there is a constant ~ D such that for any subset I ' M , all subsets of I violated by some (X; ') can be computed in D operations. Since computing the above set system can be accomplished by constructing the arrangement of the hyperplanes A i ffl (see the argument of [2] for linear programming), we have ~ Dg. The algorithm of [2] computes an optimal basis of (8) by performing mD O(D) operations and mpoly(D)+D O(D) violation tests with subsets I of size at most D. Since each violation test can be accomplished in n O(minfm;n 2 g) operations, we conclude that the derandomized algorithm still requires mn O(minfm;n 2 g) operations with ln O(minfm;n 2 g) -bit numbers. Corollary 5.3 The complexity bounds of Theorem 5.1 apply to the problem of computing an optimal basis of (8). Theorem 5.4 Given an optimal basis S of (8), in n O(minfm;n 2 g) operations over ln O(minfm;n 2 g) - bit numbers we can compute a univariate semialgebraic representation of ' , i.e., a system of univariate polynomial inequalities with integer coefficients such that ' is the only real solution of the system. In particular, ' is a root of a nontrivial polynomial h(') 2 ZZ['] such that log Proof: Assume w.l.o.g. that the given basis S coincides with M . In particular, 1)=2. From von Neumann's saddlepoint theorem and (4), it follows that for R - 0 Consider the formula where R the bound of Theorem 3.1. This formula states that ' - ' , and it can be written as follows: Now the formula in the sense that ' is the only real solution of ('). By consecutively applying Proposition 2.2 to (') and ('), the latter formula can be transformed into a quantifier operations with maxfl; log Rg(mn) O(m) - ln O(minfm;n 2 g) -bit numbers. (') is composed of univariate polynomial relations h(')40, where 4 2 f -; !; =; 6=; ?; - g: Since ' is the only real solution of ('), this formula can be transformed into an equivalent system of polynomial inequalities, which must contain a polynomial h such that h(' Remark 5.5 Under the assumption of Theorem 5.4, ' can be approximated to an accuracy operations (see Theorem 1.2 in [8]). Note that unlike the upper bound on the operations stated in Theorem 5.1, this bound depends on l. Remark 5.6 The minimal polynomial of an algebraic number ff is the primitive irreducible polynomial p(x) 2 ZZ[x] such that and the leading coefficient of p(x) is positive. The height of ff is the height of its minimal polynomial. Theorem 5.4 and the well-known inequality of Mignotte (see [5], p. 261) show that log height(' Theorem 5.4 immediately implies Theorem 4.2, whose proof was postponed in Section 4. Proof of Theorem 4.2. Suppose that system (1) is infeasible. Then ' ? 0 and by Theorem 5.4, the positive algebraic number ' is a root of a nontrivial polynomial h(x) 2 ZZ[x] with integer coefficients of bit length ln O(minfm;n 2 g) . Since Proposition 2.4 implies that height(h)), the theorem follows. The following result deals with the complexity of testing the feasibility of (2). Theorem 5.7 The feasibility of (2) can be determined in O(mn 4 operations over l is the maximum bit length of the entries of 1)=2, we can find a linearly independent subsystem of operations. We can thus assume that m - n(n+1)=2. The feasibility of (2) can be stated as the sentence By Proposition 2.3, the validity of the above sentence can be decided in n O(m) arithmetic operations with ln O(m) -bit numbers. It is easy to see that the discrepancy - of (2) satisfies h(- with a nontrivial polynomial ZZ[x] such that log This result and Proposition 2.4 imply Theorem 4.3. 6 Concluding Remarks Theorem 5.1 can be extended to the optimization version of problem where D is a given n \Theta n integral symmetric matrix. Specifically, in addition to testing the feasibility of (14); each of the following problems can also be solved in mn O(minfm;n 2 g) operations over ln O(minfm;n 2 g) -bit numbers: Boundedness. Determine whether the objective function is bounded from below on the set of feasible solutions. Infimum. Find a univariate semialgebraic representation for the finite infimum ff . Attainment. Determine whether the infimum is attained, that is, whether (14) has an optimal solution. Least Norm Optimal Solution. Find a univariate semialgebraic representation for each component of the least norm optimal solution of (14). Remark 6.1 The boundedness problem readily reduces to the feasibility problem. Although the attainment problem can also be reduced to the feasibility problem via the duality result of [7], the latter reduction polynomially increases both n and m and cannot be used in fixed dimension. Finally, for the optimization version of (G) with a given integral m-vector d: the above four problems can be solved in O(mn 4 )+n O(minfm;n 2 g) operations over ln O(minfm;n 2 g) - bit numbers. We plan to address these and some other extensions of Theorems 5.1 and 5.7 in a subsequent paper. --R A Randomized Scheme for Speeding Up Algorithms for Linear and Convex Programming Problems with High Constraints-to-Variables Ratio On Linear-Time Deterministic Algorithms for Optimization Problems in Fixed Dimension Las Vegas Algorithms for Linear and Integer Programming When the Dimension Is Small Geometric Algorithms and Combinatorial Optimization Bounds, in Buchberger An Algorithmic Analysis of Multiquadratic and Semidefinite Programming Problems An Exact Duality Theory for Semidefinite Programming and its Complexity Implications On the Computational Complexity of Approximating Solutions for Real Algebraic Formulae On the Computational Complexity and Geometry of the First Order Theory of the Reals. Convex Analysis Theory of Linear and Integer Programming --TR --CTR Xun Li , Xun Yu Zhou , Mustapha Ait Rami, Indefinite Stochastic Linear Quadratic Control with Markovian Jumps in Infinite Time Horizon, Journal of Global Optimization, v.27 n.2-3, p.149-175, November
complexity;bounds on solutions;semidefinite programming
596157
A Stochastic/Perturbation Global Optimization Algorithm for Distance Geometry Problems.
We present a new global optimization approach for solving exactly or inexactly constrained distance geometry problems. Distance geometry problems are concerned with determining spatial structures from measurements of internal distances. They arise in the structural interpretation of nuclear magnetic resonance data and in the prediction of protein structure. These problems can be naturally formulated as global optimization problems which generally are large and difficult. The global optimization method that we present is related to our previous stochastic/perturbation global optimization methods for finding minimum energy configurations, but has several key differences that are important to its success. Our computational results show that the method readily solves a set of artificial problems introduced by Mor and Wu that have up to 343 atoms. On a set of considerably more difficult protein fragment problems introduced by Hendrickson, the method solves all the problems with up to 377 atoms exactly, and finds nearly exact solution for all the remaining problems which have up to 777 atoms. These preliminary results indicate that this approach has very good promise for helping to solve distance geometry problems.
Introduction Distance geometry problems, which are concerned with determining spatial structures from measurements of internal distances, arise in the structural interpretation of nuclear magnetic resonance data and in the prediction of protein structure. Methods of calculating the conformations of biological molecules from distance constraints have become an important tool in structural biochemistry. Solving the distance geometry problem in this context would determine the three-dimension shape of the protein, which is critical for understanding its chemical and biological properties. For general reviews of the distance geometry problem and its relationship to the structure of chemical molecules, see Crippen and Havel[6], Havel[8], Kuntz, Thomason and Oshiro[10],and Br-unger and Nilges[1]. The distance geometry problem can be naturally formulated as a non-linear global optimization, where the objective function is constructed such that the distance constraints are satisfied at the global solutions of the prob- lem. A simple objective function can be defined to enforce the constraints. This optimization problem is believed to be computationally intractable in general because it has been shown to be strongly NP-complete in the one dimensional case [6, 13], and strongly NP-hard in the higher dimension case [14]. A large number of such methods for solving distance geometry problems have been proposed, such as Crippen and Havel[6], Havel[8], Hendrickson[9], Glunt, Hayden, Raydan[7], and Mor'e and Wu[12, 13]. The method we present in this paper is based on the stochastic/perturbation global optimization approach[4, 2, 3, 5] with several new features. The purpose of this paper is to describe this approach and to demonstrate its capabilities on some difficult distance geometry problems. The larger context of this research is to continue to develop and understand the capabilities of the stochastic/perturbation global optimization methodology, which we have found very successful for several large-scale global optimization problems arising from molecular chemistry. Our stochastic/perturbation algorithm combines a first, stochastic phase that identifies an initial set of local minimizers, with a second, more deterministic phase that moves from low to even lower local minimizers. In the second phase, by incorporating the partially separable structure of the problem, we work on very small-dimensional global optimization subproblems and then for local minimizers in the full-dimensional space. Both the selection of small dimensional subproblems and some other important algorithmic features are specific to the distance geometry problem. We experimented with our algorithm on Mor'e and Wu's artificial problems [12] and on the protein fragment problems from Hendrickson[9]. For the artificial problems, even our first phase can find the exact solutions with great success. For the protein fragment problems, which are considerably more difficult, we have found the exact solutions solutions for problems with up to 377 atoms (1131 parameters). Another important issue in dealing with distance constraints is that there may not exist any molecular structure satisfying the given distance con- straints, due to measurement or experimental errors. In practice, lower and upper bounds on distances are specified instead of exact distances. Therefore our algorithm is intended to deal with both exact and inexact distance geometry problems. For the larger protein problems of Hendrickson[9], with up to 777 atoms (2331 parameters), our algorithm has found approximate solutions with maximum relative error of distance at most 0.04. These computational successes on problems of considerable size appear to indicate that our algorithm is a powerful tool for solving distance geometry problems. The remainder of this paper is structured as follows. Section 2 describes distance geometry problems and current approaches developed for these problems. In section 3, we describe the stochastic/perturbation algorithm used to deal with the distance geometry problems. The framework of our algorithm is outlined, and several new features are discussed in the sec- tion. These are followed by extensive experimental results for our method on distance geometry problems in Section 4. Section 5 contains some brief conclusions and comments about future research. 2 The Distance Geometry Problems As mentioned earlier, solving the distance geometry problems is an important tool in determining the three-dimensional structure of a molecule. In the ideal case, distance geometry problems are concerned with finding positions 3 such that where S is a subset of the atom pairs, and is the given distance between atom i and j. Usually S has many fewer than m 2 =2 elements, that is only a small subset of pairwise distances is known. There may not exist any solution x to these distance con- straints, due to the error in the theoretical or experimental data. For exam- ple, this is guaranteed to happen if data for atoms violate the triangle inequality. In the more general distance geometry problem, lower and upper bounds on the distances are specified instead of exact values. In this case, the distance geometry problem is to find a set of positions x l i;j -k x where l i;j and u i;j are lower and upper bounds on the distance constraints, respectively. The distance geometry problems with constraints (2.1) and (2.2) can be naturally phrased as a nonlinear global optimization problem. The objective function is constructed so that the constraints are satisfied at the global minimizers of the optimization problem. One simple approach, which we utilize in this paper, is to penalize all the unsatisfied constraints. We formulate the distance geometry problem in terms of finding the global minimizers of the function (i;j)2S where l 2 This is the approach taken by Mor'e and Wu[13]. In order to compare results with those in Mor'e and Wu[12], we also use the following function for the exact problem: (i;j)2S where w i;j are positive weights. This function is same as (2.4) for the exact distance geometry problems if l i;j . Clearly, solves problem (2.2) if and only if x is the global minimizer of f(x) in (2.4) and problem (2.1), a solution x is equivalent to a global minimizer of the function (2.5) with function value 0. The difficulty in solving the distance geometry problem arises from several sources. First, the problem in itself is strongly NP-complete in one dimension, and strongly NP-hard in higher dimension[6, 14, 13], therefore it is very unlikely that an efficient algorithm for solving all cases of the problem can be found. Also there are huge numbers of local minimizers for the functions (2.4) and (2.5), which makes it very challenging to locate the basin of attraction of the global minimizer. It can be shown that function (2.5) has an exponential number of local minimizers[9]. The another important aspect is that most distance geometry problems are large and contain hundreds or more atoms. Thus even the calculation of a local minimizer makes heavy demands on the computer time. Most of the previous work by chemists on the distance geometry problems depends on the intrinsic properties of the chemical structure of solu- tions. Often, these methods use special heuristics based on chemical properties to generate initial configurations, and then perform local optimizations. The method of Hendrickson[9] is of particular interest to this paper because the more difficult test problems that we use are taken from that paper. Hen- drickson's method also utilizes chemical structure information. It works well on his test problems, which are generated from bovine pancreatic ribonucle- ase, a protein containing 124 amino acids, but relies on the assumption that the distances are highly accurate. In contrast, general global optimization approaches to the distance geometry problem do not utilize information about the solution structure. While utilizing problem-specific information may ultimately be desirable, it is important to understand the capabilities of general global optimization approaches for distance geometry problems. Work of this type includes the multistart method and Mor'e and Wu's continuation approach [12, 13]. In the former, a set of starting points are randomly generated and local minimization is performed from them. This method is simple and easy to implement. But it is highly unreliable in the context of distance geometry because the problems are usually large and have a huge number of local min- imizers. In the continuation approach, the original function is transformed into a smoother function with fewer local minimizers. A series of local optimization algorithms is then applied to the transformed function, tracing its minimizers back to those of the original function. The experimental results of Mor'e and Wu show that this approach is markedly superior to the multistart method. In the following section, we will describe our stochastic/perturbation algorithm for the distance geometry problem. Our algorithm is a general global optimization method, and applies to both exact and inexact problems. We do not utilize a smoothing approach as in the work of Mor'e and Wu, but our approach could readily be combined with smoothing techniques. Indeed, our experience with minimum energy conformation problems is showing that the combination of smoothing and our stochastic/perturbation global optimization approach is a promising framework (see e.g. [15]). 3 The Stochastic/Perturbation Global Optimization Method Our new algorithm for the distance geometry problems is based on the stochastic/perturbation global optimization method, which was constructed to solve large-scale global optimization problems. This approach has been successfully applied to Lennard-Jones problems [4], water cluster problems [2], and protein conformation problems [3]. The basic framework of our global optimization method for distance geometry problems is outlined in Algorithm 3.1 below. The method combines an initial phase that locates some low local minimizers with a second phase for moving from low to even lower local minimizers. During the first phase, a full dimensional random sample is generated over the domain space by randomly and independently placing each atom. The worst configurations are discarded, and the better ones are improved by selecting and moving an atom or a pair of atoms, until the function value for the configuration falls below a specified threshold level. A subset of these improved configurations is used as start points for a full dimensional local optimization algorithm. Some of the local minimizers found in this phase are passed on to the second phase for improvement. In the second phase, a local minimizer is successively selected for improvement as discussed below. A pair of atoms is chosen, and a small-scale stochastic global optimization algorithm is applied to the configuration with only these two atoms as variables and the remainder of the configuration fixed. This is followed by full dimensional local minimizations starting from the best configurations that resulted from the small-scale global optimization step. The lowest new configurations are then merged into the list of local minimizers, and this phase is iterated a fixed number of times. A key feature of this approach is that both phases make use of strategies that work on a very small subset of the atoms at once. In the initial phase this approach is used to improve the sample points by sampling on only one atom or a pair of atoms at a time in step 1b while leaving the remaining atoms temporarily fixed. When a pair of atoms is chosen, it is a pair for Algorithm 3.1 - Framework of the Large-Scale Global Optimization Algorithm for Distance Geometry Problems 1. Initial Generation of Configurations Phase : (a) Sampling in Full Domain : Randomly generate the coordinates of the sample points in the sampling domain, and evaluate f(x) at each new sample point. Discard all sample points whose function value is below a global "cutoff level". (b) One-atom/two-atom Sampling Improvement : For each remaining sample point : While the energy of the sample point is above the threshold value, Repeat: ffl Select the atom that contributes most to the function value or a pair of atoms in S that violates the ideal distance most ffl Randomly sample on the location of the selected atoms ffl Replace the atoms in the sample point with the new sample coordinates that give the lowest energy value. (c) Start Point Selection : Select a subset of the improved sample points from step 1b to be start points for local minimizations. (d) Full-Dimensional Local Minimizations: Perform a local minimization from each start point selected in step 1c. Collect some number of the best of these minimizers for improvement in Phase 2. 2. Improvement of Local Minimizers Phase: For some number of iterations (a) Select a Configuration : From the list of full-dimensional local min- imizers, select the local minimizer and a pair of atoms in S to be optimized Transform the configuration by multiplying the position of each atom relative to the center of mass of the configuration by a constant factor of between 1.25 and 2.0 or larger. (c) Two-atom Small Global Optimization : Apply a global optimization algorithm to the expanded configuration with only the two atoms chosen in step 2a as variables. (d) Full-Dimensional Local Minimization : Apply a local minimization procedure, using all the atoms as variables, to the lowest configurations that resulted from the two-atom global optimization. Merge the New Local Minimizers : Merge the new lowest configurations into the existing list of local minimizers. which a constraint distance is given in the problem formulation, that is an We will refer to this as a "constrained pair of atoms". When a constrained pair of atoms is used in step 1b, the sampling is done so that the distance between the atoms is always at the constraint value ffi i;j . In the second phase, this small subproblem approach is used to move a constrained pair of atoms in an existing configuration to new positions via the small-scale global optimization in step 2c. This small global optimization locates the best possible position for the selected atoms in the current configuration with the remaining atoms temporarily fixed. Again, when step 2c samples on the constrained pair of atoms, the distance between them is kept at the constraint value. All these small subproblem steps are very efficient due to the partial separability of the objective functions for the distance geometry problem, which make the cost of evaluating the function when only one or two atoms are moved much less expensive than the cost of a full function evaluation. The atom or pair of atoms that is chosen in each of these small-scale steps is one for which there seems to be potential for improving the overall configuration by moving this atom or pair. In phase 1, the atom which contributes most to the overall function value (and thus appears to have the greatest potential for reducing the function value) is selected. In phase 2 and alternatively in phase 1, the constrained pair of atoms that has the worst violation from the ideal distance is selected. In our early experiments, we also employed a one-atom small-scale global optimization heuristic in the second phase. The experimental results suggested that for some difficult problems the two-atom approach is more effective and that it locates lower configurations than the one-atom strategy, while for many other problems the two strategies are about equally effective. For this reason we have used the two-atom approach in phase 2. It may be that in some cases, moving only one atom is not enough to force the configuration out of the region of attraction of the current local minimizer. The choice of which local minimizer to improve in phase 2 is an important heuristic in this method. Through our experimentation on previous problems, we have concluded that a strategy that balances selecting a breadth of configurations with working on the best current configurations is most effective. This is the approach we have used for the distance geometry problem. Another important feature in Algorithm 3.1 is the expansion step, step b of phase 2, which expands the configuration around its center of mass prior to the small-scale global optimization. The expansion step was used first with minimum energy water cluster problems, where it significantly improved the ability of the phase 2 iteration to find improved configurations. This improvement apparently is due to expansion creating more room to move the atom or atoms that are the variables in the small-scale global optimization. The local optimization step then contracts the configuration again, into a new and hopefully improved local minimizer. The physical analogy of expansion is to heating in annealing. The drawback of expansion is that the local minimizations in phase 2 become more expensive. Our initial intuition was that the expansion would not be appropriate for distance geometry problems, because it leads to large violations in all the distance constraints. But we found that the algorithm without the expansion step didn't perform well in phase 2; after step 2d, most of the configurations returned to the local minimizer that the step had started with in step 2a. Possibly this is due to the relatively small number of constraints in distance geometry problems, which means that each atom is only involved in a small number of distance constraints. Therefore moving only the pair of atoms doesn't change the configuration sufficiently from the current local minimum. We found that when the expansion step is added, the small-scale global optimization step is far more successful at producing significantly different, and improved, configurations. We have experimented with different expansion factors in the range 1.25 to 2.5. For most of the problems, an expansion factor between 1.25 and 2.0 is most effective, but for some problems, an expansion factor between 2.0 and 2.5 is more helpful. In order to be consistent with the stochastic aspect of our algorithm, for most of the test problems in next section, the expansion factor is chosen randomly between 1.25 and 2.0. A final useful feature in our implementation, which is not captured in Algorithm 3.1, is a progression of problems approach. By this we mean that for some more difficult problems, we solve a sequence of problems of the form (2.4), with increasingly tight constraints (i.e. lower and upper bounds on the distances). Phase 1 is applied only to the first problem (with the loosest constraints), and phase 2 is applied to each problem, with the best configurations from the previous, looser-constrained problem forming the starting set of configurations for phase 2 for the next, tighter-constrained problem. The motivation for this approach came from experimentation. For the easier problems, applying Algorithm 3.1 to the problem directly was very successful. For the more difficult problems, we noticed that finding the global minimizer of function (2.4) directly was much more difficult than starting to work on the distance geometry problem with looser distance constraints, and then gradually tighting the constraints and applied Phase 2 until the expected accuracy. This approach is in some sense similar to the continuation methods. It has the additional advantage that for real problems, we may not know exactly how accurate the given distances are, and with this approach we can find the solutions to the best accuracy that is possible. In fact the chemists care more about finding reasonable solutions than exact solutions, and this approach coincides with this idea and yields an algorithm that can be tailored to find reasonable configurations. 4 Experimental Results We have run our algorithm on a set of artificial distance geometry problems from Mor'e and Wu [12, 13], and on a set of protein fragment test problems generated by Hendrickson[9]. In all of these problems, we are given the exact distances. In order to have both exactly and inexactly constrained problems for function (2.4), we used l 2 with values of epsilon that are given later. Most of our test problems are large, with hundreds or thousands of atoms. Therefore in our implemen- tation, we used the limited memory BFGS algorithm[11] to perform the full dimensional local minimizations, whereas the small-scale local minimizations within the small-scale global optimization step used the standard method. All the experiments were conducted on an Intel Paragon multiprocessor. In the artificial distance geometry problems from [12, 13], the molecule has atoms located in the three-dimensional lattice for some integer s - 1. The problem is to determine the structure of this molecule if we are given ffi i;j 2 S, where and r is an integer between 1 and m. Our first experiment for the artificial problem is to find the global minimizers using the exact function (2.5), using w First Problem Domain B Domain 2B m r #fval #gval #global #fval #gval #global 343 Table 4.1: Phase 1 results for Artificial Problems and Function (2.5) we ran Phase 1 for the molecules with (m; starting from the domain space and the domain 2B. (These are the same problems reported in [12, 13].) One-atom moves were used in step 1b. The numerical results are presented in Table 4.1, where # fval, #gval and #global are the number of function and gradient evaluations, the global solutions out of the total minimizers, respectively. Although Phase 1 can't find the global solution in either the smaller domain space or the larger one for the last two problems, recall that Phase 1 is simply designed to provide initial configurations for the second phase, which accounts for most of the work in the algorithm. In our experience, problems that can already be solved in Phase 1 are not particularly difficult global optimization problems. Note also that Phase 1 is in some sense similar to the multistart algorithm except for the one-atom or two-atom steps. Compared to the multistart results in [12], the phase 1 works quite well in this context. We also ran the same artificial problems using the exact version of function (2.4), i.e. with results for these runs are given in Table 4.2. Again, one-atom moves were used in step 1b. Table 4.2 shows that using function (2.4), Phase 1 already finds the global solutions for each problem in both cases. This indicates that these are not particularly difficult global optimization problems. The comparison of these to those in Table 4.1 seems to indicate that using the exact version of function (2.4) creates an easier problem than using function (2.5). We believe that this is because function (2.4) is smoother. Finally, there is a marked difference between Problem Domain B Domain 2B m r #fval #gval #global #fval #gval #global 343 Table 4.2: Phase 1 results for Artificial Problems and Function (2.4) the results in Table 4.2 and the multistart results given for these problems and function (2.4) in [13], in which the global minimizers were never found. However these results are not directly comparable because the results in [13] are for (2.4) with Tables 4.3 and Table 4.4 give the results from applying Phase 2 to the same set of artificial problems, again using functions (2.5), and (2.4) with In these tables, FLS, SLS are the total number of full dimensional and small dimensional local optimizations, respectively, and Ffval and Pfval are the number of full function evaluations and partial (only two atoms change) function evaluations, respectively. The results show that for function (2.5), applying Phase 2 allows the method to readily solve the two problems not solved in Phase 1 and to relocate the global minimizers for all the problems many times. For function (2.4), more than half the local minimizations find the global minimizer. These results show that Algorithm 3.1 is very successful on these artificial problems. A comparison with the efficiency of the method of Mor'e and Wu [12, 13] is difficult in part because they do not give costs in the second of these papers, but mainly because smoothing is a very useful technique that makes problems easier to solve, but has not been used in our algorithm. We then proceeded to a set of significantly more difficult test problems. These problems are generated by Hendrickson [9] from the bovine pancreatic ribonuclease, a typical, rather small protein containing 124 amino acids. Hendrickson derived the set of twelve test problems by using fragments consisting of the first 20, 40, 60, 80, and 100 amino acids as well as the full protein, with two sets of distance constraints for each size. The distance information given for these problems is exact, meaning that there is a global Problem FLS #Ffval SLS #Pfval #global 343 50 33624 15168 1827104 9/40 Table 4.3: Phase 2 results for Artificial Problems and Function (2.5) Problem FLS #Ffval SLS #Pfval #global 343 100 17853 39068 4097761 47/81 Table 4.4: Phase 2 results for Artificial Problems and Function (2.4) minimizer where all constraints are satisfied exactly. The problems have from 63 to 777 atoms (189 to 2331 parameters). Mor'e and Wu [13] tested their method on the smallest of these problems, using function (2.4), and reported success for the values but not below. For each of the twelve problems of Hendrickson [9], we applied the Phase 2 repeatedly to the function (2.4) with decreasing ffl until In our runs, we usually chose the values used by Mor'e and Wu [13]. For the two smallest problems, however, we started from 0.02, whereas for the largest problems, we may find the global solutions for the subproblem and then proceed to the subproblems. For each value of ffl, we generally performed 50 iterations in Phase 2. We applied Phase 1 on the first subproblem only, using the constrained pair of atom moves in step 1b. Our numerical results on the Hendrickson problems are summarized in the Table 4.5. The first two columns record the number of atoms and the given distance constraints, the third column is last ffl-problem which we can solve successfully in the progression of problems approach, and the final column is the maximum relative error of the distances with respect to the ideal distances in our best solution to this subproblem, where the maximum relative error is defined as Table 4.5 shows that for the first seven problems, which have 63 to 377 atoms (189 to 1131 parameters) we can find the global solutions to the problems. For the remaining five problems, which have 472 to 777 atoms (1416 to 2331 parameters), we solve the problems down to relative accuracy levels of between 0.01 and 0.04. These results indicate that our approach has the potential to locate exact or nearly exact solutions of quite large distance geometry problems without using any particular structural information. Finally, we briefly discuss the costs of Algorithm 3.1 on the protein fragment problems. We have analyzed the costs of two typical runs for the 63-atom and 102-atom problems. For the 63-atom problem, the entire solution process required 692,488 full function evaluations and 2,416,214 partial function evaluations. For the 102-atom problem, the entire solution process required 498,500 full function evaluations and 952,836 partial function evaluations. Detailed timings show that the full dimensional local optimization steps in Phase 1 and 2 dominate the cost of the algorithm for these problems, accounting for 80% of the total time. The other steps are relatively inexpensive. Therefore a cheaper local optimization algorithm will Number of Distance Final Final Atoms Constraints Subproblem Relative Error 236 957 0.00 8.1e-4 287 1319 0.00 8.1e-4 Table 4.5: Results on Protein Fragment Problems reduce the total cost of our algorithm. An alternative way to reduce the cost would be to work in internal coordinates. The advantage of this approach is that the number of variables is greatly reduced, generally by a factor of ten or more. Although the problem is no longer partially separable in this parameterization, the overhead in re-evaluation of the partial function and gradient values is relatively minor compared to the reduction in the local optimization cost. A disadvantage of this approach is that it already uses some structural information of the problem, namely the primary structure of the protein, which may be considered contrary to the goal of assessing the potential of general-purpose global optimization methods for distance geometry problems. 5 Summary and Future Research We have presented a new global optimization algorithm intended to solve exactly or inexactly constrained distance geometry problems. The algorithm utilizes small-scale global optimization calculations on selected subsets of the parameters, performed by a stochastic global optimization method, as a key part of its approach. Its structure is related to our previous stochas- tic/perturbation global optimization methods for molecular clusters and protein folding, but there are several important differences. In particular, three algorithmic choices helped in the success of the method and none were anticipated by us initially. The first is to expand the configuration before each small-scale global optimization. The second is to select a linked pair of atoms rather than a single atom as the parameters for the small-scale global opti- mization. The third is to solve a progression of distance geometry problems where the constraint bounds become tighter and tighter. Our computational tests on artificial problems with up to 343 atoms (1029 variables) and on more difficult protein fragment problems with up to 777 atoms (2331 variables) indicate that the method is very successful in located the configurations satisfying the given distance constraints. The results on the protein problems especially indicate that the method is quite successful in solving large and apparently difficult distance geometry problems without using any information about the solution structure. There are many directions for continuing this research that we are con- sidering. One is to use more than two atoms in the small-scale global optimization step of phase 2. While this increases the cost of the small-scale global optimization, this is not a dominant cost, and with such a strategy we were able to solve the 480 atom problem at the subproblem 0.00 level. A second possible direction is to use a buildup approach, where one finds (approximate) solutions to subsets of the problem on the way to solving the full problem. The third, and perhaps most intriguing direction, is to combine a smoothing approach like that of Mor'e and Wu [12, 13] with our stochastic/perturbation global optimization technique. Our recent research in protein folding has shown that this combination has great potential. Fi- nally, if this stochastic/perturbation approach were to be part of a production distance geometry code, one would want to find ways to combine it with the chemical structural knowledge that chemists use in the solution of these problems. Acknowledgement We thank Zhijun Wu and Bruce Hendrickson for providing the test problems, and also thank Zhijun Wu for many helpful communications about this research. --R Computational Challenges for macromolecular structure determination by X-ray crystallography and solution NMR-spectroscopy A New Stochastic/Perturbation Method for Large-Scale Global Optimization and its Application to Water Cluster Problems A Parallel Global Optimization Method for Solving Molecular Cluster and Polymer Conformation Problems A New Large-Scale Global Optimization Method and its Application to Lennard-Jones Problems Global Optimization Methods for Protein Folding Problems Distance Geometry and Molecular Conformation Molecular Conformation from Distance Matrices An Evaluation of Computational Strategies for Use in the Determination of Protein structure from Distance Geometry Constraints Obtained by Nuclear Magnetic Resonance The Molecule Problem: Exploiting Structure in Global Optimization Distance Geometry On the Limited Memory BFGS Method for Large Scale Optimization Global Continuation for Distance Geometry Problems Embeddability of Weighted Graphs in k-space is Strongly NP- hard Global Optimization for Molecular Clusters Using a New smoothing approach --TR --CTR Le Thi Hoai An, Solving Large Scale Molecular Distance Geometry Problems by a Smoothing Technique via the Gaussian Transform and D.C. Programming, Journal of Global Optimization, v.27 n.4, p.375-397, December
lennard-jones cluster;global optimization;molecular conformation
596192
A Constraint Satisfaction Approach to a Circuit Design Problem.
A classical circuit-design problem from Ebers and Moll (1954) features a system of nine nonlinear equations in nine variables that is very challenging both for local and global methods. This system was solved globally using an interval method by Ratschek and Rokne (1993) in the box [0, 10]^9. Their algorithm had enormous costs (i.e., over 14 months using a network of but they state that at this time, we know no other method which has been applied to this circuit design problem and which has led to the same guaranteed result of locating exactly one solution in this huge domain, completed with a reliable error estimate. The present paper gives a novel branch-and-prune algorithm that obtains a unique safe box for the above system within reasonable computation times. The algorithm combines traditional interval techniques with an adaptation of discrete constraint-satisfaction techniques to continuous problems. Of particular interest is the simplicity of the approach.
Introduction The transistor modeling problem of Ebers and Moll [6] is the system of nonlinear equations where the constants g ik are given by 5.2095 10.0677 22.9274 20.2153 28.5132 111.8467 134.3884 211.4823 The problem is very challenging both for local and global methods because small variations in the inputs produce large differences in the functions. Ratschek and Rokne [23] summarize various attempts to find a solution to this problem using local methods; these descriptions are not repeated here. It suffices to say that successful attempts require very elaborate procedures, sometimes combining several globally convergent algorithms. In the same paper [23], Ratschek and Rokne propose an interval method which solves the problem globally in the box [0; 10] 9 . In particular, they show that there exists a unique solution in this box and they provide a guaranteed error estimate by enclosing the solution in a box whose intervals are of width smaller than 3:2 \Theta 10 \Gamma4 . The computation times to obtain this result are, however, extremely large. No computation times were given in [23], since this was not the primary aim of the paper. However, a preliminary draft of the paper indicated that the total process took over 14 months using a network of workstations. Ratschek and Rokne also recommended that such methods for solving unstable systems should be further investigated in order to reduce the computations. Here we present a novel interval algorithm that isolates all safe boxes (i.e., all boxes which contain at least one solution) to nonlinear equation systems and show that the algorithm isolates a single safe box for the circuit-design problem within reasonable computation times. 1 The main novelty of the procedure is in the way in which constraints are used to prune the search space. The pruning techniques, some of which were presented in [24] and some of which are novel, are based on constraint-satisfaction techniques from artificial intelligence and are particularly effective when far from a solution. These techniques are thus orthogonal to traditional interval techniques, which are most effective close to a solution (when the boxes are small). The paper is organized as follows. Section 2 gives the necessary background in interval analysis. Section 3 discusses the problem description and the simplyfying assumptions. Section 4 presents a generic branch-and-prune algorithm for isolating all solutions to a nonlinear system of equa- tions, Section 5 presents the various pruning techniques, and Section 6 reconsiders the simplifying assumptions. Section 7 reports the experimental results and Section 8 concludes the paper. Preliminaries Here we review some basic concepts of interval analysis to needed for this paper. More information on interval arithmetic can be found in many places (e.g., [1, 8, 7, 17, 18, 20, 22]). Our definitions are slightly non-standard. 2.1 Interval Arithmetic We consider 1g, the set of real numbers extended with the two infinity symbols, and the extension of the relation ! to this set. We also consider a finite subset F of ! 1 containing In practice, F corresponds to the floating-point numbers used in the implementation. is the set of real numbers 1 Proving that the safe box contains a unique solution requires some experimental trial-and-error work; see Section 6 of [23] for a discussion of this issue. The set of intervals is denoted by I and is ordered by set inclusion. 2 S be a subset of !. The enclosure of S, denoted by S or 2S, is the smallest interval I such that S ' I . We often write r instead of frg for r 2 !. We denote real numbers by the letters F-numbers by the letters l; m; u, intervals by the letter I , real functions by the letters f; g and interval functions (e.g., functions of signature I ! I) by the letters F; G, all possibly subscripted. We use l to denote the smallest F-number strictly greater (resp. smaller) than the F-number l. To capture outward rounding, we use dre (resp. brc) to return the smallest (resp. largest) F-number greater (resp. smaller) or equal to the real number r. We also use ~ I to denote a box hI and ~r to denote a tuple hr is the set of rational numbers and N is the set of natural numbers. We also use the following notations: right([l; canonical interval is an interval of the form [l; l] or [l; l where l is a floating-point number. The fundamental concept of interval arithmetic is the notion of interval extension. Definition 4 [Set Extensions] Consider a set S ' ! n and a function !. The set extension of f is defined as I is an interval extension of I 0 if Example 1 The interval function \Phi defined as is an interval extension of addition of real numbers. In this paper, we restrict attention to monotonic interval extensions because of their fundamental properties and because traditional interval extensions of primitive operations satisfy these requirements naturally. Definition 6 [Monotonic Interval Extensions] An interval function F : I n ! I is monotonic in ~ I 0 if I 2 ' ~ I 1 ' ~ I 2 These intervals are usually called floating-point intervals in the literature. It is important to stress that a real function can be extended in many ways. For instance, the interval function \Phi is the most precise interval extension of addition (i.e., it returns the smallest possible interval containing all real results), while a function always returning [\Gamma1; 1] would be the least accurate. In the following, we assume fixed monotonic interval extensions for the primitive operators (for instance, the interval extension of + is defined by \Phi). In addition, we overload the real symbols and use them for their interval extensions. 2.2 Constraint Representations It is well known that different computer representations of a real function produce different results when evaluated with floating-point numbers on a computer. As a consequence, the way in which constraints are written may have an impact on the behavior on the algorithm. For this reason, a constraint or a function in this paper is considered to be an expression written in some language. In addition, we abuse notation by denoting functions (resp. constraints) and their representations by the same symbol. In the remaining sections, we assume that real variables in constraints are taken from a finite (but arbitrarily large) set fx g. Similar conventions apply to interval functions and constraints. Interval variables are taken from a finite (but arbitrarily large) set fX g. 3 Problem Description The problem considered in this paper is finding all solutions to a system of equations in a box ~ I Of course, on a computer, it is generally impossible to find these solutions exactly and interval methods aim at returning small boxes containing all solutions. Preferably, each such box is safe, meaning that it contains a solution. For the purposes of this paper, interval methods can thus be viewed as solving the following problem: assuming that F i is a monotonic interval extension of f i (1 - i - n), find all canonical 3 boxes hI I 0 satisfying This is of course a simplification, since interval methods generally use various interval extensions. However, restricting attention to this problem has the benefits of crystallizing the intuition underlying our novel pruning methods. Section 6 reconsiders this simplification. 3 In practice, one may be interested in boxes of a certain width or one may want to stop as soon as a safe box is obtained. It is easy to generalize our results to these requirements. Notation: Let S be a system of constraints of the form and let ~ I be a box hI We denote by S( ~ I) and S(I the fact that ~ I satisfies the system of interval constraints S or, in symbols, Note also that we use S to denote systems of constraints and S to denote systems of interval constraints. 4 The Generic Branch-and-Prune Algorithm The above problem description highlights the finite nature of the problem, since there are only finitely many floating-point intervals. Most interval methods are thus best viewed as global search algorithms iterating two main steps: pruning and branching. The basic schema underlying these algorithms, the branch-and-prune schema, is depicted in Figure 1. The function Search receives a system of interval constraints S and an initial box ~ returns the set of canonical boxes ~ I in ~ I 0 satisfying S( ~ I). The function Search first applies a pruning step that reduces the initial box. This pruning step is the main topic of this paper. If the resulting box ~ I is empty, there is no solution to the problem. If the box ~ I is canonical, it is returned as a result. Otherwise, the box is split along one dimension into two subboxes, ~ I 1 and ~ I 2 , which are then explored recursively using the same algorithm. A specific interval algorithm can be obtained by specifying the splitting strategy and pruning techniques. Our algorithms use a splitting strategy that consists of splitting the largest interval in two parts. The main novelty of our algorithms lies in the pruning techniques and we will define three pruning operators, Prune 0 , Prune 1 , and Prune 2 , that produce three distinct algorithms. 5 The Pruning Techniques The two pruning techniques used in our algorithm are box(1)- and box(2)- consistency. To ease understanding, these techniques are contrasted with a traditional interval technique, which we call box(0)-consistency for reasons that will become clear below. 5.1 Box(0)-Consistency It is traditional in branch-and-prune algorithms to use a relaxation of the problem at hand. If there is no solution to the easier problem, it follows that there are no solutions to the original problem. Box(0)-consistency is a weak, but very simple, relaxation used in most interval systems. Given the function Search(S, ~ I begin ~ I := Prune(S, ~ I 0 ); if Empty( ~ I) then return else if Canonical( ~ I) then return f ~ Ig else I return Search(S, ~ I 2 ); Figure 1: The Branch-and-Prune Algorithm problem of finding canonical boxes in a box ~ I satisfying a system of interval constraints S, box(0)- consistency consists of testing S( ~ I). If S( ~ I) does not hold, there are obviously no solutions to the original problem, because of the definition of interval extensions. The pruning operator associated with box(0)-consistency can thus be defined as follows: ~ I otherwise. Box(0)-consistency can in fact be viewed as a form of projection. The original problem could be stated as an existence question and box(0)-consistency approximates it by replacing each interval variable by its interval to obtain the test which reduces to testing each constraint in S independently. 5.2 Box(1)-Consistency This section presents the first pruning operator used in our algorithm. It starts with an informal discussion, then specifies the pruning operator, and presents a simple implementation. 4 4 Separating the specification from the implementation has the advantage of distinguishing what is being computed from how to compute it. There are many ways to implement the concepts described in this section, and our goal here is to focus on the concepts, not on the technical details (which can be found elsewhere [24]). 5.2.1 Informal Presentation The first fundamental idea underlying box(1)-consistency [2] is to project all variables but one or, more precisely, to replace all variables but one by their intervals. This produces a stronger pruning than box(0)-consistency but, of course, at a higher cost. The original existence problem is thus approximated by This relaxation can be tested relatively easily. Notice first that the conditions are independent. In addition, a condition of the form can be tested by considering all the canonical intervals I in I 1 and testing Our implementation tries to be more effective by using adaptations of the interval Newton method. The second fundamental idea underlying box(1)-consistency is to reduce the intervals associated with the variables. Consider the relaxation and let I l be the leftmost interval in I 1 satisfying and I r the rightmost interval in I 1 satisfying It should be clear that X 1 must range in the interval I 0 I since any interval on the left or on the right of I 0 violates one of the conditions of the relaxation. The interval associated with X 1 can thus be reduced to I 0 . This reduction is applied for each of the variables until no more reduction takes place. The resulting box is said to be box(1)-consistent. In the course of this process, it is possible to detect that no solution to the original problem exists, since the intervals associated with the variables become smaller. Note also that a variable can be considered several times in this reduction process. 5.2.2 The Pruning Operator We now formalize the informal discussion above and present the pruning operator associated with box(1)-consistency. Recall that all definitions assume that S is defined over the set of variables . The main concept is box(1)-consistency, which expresses that a system cannot be narrowed further by the reduction process described informally in the previous subsection. Definition 7 [Box(1)-Consistency] Let S be a system of interval constraints, let ~ I be a box n). S is box(1)-consistent in ~ I and S is box-consistent in ~ I if it is box(1)-consistent in ~ I wrt all i in The pruning operator associated with box(1)-consistency simply returns the largest box in the initial interval that is box(1)-consistent (or, more informally, the largest box in the initial interval that cannot be narrowed further). This box always exists because of the monotonicity of interval extensions and is unique. Of course, it can be empty. Definition 8 [Box(1)-Consistency Pruning] Let S be a system of interval constraints and let ~ I be a box. The pruning operator associated with box(1)-consistency can be defined as I 0 where ~ I 0 is the largest box in ~ I such that S is box(1)-consistent in ~ I 0 . 5.2.3 A Simple Implementation The pruning operator can be computed in many ways. Figure 2 presents a simple iterative algorithm for this purpose; see [24] for a more efficient implementation. The algorithm is a simple fixpoint algorithm that terminates when no further pruning can be obtained (i.e., ~ I p ). The body of the loop applies a narrowing operation on each of the variables and produces a new box that is the intersection of all these narrowings. The narrowing function Narrow 1 (S, ~ I,i) returns the largest box ~ I 0 in ~ I such that S is box(1)-consistent in ~ I 0 wrt i. 5.3 Box(2)-Consistency Box consistency has been shown to be effective for solving a variety of nonlinear applications [24]. For some applications, however, and for the transistor modeling problem in particular, better performance can be obtained by using a stronger local consistency condition that we call box(2)- consistency. Box(2)-consistency is in fact an approximation of path consistency [16] and is related to some consistency notions presented in [14]. function Prune 1 (S, ~ begin repeat ~ I p := ~ until ~ I = ~ I return ~ I; function Narrow 1 (S,hI begin I c is canonical and S(I return ;; else return Figure 2: Implementing Box(1)-Consistency 5.3.1 Informal Presentation Box(2)-consistency generalizes box(1)-consistency by projecting all but two variables. The original existence problem is thus approximated by 9X Once again, it is possible to test this relaxation easily, at least from a conceptual standpoint. The conditions are independent and a condition of the form can be tested by considering all pairs of canonical intervals in I 1 and I 2 for X 1 and X 2 . As was the case for box(1)-consistency, box(2)-consistency makes use of this relaxation to prune the intervals associated with each variable. Consider a condition and a canonical interval I 0 . If there is no box ~ I such that S is box(1)- consistent in ~ I 0 , then obviously 1 . It is thus possible to narrow the bounds of I 1 by using this pruning rule, and this process can be iterated for all variables until no further pruning is available. 5.3.2 The Pruning Operator The notion of box(2)-consistency is defined in terms of box(1)-consistency or, more precisely, in terms of whether a system of interval constraints can be made box(1)-consistent in a box. Definition 9 [Box(1)-Satisfiability] A system of interval constraints S is box(1)-satisfiable in ~ I 0 , denoted by BoxSat 1 (S; ~ I 0 ), if there exists a box ~ I ' ~ I 0 such that S is box(1)-consistent in ~ I. Informally speaking, box(2)-consistency says that the bounds of each variable are box-satisfiable, implying that they cannot be reduced further using the pruning rule described above. S be a system of interval constraints and ~ I be a box is box(2)-consistent in ~ I wrt i if when l i 6= u i and if The system S is box(2)-consistent in ~ I if it is box(2)-consistent in ~ I wrt all It can be shown that box(2)-consistency implies box(1)-consistency. Proposition 1 Let S be a system of monotonic interval constraints. If S is box(2)-consistent in ~ I , then S is box(1)-consistent in ~ I . Proof Assume for simplicity that ~ I . The proof is similar otherwise. Since S is box(2)-consistent in ~ I , S is box(2)-consistent in ~ I wrt all i (1 - i - n). Thus, or, more explicitly, 9 ~ I is box(1)-consistent in ~ It follows that and, by monotonicity of F i , that The proof that is similar. The result follows from the definition of box(1)-consistency. 2 The pruning operator associated with box(2)-consistency simply returns the largest box in the initial interval which is box(2)-consistent. Definition 11 [Box(2)-Consistency Pruning] Let S be a system of interval constraints and let ~ I be a box. The pruning operator associated with box(2)-consistency can be defined as I 0 where ~ I 0 is the largest box in ~ I such that S is box(2)-consistent in ~ I 0 . 5.3.3 A Simple Implementation Once again, the pruning operator can be computed in many ways. Figure 3 presents a simple iterative algorithm close to our actual implementation. The algorithm is again a simple fixpoint algorithm that terminates when no further pruning can be obtained. The body of the loop applies a narrowing operation on each of the variables and produces a new box that is the intersection of all these narrowings. The narrowing function Narrow 2 (S, ~ I,i) returns the largest box ~ I 0 in ~ I such that S is box(2)-consistent in ~ I 0 wrt i. The pruning operator of box(1)-consistency is used to compute this narrowing operator. Note that it would be easy to define any level of box-consistency at this point, since box(k-1)-consistency can be used to define box(k)-consistency in a generic way. 5.4 Related Work It is useful to relate these pruning operators to earlier work in constraint satisfaction. Projec- tions, and approximations of projections, have been used extensively in the artificial intelligence community (under the name consistency techniques) to solve discrete combinatorial problems (e.g., [16, 15]). They have been adapted to continuous problems (e.g., [5, 14]) and used inside systems such as BNR-Prolog and CLP(BNR) [21, 3] and many systems since then. The techniques used in systems like BNR-Prolog and CLP(BNR) are weaker than box(1)-consistency, since they decompose all constraints into ternary constraints on distinct variables before applying a form of box(1)-consistency. They do not scale well on difficult problems. Box(1)-consistency was introduced in [2]. It is related to the techniques presented in [10], which uses a similar idea for the splitting process. The consistency notions of [14] are a weaker, and less effective, form of box(k)-consistency: it is obtained by decomposing all constraints into ternary constraints over distinct variables and by applying a form of box(k)-consistency on the resulting constraints. function Prune 2 (S, ~ begin repeat ~ I p := ~ until ~ I = ~ I return ~ I; function Narrow 2 (S,hI begin I c is canonical and return ;; else return Figure 3: Implementing Box(2)-Consistency 6 The Branch-and-Prune Algorithm Revisited We now reconsider the assumptions of Section 3. As mentioned in that section, our algorithm uses two interval extensions, the natural interval extension and the mean value interval extension, since distinct interval extensions may produce different prunings. In particular, the natural interval extension is generally better when far from a solution, while the mean value interval extension is more precise when close to a solution. This section reviews both extensions and reconsiders the overall branch-and-prune algorithm. 6.1 The Natural Interval Extension The simplest interval extension of a function is its natural interval extension. Informally speaking, it consists in replacing each number by the smallest interval enclosing it, each real variable by an interval variable and each real operation by its fixed interval extension. In the following, if f is a real function, b f is its natural extension. Example 2 [Natural Interval Extension] The natural interval extension of the function x 1 is the interval function The advantage of this extension is that it preserves how constraints are written and hence users of the system can choose constraint representations particularly appropriate for the problem at hand. It is useful to generalize the natural interval extension to a system of constraints. Definition 12 [Natural Interval Extension of a System] Let S be a system of constraints The natural extension of S, denoted by b S, is the set of the interval constraints The following result is easy to prove by induction. Proposition 2 The natural interval extension of a function, a constraint, or a system of constraints is monotonic. 6.2 The Mean Value Interval Extension The second interval extension is based on the Taylor expansion around a point. This extension is an example of centered forms that are interval extensions introduced by Moore [17] and have been studied by many authors, since they have important properties. The mean value interval extension of a function is parametrized by the intervals for the variables in the function. It also assumes that the function has continuous partial derivatives. Given these assumptions, the key idea behind the extension is to apply a Taylor expansion of the function around the center of the box and to bound the rest of the series using the box. I be a box hI be the center of I i . The mean value interval extension of a function f in ~ I, denoted by -(f; ~ I), is the interval function d @f Let S be a system of constraints The mean value interval extension of S in ~ I , denoted by -(S; ~ I), is the system of interval constraints function Search(S, ~ I begin ~ I := Prune( I 0 ), ~ I 0 ); if Empty( ~ I) then return else if Canonical( ~ I) then return f ~ Ig else I return Search(S, ~ I 2 ); Figure 4: The Branch-and-Prune Algorithm Revisited Note that the mean value interval extensions is defined in terms of natural interval extensions. The proof of the following proposition can be found in [4]. Proposition 3 The mean value interval extension of a function is a monotonic interval extension. It is interesting to note that box consistency on the mean value interval extension of a system of constraints is closely related to the Hansen-Sengupta operator [8], which is an improvement over Krawczyk's operator [13]. Hansen and Smith also argue that these operators are more effective when the interval Jacobian of the system is diagonally dominant [9] and they suggest conditioning the system S. For the purpose of this paper, we simply assume that we have a conditioning operator use the notation - c (S; ~ I) to denote -(cond(S; ~ I); ~ I). See [11, 12] for extensive coverage of conditioners. 6.3 The Branch-and-Prune Algorithm We are now in position to reconsider our branch-and-prune algorithm. The new version, given in Figure 4, differs in two ways from the algorithm presented in Section 4. On the one hand, the algorithm receives as input a system of constraints (instead of a system of interval constraints). On the other hand, the operation Prune receives a system of interval constraints consisting of the natural interval extension and the conditioned mean value interval extensions of the original system. 6.4 Existence Proof We now describe how the algorithm proves the existence of a solution in a box. Let ff 0g be a system of equations over variables fx be a box and define the intervals I 0 n) as follows I 0 d I 0 then there exists a zero in hI 0 i. A proof of this result can be found in [19]. 7 Experimental Results This section reports experimental results of the branch-and-prune algorithms. We compare the branch-and-prune algorithm with two instantiations of the pruning operator: Prune 1 to obtain the algorithm presented in [24] and Prune 2 to obtain a novel algorithm more effective for the transistor modeling problem. The Transistor Modeling Problem The box(2)-consistency algorithm was applied to find all solutions of the transistor modeling problem. Branching was applied until a safe box or a box of width smaller than 10 \Gamma8 was obtained. The algorithm returned a unique box in the original range [0; 10] 9 , together with a proof that the box contains a solution. The algorithm performs only 118 branchings and takes 2359.80 seconds (roughly 40 minutes) on a Sun Ultra- running Solaris. This is of course a considerable improvement over the results of Ratschek and Rokne [23]. Interestingly, the box(1)-consistency algorithm performs 135099 branchings and takes 11841 seconds (roughly 3 hours and 20 minutes) on the same machine, still a considerable improvement over [23]. Benchmarks from Continuation Methods It is worth comparing the box(1)- and box(2)- consistency algorithms on some standard benchmarks from continuation methods [25]. The box(1)- consistency algorithm compares well with continuation methods on these benchmarks [24]. Table 1 reports the results and gives, for each benchmark, the number of variables, the degree of the polynomial system, the initial range of the variables, and the CPU time and number of branchings of Benchmarks v d range Box(1)-Time Box(1)-Branch Box(2)-Time Box(2)-Branch combustion chemistry 5 Table 1: Box(1)- Versus Box(2)- Consistency on Some Benchmarks from Continuation Methods. the box(1)- and box(2)-consistency algorithms. See [25, 24] for a description of the benchmarks. The intention is not to compare the two algorithms systematically but rather to make readers aware that neither of two algorithms is really superior. As can be seen, the box(2)-consistency algorithm always performs less branchings (as should be expected) but is always slower than the box(1)-consistency algorithm. On these problems, box(1)-consistency seems to give a better tradeoff between pruning and pruning time. A fundamental topic for future research is thus to determine when box(2)- consistency is more effective than box(1)-consistency and, more generally, to characterize the class of nonlinear problems for which these techniques are effective. 8 Conclusion This paper reconsidered the transistor modeling problem from Ebers and Moll [6], which consists of nine nonlinear equations and is challenging both for local and global methods. The problem was tackled by a novel branch-and-prune algorithm combining techniques from interval methods and constraint satisfaction. In particular, the algorithm enforces a local consistency condition called box(2)-consistency that strengthens the notion of box(1)-consistency introduced in [24]. The algorithm was applied to find all solutions to the transistor modeling problem and returned a unique safe box in the range [0; 10] 9 in about 40 minutes, performing only 118 branchings. The paper also indicated that box(2)-consistency may be too strong a local condition for many problems, since it is slower than box(1)-consistency on benchmarks from continuation methods [25]. An interesting avenue of research is to characterize more formally the class of applications for which box(1)- and box(2)-consistency are effective pruning techniques. Acknowledgments Special thanks to Christian Bliek for bringing the transistor modeling problem to our attention and to Yves Deville for many interesting discussions. This research was partly supported by the Office of Naval Research under grant N00014-91-J-4052 ARPA order 8225, the National Science Foundation under grant numbers CCR-9357704, a NSF National Young Investigator Award. --R Introduction to Interval Computations. CLP(Intervals) Revisited. Applying Interval Arithmetic to Real Mean Value Forms in Interval Analysis. Logical Arithmetic. An Interval Newton Method. Bounding Solutions of Systems of Equations Using Interval Analysis. Interval Arithmetic in Matrix Computation: Part II. Safe Starting Regions by Fixed Points and Tightening. Preconditioners for the Interval Gauss-Seidel Method A Review of Preconditioners for the Interval Gauss-Seidel Method Consistency Techniques for Numerical Constraint Satisfaction Problems. Consistency in Networks of Relations. Networks of Constraints Interval Analysis. Methods and Applications of Interval Analysis. Safe Starting Regions for Iterative Methods. Interval Methods for Systems of Equations. Constraint Arithmetic on Real Intervals. New Computer Methods for Global Optimization. Experiments Using Interval Analysis for Solving a Circuit Design Problem. Solving Polynomial Systems Using a Branch and Prune Approach. Homotopies Exploiting Newton Polytopes For Solving Sparse Polynomial Systems. --TR --CTR Laurent Granvilliers , Frdric Benhamou, Progress in the Solving of a Circuit Design Problem, Journal of Global Optimization, v.20 n.2, p.155-168, June 2001 Yves Deville , Micha Janssen , Pascal Van Hentenryck, Consistency Techniques in Ordinary Differential Equations, Constraints, v.7 n.3-4, p.289-315, July-October 2002 Laurent Granvilliers , Frdric Benhamou, Algorithm 852: RealPaver: an interval solver using constraint satisfaction techniques, ACM Transactions on Mathematical Software (TOMS), v.32 n.1, p.138-156, March 2006 Yahia Lebbah , Claude Michel , Michel Rueher, A Rigorous Global Filtering Algorithm for Quadratic Constraints, Constraints, v.10 n.1, p.47-65, January 2005 Yahia Lebbah , Olivier Lhomme, Accelerating filtering techniques for numeric CSPs, Artificial Intelligence, v.139 n.1, p.109-132, July 2002 Stefan Ratschan, Search Heuristics for Box Decomposition Methods, Journal of Global Optimization, v.24 n.1, p.35-49, September 2002 Frdric Benhamou , Frdric Goualard , ric Langunou, , Marc Christie, Interval constraint solving for camera control and motion planning, ACM Transactions on Computational Logic (TOCL), v.5 n.4, p.732-767, October 2004 Laurent Granvilliers , Eric Monfroy , Frdric Benhamou, Symbolic-interval cooperation in constraint programming, Proceedings of the 2001 international symposium on Symbolic and algebraic computation, p.150-166, July 2001, London, Ontario, Canada
interval methods;branch and prune;constraint satisfaction;transistor modelling;electrical circuit;global zero search
596220
Comparing Nonsmooth Nonconvex Bundle Methods in Solving Hemivariational Inequalities.
Hemivariational inequalities can be considered as a generalization of variational inequalities. Their origin is in nonsmooth mechanics of solid, especially in nonmonotone contact problems. The solution of a hemivariational inequality proves to be a substationary point of some functional, and thus can be found by the nonsmooth and nonconvex optimization methods. We consider two type of bundle methods in order to solve hemivariational inequalities numerically: proximal bundle and bundle-Newton methods. Proximal bundle method is based on first order polyhedral approximation of the locally Lipschitz continuous objective function. To obtain better convergence rate bundle-Newton method contains also some second order information of the objective function in the form of approximate Hessian. Since the optimization problem arising in the hemivariational inequalities has a dominated quadratic part the second order method should be a good choice. The main question in the functioning of the methods is how remarkable is the advantage of the possible better convergence rate of bundle-Newton method when compared to the increased calculation demand.
Introduction Hemivariational inequalities introduced by Panagiotopoulos are generalizations of variational inequalities. By means of them problems involving nonmonotone and multivalued constitutive laws and boundary conditions can be defined mathematically. In many cases hemivariational inequalities can be reformulated as substationary point problems of the corresponding nonsmooth nonconvex energy functionals. For mathematical theory and the applications of hemivariational inequalities we refer to [23],[22]. The aim of this paper is to apply nonsmooth nonconvex optimization methods for the numerical solution of hemivariational inequalities and analyse their efficiency. As a typical example of hemivariational inequalities we consider a laminated composite structure under loading when the binding material between laminae obeys a nonmonotone multivalued law. This kind of a mechanical problem has been investigated numerically in [21],[26]. There it has been replaced by a sequence of convex subproblems which have been solved by using convex minimization methods. The discretization of the considered problem is realized by the finite element method scheme for nonmonotone multivalued differential inclusions presented in [17],[18],[19]. This scheme has been proved to be mathematically well-posed: stable and convergent. Bundle methods are at the moment the most promising methods for nonsmooth optimization. Their origin is the classical cutting plane method of [4] and [7] and they base on the piecewise linear approximation of the objective function. Due to the numerical experiments the proximal bundle methods (see [10], [25] and [16]) seem to work in the most efficient and reliable way. They can be called also to diagonal variable metric methods, since a stabilizing quadratic term in form of diagonal matrix was added to the polyhedral approximation in order to accumulate some second order information about the curvature of the objective function. However, the development of "real" second order method has been fascinating the researchers of nonsmooth optimization during its whole history. Several attempts has been done in order to exploit the second order subderivative information. Already in his pioneering work [11] Lemar'echal derived a version of variable metric bundle method utilizing the classical BFGS secant updating formula from smooth optimization. Due to the disappointing numerical results in [12] this idea was buried nearly for two decades. Several modifications of the variable metric concepts has been proposed for example in [6], [13], [3] and [20]. According to very limited numerical experiments (see for example [6]) it seems that the variable metric bundle methods works fairly well. However, when proportion the results to the extra computational efforts needed with the full matrix algebra they do not offer substantial advancement in numerical solution process. More recently a new second order approach has been proposed in [15], where the bundle-Newton method was introduced. The main difference compared to the earlier methods was include second order information directly to the model, in other words the piecewise linear model was replaced by piecewise quadratic model. In numerical tests of [15] it turns out to be very effective for quadratic type of problems. The aim of this paper is to compare the proximal bundle method and the bundle-Newton method in solving hemivariational inequalities. Since the optimization problem arising in the solution of hemivaria- tional inequalities has a dominated quadratic part the second order method should be a good choice. The main question in the functioning of the method is how remarkable is the advantage of the possible better convergence rate of the bundle-Newton method when compared to the increased calculation demand. The numerical results indicate the superiority of the bundle-Newton method for the problems having a quadratic nature. The paper is organized as follows. In chapter 2 we introduce the considered optimization problem and give some definitions needed in continuation. The chapters 3 and 4 are devoted to the proximal bundle method and bundle-Newton method, respectively. In chapter 5 the test hemivariational inequality problem is introduced and finally in chapter 6 we give some numerical results. 2. Nonsmooth and Nonconvex Optimization Problem We consider the following nonsmooth and nonconvex optimization problem ae minimize f(x) subject to x 2 K; where the objective function f from R n to R is supposed to be locally Lipschitz continuous function. The feasible set K has a more specific structure, i.e. where x l and x u are the lower and upper bounds for variables, respec- tively. We suppose that at each x 2 K we can evaluate the function value f(x) and an arbitrary subgradient g(x) from the subdifferential of Clarke (see [5]) The normal cone of K at x 2 K can be defined (see [16]) by cl ae [ oe R is the distance function of K, i.e. ck Definition 1. Suppose that f : R n ! R is locally Lipschitz continu- ous. Then x is called a substationary point of the problem (P) if Now we can formulate the following necessary optimality condition. Theorem 1. Every local minimizer of the problem (P) is substation- ary. For the proof we refer to [16]. In the following chapters we describe two methods: proximal bundle method and bundle-Newton method for finding local minimizers (and thus substationary) points of the problem (P). For the convergence of the methods we need the following semismoothness assumption due to [2]. Definition 2. The function f : R n ! R is said to be upper semis- mooth, if for any x 2 R n , d 2 R n and sequences g i ae R n and t i ae (0; 1) satisfying d) and t i # 0, one has lim sup 3. Proximal Bundle Method In this chapter we shortly describe the ideas of the proximal bundle method for nonsmooth and nonconvex minimization. For more details we refer to [10], [25] and [16]. 3.1. Direction finding Our aim is to produce a sequence fx k g 1 converging to some local minimum of the problem (P). Suppose that the starting point x 1 is feasible and at the k-th iteration of the algorithm we have the current iteration point x k and some trial points y j 2 R n (from past iterations) and subgradients where the index set J k is a nonempty subset of The idea behind the proximal bundle method is to approximate the objective function below by a piecewise linear function, in other words, we replace f by so called cutting-plane model which equivalently can be written in the form with the linearization error Note, that in convex case the subdifferential can be rewritten as (see Then it is easy to proof that In other words, if f is convex, then the cutting-plane model - f k is an under estimate for f and the nonnegative linearization error ff k how good an approximation the model is to the original problem. In nonconvex case these facts are not valid anymore: ff k may have a tiny (or even negative) value, although the trial point y j lies far away from the current iteration point x k and thus the corresponding subgradient j is useless. For these reasons the linearization error ff k j is replaced by so called subgradient locality measure (cf. [8]) where is the distance measure parameter and is the distance measure estimating without the need to store the trial points y j . Then obviously fi k for all j 2 J k and min x2K min In order to calculate the search direction d k 2 R n we replace the original problem (P) by the cutting plane model subject to x k (CP) where the regularizing quadratic penalty term 1u k d T d is added to guarantee the existence of the solution d k and keep the approximation local enough. The weighting parameter added to improve the convergence rate and to accumulate some second order information about the curvature of f around x k . It was adapted from the proximal point algorithm by [24] and [1] and was first time used in [10] and [25]. Notice, that the problem (CP) still is a nonsmooth optimization problem. However, due to piecewise linear nature it can be rewritten as a (smooth) quadratic programming subproblem finding the solution subject to \Gamma fi k and 3.2. Line search In previous section we calculated the search direction d k . Next we consider the problem of determining the step size into that direction. We assume that mL 2 (0; 1), mR 2 (mL ; 1) and - t 2 (0; 1] are fixed line search parameters. First we shall search for the largest number is the predicted amount of descent and it holds due to (11). If such a parameter exists we take a long serious step Otherwise, if (11) holds but then a short serious step R d k is taken and if t k we take a null step L is such that In long serious step there occurs a significant decrease in the value of the objective function. Thus there is no need for detecting discontinuities in the gradient of f , and so we set g k+1 2 @f(x k+1 ). In short serious steps and null steps there exists discontinuity in the gradient of f . Then the requirement (12) ensures that x k and y k+1 lie on the opposite sides of this discontinuity and the new subgradient g k+1 2 @f(y k+1 ) will force a remarkable modification of the next search direction finding problem. In what follows we are using the line search algorithm presented in [16]. The convergence proof of the algorithm is assumed f to be upper semismooth (see (2)). The iteration is terminated if final accuracy tolerance supplied by the user. 3.3. Weight updating One of the most important questions concerning proximal bundle method is the choice of the weight u k . The simplest strategy might be to keep it constant u k j u fix . This, however, leads to several difficulties. Due to Theorem 4 we observe the following: If u fix is very large, we shall have small jv k j and kd k k, almost all steps are serious and we have slow descent. If u fix is very small, we shall have large jv k j and kd k k, and each serious step will be followed by many null steps. Therefore, we keep the weight as a variable and update it when neces- sary. For updating u k we use the safeguarded quadratic interpolation algorithm due to [10]. 4. Bundle-Newton Method Next we describe the main ideas of the second order bundle-Newton method. For more details we refer to [15]. 4.1. Direction finding We suppose that at each x 2 K we can evaluate, in addition to the function value f(x) and an arbitrary subgradient g(x) 2 @f(x), also an n \Theta n symmetric matrix G(x) approximating the Hessian matrix f(x). For example, at the kink point y of piecewise twice differentiable function we can take "infinitely close" to y. Instead of piecewise linear cutting-pane model (3) we introduce a piecewise quadratic model of the form ~ damping parameter. The model (14) can again equivalently be written as ~ and for all j 2 J k the linearization error takes now the form Note that now even in the convex case ff k j might be negative. Therefore we replace the linearization error (16) again by the subgradient locality measure (7) and we remain the property (see [15]) min ~ The search direction d k 2 R n is now calculated as the solution of minimize ~ d) subject to x k (CN) Note, that since the model already has second order information no regularizing quadratic terms are needed like in (CP). The problem (CN) is transformed to a nonlinear programming problem, which is then solved by a recursive quadratic programming method (see [15]). If we denote this procedure leads to a quadratic programming subproblem finding the solution (d subject to \Gamma fi k and where and - are the Lagrange multipliers of (QN) from the previous iteration k \Gamma 1. In calculations W k is replaced by its positive definite modification, if necessary. 4.2. Line search The line search operation of the bundle-Newton method is following the same principles than in section 3.2 for proximal bundle method. The only remarkable difference occurs in the termination condition for short and null steps, in other words (12) is replaced by two conditions and where supplied by the user. The bundle-Newton method is using the line search algorithm presented in [15]. The convergence proof of the algorithm is assumed f to be upper semismooth (see (2)). 5. Formulation of the problem 5.1. Continuous problem We consider a two-dimensional laminated composite structure consisting of two elastic laminae and a interlayer binding material under loading (see Fig.1). The mechanical behaviour of the binding material, the interlaminar bonding forces versus the corresponding relative displacements of the laminae, is depicted by Fig.2. This relation is typically nonmonotone and multivalued. Therefore, this kind of mechanical problems lead to the hemivariational inequalities or in the energy formulation to the substationary point problems of nonconvex, nonsmooth functionals. The similar mechanical structures have been investigated in [21],[26]. Due to the symmetry of the structure and by assuming that the forces applied to the upper and lower part of the structure are the same it is enough to study the upper lamina. Next, we formulate mathematically the problem. We denote by\Omega ae R 2 the upper lamina in its undeformed state. The Lipschitz boundary \Gamma of\Omega consists of four nonoverlapping open subsets We denote by i;j=1 the stress tensor, i;j=1 the strain tensor, F F G G G G24 Figure 1. Laminated composite structure under loading Relative normal displacement of laminae in mm Interlaminar binding force in kN/m F G2000040000 A Figure 2. Nonmonotone adhesive force between laminae the displacement, the outward unit normal vector to \Gamma and the boundary force. Assuming that the deformations are small the lamina obeys the Hooke's law of form @x l and i;j;k;l=1 is the elasticity tensor satisfying the usual symmetry and ellipticity conditions. The equation of the equilibrium state of\Omega is as follows: since there are no volume forces. Next, we define the boundary conditions. On \Gamma 1 we have given displacement and on both \Gamma 2 and \Gamma 3 given boundary forces where F is a constant force. Furthermore, on \Gamma 4 we have in the tangential direction a given boundary force On the other hand, due to the binding interlayer material it holds a nonmonotone, multivalued boundary condition (see Fig.2) expressed by means of the subdifferential @j of a locally Lipschitz function j and because of the nonpenetration of the laminae a unilateral boundary condition in the normal direction. Note that we have to scale by 1the x 1 -axis of the graph in Fig.2 in order to get the nonmonotone law of (26), because in Fig.2 we have relative displacements of the laminae. By employing Green-Gauss theorem and taking into account both the equilibrium equation (21) and the boundary conditions (22)-(27) we obtain the following hemivariational inequality: Find u 2 K such that Z Z where K is a set of kinematical admissible displacements defined by G G F GG Figure 3. 32\Theta4 triangulation Th of\Omega (\Omega\Gamma denotes the Sobolev space), j ffi is the generalized directional derivative of j and a the bilinear form of the linear elasticity defined by Z Due to the symmetry of the bilinear form a the potential energy of the considered mechanical system has the form: Z Z Then, we can formulate the following problem: Find u 2 K such that u is a substationary point (cf. (1)) of f on K, i.e. where NK (u) is the normal cone to the nonempty, closed and convex set K at u. It can be shown that the problem (28) has at least one solution (see [22]), and that the solutions of the inclusion (33) are also the solutions of the inequality (28) (see [19]). 5.2. Discrete problem We apply the finite element approximation scheme developed in [17],[18]. Let h be a discretization parameter related to the mesh size of the triangulation of\Omega (see Fig.3). The set K is approximated by the set of piecewise linear functions over the triangulation T h defined by For the approximation of the bilinear form a and the boundary inte- gral R F (v 2 \Gammau 2 )d\Gamma we use appropriate numerical integration formulae (this is standard in the finite element method). The nonmonotone term in (28) is approximated by the following numerical integration formula Z i2I where fx i g i2I is the set of the nodal points of the triangulation T h on are the coefficients of the integration formula. Let us make the identification X h with the corresponding subset of R 2n , i.e. the displacement function v is identified with the displacement vector of the nodal points (n is the number of the nodal points). Then the discrete problem in the matrix form is as follows: Find u 2 K h such that i2I where A is the stiffness matrix of the structure, F the (discrete) load vector and I the set of the indices corresponding to the x 2 -displacements of the nodal points on \Gamma 4 . Then we can formulate the discrete substationary point problem on where i2I Because of Corallary 1 of Proposition 2.3.3 in [5] and the terms j(u i ), are independent of each others we have that i2I Now it holds that every solution of (38) is a solution of the problem (37) (see [19]). Also due to the results in [18],[19] we know that the discrete problem (38) is solvable and its solutions converge in subsequences to the solutions of the continuous problem (33). Remark 1. The functional f h is upper semismooth. Indeed, let u; d 2 R 2n be given. ?From Definition 2 of the upper semismoothness we see that it is enough to study the restriction of f h onto the line Rg, on which f h is smooth except finite number of points (cf. and Fig.2). If f h j L is smooth at u, it is smooth also on some small (one-dimensional) neighborhood of u and the condition lim sup is trivially satisfied. On the other hand, if f h j L is nonsmooth in u, there exists small neighborhood of u in which u is the only nonsmooth point. Therefore, f h j L is continuously differentiable in this set except at the point u and the classical one-sided derivatives exist at u implying (41). Since the nonlinear behaviour and the constraints of the problem have an effect only on the x 2 -displacements on the nodes on \Gamma 4 , we can make our computation much more effective by applying the method of condensation of unknowns. Assuming that m components corresponding to the index set I are listed first, we have the following decomposition of the matrix A and the vectors F and u A 21 A 22 where A 11 is an m \Theta m matrix and F 1 ; . Then the elimination of u 2 (2 R 2n\Gammam ) from (37) leads to the following discrete potential functional ~ where ~ called the Schur complement and ~ . Hence, the eliminated substationary point problem is formulated as follows: Find u 1 2 K h such that where K h is interpreted as a set of R m . 6. Numerical Results The optimization algorithms has been implemented in Fortran 77 and the test runs have been performed on an HP9000/J280 (180 MHz) computer. The tested optimization codes are presented in Table I. All the codes are utilizing the subgradient aggregation strategy of [8] to keep the storage requirements bounded. The code PB is employing the quadratic solver QPDF4, which is based on the dual active Table I. Tested codes and software. Code Software Package Author(s) Method PB NSOLIB M-akel-a Proximal bundle PBL UFO Luk-san &Vl-cek Proximal bundle BNL UFO Luk-san &Vl-cek Bundle-Newton Table II. Eliminated with discretization 16 \Theta 2. Code Ni Nf CPU f Ni Nf CPU f set method derived in [9], while the codes PBL and BNL are using the solver ULQDF1 implementing the dual projected gradient method proposed in [14]. The above optimization codes have been applied to the mechanical structure of Fig.1 under eleven constant loadings (F=20000, 21000, In calculation we have used the incremental procedure: the loading of the structure is increased uniformly and as an initial guess for the solution of the next load it is used the solution of the previous load. The bigger increment from 25000 kN/m 2 to 26200 kN/m 2 is due to the fact that the structure is very sensitive to the increase of the loading between 25000-27000 kN/m 2 where the partial (branches B-G in Fig.2) and the complete delamination (branch G-H in Fig.2) take place. With the load 26200 kN/m 2 we can illustrate the case in which only the partial delamination occurs. We have applied the plane stress model Table III. Eliminated with discretization 4. Code Ni Nf CPU f Ni Nf CPU f PBL 426 434 3.15 -0.871201 3007 3063 14.16 -15.21929 Table IV. Eliminated with discretization 64 \Theta 8. Code Ni Nf CPU f Ni Nf CPU f Table V. Eliminated with discretization 128 \Theta 16. Code Ni Nf CPU f Ni Nf CPU f with the elasticity modulus the Poissons ratio 0:3 and the thickness The calculated results with different discretizations are presented in Tables II-VI, in which Ni denotes the number of iterations, Nf denotes the number of objective function (and also subgradient) evaluations, CPU the used total computer time in seconds and f is the objective function value at termination. Note that the elimination, the computation of the Schur complement ~ A and the new load vector ~ F , for the discretization 4, 64 \Theta 8 and 128 \Theta 16 is taking about 0.67, 1.35, 23.1 and 545.0 seconds, respectively. The stopping criterions are chosen to be comparable, in other words the relative accuracy in the optimal objective function value is about 6 digits. Table VI. Not eliminated with discretization 4. Code Ni Nf CPU f Ni Nf CPU f PB 4458 4474 156.97 -0.880013 40112 40128 1458.51 -15.22886 Table VII. Average error of the normal displacements on \Gamma 4 by BNL. Discretization Error in mm Error in mm 128 \Theta 16 * From the numerical results in Tables II-VI we can conclude the superiority of the bundle-Newton code BNL. In all cases it used less computing resources and found the local minimum in the most reliable way. Due to the matrix operations the individual iteration is more costly than in proximal bundle methods. However, the total amount of iterations stays very low, and thus the used CPU time of BNL is always less than with the other codes. Note, that the iteration number and the function evaluations of BNL do not depend on the dimension of the problem. On the other hand, when the size of problem is doubled, BPL needed about two times and PB about 1.6 times more iterations and function evaluations. When comparing the CPU times versus discretization the behaviour is nonlinear and the multiplier is growing exponentially. We have to remember, that especially in the large problems the elimination is taking most of the time and the differences do not seem to be so remark- able. For example for the discretization 128 \Theta 16 the optimization time of BNL is only about 2 % of the total time. By comparing the Tables III and VI we can see the influence of the elimination; the solution of the problem without elimination is taking nearly ten times more CPU time than the eliminated one. Note that especially for the proximal bundle methods the elimination is essential. Although the codes PB and PBL are realizations of the same method, the difference in their functioning is remarkable. PBL is clearly more effective when comparing the function evaluations and CPU times. However, it has slight difficulties to reach the desired accuracy in larger problems (see the function values denoted by in Tables V-VI). Notice also, that PB converges to the different local minimum with the load (see the function values denoted by in Tables III-V). When we changed the starting point to be zero in all the components it found the same optimum as the other codes. This behaviour is due the fact that the delamination takes place between the loads 25000-27000 Normal displacements in mm Interface in mm load 23000 kN/m^2 load 26200 kN/m^2 load 27000 kN/m^2 load 30000 kN/m^2 Figure 4. Normal displacements of the interface of the upper lamina making the structure very unstable and sensitive to the increase of the load. In Table VII we have listed the average errors of the normal displacements with the loads 26200 kN/m 2 and 27000 kN/m 2 obtained by BNL. As an exact solution it is used the solution obtained by the discretization 128 \Theta 16. Table VII indicates that the applied approximation scheme has a very good convergence rate and the results calculated by BNL are reliable. In Fig.4 and Fig.5 it is presented the normal displacements of the upper lamina (the relative displacements of laminae are twice larger) and the binding forces between laminae. Further, Fig.6 illustrates the progress of delamination: With the load 23000 kN/m 2 there is no damage of the binding material, all the nodes are on the branch A-B. The partial delamination occurs when the load is increased to 26200 kN/m 2 , some of the nodes are on the branch C-D, and the complete delamination when the load is greater than 27000 kN/m 2 , most of the nodes are on the branch G-H. 7. Conclusions The solution of a hemivariational inequality can be found as a sub- stationary point of some nonconvex functional being composed of a Interlaminar binding force in Interface in mm load 23000 kN/m^2 load 26200 kN/m^2 load 27000 kN/m^2 load 30000 kN/m^2 Figure 5. Nonmonotone adhesive force between laminae branch G-H F=23000 kN/m F=27000 kN/m F=30000 kN/m22 branch A-B branch C-D branch E-F Figure 6. Progress of delamination dominating quadratic part and a nonsmooth piecewise quadratic part. We have tested the functioning of different nonconvex and nonsmooth optimization methods in the solution of a laminated composite structure under loading when the binding material between laminae obeys a nonmonotone multivalued law. Due to the strong quadratic nature of this kind of problem, the bundle-Newton method based on the second order piecewise quadratic model proves to be superior when compared to proximal bundle method based on the first order polyhedral approximation. Bundle- Newton method was clearly faster and more reliable, and what is best, the iteration number and the function evaluations of bundle-Newton method do not depend on the dimension of the problem. The same trend can be seen also in the academic tests of [15]. Acknowledgements This work was supported by the grants SA-34063 and SA-8583 given by the Academy of Finland and the grant No. 201/96/0918 given by the Czech Republic Grant Agency. The research was partially done during the third author's visit at the Laboratory of Scientific Computing --R Numerical Methods for Nondifferentiable Convex Opti- mization Optimization of Upper Semidifferentiable Functions Gilbert J. Golstein A. Optimization and Nonsmooth Analysis Variants to the Cutting Plane Approach for Convex Nondifferentiable Optimization The Cutting Plane Method for Solving Convex Programs Methods of Descent for Nondifferentiable Optimization A Method for Solving Certain Quadratic Programming Problems Arising in Nonsmooth Optimization Proximity Control in Bundle Methods for Convex Non-differentiable Optimization Nonsmooth Optimization and Descent Methods Numerical Experiments in Nonsmooth Optimization An Approach to Variable Metric Bundle Methods Dual Method for Solving a Special Problem of Quadratic Programming as a Subproblem at Linearly Constrained Nonlinear Minmax Approximation A Bundle-Newton Method for Nonsmooth Unconstrained Minimization Approximation of Non-Monotone Multivalued Differential Inclusions Finite Element Approximation of Vector-Valued Hemivariational Inequalities A Quasi-second-order Proximal Bundle Algorithm Baniotopoulos C. Mathematical Theory of Hemi- variational Inequalities and Applications Hemivariational Inequalities Monotone Operators and the Proximal Point Algo- rithm A Version of the Bundle Idea for Minimizing a Nonsmooth Functions: Conceptual Idea --TR
bundle methods;hemivariational inequalities;nondifferentiable programming;nonmonotone contact problems;substationary points
596236
Global Optimization by Multilevel Coordinate Search.
Inspired by a method by Jones et al. (1993), we present a global optimization algorithm based on multilevel coordinate search. It is guaranteed to converge if the function is continuous in the neighborhood of a global minimizer. By starting a local search from certain good points, an improved convergence result is obtained. We discuss implementation details and give some numerical results.
Introduction Problems involving global optimization (traditionally usually minimization) of a multivariate function are widespread in the mathematical modeling of real world systems for a broad range of applications (see, e.g., Pint' er [18]). Many problems can be described only by nonlinear relationships, which introduces the possibility of multiple local minima. The task of global optimization is to find a solution for with the objective function obtains its smallest value, the global minimum. When the objective function has a huge number of local minima, local optimization techniques are likely to get stuck before the global minimum is reached, and some kind of global search is needed to find the global minimum with some reliability. Our Global Optimization Homepage on the World Wide Web (WWW) at the address http://solon.cma.univie.ac.at/ ~ neum/glopt.html contains a large number of commented links to information and software packages relevant to global optimization. Algorithms for solving global minimization problems can be classified into heuristic methods that find the global minimum only with high probability, and methods that guarantee to find a global optimum with a required accuracy. An important class belonging to the former type are the stochastic methods (e.g., Boender & Romeijn [3]), which involve function evaluations at a suitably chosen random sample of points and subsequent manipulation of the sample to find good local (and hopefully global) minima. A number of techniques like simulated annealing (e.g., Ingber [9, 10]) and genetic algorithms (e.g., Michalewicz [13]) use analogies to physics and biology to approach the global optimum. The most important class of methods of the second type are branch and bound meth- ods. They derive their origin from combinatorial optimization (e.g., Nemhauser & Wolsey [14]), where also global optima are wanted but the variables are discrete and take a few values only. Branch and bound methods guarantee to find a global minimizer with a desired accuracy after a predictable (though often exponential) number of steps. The basic idea is that the configuration space is split recursively by branching into smaller and smaller parts. This is not done uniformly but instead some parts are preferred and others are eliminated. The details depend on bounding procedures. Lower bounds on the objective allow to eliminate large portions of the configuration space early in the computation so that only a (usually small) part of the branching tree has to be generated and processed. The lower bounds may be obtained by using dc-methods (e.g., Horst & Tuy [8]), techniques of interval analysis (e.g., Hansen or majorization resp. minorization methods based on the knowledge of Lipschitz constants (e.g., Pint' er [18]). Unlike heuristic methods, however, these methods are only applicable if something about the analytical properties of the objective is known, since one needs to be able to compute powerful and reliable underestimating functions. The algorithm we are going to describe in this paper is an intermediate between purely heuristic methods and methods that allow an assessment of the quality of the minimum obtained; it is in spirit similar to the direct method for global optimization by Jones et al. [11]. As the latter method, our method is guaranteed to converge in the long run if the objective is continuous in the neighborhood of a global minimizer; no additional smoothness properties are required. In contrast to many stochastic methods that operate only at the global level and are therefore quite slow, our algorithm contains local enhancements that lead to quick convergence once the global part of the algorithm has found a point in the basin of convergence of the global optimum. Moreover, for all control variables in our algorithm meaningful default values can be chosen that work simultaneously for most problems. Thus fine tuning is usually not required, and since it is a deterministic method, there is no need for multiple runs. In this paper, we consider the bound constrained optimization problem min f(x) (1) with finite or infinite bounds, where we use interval notation for rectangular boxes, ng for . In the case where all bounds are infinite we obtain an unconstrained optimization problem. In direct, a finite box is normalized to [0; 1] n and a tree of boxes is constructed. Each box is characterized by its midpoint, and the side lengths of the boxes are always of the form 3 \Gammak , k 2 N 0 . A disadvantage of that algorithm is that infinite box bounds cannot be handled. Moreover, since the boundary can never be reached, it converges more slowly than necessary in cases where the minimizer lies at the boundary. For example, in the case of functions that are monotonous in each variable the optimizer is at a vertex, but direct converges very slowly, especially when the bounds are wide. Inspired by direct, we devised a global optimization algorithm based on multilevel coordinate search (MCS). In our algorithm we remedy the above shortcomings and allow for a more irregular splitting procedure. In Section 2, we first define a class of algorithms for solving the global optimization problem (1) satisfying six properties A1-A6 and prove their convergence in the long run. In addition to the splitting procedure, an enhancement is obtained by starting local searches from certain good points, which results in in an improved convergence result. In Section 3, an outline of our implementation of the MCS algorithm is given, and details are explained in Sections 4 to 6. Finally, numerical results are presented in Section 7. Acknowledgments The authors gratefully acknowledge partial support of this research by the Austrian Fond zur F-orderung der wissenschaftlichen Forschung (FWF) under grant P11516- MAT. 2. Convergence results In this section we consider a class of algorithms for solving the minimization problem (1) with six properties A1-A6 defined as follows. Each box in the tree is characterized by a base point x and its function value f(x). Moreover, we assign to each box a level s 2 g. When a box is split, its level is set to zero; the levels of non-split boxes are positive. The level of the initial box is 1. The base points of boxes with level s max and their function values are put into the base list. The levels of MCS correspond to the side lengths of direct, i.e., the boxes with small level are the 'large' boxes that have not been split very often yet. Only boxes of the same level compete for being chosen for branching; cf. A3. A2: A box is always split along a single coordinate i (to be chosen appropriately), the splitting index, into two or more (at most p - 2) parts, and the descendants of a box with level latter only when it is - s The base points of the descendants of a box are chosen in a natural way such that they differ from the base point of the parent box in (at most) one coordinate and thus this procedure of generating new function values is a variant of the standard coordinate search method. In contrast to our algorithm, direct splits a box along all coordinates in each step. A3: The branching process proceeds by a series of sweeps through the levels defining a record. A sweep is defined by the following three steps. Step 1. Scan the list of non-split boxes and define a record list containing for each level pointing to a box with the lowest function value among all boxes at level s. If there is no box at level s, set b Initialize s to the lowest level with b s 6= 0. Step 2. The box with label b s is a candidate for splitting. If it is split, mark it as split in the list and insert its children. Update the record list if any of the children yields a strict improvement of f on its level. If it is not considered to be worthwhile to split the box, its level is increased by one and possibly b s+1 has to be updated. Step 3. Increase s by 1. If new sweep. Else if b to Step 3, else go to Step 2. Clearly, each sweeps ends after at most s visits to Step 3. Like in direct, this strategy yields a combination of global and local search. The key to balancing global and local search is the multilevel approach. The fact that we start with the boxes at the lowest levels (i.e., the 'large' boxes) in each sweep constitutes the global part, and the selection of the box with lowest function value at each level forms the local part of the algorithm. A4: For each M ? 0 there exists an r(M) 2 (0; 1) independent of the choice of s max such that for each box [x that has been generated by splitting the box [x; y] with along the ith coordinate we have This assumption ensures that the descendants of a finite box will eventually become arbitrarily narrow after sufficiently many splits along each coordinate. The worst case r(M) of the shrinking factor may depend on the box bounds since we do not want to overemphasize large numbers in the case of wide box bounds. A5: The rule for splitting boxes with infinite bounds is as follows. If the ith box bounds are (\Gamma1; 1) and we split along the ith coordinate, the interval is always split at one or more values contained in a fixed interval [z 1 (independent of the choice of means of which it turns into two semibounded intervals contained in (\Gamma1; z 2 possibly additional finite ones contained in [z 1 When a box with the ith bounds [x along the ith coordinate, we obtain a box with ith bounds one or more boxes with finite ith bounds. Generally let [x be the unbounded part generated by splitting the ith bounds [x the ith coordinate. There exist functions monotonously increasing on (0; 1), such that F 2 (which implies F m all x 0 2 R . An analogous assumption holds for successive splits of (\Gamma1; x 0 ]. The lower bound F 1 guarantees that the unbounded descendants of the interval [x shrink sufficiently fast. However, they should not shrink too fast since then the bounded descendants could become arbitrarily large, and therefore we impose the upper bound Assume that a box with level s, 1 - s ! s max , is a candidate for splitting since b s points to this box, and let n j be the number of times coordinate j has been split in the history of this box. If s ? H(n; minn j ), the box is in any case eligible for splitting, and the splitting index is a coordinate i with This assumption makes sure that along each path in the tree, each coordinate is split inifinitely often. Proposition 1. In each sweep, one box leaves the lowest non-empty level and no box is added at that level. Each level will eventually become empty; in particular the splitting procedure will come to an end when all non-split boxes have level s max . Proof. Directly from A1-A3. Proposition 2. Let 2 - s 0 - s max . Then the levels up to s are empty after at most (p s defined by A2. In particular, the algorithm finishes after at most (p smax Proof. In the course of the algorithm, the root box [u; v] with level one can at most have descendants of level s, 2 - s - s max \Gamma1, since the worst case is splitting a box into p parts and advancing the levels of the descendants in steps of one. By Proposition 1, the levels 1 to s are cleared after at most As expected, this upper bound is exponential. However, in practice, the bound is very pessimistic since we do not only split a box at the lowest non-empty level (or advance its level by one) in each sweep but do that at each non-empty level. Moreover, the levels of the descendants of the boxes will sometimes advance by two, or a box will be advanced to the next level without being split. Also, good approximations to the global minimizer will usually have been reached long before all levels are cleared. Proposition 3. For a box with level 0-l-m each coordinate has been split at least m times. Proof. This is proved by induction over m. The statement is trivial for be a box with level s ? i be the number of times coordinate i has been split in the history of B 0 . Then by induction hypothesis. Assume that minn be the parent box of B its level. Since the levels advance in steps of one or two and s ? 2(m 1)n, it is possible to go back n times in the history of box B 0 , and we have s that was obtained by splitting B i according to A6. Let n 0 i be the number of times coordinate i has been split in the history of box B n . Since B 0 was generated from B n by splitting n times according to A6, we have minn 0 contradiction to the induction hypothesis. Consequently our assumption was wrong and we must have Proposition 4. For each exists an m i (ffi) such that the ith side length of the box containing a point - x 2 [u; v] with finite coordinates is less than ffi if it has been split at least m i (ffi) times along the ith coordinate. Proof. Consider the ith side length of the box containing - x. Case 1. u i is finite, implies that there exists an l 2 N such that F l after at most l splits in the ith coordinate the box containing - x has finite ith bounds contained in [u and the problem is reduced to Case 2. The case treated similarly. The case reduced to a semibounded interval contained in [z 1 or a finite interval contained in [z 1 after one split. Case 2. u i and v i finite, M := max(ju i j; jv i j). Then there exists a k 2 N 0 such that defined in A4, i.e., by A4 the ith side length of the containing - x is less than ffi if the box has been split at least k times along the ith coordinate. Propositions 2-4 yield the following convergence theorem. Theorem 5. Suppose that the global minimization problem (1) has a finite solution - R is continuous in a neighborhood of - x. Let the algorithm satisfy the assumptions A1-A6, and let " ? 0. Then there exists an s 0 such that for s after at most (p s a base point x is found with i.e., the algorithm converges if the number of levels tends to 1. In order to enhance the performance of the algorithm, we subject a base point of a box of level s max to local search before putting it into the base list, which is another local feature of our algorithm. For theoretical purposes we make here the obviously idealized assumption that the local search algorithm reaches a local minimizer after finitely many steps if it is started in its basin of attraction. Theorem 6. Let the assumptions of Theorem 5 be satisfied, assume that a local search is started from each candidate for the base list, at least when its function value is smaller than the function value of any local minimum already in the base list, and assume that there is an " ? 0 such that f(y) ? nonglobal local minimizer y and for any y 2 [u; v] with sufficiently large norm. Then there exist numbers L and S such that, for any s max - L, a global minimizer is found after at most S sweeps. Proof. By assumption there exists a with Propositions 3 and 4 there exists an L such that the side lengths of the box containing - x are smaller than ffi when its level is at least L, and by Proposition 2 the box containing - x reaches level L after at most S := (p sweeps. Let s max - L. Then after at most S sweeps the box B containing - x is at level L, and for its base point x we have f(x) - f(-x) Two cases are possible. If we have found a better point before, we have already made a local search from such a point, which finds a global minimizer by assumption. In the case that x is the current best point, at the end of that sweep a point y with f(y) - f(x) is subjected to local search, by means of which we find a global minimizer. 3. The MCS algorithm There are many ways to design algorithms satisfying A1-A6 and hence guaranteeing convergence to a global minimizer. However, trivial implementations are very slow. A number of heuristic enhancements are needed to obtain a high quality method. In the following we describe the most useful version that we could find. We first give an overview over the techniques used and discuss the details of each technique in Sections 4 to 6. Unlike in direct, the base point of a box in our algorithm is usually not the midpoint but a point at the boundary, often but not always a vertex, and a base point can belong to more than one box. Moreover, we also assign to each box an opposite point. The construction is such that the base point x and the opposite point y determine the box, and we call such a box B[x; y]. The algorithm starts with a so-called initialization procedure producing an initial tree of boxes. Whenever a box is split along some coordinate i for the first time (either in the initialization procedure or later), this is done at three or more given values x l (where function values are computed) and some adaptively chosen intermediate points, and at least four subboxes are obtained. After the initialization procedure we start with the sweeps as described in A3. When a box with level 1 - s ! s max has been selected for splitting by the procedure described in A3, we use the following splitting strategy: the box is ffl either split according to A6 (splitting by rank), ffl or it is split along a coordinate where the maximal gain in function value is expected according to a local separable quadratic model obtained by fitting 2n+1 function values (splitting by expected gain), ffl or, if the expected gain is not large enough, it is not split at all but its level is increased by one. If the box has already been split along the coordinate chosen as splitting index, only a single new function evaluation is needed, and we obtain two or three subboxes. The new function value is taken at a point that is predetermined by the box bounds in the case of splitting by rank, and in the case of splitting by expected gain, at a point where the maximal gain in function value is expected. Before putting a base point of a box of level s max into the base list, we first check whether it is in the the basin of attraction of a local minimum already in the base list, and if this is not the case, we subject it to local search. The local search algorithm used in our program essentially consists of building a local quadratic model by triple searches, defining a promising search direction with the aid of the quadratic model and a modified Cholesky factorization and making a line search along this direction. This procedure is repeated until a stopping criterion is fulfilled. Moreover, devices to handle the case that one or more coordinates of the current point are at the boundary are included. 4. Initialization Before building the tree of boxes, one first evaluates f at an initial point x 0 and sets evaluated at two or more points in [u; v] that agree with x in all coordinates k 6= i. Thus one has x l f l and x . The point with smallest function value is then renamed x before repeating the procedure with the next coordinate. The numbers x l and the indices l i are stored in an initialization list. The choice of the initialization list is left to the user, who may incorporate in it knowledge about the likely distribution of good points. A good known starting point can be defined as x 0 . Some possible choices are discussed in Section 7. From the initialization list and the corresponding list of function values, an initial tree of boxes is constructed as follows. The root box is B[x; point and as y one of the corners of [u; v] farthest away from x. Note that x need not be a vertex and that some or all coordinates of y can be infinite. For current box is split along the ith coordinate into 2L exactly one of the x l i as endpoints, depending on whether two, one or none of the x l are on the boundary, which means that in addition to x l we have to split at z l . The additional splitting points are chosen as z l is the golden section ratio and chosen such that the part with the smaller function value gets the larger fraction of the interval. The resulting subboxes get as base point the point x 0 obtained from the current by changing x i to the x l i that is a boundary point of the corresponding ith coordinate interval, so that f(x 0 i , and as opposite point the point obtained from y by changing y i to the other end of that interval. The information available so far allows us to define priorities on the coordinates. For each i, we compute the union of the ranges of the quadratic interpolant through any three consecutive take the difference of the upper and lower bound obtained as a crude measure of the variability of f with the ith component. Components with higher variability get a higher priority, and this ranking is saved in a vector - such that the component with index i has the - i th highest estimated variability. Moreover, if the x obtained after splitting the ith coordinate belongs to two boxes, the one containing the minimizer of the quadratic models is taken as current box for coordinate i + 1. The root box gets level 1. If a box of level s is split, the boxes with the smaller fraction of the golden section split get level s Thus the current box for splitting in the next coordinate is in any case one with level after finishing the initialization procedure, the first level is empty and the non-split boxes have levels which implies that it is meaningful to take Two examples for the set of boxes, their base points and their levels after the initialization procedure in the two-dimensional case are shown in Figure 1 (where 6 in both cases x It is easy to see the connection between the golden section split and the assignment of levels. When a box B with level s is split along coordinate i according to the golden section split and its larger part B 0 is again subjected to the golden section split along the ith coordinate, the larger descendant of B 0 has the same ith coordinate length as the smaller descendant of B, and these boxes both have level s 2. Moreover, the box with the better function value gets the larger fraction of the interval and the smaller r r Figure r r Figure level because then it is more likely to be split again more quickly, which was also the strategy adopted in direct. Any choice of x l including the endpoints u i , v i in the list guarantees that, in the simple case that f is monotonous in each variable, the final x of the initialization phase is already the global minimizer. 5. Splitting We do not store the box bounds for each box but the information in the tree (label of the parent box, splitting index, splitting value, a label identifying which of the many children it is etc. This keeps the amount of storage proportional to the number of function evaluations and allows us to recover information to build a separable quadratic model by going back in the history of the box. Suppose that we want to split the ith coordinate interval u t fx for is the ith component of the base point of the box being considered. In order to fulfil assuption A4, we may not split too close to x i . If y i is large, we also do not want the new component x 0 i to be too large and therefore force it to be from some smaller interval u t f- g. We choose this interval according to where the function subint is given by the following pseudo-Matlab function: function else Having selected a box with level s ! s max for splitting (by the procedure described in A3), we recover its base point x, the opposite point y and the number n i of times coordinate i has been split in the history of the box. Splitting by rank. Let s ? H(n; minn i ), where H is the function defined by A6. Then we select the splitting index i among the indices i with smallest n i as the one with lowest - i (and hence highest variability rank). the splitting is done according to the initialization list at x l and at the golden section split points, as discussed in Section 4, and the new base points and opposite points are defined as before. The boxes with the smaller fraction of the golden section split (and thus larger function values) get level min(s and all other ones get level s + 1. the ith component ranges between x i and y i , and the splitting value is chosen as z The box is split at z i and at the golden section split point, and we obtain three parts with only one additional function evaluation at the point x 0 obtained by changing the ith coordinate of x to z i . The smaller fraction of the golden section split gets level min(s and the two other parts get level s + 1. Moreover, the base point of the first child is taken to be x, the base point of the second and third child is the point x 0 defined above, and the opposite points are obtained by changing y i to the other end of the ith coordinate interval of the corresponding box. Splitting by expected gain. Let s - H(n; minn i ). In order to build a local separable quadratic model, we need two additional points and corresponding function values for each coordinate. Whenever we have split in the ith coordinate in the history of the box, we obtain values that can be used for quadratic interpolation in this coordinate. For each coordinate we take the first two points and function values found by pursuing the history of the box back to [u; v] since these points are expected to be closest to the base point x. For coordinates that have not yet been split, we obtain this information from the initialization list. Let be the local separable model for f(-) generated by interpolation at x and the 2n additional points collected as above. For each coordinate i, we define the expected gain - e i in function value when we evaluate at a new point obtained by changing this coordinate in the base point. Again two cases have to be distinguished. Case 1. In the history of the current box, coordinate i was never split, i.e., Then we split according to the initialization list at points where we already know the obtainable function differences, and therefore compute the expected gain as Case 2. If n i ? 0, the ith component ranges between x i and y i , and with the quadratic partial correction function at our disposal, we can calculate the maximal gain expected when changing the value of only. For the reasons discussed above, we choose the splitting value from u t f- where Then we compute with minimum achieved at - . If the expected best function value satisfies 1-i-n best ; where f best is the current best function value (including the function values obtained by local optimization), we expect the box to contain a better point and split, using as splitting index the component with minimal - e i . The condition (2) prevents wasting function evaluations by splitting boxes with bad base point function values; these boxes will eventually be split by rank anyway. In Case 1 we again split according to the initialization list, and the definition of the new base points and opposite points and the assignment of levels are as before. In Case 2 we use z i as splitting value and the box is split at z i (if z i 6= y i ) and at the golden section split point, and we obtain two or three parts. The larger fraction of the golden section split gets level s + 1, the smaller fraction level min(s the third part is larger than the smaller fraction of the golden section split, it gets level Moreover, the base point of the first child is taken to be x, the base point of the second and third (if z i 6= y i ) child is obtained by changing the ith coordinate of x to z i , and the opposite points are again obtained by changing y i to the other end of the ith coordinate interval of the box. If (2) is violated, we do not expect any improvement and therefore do not split but increase the level by 1. 6. Local search The theory of local optimization provides powerful tools for the task of optimizing a smooth function when knowledge of the gradient or even the Hessian is assumed. When no derivative information is available, the traditional methods are based on the employment of conjugate directions and successive line searches; cf. the direction set method of Powell [19] and a modification due to Brent [4]. These algorithms, however, do not allow the specification of bound constraints. Elster & Neumaier [6] developed an algorithm for optimization of low-dimensional bound constrained functions, based on the use of quadratic models and a restriction of the evaluation points to successively refined grids. Hoewever, the work in that algorithm grows with the dimension n as O(n 6 ) and hence is unsuitable for larger dimensions. The local optimization algorithm we are going to describe in the sequel also makes use of quadratic models and successive line searches, and devices to handle bound constraints are incorporated. We first explain the procedure of building a local quadratic model by triple searches. Since we do not want to start local searches from points belonging to the domain of attraction of the same local minimum and to put several copies of essentially the same point into the base list, we have devised a criterion to avoid that; see Step 2 in the subsection 'base list' below. We conclude with a discussion of the validity of assumptions A4 and A5 for our algorithm. Triple search. We want to use function values to construct a quadratic model best best best Assume that we have three vectors x l (with componentwise inequalities). The function values are to be taken at points x with x as follows. If x best denotes current best point in the triple search, denote by x (i;1) and x (i;2) the points obtained from x best by changing its ith coordinate to the other two values in fx l and by x ik , k ! i, the points obtained by changing the ith and kth coordinate to the ones with the smaller q(x (i) ) resp. q(x (k) ) with the current quadratic model q. Thus we obtain the following procedure that we are going to describe in more detail in the sequel: best ) compute f(x (i;1) ) and f(x (i;2) ); compute g i and G ii store x newbest but do not update x best compute q(x (k;1) ) and q(x (k;2) ) from the current model compute f(x ik ) update x newbest but do not update x best compute G ik if x newbest 6= x best , update x best , f and g 1:i When, for an computed approximations for g l and G lk , by interpolating at points that differ only in the first best is the current best point in the triple search, we obtain approximations for g i and G ii by determining these numbers such that the quadratic polynomial best best best interpolates at (x (i;j) 2. At this stage we do not yet update x best , but if min(f(x (i;1) ); f(x (i;2) best ), we store this point as x bestnew . Assume that, in addition, we have already calculated approximations for G il , 1 - l - 1. Then we can compute best best best with the current quadratic model q, and we have q(x (i;j) 2. Let x ik be defined as above. Then we choose G ki such that the quadratic model interpolates at x ik , i.e., such that the equation best best best best best best best is satisfied. Again we do not update x best but update x bestnew if x ik yields a strict improvement in function value. Only after finishing the loop over k, we reexpand the model around the best point by G kl (x bestnew best l It is easy to see that the above method gives the unique quadratic interpolant to f at distinct points; in particular, we recover the exact objective function as In a diagonal triple search we only carry out the diagonal part of the above algorithm and take the off-diagonal elements of the Hessian from the previous iteration; thus only 2n additional function values are needed. Coordinate search. To find x l use a coordinate search based on a line search routine. A Matlab version of the actual line search used can be obtained electronically from http://solon.cma.univie.ac.at/ ~ neum/ms/ls/ This univariate line search program ls0 contains a parameter smaxls limiting the number of points used for the line search. It is possible to feed other points in addition to the starting point and their function values into the program, and these points are included in smaxls. A line search with along each coordinate. The first line search is started with the candidate for the base list. After the line search along the first coordinate, we can take the best point and its two nearest neighbors on both sides (or, if such points do not exist, the two nearest neighbors on one side) as fx l g. The subsequent line searches are started with the current best point obtained from the previous line search. After a line search in a coordinate i ? 1, we take the best point, the starting point of the line search (if it is different from the best point) and, if possible, the nearest neighbor of the best point on the other side as fx l g. The ith coordinate of the old x best has to be among fx l since otherwise we would lose all points through which the surface has been fitted previously. Local search. Now we have all ingredients at our disposal to describe the steps of the local search algorithm used in our implementation of MCS. Step 1. Starting with the candidate for the base list, we make a full triple search, where fx l are found by a coordinate search as described above. This procedure yields a new point x, its function value f , an approximation g of the gradient and an approximation of the Hessian G. Step 2. Set is the modified Cholesky factorization without scaling of [15, Section 2] and the starting value " used. Set of the current x is u k or v k and p k points outward, set we make a line search with ls0 along x+ffp. Let ff min and ff max be the bounds for ff obtained from the box bounds [u; v]. If p T Gp ? 0 and \Gammag T p=(p T Gp) 2 \Gammag T p=(p T Gp); otherwise we take - as the one where we expect the smaller function value according to the quadratic model. The values ff are used as input for ls0. If G is positive definite and " is chosen, the modified Cholesky factorization is the Cholesky factorization, the search direction is g, and we have - positive definite ff is close to 1. Set old is the function value of the current point at the beginning of Step 2 and f pred is the function value predicted by the above quadratic model with ff, i.e., fac is a measure of the quality of the quadratic model, and it is 1 if the line search did not improve the function value. Step 3. Stop if some limit on the number of visits to Step 3 (per local search) or the limit on function calls has been exceeded. Also stop if none of the components of the current x are at the boundary and the last triple search was a full one and the stopping criterion is fulfilled. Step 4. If some components of the current x are at the boundary and the stopping criterion is fulfilled, we make line searches with at most smaxls points along these coordinates. If the function value was not improved by these coordinate searches, stop. Step 5. If fac ? 0:1 or the stopping criterion was fulfilled in Step 3, we make a full triple search, otherwise we make a diagonal triple search. We make these triple searches only in the coordinates i such that the component x i of the current x is not at the boundary. The set fx l is taken to consist of x i , two neighbors at distance ffi and 2ffi if x i lies at the boundary) and ", where " denotes the machine accuracy. We obtain a new point x and an approximation of its reduced gradient g and its reduced Hessian G. Step 6. We set only in the components of the current x that are not at the boundary (p make a line search along x ff as input as before. The quantity fac is defined as in Step 2, where f old is the function value at the current point at the beginning of Step 6. and go to Step 3. The stopping criterion is fulfilled if the function value was not improved in Steps 5 and 6 (resp. Steps 1 and 2) or if jgj T max(jxj; jx old is an input parameter of our program and f 0 is a 'typical' function value, e.g., the smallest function value found in the initialization procedure. Base list. The local searches are only carried out at the end of each sweep. Then all candidates for the base list that have been collected in this sweep are ordered by ascending function value, and they are gone through as follows. Let x be a candidate for the base list. Step 1. We check whether we have already made a local search from this point. (Often a point belongs to two boxes.) If this is the case, we take the next candidate for the base list and go to Step 1. Step 2. Let w be the points already in the base list, and by renumbering assume that they are sorted by their distance to x, starting with the nearest point. For we do the following. Step 2a. Compute the function value at x increase i by one and go to Step 2a. Step 2b. Compute the function value at x increase i by one and go to Step 2a. Else if points seem to lie in the same valley. However, we do not discard x for local search but set x to the value x 0 or x 00 with the smaller function value, increase i by one and go to Step 2a. Else we do not subject x to local search because the four function values are monotonous, take the next candidate for the base list and go to Step 1. Step 3. Local search. Step 4. The new point x obtained from local search is subjected to a procedure similar to Step 2 in order to find out whether we have really found a new point, and only in that case it is put into the base list. If we disregard the splits according to the initialization list, A4 is fulfilled for our algorithm with 0:001. The factor 0.9 applies to the 'regular' case where for [x 00 y. For the 'safeguarded' case of subint for large y the worst case is at most the largest of the three numbers 1( and 1). Moreover, A5 is satisfied for the choice 7. Numerical results Test functions. Jones et al. [11] gave an extensive comparison of their direct method with various methods on seven standard test functions from Dixon & Szeg- [5] and two test functions from Yao [24]. Since our MCS algorithm is based on important insights from [11], we first consider the same test set to evaluate the efficiency of MCS. For each test function, the dimensions and box bounds, used by Jones et al. but inadvertently omitted in [11], are given in Table 1. We thank Don Jones for providing us with the code of the test functions. Label Test function Dimension Default box bounds GP Goldstein-Price 2 [\Gamma2; 2] 2 C6 Six-hump camel 2 [\Gamma3; 3] \Theta [\Gamma2; 2] Table 1. Dixon & Szeg- dimensions and box bounds All but the last four lines of Table 2 are taken from one of the tables of results of [11]. The fourth last line contains results for the differential evolution algorithm DE by Storn & Price [22] from http://http.icsi.berkeley.edu/ ~ storn/code.html, using the Matlab program devec2.m with the default values for the control parame- ters. Since this algorithm operates only at the global level, it takes a rather long time to find a minimum with high accuracy and therefore we used obtaining a relative error stopping criterion. The number of function evaluations needed for convergence was averaged over 25 runs for each test function. In the case of Hartman6, one run did not converge after 12 000 function evaluations and we averaged only over the remaining 24 runs. The last three lines give results for MCS, first over the same bound constraints as in [11], then averages over strongly perturbed box bounds, and finally for the unconstrained version. Details will be given below. When assessing the results, we have to bear in mind that the first 11 algorithms already appeared in the 1978 anthology edited by Dixon & Szeg- are therefore are somewhat old. The number of function calls needed for convergence is not the only method of assessing the quality of an algorithm, but it is an important one in the case of most real life applications, where function evaluations are expensive. Termination. In the presentation of test results, methods are usually compared on the basis of their performance on problems with known solutions. The algorithm is terminated when a function value within some tolerance of the global minimum has been found, and we also adopt this strategy. However, in practical problems, one does not know the solution in advance and needs a criterion that tells the program when to stop searching for a better local minimizer. This criterion should be stringent enough that it does not waste too many function values after the global minimum has been found, but it should also be loose enough to ensure that in typical cases, the algorithm does not terminate before the global minimizer has been found. Stochastic approaches to the design of suitable stopping criteria are surveyed in Section 6 of Boender & Romeijn [3]. One of the methods proposed there consists in stopping when the number n of local searches done is larger than a function N(w) of the number w of different local minima found so far. The function N(w) depends on the assumptions, and several specific implicit definitions of N(w) are given in [3]. Method S5 S7 S10 H3 H6 GP BR C6 SHU Bremmerman [5] (a) (a) (a) (a) (a) (a) 250 Mod. Bremmerman [5] (a) (a) (a) (a) 515 300 160 Zilinskas [5] (a) (a) (a) 8641 5129 Gomulka-Branin [5] 5500 5020 4860 T-orn [5] 3679 3606 3874 2584 3447 2499 1558 Gomulka-T-orn [5] 6654 6084 6144 Gomulka-V.M. [5] 7085 6684 7352 6766 11125 1495 1318 Price [5] 3800 4900 4400 2400 7600 2500 1800 Mockus [5] 1174 1279 1209 513 1232 362 189 Belisle et al. [1] (b) 339 302 4728 1846 Boender et al. [2] 567 624 755 235 462 398 235 Kostrowicki-Piela [12] (g) (g) (g) 200 200 120 120 Perttunen [16] (c) 516 371 250 264 82 97 54 197 Perttunen- Stuckman [17] (c) 109 109 109 140 175 113 109 96 (a) Jones et al. [11] (c) 155 145 145 199 571 191 195 285 2967 Storn-Price [22] (d)(h)6400 6194 6251 476 7220 1018 1190 416 1371 All but the last four lines are taken from [11]; missing entries were not available from the literature. (a) Method converged to a local minimum. (b) Average evaluations when converges. For H6, converged only 70 % of time. (c) Convergence defined as obtaining a relative error ! 0:01 %. (d) Convergence defined as obtaining a relative error Perturbed box bounds. (f) Unconstrained optimization problem. (g) Global minimum not found with less than 12 000 function calls. Average over 25 cases. For H6, average over 24 cases one case did not converge within 12 000 function values. (i) An asterisk indicates that the first local optimization gave the local optimum. Table 2. Number of function evaluations for various methods compared to MCS This result is theoretically justified for the random multiple start method only but may serve as a guideline also for other methods that use local searches. However, with MCS we try to do very few local optimizations only, and this reasoning appears inadequate. So far, we have not yet found a useful general purpose stopping criterion for MCS. For the purposes of the numerical tests reported here, the stopping criterion for MCS was taken as obtaining a relative error ! 0:01 % in the optimal objective function value (which happens to be nonzero always), i.e., \Gamma4 , which was also the criterion used by Jones et al. [11]. For the algorithms quoted in [11], results based on the definition of convergence used by their authors are reported. MCS control parameter settings. We applied a Matlab version of MCS with is the dimension of the problem, H(n; m) to the test functions and used a simple initialization list consisting of midpoint and boundary points, i.e., 2: The limit on visits to Step 3 per local search was set to 50, and the parameter fl in the stopping criterion for local optimization was taken as (cf. the subsection 'local search' of Section 6). Note that all examples have been run with identical parameter settings, so that no tuning to the individual test problems was involved. Modified bounds. We also investigated the stability of our results for MCS with respect to random perturbations of the box bounds. Instead of the default box bounds [u; v] given in Table 1, we employed the box bounds [u where j is a random variable that is uniformly distributed in the interval (\Gamma0:5; 0:5), but a value of j was only accepted in a given problem if at least one of the global minimizers was in [u The results given in the second last line of Table 2 were taken as an average over 25 runs with different perturbed box bounds for each test function. For Hartman6, we obtained one outlier for which the algorithm had not found the global minimum after 12 000 function calls, and we report a result averaged over the 24 remaining runs. Moreover, we applied MCS to the unconstrained optimization problem for the Dixon test set and added the results to Table 2. In this case, we cannot use an initialization list consisting of midpoint and boundary points any more. For infinite or very wide bounds, we suggest to use the following safeguarded version based on the function subint and again take l 2: if else Discussion. The results show that MCS seems to be strongly competitive with existing algorithms in the case of problems with reasonable finite bound constrained. MCS with unperturbed box bounds wins in 7 of 9 test cases against every competing algorithm and is only beaten once by Perttunen resp. Perttunen-Stuckman for the remaining two test functions. MCS with perturbed box bounds still wins against all competing algorithms for 5 test functions. Only the results for Shekel's functions seem to depend heavily on the choice of the box bounds, but they are comparable with the results of some other algorithms, and we do not know whether direct is stable with respect to perturbation of the box bounds. For unconstrained problems of dimension n - 4, the performance of MCS is less satisfactory. The reason is that in the exploration of an unbounded domain, it is easy to miss the region where the global minimum lies if one has already found a low-lying nonglobal minimizer. For example, for Shekel's functions the algorithm gets caught in the second best local minimizer (1; 1; 1; 1), which is a point in the initialization list. Further test problems. The Dixon & Szeg- test set has been criticized for containing mainly easier test problems. A more challenging test set was used in the first contest on evolutionary optimization (ICEO) at the ICEC'96 conference; cf. Storn Price [23]. This test bed contains five problems, each in a 5-dimensional and a 10-dimensional version, and on these test functions, MCS showed some limitations. The names and default box bounds of the ICEO test functions are given in Table 3, and the results are shown in Table 4. The first two lines in Table 4 are results, taken from [23], of two different versions of DE. Problem Name Box bounds 1 Sphere model [\Gamma5; 5] n 3 Shekel's foxholes [0; 10] n 4 Michalewicz's function [0; -] n 5 Langerman's function [0; 10] n Table 3. ICEO test functions and their box bounds We applied MCS with three different choices of the initialization list to the ICEO test functions. MCS1 is the standard version with midpoints and boundary points. For MCS2, we took x 1 i.e., the points are uniformly spaced but do not include the boundary points. For MCS3, we generated an initialization list with the aid of line searches. Starting with the absolutely smallest point in [u; v], we made line searches with ls0 with and along each coordinate, where the best point was taken as starting point for the next line search. The parameter nloc in ls0 determines how local or global the line search is since the algorithm tries to find up to nloc minima within smaxls function values. For the line searches in the local search method described in Section 6, was taken (entirely local line search). For each coordinate i, all local minimizers found by the line searches were put into the initialization list, and if their number was less than three, they were supplemented with the values obtained from ls0 closest to u i and v i . Finally, we applied MCS to the test functions used by Storn & Price [22]. Their names and box bounds are shown in Table 5, and their definition can be found in [22]. Problem 7 is a shifted version of ICEO2 for Problems 8 and 9 have general constraints and were therefore not used here. The first four lines of Table 6 are taken from [22]. ANM denotes the annealed Nelder Mead strategy of [20] and ASA the Adaptive Simulated Annealing method by Ingber [9, 10], and DE1 and DE3 are two different versions of DE. For MCS, we used an initialization list consisting of midpoint and boundary points when there was no global minimizer among the points in the initialization list; in the latter case, a different initialization list with Problem 4 contains a random variable, and the result presented for MCS was averaged over Problem 1 Problem 2 Problem 3 Problem 4 Problem 5 5D 10D 5D 10D 5D 10D 5D 10D 5D 10D MCS3 26 51 26956 - 32077 - 1903 - values. A dash indicates that a global minimizer was not found after 100 000 function calls. Table 4. Number of function values for the ICEO functions Problem Name Dimension n Box bounds 4 Quartic with random noise 8 Zimmerman's problem 2 constraints 9 Polynomial fit, 9, 17 constraints Table 5. Test functions of Storn & Price [22], their dimensions and box bounds ANM 95 106 90258 - ASA 397 11275 354 4812 1379 3581 - discontinuous test function; 2 noisy test function. A dash indicates that a global minimizer was not found after 100 000 function calls. Table 6. Number of function values for the Storn & Price functions runs, where convergence was defined as reaching a point with function value - 15. Problem 1 has a quadratic objective function, hence is easy for MCS, and Problem 3 is easy for MCS since the objective function is monotonous. 8. Conclusions The multilevel coordinate search algorithm MCS, presented in this paper, has excellent theoretical convergence properties if the function is continuous in the neighborhood of a global minimizer. In the current implementation, our test results show that MCS is strongly competitive with existing algorithms in the case of problems with reason-able finite bound constrained. In our comparison, MCS outperforms the competing algorithms almost always on the classical test problems set of Dixon & Szeg- o. For unconstrained problems of dimension n - 4, the performance of MCS is less since in the exploration of an unbounded domain, it is easy to miss the region where the global minimum lies if one has already found a low-lying nonglobal minimizer. The same problem applies for some hard test problems with a huge number of local minima. --R Rinnoy Kan Algorithms for Minimization without Derivatives The global optimization problem: an introduction A grid algorithm for bound constrained optimization of noisy func- tions Deterministic Approaches Very fast simulated re-annealing Lipschitzian optimization without the Lipschitz constant Diffusion equation method of global minimization: performance on standard test functions Genetic Algorithms Integer and Combinatorial Optimization On satisfying second-order optimality conditions using modified Cholesky factoriza- tions The rank transformation applied to a multiunivariate method of global optimization An efficient method for finding the minimum of a function of several variables without calculating derivatives Numerical Recipes in C A multistart global minimization algorithm with dynamic search trajectories Differential evolution - a simple and efficient adaptive scheme for global optimization over continuous spaces Minimizing the real functions of the ICEC'96 contest by differential evolution Dynamic tunneling algorithm for global optimization --TR --CTR Rong Yan , Alexander G. Hauptmann, The combination limit in multimedia retrieval, Proceedings of the eleventh ACM international conference on Multimedia, November 02-08, 2003, Berkeley, CA, USA H.-M. Gutmann, A Radial Basis Function Method for Global Optimization, Journal of Global Optimization, v.19 n.3, p.201-227, March 2001 evket lker Birbil , Shu-Cherng Fang , Ruey-Lin Sheu, On the Convergence of a Population-Based Global Optimization Algorithm, Journal of Global Optimization, v.30 n.2-3, p.301-318, November 2004 J. M. Gablonsky , C. T. Kelley, A Locally-Biased form of the DIRECT Algorithm, Journal of Global Optimization, v.21 n.1, p.27-37, September 2001 Y. Wu , L. Ozdamar , A. Kumar, TRIOPT: a triangulation-based partitioning algorithm for global optimization, Journal of Computational and Applied Mathematics, v.177 n.1, p.35-53, 1 May 2005 U. M. Garcia-Palomares , F. J. Gonzalez-Castao , J. C. Burguillo-Rial, A Combined Global & Local Search (CGLS) Approach to Global Optimization, Journal of Global Optimization, v.34 n.3, p.409-426, March 2006 Jaewook Lee, A novel three-phase trajectory informed search methodology for global optimization, Journal of Global Optimization, v.38 n.1, p.61-77, May 2007 A. Ismael Vaz , Lus N. Vicente, A particle swarm pattern search method for bound constrained global optimization, Journal of Global Optimization, v.39 n.2, p.197-219, October 2007 Panayiotis G. Georgiou , Chris Kyriakakis, Maximum likelihood parameter estimation under impulsive conditions, a sub-Gaussian signal approach, Signal Processing, v.86 n.10, p.3061-3075, October 2006 Michael Bartholomew-Biggs , Bruce Christianson , Ming Zuo, Optimizing Preventive Maintenance Models, Computational Optimization and Applications, v.35 n.2, p.261-279, October 2006
local optimization;coordinate search;global optimization;bound constraints
596237
Efficient and Adaptive Lagrange-Multiplier Methods for Nonlinear Continuous Global Optimization.
Lagrangian methods are popular in solving continuous constrained optimization problems. In this paper, we address three important issues in applying Lagrangian methods to solve optimization problems with inequality constraints.First, we study methods to transform inequality constraints into equality constraints. An existing method, called the slack-variable method, adds a slack variable to each inequality constraint in order to transform it into an equality constraint. Its disadvantage is that when the search trajectory is inside a feasible region, some satisfied constraints may still pose some effect on the Lagrangian function, leading to possible oscillations and divergence when a local minimum lies on the boundary of the feasible region. To overcome this problem, we propose the MaxQ method that carries no effect on satisfied constraints. Hence, minimizing the Lagrangian function in a feasible region always leads to a local minimum of the objective function. We also study some strategies to speed up its convergence.Second, we study methods to improve the convergence speed of Lagrangian methods without affecting the solution quality. This is done by an adaptive-control strategy that dynamically adjusts the relative weights between the objective and the Lagrangian part, leading to better balance between the two and faster convergence.Third, we study a trace-based method to pull the search trajectory from one saddle point to another in a continuous fashion without restarts. This overcomes one of the problems in existing Lagrangian methods that converges only to one saddle point and requires random restarts to look for new saddle points, often missing good saddle points in the vicinity of saddle points already found.Finally, we describe a prototype Novel (Nonlinear Optimization via External Lead) that implements our proposed strategies and present improved solutions in solving a collection of benchmarks.
Introduction Many applications in engineering, decision science and operation research are formulated as optimization problems. These applications include digital signal processing, structure optimization, engineering design, neural-network learning, computer-aided design for VLSI, and chemical control processing. High-quality solutions in these applications may have significant economical impacts, leading to lower implementation and maintenance costs while improving the quality of outputs. The constrained nonlinear global optimization problems that we study take the following form subject to g(X) where f(X) is an objective function, is a set of k inequality constraints, and is a set of m equality constraints. All f(X), g(X), and h(X) are assumed to be (nonlinear) differentiable real-valued continuous functions. Global minimization looks for a solution that satisfies all the constraints and is no larger than any other local minimum. This is a challenging problem as there may not be enough time to find a feasible solution, or even if a feasible solution is found, there is no way to show that it is minimal. In practice, one only seeks as many local minima as possible that satisfy the constraints, and pick the best local minimum. Active research in the past three decades has produced a variety of methods to find solutions to constrained nonconvex nonlinear continuous optimization problems [29, 10, 7, In general, they are divided into transformational and non-transformational methods. Non-transformational approaches include discarding methods, back-to-feasible-region methods, and enumerative methods. Discarding methods [11, 15] drop solutions once they are found to be while back-to-feasible-region methods [12] attempt to maintain feasibility by reflecting moves from boundaries if such moves go out of the current feasible region. Both of these methods have been combined with global search and do not involve transformations to relax constraints. Last, enumerative methods [10] are generally too expensive to apply except for problems with linear objectives and constraints, and for bilinear programming problems [1]. Transformational approaches, on the other hand, transform a problem into another form before solving it. Some well-known methods include penalty, barrier, and Lagrange-multiplier methods [14]. Penalty methods transform constraints into part of the objective function and require tuning penalty coefficients either before or during the search. Barrier methods are similar except that barriers are set up to prevent solutions from going out of feasible regions. Both methods have difficulties when they start from an infeasible region or when feasible solutions are hard to find. However, they can be combined with other methods to improve their quality. Lagrange-multiplier methods (or Lagrangian methods) introduce Lagrange variables to gradually resolve constraints through iterative updates. They are exact methods that optimize the objective using Lagrange multipliers to meet the Kuhn-Tucker conditions [14]. In view of their advantages, we use them for constraint relaxation in this paper. Given an optimization problem with only equality constraints, minimize f(X) (2) subject to the corresponding Lagrangian function and augmented Lagrangian function are defined as is a set of Lagrange multipliers. We use the augmented Lagrangian function in this paper as it provides better numerical stability. According to classical optimization theory [14], all the extrema of (4) (called saddle points), whether local or global, are roots of the following set of first-order necessary conditions. These conditions are necessary to guarantee the (local) optimality to the solution of (2). Because a saddle point is a local minimum of the Lagrangian function L(X; ) in the original-variable (X) space and a local maximum of L(X; ) in the Lagrange-multiplier () space, it can be obtained by solving the following dynamic system of equations d dt d dt which perform descents in the original-variable space of X and ascents in the Lagrange-multiplier space of . The goal of this paper is to develop new strategies to improve the convergence speed and solution quality of Lagrangian methods. We achieve this by three approaches. First, we study in Section 2 methods to eliminate oscillations when inequality constraints are transformed into equality constraints. These help improve convergence time as compared to existing methods of adding slack variables. Second, we present in Section 3 a method to improve the convergence speed of Lagrangian methods by adaptively adjusting the relative weights between the objective and the constraints. Third, we discuss in Section 4 a global-search method that looks for multiple saddle points in a continuous trajectory, without restarting the search from random starting points. We also present our prototype Novel that integrates these strategies. Finally, we show in Section 5 new improved results on a collection of benchmarks. Handling Inequality Constraints Lagrangian methods work well with equality constraints, but cannot directly deal with inequality constraints (1), except in some simple cases where one can directly solve the first-order equations in closed form. In general, inequality constraints are first transformed into equivalent equality constraints before Lagrangian methods can be applied. 2.1 Transformation Using Slack Variables One possible transformation [14] to handle inequality constraint is to add a slack variable z i to transform it into an equality constraint g i (X)+z 2 simplification [14], the augmented Lagrangian function for problem (1) becomes where and are Lagrange multipliers. In the same way as (6), a saddle point to (7) can be reached by doing descents in the X space and ascents in the and space. The balance between descents and ascents depends on the magnitudes of Lagrange multipliers and , which play a role in balancing objective f(X) and constraints h(X) and g(X) and in controlling indirectly the convergence speed and solution quality of the Lagrangian method. At a saddle point, the forces due to descent and ascent reach a balance through appropriate Lagrange- multiplier values. To emphasize how the relative weights affect the convergence speed and solution quality, we introduce an additional weight w into (7) and get a weight on the objective. When which is the original Lagrangian function. The corresponding dynamic system is as follows d dt d dt d dt Starting from an initial point (X(0); (0); (0)), we solve the dynamic equations (9)-(11) using an ordinary differential equation solver LSODE 1 and observe a search trajectory (X(t); (t); (t)). When a saddle point is on the boundary of the feasible region (which is true for most problems in the benchmark collection [6]), the dynamic equation approaches it from both the inside and outside of the feasible region. We observe three behaviors of the search trajectory: ffl The trajectory gradually reduces its oscillations and eventually converges; ffl The trajectory oscillates within some range but never converges; ffl The magnitude of oscillations increases, and the trajectory eventually diverges. To illustrate these three behaviors (divergence, oscillations without convergence, and reduction of oscillations until convergence), we apply (9)-(11) to solve Problem 2.3 in [6], a quadratic programming problem with linear inequality constraints. The problem is given as follows: Minimize 2subject to 2x 2 It is a quadratic nonlinear programming problem with linear inequality constraints. We set 0), the initial point at 0, at the middle of the search space, i.e., x 1 LSODE is a solver for first-order ordinary differential equations, a public-domain package available from http://www.netlib.org. Objective Iterations (*1000)0.51.52.5 Violation Iterations (*1000) Figure 1: The objective and maximum violation oscillate using the slack-variable method (w = 1). -12.2 -11.4 -11.2 -10.4 -10.2 Objective Iterations (*1000)0.020.060.10 Violation Iterations (*1000) Figure 2: The objective and maximum violation converge after oscillations subside using the slack- variable method We further set the maximum (logical) time for LSODE to be t which is divided into small units of resulting in a maximum of 10 5 iterations (= t max =4t). The stopping condition for (9)-(11) is the Lyupunov condition: The dynamic system stops when it converges or when it reaches the maximum number of iterations. When quickly into infinity, meaning that the original Lagrangian method will diverge. If we scale the objective by 10 (i.e., then the objective value oscillates within the range [\Gamma17; \Gamma10], while the maximum violation v max (t) is between 0 and 0:4, as shown in Figure 1. Here, v max (t) at time t is defined as If we further reduce w to 1=15, then the oscillations subside, and the trajectory eventually converges (see Figure 2). Intuitively, the occurrence of oscillations can be explained as follows. Suppose we start from an infeasible point initially As the search progresses, the corresponding i (t) increases and pushes the trajectory towards a feasible region. At some time t, the inequality constraint g i (X(t)) 0 is satisfied for the current point X(t). At this point, d i (t) negative when Hence, the trajectory decelerates but continues to move into the feasible region even when the corresponding constraint, g i (X(t)), is satisfied. The movement of the trajectory inside the feasible region eventually stops because the local minimum is on the boundary, and the corresponding force due to descents in the objective space pushes the trajectory outside the feasible region. Likewise, when the trajectory is outside the feasible region, a force due to the constraints pushes the trajectory inside the feasible region. If these two forces are not well balanced, the search may diverge or oscillate without convergence. 2.2 Transformation Using the MaxQ Method To avoid oscillations in the method based on slack variables, we would like a search to converge to a local minimum without oscillations when it is on the boundary of a feasible region and when the trajectory is outside the feasible region. This is done by our proposed MaxQ method. Without loss of generality, we ignore equality constraints in the following discussion for sim- plicity, knowing that equality constraints are handled in the way described in Section 1. The MaxQ method transforms an inequality constraint as follows. where q i are control parameters to be determined later. For a given X , if The augmented Lagrangian function is It is important to choose suitable control parameters q i they affect the convergence speed of our method. One can easily show that, when q i 1, inequality constraint is equivalent to equality constraint Suppose that q i is a constant. Using a dynamic system similar to that in (9)-(11) to solve Lagrangian function (16), we need to evaluate partial derivative 5X p i (X), where (X) is not continuous, and the derivative of L q (X; ) is not continuous when g However, the continuity of derivatives is required by most differential-equation solvers, such as LSODE. For this reason, we require q i ? 1 in (15). One way of selecting q i is to make it very close to 1, namely, 1. At this point, the dynamic system will approach a feasible region slowly when the saddle point is on the boundary of the feasible region. This is true because p 0 independent of how far the current point X is away from the feasible region. Thus, larger control parameters q i are needed for fast convergence if the current point X is far from the feasible region. In contrast, if we choose meaning that LSODE converges very slowly towards the saddle point on the boundary of the feasible region. Taking these facts into account, in order to have fast convergence, we should adapt q i dynamically as the search goes to a saddle point. Since different inequality constraints may have different convergence rates to the saddle point, we associate inequality constraint g i (X) 0 with its own control parameter q i . Each parameter will be updated dynamically based on the value of g i (X): q i is large if g i (X) AE 0, and q i is gradually reduced to a value approaching 1 when the search is close to the saddle point on the boundary. One possible choice of q i is as follows. are two parameters that control the shape of function q i (x). When g i (X) approaches 0, q i will approach 1. The dynamic equation to solve (16) is d dt d dt where In the proof above, we assume that q i takes the form in (17), where s 2. With this choice, very fast from 1 to 0 near the saddle point as making it difficult for LSODE to find a suitable step size in order to reach the saddle point. To let the gradient change smoothly, we set s Note that (16) is similar to (7) in the sense that both use the max function. The main difference is that (16) avoids the case in (7) in which inequality constraint g i (X(t)) 0 is satisfied at time t but g i (X(t)) also appears in the Lagrangian function. When g i (X(t)) is satisfied, it is meaningful to minimize f(X) independent of the value of g i (X(t)). 2.3 Convergence Properties of the MaxQ Method There are two types of saddle points X as shown in Figure 3. When saddle point X is within a feasible region, i.e., This means that , an equilibrium point of dynamic system (18)-(19), is given by d dt dt Thus, the trajectory controlled by (18)-(19) will converge to this saddle point X . When the saddle point X is on the boundary of the feasible region shown in Figure 3b, it will be asymptotically approached from outside the feasible region (right side of Figure 3b). In order to prove this, we only need to show that X is asymptotically a regular point of the constraints transformed into an equivalent equality constraint Because X is on the boundary of the feasible region, In addition, when X \Gamma! X , 1. By taking limits, we obtain lim lim Therefore, lim X!X 5 according to (20), which means that asymptotically regular point X of p i (X) is the same as that of the original constraint g i (X) 0. Since a saddle point, if it exists, must be a regular point of g i (X) [14], X can be asymptotically reached by the dynamic system (18)-(19). feasible region feasible region (a) (b) Figure 3: Relationship between saddle point and feasible region. (a) The saddle point is within the feasible region; (b) The saddle point is on the boundary of feasible region. 2.4 Dynamic Conversion of Inequality Constraints to Equality Ones in MaxQ As discussed above, if solution X is on the boundary of a feasible region, i.e., when some equals zero, then dynamic system (18)-(19) cannot reach this point exactly. This point can only be approached with a precision proportional to the convergence condition of the LSODE solver. However, this may take a long time even if q i is changed dynamically. This happens because the violation of constraint g i (X ) is small when the current point X(t) is close to the boundary X (from the right side of Figure 3b). Consequently, the increment of the corresponding Lagrange multiplier i will be small, leading to slow progress towards the saddle point. Suppose we know that some g j (X solution. In this case, faster convergence can be achieved if we consider g j (X) as an equality constraint, rather than an inequality constraint, The difficulty, however, is that it is impossible to know in advance which inequality constraints g j (X) 0 will satisfy the boundary condition (i.e., solution on the boundary. We can utilize the property that, when X(t) is very close to the boundary with respect to a particular inequality constraint (g j (X(t)) is positive and close to zero) and convergence of this constraint is slow, it is very likely that the saddle point is on the boundary with respect to this constraint. At this point, we can dynamically convert inequality constraint g j (X(t)) 0 into equality constraint g j to improve the convergence rate. Since we solve the dynamic system using LSODE, let X and X 0 be the points of two successive iterations. The conversion of inequality constraint occurs when the following two conditions are satisfied. ffl The dynamic system converges to some point X when it changes very little for a window of 10 iterations. We define the current point to have slow convergence when #fmax our experiments). ffl The dynamic system converges to the boundary when g j (X) is very close to zero; that is, our experiments). Both conditions are very important. Without the first condition, trajectory X(t) that occasionally passes the boundary of g j (X(t)) would erroneously cause some inequality constraints to be con- verted. It makes sure that the trajectory really changes very little during a period of time. The second condition guarantees that only those inequality constraints very close to the boundary can be converted into equality ones. Note that dynamic conversion can happen to many inequality constraints at the same time as long as they satisfy these two conditions. If dynamic conversion is performed on inequality constraint g j (X) 0, then the terrain of Lagrangian function L q (X; ) will be changed and be totally different. To maintain the search direction in the original-variable space X , we have to adjust Lagrange multiplier j . Let the current point just before the conversion be (X; j ). The Lagrangian term associated with inequality constraint according to the conversion conditions. The derivative of L j (X; j ) with respect to X and j are ih be the point after the conversion. This means that we apply equality constraint at the current point before. The Lagrangian term related to g j and the derivative of " with respect to X and " j are Since the control parameter q j is close to 1, the search direction in the Lagrange-multiplier space changes very little, meaning that 5 j independent of the value j . To retain the search direction in the original-variable space X , we set 5 X L j (X; and get 2.5 Illustration of the MaxQ Method As is in the slack-variable method, we introduce an additional weight w into the original augmented Lagrangian function in the MaxQ method. a weight on the objective, is the Lagrange multiplier for equality constraints, , the Lagrange multiplier for inequality constraints, and To show how MaxQ avoids divergence and oscillations that occur in the slack-variable method, we consider the same problem 2.3 defined in (12) [6]. The starting point is at the middle of the search space, the same as that used in the slack-variable method. Three cases were tested: no scaling, scaling the objective by 10 (i.e., scaling the objective by 15. All of them converge with similar behavior. Figure 4 shows the second case. Obviously, MaxQ has smoother and better convergence property as compared to the slack-variable method. The solution to problem 2.3 is on the boundary of the feasible region as shown in Figure 3b. Since all the Lagrange multipliers are zero initially, only objective function f(X) takes effect in the Lagrangian function, causing the trajectory to move away from the feasible region. The Lagrange Objective Iterations (*1000) Violation Iterations (*1000) Figure 4: The objective and maximum violation converge smoothly using MaxQ for Problem 2.3 defined in (12) [6] (w = 1). The number of iterations required is large and is over 94,000. multipliers then increase, pushing the trajectory back towards the boundary. Hence, the objective- function value increases, and the value of the maximum violation decreases (see Figure 4). Note that there is a gap between our current implementation of MaxQ and its theoretic result, in the sense that the analytic proof requires s in (17) but LSODE uses s or 3:0. This gap causes the MaxQ method to converge slowly sometimes, like the case shown in Figure 4 that requires over 94,000 iterations. In this example, MaxQ reduces the constraint violation very quickly in the beginning, but slowly afterwards. This problem can be solved by two approaches. The first is to use another differential-equation solver that is insensitive to quick changes of gradients, making the analytic result hold during the search. This will be investigated in the future. The second is to adaptively adjust the relative weight w between the objective and the constraints during the search. As soon as we detect divergence, oscillations, or slow convergence, w is adjusted accordingly. When we apply this strategy in the slack-variable method [34], it is able to eliminate divergence, reduce oscillations, and greatly speed up convergence. We discuss the application of this strategy in the MaxQ method in the next section. 3 Adaptive Lagrange-Multiplier Method In the last section, we have studied two methods to handle inequality constraints. We have shown that the slack-variable method is sensitive to the relative weights between the objective and the constraints, and that its implementation may diverge, oscillate or converge. Although the MaxQ method does not oscillate and is not sensitive to the relative weights, careful weighting may help accelerate its convergence speed. The proper weights for fast convergence, however, are problem-dependent and are impossible to choose in advance. In this section, we describe a strategy for MaxQ to dynamically adapt the weights based on the behavior of the search progress. This is based on the strategy we have developed for the slack-variable method [34]. 3.1 Dynamic Weight-Adaptation Strategy Lagrangian methods rely on two counteracting forces to resolve constraints and find high-quality solutions. When all the constraints are satisfied, Lagrangian methods rely on gradient descents in the objective space to find local minima. On the other hand, when any of the constraints are not methods rely on gradient ascents in the Lagrange-multiplier space in order to increase the penalties on unsatisfied constraints and to force the constraints into satisfaction. The 1. Set control parameters: time interval 4t initial weight maximum number of iterations i max 2. Set window size 3. is the iteration number */ 4. while j i max and stopping condition is not satisfied do 5. advance search trajectory by 4t in LSODE to get to (X 6. if trajectory diverges then reduce w; restart the algorithm by going to Step 2 7. record useful information for calculating performance metrics 8. if ((j mod Nw ) == /* Test whether w should be modified at the end of a window */ 9. compute performance metrics based on data collected 10. change w when the conditions are satisfied 11. end while Figure 5: Framework of the dynamic weight-adaptation algorithm balance between gradient descents and gradient ascents depends on the magnitudes of Lagrange multipliers and , which play a role in balancing objective f(X) and constraints h(X) and g(X) and in controlling indirectly the convergence speed and solution quality of the Lagrangian method. At a saddle point, the forces due to descent and ascent reach a balance through appropriate Lagrange-multiplier values. Combining augmented Lagrangian functions (8) and (25), we get a general augmented Lagrangian function as follows: where w is the relative weight, is the Lagrange multiplier for equality constraints, and for inequality constraints. L ineq (; X) depends on the way to deal with inequality constraints. It can be the slack-variable method or MaxQ. Starting from an initial point (X(0); (0); (0)), the dynamic system to find saddle points is d dt d dt d dt Figure 5 outlines the algorithm. Its basic idea is to first estimate the initial weight (Step 1), measure the performance metrics of the search trajectory (X(t); (t); (t)) periodically, and adapt w(t) to improve convergence time or solution quality. Let t max be the total (logical) time for the search, and t max be divided into small units of time 4t so that the maximum number of iterations i t. Further, assume a stopping condition if the search were to stop before t max (Step 4). Given a starting point set the initial values of the Lagrange multipliers to be zero; i.e., be the point of the i th iteration, and v max (i) be its maximum violation defined in (14). Objective Iterations (*1000) Violation Iterations (*1000) Figure The objective and maximum violation converge after 756 iterations for MaxQ using dynamic weight adaptation (initial To monitor the progress of the search trajectory, we divide time into non-overlapping windows of size Nw iterations each (Step 2). In each window, we compute some metrics to measure the progress of the search relative to that of previous windows. For the q th window calculate the average value of v max (t) using (14) over all the iterations in this window, During the search, we employ LSODE to solve dynamic system (27)-(29), and advance the trajectory by time interval 4t in each iteration in order to arrive at point (X At this point, we test whether the trajectory diverges or not (Step 6). Divergence happens when the maximum violation v max (t) is larger than an extremely large value (e.g. 10 20 ). If it happens, we reduce w by a large amount, say w (= w=10, and restart the algorithm. In each iteration, we also record some statistics such as v max (t), which will be used to calculate the performance metrics for each window (Step 7). At the end of each window or every Nw iterations (Step 8), we decide whether to update w based on the performance metrics (30) (Step 9). In our current implementation, we use the average value of maximum violation v max (t). In general, other application-specific metrics can also be used. Based on these measurements, we adjust w accordingly (Step 10). As explained in section 2.5, the major problem with MaxQ sometimes is its slow convergence, which can be measured by how fast the maximum violation decreases. Therefore, we monitor the reduction of v q , the average value of the maximum violation. If it is found to decrease slowly, i.e., is a threshold (e.g. reduce weight w by half (w (= w=2). Its effect is to put more weight on the constraints, thus pushing the trajectory quickly to a feasible region. Note that when comparing the values between two successive windows and q, both must use the same weight w; otherwise, the comparison is not meaningful because the terrain may be different. Hence, after adapting w, we should wait at least two windows before changing it again. 3.2 Illustration of the Weight-Adaptation Strategy for MaxQ To use the dynamic weight-adaptation method, we set time interval size Corresponding to Figure 4, we start from the initial weight the same starting point (X(0); (0); (0)). Figure 6 shows the resulting search profile, in which the search converges using only 756 iterations, which is significantly better than the 94,000 iterations without weight adaptation. It is important to note that solution quality is the same as that in Figure 4, and both obtain the objective value \Gamma11:25 when the search converges. The remaining issue in our algorithm is to choose a good starting value for is too large, then it is difficult for the constraints to be satisfied, resulting in slow convergence for MaxQ. If w(0) is too small, then the objective is under-weighted, and the search may converge to a worse saddle point. After studying many benchmark problems [6], we found that if we set the starting points of the search to be the ones given by [6] and both the slack-variable and MaxQ methods work well. This suggests us to start with our experiments. A small number of the problems may still need further adaptation of w during the search. 4 Global Search for Saddle Points The Lagrangian method presented in the last two sections only looks for a single saddle point, behaving like a local search. To find multiple saddle points, global-search methods are needed to bring the search out of local saddle points. In this section, we present our global search strategy, followed by a description of our trace-based search method. We first describe some previous work on solving unconstrained nonlinear optimization problems since the dynamic system for solving a Lagrangian formulation can be treated as an unconstrained nonlinear optimization problem. 4.1 Previous Work on Unconstrained Nonlinear Optimization Unconstrained nonlinear optimization, in general, is multi-modal with the following features. (a) Flat regions may mislead any gradient-based methods. (b) Gradients may differ by many orders of magnitude, making it difficult to use gradients in any gradient-based search method. (c) The existence of many local minima may trap some search methods, leading to suboptimal solutions. Based on these observations, global search methods should be able to use gradient information to descend into local minima, and be able to escape from local minima once it gets there. Search methods can be classified into local and global. Local search algorithms, such as gradient-descent and Newton's methods, find local minima efficiently and work best in uni-modal problems. Global-search algorithms, in contrast, employ some heuristic or systematic strategies to look for global minima and do not stop after finding a local minimum [18, 29, 14]. Note that both gradients and Hessians can be used in local- and global-search methods [29]. Local-search algorithms have difficulty when the surface is flat where gradients are close to zero, or when gradients can be in a large range, or when the surface is very rugged. If gradients vary greatly, a search may progress too slowly when gradients are small or may over-shoot when gradients are large. If the function surface is rugged, a local search from a randomly chosen starting point will most likely converge to a local minimum that is near the starting point and results in a solution worse than the global minimum. Moreover, these algorithms may require some properly chosen parameters as incorrectly chosen parameters may cause slow or unstable convergence. To avoid getting stuck in local minima in local-search methods, many global-search methods have been developed. These methods rely on local search to determine local minima, but focus on bringing the search out of local minima when it gets there. Few deterministic methods have been developed, most of which apply deterministic heuristics to bring a search out of a local minimum (such as those in covering and generalized descent methods) [29, 24, 10]. These deterministic methods do not work well when the search space is too large to be covered adequately [29, 9, 4]. Trajectory methods rely on an internal force (such as the momentum of the trajectory) to continue to move the trajectory once it gets to a local minimum. They do not work well when gradients are steep and the search space is rugged. Probabilistic methods, on the other hand, rely on probability to make decisions. The simplest probabilistic algorithm uses restarts to bring a search out of a local minimum when little improvement can be made locally [20, 35, 36]. More advanced methods rely on probability to indicate whether the search should ascend from a local minimum. Simulated annealing is one of these methods that accepts up-hill movements [22, 30, 2, 3, 21, 13]) based on some probability. Other stochastic methods rely on probability to decide which intermediate points to be interpolated as new starting points, such as random recombinations and mutations in evolutionary algorithms [8]). All these algorithms are weak in either their local search [14] or their global search [9, 29, 24, 23]. For instance, gradient information useful in local search is not well used in simulated annealing and evolutionary algorithms. In contrast, gradient-descent algorithms with multi-start are weak in their global search. Other probabilistic methods utilize sampling to determine the terrain and to decide where to search [24, 16, 29, 31]. Such strategies may fail when the terrain is very rugged or when the search gets trapped in a deep but suboptimal basin. This happens in clustering methods whose performance is similar to that of random restarts when the terrain is rugged [24, 28]. Bayesian methods also do not work well. Most samples that they collect randomly from the error surface are close to the average error value, and these samples are inadequate to model the behavior at minimal points [27, 17, 29, 31]. In addition, they are computationally expensive and are usually not applicable for problems with over twenty variables. 4.2 Novel: A Trace-Based Search Method To find saddle points of (26), Eq. (27) performs descents in the original-variable space to locate local minima of the objective function when the constraints are satisfied, whereas (28)-(29) perform ascents in the Lagrange-multiplier space when the constraints are violated. We are interested to move the trajectory from one saddle point in a feasible space to another, without having to restart the search. To do so, we add an external force to pull the search out of a saddle point in the original-variable space continuously and escapes from it without restarts [32, 26]. It has three features: exploring the solution space, locating promising regions, and finding saddle points. In exploring the solution space, the search is guided by a continuous terrain-independent trace that does not get trapped by local saddle points. This trace is usually an aperiodic continuous function that runs in a bounded space. It continues to move over the search space independent of local gradients. In locating promising regions, our trace-based method uses local Lagrangian search to attract the search to a saddle point but relies on the trace to pull it out once little improvement can be found. Finally, our trace-based method selects one initial point from each promising local region and uses them as initial points for a Lagrangian search to find saddle points. In exploring the search space, the trace plays an important role in discovering regions with new local saddle points. A trace is a continuous aperiodic function of (logical) time that generates a trajectory. At time both the trace and the trajectory start at the same point. As the trace moves from point X 1 to point X 2 , the trajectory moves from point Y 1 to Y 2 , where Y 2 is a function of the local gradient at Y 1 and the distance between X 1 and Y 1 (see Figure 7). These two forces (descent into a local saddle point and attraction exerted by the trace) form a composite vector that represents the route taken by the trajectory. Trace function Trajectory 1 Trajectory 2 Trajectory 3 Trace Trace direction Moving direction Gradient direction Trajectory Figure 7: Novel has two phases: global search and local refinement. In the global-search phase, the trajectory combines a Lagrangian search and the pull exerted by the moving trace. In the local-search phase, the trajectory is sampled to collect starting points for pure Lagrangian searches. When dealing with constrained problems formulated by Lagrangian functions, there are two different sets of variables, the original variables X and the Lagrange multipliers and . Intuitively, there is no need of a trace in the Lagrangian space because the Lagrange multipliers are responsible to bring the trajectory to a feasible region and does not involve finding local minima. As the trace pulls the search out of a local saddle point and enters an infeasible region, the corresponding Lagrange multipliers will be automatically increased based on the dynamic system, and then push the trajectory back into the feasible region. In this sense, the Lagrange multipliers are passive since they change with the constraints. If one also uses a trace function in the Lagrangian space, the force imposed by the trace and that by the Lagrange multipliers may contradict. In short, the trace does not affect the search in the Lagrange-multiplier space defined by (28) and (29). This strategy corrects a problem in our original trace-based strategy presented in [32] that uses a trace in the Lagrangian space. The overall dynamic system for nonlinear constrained optimization is as follows. d dt d dt d dt where g and t are constants controlling the relative weights between local search and global exploration. Note that our proposed method is a trajectory-based method, but differs from existing trajectory-based methods that rely on internal forces to modify the trajectory. Instead, it uses an external force (a traveling problem-independent trace) to pull the trajectory out of local minima. The design of a good trace function TX (t) is very important because our method relies on it to travel through the solution space. Four criteria have been considered up to now. First, the trace should be aperiodic so that it does not return to the same starting points and regenerates possibly the same trajectory. Second, the trace needs to be continuous in order to be differentiable. This allows the generated trajectory to follow the terrain in a continuous manner without restarting from new starting points. Third, the trace should be bounded so that it will not explore unwanted regions. Last, the trace should be designed to travel from coarse to fine so that it examines the search space in greater details when more time is allowed. Since an analytic approach to design a good trace function is intractable, we have studied some heuristic functions and fine-tuned them [25]. In the following, we summarize our observations. ffl Our trace-based method is a fine-level global search in the sense that the search space covered by a trace grows linearly with the length of the trace (and, therefore, the time to complete the algorithm). However, a search space grows exponentially with respect to the number of dimensions. Hence, a trace-based method does not give good coverage when the search space is large. To overcome this limitation, the method may need to be combined with other coarse-level global searches, such as simulated annealing and genetic algorithms. This hybrid approach does not work well for constrained optimization problems because existing coarse-level global-search algorithms generally have difficulty in satisfying constraints. Starting from points where constraints are violated usually do not lead to good feasible solutions. ffl Our trace-based method relies on the distance between the current position of the trace and that of the trajectory to pull the trajectory out of local saddle points. When local gradients are very large, the external force due to the trace may not be enough to pull the trajectory out of local saddle points. To address this problem, we have developed a dynamic variable scaling method that scales a variable dimension when the gradient of the trajectory in this dimension exceeds a threshold (e.g. Scaling not only reduces the local gradient of the trajectory but also increases the distance between the trace and the trajectory, thereby providing adequate force to pull the trajectory out of local saddle points. By controlling the scaling factor, we are able to explore a larger search space. Based on substantial experiments, we have designed an aperiodic trace function as follows. where i represents the i'th dimension, ae is a coefficient specifying the range, and n is the dimension of the original variable space X . We have described our trace-based method using one trace function. In general, the method can be cascaded, using the output trajectory of one stage as the trace of the next stage. This bootstrapping allows trajectories in later stages to go deeper into a local region, thereby providing better starting points for the local-search phase. In Figure 7, we have shown three stages of the global-search phase, each of which outputs a trajectory based on (31). In the first stage of the global search, the user-defined trace function leads the trajectory to form Trajectory 1 in Figure 7. In the second and third stages of the global search, the trace function TX (t) is the trajectory from the previous stage. Using the trajectories output from the three stages, we identify a set of promising starting points and perform local Lagrangian searches from them. The final result is the best solution among all these local searches. In our implementation, when an output trajectory is a collection of discrete sample points of a continuous trajectory, interpolations are performed to form a continuous trace for the next stage. Generally, weights g and t can have different values in different global stages. For example, t can be set to have large values relative to g in the first stage so that global search is emphasized. In later stages, g can have larger values, and the search is more focused on a local region. In the simplest case, g and t are set to constants and have the same values in all the stages. In our experiments, we have set g a = 1 and stages. We have implemented our MaxQ, adaptive weighting, and trace-based search in a prototype Novel (Nonlinear Optimization Via External Lead) that extends our previous work in trace-based search [32, 26]. The prototype can solve both nonlinear constrained as well as unconstrained optimization problems. It has been applied to solve design problems in signal processing [33], neural-network learning [26], and benchmark problems in operations research. Results on the latter are presented in the next section. 5 Experimental Results In this section we describe our experimental results on some existing benchmarks [6]. These benchmarks are challenging because they model practical applications that have been studied extensively in the past. As a result, improvements are generally difficult. We use a common set of parameters, including step size and trace function, to solve all the problems. The reason for not tuning the parameters is to avoid any bias, as good solutions can always be obtained by sufficient tuning. We further set the starting points of Novel as those suggested in the benchmark set. The only exception is the problem-specific search range of the trace, which we set manually based on the solutions reported in the benchmarks. In practice, this is reasonable as search ranges are generally known. In cases that the search range is not available, we use trial and error, starting from a small range and gradually increasing it until no improvement in solutions can be found. In the global-search phase of Novel, we used three stages that produce three trajectories (see Figure 7). Using chose 100 starting points from each trajectory based on their Lagrangian values for Lagrangian searches. After 300 Lagrangian searches, we report the best solution. Table 1 summarizes the results found by Novel. Column 1 lists the problem identifications that appear in the benchmark collection [6]. Column 2 shows the problem-dependent search range that the trace covers. Column 3 shows the best known solutions reported in [6], and Column 4, the solutions reported by Epperly [5]. Here, symbol '\Gamma' means that the method is unable to find a solution for the corresponding problem. Column 5 shows the results obtained by Novel using the slack-variable method where the dynamic weight-adaptation strategy described in Section 3 is used. Without using adaptive weights, more than half of these problems cannot be solved due to divergence and oscillations described in Section 2.1. The last column shows the results obtained by Novel using MaxQ. Results in bold font are improved by Novel over the best known results, with improvements of up to 10%. Our results indicate that Novel is robust in discovering new regions and in escaping from local traps. 6 Conclusions In this paper, we have studied three strategies to improve Lagrangian searches for solving nonlinear constrained optimization problems. First, we have studied a new method called MaxQ to convert inequality constraints into equality constraints. A Lagrangian search using the new constraints approaches saddle points on boundaries of feasible regions without oscillations. This overcomes the problems of oscillations and divergence when inequality constraints are converted by adding slack variables. Second, we have developed a method to adaptively adjust the relative weights between the objective and the constraints in a Lagrangian formulation. We show that adaptive weighting can improve the convergence speed of Lagrangian methods, without affecting the solution quality. Finally, we have applied a trace-based search to bring a trajectory from one saddle point into another in a continuous fashion. Our method generates information-bearing trajectories in Table 1: Results on a collection of constrained optimization benchmarks [6] comparing Novel using MaxQ, Novel using the slack-variable method, and Epperly's method [5]. Search times are in CPU seconds on a Sun SS 10/51. Improved solutions found by MaxQ are indicated in bold font. Symbol '\Gamma' means that the method was not able to find a solution for the corresponding problem. Problem Novel Search Best Known Epperly's Slack Variable MaxQ ID Range Solutions Solutions Solutions Solutions 2.1 1.0 \Gamma17:00 \Gamma17:00 \Gamma17:00 \Gamma17:00 2.2 10.0 \Gamma213:00 \Gamma213:00 \Gamma213:00 \Gamma213:00 2.3 10.0 \Gamma15.00 \Gamma15.00 \Gamma15.00 \Gamma15.00 2.4 10.0 \Gamma11.00 \Gamma11.00 \Gamma11.00 \Gamma11.00 2.5 1.0 \Gamma268.00 \Gamma268.00 \Gamma268.00 \Gamma268.00 2.6 1.0 \Gamma39.00 \Gamma39.00 \Gamma39.00 \Gamma39.00 2.7(2) 40.0 \Gamma884.75 \Gamma884.75 \Gamma884.75 \Gamma884.75 2.8 25.0 15990.00 15990.00 15639.00 15639.00 3.1 3.2 50.0 \Gamma30665.50 \Gamma30665.50 \Gamma30665.50 \Gamma30665.50 3.3 10.0 \Gamma310.00 \Gamma310.00 \Gamma310.00 \Gamma310.00 3.4 5.0 \Gamma4.00 \Gamma4.00 \Gamma4.00 \Gamma4.00 4.3 5.0 \Gamma4.51 \Gamma4.51 \Gamma4.51 \Gamma4.51 4.4 5.0 \Gamma2.217 \Gamma2.217 \Gamma2.217 \Gamma2.217 4.5 5.0 \Gamma11.96 \Gamma13.40 \Gamma13.40 \Gamma13.40 4.6 5.0 \Gamma5.51 \Gamma5.51 \Gamma5.51 \Gamma5.51 4.7 5.0 \Gamma16.74 \Gamma16.74 \Gamma16.75 \Gamma16.75 5.2 50.0 1.567 \Gamma 1.567 1.567 5.4 50.0 1.86 \Gamma 1.86 1.86 6.2 100.0 400.00 400.00 400.00 400.00 6.3 100.0 600.00 600.00 600.00 600.00 6.4 100.0 750.00 750.00 750.00 750.00 7.2 100.0 56825.00 \Gamma 56825.00 56825.00 7.3 7.4 its global search based on a user-defined trace function, and samples these trajectories for good starting points in its local search. This overcomes the problem of using random restarts when the trajectory is already in the vicinity of good saddle points. We have applied MaxQ, adaptive weighting, and trace-based search in Novel [32, 26], a global optimization system we have developed earlier to solve constrained as well as unconstrained optimization problems. We have tested many benchmark problems derived from manufacturing, computed aided de- sign, and other engineering applications and have compared MaxQ to the method based on slack variables [14] and that of Epperly [5]. Our results show that MaxQ is more robust in convergence and has found solutions that are either better than or the same as existing solutions. Our future work in this area will be on finding better trace functions, parallelizing the execution on massively parallel computers, and studying other challenging applications in neural-network learning and signal processing. --R Global minimization by reducing the duality gap. Generalized simulated annealing for function optimization. Minimizing multimodal functions of continuous variables with the simulated annealing algorithm. An extended continuous Newton method. Global Optimization of Nonconvex Nonolinear Programs Using Parallel Branch And Bound. A Collection of Test Problems for Constrained Global Optimization Algorithms Recent Advances in GLobal Optimization. An introduction to simulated evolutionary optimization. Global optimization using interval analysis. Global optimization: Deterministic approaches. Adaptive Simulated Annealing (ASA). An adaptive simulated annealing algorithm for global optimization over continuous variables. Random tunneling by means of acceptance-rejection sampling for global optimization Genetic Algorithms Bayesian Approach to Global Optimization. Application of bayesian approach to numerical methods of global and stochastic optimization. Complexity in numerical optimization. Constrained Global Optimization: Algorithms and Applica- tions Pure adaptive search in Monte Carlo optimiza- tion A combined multistart-annealing algorithm for continuous global optimization Simulated annealing for constrained global optimization. On the convergence of the Baba and Dorea random optimization methods. Stochastic techniques for global optimization: A survey on recent advances. Global Search Methods for Solving Nonlinear Optimization Problems. Global optimization for neural network training. A global search method for optimizing nonlinear systems. Topographical global optimization. A Monte Carlo simulated annealing approach to optimization over continuous variables. Global optimization of QMF filter-bank design using NOVEL Improving the performance of weighted Lagrange- multiple methods for constrained nonlinear optimization Pure adaptive search in global optimization. Improving hit-and-run for global optimization --TR --CTR Global Optimization Algorithm using Lagrangian Underestimates and the Interval Newton Method, Journal of Global Optimization, v.24 n.3, p.349-370, November 2002
trace-based search method;adaptive weights;lagrange-multiplier method;oscillations;convergence speed;local search;nonlinear continuous constrained optimization;inequality constraints;global search
596250
A Geometric Approach to Global Optimization.
In this paper we consider the problem of optimizing a piecewise-linear objective function over a non-convex domain. In particular we do not allow the solution to lie in the interior of a prespecified region R. We discuss the geometrical properties of this problems and present algorithms based on combinatorial arguments. In addition we show how we can construct quite complicated shaped sets R while maintaining the combinatorial properties.
Introduction The solution of piecewise linear programs has always played a special role in optimization. In general, piecewise-linearities are often employed to give a more realistic description of costs than can be achieved by linear terms alone. Also in worst case analysis piecewise linearity arises naturally. Furthermore, piecewise-linear programs may be used for approximating convex objective functions by piecewise linear functions. Such an approximation method including error bound analysis was proposed by [BHR91]. In addition, piecewise linear programs arise in specific branches of operations research, like for example location theory, where - given a number of existing facilities - a new facility has to be located so that some objectives are optimized. The objectives are often in the form of piecewise linear functions due to the fact that typically the sum (or the maximum) of weighted distances from the existing facilities to the new facility is chosen as a criterion. Moreover, these distances are often derived from norms with a polyhedral unit ball. See [LMW88], [FLFMW92], and [Nic95] for various piecewise linear models in location theory. For an overview about methods for solving piecewise linear programs the reader is referred to [Fou85], [Fou88], [Fou92], and references therein. As a conclusion it is quite natural also to look at global optimization problems with piecewise-linearities. In particular, we are interested in the problem of optimizing a piecewise-linear objective function over a non-convex domain. These includes problems with reverse convex constraints. For more details and an introduction to global optimization the reader is referred to [HT96] and references therein. Now we will formally introducing the model we consider in this paper. 1.1 The model Consider the following piecewise linear optimization problem in and the f are K different affine linear functions. This is a well-known convex problem and can efficiently be solved by linear programming methods. We can also easily introduce a feasibility region A, where A is a polyhedral set with A := fx 2 Also (OL) with the additional restriction that x 2 A can be formulated as linear program and therefore be solved efficiently. min z If, however, the feasible region A cannot be described by linear inequalities, or is not convex, we are in the area of global optimization. In particular we consider in this paper A := cl(IR n n R), with R ' IR n . By cl(S) we denote the closure of a set S. Now the minimization problem (ROL) reads min f(x) or min where int(R) denotes the interior of R. With the help of geometric properties of this problem we will be able to give algorithms to solve (ROL) for a large class of possible sets R. The remainder of the paper is organized as follows: First we state geometrical properties of level curves and level sets. These results are used to derive a combinatorial description of the solution set in IR n for (ROL) and convex sets R. Section 4 shows how these results can be extended in the plane for more general sets R while maintaining the combinatorial character of the solution. The paper ends with some conclusions and an outview to further research. 2 The concept of cells and level sets For any set R ' IR n let ext(R) denote the extreme points of R and @R the boundary of R. Furthermore, let Kg. Then f(x) can be written as is an affine linear function and the f k are pairwise different. To denote the solution sets we introduce R := min and R g: By defining cells else we get a subdivision of IR n into not more than K nonempty cells. The set of all cells is denoted C. Some properties of C are summarized in the following lemma. Lemma 1 For the cells C i , i 2 K the following holds: 1. For i 2 K we have that either C i is a polyhedral set with dimension n or 2. ig. 3. S 4. Proof: ad 1.) From the definition of C i it follows that C i is either empty or a polyhedral set which is full-dimensional, because in that case int(C i ) 6= ;. ad 2.) Let x 2 int(C i ) and suppose there exists k 6= i such that f i As x is in the interior of C i there exists a ball U := U(x) around x such that U(x) ' int(C i ). As f i 6= f k the hyperplane H separates U into two parts U H such that for all y H . This is a contradiction since U ' C i . ad 3.) Take an x 2 IR n . We then have an index i such that f g. If i is unique, x 2 int(C i ) and we are done. If i is not unique we define I := fi Then there exists a ball U := U(x) around x such that Now define for all k; l; k 6= l; k; l 2 I: g kl := fy G as the union of all g kl . As the g kl are hyperplanes, U n G 6= ;. Take G. Then there exists a unique j 2 I such that meaning that the cell C j 6= ;. As also f j (x) - f k (x) for all k 2 K we can conclude that x 2 C j . ad 4.) This statement follows directly from 2. q.e.d. As the lemma shows, we need only to investigate cells C and we can neglect the empty ones. In the following we therefore will assume that all cells are non-empty. The concept of cells has frequently been used in optimization (see [RS72], [Ede87], [HN95], [Nic95] and references therein). From this literature the following result is well-known, but restated here for the specific context of this paper. Theorem 1 X is either a whole cell or an r-dimensional facet of a cell, r 2 Proof: We can solve (OL) by inspecting all cells. For any cell C i the optimization problem is a linear program of which the feasible set is the polyhedral set C i . Therefore the result follows from the well known properties of linear programming theory. q.e.d. The following corollary says, that, for finding one optimal solution it is enough to look at the extreme points of the cells. Corollary 1 There always exists a point x 2 X with x is the extreme point of a cell C i for an index i 2 K. We define C @ as the set of all facets of all cells C 2 C. Additionally, we define construction lines for all j. The set of all construction lines, denoted by H, constitutes an arrangement of hyperplanes which define a set of cells CH similar to the ones introduced by [Ede87]. Note that CH is a subpartition of C, i.e. denote the set of points x 2 IR n belonging to a facet of a cell or a construction line, respectively). The name "lines" refers to the 2-dimensional case where we in fact have 1- dimensional construction lines. In the n-dimensional case the construction lines are (n \Gamma 1)-dimensional sets. Example 1 We are given In Figure 1 the two cell partitions C and CH are shown. We can easily compute X checking the extreme points of all cells C 2 C. Figure 1: Illustration for Example 1. The bold part shows C @ . The normal and the bold lines together constitute H. One more definition is necessary for that paper. For t 2 IR define the level set L- (t) and the level curve L= (t) as Note that for any convex function f , L- (t) is a closed, convex set, which is nonempty for all t - t . Lemma 2 For t ? t we have that @L- Proof: For t - t the result is obvious. For t ? t we know that any convex function f with minimal value t is strict convex on the set fx 2 of all non-minimal solutions. That proves the lemma. q.e.d. Using level curves and level sets we can reformulate (OL) and (ROL). Theorem 2 a) t is the optimal objective value of (OL) R is the optimal objective value of (ROL) c) In a) and b) L= (t) can be replaced by L- (t) d) x is an optimal solution of (ROL) with R if and only if there exists a IR, such that and The proof follows easily from the definition of level curves, level sets, and Lemma 2. Lemma 3 . The level curves L= (t) are piecewise linear on C. More specific, has one of the following shapes: 1. empty 2. the whole cell C i , in this case we have that 3. an r-dimensional facet of C 4. the intersection between C i and a hyperplane H with it follows that t ? t Note that only in Case 2 and Case 4 we can say something about the optimality of t. Proof: Case 1: f i const 8x 2 IR n . Then H can either be empty (yielding the first case) or yielding that L= t. In the latter case we also have that for all x 2 C i t. Since f is a convex function, it only can be constant on a full-dimensional set if it is minimal on that set. Therefore that we can conclude and the whole cell C i is optimal. Case 2: f i (x) is not a constant function. Then H is a hyperplane in IR n and ffl If we have that H " C i ' @C i then H is a supporting hyperplane for since C i is a polyhedral set it follows that H " C i is a facet of , we can conclude that int(C i ;. In this case let C i be the two parts of the cell C i which are separated by the hyperplane H. That means, one of C i is completely contained in L- (t). Let C t . Consequently, C which is a contradiction, because C Note that as C i is convex we have that Proof: As L= there are three possibilities according to Lemma 3. ffl If L= consequently, ffl If ffl If there exists a hyperplane H with C i then suppose L= which is a contradiction to Theorem 1. q.e.d. The following theorem is a consequence of the convexity of f (see also [HT96] and [HN94]). Theorem 3 If X ' int(R) then we have X Proof: Let x 2 X and assume that there exists y 2 IR n such that y 2 X Then y 62 R, yielding that f(x) ! f(y) and therefore we know for all - 2 (0; 1] that This also holds if we choose - such that -x which proves the theorem. q.e.d. 3 Solution for convex forbidden sets in IR n In this section we look at forbidden sets R which are convex. As we will show in the following theorem, for these sets there always exists an optimal solution for the restricted problem, which is in the intersection of the boundary of R with the the boundary of a cell C 2 C: Theorem 4 Let R be a convex set, R ' IR n and X ' int(R). Then there exists x R and c 2 C @ with x Furthermore only intersections with dim(c " have to be investigated. Proof: Let x R such that x 62 C @ . Let we know that t ? t , and consequently (see Lemma is a hyperplane with L= Now we look at the following two cases. Case 1: Note that since x 62 C @ we know that x 2 Therefore we can conclude that H a hyperplane, and x 2 @R " H i . According to Corollary 2 there exists @R such that z 2 X Case 2: That means cannot be optimal for the restricted problem since Theorem 3 tells us that X Now suppose we have dim(c " This implies that c coincides locally with @R and among c the objective function is linear. Assume that there is no c every level curve touching c " @R has to be linear throughout c " @R and therefore L= 6' R, contradicting the optimality condition. It follows that for a x be optimal there must exist a c and the result follows by replacing c by c 0 . q.e.d. Theorem 5 Let R ' IR n be a convex polyhedron, and X ' int(R). The set of facets of R is denoted by F . Then there exists always an optimal solution x R in the finite set of points Cand Proof: From Theorem 4 we know that for an optimal x we have x for some H 2 H. Choose a facet F 2 F such that x 2 choose an index i such that also x 2 @C i holds. Now we have x 2 Therefore means f is linear in F " Now we can conclude from the theory of linear programming that there always exists an optimal vertex v of the polyhedron F " know that v 2 H " @R and additionally The next theorem gives a characterization of the optimal solution for all convex restricted sets. Note that in IR 2 this result is equivalent to Theorem 4. Theorem 6 Let R ' IR n be convex , and X ' int(R). Then there exists always an optimal solution x R such that x R is a zero-dimensional intersection between the boundary @R and a sufficient number of construction lines H 2 H. Proof: From Theorem 4 we know that there always exists an optimal solution x R such that x only consists of one single point, we are done. If not, we consider the following problem (ROL1) on the Note that the optimal solutions of (ROL1) are optimal for (ROL). As the restriction of f to the linear subspace H 1 remains piecewise linear and convex, R " is a convex set and are the new construction lines we can apply Theorem 4 again. We conclude that there always exists an optimal solution x R to (ROL1) (and therefore to (ROL)) such that x \Lambda1 according to the definition of H 1 . Then we know that We repeat this argument until we have a number z of hyperplanes H H such that the set only consists of one single point, which then is an optimal solution to (ROL). q.e.d. The following corollary shows how Theorems 4 and 6 can be used to derive an efficient algorithm for solving (ROL) in the plane. We use the fact that since H is easier to compute. Corollary 3 Let be a convex set such that X ' int(R). Then there exists always an optimal solution x R in the finite set of points In the algorithm we first compute the set Cand and then look for the best candidate Algorithm for solving (ROL) in the plane. R 1. Compute 2. Compute 3. For all 4. Determine t )g. 5. Output: X The complexity of the above algorithm is dominated by the complexity of Step 3 which is Int O(K 2 ) (where Int is the complexity of computing an intersection point between a line and the boundary of the forbidden set R) and the complexity of Step 4 being O(K 3 ). Overall we have a complexity of O(K 3 Example 2 We use the same objective function f as in Example 1. In addition we are given a convex set R = [\Gamma10; 14] \Theta [2; 20]. By checking the set of candidates Cand we get X objective value 38(see Figure 2). In Figure 3 the reduced candidate set based on Theorem 4 is shown. R Figure 2: Illustration for Example 2. The bold points show the set of candidates H " @R. When applying Theorems 4 and 5 for higher dimensions to derive efficient algorithms a pure enumeration of the set of Candidates is not appropriate. Therefore combined enumeration and search procedures are suggested. Instead of solv- ing R Figure 3: Illustration for Example 2. The bold points show the set of candidates @C " @R. for all facets F 2 F it is also possible to restrict the search procedure to linear subspaces of lower dimensions and solve for all and all facets F 2 F Which of those alternative solution approaches is more appropriate is dependent on the input data. 4 Extensions in the plane In this section we consider more realistic forbidden sets. At first we examine what happens, if the forbidden set is not connected, and then extend the results to so-called bumpy sets. The algorithm of the previous section can easily be modified to accommodate the case where R is the union of pairwise disjoint, convex sets R 1 In this situation we first solve the unrestricted problem (OL). Then the following result is an immediate consequence of Theorem 3. Theorem 7 Let and R i are convex sets for L. Then there exists an optimal solution x R such that either x R 62 int(R) or there exists some l with X R ' @R l . Proof: If X ' int(R) Then the convexity of X and the assumptions on the sets imply that there exists some l such that X ' R l . Hence we can replace R by R l and use Theorem 3 to conclude that X R ' @R l . q.e.d. Our results can be extended to other cases of nonconvex restricting sets R. If we review the proof of Theorem 4 it becomes apparent that we can apply a combinatorial algorithm as in the case of convex sets whenever we have only a finite set of points on @R which can lie in the (relative) interior of linear pieces of a level curve. Therefore we introduce the concept of bumpy sets, which was first mentioned in [HN95] in the area of location theory. with its set of roots Roots(R) is any set which can be constructed by a finite application of the following rules: 1. Any convex set R is a bumpy set. In this case 2. If R 0 is a bumpy set then R is a one-dimensional, connected, proper curve segment (i.e. there exists a bijective mapping OE : [0; In this case R 0 is called the base of R and R 1 the bump. The two disjoint endpoints of s are the new roots of R, i.e. Note that dependent on the construction of the bumpy set, the same bumpy set can have a different set of roots, i.e. Roots(R) is not uniquely defined. For the following algorithm, however, any set of roots leads to an optimal solution. For the running time, the smallest set of roots is preferable. In Figure 4 and Figure 5 some examples of bumpy sets are shown, and the following lemmas describe some classes of bumpy sets. Some more applications of bumpy sets will be shown at the end of this section. Figure 4: An example for a bumpy set. The dots mark the set of roots. Lemma 4 Let R be a strict convex set and R 0 a translate of R, such that int(R " set. Proof: Choose R as base and R 0 as bump. ffl R is convex and therefore a bumpy set, line segment as @R " @R 0 consists of only two points (see, e.g. [Kle97]). q.e.d. is a bumpy set. Proof: If R is convex it trivially is a bumpy set. If not we define the convex set R 0 as the base of the bump. Let be the connected components of As R is not convex, we know that R 0 consists of L connected, one-dimensional, proper curve segments. That means we Figure 5: Another example for a bumpy set. The dots mark the set of roots. can iteratively add all bumps BL to R 0 , as we know that conv(B l ) ' As mentioned, we now will extend the algorithm of the previous section to the case, where R is a bumpy set. In the following we therefore prove that to find an optimal solution to (ROL), it again is enough to evaluate a finite candidate set. In order to do this one more result about the structure of bumpy sets is needed. Lemma 6 Let R be a bumpy set and s be a line segment. If s ' R and int(s) touches @R at a unique point x then x 2 Roots(R). Proof: Induction over the number of bumps. For a convex set R and a line segment s ' R we know that the interior of s cannot touch @R from inside at a unique point. Now take any bumpy set R with base R 0 and bump R 1 . Let s be a line segment with its interior touching @R from inside at a unique point, i.e., s ' R and fxg. Now we distinguish three cases: ffl If x 62 @R 1 , we know that x 2 Roots(R 0 due to the induction hypothesis. by definition x 2 @R is one of the (new) roots of R. and therefore we again have a linear piece touching a convex set from inside at a unique point, which is a contradiction. q.e.d. The following theorem shows that in order to solve (ROL) with R is a bumpy set it suffices to examine the candidate set of the last section, i.e. the zero-dimensional intersections between the boundary of R and the construction lines additionally, the roots of the bumpy set R. Theorem 8 Let R ' IR 2 be a bumpy set such that X ' int(R). Then there always exists an optimal solution x R in the finite set of points Cand Proof: With Theorem 2 we know that t R is the optimal value of (ROL) if and only if L- (t We also know that L= (t is the set of optimal solutions. Now take any x There exists an open set U 2 IR 2 such that we have a unique point x fx As the level curve L= (t R ) consists of linear pieces (see Lemma 3) we either have ffl that x R is the endpoint of such a linear piece, in this case x for some H 2 H (see Theorem 4) or R is in the interior of such a linear piece, then, according to Lemma 6, x R is part of a linear piece s ' L= (t we know that the endpoints x of s are also in @R and clearly f(x 1 R . At both endpoints, on the other hand, other construction lines intersect (as the level curves are piecewise linear on the cells, see Lemma 3), such that we have x Algorithm for solving (ROL) in the plane with R is a bumpy set. R 1. Compute 2. Compute 3a. For all 3b. 4. Determine t )g. 5. Output: X Example 3 We use the same objective function f and forbidden region R as in Example 2. In addition we are given a convex set R 2] as a bump. By checking the extended set of candidates Cand we get the new candidates (1; 2) , loose the two former candidates ( 16; 2) and (2; 2). By computing the objective function for the candidates we get X with objective value 40(see Figure 6). Finally we give two examples for the application of bumpy sets. The first application shows that we can solve the restricted problem (ROL) for any simple polygon. These simple polygons can then be used for approximating more complex sets R (see [BHR91], [Rot92] and [KN96] for more details). Lemma 7 Any simple polygon is a bumpy set. Proof: Induction over the number of vertices n of the polygon. For a convex triangle which trivially is a bumpy set. Now take any polygon R with more than 3 vertices. Consider a triangulation of R into (such a triangulation always exists, see e.g. [Kle97], [O'R93]) and take any triangle R 1 of that triangulation such that R 1 has two edges on the boundary of R. Define R 1 as bump and R 0 := R n R 1 as base. Then the triangulation of R 0 consists of triangles, such that the number of vertices of R 0 is is a bumpy set, R 1 is convex and proper line segment, since it is the third edge of the triangle R 1 . q.e.d. If R is a polygon then the set of roots in the above construction is contained in the set of vertices of the polygon, i.e. vertex of Rg; R R Figure Illustration for Example 3. The bold points show the set of candidates H " @R plus the roots of the bump. such that the set of candidates in Step 3b of the algorithm would be Cand vertex of Rg: This can be sharpened, as the following lemma shows. Lemma 8 Let R be a polygon and let X ' int(R). Then there exists always an optimal solution x R in the finite set of points Cand Proof: This lemma can be shown by a special bumpy set construction using only reflexive vertices as roots. Another possibility is to prove the result along the lines of Theorem 8. q.e.d. Another application is the following. Suppose we want to have m solutions \Lambdam to the problem (OL), but have the restriction that these solutions are not too similar,i.e. for a given number r 2 IR and a convex distance measure d. This concept has been used by [BW96] in the context of location theory. Then we can proceed as follows: First we find the best solution x \Lambda1 to our problem. Then we forbid all solutions which are too similar to x \Lambda1 , i.e. R rg. Solving that problem we find a solution x Iterating that procedure, to find x \Lambdak we solve problem (ROL) with Then minimal solution to the above problem. This problem can be solved with our algorithm, if all the sets R k are bumpy sets. If m - 3,x Lemma 4 shows that this can be done for any strict convex distance function d in IR 2 . If we also want to have distance functions with linear pieces, e.g. l 1 or l 1 distances or any kind of gauges we can use Lemma 5 can be shown, that the sets R k are bumpy sets, if they do not have any wholes. Even if they have wholes, Lemma 6 remains true, such that the algorithm can be adopted to solve the problem. Conclusions In this paper we have shown how global optimization problems can be solved using geometrical methods and combinatorial arguments. These discretization approach proved to be very successful if the original assumptions are weakened, like in the case of bumpy sets. Moreover, this approach yields numerically stable and efficient procedures and allows us to easily compute the whole set of optimal solu- tions. Of course, this approach is also useful for general convex objective functions using approximation methods, like the ones described in [BHR91]. Therefore our future work will include the adaption of such approximation methods to (ROL). Also computational tests and bounding techniques are under consideration. --R Sandwich approximation of univariate convex functions with an application to separable convex programming. The rectilinear distance minisum problem with minimum distance constraints. Algorithms in Combinatorial Geometry. A simplex algorithm for piecewise-linear programming I: derivation and proof A simplex algorithm for piecewise-linear programming II: finiteness A simplex algorithm for piecewise-linear programming III: computational analysis and applications Combinatorial algorithms for some 1-facility median problems in the plane Restricted planar location problems and applications. Global Optimization. Algorithmische Geometrie. bounds for the approximative solution of restricted planar location problems. Facilities Location: Models Discretization of Planar Location Problems. The convergence rate of the sandwich algorithm for approximating convex functions. Introduction to piecewise-linear topology --TR
discretization;piecewise linear programs;reverse convex constraints;geometric approach
596347
Solving the Sum-of-Ratios Problem by an Interior-Point Method.
We consider the problem of minimizing the sum of a convex function and of p1 fractions subject to convex constraints. The numerators of the fractions are positive convex functions, and the denominators are positive concave functions. Thus, each fraction is quasi-convex. We give a brief discussion of the problem and prove that in spite of its special structure, the problem is \cN\cP-complete even when only p=1 fraction is involved. We then show how the problem can be reduced to the minimization of a function of p variables where the function values are given by the solution of certain convex subproblems. Based on this reduction, we propose an algorithm for computing the global minimum of the problem by means of an interior-point method for convex programs.
Introduction Nonlinear programming problems often involve objective functions that can be expressed in terms of one or several ratios. Exploiting the special structure of such fractional programs has been the subject of extensive studies in the last few decades. For an overview of fractional program- ming, we refer the reader to (Schaible, 1995) and the references given therein. Fractional programs with only a single ratio or a maximum of finitely many ratios are fairly well understood. Under suitable conditions, these problems still satisfy some form of generalized convexity, which can be exploited in algorithms for the numerical solution of such problems. For example, there are polynomial-time interior-point methods for classes of such problems; see (Freund and Jarre, 1994, 1995; Nemirovskii, 1996). y Numerical Analysis Manuscript No. 99-3-13, Bell Laboratories, Murray Hill, New Jersey, June 1999. Available on WWW at http://cm.bell-labs.com/cs/doc/99. Roland W. Freund and Florian Jarre On the other hand, fractional programs with sums of ratios are much more difficult and not as well understood; see (Schaible, 1995, 1996). Such problems possess some form of generalized convexity only in special cases, such as the ones discussed in (Schaible, 1984; Hirche, 1985), and in general, they have multiple maxima and minima. Algorithms for classes of sum-of-ratios problems are described in (Cambini et al., 1989; Chen et al., 1998; Falk and Palocsay, 1992; Konno and Kuno, 1990; Konno and Yamashita, 1998; Ritter, 1967) and in the review article (Schaible, 1996). However, most of these algorithms are for the optimization of linear ratios subject to linear constraints. The purpose of this paper is to present a suitable interior-point approach for the solution of much more general problems with convex-concave ratios and convex constraints. Our approach is based on approximating the sum-of-ratios problem by a sequence of convex minimization problems. For such convex problems, interior-point methods have become the methods of choice, both from the point of view of theoretical complexity and of practical efficiency. By using a simple warm-start strategy, the cost for solving the individual convex subproblems can be reduced to very few iterations. Finally, the interior-point method provides certain dual information needed for the overall approach. More precisely, we consider the problem of minimizing or maximizing the sum of a single function and of p - 1 ratios subject to convex constraints, and we explore the use of interior-point methods for the solution of such problems. More precisely, we study problems of the subject to x 2 and subject to x 2 S: (2) Here and in the sequel, we make the following assumptions. ASSUMPTION 1. S ae IR n is a compact convex set such that f j (x) - 0 and S. For the minimization problem (1), the functions h and f 1 are convex and the functions are concave. For the maximization problem (2), the functions h and f 1 are concave and the functions are convex. For simplicity, from now on we restrict ourselves to minimization problems (1). The results and algorithms for (1) in this paper can easily Solving the Sum-of-Ratios Problem 3 be converted to maximization problems (2) by simply exchanging "min" and "max", "convex" and "concave", and "-" and "-". In Section 2, we first discuss the simplest case, namely the sum of a convex function and only We show that this problem is NP-complete and propose a method for finding the global minimizer. In Section 3, the method is generalized to the case p - 2. In Section 4, we report results of numerical experiments. In Section 5, we make some concluding remarks. 2. Sum of one fraction and a convex function Throughout this section, we assume that 1. In this case, problem (1) reduces to the form subject to x 2 S: (3) Here, f; are functions that satisfy the conditions specified in Assumption 1, i.e., f and h are convex, g is concave, and f(x) - 0 S. For any fixed r ? 0, let ae r oe and r g, the feasible set in (4) is empty, and in this case, we set q(r) := 1. Note that x(r) is not necessarily unique, but, of course, q(r) is. From the definition of q, it is obvious that x(r ) solves (3) if, and only if, r minimizes q. Thus, problem (3) is reduced to the one-dimensional problem of minimizing the function q. Determining x(r) for a given value r ? 0 is a convex optimization problem, which can be solved by several methods. If a separation oracle for S (ae IR n ) is given, the evaluation of q for a given value of r can be done (up to a given precision) by the ellipsoid method. Here, by "separation oracle", we mean a subprogram that accepts as input any vector x 2 IR n and produces as output either the information "x 2 S", or a vector h 2 IR n , h 6= 0, with h T y - h T x for all y 2 S. In the second case, the vector h defines a hyperplane that "separates" x from S. If self-concordant barrier functions for the sets ae r - and g(x) - r oe 4 Roland W. Freund and Florian Jarre for real numbers - are known, then q(r) can also be evaluated by an interior-point method. Here, a barrier function for a convex set C is a function that is convex and finite in the interior of C, and goes to infinity as x approaches the boundary of C. The notion of self- concordance was first introduced in (Nesterov and Nemirovskii, 1994). Roughly speaking, self-concordance is defined as a local Lipschitz condition of the Hessian of the barrier function. As shown in (Nesterov and Nemirovskii, 1994), many convex sets possess easily computable self- concordant barrier functions, and the concept of interior-point methods based on self-concordance is a very general approach. We remark that for the special case of a constant function h and problem (1) can be reduced to a problem of the form (3), i.e., with only one ratio, by means of the Charnes-Cooper transformation (Charnes and Cooper, 1962); see, e.g., (Cambini et al., 1989). A self-concordant barrier function for the conic hull introduced by this transformation is discussed in (Freund et al., 1996). In general, when is constant, the Charnes-Cooper transformation can be used to reduce problem (1) to a sum-of-ratios problem with This simple reduction may be crucial for algorithms whose computational costs grow rapidly with the number of ratios. For example, given a sum-of-ratios problem with will be more efficient to first employ the Charnes-Cooper transformation and then apply the algorithm of the present paper to the reformulation with than using the same algorithm for the solution of the original problem with 2. 2.1. Properties of the function q Next, we recall some well-known properties of the function q given by (4) and (5). oet, and by the convexity of f , we have s ae f(x(r)) r oe Similarly, the convexity of h implies that By the concavity of g, it also follows that g(x) - s. Hence, x is feasible for (4), and s Solving the Sum-of-Ratios Problem 5 In spite of (6), (7), and (8), the function q is not quasi-convex, i.e., in general it may happen that Note that if q were quasi-convex, problem (3) could be solved in polynomial time by using a golden-mean search for q. In view of the above derivation, we may still ask ourselves whether the function q may "smooth out" some of the local minimizers of (3), and whether minimizing q might be easier than solving problem (3) directly (assuming that we can evaluate q and its derivatives). The observation that the function q is not necessarily simpler than (3) is illustrated in Figure 1 below, which depicts the function q for a special case where S is just a real interval. This plot shows that q may exhibit a very "irregular" behavior. 2.2. NP-completeness Next, we prove that problem (3) is "essentially" NP-complete. To this end, we show that a well-known NP-complete problem, namely the following knapsack problem, can be recast as a special instance of problem (3). Knapsack problem: Let an integer d ? 1, weights and costs c g. The problem is to find a subset I 0 ae I such that maximized subject to the constraint For a discussion of the knapsack problem and a proof of its NP- completeness, we refer the reader to (Garey and Johnson, 1979). Our result on the NP-completeness of problem (3) can now be stated as follows. THEOREM 2. Problem (3) is NP-complete in the following sense. Let the data of a knapsack problem with d 2 IN weights be given. There exists a convex, piecewise linear function f , a linear function g, and a linear function h defined on the interval such that f , g, h, and their respective derivatives can be evaluated in polynomial time, f , g, and h take values of polynomial size, and solving problem (3) is equivalent to solving the given knapsack problem. REMARK 3. The right endpoint 2 d of the interval S in Theorem 2 is not polynomial. At first sight, this might lead to the impression that the reduction of a knapsack problem to problem (3) is exponential. This, 6 Roland W. Freund and Florian Jarre of course, is not the case. Indeed, just as in the case of linear pro- grams, which may also involve non-polynomial upper or lower bounds, one only needs polynomiality in the coding length of the problem. The coding length of problem (3) is at least d, and hence the coding length of the endpoint 2 d is in fact polynomial in the coding length of the problem. Finally, note that if the function f=g were convex, then an ffl-approximation to problem (3) could be computed in polynomial time. The NP-completeness in Theorem 2 does not result from the size of the endpoint 2 d of S, but from the lack of convexity. Proof of Theorem 2. Let d 2 IN , weights w be the given data of a knapsack problem. From this data, we now construct a special problem of the form (3) that is "equivalent" to the knapsack problem. To this end, we first enumerate the 2 d subsets of I by simply counting from 1 to 2 d in the binary system. Then, for each subset I 0 of I, there exists an index 1 - k - 2 d such that I is the k-th subset of the enumeration. We can determine I k just by knowing its index k, and without looking at any other subset. We can also determine the weight w i of the k-th subset just by knowing the index k. For where c := . Solving the knapsack problem is then equivalent to finding min 1-k-2 d For later use, we note that Next, we set S := [1; 2 d ] and define functions f; The functions g and h are the linear functions given by The function f is defined as the piecewise linear interpolant through the points . Hence, on each interval k - x Solving the Sum-of-Ratios Problem 7 Using (10), one readily verifies that the function f is convex on S. Clearly, given any x 2 S, it is possible to evaluate f(x) in O(d) arithmetic operations, and the number of digits needed to represent the function values are at most 2d plus the number of digits needed to evaluate Finally, we show that for the set S and the functions f , g, and h just defined, the minimizer of (3) is the index k of a k-th subset I k ae I that solves the knapsack problem. Let 1 and consider the objective function of (3) for By (10)-(12), the second derivative of the objective function satisfies for all x 2 This shows that the objective function of (3) is concave on [k; k + 1], and thus its minimum over [k; k + 1] is attained at 1. By (11) and (12), the corresponding function values are 2 d for 2 d for Therefore, problem (3) is equivalent to (9), which in turn is equivalent to solving the knapsack problem. For the special instance of problem (3) constructed in the proof of Theorem 2, the evaluation of the associated function (5), q(r), is particularly simple. Indeed, let x - r and x 2 d dx r r r r r 0: This shows that the objective function in (4) is monotonically increasing Therefore, the minimum in (4) is attained for and the function q(r) in (5) is identical to the objective function of (3). In Figure 1, we plot the function q for the case of the knapsack problem with Figure 1 displays an example where minimizing q is identical to solving problem (3). In general, however, we may anticipate that the 8 Roland W. Freund and Florian Jarre Value of r Value of Figure 1. The objective function with random w i 's and c i 's. structure of the higher-dimensional problem (3) is far more complicated than the scalar function q. We propose an approach for solving problem (3) by evaluating q for various values of r and exploiting Lipschitz properties of q. We emphasize that in the case where the function q has very many local minimizers of approximately the same magnitude (as in the class of problems constructed in the proof of Theorem 2), any approach for solving problem (1) will necessarily be very slow (unless 2.3. A global minimization method If f; g, and h are smooth, due to the structure of S, the function q is generally a piecewise smooth function. To compute a global minimizer r of q, we construct a lower-bound function q(r) - q(r) and then minimize q. The function q depends on a partition, where we assume that r (1) - r and r (k) - r . Note that Solving the Sum-of-Ratios Problem 9 so that a value for r (1) can be obtained from a given lower bound of g on S, and a value for r (k) by solving the concave maximization problem in (13). Let some i and be given. Define a lower-bound number ae r (i+1) oe so that q i - q(r) for r 2 [r (i) ; r (i+1) ]. Note that evaluating the right-hand side of (14) amounts to solving a convex optimization problem. Let x i be a solution of the minimization problem in (14). It follows that Using these two inequalities, for all r 2 [r (i) ; r (i+1) ], we get r r r (i+1) r r (i) r r r (i) r (i) r r (i) where r (i) Note that the inequality (15) follows from r (i) =r - 1 and q i \Gammaq(r (i) The bound (16) proves left-sided Lipschitz continuity of q. Indeed, near the above bound is close to the value q(r (i) ). However, for , the bound reduces to q i , which is lower than the value q(r (i+1) ). Roland W. Freund and Florian Jarre Note that a bound of the form (15) with q(r (i+1) ) in place of q(r (i) ) is not possible. It may occur that q(r (i+1) ) AE q(r (i) ). Intuitively, this will happen when no longer contains points for which f or h are reasonably small but does. In this case, the Lipschitz constant for q from the right may be much larger than the one from the left. To determine a suitable Lipschitz property from the right, we define the function ~ ae r (i+1) oe Observe that ~ q i is convex (in r). Moreover, ~ q i satisfies ~ for r - r (i+1) , and ~ We remark that when evaluating q(r (i+1) ), the problem (17) with is solved, and the Lagrange multiplier-denoted by - g in the sequel-corresponding to the constraint g(x) - r (i+1) can also be computed. Indeed, interior-point methods can be implemented so that such a multiplier is obtained at no extra cost. The Lagrange multiplier leads to the bound for r - r (i+1) ; see, e.g., Theorem VII.3.3.2 in (Hiriart-Urruty and Lemarechal, 1993). The lower-bound function q(r) is then defined for as the maximum of the bounds (16) and (18), A simple method for solving problem (3) then proceeds as follows. Given k points a new point - r from the interval (r (i) ; r (i+1) ) that contains a minimizer of minf is chosen. Then, - r is inserted into the list (19) (thus k is increased by one), and the process is repeated. Note that the update of q(r) only involves the interval between r (i) and r (i+1) neighboring - r. This interval is split into two subintervals r] and [-r; r (i+1) ], and the minimum of q(r) is evaluated over both intervals. In particular, the effort for minimizing q merely consists of bookkeeping. Figure 2 gives an example of the bounds leading to q(r). Note that the slopes of q(r) and q(r) may be of opposite sign, so that in the interior of [r (i) ; r (i+1) ] the function q(r) may not be a good approximation to q(r). Hence, - may be a poor choice. A more reliable choice used in our numerical examples below is - r := 1 Solving the Sum-of-Ratios Problem 11 slope q 0 arg min q(r) ffl . Figure 2. The functions q and q. To keep the evaluation of q at moderate costs, it suffices to compute only approximations to q(-r) and ~ along with some error esti- mates. Interior-point methods are particularly suitable for computing an approximate solution - x(r) of the convex problem (4), along with a certified error bound of the form jq(r) ffl. The computation of - x(r) takes at most O(log(1=ffl)) iterations provided that a self-concordant barrier function for S and for the level sets of the functions f , g, and h is known. This observation is the key point for our proposed algorithm. Next, we present a statement of the algorithm. ALGORITHM 4. (Conceptual overall algorithm for INPUT. Functions f , g, h and a compact convex set S defining the single-ratio problem (3). A stopping tolerance ffl ? 0. Step 0. Determine r (1) and r (2) with Roland W. Freund and Florian Jarre If no such value r (1) exists: STOP, the problem violates Assumption 1. Otherwise, compute q(r (1) ), q(r (2) ), and the Lagrange multipliers for g. (number of "support points" r (\Delta) ). Step 1. Set - Step 2. Compute q(-r) along with the Lagrange multiplier - g for g. Step 3. Based on (18), evaluate arg min r2[r (i) ;-r] q(r) and arg min Step 4. Increase k by one, and insert - r into the list of r (\Delta) 's. Step 5. Find Step 6. If q(~r) - min 1-k q(r (') ~ r is an approximate minimizer. Otherwise, return to Step 1. We remark that, in Step 3 of Algorithm 4, the bound q(r) may either be obtained by setting q i \Gamma1, or by solving an additional problem of the form (14). The latter case is more expensive and results in a better bound for q(r) since both (16) and (18) are used. In most cases in (18) is "overly large"), it is more efficient to rely on (18) only, and not to solve (14). Note that the minimizers of the lower-bound function q(r) computed in Step 3 of Algorithm 4 can be stored in a heap, so that Step 5 merely consists of selecting the first element from this heap. Finally, we remark that, in practice, the feasible set S will usually be of the form ae oe are given convex functions. Solving the Sum-of-Ratios Problem 13 3. Minimizing the sum of several fractions In this section, we return to the general problem (1) of minimizing the sum of a convex function and p ratios. The basic idea for solving problem (1) is similar to the special case treated in Section 2. Assumption 1 be satisfied. In this case, is a vector of p parameters. In analogy to the definitions (4) and (5) of x(r) and q(r) in the case where S(r) := ae oe and Initially, we assume that vectors r (1) and r (2) are computed such that there is a minimizer r of q satisfying r (1) - r - r (2) . (As usual, the -sign is understood component wise.) Each component of r (1) and r (2) can be computed separately as in the case Now let r, r (i) , r (i+1) , and some direction \Deltar 2 IR p be given such that the relations are satisfied. The bounds (16) and (18) can be generalized to provide bounds for q(r \Deltar). We split \Deltar be the Lagrange multipliers for the constraints (20). By Theorem VII.3.3.2 in (Hiriart-Urruty and Lemarechal, 1993), a lower bound for q is given by To obtain a lower bound for q in direction \Deltar + , we define the value r (i+1) r (i) It then follows that 14 Roland W. Freund and Florian Jarre r (i) r (i+1) r (i+1) r (i) r (i+1) r (i+1) r (i+1) r (i) r (i) r (i+1) where r (i+1) r (i+1) r (i+1) r (i) r (i+1) Combining the above relations, we get r (i+1) This bound is analogous to the one for In (22), we may replace r (i) by r and r by (r to obtain the new bound with r (i+1) \Deltar r (i+1) r (i+1) \Deltar r (i+1) Solving the Sum-of-Ratios Problem 15 Based on this bound, we can define an anisotropic trust region about each point r, as long as some lower and some upper limits (like r (i) and r (i+1) in the previous derivation) are given. The union of the trust regions about all support points r (i) forms a Voronoi diagram in IR p , the vertices of which contain the candidates for the minimizer of the lower-bound function q(r). For a definition of Voronoi diagrams, their properties, and algorithms for their numerical computation, we refer the reader to (Aurenhammer, 1991; Fortune, 1997). As in the case these candidates for the minimizer of q(r) may not result in the best choice for inserting a new value - r somewhere between the known points r (i) . In addition, the computation of the vertices of the Voronoi diagram is complicated and expensive. We propose a simpler scheme based on bounds analogous to (18) where the lower-bound function q(r) is defined in the box ae oe with given vectors r (i) j . For r 2 B (i) , we obtain from (21) that Next, we summarize the resulting overall algorithm. ALGORITHM 5. (Conceptual overall algorithm for p - 2.) INPUT. h, and a compact convex set S defining the multi-ratio problem (1). A stopping tolerance ffl ? 0. Step 0. For all If r (1) j does not exist for some 1 - j - p: STOP, the problem violates Assumption 1. Otherwise, compute q(r (1) ), q(r (2) ), and the Lagrange multipliers j for the g j 's. (number of "support points" r (\Delta) ). containing arg min q(r)). Roland W. Freund and Florian Jarre Step 1. Set (the index where a split pays most) and define - r by r (i) Step 2. Compute q(-r) along with the Lagrange multiplier - j for each function g j . Step 3. Based on (21), evaluate arg min l q(r) and arg min l Step 4. Increase k by one, insert - r into the list of r (\Delta) 's splitting B (i) along the hyperplane r r l into two boxes (one for - r and one for r (i) ). Step 5. Find i - k such that ~ r := arg min r q(r) 2 B (i) . Step 6. If q(~r) - min 1-k q(r (') ~ r is an approximate minimizer. Otherwise, return to Step 1. 4. Numerical experiments Algorithm 4 for minimizing the sum of a convex function and of convex-concave fraction has been implemented in Matlab. For the solution of the convex subproblems, we use the interior-point method described in (Jarre and Saunders, 1995). As we have seen in Figure 1, the resulting problem (3) may be very complicated, and may have very many local minimizers. Nevertheless, we anticipate that the parameterization with respect to r will smooth out many of the local minimizers of (3) and thus result in a function q that is easier to minimize than the objective function of the original problem (3). In this section, we report numerical results of Algorithm 4 applied to certain examples with random data. In this case, the expectation that the function q is easier to minimize than the original problem (3) was fully met. In fact, the function q appeared to be unimodal with respect to r for these examples. Solving the Sum-of-Ratios Problem 17 Our test examples are minimization problems of the form (3), where ae oe Here, the matrices H;F;G and are constructed to be positive semidefinite. Therefore, in (23), the functions h and f are convex, the function g is concave, and the feasible set S is convex. The data for (23) is chosen randomly as follows. For each matrix D i , we first generated a random lower bidiagonal matrix L i the nonzero entries of which are uniformly distributed in [\Gamma1; 1], and then we computed i . This guarantees that each D i is a positive semidefinite tridiagonal matrix. Similarly, H, F , and G are constructed as random positive semidefinite tridiagonal matrices. In (23), h, f , g, and are vectors that were also generated randomly. Further- more, the scalars -, ', fl, and were chosen such that the interior of the feasible domain S is guaranteed to be nonempty, and such that the functions f and g are guaranteed not to have a zero in S. Finally, we have run experiments for problems (23) with values of n ranging from and values of m ranging from Note that the constraints in (23) are nonlinear, and therefore, adaptations of the simplex method for solving problem (3) with data (23) would be rather complicated. In Figure 3, we plot the function q for a typical example (23) with marks a point r at which the method has evaluated the function q in order to be able to guarantee that the final iterate is indeed an approximate global minimizer. Thus, each stands for the application of an interior-point method to solve a convex problem of the form (4). Since for each " " the interior-point method can be "warm-started" using as starting points some convex combination of almost final iterates of two neighboring problems, the overall number of interior-point iterations for each " " was less than eight in the average. The curve q(r) is of course not known, in general. (Here, it is plotted merely for illustration; its values were determined by solving a convex problem of the form (4) for some 200 evenly spaced values of r.) In Figure 4, we show a detailed enlargement of the points generated by Algorithm 4 near the global minimizer of a problem with 20. The plot shows that the distance between support points r on the right of the minimizer is much smaller than to the left, indicating that in this particular case, the Lipschitz bound (16) provides a much Roland W. Freund and Florian Jarre -5 Value of r Value of Exact function Points generated by algorithm Figure 3. The function q(r) for a random example, more accurate approximation to q(r) than (18). Thus, the algorithm did not evaluate a further refinement for the points on the left of the minimizer. In Table I, we report the number of iterations taken by our Matlab implementation to solve problem (23) with constraints and different dimensions n. The stopping criterion for these examples was chosen such that q(r final is guaranteed. Here, q(r opt ) is the unknown global optimum of (23). The numerical results are intended to provide a first rough estimate on the dependence of our algorithm on the dimension n of the space. We stress that the numbers of Newton steps and Hessian evaluations in Table I could be further reduced by a more sophisticated implementation. The number of Newton steps given in Table I refers to the sum of exact and inexact Newton steps. For inexact Newton steps, the Hessian matrix of a previous Newton step has been used in place of the current Hessian matrix. The overall computational effort is dominated by the number of Hessian evaluations. The number of r's refers to the number of support points r at which q(r) was evaluated. Solving the Sum-of-Ratios Problem 19 Value of r Value of Exact function Points generated by algorithm Figure 4. The function q(r) near the global minimum for a random example, Table I. Iteration numbers. # of r's 20 17 14 14 # of IP iterations 407 384 299 326 # of Hessians 644 584 478 548 # of Newton steps 1277 1134 980 1135 The random examples presented above exhibited only one local minimizer of the function q, as in Figure 3. We therefore constructed some small problems in three dimensions in such a way that there were several local minimizers at integer values of r. If two or more of the local minimizers have nearly the same value q(r), the method refines about both minimizers until the global minimizer has been identified. The plot in Figure 5 shows such a "worst-case" behavior where the method takes a large number of steps before identifying a point near as an ffl-global minimizer. If the stopping tolerance ffl is decreased further, then only the bounds near are refined to increase the accuracy of the global minimizer. Roland W. Freund and Florian Jarre 50.20.40.60.8Value of r Value of Exact function Points generated by algorithm Figure 5. The function q(r) for a case with several local minimia. It is needless to say that Algorithm 4 does not lend itself to solving the knapsack problem of Section 2.2. The structure of the knapsack problem is not exploited by Algorithm 4, and the Lipschitz bounds (16) and (18) are too weak to provide a sufficiently sharp lower estimate for the function q for an interval of length more than one. Hence Algorithm 4 is at least as expensive as enumerating all possible integer solutions. While the knapsack problem represents an example for which Algorithm 4 is not suitable, we believe that most applications have a structure more similar to the random problems above for which Algorithm 4 provides a reliable and reasonably fast method for identifying the global minimum. 5. Conclusions We considered the sum-of-ratios problem in IR n where the sum of a convex function and p convex-concave fractions is minimized subject to convex constraints. We proposed an approach to transform this problem to the problem of minimizing a suitably defined function q of variables. The function q can be evaluated by using an interior-point method for convex minimization. We established Lipschitz bounds for Solving the Sum-of-Ratios Problem 21 q that can also be evaluated numerically by using an interior-point method. Based on these bounds, a method was derived to find an ffl-approximation to the global minimizer of the sum-of-ratios problem. We presented numerical experiments with the proposed algorithm for the case of minimizing the sum of a convex function and of convex-concave fraction. An implementation of the algorithm for the case will be described elsewhere. Acknowledgements We would like to thank Siegfried Schaible for bringing the sum-of- ratio problem to our attention and for providing us with a copy of the technical report (Schaible, 1996). The authors are grateful to the referees and the editor for their constructive comments that helped us to improve the presentation of the paper. --R Voronoi diagrams-a survey of a fundamental geometric data structure On maximizing a sum of ratios Information and Optimization Sciences 10: Programming with linear fractional function- als Efficient algorithms and implementations for optimizing the sum of linear fractional func- tions Optimizing the sum of linear fractional func- tions Voronoi diagrams and Delaunay triangulations An interior-point method for fractional programs with convex constraints An interior-point method for multifractional programs with convex constraints On self-concordant barrier functions for conic hulls and fractional programming Computers and Intractability: A Guide to the Theory of NP-Completeness Convex Analysis and Minimization Algorithms I A practical interior-point method for convex programming Generalized linear multiplicative and fractional programming Minimization of the sum and the product of several linear fractional functions over a polytope On polynomiality of the method of analytic centers for fractional problems A parametric method for solving certain nonconcave maximization problems Simultaneous optimization of absolute and relative terms Fractional programming Fractional programming with sums of ratios --TR --CTR Takahito Kuno, A Revision of the Trapezoidal Branch-and-Bound Algorithm for Linear Sum-of-Ratios Problems, Journal of Global Optimization, v.33 n.2, p.215-234, October 2005 Yang Dai , Jianming Shi , Shouyang Wang, Conical Partition Algorithm for Maximizing the Sum of dc Ratios, Journal of Global Optimization, v.31 n.2, p.253-270, February 2005 Harold P. Benson, Using concave envelopes to globally solve the nonlinear sum of ratios problem, Journal of Global Optimization, v.22 n.1-4, p.343-364, January 2002 H. P. Benson, Global optimization algorithm for the nonlinearsum of ratios problem, Journal of Optimization Theory and Applications, v.112 n.1, p.1-29, January 2002
convex subproblem;interior-point method;sum of ratios;knapsack problem;cNcP-completeness;fractional programming;global optimum
596351
On Copositive Programming and Standard Quadratic Optimization Problems.
A standard quadratic problem consists of finding global maximizers of a quadratic form over the standard simplex. In this paper, the usual semidefinite programming relaxation is strengthened by replacing the cone of positive semidefinite matrices by the cone of completely positive matrices (the positive semidefinite matrices which allow a factorization FFT where F is some non-negative matrix). The dual of this cone is the cone of copositive matrices (i.e., those matrices which yield a non-negative quadratic form on the positive orthant). This conic formulation allows us to employ primal-dual affine-scaling directions. Furthermore, these approaches are combined with an evolutionary dynamics algorithm which generates primal-feasible paths along which the objective is monotonically improved until a local solution is reached. In particular, the primal-dual affine scaling directions are used to escape from local maxima encountered during the evolutionary dynamics phase.
Introduction A standard quadratic problem (standard QP) consists of nding global maximizers of a quadratic form over the standard simplex, i.e. we consider global optimization problems of the form subject to x 2 ; (1) Paper presented at the GO99 conference, Firenze, Sept. 28{Oct. 2, 1999. I.M.B. and M.D. thank M. Locatelli and F. Schoen for their hospitality and perfect organization. Large part of the work was done while T.T. was working at TU Delft. c 1999 Kluwer Academic Publishers. Printed in the Netherlands. I.M. Bomze, M. Dur, E. de Klerk, C. Roos, A. Quist, T. where A is an arbitrary symmetric n n matrix; a > denotes transposition; and is the standard simplex in the n-dimensional Euclidean space IR n , denotes the non-negative orthant in IR n (of course, the region fy 2 IR n can always be represented by IR n+1 , introducing a slack variable). To avoid trivial cases, we assume throughout the paper that the objective is not constant over , which means that fA; Eg are linearly independent where is the n n matrix consisting entirely of unit entries, so that x > . For a review on standard QPs and its applications, which also oers a justication for terminology see [9]. Note that the maximizers of (1) remain the same if A is replaced with A where is an arbitrary constant. So without loss of generality assume henceforth that all entries of A are non-negative. Furthermore, the question of nding maximizers of a general quadratic function x > Qx can be homogenized in a similar way by considering the rank-two update ce > in (1) which has the same objective values on . Of course, quadratic optimization problems like (1) are NP-hard [23]. Nevertheless, there are several exact procedures which try to exploit favourable data structures in a systematic way, and to avoid the worst-case behaviour whenever possible. One example for this type of algorithms is specied in this paper: the proposed procedure exploits extensively the special structure of a standard QP (e.g., that the feasible set is the standard simplex), as opposed to the general formulation of a quadratic problem. This article is organized as follows: Section 2 contains a concise exposition of primal and dual problems in copositive programming which involves copositive rather than positive-semidenite matrices, using an explicit characterization of the dual cone of the convex, non-polyhedral cone of all copositive matrices. We also shortly treat the (relaxation of copositive programming to) all-quadratic problems on the simplex as considered in [41]. In Section 3, this will be then specialized to be applied to standard QPs, which enjoy the property that the copositive programming relaxation becomes an exact reformulation of (1). Here the dual is in fact a univariate copositive-feasibility problem which can be seen as a straightforward generalization of an eigenvalue bound problem. Section 4 contains a short review on the replicator dynamics, which by now has become an increasingly popular local optimization procedure for standard QPs. This technique is combined with primal-dual search directions from general conic programming [28, 45], which are used to escape from ine-cient local solutions returned by the replicator dynamics iteration. On copositive programming and standard quadratic optimization problems 3 2. Copositive programming problems: general setup Consider the following primal-dual pair of linear programming problems over a pointed convex cone K IR d , see, e.g. [37, 17, 28, 45]: subject to where D is an m d matrix with full row rank and b 2 IR m while c 2 IR d , and subject to D > y where K Kg is the dual cone of K. In semidefinite programming, coincides with the cone P of all symmetric positive-semidenite n n matrices, which is self-dual under the usual inner product trace (SX) on the d-dimensional Euclidean space S n constructed by identifying the upper triangular half of a symmetric n n matrix with its vectorized version. However, we need not restrict ourselves to cases of self-dual cones K if we can handle the dual cone K , even if the geometry of K and K becomes more complicated. In fact, it turns out useful to study more general cases, e.g. putting K equal to the cone of copositive matrices. Recall that a symmetric n n matrix M is said to be copositive (more precisely, i.e., if the quadratic form generated by M takes only non-negative values on the positive orthant . The matrix M is said to be strictly (IR n -)copositive, if the inequality in (4) is strict whenever v 6= o. Clearly, this cone K has non-empty interior and so does its (pre-)dual cone K (see Proposition 1 below) which can be described as follows, see, e.g. [19, 43]: the convex hull of all symmetric rank-one matrices, i.e. dyadic products generated by non-negative vectors. Elements of K are called completely positive matrices. Note that dropping non-negativity requirement, we again arrive at the semidenite case. Even without constraints, checking whether or not a matrix belongs to K is co-NP-hard [36]. Some algorithms for this problem can be found, e.g. in [33, 18, 46, 5, 47, 48, 14, 6], to mention just a few. Less obvious is the primal feasibility problem (also without constraints). In fact, the authors are not aware of any nite and exact procedure to determine whether or not a given symmetric nn matrix is completely positive if See also [32, 29, 20, 4, 3, 25, 15, 49]. However, the following result may be helpful: 4 I.M. Bomze, M. Dur, E. de Klerk, C. Roos, A. Quist, T. PROPOSITION 1. Let K be as in (5), d = n+1 , and denote by X has no negative Then is a non-negative n (d Proof. In view of Caratheodory's theorem, the rst identity (cf. Theorem 1 of [32]) is obvious by taking 1g, and by noting that the middle set in (7) necessarily is a subset of K due to the fact that hM;FF > are the columns of F . But then the inclusion K+ K is also immediate. To nalize the proof, observe that x > x implies that K co K+ . Unfortunately, the cone K+ itself is not convex and therefore strictly smaller than K, as the following example shows: Example 1. The nonsingular 3 3 matrix with a =6 403 belongs to K (cf. [3]). However, its square root is approximately 0:908 0:092 0:408 0:092 0:908 0:408 0:408 0:408 1:6337 5 whence although the rank-one matrices aa > , bb > and cc > as seen above belong to K+ . A singular variant is obtained by aa > +bb > , i.e. replacing the lower right corner entry of X with 2. As a general application of the primal-dual approach given by (2) and (3) consider the so-called all-quadratic problem on which appears as a subproblem in [41]: subject to x 2 Note that also inhomogeneous quadratic constraints can be written in this form (see above), so that by introducing slacks we also can write problems with quadratic inequality constraints in the form (8). Further, additional linear constraints of the form d can be written as hD; does not change On copositive programming and standard quadratic optimization problems 5 sign over the feasible set (otherwise one has to subdivide this set accordingly). So the normalization condition in can be written as hE; Hence with K as in (5), we may view the following copositive programming problem as a relaxation of subject to hE; Indeed, linearity (in fact, convexity) of the objective ensures that one solution X to the problem (9) is attained at an extreme point of the feasible set. If X happens to lie also on an extreme ray of K, then automatically so that this condition can be dropped without loss of generality. In this case, the relaxation (9) becomes an exact reformulation of (8). Unfortunately, this is not always the case, as the following example shows: Example 2. Consider the problem (9) to maximize hA subject to and, of course, X 2 K as well as hE; 1. Obviously, the only solution to this problem is given by the rank-two matrix X with x We proceed as in the general case with the primal-dual pair (2),(3) to establish the dual problem of (9) which has m+ 1 structural variables y 0 and also d slacks contained in S: subject to y Given that we can solve the primal and dual feasibility problems with limited eort, it is possible to use the search directions for a feasible primal-dual interior point algorithm. Indeed, the following results of Nesterov and Nemirovskii [37] are valid for a general class of convex cones which include the cone K given by (5): There exist so-called self-concordant barrier functions for the cones K and K Interior point methods which converge in a polynomially bounded number of steps can be formulated using the self-concordant barriers. Unfortunately, the self-concordant barriers for K and K are not known, for an elaborate discussion on this topic see [43]. However, Tuncel [45] has recently noticed that even in this case one can still formulate a class of interior point methods known as primal{dual a-ne scaling algorithms. For ease of reference we now reproduce a generic roster for a primal-dual interior-point method from [21, 27, 34]. Of course, it is in general not harder to solve (8) to 6 I.M. Bomze, M. Dur, E. de Klerk, C. Roos, A. Quist, T. optimality than to resolve the feasibility questions (i.e., to check membership of K or there could be special instances where the procedure is still helpful. Generic Interior-Point Primal-Dual Algorithm 1. Choose an initial point (X and that 2. Until a stopping criterion is satised, repeat the following step: choose an improving feasible direction (dX; dy; dS) and step length > 0 such that still X+dX 2 int K as well as S Feasibility w.r.t. the equality constraints is maintained by the so-called primal-dual a-ne scaling (or zero-order) search direction provided by Kojima and Tuncel [28, 45]. Slightly simplied, this class of directions is the solution of the linear system in S n where H is an arbitrary positive-denite, symmetric linear operator on S n and As usual, the remaining (strict) feasibility requirements are guaranteed by a suitable choice of the step lengths. Note that a solution to (11) always exists as also QH is positive-denite provided hX; Si > 0 which is guaranteed for interior point pairs Theorem 3.3 in [45]), since we assume that fE; A are linearly independent, in correspondence with the full row rank assumption on D in (2). Thus we have the same situation as in the classical SDP case for the search direction commonly used there, cf., e.g. [16]. Kojima and Tuncel prove in [28] (cf. Theorem 3.4 in [45]) that if we choose the search directions from (12), then the duality gap decreases linearly with a factor essentially being the step length, and both primal and dual objectives will be improved, unless optimality is reached. Decisive for their arguments is the positive deniteness of H and the property that Looking at formula (12), it is evident that much would be gained if the terms containing HS vanished, which of course is impossible if H is positive-denite. We On copositive programming and standard quadratic optimization problems 7 therefore propose a positive-semidenite variant of the above-mentioned result where H has a single zero eigenvalue belonging to the direction S. Note that we no longer assume that (X; S) 2 int Kint K , but only hX; Si > 0. Recall that the latter relation characterizes non-optimality of pairs (X; S). THEOREM 2. Suppose that H is a positive-semidenite, symmetric linear operator on Consider a pair (X; S) 2 K K with hX; Si > 0 and dene the symmetric linear operator RH on S n by Then RH is positive-denite and satises Furthermore, the solution (dX; dy; dS) to the system in S n IR m+1 S n is unique and satises hX Proof. The rst argument is quite similar to that in Theorem 3.3 of [45]. For any which is non-negative and can vanish only if both hX; O this is absurd as the latter relation implies assumption whereas hX; again by assumption, hX; Si > 0. Hence the operator is positive-denite. Finally, Turning to system (14), we show that the related homogeneous system in (dX; dy; dS) has only the trivial solution. Indeed, substituting for dS in the equation yields dy and substituting then for dX in the rst m+ 1 equations of (14) gives, after changing the signs, a homogeneous system in dy with coe-cient matrix6 6 6 4 8 I.M. Bomze, M. Dur, E. de Klerk, C. Roos, A. Quist, T. which is, due to linear independence of fE; A easily seen to be positive- denite as RH is so. Thus dy = o, yielding Hence (14) has always a unique solution. To establish reduction of the duality gap, let us rst deal with the second-order term hdX; dSi which vanishes because of the feasibility conditions imposed on (dX; dy; dS) in (14): indeed, Now the rst-order terms hX; dSi because of This establishes hX Using similar arguments as in [28], one can also show that both primal and dual objectives are improved by the directions given by (14). We will establish this result directly in the next section for the special case we focus upon in this paper. Remarks In semidenite programming (SDP) where there are many possible choices of the operator RH ; only one choice is known to allow convergent algorithms, namely the Nesterov-Todd primal{dual a-ne-scaling direction, where Note that that RH is a positive denite linear operator. Also note that this choice of RH is not possible for copositive programming, since X is not positive-denite for all copositive matrices S. For SDP, the primal-dual algorithm using this search direction is globally convergent and polynomial for a suitable choice of the step length [26]. Another choice of primal-dual scaling direction is the so-called (primal) HKM a-ne- scaling direction, where . As mentioned, this search direction is not globally convergent for any choice of step length. In particular, it can converge to a non-optimal point [35]. Moreover, it cannot be used for copositive programming because a copositive matrix S can be singular despite hX; Si > 0 for all , which is linearly independent from E. Finally, a primal-dual a-ne scaling direction for SDP which is also dened for copositive programming is the so-called dual HKM direction, which is given as the On copositive programming and standard quadratic optimization problems 9 solution of As with the (primal) HKM direction, no primal-dual SDP algorithm using this search direction is globally convergent [35]. The preceding observations prove two things: Using only primal-dual a-ne scaling directions in interior point methods for conic programming does not necessarily lead to a globally convergent algorithm; One cannot guarantee a xed feasible step length for all primal-dual a-ne scaling directions (even in the SDP case); in other words, 'jamming' can occur. For these reasons we will discuss a hybrid algorithm in Section 4 which uses primal-dual a-ne-scaling steps only as an escape strategy. 3. Standard quadratic optimization and copositive programming First note that the standard QP (1) is a special case of the all-quadratic problem (8) with no quadratic constraints and A A. Hence in this case we arrive at the copositive programming problem (9) with a single constraint: subject to hE; so that the dual has only one structural variable subject to yE This amounts to search for the smallest y such that yE A is copositive. In this sense, the dual problem (16) is related to the question of eigenvalue bounds (replace E with the identity matrix I and \copositive" with \semidenite"). Further observe that in this case, (15) is no relaxation but indeed an exact reformulation of the standard QP (1), due to the following LEMMA 3. The extremal points of the feasible set of (15) are exactly the rank-one matrices 1 From this point on y will denote a scalar variable. I.M. Bomze, M. Dur, E. de Klerk, C. Roos, A. Quist, T. Proof. Of course all belong to 1g. Now suppose that for a vector x 2 , we have xx and some with 0 < < 1. Choose an orthogonal basis fx 1 of IR n with . Then since Z and U also are positive semidenite, we get from that x > both Z and U have rank one. As both belong to K, we thus obtain . But then we obtain x > Z and U must be positive multiples of xx > . The requirement hE; shows that To show the converse, suppose that X is an extremal point of M K. Then as well as i. Hence is, due to (17), a convex combination of matrices U i in M, whence by the extremality assumption is of the form stated. In principle, the roster of the algorithm of Section 2 applies, but the update equations now reduce to which for the dual part means simply that we have to continue the line search for the generalized eigenvalue bound of A as in (16). Of course, a similar reduction applies to the Kojima-Tuncel search directions from (11), where QH replaces RH . Now let us calculate the update steps explicitly, in order to avoid unnecessary numerical complications. Remember that we have still freedom in choosing the positive- semidenite operator H as long as S gives the unique direction to the zero eigenvalue of H (note that by assumption on linear independence of fA; Eg, the matrix never can vanish regardless whether it belongs to K or not). For instance, we may assume that the orthoprojection of E onto the orthogonal complement of S in S n is also an eigenvector of H with a suitably chosen eigenvalue > 0. As this is On copositive programming and standard quadratic optimization problems 11 equivalent to imposing THEOREM 4. Put and denote by (dX; dy; dS) the solution of (18). Then (19) implies Proof. From Inserting further obtain the result for dy, observing that hE; Similarly, we derive hX;Si X, and the proof is complete. For further formulation, it may be convenient to write X with We now directly show that both objectives are indeed improved by the chosen directions. THEOREM 5. Assume that (X; S) 2 KK with hX; Si > 0. If the improving feasible direction (dX; dy; dS) is chosen as in Theorem 4, then for > 0 both primal and dual objective function improve strictly, i.e. Proof. First, dy is strictly negative, since n 2 hE;Si 2 by the Cauchy-Schwarz inequality (note that also fE; Sg are linearly independent). To see the strict monotonicity in the primal objective function, compare the reduction of the duality gap with the improvement in the dual objective. From Theorem 2, we know that the reduction of the duality gap is hX; Si. Therefore, to show that also the primal objective contributes to this reduction, we have to show that dy < hX; Si. But since the denominator of the above fraction is a positive number bigger than 1. 12 I.M. Bomze, M. Dur, E. de Klerk, C. Roos, A. Quist, T. For the sake of completeness, we now also provide explicit update formulae for the original Kojima/Tuncel search direction, i.e. for the solutions to the system with H now again positive-denite but otherwise arbitrary, and QH from (12). Of course, for concrete implementation it remains to specify the values HE and HA. THEOREM 6. Assume that (X; S) 2 K K with hX; Si > 0. Put denote by (dX; dy; dS) the solution of (22). Then Furthermore, both primal and dual objectives are strictly improved if > 0. Proof. The arguments are very similar to that of Theorems 4 and 5, and therefore omitted. 4. A hybrid method: replicator dynamics and primal-dual escape steps To nd local solutions to the standard QP (1), we propose to use replicator dynamics. For the reader's convenience, we here provide a short overview, and refer for more detail to [7, 11, 12]. Consider the following dynamical system operating on : _ where a dot signies derivative w.r.t. time t, and a discrete time version ng : (25) Note that x(0) 2 IR n implies since A is nonnegative by assumption. The stationary points under (24) and (25) coincide, and all local solutions of (1) are among these. Of course, there are quite many stationary points, e.g. all vertices of . However, it can be shown [7] that x is a strict local solution if and only if x is On copositive programming and standard quadratic optimization problems 13 asymptotically stable, i.e. every solution to (24) or (25) which starts close enough to x, will converge to x as t %1. Both (24) and (25) arise in population genetics under the name selection equations where they are used to model time evolution of haploid genotypes, A being the (symmetric) tness matrix, and x i (t) representing the relative frequency of allele i in the population. The Fundamental Theorem of Selection states that average tness, i.e. the objective function x(t) > Ax(t) is (strictly) increasing over time along trajectories [13, 22], and moreover every trajectory x(t) converges to a stationary point [31, 22]. Furthermore, one can prove [7, 12] the following facts: if no principal minor of vanishes, then with probability one any trajectory converges to a strict local solution x of (1); further, if ng with y 6= x; and - is contained in the basin of attraction of x, where for a subset we shall denote the face of corresponding to by and its relative interior by The dynamical systems (24) and (25) are frequently called replicator dynamics, and are well suited for implementation in practical applications, see [40, 11, 8]. This is re ected also in theory by the result that (24) is most e-ciently approaching xed points in the sense that it is a Shahshahani gradient system [44]. The discrete time version (25) also corresponds to a particular instance of an algorithm widely popular in computer vision. These relaxation labeling processes are closely related to articial neural network learning systems, and have found applications in a variety of practical tasks, e.g. to solve certain labeling problems arising in the 3-D interpretation of ambiguous line drawings [42, 24, 39]. Furthermore, the dynamics (25) belongs to a class of dynamical systems investigated in [1, 2], which has proven to be useful in the speech recognition domain [30]. Although strictly increasing objective values are guaranteed as we follow trajectories under (24) or (25), we could get stuck in an ine-cient local solution x of (1). From the preceding results, then necessarily x One possibility to escape from x is by the G.E.N.F. approach [12]. An alternative is to merge the replicator dynamics method with the usual interior-point steps borrowed from semidenite programming, and this will be described in the sequel. But given any escape procedure, we are now ready to describe the principal algorithm for solving (1) globally. Note that this procedure stops after nitely many repetitions, since it yields strict local solutions (in every 14 I.M. Bomze, M. Dur, E. de Klerk, C. Roos, A. Quist, T. there is at most one of these) with strictly increasing objective values: Replicator Dynamics Algorithm 1. Start with n e or nearby, iterate (25) until convergence; the limit lim t!1 x(t) is a strict local solution with probability one (provided all principal minors of A do not vanish); 2. call an escape procedure to improve the objective, if this is still possible; denote the improving point e x; 3. repeat 1., starting with x. Now we are ready to present a combination of the above procedure and the interior-point method yielding improving direction, in the hope that it will be possible to escape from ine-cient local solutions. A Hybrid Algorithm 1. Initialisation: choose n e or nearby. Put y 2. Replicator dynamics for fast primal updates: starting from x(0) iterate (25) until convergence; the limit is a strict local solution with probability 3. Dual update: check copositivity of yE A via shortcuts (cf. Fig. 1 in [6], for a special case see [10]). In the a-rmative, x is the global solution of (1), since the duality gap is zero (cf. also Theorem 7 in [7]); stop. If however a point e is found such that e x > (yE A)ex < 0, then e x improves the objective; repeat step 2 starting with this point. Else (no decision), keep the old value of y 0 , and proceed to step 4. 4. Step back from the boundary: Choose > 0 so small, that the point and the matrix I satises (with b x the previous iterate, so that x > Ax b x > Abx is the previously obtained improvement) x > Abx 1 This is a quadratic inequality for . Note that by construction, X is both positive- denite and has a non-negative square root: r On copositive programming and standard quadratic optimization problems 15 5. Since X is both positive-denite and has a non-negative square root, one can choose Y as in (21) and > 0 su-ciently small such that e the same properties. This is possible because the mapping X 7! X is Holder continuous around X. Hence primal feasibility of e X is maintained (cf. Lemma 1), and we get an explicit positive (square root) factorization of e are all non-negative. Thus e with i 0 and x 1g. 6. Primal update: Now x > Ax hA; Xi < hA; e possible, choose e x such that x > This will always be possible if hA; e Repeat from step 2, starting with The following small example illustrates the ideas behind the hybrid algorithm. In par- ticular, the example is meant to illustrate how the escape strategy in steps 4 through 6 works. Example 3. Let A = and suppose we arrived via replicator dynamics starting at already at the (suboptimal) local solution x . Then and improvement As x is not the global solution, the matrix is not copositive. Following step 3 of the hybrid algorithm, we return to the old y arrive at the dually feasible (i.e., copositive) matrix which incidentially coincides with the optimal S (see below). Note that although neither X nor S is interior, we have hX; Suppose for the moment we ignored step 4 above and tried to proceed directly in forming the matrix Y (the signshall emphasize the preliminarity of this trial) for the escape step along (20) and (21). The key quantities in (20) and (21) are hE; 9. Hence I.M. Bomze, M. Dur, E. de Klerk, C. Roos, A. Quist, T. and, furthermore, Therefore dy 9 which is positive-denite but has negative o-diagonal entries so that Y is infeasible for all positive . This shows that the step back from the boundary, i.e. step 4 in the hybrid algorithm is really necessary. Note that ignoring primal feasibility in this respect, one could investigate instead whether the vector Y e belongs to (i.e., has no negative coordinate, as automatically e > improves the objective. With regard to the latter aim, it is desirable to take as large as possible (recall that x is locally optimal and the objective is quadratic so that the improvement will be largest for the largest possible distance { if there is one at all). In our case, this means considering as a candidate for an improving point in the vector 3+9 38 as and indeed we manage to escape because we get an improvement if is chosen large enough, as ( But let us return to the hybrid algorithm as proposed above: choose, e.g., step 4, so that 0:9 0:1 and xx 0:81 0:09 0:09 0:01 as well as I 0:748 0:072 0:072 0:108 with so that the duality gap will be slightly increased, as expected. Note that S and the update part (26) of Y remain the same as the dual variable y does not change, and observe that, as required in step 4, the choice of Now Motivated by the trial with above, we choose large enough to enable an escape, e.g. and arrive at the positive-denite matrix On copositive programming and standard quadratic optimization problems 17 Hence the primal feasibility requirement e will be met if and only if all entries of the latter matrix are non-negative, which means 0:864 1:792 0:482. A typical choice of in step 5 would be 2 (cf.[45]), but for simplicity we choose here 3 . Then e 0:6306 0:0222 0:0222 0:3250 int K with 0:7939 0:0163 0:0163 0:5699 and h e In step 6 of the hybrid algorithm we therefore obtain by normalizing the last column of X, with objective value e x > Aex 2:8912 > 2. The last steps in the hybrid algorithm are as follows: use the improving point e x as the starting vector for the replicator dynamics iteration, which nally leads to the global solution . For the nal check for optimality we now calculate X 5. Conclusions The problem of maximizing a quadratic form over the simplex has an exact reformulation as a copositive programming problem, i.e. a conic programming problem over the cone of copositive matrices. The advantage of such a reformulation is that successful ideas from the theory of interior point methods can thus be applied to nonconvex quadratic optimization. In particular, primal-dual a-ne scaling directions can be used in escape strategies if ine-cient local solutions are obtained from local optimization procedures like replicator dynamics. --R An inequality with applications to statistical estimation for probabilistic functions of Markov processes and to a model for ecology. Complete positivity. Remarks on the recursive structure of copositivity. An introduction to population genetics theory A recursive algorithm to detect (strict) copositivity of a symmetric matrix. Completely positive matrices associated with M- matrices Exploiting sparsity in primal-dual interior-point methods for semide nite programming On copositive matrices. An interior-point method for semide nite programming The Theory of Evolution and Dynamical Systems. Introduction to Global Optimization On the foundations of relaxation labeling processes. On nonnegative factorization matrices. Polynomial primal-dual a-ne scaling algorithms in semide nite programming Square triangular factorizations of completely positive matrices. An introduction to the application of the theory of probabilistic functions of a Markov process to automatic speech recognition. Factorizations of completely positive matrices. Copositive matrices and de Some NP-complete problems in quadratic and linear program- ming Interior point methods in convex programming: theory and applications. A new semide On the dynamics of relaxation labeling processes. Nonconvex all-quadratic global optimization problems: solution methods Scene labeling by relaxation operations. Copositive relaxation for general quadratic programming. Game dynamics Notes on completely positive matrices. --TR --CTR Kurt M. Anstreicher , Samuel Burer, D.C. Versus Copositive Bounds for Standard QP, Journal of Global Optimization, v.33 n.2, p.299-312, October 2005 Immanuel M. Bomze , Etienne De Klerk, Solving Standard Quadratic Optimization Problems via Linear, Semidefinite and Copositive Programming, Journal of Global Optimization, v.24 n.2, p.163-185, October 2002 Immanuel M. Bomze, Branch-and-bound approaches to standard quadratic optimization problems, Journal of Global Optimization, v.22 n.1-4, p.17-37, January 2002
positive semidefinite matrices;copositive programming;standard quadratic optimization;global maximization
596362
Global Optimization of Marginal Functions with Applications to Economic Equilibrium.
We discuss the applicability of the cutting angle method to global minimization of marginal functions. The search of equilibrium prices in the exchange model can be reduced to the global minimization of certain functions, which include marginal functions. This problem has been approximately solved by the cutting angle method. Results of numerical experiments are presented and discussed.
Introduction Marginal functions, that is, functions of the form where a is a set-valued mapping, have been intensively studied by many researches, (see, for example, [11, 16, 14] and references therein). Note, that the so-called max function: is the simplest example of a marginal function. Here Y can be considered as the image of a constant mapping a. We assume that the set Y in (1.2) is nite dimensional, innite and compact. Marginal functions arise in the study of many problems of mathematical economics. Global minimization of marginal functions (in particular, max functions) is a very complicated problem. Indeed, almost all known methods of global minimization require This research was supported by the Australian Research Council to compute values of the objective function many times. However, the value of marginal function of (1.1) at a point p can be found only by solving the optimization problem subject to y 2 a(p); so the calculation of the value of is very time consuming. Recently the so-called cutting angle method for global optimization of Lipschitz function has been developed (see [2, 23, 6, 7]). We need to compute only a few values of the objective function at each iteration of the cutting angle method, so we can hope that this method is suitable for minimizing some marginal functions. In this paper we apply the cutting angle method to solving some problems of global optimization, which arise in the theory of economic equilibrium. The objective functions of these problems are not always Lipschitz, so we need to transform objective functions in order to obtain Lipschitz continuity. Note that the data in models of economic equilibrium often is not very precise. Thus, we can restrict ourselves to the search of approximate solutions of corresponding problems of global optimization. We show that the cutting angle method, which can nd an approximate solution fairly quickly, can serve for solving these problems. Consider a market economy with m consumers (agents) and n goods. The consumer j has a utility function U j and a vector of initial endowments ! j . It has been shown in [20] that, under some natural assumptions, a vector of equilibrium prices of this model can be found as a solution of the following problem: subject to p 2 ri with are special marginal functions (see Section 5). Problem (1.3) has been studied from various points of view in [1] and [5]. The similar approach for Arrow-Debreu equilibrium model was proposed in [21]. The equilibrium exists if and only if the value of problem (1.3) is equal to zero. It is well known that the equilibrium does exist if the utility function U j is quasiconcave and j is a strictly positive vector for all j. If these conditions hold, we need to solve problem (1.3) with the known value of global minimum. If at least one of these conditions is not valid, problem (1.3) can serve for recognition of the existence of the equilibrium. The structure of this paper is as follows. In Section 2 we discuss problems, which arise under minimization of marginal functions. In Section 3 we recall brie y the cutting angle method and in Section 4 we recall the exchange model of economic equilibrium. Section 5 provides a discussion of a reformulation of an equilibrium problem as a special optimization problem. In Section 6, we study Lipschitz continuity of the objective function of this optimization problem. Section 7 provides a discussion of the calculation of equilibrium prices by the cutting angle method. We record results of numerical experiments in this section. Appendix contains data, which were used for numerical experiments. 2 Minimization of marginal functions Let a be a set-valued mapping transforming a set P into the set of all non-empty subsets of a set Y . Consider a function function dened on P by is called a marginal function. In this paper we assume that P and Y are closed subsets of nite-dimensional spaces and images of the mapping a are compact sets. Assume also that the function y 7! f(p; y) is upper semicontinuous for all . Then the supremum in (2.1) is attained. Continuity, Lipschitz continuity and directional dierentiability of marginal functions have been intensively studied (see, for example, [4, 11, 14, 16] and references therein). A well-known example of marginal functions is delivered by a parametric problem of mathematical programming: subject to g i (p; y) 0 (i 2 I); h j (p; Here J)g. It is assumed that the set a(p) is nonempty and compact for all Marginal functions arise in the study of many problems of mathematical economics. Assume, for example, that a is a production mapping of a producer, that is, a(x) is the set of all outputs, which can be produced by the producer from an input x. Let p be a price vector (that is a vector with positive coordinates). Then the maximal prot (x), which can be obtained by the producer, is equal to (max y2a(x) [p; y]) [p; x], where [u; v] is the inner product of vectors u and v. The marginal function max y2a(x) [p; y] describes the revenues of the producer. A dierent kind of marginal functions arises in the study of a consumer behaviour. Assume that a consumer has a utility function U . Then the maximal utility, which can be obtained by the consumer is where B(p) is the set of vectors of goods, which are available for the consumer, if there is a price vector p at the market. are marginal functions and let P be a compact set. Consider the following problem of global optimization subject to p 2 P: (2.2) Many problems from various elds of mathematics and its applications can be represented in form (2.2). We mention here only bilevel programming (see for example [24] and references therein) and problems of economic equilibrium (see [20]). Consider the simplest version of (2.2), namely the problem subject to p 2 where is a marginal function dened by (2.1). A very special case of problem (2.3), where the mapping a in (2.1) is constant: known as a minimax problem. Thus the minimax problem has a form: to nd a point (~p; ~ Y such that If Y is nite, we have a discrete minimax problem, if Y is a compact innite set, we have a continuous minimax problem. Discrete minimax problems have been studied by many authors (see, for example [10, 17] and references therein). We consider here only continuous minimax problems, which are much more complicated. We mention three numerical methods, which were proposed for solving continuous minimax problems in the seventies and earlier: Arrow - Hurwicz method (see [3, 15]), the net method (see [10]); and the method of extremal basis (see [9]). Modern approaches to solving these problems can be found in the book [17] by E. Polak. The main attention in this book is paid to the minimization of convex functions of the form The net method and the method of consistent approximation [17] can be applied for the search for local minima of , when this function is non-convex. These methods are based on an approximation of the given set Y by a nite set Y 0 . Having such an approximation we can substitute a continuous minimax problem for a sequence of discrete minimax problems and then solve these problems by known algorithms (see, for example [10, 17]. These methods require to use large nite sets Y 0 , hence we need to minimize the maximum of a large number of functions. The calculation of local minima of a non-convex marginal function is very time- consuming. Indeed, the application of a majority of known numerical methods for local optimization is based on the calculation of the objective function and its subgradients (in a certain sense) in many dierent points. Sometimes it is possible to use only approximate values of the function and its subgraients, however, in order to nd an approximate value of the function at a point p we need to nd an approximate solution of the problem subject to y 2 a(p): (2.5) Thus the search for a local minimum of the function requires to solve (approximately) problem (2.5) very often. Note that as a rule the calculation of subgradients of a non-convex marginal function is more complicated than the calculation of a value of this function. It leads to the following conclusion: optimization methods, which are based on the calculation both functions and its subgradients, as a rule are not applicable for the local minimization of non-convex marginal functions. Methods, which are based only on calculation of values of the objective function (so-called derivative-free methods), are more preferable. any convexity assumptions a marginal function may have a lot of local mini- mizers, so we need to use methods of global optimization for solving problem (2.2). How- ever, the majority of these methods (branch-and-bound, random search etc) require very many objective function evaluations. So, these methods are not applicable for minimization of marginal functions (and even for solving minimax problems) if the dimension of the problem is suciently high. For global optimization of marginal functions we need to nd methods, which require a small amount of objective functions evaluations. One of such methods is the so-called cutting angle method, which uses only few function evaluations at each iteration. So we can propose that the cutting angle method is applicable for the global minimization of marginal functions. Cutting angle method ng. Consider the space IR n of all vectors We shall use the following notations. is the i-th coordinate of a vector x 2 is the inner product of vectors l and x. A function f dened on the cone IR n of all n-vectors with nonnegative coordinates is called an IPH function if f is increasing (x y =) f(x) f(y)) and positively homogeneous of the rst degree 0). The following result holds (see [23]). Theorem 3.1 Let f be a Lipschitz function dened on the unit simplex 1g. Then there exists a constant c > 0 such that for all c 0 c the function is the restriction of a certain IPH function g to S. Thus the minimization of a Lipschitz function f over S is reduced to the minimization of an IPH function g over S. The minimization of a Lipschitz function subject to linear constraints can be transformed to the minimization of another Lipschitz function over the simplex (see [23] for details), hence the minimization of a Lipschitz function subject to linear constraints is reduced to the minimization of an IPH function over the simplex. The cutting angle method was proposed (see [2] and also [22]) for the minimization of a so-called ICAR (increasing convex-along-rays) function dened on IR n over a compact subset of IR n . We consider here only a version of this method, which is suitable for the minimization of an IPH (increasing positively homogeneous of degree one) function over the unit simplex. This version has been proposed and discussed in detail in [6, 7]. It follows from monotonicity of an IPH function f that f(x) . We assume in the sequel that f(x) > 0 for all x 6= 0. For x 2 IR n we shall use the following c=x is the vector with the following coordinates: x The cutting angle method is based on the following result (see [22] and references therein). Theorem 3.2 ! IR be a function such that f(x) > 0 for all x 6= 0. Then f is IPH if and only if there exists U IR n f0g such that l2U min l min l and min l be the unit vector such that I(e k First we present the simplest version of the cutting angle method. The cutting angle method for minimizing an IPH function over the unit simplex Dene the function h kn min l k kn l k and set Step 1. Find a solution x of the problem subject to x 2 S: (3.1) Step 2. Set Step 3. Compute dene the function l j min l k and go to Step 1. The convergence of the cutting angle method has been proved under very mild assumptions (see [16], where the convergence of much a more general method was established, and also [22]). Remark 3.1 The solution of the auxiliary problem (3.1) is the most dicult part of the algorithm. The special method for the solution of this problem has been developed in [6, 7]. We do not discuss this method here. Dierent approaches are possible in lower dimensions ([2, 23]): in particular, the auxiliary problem can be reduced to a mixed-integer linear programming problem, which can be solved by standard optimization packages (for example, CPLEX). Remark 3.2 Only one value of the objective function f should be calculated at each iteration. Remark 3.3 A more advanced version of the cutting angle method was proposed in [7]. All approximate solutions of the problem (3.1) are considered in this version. We shall use the advanced version in this paper, since it allows us to fasten the search for a global minimizer (see [7] for details). The advanced version may require a few calculation of the objective function at each iteration. Remark 3.4 Let It follows from Theorem 3.2 that h j (x) f(x) for all x 2 S. Hence Thus j is a lower estimate of the global minimum of f over S. Let It can be shown (see, for example [22] and references therein) that j is an increasing sequence and j we have a stopping criterion, which enables us (at least theoretically) to obtain an approximate solution with an arbitrary given tolerance. Remark 3.5 The cutting angle method can be considered as a special case of many well-known algorithms (see [22] for a short survey of some of these algorithms). However, the numerical implementation of the cutting angle method has demonstrated that it works much better than many other versions of these algorithms. 4 The equilibrium model We study the so-called exchange models of economic equilibrium. First we consider a classical version of this model. The classical exchange model describes a market, where n goods are circulated. Let I ng. There is a nite number, say m, of economical agents, which are called consumers, at the market. Let is described by a pair (U is a utility function (objective function) of the consumer is a vector of her initial endowments. A state of economy is a vector . A state X is called feasible if it belongs to the set where is the vector of all products available on the market. A vector p 2 IR n n f0g is called a price vector. Having a price vector p, the consumer can sell her initial endowment ! j and can buy a vector x j from the set This set is called the budget set of the consumer j. Note that B j so we can assume without loss of generality that the mapping B j is dened only on the unit simplex 1g. We shall denote the model under consideration by M: A semi-equilibrium of the model M is a pair (p; X), where p is a price vector and is a feasible state, such that x j is a solution of the consumer problem: subject to x The vector p is called equilibrium prices and the vector x j is called an equilibrium state of the agent j. It is well known that a semi-equilibrium exists if functions U j are quasiconcave and vectors ! j are strictly positive (see, for example, [13] and references therein). We assume in the rest of this paper that We shall consider both models with quasiconcave U j and without quasiconcave U j . Let (p; X) be a semi-equilibrium of the model M. Since X is a feasible state, the inequality is dened by (4.2). A semi-equilibrium (p; X) is called an equilibrium if It is well-known (and easy to check), that each semi- X) with is an equilibrium. We accept the following Assumption 4.1 There exists j 2 J such that the utility function U j is non-satiated in the following sense: for each x 2 IR n and each i 2 I there exists i > 0 such that is the i-th unit vector. Then each vector of equilibrium prices p is strictly positive. Indeed, assume that Let x j be an equilibrium state of the consumer j. Then the budget set B of this consumer contains the ray fx there exists (p) such that U j which contradicts the denition of the semi-equilibrium. Thus Assumption 4.1 allows us to consider only strictly positive equilibrium prices (therefore, an equilibrium is guaranteed to exist by the above). We also consider one more type of exchange models, namely a model of exchange with xed budgets. In contrast with the classical version, it is assumed that each agent j has a xed budget, that is, a special sum of money d j , which does not depend on market prices . The budget mapping B j (p) should be replaced for the mapping ~ Having mappings ~ (p), we can dene the semi-equilibrium and equilibrium in this model in the same way as in the classical case. Note that in contrast with the classical model, ~ so we need to consider mappings ~ on the cone IR n f0g. The existence results for the model with xed budgets are similar to those for the classical model. If Assumption 4.1 holds then equilibrium prices are strictly positive. Remark 4.1 It is known [18, 13] that an equilibrium for a model with xed budgets and positively homogeneous utility functions can be found by solving a convex programming problem. However this does not hold if utility functions are not positively homogeneous. In this paper we consider a search for an economical equilibrium as an example of application of the cutting angle method. From this point of view models with xed budget is of a special interest since they can be reduced to a problem of global optimization, which is dierent from that for classical exchange models. 5 Equilibrium prices as a solution of a special optimization problem It is not hard to formulate a bilevel problem such that its solution set coincides with the set of equilibrium prices. Let Assumption 4.1 hold. Consider the set ri ++ . Assume for the sake of simplicity that the utility functions U are strictly concave. Then the consumer's problem subject to x has a unique solution for each p 2 ri S. Denote this solution by x j (p). Consider a function only if p is a vector of equilibrium prices. Thus equilibrium prices can be found as a solution of the following bilevel problem: min subject to p 2 ri S (5.2) where x j (p) is a solution of problem (5.1). The optimal value of problem (5.2) is known (and equal to zero). However, this problem is very complicated. Note that the set-valued mapping B j is not Lipschitz on ri S. (See, for example, [4] for the denition of Lipschitz set-valued mappings.) So we cannot hope that the mappings x j (p) (j 2 J) and the function (p) are Lipschitz for an arbitrary functions U j . Hence we cannot guarantee that problem (5.2) can be solved by the cutting angle method. We now consider a certain dierent type of optimization problems, which can serve for the search of economic equilibrium. Such problems were suggested in [20]. For each ri S consider sets A Y and Both sets A (p) and A(p) are compact for p 2 ri S, so the following functions are well ri S (5.5) and ri S: (5.6) 0: The following assertion holds (see [20]). Lemma 5.1 Let Assumption 4.1 holds. and let ri S. The equality if and only if p is a vector of equilibrium prices. Thus in order to nd a vector of equilibrium prices we need to solve the following problem of global optimization: subject to p 2 ri S: (5.7) The objective function of this problem is the dierence of two marginal functions H 1 and . If the equilibrium exists then the minimal value of this problem is known and equal to zero. We accept the following assumption. Assumption 5.1 (1) For each j 2 J there exists a point a j 0 such that the utility function U j is twice continuously dierentiable on the set a ++ and ++ and y j 6= 0: (5.8) (2) The utility function U j (j 2 J) is increasing in the following sense: if x 1 x 2 then It follows from (5.8) that U j is a strictly concave function. If Assumption 5.1 holds then for each p 0 and each j 2 J the consumer's problem subject to x has a unique solution x Hence, the problem subject to which serves for the denition of the function H 1 , has a unique solution The problem X subject to which serves for the denition of H 2 also has a unique solution ~ It has been proved in [21] that the function H 1 is Frechet dierentiable with the piece-wise mapping rH 1 . We have is an arbitrary index belonging to I j (p) := 0g. The function H 2 is directionally dierentiable and, under some additional assumptions, also Frechet dierentiable with the piece-wise C 1 gradient mapping. Assume that there exist I and a sequence p (k) 2 ri S such that It follows from (5.11) that So we cannot hope that H(p) is a Lipschitz function on the set ri all quasiconcave utility functions U j and we cannot guarantee that the global minimizer of H over the simplex S can be found by the cutting angle method. In order to get a Lipschitz function we shall slightly change the denition of the function H. First we note that the function H 2 is bounded on ri S. Indeed, let p 2 ri S and let Since the utility functions U j are increasing, we have where However the function H 1 can be unbounded on S. Since we are interested in the global minimum of the function points where the function H 1 is very large. For this purpose we consider the function where d is an arbitrary positive number, which is bigger than a global minimum of the function H. Clearly global minimizers of functions H and L coincide so if the equilibrium exists, then a point p 2 ri S is an equilibrium prices if and only if Thus a vector of equilibrium prices can be found as a solution of the following problem: subject to p 2 ri S: (5.14) 6 Lipschitz continuity of the function L. In this section we shall check that the function L is Lipschitz. Let dg. It is sucient to verify that L is Lipschitz on the set E. First we prove the following simple Proposition. Proposition 6.1 The function H 1 is bounded on the set E. Proof: If dened by (5.12). 2 Corollary 6.1 There exists a vector a 2 IR n ++ such that x j (p) a for all Indeed, U j (x(p)) H 1 (p). So Proposition 6.1 and Assumption 5.1(3) imply the boundness of the set fx Jg. We need the following theorem which is a special case of a general result from [19]. (This special case and its relation with the mentioned general result were discussed in Theorem 6.1 Let P be an open subset of a nite dimensional space and x (p) be a solution of the following parametric convex programming problem: subject to where K and R are nite sets, a k (k 2 K); h r (r 2 R) are vectors and A r (r 2 R) are matrices. Assume that T is a twice continuously dierentiable concave function dened on an open set such that Assume also that the Mangasarian-Fromovitz constraint qualication holds at the point x (p), that is, there exists y 2 IR n such that where Then the function x (p) is piece-wise continuously dierentiable near the point p. In particular, x (p) is a locally Lipschitz function near the point p. In order to apply this theorem, we need to consider an open set P S such that for all Such a set exists since ! j 0 for all j 2 J . For the following two optimization problems. Problem subject to [p; x] where a is a vector from Corollary 6.1. Problem P 1 has a unique solution x (p). Problem subject to This problem has a unique solution ~ (p)) j2J . Proposition 6.2 1) For each p 2 P the Mangasarian-Fromovitz constraint qualication holds for problem P 1 at the point x j (p). 2)For each p 2 P the Mangasarian-Fromovitz constraint qualication holds for problem P 2 at the point ~ We leave the simple but cumbersome proof of this proposition to the reader. Proposition 6.3 Let Then the function H is locally Lipschitz on the set P . Proof: It follows from Theorem 6.1 and Proposition 6.2 a that functions x (p) and ~ x j (p) are locally Lipschitz. Functions U j are locally Lipschitz as well. Hence H (p) is locally Corollary 6.2 The function H (p) is Lipschitz on the compact set S. Theorem 6.2 The function L dened by (5.13) is Lipschitz on the simplex S. Proof: It is sucient to prove that the function H is Lipschitz on the set H(p) dg. Since x j (p) a for (p) for (p) for all p 2 S. Hence The result follows from Corollary 6.2. 2 7 The calculation of equilibrium prices by the cutting angle method In this section we discuss results of numerical experiments, which were carried out in order to nd an approximate equilibrium prices or to verify that the equilibrium does not exist. The problem (5.14): subject to p 2 ri S has been solved by the cutting angle method. First we give some remarks. 1. We consider a search of an economical equilibrium as an example of applications of the cutting angle method to approximate global optimization of a complicated function, which is the dierence of two marginal functions. Currently the developed approach can mainly be used for research purposes. We hope that the further development of the cutting angle method will allow one to use this approach for many real problems of economic equilibrium. 2. Assume that the equilibrium exists. We have actually we are looking for a global minimum of the function H, which is equal to zero. We also use the function H for determining the precision of results of calculations. This function is dened with the help of utility functions U j , which describe the preferences of agents. Note that the function j U j with j > 0 and function U j describes the same preferences (see [20] for a corresponding discussion). Thus if we replace the function H for the function H with arbitrary , we obtained the same approximate global minimizers, however the measure of the precision will be changed. In order to avoid this situation, we need to consider only functions U j , which are normalized in a certain sense. All numerical experiments were carried out with utility functions of the form are dened by maximization of the sum we consider the following normalization of function H: 3. The cutting angle method is suitable for the search of approximate solution of a global optimization problem. If there exists the equilibrium of the model under consideration, then the optimal value of the normalized function H is equal to zero. We search for vectors p such that H(p) 0:0010:003. Assume that we consider a model, for which the existence of the equilibrium are not proved and we want to learn whether the equilibrium exists. Recall (see Section 3) that the cutting angle method produces the lower estimates j of the global minimum. If j > 0 for some j, then the equilibrium does not exist. 4. We used the exact penalty method for solving internal problems, which are problems of convex programming (maximization of a concave function subject to linear constraints). Thus, the exact penalty functions have been constructed; for their minimization we used the so-called discrete gradient method ([5]). Since the precision of the results obtained by the cutting angle method is approximately 0:001 0:003, internal problems were solved with the precision 10 4 . 5. We used a computer IBM Pentium-S CPU 150 MHz. Problems with were mainly considered. It takes approximately 15-18 minutes to nd a solution of such a problem with the precision 0:001 0:003. Results of numerical experiments show that the solution of internal problems takes the main part of CPU time. Our aim is to show that the cutting angle method can be successfully applied for solving the problem (5.14). Using more eective methods of convex optimization for solving internal problems and more iterations of the cutting angle method we can nd a more precise solution of problem (5.14). Classical exchange models with concave utility functions First we consider a classical exchange model such that Assumption 4.1 and Assumption 5.1 hold. Then the equilibrium does exist, so the value of problem (5.14) is known and equal to zero. Since the function L is Lipschitz over the unit simplex, the cutting angle method can be applied. Example 7.1 Consider the economical system with 8 consumers and 4 goods. The utility functions are dened as follows: Vectors are the rows of the matrices are rows of matrix 1 . Coecients are coordinates of the vector c 1 . Matrices A 1 and vector c 1 can be found in Appendix. Numerical results for Example 7.1. The point with found by the cutting angle method after iterations. It takes 22 objective function evaluations. Example 7.2 We again consider the same economical system. Utility functions have the same form with the same vector of coecients c 1 and the same matrix B 1 . However coecients j are dened in dierent way. We assume now thatX Vectors are the rows of the matrix A 0 1 , which is in Appendix . Numerical results for Example 7.2. The point with found by the cutting angle method after 17 iterations. It takes objective function evaluations. Exchange model with xed budgets and concave utility functions Consider a model with xed budgets such that Assumption 4.1 and Assumption 5.1 hold. We now use the approach similar to that for the classical model. Let p 2 IR n ++ . Consider set-valued mappings ~ A ~ where ~ dened by (4.5). Let ~ ~ Further, let ~ d), where d is an arbitrary positive number. The same argument as in Section 5 demonstrates that a vector of equilibrium prices can be found as a solution of the problem: ~ subject to p 2 IR n In order to apply the cutting angle method we need to transform this problem to a problem with a simplicial constraint. For this purpose we introduce constraints p i M where M is a suciently large number. (It is assumed that there exist a vector p of equilibrium prices such that Consider the set i2I Clearly this set contains the set fp Adding a new variable p n+1 and replacing p i =t by q i we transform the set D to the unit simplex . Example 7.3 Consider the economical system of Example 7.2. It is assumed now that each consumer j has the xed budget d j and her budget set ~ (p) has the following form ~ Vector can be found in Appendix. Numerical results for Example 7.3. The point with found by the cutting angle method after 5 iterations. It takes 24 objective function evaluations. Note that 3. Classical exchange model with not necessarily concave utility functions The existence of the equilibrium in classical model can be proved if utility functions U are quasiconcave. If at least one of these functions is not quasiconcave we cannot guarantee that the equilibrium exists. However, we also cannot guarantee that the equilibrium does not exist. In order to check the existence of the equilibrium we can solve problem (5.14). If the value of this problem is equal to zero then the equilibrium exists and a global minimizer of the function L is a vector of equilibrium prices. Otherwise, the equilibrium does not exist. Some numerical experiments have been carried out in order to check the existence of equilibrium. We consider utility functions, which can be represented as the maximum of two concave functions. (Functions of such structure arise when indivisible goods are considered.) We now describe numerical experiments which were carried out. 1. A number of models with two goods and two consumers has been examined. Each consumer has a utility functions, which is the maximum of two functions of the form (*). Coecients c j and b i were chosen at random. Coecients j were chosen also at random, however it is assumed that either 0:8. A model with three goods and three consumers, which have utility functions of the same form with chosen at random coecients, also has been examined. Numerical experiments showed that all models under consideration possess equilibrium. These results allow us to consider the following conjecture: if the number of goods is equal to the number of consumers and a utility function of each consumer is the maximum of two functions, for which Assumption 4.1 and Assumption 5.1 hold, then the classical exchange model has an equilibrium. This conjecture was discussed with Professor J.-M. Bonnisseau, who suggested the following counter-example. Example 7.4 (J.-M. Bonnisseau, [8] Consider a classical exchange model with two goods and two consumers. Utility functions U j and vectors ! j (J =1,2) of consumers have the following The demand of the consumers can be explicitly calculated. This calculations showed that an equilibrium does nor exist. Numerical experiments, based on the approach suggested in this paper, also conrmed that the model under consideration and also some similar models have no equilibrium. Nevertheless our numerical experiments which were performed by a random choice of coecients show that there exists many models, which have the described above form and possess an equilibrium. 2 A model with two products and three consumers has been examined. Each consumer has a utility function, which is the maximum of two functions of the form (*) with chosen at random coecients. This model has no equilibrium, however each its submodel with two consumers possesses an equilibrium. We now present one of the corresponding examples. Example 7.5 Consider two economical systems, one of them has 3 consumers and 2 goods, the other has 3 consumers and 3 goods. Assume the consumer j has a utility function where U jk Y jk 2: For the rst system vectors are the rows of matrices A 3 and respectively, which can be found in Ap- pendix. Coecients (c 1 3 are coordinates of the vector c 2 (see Appendix). For the second system vectors and are the rows of matrices A 3 and 3 , respectively (see Appendix). Coecients 3 are coordinates of the vector c 2 . First we describe our approach for solving internal problems. Note that the function H 1 has the form ri S: Hence, we can nd the value H 1 (p) of the function H 1 at a point p by solving the convex programming problems: U ji We also have for p 2 ri S: In order to nd the value H 2 (p) of the function H 2 at the point p we consider all possible combinations: 2: Then we solve the convex programming problems : and calculate H 2 (p) by (7.4). Solving problems (5.14) for the described models, we found that the rst economic system does not possesses the equilibrium. However, all its subsystems consisting of 2 consumers and 2 goods have the equilibrium. The second system also has the equilibrium. Appendix In this Appendix we report the data, which were used for numerical experiments. Examples 8.1 { 8.3 0:05205 0:35366 0:24230 0:15199 0:23562 0:08699 0:30802 0:16937 0:25335 0:17358 0:10888 0:26420 0:09515 0:33689 0:18525 0:18271 0:21114 0:13245 0:32138 0:13503 0:25861 0:14220 0:14025 0:25893 0:13280 0:32224 0:13540 0:20955 0:18448 0:18196 0:33593 0:09763C C C C C C C C C C C C A 0:06506 0:44207 0:30288 0:18999 0:29452 0:10874 0:38502 0:21171 0:31668 0:21697 0:13610 0:33025 0:11893 0:42112 0:23156 0:22839 0:26393 0:16556 0:40172 0:16879 0:32326 0:17775 0:17532 0:32367 0:16601 0:40281 0:16925 0:26194 0:28793 0:06658 0:23274 0:20954 0:24972 0:22317 0:10869 0:29905 0:12910 0:22351 0:19709 0:20271 0:28129 0:05614 0:27284 0:17613 0:26395 0:16693 0:16528 0:28286 0:09218 0:19242 0:20010 0:20564 0:25339 0:06827 0:29563 0:12479 0:22732 0:17509 0:41652 0:27501 0:03812 0:02936 0:06985 0:04612 0:20445 0:15747 0:37461 0:24734 0:25905 0:19953 0:47466 0:31340 0:07548 0:05814 0:13831 0:09132 0:26728 0:20586 0:48973 0:32334C C C C C C C C C C C C A Example 8.4 0:12829 0:87171 0:73035 0:26965 0:59343 0:40657 0:22022 0:77978 0:61452 0:38548 0:64521 0:35479C A 0:28793 0:06658 0:24972 0:22317 0:12910 0:22351 0:28129 0:05614 0:26395 0:16693 0:09218 0:19242C A 0:22732 0:17509 0:03812 0:02936C A A 3 =B @ 0:08032 0:54576 0:37392 0:37362 0:13794 0:48843 0:47284 0:32395 0:20321 0:15414 0:54576 0:30010 0:31752 0:19918 0:48330 0:47796 0:26282 0:25922C A C A Acknowledgement The authors wish to thank Professor J.-M. Bonnisseau for very useful discussions and two anonymous referees for valuable comments. --R A parametric approach to global optimization problems of a special kind Cutting angle method in global opti- mization Studies in Linear and Nonlinear Program- ming Global optimization of increasing positively homogeneous functions over the unit simplex Method of extremal basis Constructive nonsmooth analysis Mathematical Theory of Economic Dynamic and Equilibria Numerical Methods for Deterministic and Stochastic Minimax Prob- lems Foundation of Mathematical Optimization Economic equilibrium and optimum Directional derivatives of the solution of a parametric nonlinear program On some problems of non-smooth optimization in economic theory Reformulation of a problem of economic equilib- rium Kluwer Academic Pub- lishers Bilevel and multilevel programming: a bibliography review --TR --CTR A. M. Bagirov , A. M. Rubinov, Cutting Angle Method and a Local Search, Journal of Global Optimization, v.27 n.2-3, p.193-213, November
global optimization;exchange model;cutting angle method;marginal function;economic equilibrium
596473
Solving Standard Quadratic Optimization Problems via Linear, Semidefinite and Copositive Programming.
The problem of minimizing a (non-convex) quadratic function over the simplex (the standard quadratic optimization problem) has an exact convex reformulation as a copositive programming problem. In this paper we show how to approximate the optimal solution by approximating the cone of copositive matrices via systems of linear inequalities, and, more refined, linear matrix inequalities (LMI's). In particular, we show that our approach leads to a polynomial-time approximation scheme for the standard quadratic optimzation problem. This is an improvement on the previous complexity result by Nesterov who showed that a 2/3-approximation is always possible. Numerical examples from various applications are provided to illustrate our approach.
Introduction A standard quadratic optimization problem (standard QP) consists of nding global minimizers of a quadratic form over the standard simplex, i.e. we consider global optimization problems of the form x > Qx (1) where Q is an arbitrary symmetric n n matrix; a > denotes transposition; and is the standard simplex in the n-dimensional Euclidean space IR n , denotes the non-negative orthant in IR n . To avoid trivial cases, we assume throughout the paper that the objective is not constant over , which means that fQ; En g are linearly independent where ee > is the n n matrix consisting entirely of unit entries, so that x > on . For a review on standard QPs and its applications, see [3]. We only mention here that this problem is known to be NP-hard, and contains the max-clique problem in graphs as special case. Note that the minimizers of (1) remain the same if Q is replaced with Q+ En where is an arbitrary constant. So without loss of generality assume henceforth that all entries of Q are non-negative. Furthermore, the question of nding minimizers of a general quadratic function x > Ax can be homogenized considering the rank-two update ce > in (1) which has the same objective values on . ISDS, Universitat Wien, Austria. E-mail: immanuel.bomze@univie.ac.at y Faculty of Information Technology and Systems, Delft University of Technology, P.O. Box 5031, 2600 GA Delft, The Netherlands. E-mail: E.deKlerk@ITS.TUDelft.NL In this paper we will show how to derive approximation guarantees for this problem via semidenite programming (SDP). The main idea is as follows: we can give an exact reformulation of the standard quadratic optimization problem as a copositive programming problem, and subsequently approximate the copositive cone using either linear inequality systems, yielding LP relaxations; or, more rened, systems of linear matrix inequalities (LMI's), yielding an SDP formulation. Both SDP and copositive programming problems are examples of conic programming problems, and we begin by reviewing these concepts. 1.1 Preliminaries; conic programming We dene the following convex cones: The n n symmetric matrices: the n n symmetric positive semidenite matrices: the n n symmetric copositive matrices: the n n symmetric completely positive matrices: the n n symmetric nonnegative matrices: the n n symmetric doubly nonnegative matrices: We consider the usual inner product hX; Y i := Tr(XY ) on Sn and recall that the completely positive cone is the dual of the copositive cone, and that the nonnegative and semidenite cones are self-dual with respect to this inner product. Furthermore, the dual cone of Dn is D a cone which is contained in Cn and which will play an important role in the relaxations to follow. For a given cone K and its dual cone K we dene the primal and dual pair of conic linear programs: (D) d := sup y (2) If n we refer to semidenite programming, if to linear programming, and if copositive programming. The well-known conic duality theorem, see, e.g., [15], gives the duality relations between (P ) and (D). Theorem 1.1 (Conic duality theorem) If there exists an interior feasible solution X of (P ), and a feasible solution of (D) then p and the supremum in (D) is attained. Similarly, if there exist y feasible solution of (P ), then the inmum in (P ) is attained. As is well known [10], optimization over the cones S n and Nn can be done in polynomial time (to compute an -optimal solution), but copositive programming is NP-hard, as we will see in the next section. 1.2 Standard quadratic optimization via copositive programming In [4] it is shown that we can reformulate problem (1) as the copositive programming problem Problem (3) is called a copositive program because of its dual formulation (note that the optimal values of both (3) and (4) are attained and equal due to Theorem 1.1; see [4]). The reformulation makes it clear that copositive programming is not tractable (see e.g. [16, 4]). In fact, even the problem of determining whether a matrix is not copositive is NP-complete [9]. In [4], some ideas from interior point methods for semidenite programming are adapted for the copositive programming case, but convergence cannot be proved. The absence of a computable self- concordant barrier for this cone basically precludes the application of interior point methods to copos- itive programming. A solution to this problem was recently proposed by Parillo [12], who showed that one can approximate the copositive cone to a given accuracy by a suciently large set of linear matrix inequalities. In other words, each copositive programming problem can be approximated to a given accuracy by a suciently large SDP. Of course, the size of the SDP can be exponential in the size of the copositive program. In the next section we will review the approach of Parillo, and subsequently work out the implications for the copositive formulation of the general quadratic optimization problem by applying the approach of De Klerk and Pasechnik [5]. The basic idea is to replace the copositive cone in (4) by an approximation: either a polyhedral cone or a cone dened by linear matrix inequalities. In this way we obtain a tractable approximation problem. 2 Approximations of the copositive cone Since any y 2 IR n can be written as indicates the componentwise (Hadamard) product, we can represent the copositivity requirement for an (n n) symmetric matrix M as There are many possible representations of the polynomial P as a homogeneous polynomial of degree four, if we allow for nonzero coecients of terms like In particular, if we represent M is a symmetric matrix of order n+ 1n(n 1), then f M is not uniquely determined. The non-uniqueness follows from the identities: It is easy to see that the possible choices for f dene an ane space (see below for a closer description of that space). 2.1 Sum-of-squares decompositions Condition (5) will certainly hold if the polynomial P can be written as a sum of squares (s.o.s., in for some polynomial functions f sum of squares decomposition is possible if and only if a representation of P (x) exists where f M in (6) is positive semidenite. We will show this for any homogeneous polynomial of degree 2r below, but introduce some convenient notation beforehand: for any x 2 IR n and any multi-index m 2 IN n (with IN i the corresponding monomial of degree jmj. Also, denote by I n sg the set of all possible exponents of monomials of degree s (there are s of them) and, as usual, 2I n (s)g. Finally, given a set of multi-indices I , we dene [x] m2I as the vector with components x m for each m 2 I . Lemma 2.1 If P (x) is a homogeneous polynomial of degree 2s in n variables has a representation l for some polynomials f i (x) then there are polynomials h i (x) which are homogeneous of degree s for all i such that P has a s.o.s. representation as above if and only if there is a symmetric positive-semidenite matrix d d matrix f d such that s and e Proof: It is easy to see that the degree of f i is at most s for each i: if we assume to the contrary that f j has maximal degree of all f i and its degree exceeds s, then the square of its leading term which appears in the s.o.s. will never cancel out, since there can be no monomials of the same degree with negative coecients. We can therefore decompose each f i as f is homogeneous of degree s (or zero, but without loss of generality we assume that this happens only if t < i l, including the possibility of and the degree of g i is less than s. Now that g has degree less than 2s while h, as P itself, is homogeneous of degree 2s. Thus that any homogeneous h i can be written as h i a i a > d . The converse is obvious via spectral decomposition of f Next let us characterize all the matrices f which allow for a representation x >f Me x for a given homogeneous polynomial Lemma 2.2 Let 2m be a homogeneous polynomial of degree 2s in n variables d as in Lemma 2.1. Then only if (j;k)2[I f (j;k)2[I f Proof: Observe that f . The assertion now follows by equating the corresponding coecients of the polynomials e x >f Me x and Parillo showed [12] that P (x) in (5) allows a sum of squares decomposition if and only if M which is a well-known sucient condition for copositivity. For completeness of presentation, we give a new proof here, which follows easily from the preceding lemmas. Let us dene the cone K 0 n , the cone dual to that of all doubly nonnegative matrices. Theorem 2.1 (Parillo [12]) P allows for a polynomial s.o.s. if and only if Proof: In this case, the degree 2s of P equals four, so that Obviously, A 2e i of M . Therefore Lemma 2.2 yields f M is assumed to be symmetric. Note that we may and do assume that f M is positive-semidenite by Lemma 2.1. Now diagonal elements of f cannot be negative. Further, for all i; j, which means that S is a principal submatrix of the positive-semidenite matrix f M . Hence also S n , which shows the necessity part of the assertion. To establish suciency, observe that (x x) > S(x x) is, by spectral decomposition of S, even a s.o.s. in the variables z obviously a s.o.s. Hence (x x) > M(x x) is, as the sum of two s.o.s. decompositions, itself a s.o.s. 2 Higher order sucient conditions can be derived by considering the polynomial: and asking when P (r) (x) has a sum of squares decomposition. It is clear from Lemma 2.1 that the set of matrices M which satisfy this condition forms a convex cone. Denition 2.1 (De Klerk and Pasechnik [5]) The convex cone K r n consists of the matrices for which P (r) (x) in (10) allows a polynomial sum of squares decomposition. Obviously, these cones are contained in each other: K r n for all r. This follows from By explicitly calculating the coecients Am (M) of the homogeneous polynomial P (r) (x) of degree summarizing the above auxiliary results, we arrive at a characterization of K r n which has not appeared in the literature before. Theorem 2.2 Let n; r 2 IN and . Further, abbreviate m(i; and introduce the multinomial coecients For a symmetric matrix M 2 Sn , dene n if and only if there is a symmetric positive-semidenite dd matrix f d such that (j;k)2[I f (j;k)2[I f Proof: By the multinomial law, c(k)x 2k The last identity follows by setting Hence Am (M) as given by (12) are the coecients of P (r) , and the assertions follow by observing with the help of Lemma 2.1 and Lemma 2.2. 2 The following auxiliary result simplies the expressions Am (M) considerably. Lemma 2.3 Let M be an arbitrary n n matrix, and denote by diag the vector obtained by extracting the diagonal elements of M . If Am (M) is dened as in (12), then diag M for all m 2 I n Proof: Note that by denition, c(m(i; case i 6= j while even c(m(i; so that nonzero coecients of M ij occur only for some (i; pairs depending on m. Hence straightforward calculation shows, using which exactly corresponds to (15). 2 Observe that for we have, from m > Parillo [12] showed that M 2 K 1 n if the following system of linear matrix inequalities has a solution. where M (i) 2 Sn for The converse is also true: if M 2 K 1 n then the system of LMI's (17){(20) has a solution; this was used by De Klerk and Pasechnik [5] without giving a rigorous proof. We will now give a complete proof, by using our new characterizations of the cones K r n in Theorem 2.2 for 1. Note that in this case. We will use a shorthand notation where ijk as a subscript indicates the multi-index Theorem 2.3 M 2 K 1 n if and only if there are n symmetric nn matrices M (i) 2 Sn for such that the system of LMI's (17){(20) is satised. Proof: First assume that M 2 K 1 . By Theorem 2.2 there exists a f d satisfying (13) such that where e By (15), we have A iii n. Similarly, the left-hand sides of (13) read in case f f f Now put S (i) M ijj;ikk for all triples (ijk). Then S (i) since it is a principal submatrix of the positive-semidenite matrix f M . Hence setting M we see that condition (17) is satised. It remains to show that (18){(20) hold. Now and similarly ii 2S (i) whereas f f because the diagonal entries of f M cannot be negative. Thus we have constructed a solution to the system of LMI's (17){(20). Conversely, assume that a solution to (17){(20) is given. Observe that The rst sum is obviously a s.o.s., since M M (i) n for every i. The second sum can likewise be written as a s.o.s. because of i<j<k ii x 3 i<j<k rh where we have used the non-negativity condition (20) to obtain the last equality. Note that the rst two sums of the last expression vanish due to (18) and (19). Thus P (1) (x) is represented as a s.o.s. 2 By closer inspection of the preceding proof we see that the condition (17) can be relaxed, to arrive at a (seemingly) less restrictive system of LMI's, namely: Indeed, the rst sum in (22) is still a s.o.s., since M M (i) n for every i, and because of Theorem 2.1. Hence (24){(27) constitute an alternative characterization of K 1 n , which in the next section will turn out to be quite insightful. There we will also specify an (apparently) even more relaxed characterization of K 1 n , see (40)|(43) in Subsection 2.2 below. With slightly more eort, one could derive similar systems of LMI's for the cones K r 2. However, d then increases so rapidly with n (recall that that the resulting problems become too large for current SDP solvers | even for small values of n. We therefore change our perspective in the next subsection, to arrive at a series of LP approximations of the copositive cone. These approximations are weaker than the SDP ones, but can be solved more easily. 2.2 LP relaxations yielded by nonnegativity We start with a simple observation: If the polynomial P (r) (x) has only nonnegative coecients, then it is already allows a sum-of-squares decomposition. This motivates the following denition. Denition 2.2 (De Klerk and Pasechnik [5]) The convex cone C r n consists of the matrices for which P (r) (x) in (10) has no negative coecients. Hence for any r, we have C r Again, we obviously have C r n for all r. We can immediately derive a polyhedral representation of the cones C r this characterization has not appeared in the literature. Theorem 2.4 For any m 2 IR n , dene Diag m as the n n diagonal matrix containing m as its diagonal, i.e., satisfying diag (Diag m) = m. Then for all r 2 IN 0 and n 2 IN, Proof: Follows from (14) in Theorem 2.2 and Lemma 2.3, with the help of the basic relations Note that C 0 n if and only if M 2 Sn with This follows from Theorem 2.4 by the same arguments as in Theorem 2.3. We can also establish an alternative characterization of C 1 n similar to that in Theorem 2.3: Theorem 2.5 M 2 C 1 n if and only if if and only if there are n symmetric n n matrices M (i) 2 Sn such that the following system of linear inequalities has a solution: Proof: Suppose that M 2 C 1 n and dene N (i) jk as follows: Then N (i) 2 Sn and because of (28) and (29) we get N (i) 2 Nn . Further, M Finally, (30) implies (34) because M (i) contains three distinct elements due to the denition of N (i) . The converse follows as in the proof of Theorem 2.3, without taking square roots in (23). 2 By comparing (31){(34) to (24){(27) we see that merely S in (24) has been shrunk to Nn in (31). This re ects the fact that C 1 Further, the two equalities (32) and (33) can be replaced with inequalities, without changing the characterization: n if and only if there are n symmetric nn matrices M (i) 2 Sn for such that the following system of linear inequalities has a solution: Similarly, also M 2 K 1 n if and only if there are n symmetric n n matrices M (i) 2 Sn for such that the following system of LMI's has a solution: ii 0; Indeed, we may use non-negativity via (41), (42) and (43) in (23) to obtain the desired s.o.s. decomposition there, and an analoguous argument without taking square roots applies to establish suciency of (36)|(39). Every strictly copositive matrix M lies in some cone C r suciently large; this follows from a famous theorem of Polya [13]. In summary, we have the following theorem. Theorem 2.6 (De Klerk and Pasechnik [5]) Let n +Nn be strictly copositive. Then there are integers r K (M) and r C (M) with 1 r K (M) r C (M) < +1, such that for all r r K (M) while n , and similarly for all r r C (M) while n . The rst part of the theorem (concerning the cones K r already follows from arguments by Parillo [12]. In order to bound the number r C (M) from above we need the following result from De Klerk and Pasechnik [5], which in turn is a consequence of a result of Powers and Reznick [14]. Theorem 2.7 (De Klerk and Pasechnik [5]) If a symmetric (n n) matrix M is strictly copos- itive, the function has no negative coecients, and hence allows for a sum of squares decomposition, if r > L= 2; where and z > Mz: (45) In other words, if a strictly copositive matrix M satises L= 1 r for some integer r, where L and are respectively dened in (44) and (45), then M 2 C r equivalently, r K (M) r C (M) dL= 1e : Note that is a 'condition number' of M which can be arbitrarily small in general, and which is NP-hard to compute. 3 Approximation results In this section we consider families of LP and SDP approximations to p prove a bound on the quality of these approximations. 3.1 LP-based approximations Let us dene: for which has dual formulation Note that problem (47) is a relaxation of problem (4) where the copositive cone is approximated by n . It therefore follows that p (r) C p for all r. We now provide an alternative representation of p (r) C . This representation uses the following rational grid which approximates the standard simplex: r+2 I n (r fy A naive approximation of problem (1) would be The next theorem shows that there is a close link between p (r) C and the naive approximation p (r) . In particular, one can obtain p (r) C in a similar way as the naive approximation p (r) is obtained, i.e. by only doing function evaluations at points on the grid (r). Theorem 3.1 For any r 2 IN 0 consider the rational discretization (r) of the standard simplex from (48). If Q is an arbitrary symmetric n n matrix and q r+2 diag Q, then Proof: First we use the representation of C r n from Theorem 2.4, putting observing that from (15) and (16), diag Q Then by (47) we have diag Q which gives (49) by putting 3.2 SDP-based approximations Similarly as in the denition of p (r) C , we can dene SDP-based approximations to p using the cones n instead of C r for which has dual formulation It may not immediately be clear that the optimal values in (50) and (51) are attained and equal. However, this follows from Theorem 1.1 by noting that feasible in problem (50) as well as strictly completely positive and therefore in the interior of (K r similarly, feasible solution of (51) in the interior of K 0 n , and consequently in the interior of K r n for all Note that p (r) C for all n . 3.3 A bound on the quality of approximation We will show that the dierence p p (r) C can be bounded in terms of r and of the range Qmax Q min of the data, where we put Recall that we assumed linear independence of fQ; En g which implies in particular Q min p < Qmax (indeed, otherwise we got 0 z > (QmaxEn Q)z Qmax p which were only possible Theorem 3.2 Consider an instance of the standard quadratic optimization problem (1). For a given integer r 1, dene p (r) C as in (47), and Qmax and Q min as in (52). One has Proof: First note that the assertion is trivial in the exact case p (r) whence we assume that in the sequel. Further, as Q Q min En 2 Nn C r C Q min . Next, without loss of generality we may and do assume that p Qmin if the contrary were the case, then we are done because of If now p Qmin holds, then in particular necessarily p > Q min . Let us consider the function which is continuous, positive and strictly increasing on the interval [Q min while '(p) % +1 as p % p . Hence there is a p 2]Q min ; p [ such that pEn . We claim that M 2 C r . To this end, we proceed to bound the parameters and L in Theorem 2.7 for this matrix M : z z > Qz and also Hence and by Theorem 2.7, we obtain M 2 C r n . Invoking again (47), we deduce C < p . Now ' is increasing, thus '(p (r) Finally, the obvious relation Qmax Q min maxfQmax p (r) yields '(p (r) which in turn immediately implies the assertion via (54). 2 Finally we notice that a direct derivation of this approximation result via the alternative grid representation in Theorem 3.1 does not seem to be straightforward. 4 Comparison with known approximation results Consider the generic optimization problem: := for some nonempty, bounded convex set S, and let Denition 4.1 ([11]) A value approximates with relative accuracy 2 [0; 1] if We say approximates with relative accuracy 2 [0; 1] in the weak sense if The approximation is called implementable if . It is known from Bellare and Rogaway [1] that there is no -approximation of the optimal value of the problem for some 2 0; 1 in the weak sense, unless Using semidenite programming techniques, the problem can be approximated with the relative accuracy (1 O(1= log(n))), see [11] and the references therein. Note the the standard quadratic optimization problem (1) is a special case of this problem where we pass from minimization to maximization. The result in Theorem 3.2 can be restated as: p (r) C is a -approximation of p in the weak sense for r Qmax Q min 1: Thus p (r) C is only a polynomial-time -approximation of p for instances where Qmax Q 4.1 Special case: the maximal stable set problem E) be a simple graph and let (G) denote its stability number (cardinality of the largest stable set). It is known from Motzkin-Straus [8] (see also De Klerk and Pasechnik [5]) x where A is the adjacency matrix of G and I n denotes, as usual, the n n identity matrix. The stability number (G) cannot be approximated in polynomial time to within a factor jV j 1 2 for any > 0 unless or within a factor jV j 1 for any > 0 unless If we denote I n and dene C is dened in (46), then we may assume without loss of generality (r) 2 (assume G is not a complete graph). This means p (r) 2 and hence with and ' from (53), we get (p p (r) Now we apply the inequality (54), to arrive at(G) (r) (r) If we set then we can rewrite (56) as the equivalent inequality, by isolating (r) and subtracting (G): Hence we get (r) (G) < 1 which implies b (r) This was the main result by De Klerk and Pasechnik [5]; in other words, Theorem 3.2 generalizes this result. 5 Examples Here we give some examples for problems of various origin. Example 5.1 Consider an instance of the standard quadratic optimization problem (1), where the matrix Q is given by: This example corresponds to computation of the largest stable set in a pentagon (see Section 4.1 and [5]). We have p (0) 5 0:44721, and p (1) 1. The proof that p (1) 1requires the observation that the matrix is known to be in K 1 (but it is not in K 0 n +Nn ); for a proof, see [12]. Example 5.2 Consider an instance of the standard quadratic optimization problem (1), where the matrix Q is given by: This example corresponds to computation of the largest stable set in the complement of the graph dened by the skeleton of a dodecahedron (see Section 4.1 and [5]). Here p 0:309 but the approximation yield the trivial lower bound: p (1) This example shows that { even though the approximation p (1) is much more expensive to compute than p (1) (see Section { it can yield a much better lower bound on p . This example | rst considered in [5] | is the smallest problem we know where the p (1) approximation to p is not exact. It remains an open problem to nd the smallest value of n where the p (1) K approximation to p is not exact. Example 5.3 This example is from a mathematical model in population genetics [18], where, among others, the following matrix 410A is considered: and the objective is to maximize x > Qx subject to x 2 . There are 5 dierent local solutions to this problem. The globally optimal value here is 16 1which corresponds to changing to a minimization problem of the form (1) with a nonnegative coecient matrix, we obtain the upper bound 21 for the optimal value via computation of p (1) , while the approximation via computation of p (1) K is exact. Example 5.4 This example deals with portfolio optimization and is taken from Berkelaar et al. [2]. Here, x 2 corresponds to a portfolio: x i is the fraction of your capital to be invested in investment i. Given a portfolio x 2 there is a risk x > Qx associated with the portfolio which should be minimized, and an expected return r > x to be maximized. An example from [2] has data: 0:82 0:23 0:155 0:013 0:314 0:23 0:484 0:346 0:197 0:592 0:155 0:346 0:298 0:143 0:419 0:013 0:197 0:143 0:172 0:362 and (note that the matrix Q is not positive semidenite). We can formulate this (multi-objective) problem in the form (1) as follows: min x2 x > for some parameter c > 0 measuring risk-aversion; this problem is now of the form (1) if we set . For K 0:4839 and p (1) yields the objective value 0:3015, which suggests that the SDP relaxation is exact. 6 Numerical results We rst compare the LP approximation p (1) C of p with the stronger approximation p (1) K . To this end, we generated 1000 instances of (1) with and where the matrices Q are symmetric with entries uniformly distributed in [0; 1]. In Figure 1 we show a histogram with the 'distribution' of the ratios p (1) K for the 1000 test problems. Note that in about 3 4 of the cases the ratio is (close to) unity. This shows that the LP-based approximations are comparable to the SDP ones in this sense. However, the computation of C is much cheaper than that of p (1) K | in Table some typical solution times are given for growing values of n. In all cases the computer used was a Pentium II (450 MHz). The LP and SDP solver was SeDuMi [17] by Jos Sturm running under Matlab 5.3. K (sec) 25 4.12 not run 6.86 not run 11.2 not run 40 26.5 not run Table 1: Typical CPU times for solving the respective LP and SDP relaxations for growing n. The next set of experiments we performed was to approximate the stability number of random graphs by computing the p (1) and p (1) approximations to the optimal value of problem (55). For this purpose we generated 20 graphs on 12 vertices to have stability number 6. The edges outside the maximal of LP:SDP approximation Figure 1: Histogram showing the distribution of the ratio p (1) K approximation values for 1000 random instances of (1) with stable set were generated at random with the probability of including an edge between two given vertices outside the maximal stable set being 1. In all cases the p (1) approximation was exact while the p (1) approximation gave the trivial zero lower bound. 1 This indicates that Example 5.2 is quite special | it is dicult to nd a graph on 12 vertices where p (1) K does not give the stability number. It also shows that | although the p (1) C approximation is quite useful for random instances of (1) | it fails for most maximum stable set problems. Conclusions We have suggested a novel approach to approximating the optimal value of the standard quadratic optimization problem (1) | the problem (1) is rst rewritten as a copositive programming problem, and subsequently the copositive cone is approximated by using systems of linear or linear matrix inequalities. The resulting approximations are therefore LP's or SDP's, depending on which approximation scheme is used. Higher order approximations are also possible, where the resulting LP's and SDP's become larger, but the approximation is also provably better. In particular, we have quantied the quality of the approximation as a function of the order of approximation (Theorem 3.2). Moreover, we have presented numerical evidence showing the quality of the approximations. Acknowledgement The authors would like to thank Kees Roos for his comments on a draft version of this paper. 1 It is possible to see a priori that the p (1) approximation will be useless, by using Theorem 2.4; see also De Klerk and Pasechnik [5]. --R The complexity of approximating a nonlinear program. Sensitivity analysis in (degenerate) quadratic programming. On standard quadratic optimization problems. Approximation of the stability number of a graph via copositive programming. Improved approximation algorithms for maximum cut and satis Clique is hard to approximate within jV 1 Maxima for graphs and a new proof of a theorem of T Some NP-complete problems in quadratic and linear programming Interior point methods in convex programming: theory and applications. Nonconvex quadratic optimization. Structured Semide A new bound for P A mathematical view of interior-point methods in convex optimization Copositive relaxation for general quadratic programming. Using SeDuMi 1.02 Patterns of ESS's I. --TR --CTR Kurt M. Anstreicher , Samuel Burer, D.C. Versus Copositive Bounds for Standard QP, Journal of Global Optimization, v.33 n.2, p.299-312, October 2005 E. Klerk , D. V. Pasechnik, A linear programming reformulation of the standard quadratic optimization problem, Journal of Global Optimization, v.37 n.1, p.75-84, January 2007 Etienne de Klerk , Monique Laurent , Pablo A. Parrilo, A PTAS for the minimization of polynomials of fixed degree over the simplex, Theoretical Computer Science, v.361 n.2, p.210-225, 1 September 2006
linear matrix inequalities;approximation algorithms;stability number;standard quadratic optimization;semidefinite programming;copositive cone
596513
Active Management Framework for Distributed Multimedia Systems.
The successful deployment of next-generation distributed multimedia systems is significantly dependent on the efficient management support that improves the performance and reliability of such applications at run-time. In this paper, we describe an active management framework based on programmable monitoring agents and event-filter-action recursive model. Active management enables users to define reconfigurable and self-directed monitoring tasks that can be automatically customized at run-time in order to track the system behavior. Using this active management framework, the monitoring agents can be programmed to modify their monitoring tasks dynamically based on observed events information, and initiate the appropriate management actions accordingly. This paper also emphasizes the importance of active management for supporting a scalable, highly responsive and nonintrusive distributed management infrastructure. The presented framework, which is referred to as HiFi monitoring system, supports a comprehensive environment including code instrumentation, user subscription, agents administration, event filtering and action service. Examples of using HiFi in managing large-scale distributed multimedia systems are also shown.
Introduction The next-generation distributed multimedia systems are large-scale, resource intensive and highly dynamic. With the increasing demands of deploying large-scale distributed multi-media (LDM) systems, an efficient on-line monitoring and control has become an essential service for improving the performance and reliability of such complex applications. Examples of LDM systems include large-scale collaborative distance learning, video telecon- ferencing, distributed interactive simulation, and reliable multi-point applications. In an LDM environment, large numbers of events are generated by system components during their execution and interaction with external objects (e.g. users or processes). These events must be monitored to accurately determine the run-time behavior of an LDM system and to obtain status information that is required for management operations such as steering applications or performing a corrective action [7, 16]. However, the manner in which events are generated in an LDM system is complex and represents a number of challenges for an on-line monitoring system. Correlated events are generated concurrently and can occur from multiple locations distributed throughout the environment. This makes monitoring an intricate task and complicates the management decision process. Furthermore, the large number of entities and the geographical distribution inherent with LDM systems increases the challenge of addressing important issues, such as performance bottlenecks, scalability, and application perturbation. HiFi active monitoring system is an attempt to deliver a management architecture that explicitly addresses the challenges and requirements associated with managing large-scale distributed multimedia systems. HiFi active monitoring system supports dynamic and automatic customization of the management operations as a response to changes in LDM systems behavior [19, 24]. This is achieved through programmable monitoring agents that re-direct their monitoring activities on-the-fly upon users' requests and based on the information (events) collected during the monitoring operation. For instance, instead of monitoring all events and processes in the system, the agents monitor a subset of events/processes and the monitoring activities expand based on the information of the generated events. Therefore, active monitoring reduces the monitoring space significantly and offers a scalable management architecture. The monitoring intrusiveness is also minimized because this architecture enables initiating few monitoring tasks (targets) at the proper time. In addition, HiFi active monitoring enables the agents to react spontaneously (e.g., corrective actions) which improves the management operations response time compared with human- in-the-loop model [21]. A number of monitoring approaches and systems for monitoring distributed systems have been proposed (e.g., [5, 11, 15, 17, 21, 22, 23]). Although some of these systems provide mechanisms for modifying the monitoring requests dynamically, these mechanisms are manual (i.e., they require human intervention) and insufficient to support a programmable or self-directed monitoring tasks (actions) as described in this paper. In addition, they do not support a scalable and fine grain event filtering mechanism which is significant for monitoring "large-scale" distributed systems such as Internet-based applications. This paper is organized as follows: Section 2 explains the monitoring model and language specifications, Section 3 gives an overview of HiFi monitoring architecture and process, Section 4 describes our active monitoring approach and its impact on the management infrastructure, Section 5 provides an overview of the monitoring system implementation, Section 6 illustrates, by example, the effectiveness of using HiFi for steering distributed multi-point applications, Section 7 shows the evaluation of HiFi performance, scalability and system perturbation, and Section 8 presents the summary and concluding remarks. Monitoring Model In order to present a complete abstraction of the active monitoring architecture, our work must include modeling the application behavior, the monitoring demands, and the monitoring mechanism with considering the design objectives presented in Section 1. The program behavior can be expressed in a set of events revealed by the application during execution. In our monitoring model, we call the monitored programs event producers which continuously emit events that express the execution status. An event is a significant occurrence in a system that is represented by a notification message. A notification message typically contains information that captures event characteristics such as event type, event values, event generation time, event source, and state changes. Event signaling is the process of generating and reporting an event notification. We classify two types of events used in our model: primitive events which are based on a single notification message, and composite events which depend on more than one notification message. In the monitoring language, the event format (notification) is a variable sequence of event attributes determined by the user but it has a fixed header used in the monitoring process. An event attribute is a predicate that contains the attribute name which typically represents a variable in the producer (i.e, program) and a value. The event format also determines the type of event signaling: Immediate to forward the generated event immediately, or Delayed to allow buffering or batching events in the producer before sending them. Table I shows the High-level Event Specification Language (HESL) in BNF. This event abstraction enables consumers (1) to specify any arbitrary event format in a declarative way, and (2) to construct a complex (multi-level) abstraction of a program behavior using composite events. In addition, the event abstraction enables the consumers/users to assign values to the event attributes and does not require specifying attribute type (e.g., int, float or string). This provides a simpler interface than CORBA IDL [11]. We call the monitoring objects (e.g., human or software programs) event consumers since they receive and present the forwarded monitoring information. The consumers specify their monitoring demands via sending a filter program via the subscription process which configures the monitoring system accordingly (see Section 3). The monitoring operation is an event-demand-driven model. In other words, the producer behavior is observed based on the event generated (event-based) and on the monitoring requests (subscription-based). Therefore, as illustrated in Figure 1, events received in the monitoring system are classified based on exiting filters. If an event is detected, the action specified in the filter is performed such as forwarding the monitoring information to the corresponding consumers. The filter and action specifications are described in Section 4. We show in Figure 1 three management applications of this model: manual fault recovery that requires forwarding the monitoring information, automatic fault recovery that permit the monitoring system to perform actions on the event producers, and application steering in which the application acts as an event producer and consumer in order to tune its parameters. 3 Overview of HiFi Monitoring Architecture HiFi employs a hierarchical event filtering-based monitoring architecture to distribute the monitoring load in application environment. Based on a user's monitoring requests, the monitoring system determines the appropriate agent or set of agents within the hierarchy to be tasked with inspection and evaluation of application events. The monitoring system uses fine grain decomposition and allocation mechanisms to ensure that filtering tasks are efficiently distributed among the monitoring agents and prevent events propagation in the network. Since our focus in this paper is on the programmable monitoring environment in HiFi, we give a brief overview of the HiFi system and refer to [1, 3, 4] for more details. Hierarchical Monitoring Agents: In HiFi monitoring system, the task of detecting primitive and composite events is distributed among dedicated monitoring programs called monitoring agents (MA). MA is an application-level monitoring program that runs independently of other applications in the system and that communicates with the outside world (including producers and consumers) via message-passing. HiFi has two types of MAs: local monitoring agents (LMA), and domain monitoring agents (DMA) (see Figure 2). The former is responsible of detecting primitive events generated by local applications in the same machine while the latter is responsible of detecting composite events which are beyond the LMA scope of knowledge. One or more producer entities (i.e., processes) are connected to a local LMA in the same machine. Every group of LMAs related to one domain (geograph- ical or logical domain) is attached to one DMA. These DMAs are also connected to higher DMAs to form a hierarchical structure for exchanging the monitoring information. Because of the different roles of LMA and DMA, LMAs use Direct Acyclic Graph (DAG) [6], how- ever, DMAs use Petri Nets (PN) in order to record and track the event history [9]. Subscription Process: The monitoring process starts by a consumer sending a filter program that describes the monitoring request to the local MA. The filter is validated and decomposed into subfilters (e.g. F1, F2,.,Fn) using the decomposition algorithms in such a manner that each one represents a primitive event [1]. Then, each decomposed subfilter is assigned to one or more LMAs using the allocation algorithms based on the event sources and application distribution. The decomposition and allocation algorithms are described in [1]. The monitoring system also determines the proper DMAs for evaluating the event and the filter expression of the filter program. When MAs receive delegated monitoring tasks (subfilters) [10], they configure themselves accordingly by inserting this subfilter in the filtering internal representation which is a direct acyclic graph (DAG) for LMAs and Petri Nets (PN) for DMAs [4]. This architecture alleviates any performance bottlenecks or scalability problems by distributing the monitoring load among MAs and limiting the events' propagation to the originating sources [1]. 4 Techniques for Active Monitoring The main goal of active monitoring is to offer dynamically customizable monitoring tasks. This provides a flexible management infrastructure that scales well with number of producers and causes a minimal overhead in the application environment as shown in Section 7. In addition, active monitoring, in HiFi, reduces the monitoring latency because monitoring agents can customize their monitoring tasks automatically without users involvement. In this section, we describe the HiFi active monitoring architecture and its impact on improving scalability and reducing the perturbation of the monitoring system. 4.1 Filter-based Programmable Agents In HiFi, users (or event consumers) describe their monitoring demands via programs called filters submitted to the monitoring system at run-time. A filter is a set of predicates where each predicate is defined as a boolean-valued expression that returns true or false. Predicates may be joined by logical operators (such as AND and OR) to form an expression. In our model, the filter consists of three major components: the event expression which specifies the relation between the interesting event, filter expression which specifies the attributes value or the relation between the attributes of different events, and the action to be performed if both event and filters expressions are true. Table II shows the High-level Filter Specification Language (HFSL) in BNF. Consumers may add, modify or delete filters on-the-fly through the subscription component interface described in Section 5. When a consumer performs filter subscription, the monitoring agents reconfigure itself accordingly by updating their internal filtering representation [4]. Different consumers can potentially send different filters simultaneously. However, since the dynamic subscription may create an inconsistency in the monitoring environment, HiFi uses the subscription protocol, described in [3], to ensure an atomic state update and synchronization among the monitoring agents. As the event abstraction emphasizes the declarative aspect of the model, the filter-based programming abstraction enables consumers to describe the relation expression between different events and their attributes as well, which improves the expressive power and the usability of the monitoring language. This feature permits the agent to perform fine-grain filtering based on regular expressions. For example, assume agents have been configured through filter to detect warning events, AudioWarning and VidWarning, generated from Audio and Video processes, respectively. Consumers can limit the monitoring granularity by re-programming the MAs at run-time to detect only the event correlation between these two events such that they are both generated by the same machines via sending the following filter. [FORWARD];Warnings Correlation Filter. After decomposing and allocating this filter, the DMA will only forward Audio and Video warning events that are generated in the same machine. Consumers can also deactivate or modify an existing filter using DEL and MOD in the action part (see Table III). In addition, the programming environment permits consumers to overload the attributes values in the events in order to create a different event instance in the filter program. 4.2 Event Incarnation Actions in the monitoring model can be simply executing a program (local or remote) or forwarding the detected event to the corresponding consumers which are both necessary for automatic fault recovery and application steering, respectively. In order to improve the dynamism and the expressive power of the monitoring system, the model provides more complex actions: event and filter incarnation. In HiFi, generating new events as an action is called event incarnation in HiFi. This feature improves the performance, expressive power, and usability of the active management system as follows: ffl The event incarnation enables the consumer to activate a sequence of monitoring operations automatically using event-filter-action programming model (see Figure 1) without having the user to intervene in the monitoring process. This new event may trigger other filters which in turn causes performing further actions such as checking the status of other running processes. For example, a failure may occur in a producer (process) as a result of abnormal close of a communication channel (primitive event). In this case, the management operation involves failure recovery as well as sending an event that triggers a new filter to further diagnose the process that closed the connection. ffl An action could generate a "summary" event which summarizes the information of detecting a composite event (e.g., the event expression consists of multiple events). This enables suppressing the information of multiple events into one event (summary thereby avoiding event report implosion and reducing the event traffic. For example, a monitoring agent may be requested to generate a summary event that conveys the drop rate average of a set of receivers. We will present an application example of summary events in Section 6. ffl Performing an action such as executing a program may change the state of a running program. Therefore, sending an event that reveals the state change to the monitoring system is important to allow re-observing the behavior, and enabling automatic application steering. This significantly important for attaining system stability during the steering process as shown in Section 6. 4.3 Filter Incarnation In addition to the manual reconfiguration via dynamic users subscription described in Section 3, HiFi active monitoring also supports programmable agents that reconfigure themselves automatically based on events occurrences. A filter action can be a filter manipulation (typically, adding a new filter, deleting a filter, and modifying a filter). Thus, for example, another new filter can be activated in the monitoring environment as a result of detecting an event. We call this filter incarnation and shown in Figure 1. Filter incarnation is formally defined in Table III of the monitoring language. Adding a new filter means activating a pre-defined filter that has not been submitted to the system. This is specified in the monitoring language using a special reserved word (ADD) with the pre-defined filter name. On the other hand, deletion or modification must be performed on an existing filter that consumers subscribed to. This is specified using the reserved words, MOD and DEL, with an active filter name. When modifying an active filter, consumers must specify which parts to modify: event expression (EX), filter expression (FX), or both. This can be designated by appending the filter name as a prefix to EX and/or FX. The resulting EX and/or FX are the effective filter parts after the subscription is completed. Filter incarnation enables users to define "general" monitoring tasks that can be automatically customized by the agents at run-time in order to track and diagnose specific system behavior such as failures or performance bottlenecks. This avoids overwhelming the system by a large number of "static" (hardwired) monitoring tasks to observe all or most system activities. Consequently, consumers can initiate filters that can modify (change or expand) their monitoring scope in order to include other monitoring targets (events and processes) whenever certain events are detected. In the following, we describe various applications of filter incarnation in active monitoring. Adding/Deleting Filters for controlling monitor timing: Consumers can specify start and end times for any given monitoring activity based on events. In other words, consumers can specify to start/stop a monitoring activity when a certain event (primitive or composite) is detected. This minimizes the monitoring overhead and produces concise event traces. For example, assume a consumer wants to monitor the drop rate in the "receiving" events (RecvEvent) of bar program only when the transmission rate in the "transmission" event (TransEvent) of foo program drops below a certain threshold (STHRESHOLD). In this case, the consumer can specify a filter (MonitorSender) that monitors the "transmis- sion" events of foo which will trigger another filter (MonitorReceiver) that monitors the "receiving" events of bar if the transmission rate drops below the threshold. The filters example is shown below. [ADD MonitorReceiver]; MonintorSender: [F ORWARD]; MonintorReceiver: The MonitorReceiver filter monitors "receiving" events (RecvEvent) from foo and forwards them to consumers if their drop rate exceeds RTHRESHOLD. Similarly, the MonitorReceiver filter can be deactivated (deleted) based on the transrate value changes in the TransEvent. This permits activating MonitorReceiver filter at the proper time automatically and minimizing the monitoring perturbation in the application environment. Modifying Filters Specifications: Usually, monitoring tasks are static and defined prior to any monitoring operation. However, using filter incarnation, the filter information (e.g., attribute values) can be determined during the monitoring process itself based on the contents of the detected events. For this purpose, the HASL provides a set of virtual registers called filter registers that consumers can use for loading/restoring variables in/from monitoring agents. These registers are used by MAs to restore attribute values of received events. Consumers can simply assign the attribute value of an event used in EX or FX to a filter register and vice versa for this purpose. For example, consumers may want to monitor all events (trace) of the processes that have generated a security warning event (WarningEvent). In this case, the monitoring agents can not identify the module name information from the monitoring request (i.e., filter subscription). However, the monitoring agents can determine the module name during the monitoring operations. This is achieved by using the filer registers to save and restore the event information. In the following, the DynamicErrorTrace is a filter specification example that uses filter registers and filter incarnation to program the agent operations dynamically. In the following example, ThisMod is a filter register that restores the module name (ModuleName) after the occurrence of WarningEvent of type SECURITY. Then, the filer incarnation is used to modify the filter expression of TraceProcess filter in order to monitor/trace all events from this particular process/module only. MOD T raceP rocess:F TraceProcess: The TraceProcess is a "generic" filter that monitors and forwards all events from all modules to the corresponding consumers (Notice that "ANY" is a language keyword that indicates any string value). However, this general filter is customize by DynamicErrorTrace filter in order to perform a specific monitoring operation. As a result, this technique enables activating/deactivating the appropriate monitoring operations (or filters) at the right time and avoiding the overhead of launching multiple filters or monitoring requests simultaneously. It also reduces the monitoring latency since monitoring agents can be programmed to react spontaneously without consumers intervention. Moreover, the filter incarnation feature provides an extendible programming environment utilizing the power of the event-filter-action recursive model as shown in Section 6. 5 HiFi Monitoring System Implementation HiFi consists of four main components: Instrumentation, Subscription Service, Event Filtering and Control. Figure 3 shows the design of the monitoring system in component level and their interactions. Here, we provide an overview of the HiFi design components but detailed description is presented in [4] and [1]. Instrumentation Component: The process of inserting monitoring instructions inside the code of observed programs is called the instrumentation process. The instrumentation component utilizes the event information (HESL) supplied by the subscription component to construct low-level event formats called Event Reporting Criteria (ERC) that contains events information such as event location, reporting mode, event attributes, and the type of each attribute. The ERC is then used by the Event Reporting Stub (ERS) for constructing and reporting the event notification. The ERS is a HiFi library linked with the monitored application during compilation to facilitate event reporting. The main function of the instrumentation component is to facilitate the process of inserting the monitoring instruction (sensors) inside the program code. In many monitoring systems [17, 20], programmers write a considerable size of code for each generated event. This makes the instrumentation task tedious and error-prone. In HiFi, users only insert the user sense that specifies just the event name (e.g., ReportEvent(WarningEvent)). The instrumentation component then pre-processes the instrumented code and replaces user sensors with extended system sensors that contain all events related information such as machine name, process name, report mode and the attributes types and values. When the program starts executing, the control is transferred to the ERS which creates (i.e., forks) the LMA and initiates the agents organization protocol (described in [3]) in order to establish the agents' hierarchy described in Section 3. Subscription Service Component: The monitoring subscription is the process by which the consumers would express their monitoring demands represented in filter programs using HFSL. There are two major subcomponents in the subscription service: (1) Monitoring Language Processor which is used for validating, parsing, and constructing the monitoring- knowledge necessary for distributing events and filtering tasks, and (2) Monitoring Information Processor which is used for decomposing and allocating such information to generate monitoring tasks executable by the monitoring agents, and packaging and disseminating the produced information such as event and environment information to the agent networks using reliable multicasting. The subscription component is also responsible for receiving and presenting the event notifications forwarded from the MAs. Event Filtering Component: The event filtering component is the core component of the monitoring system and it constitutes the internal architecture of the monitoring agents, LMA or DMA. Its main functionality is (1) receiving and processing filtering tasks delegated from the subscription component after decomposition and allocation, and (2) inspecting in-coming events based on the event attributes and the filters information (i.e., filter internal representation) to determine if this event is interesting (detected) or irrelevant (rejected) [4]. This component operates on the event filtering internal representation that represent the monitoring information such as consumers' subscriptions and event specifications. The internal filtering representation can be Direct Acyclic Graph (DAG) or Petri Nets (PN). In particular, because of the different roles of LMA and DMA, LMAs use DAG, however, DMAs use PN in order to keep track of the event history. In other words, each filter is implemented in the monitoring agents as sequence of DAG and PN nodes. Control Component: The control component is provided to support reactive control management applications. The main function of this component is to perform the actions specified in a filter program. There are four types of actions supported by the monitoring architecture: program execution, information dissemination, event generation (incarnation) and filter incarnation (see Table III). As shown in Figure 3, the control component has two major subcomponents: dissemination service which uses multicasts information to the corresponding consumers, and action service which is used to execute local or remote programs, send a new event and/or perform filter incarnation (i.e., adding, modifying or deleting an existing filter). 6 Active Management for Steering Distributed Multi-point Applications The HiFi monitoring system is being used in a number of management applications such as application steering, fault recovery and debugging of distributed multimedia systems. In this section, we present an example of using HiFi for monitoring and steering Reliable Multicast Server (RMS) [2]. One of the known problems in some reliable multicasting, is the effect of slow members (e.g., machines) in group communication. A machine is described as a slow machine if its receiving rate is "much" less than the other members in the group. In this case, a slow machine could typically slow down the communication to the entire group because the sender transmission rate, in RMS, eventually adapts to the rate of the slowest receiver. Developing a solution for slow members in multicast groups is beyond the scope of this paper. However, this application example is to show the effectiveness of HiFi active monitoring approach in supporting a dynamic discovery mechanism of slow members (or machines) during a multicast session, and providing an automatic feedback to the RMS senders which make the proper steering management decision accordingly. The criteria of slow members are defined based on the user specifications. For example, the user (or manager) may define a slow member whose performance is below a certain threshold. In our example below, the RMS sender acts as a manager and sends the threshold information. Figure 4 shows the event (HESL) and the filter (HFSL) specifications used to discover slow members in multicast groups, and Figure 5 shows the Petri Nets of these filters as constructed in HiFi agents. Each RMS receiver is instrumented using HiFi instrumentation component to send McastRec event that contains the machine name, the domain name, the multicast group name, total bytes received (KBrec) so far, and number of NACKs scheduled (NackSch) which equals to the number of Nacks sent plus number of Nacks cancelled. Because of NACK suppression mechanism [8], the number of NackSch gives more accurate estimation of the drop rate than number of Nacks sent. The McastRec event is sent periodically based on time limit or maximum number of bytes received. And the RMS senders send McastSend events to indicate two things: the transmission rate (TransRate), and the drop rate threshold (threshold) for receivers in the group. Thus, the McastSend and McastRecieved events convey the status of the senders and the receivers, respectively. In order to activate this management operation at the proper time and avoid unnecessary use of resources, the slow members discovery process should be initiated only when the sender suffers some performance degradation (i.e., low transmission rate) due to the existence of slow members. This sender is called the unhappy sender and represented with US event in Figure 5. For this reason, the TransRate (in McastSend event) is first checked by the MonMcastSender filter (T1 transition in Figure 5) in order to identify the unhappy senders and their group names, and consequently activate other slow members discovery filters that monitor the receivers and update the threshold. If TransRate is found below the STHRESHOLD, then the Slow Members filter is modified to get the GrpName value stored in the filter register as described in Section 4, and both Slow Members and Update Threshold filter are activated is fired in Figure 5). The activated Slow Members filter compares the NackSch in McastRec and threshold in McastSend to identify slow members. However, because the threshold value is dynamic and may be determined from the overall performance of the participants, another filter (Update Threshold) is used to provide a feedback on the overall drop rate average to senders which consequently re-adjust the threshold value accordingly. Each LMA forwards McastSend and McastRec primitive events to its DMA that evaluates the filter expression upon receiving both events. The second filter, Slow Members, waits to receive one McastSend and McastRec events from all LMAs in the domain before the filter expression is evaluated. The ctr and LMAs are HiFi reserved key words and used to denote the number of the event occurrences and the number of LMAs in the domain, respectively. The filter expression evaluates to true if all RMS receivers in the domain send McastRec event from the indicated group name (GrpName) and the NackSch of one receiver or more is higher than the threshold. If the filter expression becomes true, then T2 (in Figure 5) fires and three actions are performed: (1) the average of scheduled Nacks for receivers in same domain is calculated by CalcAVG action, which fires T3 transition, (2) the DomAVG, which represents a summary event, is sent to the containing DMA to reveal the domain average, and (3) the McastRec event that matches the slow member criteria represented in the filter expression (i.e., NackSch ! threshold) is forwarded to the manager (RMS sender). The third filter, Update Threshold, receives the DomAVG events from the DMAs and then calculates the total NackSch average, updates the threshold and sends the McastSend with new threshold to the LMAs/DMAs again. This causes T4 and T5 transitions to fire and an event S (in Figure 5) to be sent with the new threshold. This filter can be a DMA task, instead of RMS senders. However, users must provide the action UpdateThreashold to the DMA which then can update the threshold and notify the RMS senders automatically. Since senders or receivers could be members in various multicast groups, the group name (GrpName) is used to limit this management activities (filters) on the multicast groups suffering from slow members. In addition, another filter can be used to de-activate Slow Members filter when the TransRate becomes less then STHRESHOLD. This way the slow members filter can activated and deactivated dynamically based on the condition of the multicast group. So in summary, three filters are used in this application: MonMcastSender that discovers unhappy senders and initiate the discovery process, Slow Members that identifies receivers below threshold, and Update Threshold that update the threshold value continuously base on the overall session performance. The slow members and NackSch average information are collected from each receiver via LMAs and then combined and propagated in hierarchical fashion via DMAs to the RMS of the sender. This mechanism is scalable because it avoids the notifications implosion that may occur when McastRec are forwarded to one RMS sender from group of receivers. Furthermore, distributing the processing load such as calculating the average drop rate contributes to the monitoring performance. 7 Performance Evaluation This section describes a performance evaluation study of the HiFi monitoring system. We conducted number of benchmarking and simulation experiments to assess the application perturbation, scalability and latency of HiFi. This section presents the numerical results as well as the conclusions obtained from this performance study. Application Perturbation Measurements: The application perturbation can be measured by the execution time overhead caused by the monitoring operations including the ERS event reporting process, and monitoring agents (LMAs and DMAs) operations. The first experiment is performed to evaluate the ERS overhead with respect to the event length. In order to show the effect of EventReport() function in ERS, used for constructing and reporting events, we compare its overhead with the traditional C printf() function which is frequently used by programmers as a simplest way for debugging and inspecting the program state and behavior. Figure 6 shows that the overhead of EventReport() is very comparable with printf() and ranges between 100 to 200 microseconds based on the event length. This figure also shows that ERS overhead grows slowly (about 2% per event attribute) when the event length is increased. We then measured the actual overhead caused by the monitoring system including reporting time (ERS processing), primitive event filtering time (LMAs Processing), event correlation time (DMAs Processing), UNIX socket communication and the RMS communication. In order to measure this experimentally, we use a filter correlation example called HelloWorld filter. In this experiment, two event emulation processes called Random Event Generator (REG) are used to generate up to 5000 events randomly using Bernoulli distribution. Both REG processes are located in different hosts (Sun Sparc 5 running Solaris 2.5) in the same Mbps Ethernet LAN. At each event time, REG process may choose to generate or not generate Hello or World events, which are 8 attributes long, with a probability of P . Each REG program is connected to an LMA residing in the same machine. The DMA sends a notification message to the manager if both Hello event and World event have the same sequence number (TStamp) but generated from two different REG processes (i.e., machines). In these experiments, REG programs are fist run without instrumentation or monitoring. Then programs are instrumented and run in HiFi environment to detect the event of interest. Figure 7-A depicts the results of this experiment with various event generation probabilities are developed to minimize the application perturbation including dynamic signaling and events batching. Using dynamic signaling, ERS only generates the events which are contained in the expressions of any existing filter. Other events which are defined but not used in any filter program are suppressed by ERS, even if they are invoked by event producers. The event batching is a mechanism to buffer more events before sending them and it is used if Delayed mode is selected in HESL as described in Section 2. To measure the effect of dynamic signaling, the REG programs were changed such that 50% of the generated events are filtered out by ERS. And to measure the impact of both dynamic signaling and event batching with maximum of 5 events, REG and ERS programs has been changed to reflect this effect. Figure 7-A shows a substantial improvement in reducing the application perturbation. These experiments also show a low perturbation figure, less than when P as low as 0.1 which what we believe a typical program generates in distributed environment. It is important to mention that agents communication primitives are the primary source of overhead in the monitoring operation [1]. Scalability Analysis: Our approach for evaluating the scalability of HiFi is by measuring the impact of increasing the event frequency and number of event producers on the mean response time or the monitoring latency. In this paper, we present the later case since the results are similar. The monitoring latency is the elapsed time between the event occurrence and the manager notification. In these experiments, we developed simulation routines to compare the Mean Response Time (MRT) [14] (monitoring latency) of hierarchical filtering approach with the centralized and decentralized monitoring approaches [5, 12, 13, 15, 17, 18, 22, 25]. The simulation routines assume that event arrivals (-) are exponentially distributed, and the average monitoring/filtering service rate (-) of an agent is 8000 events/second which was experimentally derived from a benchmarking experiments of LMA filtering. Therefore, the mean response time (MRT) for the centralized architecture is as follows: is the event frequency and N the number of event producers. In decentralized monitoring architecture, there are two levels of processing/filtering: in the producer agent, and in the centralized monitoring node. Thus, However, in HiFi which is hierarchical filtering-based monitoring, the latency is: dlog x (N)e such that x: branching factor and N : number of producers. The EGT and C can be neglected since it is a comparison study. Thus, the mean response time can be expressed as follows: The first factor represents the LMA filtering and the second represent the DMA filtering. Notice that the LMA receives only 0.5*f of the events since 50% of such event on average are filtered by ERS. Since not every event is necessarily forwarded up all the way in the DMA hierarchy, we calculate the MRT considering three different probabilities for forwarding an event: 10%; 50% and 90% of the events are forwarded. We also assume that there are 10 LMAs maximum are connected to one DMA in any domain (i.e., The simulation results of MRT verses number of producers of three approaches are depicted in and Figure 7-B. The event frequency (f) is considered 20 events per second. The figure shows the high superiority of the response time (latency) of the hierarchical architecture over the centralized and decentralized ones. The saturation points in the figures indicates a buffer overflow and indefinite response time since ae ? 1 in this case. The hierarchical architecture with probability 0.9 is still superior over the other architecture because of the use of dynamic signaling. This figure also shows that the MRT of hierarchical architecture grow slowly with respect of number of producers. In fact, the "jump points" in hierarchical graphs in Figure 7-B represent creating a new level in the hierarchy to accommodate additional producers. These figures also show that the centralized and decentralized approaches have a better MRT than the hierarchical approach when event frequency is low and very small number of event producers exist in the system. 8 Conclusion and Future Work This paper describes a novel active monitoring architecture (called HiFi) for distributed multimedia systems. The presented monitoring system organizes the monitoring agents in a hierarchical structure that distributes the monitoring load and limits the event propagation. The monitoring agents are programmable and can be reconfigured manually through users' interactions, or automatically by the agents based on the detected events. Users utilize a simple language interface, called filter, to define their monitoring demands and associated actions. Users can specify "general" monitoring tasks that can be customized dynamically by the monitoring agents in order to perform special monitoring tasks. We developed several techniques to support a programmable agents environment for active monitoring. This includes event incarnation to enable event-filter-action programming model, filter incarnation and filter registers to enable self-configurable monitoring operations, and dynamic subscription that enables users to add, delete or modify their requests at run-time. In this paper, we also demonstrate number of examples for using HiFi in monitoring and steering large-scale distributed multimedia systems. The active monitoring architecture offers significant advantages in the scalability and performance of the monitoring systems. It also enables the consumer to control the monitoring granularity, and thereby minimizing its intrusiveness. The scalability testing results shows an improvement of 37% and 29% (on average) of HiFi mean response time over centralized and decentralized architecture, respectively, with the increase of event producers. The primary source for these improvements is attributable to the programmable agents environment that permits distributing of monitoring load among the LMA and DMA groups and increasing concurrency of monitoring tasks. The HiFi active architecture also enables the localization of event filtering and classification in the area from which the events originate, which is necessary for reducing monitoring intrusiveness. HiFi's dynamic signaling inherent with the active monitoring, and event batching techniques significantly minimize the application perturbation to about 50%. Although HiFi was developed and used to monitor existing applications such as RMS, it remains a prototype monitoring system. Many open issues remain to be addressed by our research plan. These include: the integration of fault tolerance to make HiFi resilient to network and application failures, improved filter incarnation mechanisms to provide higher abstraction such that users can specify the ultimate monitoring target without having to specify intermediate monitoring tasks, and extended monitoring language and architecture that addresses temporal events. --R HiFi: A New Monitoring Architecture for Distributed System Management. Dynamic Monitoring Approach for Multi-point Multimedia Systems Adaptive Object-Oriented Filtering Framework for Event Management Applications High Speed and Robust Event Correlation. A Pattern-Based Packet Classifier Managing Multimedia Network Services. A Reliable Multicast Framework for Light-weight Sessions and Application Level Framing Detecting Composite Events in Active Database Systems Using Petri Nets. Network Management by Delegation - the MAD approach Object Management Group. Distributed Performance Monitoring: Methods The Arts of Computer Systems Performance Analysis. Monitoring Distributed Systems. A Generic Model for Fault Isolation in Integrated Management System. Tools for distributed Application Management. Debugging Concurrent Programs. The tempest: A practical framework for network programmability. An architecture for monitoring Network and Distributed System Management. Managing Communication Networks by Monitoring Databases. --TR --CTR Salvatore Gaglio , Luca Gatani , Giuseppe Re , Alfonso Urso, A Logical Architecture for Active Network Management, Journal of Network and Systems Management, v.14 n.1, p.127-146, March 2006 Al-Shaer , Yongning Tang, QoS Path Monitoring for Multicast Networks, Journal of Network and Systems Management, v.10 n.3, p.357-381, September 2002
multimedia systems;application steering;active networks;event correlation;distributed systems management
596514
Open Extensible Network Control.
Network control is decomposed in six parts: switch control, resource partitioning, virtual network building, virtual network control, generic services, and data-path components. Each of these parts can benefit from support for dynamically loadable code, which allows users to extend and customize the basic functionality. This is related to active networks, exept that dynamic code exercises control at the granularity of connections (flows), rather than individual packets and all aspects of network control are explicitly considered. Network resources are recursively partitionable, so that dynamic code is able to control partitions of virtual networks in any way it sees fit. Policing these partitions may occur at varying levels of "strictness".
Introduction Recent years have seen a flurry of papers on new directions in network control, either by opening up network control [1, 2], or by allowing packets on the data-path to program the network nodes, as proposed by the active networks community [3, 4]. Drawing from experience in both of these areas, we designed and implemented a framework that enables clients to program all aspects of network control in a controlled fashion. This is done by dynamically injecting code into the various entities that constitute the network control system, subject to access control restraints. This will be defined as elastic network control, an implementation of which is called Haboob 1 . 1 The term elastic was borrowed from [5]. 1.1 Contribution The Haboob not only opens up the interfaces used to control and manage networks, but also allows extension and modification of existing components that exercise such control. These components need not necessarily reside on routing/switching nodes, as is generally the case in active networks. While the key idea in the Haboob is also 'activating the network' (and as such could be placed under the active networks umbrella), the location and role played by the active code is different from that in most active networks (see also Section 3). As such, the Haboob is an attempt to bridge the gap between active networks and open control. Moreover, by now we consider the problem of making networks programmable by dynamically loadable code (or 'active packets') to be a well-established research area for which many solutions have been proposed. Although certain issues still need to be resolved, we are more interested in the next step: how best to exploit such programmability. For this purpose, an approach is taken at a slightly higher level than how active code is classified, demultiplexed, run by the appropriate execution environment, etc. However, it's also not our intention to discuss specific applications or services built with dynamically loadable code (an approach that was taken in [6, 7] and others). Instead, this paper presents both a framework and its implementation in which such questions are addressed as: which aspects of network control and management can be distinguished and how can they benefit from programmability by dynamically loadable code? It should be noted that the programming of network nodes is nothing new: switch vendors do it all the time and reprogramming happens whenever network operators update switch software by loading in new images. What makes work like the Haboob different from these traditional approaches is that it (1) allows reprogramming at a much shorter time scale, (2) without having to take the controller out of service and (3) with access control at a very fine granularity (instead of only allowing the system administrator to update the control software, it is possible to specify the amount of customisation of network control a client is allowed to do on a per-client basis). Although the implementation is ATM-based (combined with IP over ATM), the issues may be relevant to any network technology with end-to-end resource reservation. 1.2 Overview In Section 2, we clarify the terminology and overall view of network control and management maintained in this paper. In Section 3, the execution environment for dynamically loadable code is dis- cussed. Sections 4-6 discuss how such code is loaded into the heart of various components that make up network control and management. In Section 4, it will also be explained how resource partitioning and differentiation of such policies as policing and interoperability have been implemented in a specific control system. We will show how more traditional active networks are implemented in the Haboob in Section 7. As an application of dynamically loadable code a continuous media video server was built. This is discussed in Section 8. Some performance figures are given in Section 9. Related work is mentioned in Section 10. Finally, in Section 11, we summarise and draw conclusions. Henceforth, the term active networks as used in this paper, refers to such approaches as advocated in either the ANTS [3], or the SwitchWare [4] projects, i.e. where all or some packets on the data-path carry programs specifying how to deal with these or subsequent data packets. Using the term in this narrow sense, allows us to contrast it with the Haboob more easily. 2 Terminology and model We define control architecture as the set of protocols, policies and algorithms used to control a (partition of a) network. Traditionally, telecommunication networks' control architectures have been monolithic and closed, but recent research aimed at opening up control and management has achieved promising results [1], allowing, for example, multiple independent network operators to be present in the same physical network and a speedup of the time in which new services can be developed and deployed in the network. The goals of this approach are what we will call open network control, i.e. network control based on public interfaces that can be used by any party to implement new services, independent of vendor-specific solutions. We argue that most current implementations offer only limited openness, namely at the interface level. This paper shows that it is useful also to allow dynamic extensions to these interfaces, and customisation of the components that implement them. Only then all aspects of control and management are truly open. 2.1 Decomposition of control We identify six more or less orthogonal levels related to open network control. These levels have been implemented as separate components, allowing for independent development. The first level in the model considers the switch control interface, the interface between control software and physical switch. In traditional telecommunication systems this interface is closed, but it was shown in the Tempest [1] how an open, generic switch control interface called Ariel allows for non-proprietary network control systems, i.e. that work with any switch that implements the interface. Ariel is sufficiently simple to be implementable for even the simplest switches. Similar interfaces have been proposed elsewhere, e.g. by the IEEE P1520 Standards Initiative [8]. The second level comprises the partitioning of network resources by a process called the divider. The divider is capable of making hard partitions on a switch where each of the partitions may be controlled by a different control architecture. These partitions, known as switchlets, all have their own VCI space on switchports and other resources as well. In fact, to the software that is controlling the switchlet, it seems as if it is controlling its own (albeit smaller) switch, because the divider exports an Ariel interface for each of the switchlets. Switchlets allow a number of different control architectures to be active simultaneously on the same physical switch. At the third level we find the netbuilder, the process responsible for combining switchlets to create virtual networks. Again, such virtual networks are hard partitions of the network resources and each virtual network looks exactly like a somewhat smaller physical network. The fourth level, considers the control architecture. This is the software that sets up and tears down connections, and in short, actually controls the virtual network. Figure 1 shows three control architectures that are simultaneously active on two switches. One of them controls two switchlets on two switches. Whatever happens in these switchlets will never interfere with the other two switchlets: the operations of each control architecture are limited to its own partition of the resources. Ariel Ariel Ariel Ariel Ariel Ariel Ariel Ariel Switch Divider Switch Divider Manager Manager RSVP Experimental Control Architecture P-NNI ATMF P-NNI ATMF The manager interface is used by the netbuilder to create switchlets and other switchlet management tasks Physical Switch Figure 1: Multiple control architectures in a single physical network, partitioned by dividers The general sequence of events in our implementation of this model is as follows. At start of day, the control architecture requests a virtual network from the netbuilder. The netbuilder in turn requests the corresponding switch dividers to create the appropriate switchlets on the switches, making sure that the virtual network is sensible (e.g. that the VPI/VCI ranges on neighbouring switchlets overlap, etc. The dividers create the switchlets and return switchlet Ariel interface references to the netbuilder, which returns them to the control architecture. The control architecture subsequently makes calls directly over the switchlets' Ariel interfaces to set up and tear down connections, etc. Finally, the control architecture makes a call to the netbuilder to destroy the virtual network. The fifth level in our model concerns generic services: processes providing functionality that may be used by a wide variety of clients. An example of generic services is a federated trader. A trader is used by servers to export offers for their services, which can subsequently be imported and bound to by clients. Another example of generic servers is the traffic server, discussed in Section 4.1.5, which provides applications flexible access to measurements taken from a switch. The sixth and last level comprises data-path components, i.e. components that reside on the data-path and participate in network control. Examples include filters, transcoders, and other datapro- cessing engines. Unlike such active networks approaches as ANTS [9] and SwitchWare [4], we treat data-path components separately from the other components. Normally, there will be no interference with the data-path at all. Only in the (exceptional) event that data processing is needed, will data-path components be instantiated. An example is given in Section 7. We now have a complete decomposition of network control and management. The model can easily be extended if necessary in the future. The question that we ask ourselves is: how can each of these levels benefit from dynamically loadable code and what are the issues? This paper argues that it is useful to provide such dynamic customisability in all of these levels. 2.2 Lack of flexibility Open control such as described above (and in projects such as xbind [10]) allows for a speedup in the development and introduction of new services compared to traditional control of communication systems. However, so far the functionality of each of the levels in the model is still fixed (specific problems resulting from this rigidness will be discussed throughout the text). This paper argues that it is useful to allow clients to extend and override the functionality of all levels of network control and management in a controlled way. Furthermore, although in principle it is possible to build and use the ideal control architecture for every application that has specific network needs, it seems unlikely in practice that there will be more than a few of these control architectures active at the same time. The reason is that rather than implementing a special-purpose control architecture, most application writers will be content with using an existing control architecture that is closest in functionality to what the application really needs. Allowing clients to customise existing network control solutions allows them to provide this extra functionality themselves. 3 Loadable code The active networks community states that instead of standardising the computation performed on every packet, we should standardise the computational model [3]. No distinction is made between control and data. Although we agree with the statement in principle, we think that it should be extended explicitly to the control plane. As mentioned above, the Haboob treats processing on the data-path as the exception rather than the rule and therefore the role played by customised computation on individual data packets at every node in the network is relatively minor. Although we do allow such customised computation, as will be shown in Section 7, its applicability is considered to be rare, as most packets will probably want to be switched at the highest possible speed (i.e. without interference by application-specific computation). For the control plane, however, the inverse is true. As the execution of control code has no direct impact on the speed at which the data travels, the control plane is the ideal location to allow application-specific code to implement new services. The granularity of control by dynamic code is the connection (or flow) rather than individual packets that make up these flows (except for data-path components). The active networks' approach is to treat every packet on the data-path as a program (or capsule) as in [3], or at least allow some of the packets to carry program code as in [4]. Instead, we propose not to interfere with the data path at all. Also, if the network technology allows the separation of control from data, this allows important signalling messages to get through even in the face of congestion. These messages may contain the code (in the form of DLAs) to deal with the congestion. This separation is a standard feature of ATM, but requires some work in technologies like IP (e.g. different queues for these two classes of traffic). Fortunately, quite some research is currently directed towards achieving this, e.g. IntServ [11]. Moreover, instead of simply allowing clients to push code into routers, we consider each of the tasks in Section 2. In other words, we address such questions as: how can we allow clients to extend the functionality of the switch interface, who should we allow to do this, what does it mean when we allow clients to program a control architecture and to repartition resources, how much freedom do we give them when controlling these resources (e.g. with respect to policing), how can we permit an application A to specify the way in which interoperability between its own control architecture and neighbouring control architectures should be handled (for A's flows only and not for all others), etc. We will show that this approach is also different from what is commonly known as intelligent networks (INs) [12], in that the code can be both active and passive (e.g. waiting for a trigger). Additionally, elastic control operates at all levels of network control, rather than just the call state. Another way of saying this is that INs are easy to implement in the Haboob, while an inverse mapping is not possible. 3.1 The Sandbox The elastic behaviour starts with the instantiation of an execution environment for the DLAs. This environment will be called the Sandbox and is implemented as a single object. The Sandbox is language and implementation-independent and only its interfaces are specified. Apart from the default language support in a Sandbox, functionality can be added in the form of modules. A full implementation of the Sandbox comes with at least one module, which implements the simple uniform framework for interaction (SUFI). The SUFI provides operations that allow DLAs in different Sandboxes to communicate, whilst preserving location transparency, either via remote evaluation [13], or via more traditional remote procedure calls (RPCs). These SUFI operations are extremely simple. For example, parameters in remote evaluations are always treated as byte sequences. The SUFI relies on adaptors to do the marshalling, unmarshalling and type checking of the parameters at runtime. In the case of RPC, adaptors could be generated from interfaces automatically and prepended to a remote evaluation call that encapsulates the RPC (as a remote evaluation). This is currently not done. The SUFI also provides operations to the DLAs for: ffl exporting a DLA's interfaces; access control to a DLA's operations (currently based on capabilities); ffl security: for example, it provides a method for establishing trust using a trusted third party; ffl reflection: DLAs can find out information about the Sandbox and also which interfaces and operations it supports and how to use them; ffl threading and semaphores for intra-DLA interaction. It is possible to register new modules with the Sandbox, the result of which is that the commands contained in the module's interface are made available to the DLAs. The remote evaluation paradigm is quite powerful and allows clients to extend the functionality of a Sandbox. It works as follows. A client with the appropriate access rights presents a bit of code, called a granule, to a remote Sandbox for evaluation. Granules may be active or passive. For example, an active granule may iterate over a switch's management information base to see whether particular connections exist and send back the result. Passive granules may simply install the function for doing the iteration; the code will not be activated until it is explicitly called. DLAs can be thought of as being made up of granules. Granules are defined as fragments of code that can be autonomously evaluated. Remote evaluation of granules, combined with the ability to destroy granules, provides a way to implement replicating and migrating DLAs. Whether or not the state is automatically sent as part of the migration is beyond the scope of the SUFI. This depends on the language environment. The current implementations require programmers to explicitly gather the state that they would like to use at the remote Sandbox. Note that granules and migration also allows for disintegration and integration of DLAs. We speak of disintegration when a DLA is decomposed in a number of granules which migrate to different sites ("spreading out" the DLA). Integration is the opposite effect. 3.2 Implementation Sandboxes have been implemented for Tcl 2 and Java, while work on Python is underway. In each case, the underlying Sandbox for the language was implemented as a single C++ class, so it can be easily incorporated in existing code that was not written with elastic functionality in mind. Note that remote evaluation of Java granules by a Tcl Sandbox is bound to fail. Instead, the SUFI demands that in the case of different language environments, RPC can still be used for interaction according to a traditional client/server paradigm. Quite some thought has been given to safety and security. Careful shielding, strict access control, trust establishment and controlled access to resources are part of the full Sandbox implementation. Controlled access to resources should also prevent DLAs from using too much CPU time, memory, etc. This can be enforced by using an operating system such as Nemesis [14], which allows one to partition resources in the operating system itself. The current implementation, however, still uses Solaris Unix, which is not capable of strict resource control. A thorough discussion of security issues is beyond the scope of this paper. 4 Elastic control architectures There are many ways to build a control architecture, each of which may be useful in certain envi- ronments, but not in others. Another way of saying this is that if control architectures represent the boundary between client applications and the network and its primitives may be considered the API for dealing with the network, this API is tailored to a certain class of applications. Even so, we have made an attempt at a general solution for control architectures in the Noman class of control archi- tectures. Noman will be discussed in Section 4.2. First we will deal with some aspects of a specific control architecture called the Sandman, as it demonstrates some of the power of elastic control. 4.1 Sandman Sandman is a distributed control architecture, built on top of a CORBA compliant DPE. Figure 2 shows the components that make up the basic Sandman. An hierarchical trading mechanism allows clients to find the control architecture and other services. At start of day, the Sandman obtains a virtual network from the netbuilder in the form of interface references for switchlets. It performs control operations, such as connection setup, teardown, etc., by calls over the Ariel interfaces of these switchlets. The important control decisions are made by the connection manager and its support 2 a parameter supplied to the constructor decides whether to use Safe-Tcl instead of standard Tcl modules, such as call admission control (CAC), routing, etc. The switch fabric components are responsible for actual connection setup across individual switches, while the local host managers represent the local interface to the control architecture. Clients make requests to the host manager, which tries to reserve the local resources and if successful, forwards the request to the connection manager. Once a connection has been successfully established the host manager calls back the client, providing it with information about the established connection (e.g. the VPI/VCI pair in the ATM implementation). As we shall see, the Sandman has strong support for advance reservations. Endpoint CAC Router . Manager Local Host Connection Local Host Manager Endpoint Switch Fabric Component Switch Fabric Component Manager Figure 2: Components of a control architecture 4.1.1 Basic operations In its default configuration, the Sandman offers only a few commonly used operations that are expected to be sufficient for a large class of applications: 1. Unicast connection. The simplest and probably most common operation is the connection from source to one or more sinks for a particular time interval with particular characteristics (e.g. a peak rate of 1000 cells/s). A clients submits such a request to the control architecture and if the call admission control (CAC) module accepts the request, the client is guaranteed that the connection will be set up in that time interval. Traditional, immediate, connections simply leave out the interval, in which case it defaults to [now; 1i. 2. Multi-source, multi-sink connections. There is also a small number of more complicated types of connections, such as a connection that is time-shared by multiple sources and which may have multiple sinks each with its own and possibly overlapping time interval. For more details, see [15]. 3. Information gathering. This is a rather wide-ranging class of operations to discover certain things about the state of the network, the topology, routes, available bandwidth on a certain switch port, etc. 4. Loading application-specific code. Allowing users or applications to load their own code into the Sandman allows them to exploit application-specific knowledge at a very low level. We will discuss loadable code in Section 4.1.3. 5. Reservation of arbitrary sets of resources. The reservation of arbitrary sets resources by partitioning (and repartitioning) existing sets is described in Section 4.1.4. The first two of these operations allow for reservation in advance, so that guarantees about the availability of the required resources at some time in the future can be given. These are probably also the most commonly used operations that are expected to be sufficient for the majority of applications. We call these the primary operations. All other operations mentioned above are called secondary operations. 4.1.2 Call admission control Admission control in the Sandman is based on the conviction that it is prohibitively difficult to accurately characterise the behaviour of unknown sources. This means that statistical multiplexing cannot be derived from careful analysis of the supplied traffic descriptors. Instead, we base our CAC decision on a new connection's supplied peak rate only. To alleviate this potentially over-conservative reservation mechanism, we extend it with measurement-based effective bandwidth estimation, based on [16]. For this purpose, we introduce traffic servers. Traffic servers take continuous measurements of the traffic on the ports of a switch and compute the effective bandwidth of connections belonging to the Sandman's virtual network. This effective bandwidth is subsequently used for CAC decisions. In other words, for unknown sources, CAC is based on peak rate alone, but once connections become active an estimation is made of how much bandwidth they really need and this is used for any further CAC decisions. Traffic servers may vary from switch to switch and can be plugged in on the fly. It is conceivable that traffic servers are plugged in for certain switches and indeed certain switch ports, but not for others (in which case the default conservative CAC is used). A more detailed explanation of the Sandman's CAC can be found in [15]. 4.1.3 Dynamic code loading Dynamic code loading is supported by the instantiation of Sandboxes in the control architecture into which clients can load their DLAs. The DLAs have access to all primary and secondary operations. Observe that this does not really alter the functionality of the Sandman. DLAs simply act as normal Sandman clients that have moved into the Sandman's address space, so all it potentially gains is an improvement in performance. Next, we will demonstrate how the functionality can be easily extended using the same principle. It is important to make a distinction between an application's resource manipulation behaviour and its resource reservation behaviour. Resource manipulation behaviour is defined by the application's actions and operations on resources under the implicit assumption that the resources are available for it to use. This type of behaviour includes allocating resources to connections, freeing a connection's resources, etc. In resource manipulation behaviour, applications generally don't worry about the availability of the resources. Resource reservation behaviour on the other hand concerns itself solely with ensuring that certain resources are available at certain times. In general, it does not care what the resources are used for (or whether they are used at all) 3 . Due to space limitations, this paper focuses on how DLAs can be loaded that specify mainly a client's resource manipulation. However, the Sandman also gives applications complete freedom in specifying their reservation behaviour, again in the form of DLAs. For example, it is possible to specify that every day from 9am to 5pm a network should be reserved with a capacity equal to f(t) bps, where f(t) may be as arbitrary as "the number of seconds remaining until the year 2000". Note that such DLAs give rise to issues to do with scheduling; for example because a resource reservation DLA has to be evaluated each time a new reservation is made for any of the resources associated with it. In practice therefore, one will probably want to restrict both the amount of resources (e.g. CPU time) that a DLA can use and the expressiveness of the language for this type of DLAs. 4.1.4 Recursively partitioning networks The primary operations are expected to be sufficiently expressive for a large number of applications. Some applications, however, have specific needs, so in order not to restrict them, we propose to give these applications a number of resources which are theirs to use as they please (i.e. without any connections imposed on them). For this purpose, the Sandman includes the possibility to make (possibly advance) reservations for what we have called netlets, i.e. small virtual networks in a larger virtual 3 Note that many of the basic operations, such as to set up end-to-end connections incorporate both reservation and manipulation behaviour network. Netlets consist of (a share of) an arbitrary number of resources within the encompassing virtual network. The Sandman exports low-level primitives to control the individual resources of a netlet. For example, it is possible to create arbitrary connections across a single switch (assuming the resources for this connection are part of the netlet). This is quite different from traditional control primitives in communication systems, which allow clients to set up connections between endpoints only. We call these low-level primitives the Sandman's tertiary operations. Tertiary operations are currently only accessible to DLAs. It is possible to associate a DLA with a netlet, which means that DLAs can control the netlet's resources at any level of granularity it chooses. It can implement application-specific functionality in an efficient manner (all communication takes place in the same address space rather than across the network). In this way, DLAs can also implement new functions (e.g. to create up a new type of connection), which it subsequently exports. From then onwards, external clients with the appropriate access rights, can call these functions, as if they are part of the normal control archi- tecture's repertoire. This is one of the things that distinguish netlets from normal virtual networks (and from such parent-child virtual networks as discussed in Genesis [17]). The interfaces and DLAs are illustrated in Figure 3. load DLA TERTIARY OPERATIONS PRIMARY OPERATIONS Figure 3: DLAs in the Sandman Netlets can be created recursively. In other words, it is possible to create netlets in netlets, which enables applications to repartition network resources in an almost unrestricted manner. In fact, the encompassing virtual network of Section 2 itself is really a netlet (the so-called level-0 netlet). Repartitioning network resources extends the idea of switchlets into the control architecture. Additionally, netlets present new opportunities for specialising policies in virtual networks. As examples, we will consider policing and interoperability. 4.1.5 Application-specific policing netlets must be policed, because misbehaviour in one level-0 netlet N should not affect connections in any of the other level-0 netlets. Given this policing, however, we can decide not to police at the netlet level (or police more loosely), for example because we trust the applications in the netlet to behave. The reason for this is that even if connections in a netlet misbehave, the problems will be limited to N only and not propagate to the outside world. Netlets therefore allow us to differentiate the policing policy in the network. Given hard (in-band) policing for level-0 networks, we can now decide to police specific higher-level netlets only very loosely (e.g. by periodically taking measurements from switches to see if they have exceeded their allocated bandwidths) and certain other netlets not at all. In fact, the looseness may vary from netlet to netlet 4 . For loose policing, we modified the traffic servers of Section 4.1.2 by extending it with a Sandbox. A client of the traffic server (such as the Sandman) is able to load DLAs into this Sandbox, which determines in what sort of statistics the client is interested. For example, it is possible to load code that specifies (1) that a certain netlet should be checked every second to see if its connections stay within the capacity allocated to the netlet, and (2) that the control architecture be informed if this is not the case. This is an example of loose, or a posteriori, policing. Elastic traffic servers allow clients to program exactly what sort of information they receive about the measured traffic. In other words, the elasticity makes them much more general-purpose than the simple CAC tools introduced in Section 4.1.2. Such traffic servers are examples of elastic generic services. 4.1.6 Interoperability Interoperability between control architectures is extremely important in an environment where multiple control architectures are expected to coexist. Existing solutions simply map the primitives of one control architecture to their nearest match in the neighbouring control architecture. For example, suppose that two endpoints A and B in neighbouring control architectures CAA and CAB need to communicate and that CAA supports advance reservations and CAB doesn't. In all probability, an advance reservation request from A in CAA will then be mapped onto a direct reservation in CAB , so the availability of the resources at the time when they are needed is ensured. We call this the simple hop-by-hop solution. There are two problems with hop-by-hop interoperability. The first problem concerns the functionality degradation at the domain boundaries. This is a serious problem if two feature-rich control architectures communicate via a third control architecture that has a less sophisticated API. Following through the example of advance reservations, if CAA and CAB both support advance reservations, but some intermediate control architecture CAC (located between CAA and CAB ) doesn't, an advance reservation request initiated in CAA will be translated to an immediate reservation in CAC and will never be "upgraded" to an advance reservation again. In other words, the functionality degrades as domain boundaries are crossed. Our solution involves the establishment of 4 In other words, netlets are light-weight virtual networks. In this sense the relation between a higher-level netlet and a virtual network is similar to that between a thread and process. explicit inter-domain signalling channels between CAA and CAB using the simple hop-by-hop solution. The inter-domain signalling channel is used to carry higher-level control messages between feature-rich control architectures. The control messages are simply forwarded by CAC (tunnelling). A complete description of inter-domain signalling channels is given in [18]. The second problem concerns the rigidness of existing interoperability solutions. For example, in the above it was assumed that an advance reservation is always translated to an immediate reservation in CAC . This may be the right solution in some cases, but it is not necessarily the best solution in all cases. Rather than making an immediate reservation to emulate an advance reservation, pinning down resources for a very long time, an application may prefer not to prereserve in CAC at all and simply hope for the best, i.e. only try to reserve the resources when they are needed and accept the consequences if this fails. This would be a very good solution if the application knows that bandwidth will not be a problem in CAB in the requested time interval. The point is that only applications know which mapping is best for them. To overcome the problem of fixed primitive mapping, the Sandman allows applications to load in their own specific mapping policies which overrule the default mapping. Observe that this is not without risks. For example, we must guard against an application overriding the default operation mapping with its own application-specific mapping (which may well be faulty), thereby affecting the interoperability of all applications. This is where netlets are useful: we restrict the application-specific mapping to a specific netlet. Other netlets or connections outside the netlet are not affected by the new policies at all. Moreover, DLAs can be employed to extend the functionality of the control architecture, by creating innovative connection types with the help of netlets. These functions can then be exported and called from external clients. It is now possible to specify the interoperability of particular new connection types in application-specific manners, which would be hard to do without netlets. 4.2 Noman control architectures The Sandman is an example of a home-grown control architecture. It is useful in environments where such things as resource repartitioning (netlets) with DLAs, or advance reservations are required, but maybe not in many others. It is very difficult to solve the control architecture problem in a general way, without referring to a context. Noman, our attempt at finding a general solution, therefore, does not aim at designing the ultimate control architecture with functionality that will satisfy all. In fact, Noman offers no control architecture functionality at all, instead providing clients with the means to implement their own functionality. Noman consists of a Sandbox which is extended with a network control module, known as the netcontrol module. The module provides DLAs with a simple API which allows them to build, expand, control and destroy virtual networks. So on top of the Noman, a DLA can be loaded which contacts the netbuilder and creates a virtual network, consisting of a set of switchlets. The DLA obtains handles on the interface references for the switchlets that are returned as a result of this. These handles can subsequently be used to invoke Ariel operations on the switchlets, e.g. to set up or tear down connections. When the DLA is done with the virtual network, it calls an operation in the netcontrol module to remove it. In other words, Noman is not really a control architecture at all. It is merely a component that allows rapid development of new Noman-based control architectures. The dynamically loadable control architectures will be called Noman control architectures. Using Noman, clients are able to upload their favourite control architecture functionality ("dial a control architecture") and even modify, or add to it, on the fly. This is attractive for at least two reasons. First, it is possible to generate libraries of various control architecture components (e.g. routing, connection setup and CAC components), out of which components can be picked and added to the Noman in a mix-and-match fashion, in order to create the control architecture of choice. This is related to the way components are bound together in xbind [2], but is slightly more general as components can be customised and replaced on the fly. Second, it allows for vertical integration. For example, it is possible to load the application on top of a set of Noman components to make an application manage its own resources. Application and control architecture tasks are then integrated in the same address space, resulting in considerable performance improvement. The underlying Noman components could be shared by a number of applications or they could belong exclusively to a single application, in order to prevent something similar to what in the context of operating systems is called crosstalk in shared components [14]. 5 Elastic dividers and switch interfaces It is the switch divider's task to partition resources on a switch. In the simplest case, this means that each partition (switchlet) obtains its own VPI/VCI space (and possibly some other simple resources as well). It interacts with the outside world via two types of interface: (1) the management interface, which provides operations for the creation or destruction of switchlets, and (2) Ariel interfaces (one per switchlet) to perform control operations. There are therefore also two places where the instantiation of Sandboxes would be useful. The first is in the management component where new switchlets are created. The second is in the Ariel instantiation itself, i.e. in the switchlet. 5.1 DLAs in an Ariel Sandbox Placing a Sandbox in a switchlet itself allows one to load application-specific code as close as possible to the switch (this could be on the switch itself). By using a new module, known as the Sandbox Ariel module, this code is able to call any switch control operation available to a traditional control architecture, the only difference being that the call need not be transmitted over the network. If the divider runs on the switch, this could even be a simple function call in the same address space. Such active switchlet DLAs are called micro control achitectures. Micro control architectures are either controlled from external applications, or in a more extreme case, contain their own application code. The latter case will be termed application-specific micro control achitecture. Micro control achitectures that are controlled from external applications can easily extend Ariel's functionality by implementing for example such things as batch connection setup or teardown. The external application simply calls the new operation once with the appropriate parameters and the micro control achitecture sets up (or tears down) the appropriate connections. Furthermore, the Ariel Sandbox supports an override operation, which allows it to safely replace, modify, or extend the standard Ariel operations. Such functionality enables one to change the behaviour of a switchlet. As a trivial example, consider a DLA which overrides the connection setup and teardown operation with the empty operation. This effectively removes these operations from the switchlet's repertoire. The control architecture controlling the switchlet can still perform all other Ariel operations (e.g. gather port and connection statistics), but it is not allowed to create or delete connections. A more interesting example that has been implemented, deals with VPI/VCI space virtualisation in an ATM network. Recall that switchlets constitute a partitioning of (among other things) the VPI/VCI space. This means that CAA obtains a different VPI/VCI range than CAB . However it is well-known that certain control architectures depend on specific VPI/VCI values for signalling. UNI signalling, for example, uses VCI 5 on VPI 0 for signalling messages [19], while both IP Switching [20] and FORE SPANS [21] depend on VPI 0 and VCI 15. Problems arise when one wants to start up a control architecture that relies on a particular VPI/VCI value V , when V was already assigned to another switchlet. This can be solved easily, using the override operation in the Ariel Sandbox. All operations may be overridden to map V onto one of the switchlet's own VPI/VCI values V . This is transparent to the control architecture, which sends the usual control messages for VPI/VCI value V . In fact, it is possible to completely virtualise the VPI/VCI space. In other words: allow the control architectures to use the entire spectrum of VPI/VCI values supported by the switch (or even more) and install a DLA at the switchlet, which maps these virtual VPI/VCI values to real VPI/VCI values in the switchlet's address space (if needed, the Haboob allows real VCIs to be pinned down at the edge of the network). If control architectures span multiple switches, this means that for each of these the virtualisa- tion DLA needs to be installed. This is a responsibility for the netbuilder (discussed in Section 6), which knows about the allocated resources of netlets and is therefore able to install automatically the appropriate mappings in the switchlets at network creation time. 5.2 DLAs in the switchlet management component The need for elastic runtimes in the process of creating switchlets itself may be less obvious. One simple use is that it allows network administrators to install simple extensions to the default network functionality by automatic instantiation of an appropriate DLA in an Ariel Sandbox whenever a switchlet is created. This DLA can then extend or override the Ariel operations as described in the previous section. For example, the management DLA may be triggered each time a "create switchlet" operation is called (i.e. it has overridden the method to create switchlets). It can then examine the request and depending on certain conditions (e.g. the identity of the requester) it may create a new switchlet with an Ariel DLA which overrides the operations to create and delete connections with the empty operation. This ensures that this switchlet's control architecture is allowed to do anything on the network, except create and delete connections. A more interesting use of management DLAs implements switchlet aggregation, whereby the creation of a switchlet really creates a whole cluster of switchlets which are presented to the control architecture as a single (large switchlet). These aggregate switchlets are used to improve scalability. Due to space limitations we will not discuss aggregate switchlets any further 6 Elastic network builders Netbuilders are responsible for combining switchlets into virtual networks (level-0 netlets). Because they are essentially shared servers with very little functionality that is specific to any particular control architecture, netbuilders are interesting when it comes to adding elasticity to them. The question arises: who decides how virtual networks should be created? Until now, this is decided once and for all by the implementers of the netbuilder. The static netbuilder has a fixed functionality-virtual networks are built by associating switchlets in a very implementation-specific manner. This paper argues that this is unnecessarily restrictive. Although it is beyond the scope of this paper to discuss the netbuilder in much detail, we briefly outline the issues and show where the use of DLAs has lead to improvements compared to implementations that only use static code. For example, there is no reason why control architectures should not be able to push code into the netbuilder when this code is restricted to using only those operations that were exported by the netbuilder anyway. In other words, it should be possible to load active or passive DLAs into a netbuilder which, besides the SUFI, offers the exact same API as offered to control architectures. Doing this may optimise performance (especially since netbuilders may run very close to or even on a switch as well). The DLAs in turn are able to build new services by combining existing netbuilder services and exporting interface references for these services. however, only the ability to allow DLAs to (re-)combine existing netbuilder functionality has been discussed. A more interesting question is: who is allowed to add to, remove from or modify the netbuilder's default functionality? It seems reasonable to give at least the system administrators such powers. After all, they may want to override certain operations to make them include extensive logging, or to disable certain functions, or even to preclude the use of certain switches for certain operations. The system administrators then, should also be allowed to grant these privileges to other network users (in the same way as a system administrator is allowed to grant the root password of machines to certain, trusted users). The new functionality that is thus created by the system administrators can subsequently be used by non-privileged DLAs that present the appropriate capability. In this way, the netbuilder's functionality can be changed and extended on the fly in a safe manner. In the Haboob, the netbuilder allows clients to load code into a shared Sandbox. The code is able to use all the basic operations offered by the netbuilder, as well as modified or extended operations implemented by the system administrator's DLAs (provided the right capability is presented). A very simple example of a netbuilder extension is one which automatically pushes DLAs in the switchlets that are created on its behalf by the divider. In this way, we made a netbuilder push 'virtualisation DLAs' in each of the switchlets, so that the VPI/VCI spaces of these switchlets were completely virtualised (as discussed in Section 5.1). The netbuilder makes sure that each virtual VCI is mapped on the appropriate 'real' VCI (taking into account that neighbouring switchlets need the same mapping). 7 Active Networks We recognise the value of in-band computation, as advocated by the active network community, in certain cases. For example, [7] shows as an example how a protocol "booster" can extend the functionality of a protocol on the fly, by adding forward error correction (FEC) on the datapath when it is found that one of the links is very slow and unreliable. The booster only affects that particular link. Similarly, [6] shows how an "active bridge" between two ethernet segments allows for upgrading to new versions of network software with minimal disruption. So how do we deal with these problems in the Haboob? The answer is simple: if and when necessary we build a simple active network node on the fly. Consider the example of a protocol booster. Whenever the decision is made that a particular link is unreliable and needs FEC, it is simple to dispatch a DLA that reroutes the existing connection to a control entity near the link. The control entity adds the FEC and returns the code to the data path. The FEC addition can be done by a DLA running in a Sandbox in the control entity. The protocol booster mechanism is illustrated in Figure 4. Note that this does not differ in any way from the protocol boosters in [7]. In fact, we have built a very simple active node. This is an example of an elastic datapath component. remove FEC (Lossy A add FEC Figure 4: Active protocol booster in the Haboob 8 A Distributed Video Server As a further demonstration of the Haboob, a distributed video server (DVS) was built on top of Sandman. The aim of the DVS was to provide a virtual big disk to low-end users that do not have enough storage space to store, for example, feature-length films on their local disks. The DVS, which is entirely implemented in dynamically loadable code, is able to use all disks in the network, not just the user's local disk, for storing very long video files. In recording mode, the video data is automatically striped over the various disks (using coarse-grained striping), according to a user-defined load-balancing policy. This means that the appropriate connections are set up at the appropriate times to each disk on which a particular segment should be stored. In playout mode, the DVS finds all the segments that make up the video and sets up connections from the corresponding disks at the appropriate fashion to ensure smooth playback. The sink is not aware of this. It simply receives on a single VPI/VCI pair. The sources are merged in a resource efficient manner, by reusing as much of the existing connection as possible. The DVS is described in detail in [22]. 9 Performance Because of space limitations, we are not able to discuss results in great detail. As an indication of performance we measured the connection setup and teardown times in a configuration where the endpoints of communication reside on Sun UltraSparcs and the control architecture, divider and netbuilder all run off-switch (also on UltraSparcs). As shown in Table I, a setup across a single FORE ASX-200 switch using Sandman took as little as 7:5 ms 5 . A teardown took on average 2:6 ms. This compares favourably with both ATMF UNI signalling as reported in [23] and xbind open control as reported in [10] (see Section 10). This should be taken as a rough comparison only, as the configuration and load of the switches in the various scenarios were almost certainly different (which has significant effect on the performance). Replacing Sandman with a Noman control architecture adds approximately 1 ms to the setup and teardown time. These figures are very good and prove that off-switch control using DLAs need not result in poor performance. switch software location t (ms) Unknown Xbind/GSMP off-switch DEC AN/2 UNI (Q.Port) off-switch 21 ASX-200WG UNI on-switch 20 ASX-1000 Sandman/GSMP off-switch 15 ASX-200 Sandman/SNMP off-switch 7:5 Table 1: Comparison of switch setup times Related work Several solutions to make networks programmable have been proposed in the last few years. We briefly discuss active networks, open control and intelligent networks. Active Networks are packet-switched networks where each packet may carry executable code. In [3] these packets are called capsules, i.e. little programs with embedded data that are evaluated in a transient execution environment, allowing network nodes to process the data in an application-specific way. The execution of the packets in the data path is strongly related to the speed with which these packets can travel through the network. Note that the Operation, Administration and Maintenance (OAM) cells in ATM offer a similar, albeit much more restricted functionality (the OAM cell effectively carries one of a finite number of pre-defined programs). 5 The GSMP server used for the ASX-1000 was a rather buggy experimental implementation, which explains why it is so much slower than the SNMP server on the ASX200. In addition, its control processor, a i-960, is much slower than the Pentium II used for the ASX200. The Switchware project [4] at the University of Pennsylvania, is similar to this, but does not require each packet to be a capsule (although some packets may carry code). SwitchWare is designed as a three-level architecture, where the first level comprises the active packets, the second level concerns so-called active extensions (programs that can be dynamically loaded over the network and offer functions that can be used by active packets) and the third layer is the infrastructure that enforces the rules for dynamic code loading and takes care of resource allocation. Formal methods are used to prove security properties of the dynamically loadable programs. In [5] a solution to network management using delegated agents is proposed, where the agents are dynamically loadable code that can be dispatched using a so-called delegation protocol to an executing elastic (extensible) server. This helps prevent the explosion of management traffic from all over the network to a central site. Delegating management also makes the control loop (from managing code to managed device) smaller, decreasing the probability of failure at times when there are problems (and management is needed the most). The delegated agents approach has been quite influential. For example, another project that investigates the use of agents for network and systems management is INSERT [24], where Java agents are used to implement managed objects (similar to those of SNMP and CMIP) and can be instantiated dynamically. The second approach to making networks programmable is known as open control. Early implementations of such control are xbind [10] and the Tempest [1] (of which the Haboob is a continuation). Xbind constitutes a framework for the creation, deployment and management of multimedia services on ATM networks with end-to-end QoS guarantees. Services are created by interconnecting (binding) abstract representations that model the local states of network resources including links and switches. For this purpose, these abstractions are stored as an organised collection of interfaces, called the Binding Interface Base (BIB). Closely related to netlets is the more recent Genesis project [17] (netlets were introduced in a primitive form in [15]). Like the research presented here, Genesis enables recursive partitioning of network resources. The partitions of network nodes, called "routelets", are created by a parent control program, but controlled by an independent child control program. However, the Genesis model is much more complex than that of netlets. More serious is that the hierarchy of nested routelets seems reflected in the datapath. Each routelet consist of a myriad of components and services and part of this comprises the routelet's transport module, which resides on the datapath. The transport module includes a virtual input port, a forwarding engine (that processes the data) and a virtual output port. An incoming packet is classified, for example, as belonging to a specific (higher-level) routelet and demultiplexed to the routelet's transport module accordingly. It arrives at the routelet's virtual input port, is processed in the forwarding engine and sent out on the virtual output port. However, it is not sent to the next node yet. It first arrives at the virtual input port of its parent routelet. Here the process is repeated and so on, until it arrives at the root routlet's input port. Here it is processed for the last time and finally forwarded to the next node in the network. This is rather inefficient as the data is touched, not just once, but potentially many times. This is an example of multiplexing at many layers. Contrast this with the "flat" model of netlets. Here netlets form simple partitions of lower-level netlets that can be controlled in arbitrary ways, e.g. by setting up connections from one endpoint to another. The data simply flows through these connections and is not touched at all. Multiplexing happens once and at the lowest possible level. Intelligent Networks (IN) [25] allow the introduction of new services by associating them with signalling endpoints. Basic calls are separated from IN-based calls. For example, dialling an 0- 800 number will trigger a temporary suspension of call-processing and initiate a series of transactions between the Service Switching Point (SSP) and the Service Control Point (SCP), which is essentially a real-time database. A lookup in this database (e.g. for a 0-800 number) tries to find the corresponding service logic, i.e. the code which is then executed. The code sends back instructions to the SSP on how to process the call. The bulk of current IN transactions consist of translating the number dialled by the caller into another number. Conclusions We have discussed the Haboob, which represents a direction in network control, called elastic control, that combines elements from active networks and open control. It is like open control (and intelligent networks) in that it maintains a separation between control and data and attempts to open up the traditionally closed communication systems, and like active networks in the ease with which new code can be added to existing network control by loading it up in an execution environment known as the Sandbox. It also tries to standardise a computational model and defines the Sandbox SUFI to provide DLAs with a uniform interface for communication, code loading, security and other aspects. Above all, the Haboob considers the various aspects of network control independently and enables extensibility and customisability for all of them. Acknowledgements Thanks to Ian Leslie, Rebecca Isaacs and Richard Mortier of the Cambridge Computer Laboratory for their useful comments on earlier versions of this paper. --R Open Service Support for ATM. "Realizing a foundation for programmability of ATM networks with the binding architecture," "Towards an active network architecture," "The SwitchWare active network implementation," "Delegated Agents for Network Management," "Active bridging," "Protocol boosters: Applying programmability to network infrastructures," "The IEEE 1520 Standards Initiative for Programmable Network Interfaces," "ANTS: A Toolkit for Building and Dynamically Deploying Network Protocols," "Programming telecommunication networks," "Specification of Guaranteed Quality of Service (RFC 2212)," "Intelligent Network Overview," "Implementing remote evaluation," "The Design and Implementation of an Operating System to Support Distributed Multimedia Applications," "ATM Admission Control based on Measurements and Reservations," "Measurement-based connection admission control," "The Genesis Kernel: A Virtual Network Operating System for Spawning Network Architectures," "Application-specific policies: Beyond the domain boundaries," "ATM User-Network Interface Specification - Version 4.0," "General switch management protocol specification-version 1.1." "SPANS UNI: Simple protocol for ATM signalling, release 3.0," "Building a distributed video server using advanced ATM network support," "Per- formance benchmarking of ATM networks," "Mobile Agent based management in the INSERT project," "Recommendation M.3010. Principals for a Telecommunications Management Network," --TR --CTR Simon Crosby , Sean Rooney , Rebecca Isaacs , Herbert Bos, A perspective on how ATM lost control, ACM SIGCOMM Computer Communication Review, v.32 n.5, November 2002
policing;virtual networks;active networks;open control;elastic control
596687
Edge Provisioning and Fairness in VPN-DiffServ Networks.
Customers of Virtual Private Networks (VPNs) over Differentiated Services (DiffServ) infrastructure are most likely to demand not only security but also guaranteed Quality-of-Service (QoS) in pursuance of their desire to have leased-line-like services. However, expectedly they will be unable or unwilling to predict the load between VPN endpoints. This paper proposes that customers specify their requirements as a range of quantitative services in the Service Level Agreements (SLAs). To support such services Internet Service Providers (ISPs) would need an automated provisioning system that can logically partition the capacity at the edges to various classes (or groups) of VPN connections and manage them efficiently to allow resource sharing among the groups in a dynamic and fair manner. While with edge provisioning a certain amount of resources based on SLAs (traffic contract at edge) are allocated to VPN connections, we also need to provision the interior nodes of a transit network to meet the assurances offered at the boundaries of the network. We, therefore, propose a two-layered model to provision such VPN-DiffServ networks where the top layer is responsible for edge provisioning, and drives the lower layer in charge of interior resource provisioning with the help of a Bandwidth Broker (BB). Various algorithms with examples and analyses are presented to provision and allocate resources dynamically at the edges for VPN connections. We have developed a prototype BB performing the required provisioning and connection admission.
INTRODUCTION There is a growing demand that since private networks built on using dedicated lines offer guaranteed bandwidth and latency, similar guarantees be provided in IP based Virtual Private Networks (VPNs) [9], [14]. While the internet has not been designed to deliver performance guar- antees, with the advent of differentiated services [3], [2], IP backbones can now provide various levels of quality of service.Recently proposed Expedited Forwarding (EF) [10] Per Hop Behaviour (PHB) is the recommended method of build such an Virtual Leased Line Line (VLL) type point-to-point connection for VPN. To provide such service we have (and others, for example [16],[17]) recently implemented [12], [13] a Bandwidth Broker [15] that allows an user to specify a single quantitative value (i.e 1 Mbps or 2 Mbps etc.) and based on this specification the edge routers establish VPN connections dynamically. However, it is expected that users will be unable or unwilling to predict load between VPN endpoints [7]. From the provider's point of view also, guaranteeing exact quantitative service might be a difficult job at the beginning of VPN-Diffserv deployment [2]. We, therefore, propose that users specify their requirements as a range of quantitative service. For example, a user who wants to establish a VPN between stub Networks A and B (Figure 1), and is not sure whether he needs 0.5 Mbps or 0.6 Mbps or 1 Mbps, and only knows the lower and upper bounds of his requirements approximately, can specify a range 0.5- 1 Mbps as his requirement from the ISP when he outsources his service to the latter. An ISP can offer multiple such options via a website (Figure 6) to help customers to select any suitable option to activate services dynamically on the fly. This has several advantages: Users do not need to specify the exact capacity but it gives the flexibility to specify only a range. The price that customers have to pay is higher than one pays for the lower bound capacity but lower than what is normally needed to be paid for upper bound capacity. During low load it is possible that users might enjoy the upper bound rate (say 1 Mbps in the example) without paying anything extra. This kind of pricing might be attractive to users and ISPs can take advantage of that to attract more customers without breaking the commitment. This, however, poses significant challenge to the ISPs as they would need to deploy automated provisioning system that are able to logically partition the capacity at the edges to various classes (or groups where each group is identified from it's offer, for example 0.5- 1 Mbps could represent one group, 1-2 Mbps could represent another) of VPNs and manage them efficiently to allow resource sharing among the groups in a dynamic and fair manner.Also, they must provision the interior nodes in the network to meet the assurance offered at the boundaries of the network We have, therefore, proposed a two-layered model in section II to provision such VPN-Diffserv Networks where the top layer is responsible for edge provisioning and drives the lower layer in charge of interior resource provisioning with the help of a Bandwidth Broker (BB). We have restricted this paper to edge provisioning only considering the fact that most of complexities lie at the boundaries of the network and is the main driving force for overall provisioning. In section III various algorithms with examples and analysis have been presented to provision and allocate resources dynamically at the edges. Fairness issues while allocating unused resources have been addressed in section III-D. A prototype BB performing the required provisioning and connection admission has been described in section IV. Section V concludes the paper with a summary of our contributions and a discussion of future research directions. Edge Router 3 Interior Router 1 Interior Interior Edge Router 1 Edge Router 3 Edge Router 5 Edge Router 4 ISP Stub Network Customer Customer Stub Network A 1A Network Stub Customer Customer Stub Network C 1C Fig. 1. VPN Diffserv deployment scenario II. PROVISIONING REQUIREMENTS FOR VPN-DIFFSERV NETWORKS: A MODEL Provisioning in Diffserv Networks does not only mean determination and allocation of resources necessary at various points in the network, but also modification of existing resources to be shared dynamically among various VPN classes (i.e. groups). Both quantitative, as it is the case with VPNs, and qualitative traffic (some assured service) are required to be provisioned at the network boundaries and in the network interior. Determination of resources required at each node for quantitative traffic needs the estimation of volume of traffic that will traverse each network node. While an ISP naturally knows from the SLA the amount of VPN quantitative traffic that will enter the transit network through a specific edge node, this volume cannot be estimated with exact accuracy at various interior nodes that will be traversed by VPN connections if we do not know the path of such connections [1]. However, if the routing topology is known, this figure can be almost accurately estimated. If the default path doesn't meet the requirements of an in-coming connection, alternate and various QoS routing [6], [5] can also be used to find a suitable path and enforced by MPLS techniques [8]. A. Role of Bandwidth Broker for Automated Provisioning Based on the basic needs of provisioning a VPN-Diffserv network to support quantitative service we consider the provisioning as a two layered model - the top layer responsible for edge provisioning and driving the bottom layer which is in charge of interior provisioning (Figure 2). As we seek to provide a system where VPN services are available on demand, we find that Bandwidth Broker [15],[17] is the right choice, because it is not only capable of performing dynamic end-to-end admission control to setup a leased line like VPN by maintaining the topology as well as policies and states of all nodes in the network, but also capable of managing and provisioning network resources of a separately administered DS domain and cooperating with other similar domains. Edge Provisioning Interior Provisioning Bandwidth Broker CR CR CR ER CR CR CR ER Fig. 2. Layered Provisioning view of VPN-Diffserv Networks B. A Novel Bandwidth Specified as an Interval To overcome users' difficulty in specifying the exact amount of quantitative bandwidth required while outsourcing the VPN service to ISP our model supports a flexible way to express SLAs where users specify a range of quantitative amounts rather than a single value. Although it has several advantages, this also makes the edge and interior provisioning difficult. This complexity can be explained with a simple ex- ample. Referring to Figure 1 once again, assume that edge router 1 has been provisioned to provide 20 Mbps quantitative resources to establish VPN connections elsewhere in the network and ISP has provided two options via a web interface to the VPN customers to select the rate of the connections dynamically: 1 Mbps or 2 Mbps. It is easy to see that at any time there can be 20 connections each having 1 Mbps, or 10 connections each enjoying 2 Mbps, or even a mixture of the two (e.g. 5 connections with connections with 1 Mbps). When a new connection is accepted or an active connection terminates, maintaining the network state is simple and doesn't cause either reductions or forces re-negotiations to existing connections. If there are 20 connections of 1 Mbps, and one connection leaves then there will be simply 19 connections of 1 Mbps. Admission process is equally simple. Now if the ISP provides a new option ( for example, as shown in Figure 6(b)) by which users can select a range 1Mbps - 2 Mbps (where 1 and 2 are the minimum and maximum offered guaranteed bandwidth), maintaining the state and admission control can be difficult. A detailed example can be found in section III-B.When there are up to 10 users each connection would get the maximum rate of 2 Mbps, but as new connections start arriving, the rate of existing connections would de- crease. For example, when there are 20 connections this rate would and then at that stage if an active connection terminates the rate of every single connection would be expanded from 1 Mbps Mbps. This is a simple case when we have a single resource group supporting a range 1Mbps-2 Mbps. In reality, we might have several such groups as shown in Figure 6(b). In such cases, renegotiation for possible expansion of existing connections, admission control and maintenance of network states will not be simple. The idea presented here is illustrated in figure 3. BE traffic EF traffic C user_max(i) C user_min(i) user(i) user(i) VPN Pipe VPN connection demand (a) (b) Fig. 3. The SLA (a) Bandwidth is specified as an interval of C user min(i) and C user min(i) for any group i. Actual rate of a VPN connection C user(i) varies between this range but never gets below C user min(i) . (b) C user(i) is the rate that is configured in the edge router as the policing rate. Traffic submitted at a rate higher than this rate is marked as best effort traffic or dropped depending on the policy C. The Model and Notations In our model, we address this novel approach to SLA and provide policies and algorithms for automated resource provisioning and admission control. However, to support such provisioning, we first start by allocating a certain percentage of resources at each node (edge and interior) to accommodate quantitative traffic. At the edge this quantitative portion is further logically divided between dedicated VPN tunnels (i.e. require 1Mbps or 2 Mbps explicitly) and those connections that wish to have rates defined by a range (i.e 0.5-1 Mbps or 1-2 Mbps etc. This top level bandwidth apportionment is shown in Figure 4. The notations are : ffl CT is the total capacity of a node interface. ffl Cded is the capacity to be allocated to VPN connections requiring absolute dedicated service ffl Cshared is the capacity apportioned for those VPN connections who describe their requirement as a range. ffl C qual is the remaining capacity for qualitative traffic. ffl Cquan is the capacity provisioned for quantitative traffic and is equal to (Cded+ Cshared ). C quan C quan shared ded C qual z. shared y. a. C qual C qual (b) (a) b. Fig. 4. Top level Bandwidth Apportionment: (a) logical partitioning at the edge, (b) logical partitioning at an interior While at the edge Cquan is rate controlled by policing or shaping, at the interior this Cquan indicates that this amount of capacity will be allocated (actually protected) to quantitative traffic if need arises. All the values can be different at different nodes. This kind of logical partitioning is helpful because capacity is never wasted even if portions of resources allocated to quantitative traffic are not used by VPN connections. Unused capacity naturally goes to qualitative portion and enhances the best effort and other qualitative service. This is true both at the edge and in the interiors. Cshared , as shown in Figure 4, can be logically divided to multiple groups where each group supports a different range (Figure 5). As there might be multiple of such groups, for any group i we define the following notations: ffl C base(i) is the the base capacity for group i which is shared by the VPN connections belonging to that group. ffl C user min(i) is the ISP offered minimum guaranteed bandwidth that a user can have for a VPN connection. ffl C user max(i) is the ISP offered maximum guaranteed bandwidth that a user can have for a VPN connection. ffl N shared(i) is the current number of shared VPN connections in ffl C shared(i) is the amount of capacity currently used by group i. ffl C user(i) is the actual rate of active connections in group i and is equal to C shared(i) (in section 3). Cshared unused is the total unused bandwidth from all shared service groups. shared-unused C base(i) ded shared shared qual C00000000000000000001111111111111111111Unused Fig. 5. Microscopic View of Bandwidth Apportionment at Edge There are numerous sharing policies that we can apply to these shared service groups. We call them shared service groups because in reality the base capacity is shared by a certain number of VPN connections and sharing policy might allow a group to share it's resources not only among it's own connections, but also share with other groups' VPN connections in case there is some unused capacity. This may also apply to dedicated capacity. Priority can be given to certain groups while allocating unused resources. Actually, fair sharing is a challenging problem, and we will address all these issues in the following sections while developing provisioning mechanisms. III. EDGE PROVISIONING POLICIES: ANALYSIS AND ALGORITHMS Based on the model described in section II, various allocation policies could be adopted by the ISPs at the ingress point to allocate capacity dynamically to maintain and guarantee the quality of service of various types of incoming and existing VPN connections as we will have multiple classes of VPNs each supporting different bandwidth specifi- cations. Some suitable policies are : ffl Policy I: Capacity unused by one group cannot be used by any other groups. This means that if we have multiple shared service groups, one group whose resources have been exhausted while supporting numerous connections doesn't borrow resources from others even when those groups have unused capacity.Also, none of the groups are allowed to use unused capacity of dedicated service group. ffl Policy II: Capacity unused by one shared service group can be borrowed by another shared service group. However, like the previous pol- icy, they are not supposed to borrow from the dedicated service group. ffl Policy III: Capacity unused by dedicated service group can be borrowed by tunnels of shared service groups. Also, these groups can share resources among themselves. In this section, we will start with VPN Connection Acceptance algorithms at Ingress point where all admission complexities lie. This complexities are introduced because of the need to partition and share resources to support our model and policies presented above. Further analysis with examples of algorithms for Policy I,II and III clarify those in detail. A. VPN Call Acceptance at Ingress The job of admission control is to determine whether a VPN connection request is accepted or rejected. If the request is accepted, the required resources must be guaranteed. For any group i a new VPN establishment request is admitted only if at least the minimum bandwidth as stated in the offer can be satisfied while also retaining at least the minimum requirements for the existing users, i.e. if C user min(i) a new VPN connection request can be accepted. This ensures that, an admitted VPN connection will always receive at least the minimum offered bandwidth C user min(i) in group by restricting the number of maximum connections that can join the group. How much capacity the accepted connection will actually have is decided by connection state in that group and sharing policies that we are going to discuss in the next subsections. B. Capacity Allocation with no sharing among groups: Policy I The base capacity allocated to a group is solely used by the VPN connections belonging to that group only. Under no circumstance resources assigned to one group can be borrowed by others, even if that capacity is unused. This makes allocation simple not only at the edges, but also in the interior and from an implementation point of view it is simple. Since the unused capacity is not used by any other groups, qualitative services, as we mentioned earlier, are also enhanced. If a VPN connection is accepted the system checks if that connection can be allocated the maximum rate. This is possible if the base capacity C base(i) is enough to assign all the existing connections the maximum rate C user max(i) . Otherwise, the base capacity is shared among all the existing and new VPN connec- tion. Therefore, we can express this admission policy as follows: Example 1: For the following example assume that the total link bandwidth Mbps and there is only one ( shared user group .Also assume that ISP offers this group as C user Base capacity C base(1) allocated to this group is 20 Mbps. Mbps Calls are accepted as long as the condition C user min(i) of section III-A is met. When the number of calls exceed C user min(i) a new arriving call is rejected. For example, if the 21st call in the example were accepted then C user(1) would have been21 , and the minimum bandwidth could no longer be guaranteed. There- fore, the call is rejected. C. Capacity Allocation with sharing among groups: Policy II If the capacity allocated to a group is not fully used by VPN con- nections, then this capacity can be borrowed by connections of other shared service groups if needed. However, borrowed capacity must be relinquished when needed by the group from which capacity was borrowed. Although this borrowing and deallocation adds some complexity in edge provisioning, connections from various groups, how- ever, have better chances of enjoying higher rates. In the following we present algorithms regarding VPN connection arrival, termination and possible expansion of existing connections as a result of the termination of a connection from a shared service group. C.1 VPN Connection Arrival Like the previous case, VPN connection arrival essentially involves checking the availability of resources that can be used by the new con- nection, and if available, allocating this capacity to an incoming call. Even if the base capacity of a certain group allows the new connection belonging to that group to assign maximum ISP offered rate (i.e. - C user max(i) ), because of the resource sharing among various groups it might happen that resources from that group has been borrowed by other group(s) not leaving the required resources (i.e. Cshared unused ! C user max(i) ). In such a case resource must be relinquished from the appropriate groups(s). Any such de-allocation from existing connections leads to rearrangement of capacity of those connections. It should be noted that capacity should be relinquished the way it was borrowed. There are numerous ways unused capacity can be borrowed by competing groups which we will see in sections III-C.3 and III-D. For the sake of simplicity, the group which has the maximum excess bandwidth, C should release first, and then the next, and so on. /* if the group has enough base capacity to support a new connection with max. offered rate. */ if hi /* if the shared unused capacity is also enough to support the new connection with max. offered rate. See Example 2 */ if Cshared unused - C user max(i) /* if the shared unused capacity has been borrowed then capacity is relinquished from borrower(s). See Example 3 */ else n relinquish C user max(i) from group(s) which has max excess bw rearrange bandwidth of that group(s) We have just mentioned that capacity can be borrowed from one group by the others. When does one group borrows resources? Natu- rally, when the base capacity is less than what is needed i.e How much can one group borrow? This depends on how much unused resources are available. If this is at least equal to the maximum offered rate C user max(i) , then that amount is allocated, otherwise (i.e. Cshared unused ! C user max(i) ) the whole unused resource goes to the group in question and is then divided among all the connections in that group /* if the shared capacity is equal to or has exceeded the base capacity */ if hi /* but the unused capacity can still support the new connection with rate. Capacity is then borrowed. See Example 4 */ if Cshared unused - C user max(i) /*if the unused capacity is less than the max. rate. Capacity is then shared by existing and the new connection. See Example 5 */ else n Cshared unused We will now consider several numerical examples in this section to clarify the algorithms and analysis presented above. For all the following examples we assume that the total link bandwidth Mbps and there are only two shared users groups i.e. 2. For group 1 C user C user C user Prior to VPN connection request in group 1: Here, for group 1, C and C user C user max(1) . Also, Cshared unused which is greater than C user max(1) . Hence, C Prior to VPN connection request in group 1: In this example, C is greater than C user Mbps. This means that group 1 hasn't used all it's base bandwidth and a new connection can have the maximum offered bandwidth 1 Mbps. However, Cshared unused at the time of request arrival is Cshared \Gamma Mbps. This indicates that another group has has borrowed capacity from group 1. If that group had left at least C user Mbps then the request could have been assigned the desired amount of resource. Therefore, the only option left is to relinquish 1 Mbps from the group that has borrowed it. Searching the table we find that the only other group 2 has taken that bandwidth. Therefore, we need to deduct 1 Mbps from group 2 and recompute the individual share of a VPN connection as C C shared(2) \GammaC user max(1) Mbps. Obviously, C Mbps. Prior to VPN connection request in group 2: This is a case where one group has used it's full allocated base capacity but can borrow resources from the other group which has left some spare capacity. Here, C the total spared capacity Cshared unused and this value is greater than C user max(2) (i.e 2 Mbps). Therefore, the new VPN connection request can be allocated the maximum offered value (i.e. 2 Mbps) by even exceeding the base capacity of group 2. Example 5 :Prior to VPN connection request in group 2: The example here depicts a scenario where one group which has already exceeded it's base capacity and has to accommodate a new connection request when there is no unused resource left by other group(s).Here, even before the new call arrival, Group 2 has borrowed Mbps. So, the current capacity allocated to group will have to be equally distributed among all the existing and the new arriving VPN connection. Therefore, C Mbps. C.2 VPN Connection Termination When a VPN connection terminates, resources might have to be released from the relevant group depending on the current rate every connection is enjoying in that group. If the rate is less than or equal to the maximum offered rate then no capacity is released from the groups current share and as a result all the connections in that group increases equally. This is because the same capacity is shared by a lower number of connections. If, however, the current rate of every connection is already equal to the maximum offered rate, then this termination would trigger a deduction of C user max(i) from the shared resource C shared(i) . If all the connections were already enjoying C user max(i) , no rate change occurs in any of the existing connections.The algorithm is stated as follows: if /*See Example 6 */ Cshared unused = Cshared unused if Cshared unused = Cshared unused To clarify the VPN connection termination process will now consider similar examples as presented in the previous section. Example connection termination from group 1: Here, C shared(1) user 1. This means that the capacity used by this group before the connection termination will remain unchanged even after the termination. So, the new value of C shared(1) is also 10 Mbps and each VPN connection will equally share this capacity which is C shared(1) capacity is deducted from this group, the total unused shared capacity will also remain unchanged. Example 7: Before VPN connection departure from group 1: In this example, C shared(1) user states the fact that prior to this departure all active VPN connections were using the maximum possible offered bandwidth C user 1 Mbps and in total were having Hence, the departure should trigger a deduction of C user Mbps from the total capacity used by this group prior to the departure as the capacity even after the deduction will be good enough to satisfy active connections offering highest possible rate of 1 Mbps. Therefore, C Mbps and and each VPN connection will receive C shared(1) the termination process triggers deduction of C user max(1) from the capacity used by group 1, the unused shared capacity will increase by the same value. So, Cshared unused C.3 VPN Capacity Expansion Unused shared capacity left by some groups can be distributed among others. Priority can be given to certain groups while allocating unused capacity. In the next section we will present various policies to allocate unused dedicated capacity and those might apply here as well. Here we consider only one case where preference is given to the needy groups where need is determined from the ratio C user(i) C user max(i) . So, we order the groups according to this ratio where in reordered groups the first one has the lowest C user(i) C user max(i) and the last one has the highest C user(i) C user max(i) . Once reordering has been done the expansion algorithm starts allocating unused bandwidth to the first group, then the next, and so on based on the availability of resources. This can be stated as : if shared unused shared Cshared unused \Gamma [N shared(i) :C user if shared unused +Cshared unused Cshared unused Example 8: Before VPN connection termination from group 2: After the termination of a VPN connection from group 2, Cshared unused Mbps. If there is need of resources by other group(s), this capacity can be used partly or fully. We find that group 1 has need for this resource since C user(1) user max(1) 1. Now it remains to be seen to what extent we could use this unused capac- ity. Here, C shared(1) +C shared unused 11 and is greater than C user max(1) which is 1 Mbps. Therefore, capacity for group 1 can be expanded to N shared(1) :C user Mbps allocating each existing connection C user The remaining unused capacity will be reduced to Cshared unused \Gamma [N shared(1) :C user Mbps. Example 9: Before VPN connection departure from group 2: Unlike the previous example where group 1 only needed to use a portion of the unused resources, all the remaining capacity can be allocated to existing group 1 VPN connections in order to enhance the service. C shared(1) will be increased to 10+ 2=12 Mbps and each existing connection will receive C shared(1) 12Mbps. D. Fair Allocation of Unused Dedicated Resources: Policy III In the previous section we have discussed methods where one shared service group can borrow resources from another similar group. In this section, we will discuss the possibilities of sharing the unused dedicated resources among various shared service groups. If the shared service groups are allowed to borrow resources from unused dedicated resources, we then define a new term: shared Cded unused The question here is how we can allocate the unused dedicated resources fairly among the competing groups. If all VPN tunnels want the maximum bandwidth as offered in ISP policy offer, then it is possible that at some point: shared If shared , the quantity that is needed to allocated the maximum possible offered rates to all connections even after allowing the unused dedicated resources to be used by shared service groups, is greater than 0, we need to define a fair set of user throughput values (i.e. C user(i) ) given the set of maximum offered loads C user max(i) and C shared . In other words, we need to divide this extra capacity Cded unused among all the needy groups in a fair manner. However, fair sharing of extra resources is not a trivial issue and was addressed by others for different network situations [20], [11], [18], [19] . Some proposals [11] are in favour of sharing the bottleneck capacity equally among users independent of their requirements , and others [20], [18] advocate to penalize users causing overloads. While we do share the resources among VPN connections in each group, equal sharing of unused dedicated capacity will not help much to some groups where connections are already enjoying rates close to C user max(i) . At the same time it also doesn't alleviate the problem of other groups having rates above C user min(i) but much less than C user max(i) . The fairness criterion of [20] also doesn't fit here as that would deprive the heavy user groups to gain share from unused dedicated resources even when they are enjoying rates much below C user max(i) . Our case is further complicated by the fact that while penalizing heavy user groups we cannot reduce their current share, and this is what might happen in certain cases while trying to maximize the rates of lower user groups. In the following sections we will discuss various fair sharing methods at the edges. D.1 Allocation of unused resources to lower user groups first In this case, we first need to order the user groups based on their C user max(i) values. The objective is to satisfy the lower user groups first by trying to allocate maximum offered values while higher user groups have less chances to acquire resources left by dedicated service group. The rationale behind this is that more VPN users can be satisfied and allocating to higher user groups might bring little changes in many cases if sufficient extra resource is not available. If the ordering leads to service groups it is possible that if we expand K groups the VPN tunnels belonging to those group will enjoy the maximum offered band- th group receives rest of unused dedicated resource, and other tunnels remain unchanged. The total enhanced shared capacity can then be computed as follows: shared N shared(i) :C user max(i) Cded unused [N shared(i) :C user The above computation helps us to view how C shared is shared by different groups. However, this general case is true when K - 2. The other cases are: shared C shared(1) +Cded unused if Cded unused In practice, when there is unused dedicated capacity the process starts by asking the first group if the unused capacity is enough to satisfy all the VPN connections. If so, each connection receives a maximum value C user max(i) and then queries the second group. Other- wise, the whole amount of capacity is allocated to the first group and divided among the competing connections. The process continues as long as unused capacity is a positive figure. Assume a situation where we have 3 groups where VPN connections in each of them were having capacity below their respective C user max(i) . Also, Cshared = Mbps, and for group 1: C 0:25 Mbps, for group 2: C Mbps, C user Mbps, and for group 3: C Mbps, C user Mbps. Prior to the availability of Cded unused = 7 Mbps we had : Here the groups are already ordered. Applying the algorithms we see that the first two groups can be allocated the maximum rates. Therefore, they are both expanded to 15 \Theta and 12 \Theta respectively. Rest of the unused capacity Cded unused \Gamma Mbps goes to the third group. D.2 Allocation of unused resources to highest needy groups first This is much like the process as described above with the only difference that groups are ordered based on their needs. Apportionment mechanisms and algorithms remain the same. Here, need is determined from the ratio of C user(i) C user max(i) . So, groups with lower ratios get preference over groups with higher ratios. Therefore, the process starts feeding the most needy group and continues as long as it has some unused capacity. of previous section: C user max(3) C user max(1) 0:67, and C user(2) C user max(2) 0:83. Clearly, group 3 is the most needy group. If we have Cded unused Mbps then that can serve the the most needy group 3 and enhance it's service. The new C C user 0:67. In the previous examples, this group never had the chance to grab portion of the unused bandwidth, but the new policy here allows it to improve the service substantially. D.3 Allocation of unused resources based on proportional needs Although the above mechanism seems to be fair since it allocates based on the group's need, but in many cases there will be several needy groups with little differences in their needs, and in such cases the apportionment might not be always fair if unused dedicated resources are exhausted while trying to feed first few groups and other remain deprived to get a share. In this section, we there- fore, present a way to allocate unused resources based on proportional need. Any group that is in need of resource, i.e, having ratio C user(i) C user max(i) receives a portion of unused resource that is proportional to the group's need. Therefore, any group i, after receiving the extra resource based on this proportional need, is expanded to ded unused :C shared excess(i) shared excess need for group i C shared excess(i) , is actually excess quantity that is needed to offer all connections in that group the maximum value C user max(i) . Therefore, C shared Example 12: Once again, let us consider example to illustrate the use of proportional need. No ordering is needed here as allocation of extra capacity is solely based on proportional need. Here for group 1: C user(1) C user max(1) 0:67, for group 2: C user(2) C user max(2) and for group 3: C user(3) C user max(3) 0:5.Application of this allocation policy will expand the capacity of group 1 to: C 7[(0:5)15\Gamma5] connections are improved with new C C user max(1) 0:79. Similarly, for group 2: C Mbps, C C user max(2) 0:89 and for group 3: C user 0:68. This clearly shows that proportional sharing fairly enhances the rate of most needy group 3. This wouldn't have been the case had we applied other fairness methods. IV. IMPLEMENTATION OF BANDWIDTH BROKER FOR DYNAMIC CONFIGURATION As the underlying network may provide different classes of service to satisfy various VPN customers, by identifying the generic functionality provided by any resource and policy options, we present the BB with a standard WEB interface as shown in Figure 6(b). The Band-width Broker manages the outsourced VPNs for corporate customers that have Service Level Agreements (SLAs) with their ISPs and allows one such user to specify demand through a WWW interface to establish a VPN with certain QoS between two endpoints. A. The Essential Components of Bandwidth Broker The BB needs to keep track of existing connections and available resources and update relevant databases to reflect the most recent net-work state. The BB interacts with specialized configuration daemons (CD) when a certain user request arrives to setup a tunnel and the BB has to decide whether it can allocate enough resources to meet the demand of that tunnel. While the BB invokes a SLA database to check the validity of the user request, it essentially needs to maintain a connection database that contains a list of currently active VPNs and an edge resource database to keep track of records of quantitative resource available (base capacity) and current resource consumption of various router interfaces. Fig. 6. BB WEB interface for Users B. Examples of Dynamic Configuration SWITCH Network Public Internet 130.92.66.22 Univ. of Geneve Univ. of Berne Univ. of Berne Fig. 7. Experimental Setup of VPN A resource controller in the Bandwidth Broker checks resource and connection databases whenever there is any new connection arrival or departure that might trigger modification of rates of existing connec- tions. For better understanding of how edge routers are dynamically configured to meet the user demand and conform SLA we will now demonstrate some examples of dynamic rate allocations of VPN connections in commercial Cisco routers. By considering similar examples as detailed in section III we will see how the simple algorithms are really applied to the edge devices. Let us consider an experimental setup Figure of Difserv-VPNs where we have three VPN and QoS capable edge routers each having private network behind them. Configuration 1: User 'A' wants to establish a VPN connection for source 172.17.0.100 and destination 172.20.0.100 and chooses a menu (1-2 Mbps) from ISP provided website and submits his request. The resource group definition and edge resource database entries are as shown in Figure 8. Applying algorithm presented in section 3, the policing rate C user(1) that is configured in edge router 130.92.70.101 is C Mbps. If user 'B' chooses the same menu he also gets C since capacity in group 1 has the ability to support that. Assume that two more users 'C' and 'D' decide to have VPN connection with capacity varying between 0.5 and Mbps. Group 2 can support both the connections with the maximum available rate of 1 Mbps. Therefore, C Mbps is also configured in the router for these connections as we see in the following: /*policing individual VPN connection at the inbound with C for users 'A' and 'B' and C users 'C' and 'D'*/ rate-limit input access-group 140 2000000 2000000 8000000 conform-action set-prec-transmit 1 exceed-action set-prec-transmit 2 rate-limit input access-group 141 2000000 2000000 8000000 conform-action set-prec-transmit 1 exceed-action set-prec-transmit 2 rate-limit input access-group 142 1000000 2000000 8000000 conform-action set-prec-transmit 1 exceed-action set-prec-transmit 2 rate-limit input access-group 143 1000000 2000000 8000000 conform-action set-prec-transmit 1 exceed-action set-prec-transmit 2 /*Classifying the requested VPN traffic/ access-list 140 permit ip host 172.17.0.100 host 172.20.0.100 access-list 141 permit ip host 172.17.0.101 host 172.20.0.100 access-list 142 permit ip host 172.17.0.102 host 172.20.0.100 access-list 143 permit ip host 172.17.0.103 host 172.20.0.100 Here, we show only the ingress router policing and marking since diffserv is unidirectional. We assume that bit precedence 1 is used for EF traffic marking and traffic that exceed the specified rate are marked as best effort (bit precedence 2). Users not familiar with Cisco routers, should only notice the first of the traffic rate parameters (for example 2000000 in '2000000 2000000 8000000') in rate-limit policing and marking commands. This is the rate that we refer to as C user(i) for any group i. The other two are burst parameters. Controller Resource A 172.17.0.100 140 172.20.0.100 resource definition connection database edge resource database Network Elements (Edge Routers) Request from WEB interface Edge Router Group Base Capacity Used Capacity Resource Group user-min user-max User ID Source Address ID Time ID Source Tun. Dest Address Dest Tun. Resource Group Usage Current Activation Fig. 8. Partial entries of Connection and Resource Databases.A scenario when all connections receive the maximum offered value Configuration 2: Now if users 'A' and 'B' also want to establish connections from the same sources to 172.18.0.100 and 172.18.0.101 respectively and choose a menu (0.5 - 1 Mbps) i.e. group 2, we see that capacity is exhausted in group 2, and therefore, these two new connections and other two existing connections share the base capacity of 2 Mbps and each connection is configured with C C user Mbps. This is shown in Figure 9 and the new configuration script that is used at this point is as follows: rate-limit input access-group 140 2000000 2000000 8000000 conform-action set-prec-transmit 1 exceed-action set-prec-transmit 2 rate-limit input access-group 141 2000000 2000000 8000000 conform-action set-prec-transmit 1 exceed-action set-prec-transmit 2 rate-limit input access-group 142 500000 2000000 8000000 conform-action set-prec-transmit 1 exceed-action set-prec-transmit 2 rate-limit input access-group 143 500000 2000000 8000000 conform-action set-prec-transmit 1 exceed-action set-prec-transmit 2 rate-limit input access-group 144 500000 2000000 8000000 conform-action set-prec-transmit 1 exceed-action set-prec-transmit 2 rate-limit input access-group 145 500000 2000000 8000000 conform-action set-prec-transmit 1 exceed-action set-prec-transmit 2 access-list 144 permit ip host 172.17.0.100 host 172.18.0.100 access-list 145 permit ip host 172.17.0.101 host 172.18.0.101 V. AND CONCLUSION In this paper, we have proposed a novel range based SLA that allows customers to specify their requirements as a range of quantitative service for VPN connections since they are unable or unwilling to predict load between the VPN endpoints. To support such services we have proposed and developed a prototype Bandwidth Broker (BB) that can logically partition the capacity at the edges to various service classes (or groups) of VPNs and manage them efficiently to allow resource shar- Controller Resource A 172.17.0.100 140 172.20.0.100 resource definition connection database edge resource database Network Elements (Edge Routers) Request from WEB interface Edge Router Group Base Capacity Used Capacity Resource Group user-min user-max User ID Source Address ID Time ID Source Tun. Dest Address Dest Tun. Resource Group Usage Current Activation A 172.17.0.100 144 172.20.0.100 151 2 0.5 Mbps 17:20 Fig. 9. A scenario when rate of existing connections are reduced to accommodate new connections ing among the groups in a dynamic and fair manner. Various algorithms with examples and analysis have been presented to provision resource dynamically at the edges to support QoS for VPN connections. One obvious advantage of our system is the pricing gain. The price that customers have to pay is higher than one pays for the lower bound capacity but lower than what is normally needed to be paid for upper bound capacity. During low load it is possible that users might enjoy the upper bound rate without paying anything extra. Such pricing might be attractive to users and ISPs can take advantage of that to attract more customers. With all these advantages we believe that our model can be quite attractive to the ISPs willing to deploy it in a real world scenario. VI. ACKNOWLEDGEMENT The work described in this paper is part of the work done in the project Charging and Accounting Technologies for the Internet (CATI) [4] funded by the Swiss National Science Foundation (Project no. 5003-054559/1 and 5003-054560/1), the SNF R' Equip project no. 2160-053299.98/1 and the foundation F-orderung der wissenschaftlichen Forschung an der Universit?t Bern --R Routing Guidelines for Efficient Routing Methods. A Framework for Differentiated Services. An Architecture for Differentiated Services An Overview of Quality of Service Routing for Next-Generation High-Speed Networks: Problems and Solutions A framework for qos-based routing in the internet A Flexible Model for Resource Management in Virtual Private Networks. MPLS Support of Differentiated Services. A Framework for IP Based Virtual Private Networks. An Expedited Forwarding phb Bottleneck Flow Control. Implementation of a Bandwidth Broker for Dynamic End-to-End Resource Reservation in Outsourced Virtual Private Networks Implementation of a Service Broker for Management of QoS enabled VPNs. Core MPLS IP VPN Architecture. Benjamin Teitelbaum and Fairness in Window Flow Controlled Computer Networks. A study of fairness in packet switching networks. Fairness in ATM networks. --TR --CTR Raffaele Bolla , Roberto Bruschi , Franco Davoli, Capacity planning in IP virtual private networks under mixed traffic, Computer Networks: The International Journal of Computer and Telecommunications Networking, v.50 n.8, p.1069-1085, 6 June 2006 Haci A. Mantar , Ibrahim T. Okumus , Junseok Hwang , Steve J. Chapin, A scalable intra-domain resource management architecture for DiffServ networks, Journal of High Speed Networks, v.15 n.2, p.185-205, January 2006
fairness;resource provisioning;Bandwidth Broker BB;connection admission;dynamic configuration;Service Level Agreement SLA;Virtual Private Network VPN;Quality-of-Service QoS;Differentiated Services DiffServ