text stringlengths 128 2.05k |
|---|
On the same path lies the study of the order-chaos phase transition in random Boolean networks (RBNs), which have been used as models of gene regulatory networks |
, carried out by Wang et al. . This work characterizes the RBN dynamics via a phase diagram obtained with respect to Fisher information. This novelty offers a natural interpretation of the phase diagram—through a generic measure capturing information-theoretically how much system dynamics can reveal about its control p... |
Another paper that uses RBNs is that of Poblanno-Balp and Gershenson , where the effect of a modular topology is studied on the properties and dynamics of RBNs. Most RBN studies have been made on homogeneous or normal topologies. However, it is known that real genetic regulatory networks have a modular structure. A gen... |
Droop and Hickinbotham use an artificial chemistry to study the properties of networks constructed from mutation patterns observed in nature. The resulting small-world networks offer a balance between random and regular topologies, resonating with the results of Lizier et al. |
. This balance is advantageous for the exlporation of evolutionary space, i.e. evolvability. A second contribution by Brede studies the popular prisoner’s dilemma |
on regular and heterogenous networks with heterogeneous payoff landscapes. Results illustrate the non-trivial relations between network topology and the facilitation of the evolution of cooperation. |
The final paper of the issue presents a study of complex networks that utilizes colored motifs as the building blocks of the networks. Again, the approach turns to information theory as a generic method of choice: the motifs are used to define the information content of the network. Importantly, the colored motif infor... |
The selection of contributions shows that there are many open research avenues in ALife that involve complex networks, such as evolvability, artificial chemistries, social networks, game theory, genetic regulatory networks, and information theory. We hope that the papers in this special issue will further motivate the ... |
Acknowledgements We should like to thank all the reviewers for this special issue for their timely responses and useful comments: Chris Adami, Lee Altenberg, Alain Barrat, Randall Beer, Hugues Bersini, Johan Bollen, Markus Brede, Mikhail Burtsev, Alan Dorin, Nic Geard, Carlos Gershenson, Mario Giacobini, Juan Luis Jimé... |
# Source: arxiv 1105.4971 # Title: Distributed Evolutionary Computation using REST # Sections: all # Downloaded: 2026-03-03T01:57:18.593727+00:00 |
Distributed Evolutionary Computation using REST Abstract This paper analises distributed evolutionary computation based on the Representational State Transfer (REST) protocol, which overlays a farming model on evolutionary computation. An approach to evolutionary distributed optimisation of multilayer perceptrons (MLP)... |
Introduction Service Oriented Architecture (SOA) is a paradigm for organizing and utilizing distributed computational resources, called services. Using this paradigm, the service providers publish the descriptions (or interfaces) of the services they offer in a service registry, so that the service requesters can disco... |
Although there are several technologies for developing web services (SOAP, REST or XMLRPC among others ), nowadays the main approaches are SOAP (Simple Object Access Protocol) |
and REST (Representational State Transfer) SOAP is the traditional, standards-based approach, but the majority of the web services with public API offer REST interfaces, while some of them offer both REST and SOAP and very few offer just SOAP. |
All of the major Web Services providers use REST: Twitter, Yahoo’s, Flickr, del.icio.us, pubsub, bloglines, technorati, and several others. Both eBay and Amazon have Web Services for both REST and SOAP. |
On the other hand, SOAP Web Services are used in lots of enterprise software as well; for example, Google implements their Web Services using SOAP, with the exception of Blogger, which uses XML-RPC, an early and simpler pre-standard of SOAP. |
The philosophies of SOAP and RESTful Web Services are very different. Strictly, SOAP is a protocol for distributed computing, whereas REST adheres much more closely to a web-based design. SOAP requires a greater implementation and understanding effort from the client side in difference to REST based APIs, which focuses... |
It is important to note that one of the advantages of SOAP is the use of a ”generic” transport. While REST today uses HTTP/HTTPS, SOAP can use almost any transport to send the request. However, one perceived disadvantage is the use of XML because of its verbosity, and the time necessary to parse it. |
This work continues with our previous research in service oriented algorithms, as previously stated in , where a service-oriented platform was presented, or |
, where studies about P2P distributed evolutionary algorithms were performed. In this paper we propose using REST for distributed computation, demonstrating how it could be used for evolutionary computation. Our aim is to implement a distributed evolutionary algorithm (EA) using Perl and REST, to solve a costly problem... |
(QP) to solve classification problems. This paper continues the research on evolutionary optimisation of MLP ( G-Prop method) presented in |
This method leverages the capabilities of two classes of algorithms: the ability of EA to find a solution close to the global optimum, and the ability of the back-propagation algorithm (BP) to tune a solution and reach the nearest local minimum by means of local search from the solution found by the EA. Instead of usin... |
The main idea of this paper, which is basically a proof of concept, is to see what are the possibilities of this setup as a meta-computer by implementing an EA using it, and then measuring the speedup when several computers are used at the same time. The problem we will attempt to solve is a costly classification probl... |
The rest of this paper is structured as follows: Section II presents a comprehensive description of REST technology. Main paradigms of parallel and distributed evolutionary computation are reviewed in Section III Then, section IV describes the proposed method, based on a farming model. Section details the experimental ... |
II REST: Representational State Transfer After some years, Internet architects have found an alternative method for building web services in the form of Representational State Transfer (REST) |
REST is a style of software architecture for distributed hypermedia systems such as the World Wide Web. The term Representational State Transfer was introduced and defined in 2000 by Roy Fielding in his doctoral dissertation |
. Fielding is one of the principal authors of the Hypertext Transfer Protocol (HTTP) specification versions 1.0 and 1.1 REST-style architectures consist of clients and servers. Clients initiate requests to servers; servers process requests and return appropriate responses. Requests and responses are built around the tr... |
Although REST was initially described in the context of HTTP, is not limited to that protocol. RESTful architectures can be based on other Application Layer protocols if they already provide a rich and uniform vocabulary for applications based on the transfer of meaningful representational state. RESTful applications m... |
In a REST environment, clients are not concerned with data storage, which remains internal to each server, so that the portability of client code is improved. Servers are not concerned with the user interface or user state, so that servers can be simpler and more scalable. Servers and clients may also be replaced and d... |
The client-server communication is further constrained by no client context being stored on the server between requests. Each request from a client contains all of the information necessary to serve the request, and any session state is held in the client. The server can be stateful; this constraint merely requires tha... |
Main REST web services features are: Simple and lightweight (not a lot of extra XML markup) Human readable format Easy to build (no toolkits required) |
High performance III Parallel and Distributed Evolutionary Algorithms We are concentrating on parallel and distributed evolutionary computation applications, which has already been adapted to several paradigms of parallel and distributed computing (for example, Jini |
, JavaSpaces , Java with applets , MPI , service oriented architectures and P2P ). There are many ways to implement a distributed EA, one of which is the island model (migration): the population is divided into small subpopulations of the same size assigned to different processors. From time to time each processor sele... |
, Pettey et al. , Cantú-Paz and Goldberg ). Another alternative implementation is global paralelization ( farming , in which individual evaluation and/or genetic operator application are parallelized. The global model does not divide the population. Instead, such an approach employs the inherent parallelism of evolutio... |
Although many approaches to distributed EAs can be found in bibliography, in this paper we do not intend to innovate in that sense, but in the implementation (because implementation matters |
). IV Master-Slave based EA implementation using REST and Perl An ideal client-server implementation of a distributed EA could be a server process with several threads. Each thread would include a population, and would communicate with other threads through the shared code among them. Each thread would use an own tail ... |
However, as we cannot use a threaded version of the Perl modules, our implementation will focus on the most time consuming operation in G-Prop: the fitness function evaluation. The whole evolutionary algorithm is run on the master and only the objective function is sent to the slaves for evaluation (as shown in Figure ... |
The whole system can be sketched as follows: 1. The EA process sends the fitness function code to the REST server and creates the EA population. |
2. Some clients connect the REST server and load the fitness function sent as Perl code from the server (Figure shows an example of server and client processes implementations to upload and download the fitness function source code). |
3. The EA process sends non-evaluated individuals to the server. 4. The clients ask for individuals to the server in order to evaluate them. |
5. The clients evaluate individuals and send the result back to the server. 6. The EA process obtains evaluated individuals from the server and continues the evolutionary loop. |
7. The EA terminates after a fixed number of generations (it sends a termination message throughout the REST server to the clients that remain ready to attend new workloads). |
The server in these experiments is mainly used for scheduling and balancing the tasks among the different clients; the network itself is used for communication, but all the interchange of information among clients must be cleared by the central server. However, one of the objectives of the work presented in this paper ... |
Implementation was carried out using the Perl Dancer module for the Perl programming language, for its stability and the familiarity of the authors with this language |
. In addition, servers are easy to implement and deploy. As an example, Figure shows the source code of a REST server that deploys three services, and a client that obtains the fitness function (as Perl source code) by calling the corresponding service (and then evaluates that function). |
The evolutionary algorithm has been implemented using the Algorithm::Evolutionary (A::E) library . Version 0.76.2 is used in this work, available at |
under GPL license. The full source code (servers, GA and evaluators) and experiment data are available under GPL at: In this work, we adapt G-Prop as a distributed EA using REST following the detailed structure. |
G-Prop method has been fully described and analysed out in previous papers (see ), thus we refer to these papers for further details. In most cases, evolved MLP should be coded into chromosomes to be handled by the genetic operators, however, G-Prop uses no binary codification, instead, the initial parameters of the ne... |
Only “default” parameters have been used (genetic operators were applied using the same application rate). No parameter tuning has been done, since we do not intend to find the optimal ones, but to prove feasibility of the implementation. |
Experimental setup and results The tests used to assess the accuracy of a method must be chosen carefully, because some of them (toy problems) are not suitable for certain capacities of the BP algorithm, such as generalization |
. Our opinion, along with Prechelt , is that, in order to test an algorithm, real world problems should be used. V-A The ”Glass” Classification Problem |
This problem consists of the classification of glass types, and is also taken from . The results of a chemical analysis of glass splinters (percent content of 8 different elements) plus the refractive index are used to classify the sample to be either float processed or non float processed building windows, vehicle win... |
The main data set was divided into three disjoint parts, for training, validating and testing. In order to obtain the fitness of an individual, the MLP (in the slave processes) is trained with the training set and its fitness is established from the classification error with the validating set. Once the EA (in the mast... |
Up to 4 computers have been used to run the algorithm and to obtain results both in sequential and parallel versions of the program. Experiments were conducted running the server process on a Ubuntu/Linux machine, while the clients were run on a Windows 7 with the Cygwin environment and on Ubuntu/Linux machines. Comput... |
As stated before, the EA was executed using the ”default” parameter values (shown in Table ). V-B Obtained Results Time was measured using the ”gettimeofday” function in order to achieve a good precision. Time taken to run the EA is reported in Table II Sequential version of the program was run in the faster machine; a... |
Results obtained can be shown in Table II Classification errors show a comparable algorithmic result. However, better results in time are obtained parallelizing the problem between several computers. |
Figure shows that speedup does not equals the number of computers used; however, simulation time is improved using several computers. Thus, as adding new evaluators (heterogeneous computers running a Perl process) is an easy and costless task, we could take advantage of this system structure to solve costly optimizatio... |
VI Conclusions and Work in Progress This paper presents a new parallel-distributed computation implementation using REST and web services that shows the useful this new technology can be in the field of evolutionary computation. |
To implement and use communications using REST it is not necessary running virtual machines (as in Java programming), nor daemons, just only to install several libraries available for almost any programming language. Moreover, an arbitrary number of computers (clients-evaluators) can be added to the system, making it m... |
In these experiments, we have demonstrated that REST can be used as communication protocol for distributed evolutionary computation, obtaining a good speedup. Results could improve using a dedicated communication network instead of the overloaded network of the university. |
REST provides a common interface that can be called from almost any programming language. Thus, programs can be written in any language and can share data without the need of worrying about the message formats or communication protocols. |
At the same time, it does not overload too much the network. Using other distributed systems, such as Jini , the network traffic is so high that when a high number of computers are used, communication becomes difficult. |
A future in which different remote computers offer services to the scientific community can be imagined: for example, all the services available at the moment by means of HTML forms could be implemented easily as services. |
As future research, it is very important adding support for REST to existing distributed EA libraries in order to allow the implementation of multi-language EAs. Another possibility is to test P2P architectures, where each computer communicates only with one or two computers in the network. It would be very interesting... |
Acknowledgements This work has been supported in part by the CEI BioTIC GENIL (CEB09-0010) MICINN CEI Program (PYR-2010-13) project, the Junta de Andalucía TIC-3903 and P08-TIC-03928 projects, and the Jaén University UJA-08-16-30 project. |
# Source: arxiv 1105.6205 # Title: Cloud-based Evolutionary Algorithms: An algorithmic study # Sections: all # Downloaded: 2026-03-03T01:56:48.841042+00:00 |
Cloud-based Evolutionary Algorithms: An algorithmic study Abstract After a proof of concept using Dropbox™, a free storage and synchronization, showed that an evolutionary algorithm using several dissimilar computers connected via WiFi or Ethernet had a good scaling behavior in terms of evaluations per second, it remai... |
Introduction The main objective of this research is to find easily available means to either use or connect computing nodes in a distributed evolutionary computation experiment, and this often means resorting to free and readily available services. Dropbox™is one of these services: it is commercialized as a cloud stora... |
In the experiments we are performing, we are interested in its use as a file synchronization service. When one file in one of the folders that is monitored by Dropbox is changed, it is uploaded to Dropbox servers and then distributed to all the clients that share the same folder. It is interesting, however, to note tha... |
In previous experiments we measured whether adding several computers to an experiment of this kind resulted in an increase in the number of simultaneous evaluations. In this paper we will measure whether besides an increase in speed, the algorithm profits from the distribution and asynchrony of the particular instance ... |
The rest of the paper is organized as follows: after a brief section presenting the state of the art in voluntary and pool-based evolutionary computation, we describe the algorithm, the experimental setup and the implementation in Section |
III ; results of these experiments will be briefly presented in Section IV , to be followed by the conclusion, discussion and future lines of work in Section |
II State of the Art Cloud computing is an emergent technology, and as such research related to it is just recently emerging. Research addressing cloud storage is mainly related to content delivery |
or designing data redundancy schemes to ensure information integrity . However, its use in distributed computing has not been addressed in such depth. Even if it is related to data grids |
, in this paper we address the use of free cloud storage as a medium for doing distributed evolutionary computation, in a more or less parasitic way |
, since we use the infrastructure laid by the provider as part of an immigration scheme in an island-based evolutionary algorithm |
Thus we will have to look at pool-based distributed evolutionary algorithms for the closest methods to the one presented here. In these methods, several nodes or islands share a pool where the common information is written and read. To work against a single pool of solutions is an idea that has been considered almost f... |
were proposed in the early nineties as a cooperative scheme for autonomous agents. The basic idea is to create a work-flow on a set of solutions and apply several heuristic techniques to improve them, possibly including humans working on them. This technique is not constrained to evolutionary algorithms, since it can b... |
The A-Team method does not rely on a single implementation, focusing on the algorithmic and data-flow aspects, in the same way as the Meandre |
system, which creates a data flow framework, with its own language (called ZigZag), which can be applied, in particular, to evolutionary algorithms. |
While algorithm design is extremely important, implementation issues always matter, and some recent papers have concentrated on dealing with pool architectures in a single environment: G. Roy et al. |
propose a shared memory multi-threaded architecture, in which several threads work independently on a single shared memory, having read access to the whole pool, but write access to just a part of it. That way, interlock problems can be avoided, and, taking advantage of the multiple thread-optimized architecture of tod... |
, who present a design pattern for persistent and distributed evolutionary algorithms; although their emphasis is on persistence, and not performance, they try to present several alternatives to decouple population storage from evolution itself ( traditional evolutionary algorithms are applied directly on storage) and ... |
is also similar, using for persistence a small database accessed through a web interface, but only for the purpose of interchanging individuals among the different nodes, not as storage for the whole population. |
In fact, the efforts mentioned above have not had much continuity, probably due to the fact that there have been, until now, few (if any) publicly accessible online databases. However, given the rise of cloud computing platforms over the last few years, interest in this kind of algorithms has bounced back, with impleme... |
having been recently published. III Description of the algorithm and implementation A pool based evolutionary algorithm can be described as an island model |
without topology; in fact, it is closer to the island metaphor since migrants are sent to the sea (pool), and come also from it, that is, the evolutionary algorithm is a canonical one with binary codification, except for two steps within the cycle that (conditionally) emit or receive immigrants. A minimum number of eva... |
During the evolutionary loop, new individuals are selected using 3-tournament selection and generated using bit-flip mutation and uniform crossover. Migration is introduced in the algorithm as follows: after the population has been evaluated, migration might take place if the number of generations selected to do it is ... |
Migrants, if any, are incorporated into the population substituting the worst individual, along with the offspring of the previous generation using generational replacement with a 1-elite. Population was set to 1000 individuals for all problems, and the minimum number of evaluations has been four million. Several migra... |
One of the advantages of this topology-less arrangement is the independence from the number of computers participating in the experiment, and also the lack of need from a central server, although it can be arranged so that one of the computers starts first, and the others start running when some file is present. Adding... |
Two representative functions have been chosen to perform the tests; the main idea is that they took a long enough time to make sense in a distributed environment, but at the same tame a short enough time that experiments did not take a long time. One of them is |
P-Peaks , a multimodal problem generator proposed by De Jong in ; a P-Peaks instance is created by generating [MATH] random [MATH] -bit strings where the fitness value of a string [MATH] is the number of bits that [MATH] has in common with the nearest peak divided by [MATH] |
[EQUATION] where [MATH] is the Hamming distance between binary strings [MATH] and [MATH] . In the experiments made in this paper we will consider [MATH] and [MATH] . Note that the optimum fitness is 1.0. |
The second function is MMDP , which is a deceptive problem composed of [MATH] subproblems of 6 bits each one ( [MATH] ). Depending of the number of ones (unitation) [MATH] takes the values depicted next: |
[MATH] [MATH] [MATH] [MATH] [MATH] [MATH] [MATH] The fitness value is defined as the sum of the [MATH] subproblems with an optimum of [MATH] (equation ). Figure represents one of the variables of the function. The number of local optima is quite large ( [MATH] ), while there are only |
[MATH] global solutions [EQUATION] In this paper we have used [MATH] , in order to make it difficult enough to need a parallel solution. |
IV Experiments and results In this occasion the experiments were done in several different computers connected also in different ways; however, computers were added to the experiment in the same order; the problems were solved first in a single computer, then on two and finally with four computers. Total time, as well ... |
The only parameter that was changed during experiments was migration rate. We were interested in doing this, since network performance will be impacted negatively with migration rate: bandwith usage (and maybe latency) increases with the inverse of the migration rate. On the other hand, evolutionary performance will in... |
To keep (the rest of) the conditions uniform for one and two machines, all parameters were fixed but for the population, which was distributed among the machines in equal proportions: all computers maintained a population of 1000, so that initial diversity was roughly the same. Further experiments will have to be made ... |
Finally, Dropbox itself was used to check for termination conditions: a file was written on the folder indicating the experiment had finished; when the other node read that file, it finished too; all nodes were kept running until the solution was found or until a maximum number of generations were reached. That is why,... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.