text stringlengths 128 2.05k |
|---|
Breeding in our model occurs by sexual reproduction so breeding is a social activity. In our environment two agents must be at the same site at the same time to breed. If both agents are performing the breed action for overlapping periods of time then a successful sexual reproduction occurs. We can see that more time s... |
During sexual reproduction an offspring’s genome is created as a recombination of its parents’ genomes. Recombination uses the longest common subsequence of the two genomes. The offspring’s genome also has an opportunity to mutate in this process (see Marriott and Chebib, 2015b for more details of genetic mechanisms). ... |
At birth each agent’s genome creates a memome. A memome consists of a collection of memeplexes. Each memeplex in our model represents a possible set of activities for a single day. Memeplexes are subsequences of a genome. During memome generation we start at each gene in a genome and we copy gene by gene into a memeple... |
Additionally, segments are copied in a backwards direction from every gene. This means every gene in a genome is responsible for two memeplexes in its memome except the endpoints that are responsible for only a single memeplex. Notice that since every site in the environment is not necessarily represented in a genome t... |
Memome Our agent’s cognitive model is inspired by the pandemonium model (Jackson,, 1987 ; Franklin,, 1997 ; Marriott et al.,, 2010 . Each memeplex is a sub-path of a genome and thus is a path in the random geometric network. A memeplex represents a single day’s worth of activities. In the pandemonium model a memeplex i... |
Behavior selection is also quite similar to the MAP-elites strategy for multi-objective evolutionary optimization (Mouret and Clune,, 2015 . We have memeplexes organized in the memome based on the starting site. An agent’s day begins by selecting all memeplexes in its memome that begin at the agent’s current site. Reca... |
This is the primary force of cultural selection in our current implementation. It means that memeplexes with maximum resource-to-energy ratio are selected as behaviors. Since in our social learning mechanism agents that engage in social learning share only the current day’s memeplex it means this selection mechanism is... |
An agent can engage in individual learning only if its selected memeplex includes at least one meme that has a non-zero learning component. This means that an agent must spend time engaged in learning at at least one site during a day. When this occurs an agent will clone its memeplex for the day and apply a mutation. ... |
We can see that having non-zero learning components in memes would benefit the agent. However, spending time learning during the day comes at an energy cost, as with breeding. Further, as our current implementation only allows a single learning event in a day it is not beneficial for an agent to spend more than the min... |
An agent can engage in social learning once per day. The process for social learning is very similar to the process for sexual reproduction. Social learning can only occur if an agent spends time engaged in a social learning action at least one site during a day. However, for social learning to occur, another agent mus... |
Again we see a benefit to having social learning in memes as this will increase the chances of an exchange of memeplexes. However, as with learning and breeding, an agent cannot afford to spend too much time performing social learning during a day. |
Both of these learning mechanisms allow for new memeplexes to be added to the memome which means an agent can adapt its behavior. When it begins its day at the same site again it may select one of the new memeplexes instead. |
Experimental Setup We conduct three similar experimental runs with agents of different capabilities. The first control group we call breeders, and while they still have memomes, their learning and social learning mechanisms have been turned off. In these agents, learning and social learning components of genes/memes ar... |
Each run is seeded with one hundred randomly generated agents. All genes in the randomly generated genome have learning and social learning components initialized to zero. Since it would be impossible for agents to breed if this were true of breeding components, we instead have a chance of initializing breeding compone... |
We gather data on many aspects of our agents’ lives. In particular we gather information on the proportion of a genome or memeplex devoted to breeding, learning or social learning. Recall that genomes and memeplexes are both paths of sites and they can vary in length from agent to agent. We can measure the length of a ... |
We use a slightly different method to calculate the optimization of a genome/memeplex. For each site there is a gathering action. Recall the action will cost energy at least the number of resources rewarded at a site (one, two or three) and at most five. We can count all energy used above the minimum as wasted energy. ... |
The control groups of breeders and learners should not display divergent cumulative cultural evolution. Among the control groups we expect differences in genome and memeplex measurements but we expect these differences to be small. In the socializers we expect to see evidence of divergent cumulative cultural evolution.... |
We also expect to see divergence under competitive selection pressures (i.e. when they pull in different directions). We expect that genetic selection will select for breeding while also having indirect selection for learning and social learning. We expect that there will remain some selection pressure for social learn... |
Observations and Discussion Our experiment was replicated 130 times on a variety of random geometric networks. All data presented in this section is averaged over these 130 runs. |
We wish to begin with a discussion of an apparent slowdown of genetic evolution caused by cultural evolution. When we first investigated the breeding, learning and social learning genes in the genomes of socializers we found that gene concentrations for these components grow at the same rate for our two control groups ... |
We saw a similar effect on genome length over time and other data we gathered. However with further investigation we were able to determine the source of the slowdown. In both control groups the generation of agents increased at identical rates. The generations of socializers increased at about half the rate (Figure ).... |
We think this evidence suggests that cultural evolution can slow genetic evolution over time, but possibly not over generations. That is, the shielding of genetic selection pressure is not really there. Instead there is a selection pressure for longer generations which has the result of slowing genetic evolution over t... |
To confirm this we plotted gene concentrations over generations instead of days (Figure ). We can see that the socializers actually have weak acceleration of evolution over time according to this plot. These differences are small. We suspect that a greater significant difference might occur if we allowed our simulation... |
Now we wish to consider evidence for divergence under cooperative selection pressures. So we turn to a discussion of the relative optimization of genomes and memeplexes. Breeders and socializers showed a slight trend to less optimized genomes over time (Figure ). This is not unexpected as much of the genome is not actu... |
In the socializers there is also an early stage of optimization before stagnation. However the stage of optimization is considerably greater in socializers than non-socializers. Stagnation in the control runs is in part due to a weak genetic selection pressure for optimization. Selection pressure for optimization in th... |
Now let’s consider divergence under competitive selection (see Fig ). The strongest competitive pressure is relative to breeding actions. Treating memeplexes as a type of parasitic organism we can see that they are only concerned with their replication into new hosts. They are not concerned with their host’s reproducti... |
We observe that soon after social learning emerges in the population memeplexes diverge from the genome. Breeding actions in a genome continue to be selected for, but breeding actions in a memome are selected against and their concentration decreases before stagnating. Notice again this stagnation is due in part to the... |
We observe that in many socializers, breeding actions in their memeplexes are reduced to zero. This is also clear by noting that about 54% of breeders and learners have children while only 36% of socializers do. Memeplexes co-opt the agent for their reproductive ends instead of the genome’s reproductive ends. |
Interesting cases of collapsed colonies occur when these memeplexes spread to all agents before they can breed. In many runs of socializers we witness isolated colonies completely dying out. Out of 100 runs 21 ended when all agents died out before the 5000 day limit is reached. This only occurs in socializers. |
In breeders and learners, colonies can face extinction due to a shortage of resources caused by overpopulation. In these situations agents can’t get enough resources to reproduce and in some cases can’t get enough to survive. However this situation cures itself as agents die out. As agents die, they no longer collect r... |
Consider learning and social learning actions. There is strong early selection pressure in genomes and memomes for these actions. However as social learning kicks in, memetic selection against wasting time on these action takes over. Remember it is beneficial to spend as little time as possible on learning and so optim... |
It is also beneficial to not waste time on social learning. As above when social learning begins optimization places a negative selection pressure on social learning actions. However it is a memeplex’s responsibility to spread itself. A memeplex that evolves to spend no time on social learning will never be spread (but... |
We notice for all breeding, learning and social learning actions as well as for optimization, cumulative cultural evolution causes a divergent effect from genetic evolution. In our simulations, this period of time occurred early on (in the first 500 days) and then social learning maintains an optimized culture through ... |
To track cumulative cultural evolution we also assigned a generation to each memeplex. All original memeplexes created at birth are assigned generation zero. Whenever it is cloned in learning or social learning the clone is assigned generation one higher than its parent. |
We can detect cumulative cultural evolution by detecting an increase in memeplex generation over time, especially from generation to generation (Figure ). We can indeed confirm cumulative cultural evolution in our socializers by this method. Of course breeders can only have a memeplex generation of zero. Learners howev... |
Interestingly, in the socializers we do notice two stages in memeplex generation growth. The initial stage is more rapid and the second stage grows slowly at a fixed rate over time. We note that the period of rapid growth coincides with the period in which the memeplexes are being optimized prior to stagnation. The slo... |
Conclusions Our implementation incorporated two information stores, one for genetic information and one for cultural information. It also had mechanisms of horizontal transfer of cultural information between agents of multiple generations. Our implementation is an example of our dual inheritance model of cultural evolu... |
We can see that populations of agents that participate in the dual inheritance model can accelerate optimization relative to selection pressures that are cooperative between the genetic and memetic world. This is due to the potential for many cultural generations in a single biological generation. Thus, optimization ca... |
We also see cases where the selection pressure on the genome operates in an opposite direction for the memome. Genes and memes care only for spreading themselves. For genes, spreading occurs through reproductive events but for memes spreading occurs through social learning events. So it is not surprising that genes tha... |
The contrary is not necessarily true. Genes that increase the number and success of social learning events may be selected for by biological evolution if social learning also helps improve survival or reproductive success. This is true in our simulation. Memes that increase the number and success of reproductive events... |
Finally, we still see an interesting divergence in behavior of young (inexperienced) agents and old (experienced) agents. Young agents have had very little or no time to adapt their initial set of memeplexes either through individual optimization or through learning from others. Thus their behavior is still largely det... |
The interesting impact of these trends is that young agents are more likely to be parents (i.e. before they learn better). They are more likely to learn from the environment more than older experienced agents. Finally, they are also more social. They are more likely to seek out social learning events than old agents. W... |
Finally we believe that the observed divergent cumulative cultural evolution is due to a critical component of the dual inheritance model. Specifically we think it is critical to keep genetic and cultural information separate from one another even if they store the same kinds of information (as in our implementation). ... |
Acknowledgments The authors would like to thank the advice of anonymous reviewers. Jobran Chebib was supported by the Swiss National Science Foundation (grant PP00P3_144846/1 awarded to Frédéric Guillaume). |
# Source: arxiv 1612.00480 # Title: A Scalable and Adaptable Multiple-Place Foraging Algorithm for Ant-Inspired Robot Swarms # Sections: all # Downloaded: 2026-03-03T02:01:23.800379+00:00 |
A Scalable and Adaptable Multiple-Place Foraging Algorithm for Ant-Inspired Robot Swarms* Abstract Individual robots are not effective at exploring large unmapped areas. An alternate approach is to use a swarm of simple robots that work together, rather than a single highly capable robot. The central-place foraging alg... |
INTRODUCTION A large number of simple individual robots working together has the potential to be useful for tasks which a traditional single expensive, specialized and complicated robot is not able to handle, such as searching in large unmapped areas |
, distributed contaminant cleanup, and rescue . Robot swarms can also be involved in sophisticated problem solving, including cooperative transportation, de-mining, and space exploration |
We focus on developing a scalable, decentralized search-and-collection algorithm based on ant-like foraging . The swarm can adapt to changes in swarm size and the number of targets through real-time response to conditions without external or off-line intervention. Each robot in the swarm makes real-time in-situ decisio... |
are restricted to foraging in short-time windows during which not all available targets can be collected; So they are designed to collect as many targets as possible, but not for optimal complete collection |
Here, we present the multiple-place foraging algorithm (MPFA) with multiple nests that robots depart from and return to. The robots make on-line decisions to switch to new collection zones based on proximity to their last-found target. The MPFA was presented in our recent work |
and it shown that distributing 2, 4, or 8 nests in the MPFA produce higher foraging rates and lower average travel time compared to the central-place foraging algorithm (CPFA) developed by Hecker and Moses |
. Here we compare the scalability and adaptation of the MPFA to the CPFA when increases the number of robots and the number of targets. In the MPFA we deploy 4 nests uniformly in the same size of search arena. A set of real-valued parameters specifying the individual robot controllers is evolved by a genetic algorithm ... |
. Every robot in the swarm uses the same controller. We evolve foraging strategies for different swarm sizes (4, 8, 16, 32 and 64) and number of targets (128, 256, 512, 1024 and 2048). We observe the average foraging rate, collision time, travel and search time change as swarm size and the number of targets increase. |
The remainder of this paper is organized as follows. Section II introduces related work. The design of the MPFA and the description of evolution are provided in Section III and Section IV . The configuration of the MPFA in ARGoS and the experimental results are in Section and Section VI . Section VII discusses the conc... |
II RELATED WORK Central-place foraging is commonly studied in swarm robotics . Hecker and Moses utilized and formalized the behaviors from Flanagan and Letendre’s ant field studies |
to create the CPFA. The algorithm is well designed and applied to real physical robots, which are designed on the iAnt robots platform |
. The error-tolerance, flexibility, and scalability were evaluated on both simulated and real robot swarms . However, the simulated robots were not physics-based and collisions between robots were not considered. |
The studies on task allocation by Hsieh et al showed that a bio-inspired approach to the deployment of a homogeneous swarm of robots to multiple sites. The robots autonomously redistribute themselves among the candidate sites to ensure task completion by optimized stochastic control policies. It models the swarm as a h... |
There are few studies on multiple-place foraging in biological systems. The polydomous colonies of Argentine ants are comprised of multiple nests spanning hundreds of square meters |
. A study by Chapman et al showed that a community of spider monkeys can be considered as multiple central place foragers (MCPF). They select a sleeping site close to current feeding areas, and the MCPF strategy entails the lowest travel costs. A study by Tindo et al |
showed that wasps living in multiple nests have greater survival rate and increased productivity. However, multiple-place foraging has not been systematically compared to central placed foraging in robotic swarms which we do here. |
III THE DESIGN OF THE MPFA In the MPFA, robots are evenly distributed around nests. They start from a nest but return to the closest nest to their position after finding a target or giving up the search. The use of multiple collection points is the fundamental difference between the CPFA and the MPFA; all other compone... |
The behavior of an individual robot in an MPFA foraging round is shown in Fig. . Each robot transitions through a series of states as it forages for targets. This differs from the CPFA |
in how the robots return to nests which are in steps 4 and 5. In the MPFA, robots initially disperse from the nests closest to them, followed by random selected travel paths (step 1). An uninformed correlated random walk is used to search targets when robots stop to follow the paths (step 2) |
. Robots navigate home to nests closest to them when they retrieve targets or give up search (step 4 and 5) . Robots that find targets will detect the local target density before return to nests (step 3) |
. Robots that are more likely to return to previously found sites using site fidelity or pheromone recruitment (step 6), then they search the sites thoroughly with informed walk (step 7). |
In our design, the robots search globally just as in the CPFA – they can travel in the entire arena. As in the CPFA, pheromone trails are simulated using a list of pheromone-like waypoints to identify target-rich areas. When a robot returns to a nest, it will probabilistically select a waypoint from the nest’s list and... |
Since robots always return to the closest nest with a found target, the sensed information relevant to a given target neighborhood is always associated with the nest closest to the position of the identified neighborhood. Thus, if a robot follows a pheromone waypoint from a nest, then the distance from the nest to the ... |
IV The GA EVOLUTION We implement the CPFA and MPFA on a foraging task for different experiments in ARGoS. Furthermore, we use a GA to identify MPFA parameters that maximize foraging strategy. We implement our GA using GAlib |
following parameters described by Hecker and Moses . The set of seven MPFA parameters is identical to the set of CPFA parameters. The movement, sensing, and communication of each single robot are evolved and evaluated. The parameters are described in the following, |
Probability of switching to search: The robot has the probability of switching from travel to uninformed random search. The probability is initialized from a uniform random distribution, [MATH] |
Probability of returning to nest: The robot has the probability of giving up search and returning to nest. It is initialized from a uniform random distribution, [MATH] |
Uninformed search variation: If the robot searches using a correlated uninformed random walk, the successive turning angles [MATH] is defined by [MATH] , where [MATH] is the turning angle in the current step, and [MATH] is the standard deviation or uninformed search variation, which determines the turning angle of the ... |
Rate of informed search decay: If the robot searches using an informed correlated random walk, the standard deviation of the successive turning angles [MATH] decays as a function of time [MATH] [MATH] , where [MATH] is the rate of informed search decay. [MATH] is initialized from an exponential decay function [MATH] |
Rate of laying pheromone and rate of site fidelity: The information decisions are governed by parameterization of a Poisson cumulative distribution function as defined by [MATH] , where [MATH] is the likelihood of detecting at least [MATH] additional resources, and [MATH] is the rate of laying pheromone or the rate of ... |
Rate of pheromone decay: The pheromone decays exponentially over time [MATH] as defined by [MATH] , where [MATH] is the rate of pheromone decay. It is initialized from an exponential decay function [MATH] |
We repeat the evolutionary process 10 times for the CPFA as well as for the MPFA, in order to generate 10 independently evolved foraging strategies for each experimental configuration. |
In summary, using a swarm size of 40 robots, we evaluate each swarm 8 times on different random placements of targets in the partially clustered distribution to determine their fitness. We use a [MATH] uniform crossover rate and a [MATH] Gaussian mutation rate with a standard deviation of [MATH] . We use elitism to kee... |
We altered the termination criteria of the GA in order to hasten parameter convergence and ran the GA for a maximum of 100 generations. The GA terminates based on three criteria: the number of generations, the convergence of fitness, and the diversity of swarm sizes, which are introduced in GAlib |
. The GA will stop if the fitness is convergent and the diversity of the population is low. Otherwise, it will stop after 100 generations. Our code is available on GitHub |
In our GA, [MATH] of the evolution terminates on the convergence of fitness and the diversity of swarm sizes. Across 10 independent evolutionary runs, all evolved parameter sets were nearly equally fit: The standard deviation in fitness was at most [MATH] of the mean fitness value. The fitness of the best parameter set... |
EXPERIMENTAL CONFIGURATION IN ARGoS Table shows the experimental configuration in ARGoS. To test scalability, the number of targets is always 1024, and the number of robots is scaled to be 4, 8, 16, 32 or 64. We set different foraging time windows for each swarm, depending on the swarm size. The selected times allow th... |
To test adaptation, the number of robots is always 32. The number of targets is 128, 256, 512, 1024 or 2048. The foraging time is set independently for each experiment so that approximately [MATH] of the targets are collected by the best evolved strategy. All experiments are replicated 100 times. The locations of targe... |
The targets are placed in a partially clustered distribution. This distribution has various sizes of square clusters. The targets are placed either in a large cluster, a medium cluster or individual targets in a uniform random distribution (see Fig. ). Both algorithms are tested in a simulated arena size of [MATH] . Th... |
VI RESULTS We compare the efficiency of the CPFA and the 4 nest MPFA on foraging rate, collision time, and travel and search time when the swarm sizes and the number of targets are scaled up. We identify statistical differences using a t-test, and we identify whether performance varies systematically by calculating a l... |
VI-A Foraging Efficiency The total foraging rate of each swarm is the sum of the total collected targets in the swarm. We measure the average foraging rate, which is the number of targets per robot collected in every minute. Fig. shows the average foraging rate as the swarm size increases. The average foraging efficien... |
Fig. shows the average foraging rate as the number of targets increases. The average foraging efficiency of the MPFA exceeds that of the CPFA in all cases, by up to [MATH] in the case of 2048 targets. |
VI-B Collision Efficiency In our simulation, if the distance between two robots is less than 0.25m, each robot will detect a collision. Each robot senses the location of the other and turns left or right in order to avoid a collision, moving approximately 8 cm before resuming traveling. |
The collision time is the time required to avoid a collision. The total collision time of each swarm is the sum of the total collision time for all robots in the swarm. We measure the average collision time, which is the collision time per robot in collecting a target. The ”per robot, per target” collision makes the co... |
The collision time for the MPFA is less than the collision time for the CPFA. We also see that the collision time for the MPFA is reduced as the number of targets increases (see Fig. ). |
VI-C Travel and Search Efficiency Foraging time is composed of two distinct activities. When a robot departs from its nest, it travels to a location where it starts to search for targets. Once at the destination, the robot engages in a localized search. Once a target is discovered, the robot takes approximately the sam... |
We measure the average travel time and search time spent to collect one target by a robot. The average travel time for the MPFA (see Fig. ) is less than the CPFA for all swarm sizes. |
The travel time for the MPFA (see Fig. ) is also less than the CPFA as the number of targets increases. Fig. shows that the average search time decreases as the number of robots increases. The search time for the MPFA is less than the CPFA. The search time for the CPFA decreases faster than the MPFA. The improvement is... |
The search time decreases as the number of targets increases (see Fig. 10 ). The search time for the MPFA decreases faster than the CPFA. The improvement goes up to [MATH] in the last case. |
VII DISCUSSION This paper explores how swarm size and the number of targets affect foraging rates, collision time, travel and search time. Not surprisingly, increasing the swarm sizes or the number of targets decreases the average foraging rate (see Fig. and Fig. ), but decreases slower for the MPFA. This implies that ... |
The average collision time for the MPFA is much less than the CPFA as the swarm size or the number of targets increases (see Fig. and Fig. ). The collision time for the CPFA increases faster as the number of targets increases (see Fig. ). We hypothesize that the more targets there are, the harder robots will spread out... |
The increase of swarm sizes makes the average travel time for the MPFA decrease faster than the CPFA (see Fig. ). This shows that the MPFA has the advantage of reducing travel time as the swarm size increases. We hypothesize that the evolved probability of returning to nest increases faster as the swarm size increases.... |
It is obvious that the addition of more nests makes the travel time less for the MPFA. However, the information (pheromone waypoints) is distributed to multiple nests. In contrast to the CPFA, pheromone waypoints are globally available to all robots. So, there are tradeoffs among communication (and therefore search tim... |
The search time for the MPFA decreases faster with increasing numbers of targets (see Fig. 10 ). This shows that the MPFA has the advantage of reducing search time in an environment with large number of targets. We hypothesize that the evolved probability of laying pheromone increases and it is higher for the MPFA, or ... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.