Memory Induced Aggregation in Collective Foraging

Balancing Collective Exploration and Exploitation in Multi-Agent and Multi-Robot Systems: A Review

“…Conversely, higher values of μ cause an agent to preferentially carry out shorter walks between changes of direction, characteristic of higher levels of exploitation (Viswanathan et al, 2000;Viswanathan et al, 2008). The use of the Lévy parameter has been used as a method of quantifying and controlling the level of exploration and exploitation of various MRS applied in different tasks, ranging from foraging (Zedadra et al, 2019;Nauta et al, 2020a;Nauta et al, 2020b;Nauta et al, 2020c), to target search (Senanayake et al, 2016;Harikumar et al, 2019;Pang et al, 2019), to area mapping (Ramachandran, 2018;Kegeleirs et al, 2019;Ramachandran et al, 2020).…”

“…In addition to the use of Lévy distributions, action probabilities can be used to set the level of exploration and exploitation carried out by an MRS. To search for a target Matignon and Simonin (2018), used action probabilities to bias an agent in favor of carrying out either exploratory or exploitative actions. In a similar fashion, Falcón-Cortés et al (2019), Nauta et al (2020a) and Nauta et al (2020b) uses the probability of an agent using information in its memory as means of determining and controlling the level of exploration and exploitation of the system. In their best-of-n problem Prasetyo et al (2019), used an opinion switching probability threshold as a proxy to measure and control the exploration and exploitation dynamics of their swarming system.…”

“…When deploying changes in agent states, a large change in an agent's behavior is effected in response to a given stimulus observed by either an agent, a subset of agents or the entire MAS itself. This gives agents the ability to switch from an exploratory to an exploitative behavior, and vice-versa allows Nauta et al (2020a), Nauta et al (2020b) Lévy parameter and memory utilization probability Resource Foraging de Lope et al (2015), Lee and Kim (2019), Kazakova et al (2020), Lee et al (2020) Task switching probability threshold Task Allocation Viswanathan et al (2000), Viswanathan et al (2008) for them to better adapt to their local environment, with the aim to improve the overall performance of the system. These strategies tend to base their exploration strategies on random walks (Zedadra et al, 2017), followed by information exploitation driven by either a metaheuristic algorithm or direct travel towards a point of interest should one be detected.…”

“…In their review, Stadler et al (2014) posited that exploration and exploitation are associated with a certain level of learning. Therefore it can be said that an agent is carrying out exploitation if it uses previously acquired knowledge and carrying out exploration if it is not doing so (Gupta et al, 2006;Nauta et al, 2020b). Again, this is another common method of defining exploration and exploitation in MAS, where certain practitioners have developed dedicated behaviors that facilitate exploration and exploitation separately (see Section 3.1.1.1).…”

Tracking Multiple Fast Targets With Swarms: Interplay Between Social Interaction and Agent Memory

“…It was previously determined that the use of memory was counterproductive as its usage resulted in the exploitation of outdated information, causing swarm aggregation in a location at which the target is no longer present (Coquet et al, 2019;Kwa et al, 2020a). Despite these disadvantages, memory usage has been shown to encourage the aggregation of agents around high quality target patches in static non-destructive foraging tasks (Falcón-Cortés et al, 2019;Nauta et al, 2020b). In the pursuit of an evasive target that moves faster than any individual agent, the use of agentbased memory gives the swarm a longer lasting point of attraction.…”