Gal A. Kaminka: Publications

Sorted by DateClassified by Publication TypeClassified by TopicGrouped by Student (current)Grouped by Former Students

Competitive Multi-Swarm Systems

Karen Katz. Competitive Multi-Swarm Systems. Master's Thesis, Bar Ilan University,2023.

Download

[PDF]6.2MB  

Abstract

A competitive multi-swarm system, is a system with two or more distinct swarms of simple agents, with limited local knowledge, sharing the same environment and resources, where each swarm's goal is to outperform the other swarms.In this work, we developed a general model for multiple competitive swarms from game theoretic perspective. We formulated the individual and global utilities for K-swarms competition, based only on a single assumption of zero-sum game between two individual players from different swarms. A special case of two swarms competition is shown to be a zero-sum game, and a possible extension to zero-sum game for the K-swarm case is presented. To show the applicability of the model, the theory is applied into the field of competitive robot swarms. Global and individual utilities, and the estimation of the individual player's impact on its surroundings are presented as a function of times, to support applicability to any generic task. One result from this approach is that a robot can increase its swarm's utility not only by performing its original task, but also by interfering in its opponents' performance of their tasks.We propose a learning process for each individual robot in multi-swarm competition, by calculating its own reward, and providing a general way for evaluation and selection of its possible actions. The proposed learning model tries to overcome the gap due to the partial information known to each robot, by considering the swarm identity of the other robots during each interaction, and approximating differences between the swarms.As an example, the general model is applied for the more specific sub-field of multi-swarm competitive foraging. It examines the model on the unexplored problem of how robots in a competitive multi-swarm environment should interact during spatial conflicts, in order to outperform the other swarms.The proposed model has been validated and tested through an extensive series of simulated experiments, including two- and three-swarm competitions, in various densities, with and without learning. Part of the experiments were expanded for cases of a learning swarm with initial disadvantages. The results show that a learning swarm performed at least equally and usually better than a non-learning swarm, which uses a predefined policy. Surprisingly, in many cases, the overall score of all the swarms together increased when competition was involved.

Additional Information

BibTeX

@mastersthesis{karen-msc,
  author = {Karen Katz},
  title = {Competitive Multi-Swarm Systems}
  school = {{B}ar {I}lan {U}niversity},
  year = {2023},
  OPTkey = {},
  OPTtype = {},
  OPTaddress = {},
  OPTmonth = {},
  OPTnote = {Available at \url{http://www.cs.biu.ac.il/~galk/Publications/b2hd-karen-msc.html}},
  OPTannote = {},
  wwwnote = {}, 
  abstract = {
   A competitive multi-swarm system, is a system with two or more distinct swarms of simple agents, with limited local knowledge, sharing the same environment and resources, where each swarm's goal is to outperform the other swarms.
In this work, we developed a general model for multiple competitive swarms from game theoretic perspective. We formulated the individual and global utilities for K-swarms competition, based only on a single assumption of zero-sum game between two individual players from different swarms. A special case of two swarms competition is shown to be a zero-sum game, and a possible extension to zero-sum game for the K-swarm case is presented. 
To show the applicability of the model, the theory is applied into the field of competitive robot swarms. Global and individual utilities, and the estimation of the individual player's impact on its surroundings are presented as a function of times, to support applicability to any generic task. One result from this approach is that a robot can increase its swarm's utility not only by performing its original task, but also by interfering in its opponents' performance of their tasks.
We propose a learning process for each individual robot in multi-swarm competition, by calculating its own reward, and providing a general way for evaluation and selection of its possible actions. The proposed learning model tries to overcome the gap due to the partial information known to each robot, by considering the swarm identity of the other robots during each interaction, and approximating differences between the swarms.
As an example, the general model is applied for the more specific sub-field of multi-swarm competitive foraging. It examines the model on the unexplored problem of how robots in a competitive multi-swarm environment should interact during spatial conflicts, in order to outperform the other swarms.
The proposed model has been validated and tested through an extensive series of simulated experiments, including two- and three-swarm competitions, in various densities, with and without learning. Part of the experiments were expanded for cases of a learning swarm with initial disadvantages. The results show that a learning swarm performed at least equally and usually better than a non-learning swarm, which uses a predefined policy. Surprisingly, in many cases, the overall score of all the swarms together increased when competition was involved. 
  },
}

Generated by bib2html.pl (written by Patrick Riley ) on Thu Feb 22, 2024 11:36:58