Learning useful communication structures for groups of agents

Abstract

Coordination of altruistic agents to solve optimization problems can be significantly enhanced when inter-agent communication is allowed. In this paper we present an evolutionary approach to learn optimal communication structures for groups of agents. The agents learn to solve the Online Partitioning Problem, but our ideas can easily be adapted to other problem fields. With our approach we can find the optimal communication partners for each agent in a static environment. In a dynamic environment we figure out a simple relation between each position of agents in space and the optimal number of communication partners. A concept for the establishment of relevant communication connections between certain agents will be shown whereby the space the agents are located in will be divided into several regions. These regions will be described mathematically. After a learning process the algorithm assigns an appropriate number of communication partners for every agent in an - arbitrary located - group. © 2006 International Federation for Information Processing.