Relating the entropy of joint beliefs to multi-agent coordination

Abstract

Current approaches to activity coordination in multi-agent systems (teams) range from strictly top down (plan-based coordination) to purely emergent (reactive coordination), with many hybrid variants, each having its specific advantages and disadvantages. It appears to be extremely difficult to rigorously compare various hybrid approaches to multi-agent coordination (and communication), given the lack of a generic semantics or some guidelines. In this paper, we studied some intuitive inter-agent communication policies and characterised them in terms of generic information-theoretic properties. In particular, the relative entropy of joint beliefs was suggested as an indicator of teams coordination potential. Our novel behaviour-based agent architecture (based on the Deep Behaviour Projection framework) enabled consistent reasoning about belief change, including beliefs about other agents. This allowed us to examine some of the identified communication policies empirically. The obtained results confirmed that there are certain interesting invariants - in particular, a change in team coordination (and overall performance) was shown to be within the boundaries indicated by the relative information entropy. © Springer-Verlag Berlin Heidelberg 2003.