A multi-agent architecture for distributed constrained optimization and control

This article is meant as a contribution to the debate about the meaning and the prediction of emergent behaviour. By means of a number of examples, we show that reactive multi-agent systems have much in common with the statistical mechanical systems studied by physicists. The PHAMUS (PHysics-based Approaches to MUltiple computing Systems) model of an agent is an entity which possesses a number of internal states about which it can communicate with other agents. PHAMUS agents pass through a sequence of time frames, at each one being able to update their internal states as a result of interaction with other agents or with the external environment. By showing that societies of such agents are generalizations of statistical mechanical systems, we demonstrate the extreme difficulty of being able to reason about and to prove the collective behaviour of multi-agent systems from the only knowledge of the agents and their interactions. Simulation must be the main tool for assessing whether a particular agent design will lead to a multi-agent system capable of solving a particular application efficiently. This rather negative result is balanced by our promotion in this paper of a revisited methodology far studying reactive multi-agent systems that is close to the one used by physicists in statistical mechanics. Lu particular, and given our previous experience, this analogy suggests to us that the problem of simulating the behaviour of large collections of agents can be efficiently executed on multi-processor computing systems.