Analyzing and Taming Collective Learning of a Multiagent System with Connected Replicator Dynamics

This paper analyzes complex collective behaviors of a multiagent system, which consists of interacting agents with evolutionary learning capabilities. The interaction and learning of the agents are modeled by the concept of Connected Replicator Dynamics expanded from evolutionary Game Theory. The dynamic learning system we analyze shows various behavioral and decision changes including bifurcation of chaos in the sense of physical sciences. The main contributions of the paper are summarized as follows: (1) In amultiagent system, the emergence of chaotic behaviors is general and essential, even if each agent does not have chaotic properties; and (2) However, a simple controlling agent with the Keep-It-Simple-Stupid (KISS) principle, or a sheep-dog agent, is able to domesticate or tame the complex behaviors.