Evolution of coordination in reactive multiagent systems.

摘要

A well coordinated multi-agent system takes advantage of the many benefits of distributed solutions. Although distributed solutions are well suited to many domains, coordinating the distributed components remains a research problem. Common approaches to designing a multi-agent system involve creating sophisticated agents and strictly defining the agent interactions. Yet, sophisticated agents demand more effort in design and often result in brittle solutions. This thesis questions that approach, opting for less capable agents implementing crude, yet effective, coordination schemes.; This work documents the design and analysis of some primitive agents and their environment. These agents have limited sensory-motor systems and a reactive control architecture. They do not have the ability to maintain an internal state or memory, nor do they implement any learning algorithms. Instead, the agent behaviors adapt over time via evolutionary algorithms. Numerous strategies are evolved which capitalize on the few strengths of the agents. For instance, an unexpected credit assignment architecture evolves. The credit assignment system allows the reactive agents to prioritize their responses, generating more reliable behavior. The agents also evolved various coordinated behaviors which utilized their undirected, broadcast signalling system. The coordination schemes differ between the types of signals emitted, the contexts that the signals are emitted in, and the responses elicited. In fact, multi-signal schemes evolved.; To identify critical traits of the environment, the agents, their tasks, and the strategies employed by the agents, many aspects of the artificial world are examined. Coordinated and uncoordinated strategies are tested for characteristics displaying sensitivity to the general hostility of the artificial world, the effectiveness of competitors, and the multi-agent population size. Once found, the critical traits provide a framework for further research towards creating other distributed applications implemented with reactive agents.