A Method for Multi - Agent Coordination Based on Distributed Reinforcement Learning

Abstract

The emphasis of research on multi - agent system is that the individual agents apply their negotiation, coordination, and cooperation to accomplish the complicated task or resolve the complex problem. With analysis and research on distributed reinforcement learning, a method for multi - agent cooperation is proposed. Coordination level decomposes the complicated task and the central reinforcement learning is used to assign the subtask by coordination agent. In behavioral level, the task agents receive the sub - tasks and adopt the individual reinforcement to choose the effective action and accomplish global task cooperatively. With the application and experiment in Robot Soccer simulation game, this method shows better performance than that of the conventional reinforcement learning.