The Research On Reinforcement Learning Based On Cooperative Multi-agent

In multi-agent system,the environment is dynamic and the behaviors of other agents are unknown,therefore the multi-agent system and each agent in the system should be ability to learn or self-adapt.As a machine learning that doesn’t need the environment model,reinforcement learning has been the hotpot in the multi-agent system.At the same time,because the resource and ability of single agent are limited,it needs cooperation of several agent to complete the task together.The main research of this thesis is as follows:This thesis firstly introduces the research foundation of agent and multi-agentsystem, then introduces briefly the essence knowledge of multi-agent learning method,multi-agent cooperation mechanism and reinforcement learning.An improved multi-agent cooperation learning method is proposed with blackboard model,fusion algorithm and reinforcement learning algorithm unified.Inthe method,the blackboard is a memory region that may realize information sharing;the fusion algorithm is used to fusion to the shared information,and reinforcement learning algorithm is used to select action with the fused result.Pursuit game problem is a multi-agent system and simultaneously has the cooperation and competition among multi-agents,so it is widely used to test the new learning algorithms in the artificial intelligence field.This thesis makes example analysis and emulation validation to the improved method through pursuit game problem,the experimental result shows that the method can efficiently improve the cooperation learning ability of agents in the multi-agent system.