Research And Application Of Reinforcement Learning In Multi-agent Collaboration

With the development of computer science,especially the extensive application of reinforcement learning,the application of reinforcement learning in multi-agent system become more and more popular.Reinforcement learning provides researchers with a way to solve multi-agent control problems by simulating the process of obtaining rewards in human activities and maximizing rewards.In the multi-agent system,researchers can use reinforcement learning to carry out adaptive learning and deal with dynamic environment.As multi-agent system has found an increasingly wide utilization in all fields,especially more and more collaborative scenarios of multi-agent system,using reinforcement learning to design flexible algorithms in multi-agent system is also a research hotspot.In order to achieve this goal,this thesis combines the existing reinforcement learning algorithm in multi-agent system,and studies the reinforcement learning method in multi-agent collaborative environment,the main work is as follows:(1)A Multi-Agent Reinforcement Learning Algorithm for global observation environment is proposed.The algorithm proposed in this thesis uses the attention mechanism to select the agent information in the environment adaptively and aggregate the selected agents' information with attention mechanism.And it can replace the joint state and joint action in the traditional multi-agent reinforcement learning.This algorithm is applied to multi-agent collaborative environment and compared with MADDPG algorithm.The result shows the efficiency of the algorithm in the simulation experiment.(2)A Multi-Agent Reinforcement Learning Method Based on graph network is proposed.In some observable multi-agent environments,graph networks are used to transfer information between agents,so that each agent in the multi-agent system can perceive the global information.This method is combined with reinforcement learning,and if is applied to the multi-agent collaborative environment,compared with the multi-agent algorithm with global observation.The result in simulation experiments show the effectiveness of the algorithm.(3)In multi-agent collaborative environment,this thesis uses the algorithm proposed in this thesis to apply to some environment where the number of agents is variable.The algorithm proposed in this thesis is combined with the idea of course learning to simplify the complex learning process,the experiments are carried out to verify the effectiveness in those environmentTherefore,this thesis designs the corresponding algorithm in the multi-agent collaborative environment through reinforcement learning,and applies it to the simulation experiment in the multi-agent collaborative environment.