Research Of Exploration Policy For Multi-Agent Reinforcement Learning And Implementation Of Game Decision System

As a subfield of reinforcement learning,multi-agent reinforcement learning is closer to the real world so it attracts more and more people to research.However,compared to the single-agent cases,the logic of multi-agent system is more complex,which makes many problems become more difficult in multi-agent cases,such as the trade-off between exploration and exploitation.The trade-off between exploration and exploitation is an important problem in reinforcement learning,and many excellent works have been proposed for single-agent settings,but there are a few related works in the multi-agent domain,and most of the existing methods are subject to the mediocre performance and the poor scalability.Furthermore,although multi-agent reinforcement learning becomes more and more popular,there are a few algorithms libraries in this domain,the most popular library just provides a few algorithms and is complex to use.Besides,multi-agent reinforcement learning is still facing the problem of the difficulty in landing,the training of different tasks requires users to connect the algorithm with the tasks from the level of code,so it is impossible to directly train with the algorithm for the users without the knowledge of multi-agent reinforcement learning.This paper will firstly propose an exploration method named EMC in multiagent reinforcement learning,which bases on the curiosity and the episodic memory.EMC is composed of the curiosity module and the episodic memory module,the former encourages agents to make a deep and coordinated exploration,and the later is responsible for the efficient use of the collected samples,the combination of these two modules makes EMC achieve the state-of-the-art performance in experiments.Based on EMC and other algorithms,we propose an algorithms library named MAA and make a detailed introduction for its structure and running process.MAA provides several training algorithms,network structures and exploration methods in multi-agent reinforcement learning,and it is more concise and readable,the user can uses it jut by a simple command.Finally,based on MAA,we will introduce a game decision system and make a detailed analysis and design for it.This system is proposed to lower the threshold for ordinary users to use multi-agent reinforcement learning.For different tasks,users can directly train them with multi-agent reinforcement learning by our system,and our system will evaluate the models and explain the policies of agents to help users choose the best model and deeply understand the policies.