基于多智能体强化学习的博弈综述  

Multi-agent Reinforcement Learning Based Game:A Survey

作　　者：李艺春 刘泽娇 洪艺天 王继超 王健瑞 李毅[1,2] 唐漾[1,2] LI Yi-Chun;LIU Ze-Jiao;HONG Yi-Tian;WANG Ji-Chao;WANG Jian-Rui;LI Yi;TANG Yang(Key Laboratory of Smart Manufacturing in Energy Chemical Process,Ministry of Education,East China University of Science and Technology,Shanghai 200237;School of Information Science and Engineering,East China University of Science and Technology,Shanghai 200237;School of Mathematics,East China University of Science and Technology,Shanghai 200237)

机构地区：[1]华东理工大学能源化工过程智能制造教育部重点实验室,上海200237 [2]华东理工大学信息科学与工程学院,上海200237 [3]华东理工大学数学学院,上海200237

出　　处：《自动化学报》2025年第3期540-558,共19页Acta Automatica Sinica

基　　金：国家自然科学基金(62233005,U2441245);中国博士后科学基金(2024M750904)资助。

摘　　要：多智能体强化学习(Multi-agent reinforcement learning,MARL)作为博弈论、控制论和多智能体学习的交叉研究领域,是多智能体系统(Multi-agent systems,MASs)研究中的前沿方向,赋予智能体在动态多维的复杂环境中通过交互和决策完成多样化任务的能力.多智能体强化学习正在向应用对象开放化、应用问题具身化、应用场景复杂化的方向发展,并逐渐成为解决现实世界中博弈决策问题的最有效工具.本文对基于多智能体强化学习的博弈进行系统性综述.首先,介绍多智能体强化学习的基本理论,梳理多智能体强化学习算法与基线测试环境的发展进程.其次,针对合作、对抗以及混合三种多智能体强化学习任务,从提高智能体合作效率、提升智能体对抗能力的维度来介绍多智能体强化学习的最新进展,并结合实际应用探讨混合博弈的前沿研究方向.最后,对多智能体强化学习的应用前景和发展趋势进行总结与展望.Multi-agent reinforcement learning(MARL),which stands at the intersection of game theory,cybernetics and multi-agent learning,represents the cutting-edge domain within the realm of multi-agent systems(MASs)research.MARL empowers the agents with the capability to complete a variety of complex tasks through interaction and decision-making in the dynamic multi-dimensional and complicated practical environment.When progressing towards the openness of application objects,the embodiment of application issues and the complication of application contexts,MARL is gradually becoming the most effective tool for solving game and decision-making problems in the real world.This paper systematically reviews the game based on MARL.First,the basic theory of MARL is introduced,and the development process of MARL algorithms and the baseline testing environment have been introduced and summarized.Then,we focus on three types of tasks within MARL,which are cooperation,competition and mixed tasks.The latest progress in MARL is introduced by concentrating on improving the cooperative efficiency and enhancing the adversarial abilities of agents,and the most recent researches on mixed games,in combination with their practical applications,are investigated.Finally,the prospects of application and the trends of development for MARL are summarized and prospected.

关 键 词：多智能体强化学习 多智能体系统 博弈决策 均衡求解 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]