摘要
计算机博弈是人工智能的果蝇和通用测试基准.近年来,序贯不完美信息博弈求解一直是计算机博弈研究领域的前沿课题.围绕计算机博弈中不完美信息博弈求解问题展开综述分析.首先,梳理计算机博弈领域标志性突破的里程碑事件,简要介绍4类新评估基准,归纳3种研究范式,提出序贯不完美信息博弈求解研究框架;然后,着重对序贯不完美信息博弈的博弈模型和解概念进行调研,从博弈构建、子博弈和元博弈、解概念以及评估3方面进行简要介绍;接着,围绕离线策略求解,系统梳理算法博弈论、优化理论和博弈学习3大类方法,围绕在线策略求解,系统梳理对手近似式学习、对手判别式适变和对手生成式搜索3大类方法;最后,从环境、智能体(对手)和策略求解3个角度分析面临的挑战,从博弈动力学和策略空间理论、多模态对抗博弈和序贯建模、通用策略学习和离线预训练、对手建模(剥削)和反剥削、临机组队和零样本协调5方面展望未来研究前沿课题.对于当前不完美信息博弈求解问题进行全面概述,期望能够为人工智能和博弈论领域相关研究带来启发.
Computer games are the drosophilae and universal benchmarks for artificial intelligence.In recent years,sequential imperfect information game solving has always been a frontier topic in the field of computer games research.Therefore,a comprehensive analysis of the imperfect information games solving problem in computer games is carried out.Firstly,it sorts out the milestones of the landmark breakthroughs in the field of computer games,briefly introduces four new evaluation benchmarks,summarizes three research paradigms,and deeply analyzes the challenges faced by games with imperfect information.Secondly,it focuses on investigating the game model and solution concept of a sequential imperfect information games,and briefly introduces it from three aspects:game formulation,sub-game and meta game,solution concept and evaluation.The offline strategy solving methods are systematically sorted from three perspectives of algorithmic game theory,optimization theory,and game theoretic learning.The online strategy solving methods are systematically sorted from three perspectives of opponent approximate learning,opponent discriminant adaptation,and opponent generative search.Finally,the challenges faced are analyzed from three perspectives:environment,agent(opponent)and strategy solving.The future research frontiers and prospects are given from five aspects:game dynamics and strategy space theory,multi-modal adversarial game and sequential modeling,general strategy learning and offline pretrain,opponent modeling(exploitation)and anti-exploitation,ad-hoc teamwork and zero-shot coordination.This paper provides a comprehensive overview of current imperfect information game solving,hoping to inspire related research in the field of aritificial intelligence and game theory.
作者
罗俊仁
张万鹏
苏炯铭
魏婷婷
陈璟
LUO Jun-ren;ZHANG Wan-peng;SU Jiong-ming;WEI Ting-ting;CHEN Jing(College of Intelligence Science and Technology,National University of Defense Technology,Changsha 410073,China)
出处
《控制与决策》
EI
CSCD
北大核心
2023年第10期2721-2748,共28页
Control and Decision
基金
国家自然科学基金项目(61806212)
湖南省研究生科研创新项目(CX20210011)。
关键词
计算机博弈
不完美信息博弈
扩展式博弈
反事实后悔最小化
在线凸优化
无悔学习
对手建模
computer games
imperfect information game
extensive form game
counterfactual regret minimization
online convex optimization
no-regret learning
opponent modeling