一种结合策略价值网络的五子棋自博弈方法研究  

Research on a self-play method of Gobang combined with a strategic value network

作　　者：刘溜 张小川 彭丽蓉 田震[4] 万家强 任越 LIU Liu;ZHANG Xiaochuan;PENG Lirong;TIAN Zhen;WAN Jiaqiang;REN Yue(School of Artificial Intelligence,Chongqing University of Technology,Chongqing 401135,China;Institute of Artificial Intelligence System,Chongqing University of Technology,Chongqing 400054,China;School of Artificial Intelligence and Big Data,Chongqing Industry Polytechnic College,Chongqing 401120,China;Chongqing Nankai Liangjiang Middle School,Chongqing 401135,China)

机构地区：[1]重庆理工大学两江人工智能学院,重庆401135 [2]重庆理工大学人工智能系统研究所,重庆400054 [3]重庆工业职业技术学院人工智能与大数据学院,重庆401120 [4]重庆市南开两江中学校,重庆401135

出　　处：《重庆理工大学学报（自然科学）》2022年第12期129-135,共7页Journal of Chongqing University of Technology：Natural Science

基　　金：国家自然科学基金项目(60443004)。

摘　　要：针对传统蒙特卡洛树搜索算法存在“难以在节点的探索和利用之间做出平衡;难以聚焦重要搜索分支”等问题,提出使用策略价值网络完成棋局评估与落子着法生成,将策略价值网络与蒙特卡洛树搜索相结合。策略价值网络指导搜索树的展开,搜索结果用以持续更新网络参数,形成一种自博弈方法,在多轮自博弈中实现算法的迭代优化。实验表明:相较于各种经典搜索算法,所提算法在平均落子时间上降低了约95%,平均对局胜率达到80%以上。In view of the problems in the traditional Monte Carlo tree search algorithm,such as“difficulty in balancing the exploration and utilization of nodes,difficulty in focusing on important search branches and so on”,this paper proposes a strategic value network combined with Monte Carlo to complete chess evaluation and generation of moves in the game.The application of Monte Carlo is guided by the strategic value network,and the search results are used to continuously update network parameters,so a self-play method is formed to realize iterative optimization of the algorithm in multiple rounds of self-play games.In accordance with the final experiment,in contrast with various classic search algorithms,this method reduces the average dropping time of chess pieces by about 95%,and the average game winning rate reaches more than 80%.

关 键 词：蒙特卡洛树搜索 深度神经网络 五子棋计算机博弈 自博弈 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]