基于强化学习的爱因斯坦棋的算法研究  

An algorithm Research of Einstein Chess Based on the Reinforcement Learning

作　　者：向宇涛 朱道易 王忠桃[1] 董羽 XIANG Yu-tao;ZHU Dao-yi;WANG Zhong-tao;DONG Yu(Chengdu University of Technology,Chengdu 610000,China)

机构地区：[1]成都理工大学,四川成都610000

出　　处：《电脑知识与技术》2020年第22期179-181,共3页Computer Knowledge and Technology

基　　金：四川省成都理工大学2019年大学生创新创业训练项目,省级项目,项目编号S201910616131,所属编码0809.(2017YFC0602101)。

摘　　要：爱因斯坦棋是我国全国大学生指定计算机博弈比赛项目之一,在本文发表之前,多数队伍使用了搜索树算法来参加比赛,而为了提高算法的表现,多数算法需要加入具体的模板处理。然而,如果需要得到最优表现,则要加入大量模板并尽可能地搜索更深的博弈树。而使用深度强化学习并结合使用树搜索不仅不需要加入过多的模板处理、搜索深层树节点,还能随着接收不同的算法和样本,智能体可以逐步地改善表现,并进一步提高胜率。因此,本文通过使用了深度强化学习中的DQN算法,并结合了极大极小树搜索算法和随机生成行为法测试了该算法的胜率,最后取得了远优于这两种方法的结果。The Einstein chess is one of the designated National computer games for Chinese Collage students.Before the article’s published,most teams used the tree-search algorithm to participate in the competition.In order to improve the performance of their algorithms,the specific templates are always needed for the algorithms.However,if they need to get the best performance,they need the more templates and deeper tree nodes as much as possible.Using the deep reinforcement learning combined the treesearch algorithm does not need more templates and deeper nodes,and with receiving different algorithms and samples the agent could improve its performance and win rate.Therefore,this paper uses the DQN algorithm and combines the minimax tree-search algorithm and random generating method to test the win rate of this algorithm.Finally,the results are much better than other two al⁃gorithm.

关 键 词：强化学习 爱因斯坦棋 人工智能 计算机博弈 算法 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]