摘要
针对UCT算法的准确性受搜索次数影响较大的问题,提出一种结合神经网络的改进UCT算法。利用神经网络输出每一步的平均行动价值Q,结合改进的UCT算法寻找搜索过程中的高潜力节点。将传统UCT搜索改进为3个阶段:首先,通过已训练好的神经网络模型和UCT算法对当前所有子节点进行初次搜索,获得高潜力子节点;其次,利用剪枝操作去掉部分子节点,提升被搜索节点的质量;最后,二次搜索保留的高潜力子节点获得最优策略。另外,在分次搜索的过程中引入节点保留数量因子R和搜索比例因子P,辅助分次搜索,增加搜索的有效性。将其引入国际跳棋游戏中,实验结果表明:改进后的算法与其他算法相比胜率有所提升,验证了该算法的可行性。
Aiming at the problem that the accuracy of UCT algorithm is greatly affected by the number of searches,an improved UCT algorithm combined with neural network is proposed.The proposed algorithm uses the average action value Q of each step,which is output by the neural network,in combination with the improved UCT algorithm to find the high potential nodes in the search process.The traditional UCT search has been improved into three stages.Firstly,the trained neural network model and the UCT algorithm are used to conduct the initial search of all the current child nodes for obtaining the high-potential child nodes.Secondly,pruning is used to remove part of the child nodes and the quality of the searched nodes is improved.Finally,the optimal strategy is obtained by searching reserved high-potential child nodes again.In addition,in the process of sub-search,the number of nodes retained factor R and the search scaling factor P are introduced to assist sub-search and increase the effectiveness of search.By introducing the method into the checkers game,the experimental results show that the improved algorithm has a higher winning rate than other algorithms,which verifies the feasibility of the proposed algorithm.
作者
王亚杰
祁冰枝
张云博
丁傲冬
WANG Yajie;QI Bingzhi;ZHANG Yunbo;DING Aodong(Engineering Training Center,Shenyang Aerospace University,Shenyang 110135,China)
出处
《重庆理工大学学报（自然科学）》
CAS
北大核心
2021年第7期259-265,共7页
Journal of Chongqing University of Technology：Natural Science
基金
辽宁省兴辽英才计划项目(XLYC1906003)。
作者简介
王亚杰,女,博士,教授,主要从事模式识别、图像融合、机器博弈研究,E-mail:wangyajie@sina.com;通讯作者:祁冰枝,女,硕士研究生,主要从事机器博弈研究,E-mail:qbz1691@163.com。
