基于卷积神经网络的“拱猪”博弈算法  

Algorithm for“Hearts”game based on convolutional neural network

作　　者：吴立成[1] 吴启飞 钟宏鸣 王世尧 李霞丽[1] WU Licheng;WU Qifei;ZHONG Hongming;WANG Shiyao;LI Xiali(School of Information Engineering,Minzu University of China,Beijing 100081,China)

机构地区：[1]中央民族大学信息工程学院,北京100081

出　　处：《智能系统学报》2023年第4期775-782,共8页CAAI Transactions on Intelligent Systems

基　　金：国家自然科学基金项目(61773416,61873291)。

摘　　要：“拱猪”又称“华牌”,是一款极具特点的牌类游戏,属于非完备信息博弈,由亮牌和出牌2个阶段组成,整个游戏过程具有极强的反转性。为了研究“拱猪”计算机博弈算法,本文提出了一种基于深度学习的“拱猪”博弈算法,包含亮牌和出牌2个神经网络,分别用于亮牌和出牌阶段。亮牌和出牌网络均采用卷积神经网络(convolutional neural network,CNN)来构建,根据功能特点分别设计为不同的网络结构。采用11000局人类高级玩家的真实牌谱按比例生成训练数据和测试数据,对2个CNN网络进行了训练、测试和分析。结果表明,亮牌和出牌网络分别达到了88.4%和71.4%的准确率。对亮牌和出牌的一些具体例子进行的分析表明,本文算法能够产生合理的亮牌和出牌策略。“Hearts”,also known as“Chinese card game”,is a very characteristic poker game,which belongs to incomplete information games.It consists of two stages of card showdown and card playing,and there is strong reversality throughout the game.In order to study the computer game algorithm of“Hearts”,this paper proposes a“Hearts”game algorithm based on deep learning,which includes two neural networks,namely,card showdown and card playing,which are used in card showdown and card playing stage respectively.Both the card showdown network and card playing network are constructed by convolutional neural network(CNN),which are designed into different network structures according to their functional characteristics.Two CNN networks are trained,tested,and analyzed by using the real card playing patterns of 11,000 human advanced players to generate training data and test data proportionally.The results show that the accuracy of card showdown and card playing network reaches 88.4%and 71.4%respectively.The analysis of some specific examples of card showdown and card playing shows that the algorithm is able to produce reasonable card showdown and card playing strategies.

关 键 词：人工智能 非完备信息博弈 深度学习 卷积神经网络 拱猪 华牌 亮牌 出牌 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] G892[自动化与计算机技术—控制科学与工程]