【作者】 吴松；

【导师】 王轩；

【作者基本信息】 哈尔滨工业大学 ， 计算机科学与技术， 2013， 硕士

【摘要】 随着人工智能技术的发展，通过有效的博弈树搜索技术已经能够处理大部分完备信息博弈问题，而非完备信息博弈在很长的一段时间内发展缓慢。人工智能技术的逐渐成熟和计算机硬件水平的提高为解决非完备信息博弈问题创造了条件。与完备信息博弈不同，在非完备信息博弈中参与者只能观察到部分博弈信息，与此同时解决非完备信息博弈还需要处理随机性、风险管理、对手建模、欺诈、信息不可靠等问题。非完备信息博弈的决策过程与现实社会的决策过程更加相似，研究非完备信息博弈问题对于解决现实博弈问题意义重大。对手建模是非完备信息博弈中最难解决的问题之一，也是一个智能的非完备信息博弈程序不可或缺的一部分。对手建模的作用是通过可观察到的玩家行为对博弈中不可知的信息进行预测，将对手建模对未知信息的预测与非完备信息博弈树搜索等技术相结合可以有效地对博弈中的各种行为进行评价，从而在博弈过程中做出对自己最有利的决策。本课题主要研究了在非完备信息博弈中进行对手建模的方法，通过不同的对手建模方法对玩家进行分类并预测博弈中的未知信息。课题以德州扑克作为具体研究对象，通过将对手建模方法与手牌评估算法相结合实现了一个拥有较高智能水平的扑克博弈程序并参加了2013年AAAI举办的计算机扑克程序大赛，在二人组比赛中获得第8名在三人组比赛中获得第4名的成绩，从而验证了对手建模方法在非完备信息博弈中应用的效果。本课题的另一个工作是对德州扑克中手牌评估算法进行了改进，提高了算法的执行速度。手牌评估算法是德州扑克博弈程序的重要组成部分，也是本文中的对手建模算法实现的基础。更多还原

【Abstract】 With the development of Artificial Intelligence, most games with perfectinformation can be solved with effective game tree search technology, but imperfectinformation games have been ignored for a long time. The emergence of newmethods in the area of Artificial Intelligence and the improvement of computerhardware make it possible to deal with imperfect information problems.Unlike players in perfect information games, participants in imperfectinformation games can only observe partial information of the whole game. At thesame time, we have to deal with stochastic results, risk management, opponentmodeling, deception, and unreliable information when facing imperfect information.The process to deal with imperfect information game is much like thedecision-making application in the real world, the research on imperfect informationgame will contribute to competition in the real world. Opponent modeling is one ofthe most challenging aspects in imperfect information games, it is also an importantcomponent for a good imperfect information game program. A good opponent modelwill describe how the opponent act in the game, and from the information we canpredict the unknown information and what the player will do in the future.Combining the results of opponent modeling with the technology of game treesearch for imperfect information game will help us make the most beneficialdecision in games.In this paper we will introduce how to model the opponent in imperfectinformation games, and how to use opponent modeling algorithms to classifyplayers and predict unknown information in games. We chosen Texas Holdem as anexperimental platform, by using opponent modeling algorithms and hand evaluationmethods we built a poker program which participated in the2013Annual ComputerPoker Competition held by AAAI, finally we took eighth place in heads-up limitTexas Holdem and took fourth place in3-player limit Texas Holdem, this validatedthe effectiveness of opponent modeling algorithms in imperfect information games.Another work in this paper is the improvement in hand evaluator for Texas Holdem,an effective hand evaluator is an important component of a strong Texas Holdemprogram, it is also the foundation of the opponent modeling algorithm for TexasHoldem.更多还原