Research On Incomplete Information Machine Game Algorithm And Opponent Model

With the development of computer technology,a series of outstanding achievements have been generated in the field of artificial intelligence.Especially in the complete information game,most of the complete information game problems can be solved by the classical methods,such as game tree search,dynamic programming,AlphaBeta pruning algorithm and so on.But different from the complete information game,the participants cannot get all the opponent information in the incomplete information game,and the uncertain factors in the game,such as the random risk,the adjustment of the opponent's strategy,the adversary's fraud and so on,have brought many difficult problems to the research work.In this paper,we take Texas Hold'em Poker as an object of research,and analyze the algorithm of incomplete information machine game.Firstly,based on the game characteristics of Texas Hold'em Poker,the state space complexity of the game is analyzed,and the classic machine game algorithm is introduced.It is concluded that the classic game algorithm is difficult to apply to the machine game of Texas Hold'em Poker.The modern mainstream research methods are classified as the algorithm based on Nash equilibrium strategy and the one based on opponent modeling method.The former mainly aim at approaching the Nash equilibrium,and the counterfactual regret minimization algorithm is currently one of the most popular algorithms.The latter aims at maximizing the profit in the game by exploiting the weakness of the opponent.Then,the counterfactual regret minimization algorithm is analyzed deeply,and the algorithm is modified by temporal difference learning to improve the efficiency of the algorithm.In order to solve the problem that the scale of state space is too large in the game of Texas Hold'em Poker,the abstraction of the card and the hand card evaluation algorithm are proposed to simplify the state space,and the machine game framework is established on the base of the improved algorithm,and the detailed process is introduced.At the same time,we deeply analyze the traditional opponent modeling method.The result shows that the traditional opponent modeling method is difficult to reflect the opponent's strategy characteristics and weaknesses.Therefore,this paper uses the new data features to improve the strategy biased opponent modeling method.The improved modeling method can not only reflect the opponent's weakness but also deduce the opponent's hand range.Based on the improved modeling method,an opponent decision model is established,and the corresponding strategy is optimized through expectation evaluation.Finally,a machine game experiment is designed to verify the proposed algorithm and the classical algorithm.The experimental results show that the improved counterfactual regret minimization algorithm is efficient,and the income value is increased by 25.8% in the face of the fixed opponent.Meanwhile,the improved strategy bias is 2.23 times that of the traditional method,which is the highest in all kinds of algorithms.In addition,the two improved algorithms proposed in this paper have good benefits in the one to one game with traditional algorithms.