Research On Streategy Of Imperfect Information Game Based On Real-time Heuristic Search

Machine game is an important research direction in the field of artificial inteligence.The imperfect information game is a subfield of the machine game.In imperfect information game,there are characteristics of information hiding and information asymmetry.Compared with the perfect information game,the imperfect information game is closer to real life.or example,information hiding and information asymmetry exist in real problems such as bidding,auction,and stock trading.Therefore,it is more practical to study imperfect information game.The Texas Hold'em game contains important features such as information hiding,information asymmetry and random events.It is a typical imperfect information game.The traditional method for solving the Texas Hold'em game is abstraction based solution method.This method combines the state space abstraction and the Counterfactual Regret Minimi-zation(CFR)to solve the poker game strategy offline,and cannot solve the game strategy in real time.This paper takes Texas Hold'em as the research object,combines the subgame strategy solving method and the heuristic game tree search method to study the real-time strategy solving algorithm of the two-player no-limit Texas Hold'em game.Aiming at the subgame strategy solving problem of poker game,the game decomposition problem and the solution method of subgame are studied.Drawing on the decomposition idea of the perfect information game,the game tree of the incomplete information game problem is decomposed into trunk and a series of game subgames,and the concept of the endgame in the complete information game is extended to the incomplete information game.Aiming at the subgame of imperfect information game,this paper studies the construction algorithm of the subgames,makes the subgames conform to the definition of extensive form game,and uses CFR and its improved algorithm to solve the local strategy of the game.Aiming at timeliness of poker game strategy solving problem,the heuristic game tree search algorithm and the subgame value estimation algorithm are studied.In order to improve the running efficiency of the algorithm based on game tree search,the subgames value estimation are used to limit the searching depth,and the action abstraction method is used to limit the searching breadth.At the same time,for the Texas Hold'em game,a set of solutions from the poker subgame training data generation,the poker subgame training data encoding to the poker game training data fitting is given.In order to verify the effectiveness of game strategy solving algorithm based on heuristic game tree search proposed in this paper,this paper designs and implements the two-player no-limit Texas Hold'em poker agent according to the rules of international computer poker competition,and analyzes the effectiveness of the algorithm through experiments.