Research And Application Of Incomplete Information Game Based On Deep Monte Card Method

Machine game is one of the important research directions in the field of artificial intelligence,which mainly involves the planning,decision and learning of the agent.By developing machine game algorithms,machines can be equipped with thinking and decision-making abilities similar to human beings,which can be applied in many practical scenarios,such as automatic driving,financial transactions and security protection.Because the incomplete information game cannot obtain complete information,the traditional search tree algorithm is difficult to traverse the deepest node.DQN algorithm is usually used to evaluate the rationality of the strategy itself,but it also brings great difficulties to feature extraction,combination and model selection.Monte Carlo,as a statistical simulation method,can transform high-dimensional problems into corresponding statistical simulation,and use statistical analysis of random samples to approximate the solution of the problem.It is suitable to solve such as mahjong state space and randomness,high information density of incomplete information game problems,but still need a large number of samples to ensure the accuracy of the results.Mahjong game itself has certain symmetry,using symmetry to improve the algorithm will help to improve the strength of the model and save computing resources,but the existing research rarely involves the optimization of symmetry.Based on the above questions,the work content of this paper is as follows:1.A Monte Carlo incomplete information game method based on parallel computing is designed,a virtual sub-game analysis method is proposed,and a state evaluation model is established to realize the intelligent decision of incomplete information game.This method uses the situation information in the virtual game environment to replace the evaluation of the decision itself,so as to integrate the models of different types of decisions such as eating,touching,barging and discarding cards into a comprehensive model.In addition,symmetry of game data is also used to compress and encode,so as to obtain better performance under the condition of using less training data,less storage space and lower training cost,and effectively improve training efficiency through parallel computing.2.Aiming at the research requirements of mahjong’s special rules and deep Monte Carlo method,a set of complete,efficient and expansible mahjong game environment is designed.The environment is compatible with different game agents,and realizes the recording and transmission of game data,at the same time adds the game game playback function,so as to analyze the game strategy of the agent.