Research And Implementation Of Computer Game Algorithm For "Pig Arching

Computer games can be divided into complete information game and incomplete information game according to whether the information observed subjectively is complete or not.Compared with complete information game,the research of incomplete information game is more difficult and closer to life.Therefore,the research of incomplete information game has great practical significance."Gongzhu",also known as "Huapai",is a very characteristic card game,which belongs to incomplete information game.At present,there is no relevant literature on the research of "Gongzhu" game.Although the Texas Hold’em AI has been able to defeat the human professional players,and the "Doudizhu" AI is gradually approaching the level of human experts,but they have high requirements for computing power.It is impossible to achieve without enough powerful hardware resources,and the game algorithm used cannot be directly applied to the "Gongzhu".In this thesis,the game algorithm and platform design of "Gongzhu" are studied.The main research work completed includes the following parts:1.This thesis proposes a game algorithm of "Gongzhu" game based on Convolutional Neural Networks(CNN).According to the characteristics of the"Gongzhu" game,the CNN structure of the card-showing stage and the card-playing stage are respectively designed,and the existing 11,000 pieces of real player data are used for supervised learning,and the results of card-showing and card-playing predicted by "Gongzhu" AI are fitted with the real results of card-showing and cardplaying of human players.The accuracy of the card-showing on the test set is 88.4%,and the accuracy of the card-playing on the test set is 71.4%.Through the analysis of concrete examples,it can be seen that the "Gongzhu" AI based on CNN has certain card-showing and card-playing strategies.2.This thesis proposes a game algorithm of "Gongzhu" game based on Deep Monte-Carlo(DMC).The DMC adopts the method of self-play to simulate and evaluate,and uses the depth Q-network to replace the Q-table to update the Q-value,and efficiently explores and utilizes the "Gongzhu" strategy.The method of distributed parallel computing can effectively improve the training efficiency.Compared with the traditional Monte-Carlo method,it can effectively solve the problem of high variance.The constructed agent AI has played 10,000 simulated games with the CNN-based "Gongzhu" AI,and can achieve a winning rate of 78.3%.Through the multiple analysis of the experimental results,it can be seen that the"Gongzhu" AI based on DMC has stronger performance.3.Design and implement a game platform for the "Gongzhu".In view of the problem that there is currently no platform available for "Gongzhu" AI simulation and testing,it is developed based on Java language,using Model-View-Controller,(MVC)model to build the overall architecture of the system,designing three functions of game playback,human-machine game and intelligent AI game,and verifying the strong robustness and stability of the system by designing special test cases and simulated games.