| In the development of artificial intelligence,computer games,as a branch field,have played the role of drosophilid fly and are highly sought after by academia and industry.Dou Dizhu(a.k.a.Fighting the Landlord)is an imperfect information game,which has a broad mass base in China and is loved by people.However,Dou Dizhu is easy to learn,difficult to master,has a lot of hidden information,a large game space,difficult decisionmaking,and difficult cooperation among peasants.All these difficulties have brought great challenges to the research of Dou Dizhu computer game.The thesis takes Dou Dizhu as the research object,and aims at the two major problems of bidding strategy and peasant cooperation,respectively proposes a bidding model combined with attention mechanism and a peasant role game algorithm based on HIEMCTS.In order to improve the search efficiency of the game tree,a new heuristic pruning method is further proposed and a self-play training framework is introduced in the training process.Finally,the development of the game system is completed.The specific work includes the following four parts:(1)Questions and Analysis.Dou Dizhu is a three-player game in which two peasants cooperate to fight against the landlord.The game includes two stages: bidding and playing.Different bidding will affect the overall result of the game.In addition,Dou Dizhu has a large action space,with a total of 27,472 possible combinations of card types.Every time playing,there may be hundreds of possible card type combinations,which brings challenges to players.(2)Establish a Dou Dizhu bidding model based on attention mechanism.Analyzed the difficulty of the bidding problem in the Dou Dizhu in detail,split the initial handcards of the game,completed the discretization process of the Dou Dizhu card type combination,extracted some features from the situation information as the input of the model,combined with the attention mechanism and Res Net,the bidding model was constructed.It can automatically learn and calculate the contribution of input features to bidding results,so as to make the optimal bidding strategy.(3)Proposed Hidden Information Estimation Monte Carlo Tree Search(HIEMCTS)algorithm.Aiming at the problem of the peasant role game in the game of Dou Dizhu,dismantling the player’s actions and order,discretizing the game process,constructing the game tree of the peasant role in the Dou Dizhu,combining the strategy model and the opponent model,establishing a new HIEMCTS algorithm for the peasant role game.Finally,with the idea of quantification,the game process is defined by the shanwin distance and the pruning of the game tree is realized,which improves the game efficiency.(4)Develop an intelligent agent for the Dou Dizhu game.Based on the Dou Dizhu competitive platform of JJ World Network Technology Co.,LTD,analyze and design the structure of agent of the Dou Dizhu,establish an interface with the platform,use python3.8as the development language and Visual Studio Code as the IDE,complete the coding and deployment of the agent,and through training and optimization of the game agent,won the championship in the Dou Dizhu computer game of 2022 IEEE Conference on Game,and won first prize in 2022 China Computer Game Championship. |
