| Game problems are ubiquitous in real life,and wherever there are disputes over interests,there exist games.Incomplete information games are a typical representative of game problems,such as board and card games,item auctions,and commercial pricing,all of which belong to incomplete information game problems.Therefore,the research on game-related fields has great practical significance and social value.Different from complete information games,incomplete information games have the characteristics of information asymmetry and information non-disclosure,which leads to an exponential increase in the complexity of the game.It is difficult to directly apply traditional game methods to incomplete information games.Making the optimal decision in the case of incomplete scene information acquisition is the research focus of this paper.In this paper,Shangrao mahjong is taken as an example,and the semi-definite Monte Carlo tree search algorithm is combined with the residual network and the model-simplified heuristic tree search algorithm to study the incomplete information game problem.The main work and innovations of this paper are as follows:（1）Combining game knowledge and historical game data,the Monte Carlo simulation method is used to construct an opponent model to predict the distribution of opponent cards and convert it into relative probability.The huge hidden information in the incomplete information game may lead to large deviations in decision-making.In this paper,the Monte Carlo method is used to simulate the opponent’s hand,and the simulation results are adjusted by combining the knowledge of the game field,the current situation and the statistical results of the game data.It is more in line with the actual scene.The opponent model is used to provide important help for subsequent search path-related weight calculations and information completion.（2）According to the Mahjong game rules,a simplified heuristic information tree search algorithm is designed to explore the winning path.First,the residual network is used to learn the master data to obtain the card selection model,and the search space is reduced by searching for the best card.Then,based on the calculation of heuristic combination information of the optimal winning hand,the search tree is expanded using a draw strategy based on the combination information.Essentially,this process updates the hand cards to the winning state and sets the last useless card as the discarded card for the exploration path.（3）A mahjong decision-making program based on the improved semi-definite Monte Carlo tree search algorithm（SDMCTS）is realized.Due to the existence of unknown information in the field of incomplete information games,the effect of Monte Carlo tree search algorithm（MCTS）is not ideal.This paper proposes a decision-making procedure based on the improved SDMCTS algorithm.First,in the game tree construction stage,this paper uses the opponent model to predict the hidden information of the game,and transforms the incomplete information situation into a complete information situation.After that,in the node expansion stage of the game tree,this paper uses the game agent with decision-making ability to guide the expansion of nodes,and uses the game data of human master players to obtain the strategy networkof mahjong decision-making by using deep residual network training,and usesto guide The vertical expansion of the game tree improves the efficiency of the search.Finally,in the simulation stage,the simplified heuristic information tree search modelis used to quickly simulate to the end of the game and calculate the expected value of the current game situation.The experimental results show that the mahjong decision-making program implemented in this paper is higher than other decision-making programs in terms of winning percentage,average points per game and average points per game. |
