Research And Application Of Incomplete Information Game Decision Based On Game Tree And Deep Learning

Incomplete information game problems exist widely in real life,such as commercial negotiation,advertising pricing,military strategy,financial transactions and so on,which belong to the field of incomplete information game,so it is of great significance to study the related fields.Incomplete information game has the characteristics of asymmetric information,and the complexity of the problem is higher than that of complete information game.It is difficult to solve this kind of problem through traditional game methods.How to obtain the optimal solution through limited known information is the focus of this paper.In this paper,taking four players mahjong as an example,a method of combining game tree with machine learning is proposed to solve the problem of incomplete information game.The main work and innovation of this paper are as follows1)Combined with the game rules of mahjong,the game process is abstracted into a game tree.Combined with the experience of experts,the expansion mode of the search tree is optimized,and the simulated expansion process of the search tree is transformed into the reverse derivation of the winning path according to the winning conditions,so that the complexity of the search is reduced from the exponential level to the polynomial level.In the design of the valuation function,the score and winning probability are integrated,and the winning expectation is used to express the advantages and disadvantages of the Hu card path,which makes the valuation more accurate and takes into account the Hu card tendency of high score and high winning rate.2)Monte Carlo simulation is used to establish the adversary model.The hidden information of adversary is inferred by combining the data distribution characteristics and the public known information in the current situation.Probabilistic random simulation is carried out according to the rounds and the current situation.The known information of adversary is used to verify the credibility,and the simulation results are adjusted by combining with historical data statistics to make it closer to the actual situation.The opponent model obtained by Monte Carlo simulation is used to improve the evaluation accuracy of winning expectation in the process of game tree expansion.3)Deep learning is highly dependent on the quality of samples.However,in incomplete information game,it is difficult to obtain high-quality data and evaluate the quality of data samples.Therefore,this paper uses the well-designed game tree method,inputs perfect information for decision-making,generates high-quality sample data through the self match of game tree,extracts features and uses deep residual network training.Compared with the general game tree method,this method makes better use of the hidden information;Compared with the deep learning method of learning human samples directly,it only needs fewer features to achieve better prediction results.4)Design and build a test platform.Through the simple configuration of the test platform,it can realize the mutual match between robots,and save the test data to the database;Completed the construction of the playback system,and through the web side to visualize the game data.Through the test platform,data statistics,decision analysis and game behavior analysis can be carried out.