Research On Multi-Player Imperfect Information Computer Game Based On Residual Network And SDMCTS

Machine gaming is one of the research and challenging fields in the field of artificial intelligence,and the academic community has taken this field as an important research direction.According to whether the object of the game can completely obtain the current situation to classify,the machine game can be non-complete information machine game and incomplete information machine game.Among them,the complete information machine game has relatively mature game algorithms for analysis and prediction,while the noncomplete information machine game has the characteristics that the game information cannot be fully understood by the game object due to its existence,and the game situation is complex and changeable,which is closer to real life.Game scenarios such as energy scheduling,financial forecasting,chess and card games,etc.Therefore,the research on incomplete information machine games has more commercial and practical application value.In recent years,with the development of deep learning,researchers have applied deep learning networks to incomplete information machine games,and have achieved certain results.But the game agent based on deep learning deeply depends on the quality of the game data,the prediction accuracy of the network and the feature extraction ability of the game data.The Monte Carlo tree search algorithm is another decision strategy applied in the field of incomplete information machine games,but the algorithm is based on permutation and combination sampling and has a larger search space.In the vertical simulation stage of the game tree,the game object is randomly selected.The strategy of the game has greater randomness,and the agent based on this algorithm needs to take a lot of time to search in order to obtain the optimal decision.In response to the above problems,this paper studies and implements a multi-player incomplete information machine game algorithm based on residual network and semidefinite Monte Carlo tree search algorithm.This algorithm can simultaneously exert the immediacy and immediacy of optimal strategy search based on residual network.The effectiveness of the semi-definite Monte Carlo tree search algorithm.This algorithm improves the Monte Carlo sampling algorithm,and changes the random Monte Carlo sampling algorithm to the Monte Carlo sampling algorithm based on the residual neural network.The sampling algorithm can effectively reduce the randomness of the situation decomposition.It is Monte Carlo The tree node expansion phase provides prerequisites.In the Monte Carlo tree node expansion stage,this article aims at the problem of game randomness in the expansion of sub-nodes using random strategy games between the two sides of the game.The opponent's behavior is modeled,and the opponent's decision-making behavior is added to the Monte Carlo tree node expansion stage.In this method,this method can effectively reduce the randomness of the depth expansion of the game tree node,and reduce the time for the game tree to expand the simulation.In addition,in order to further improve the search efficiency of the game tree,this paper adds the improved Monte Carlo tree search process to the UCT algorithm.In the process of tree node expansion,the balance point of exploration and development is sought to improve the efficiency and accuracy of the vertical expansion of Monte Carlo tree nodes.In the training phase of the residual network model,according to the two-on-one poker game rules,different roles have a cooperative and competitive game relationship.This paper uses different data sets to separate the strategy network,card network and card guessing network of different roles.Carry out training,and finally verify the game ability of the agent based on the residual network by means of a rule-based game.In this paper,the residual network and the semi-definite Monte Carlo tree search algorithm are combined,and the algorithm is applied to the two-on-one poker game,and the agent based on the algorithm is realized.In order to verify the game learning ability of the residual network,it will be based on the residual The agents implemented by the difference network and semi-definite Monte Carlo tree search algorithms respectively play games with the agents based on the rule system.The experimental results show that the agents based on the algorithm have a high level of gameplay,which verifies the effectiveness of the algorithm...