Research On Texas Hold'em Computer Game Intelligent Decision Model

Computer games is one of the most challenging research directions of artificial intelligence.It is also called machine game which can be divided into two types: complete information computer game and incomplete information computer game.Among them,complete information computer game refers to the game in which all the players are completely transparent in the game state and process,but incomplete information Computer game refers to the game in which all players are in the state or process of the game with opaque or incomplete information.The paper's Texas Hold'em game is an example of incomplete information computer game.The paper takes it as the research object to discuss the application of artificial intelligence technology in incomplete information game.Machine learning is a research hot spot in the field of artificial intelligence.The Go game program represented by Google Alpha-Go has shown a strong game ability.Go is just a chess game with complete information for two players,and Texas Hold'em belongs to two or more people.Therefore,how to combine reinforcement learning and neural network with Go game method and apply them to the Texas Hold'em with incomplete information which is an urgent problem in the field of computer game.In particular,reinforcement learning faces problems such as incomplete information,high-dimensional state space,overestimation problems,difficulty in finding optimal solutions quickly,and difficulty in convergence in the game of Texas Hold'em,which greatly affects the combat power of the game program.Aiming at the above-mentioned problems,the thesis adopts a combination of artificial neural network and reinforcement learning to improve reinforcement learning algorithm,convolutional neural network,develop Texas Hold'em decision model,and verify that decision model achieves the expected results.The specific research content of this paper are as follows:(1)According to the network structure of the decision model,a poker data representation method of the Texas Hold'em decision model is proposed to make the data better convolution.(2)Based on zero-sum game theory,design the reward function of the decision model based on the UCT algorithm.According to the difference between the actual revenue generated by the game and the expected revenue output by the UCT algorithm,the rewards and punishments of the decision model are determined.When the actual return is greater than the expected return,the decision model will be rewarded,When the actual return is less than the expected return,the decision model will be punished,in order to update the decision model,to update the decision model.(3)In order to solve the over fitting problem of convolution neural network,dropout function is used to optimize the neural network and increase the randomness and sparsity of neural network connection.Improve the activation function of the convolutional neural network,combine the LReLU function with the Softplus function to construct the L-S function,and improve the convergence of the convolutional neural network.(4)According to the advantage learning,the evaluation function of DQN algorithm is improved by introducing correction function method,and the action selection strategy of the DQN algorithm is optimized.Then,by fusing the update target of SARSA algorithm,it dynamically combines the advantages of DQN algorithm and SARSA algorithm.Finally,the DQN-S algorithm is proposed to improve the learning efficiency of the algorithm.(5)Applying the above research results to the Texas Hold'em decision model,using Python language and Tensorflow framework,the Texas Hold'em game system is realized.Finally,an experimental verification was conducted.Compared with the improved Texas Hold'em decision-making model,the improved Texas Hold'em decision-making model won more chips and the decision-making ability of the decision model was significantly improved.The Texas Hold'em system implemented in this article participated in the National College Student Computer Game Competition held in Beijing in 2019,and won the first prize,which verified the effectiveness of the system.