Research On Chess Game Based On Deep Reinforcement Learning

Taking AlphaGo Zero and Alpha Zero as examples,deep reinforcement learning has made significant achievements in the board game,but the current related algorithms rely on the support of computing power.With the aim of reducing the dependence on computing power and improving the performance of algorithms,this paper mainly discusses how to improve the feedback mechanism and the neural network of deep reinforcement learning algorithms in computer chess games,and the impact of related improvements on network performance.In this paper,a hybrid deep reinforcement learning model is presented,which combines Q-Learning with Sarsa(lambda)updates and uses ResNet18 improvements.The UCT algorithm is combined in the game environment of Go and Jiu chess.Compared with Q-Leaming or Sarsa(lambda)algorithm alone,the learning model proposed in this paper achieves higher learning efficiency.In the experiment of Jiu chess game,the validity of the algorithm in Jiu chess is verified by comparing several parameters such as the rate of loss,the total number of games,the time of self-play and the important chess shapes built during the period of self-play.In the game experiment of Go,the algorithm proposed in this paper has been trained 60 times with the Q-Learning algorithm and the Sarsa(lambda)algorithm separately.Then the game is matched with the program based on the algorithm proposed in this paper and the program based on the Q-Learning algorithm and the program based on the Sarsa(lambda)algorithm.The result of the game shows that the algorithm proposed in this paper is also effective in the game of Go.A new neural network structure called maximum-average output layer is also proposed to replace several convolution layers in the middle of the CNN.Using the replaced network structure programming,a Go program based on deep Q learning is implemented,and ResNet18 improved program with the same number of layers is trained and played separately under the same framework of reinforcement learning model and game program.Results The chess program based on the new network structure defeated the reference program 7:3 in 10 matches,verifying the performance of the network structure with Maximum-Average Output Layer.Based on Microsoft's.Net Framework 4.7.2 framework and using Microsoft's Cognitive Toolkit deep learning library,this paper designs and implements a Go program based on deep reinforcement learning and a Jiu chess program based on deep reinforcement learning,respectively.