Design of Amazon Chess Game System Based on Reinforcement Learning
Ding Meng, Jianbo Bo, Qi Yizhong, Yao Fu, Shuqin Li · 2019
Computer Chess game is an important research field of artificial intelligence, and amazon chess is a kind of chess with complicated strategies. In this paper the author designed an amazon chess game system contains the description of the situation, the method to generate and strategy - the value evaluation. Among them, the strategy - value assessment was used in the reinforcement learning technology, relying on the neural network output and move later distribution situation valuations to assist in Monte-Carlo search trees. Monte Carlo tree search achieves self-confrontation by continuously transforming roles and performing a series of simulations. The system iterates according to the results of self-matching and continuously improves the game strategy, thus improving the chess ability of the system. In order to improve the training efficiency of neural network, in this paper, the author designed a method for the division of chess stage according to the stage characteristics of Amazon chess. At different stages, the system is encouraged to use different strategies to explore to improve training efficiency. Finally, it is verified by experiments that the performance of this kind of reinforcement learning based Amazon chess game system can be self-improved, and the training efficiency is improved compared with the general Amazon learning game system based on reinforcement learning.