The application of improved UCT combined with neural network in Tibetan JIU chess

Abstract

Tibetan JIU chess is a kind of ethnic game, which is mainly popular in Tibetan areas of China. Because of the large state space, it is impossible to use the ready-made Upper Confidence Bound Apply to Tree (UCT) game model. Given the above problem, an algorithm combining neural network and UCT is proposed. First, the strategy value function is designed to assist the simulation of UCT to prevent the inaccuracy of random simulation results. Second, the multi-process method is introduced to improve the search efficiency of the game model. Third, we simplified the structure of the neural network and improved the two-dimensional Gaussian distribution matrix to facilitate the training of the network. The experimental results demonstrate that the win rate of this reinforcement learning model has been improved, which verifies the feasibility of the proposed method.