A Game Model for Gomoku Based on Deep Learning and Monte Carlo Tree Search

摘要

Alpha Zero has made remarkable achievements in Go, Chess and Japanese Chess without human knowledge. Generally, the hardware resources have much influence on the effect of model training significantly. It is important to study game model that do not rely excessively on high-performance computing capabilities. In view of this, by referring to the methods used in AlphaGo Zero, this paper studies the model applying deep learning (DL) and monte carlo tree search (MCTS) with a simple deep neural network (DNN) structure on the Game of Gomoku Model, without considering human expert knowledge. Additionally, an improved method to accelerate MCTS search is proposed on the base of the characteristics of Gomoku. Experiments show that this model can improve the chess power in a short training time with limited hardware resources.