The research of pattern symmetry problem in learning in computer Go

Abstract

Pattern is a very effective approach to improve the Monte-Carlo tree search. Due to the limitation of affordable memory space, the two typical pattern sizes are 3*3 or 4*4. Pattern libraries may be constructed by hand-craft or machine learning, which are all suffered from pattern symmetry problem. This paper elaborates and classifies the pattern symmetry problem in 3*3-pattern and 4*4-pattern, and introduces the solution for solving it in learning procedure. The experimental results show that the solution is effective and the learning results are improved through solving pattern symmetry problem.