Reinforcement Learning for Creating Evaluation Function Using Convolutional Neural Network in Hex
Kei Takada, Hiroyuki Iizuka, Masahito Yamamoto · 2017
An evaluation function in the board game decides the next move for computer AIs, and a high accurate evaluation function leads to a strong computer AI. Recently, the evaluation function using convolutional neural network(CNN) by supervised learning shows high evaluation accuracy. Supervised learning cannot exceed the teachers, but there must be a possibility to create a more accurate evaluation function by using reinforcement learning. In this paper, we proposed an evaluation function using CNN and reinforcement learning with games of self-play in Hex. The proposed evaluation function is tested with the previous evaluation function and world-champion program MoHex2.0. The results show that evaluation accuracy of the proposed evaluation function is higher than the previous evaluation functions, and proposed computer Hex algorithm EZO-CNN obtained a win rate of 60.0% against MoHex2.0 even though the search time of EZO-CNN is shorter than MoHex2.0.