Computer Hex Algorithm Using a Move Evaluation Method Based on a Convolutional Neural Network

Abstract

In recent years, a move evaluation model using a convolutional neural network (CNN) has been proposed for Go, and it has been shown that CNN can learn professional human moves. Hex is a two-player connection game, which is included in the Computer Olympiad. It is important to consider cell adjacency on the board for a better Hex strategy. To evaluate cell adjacency from various perspectives properly, we propose a CNN model that evaluates all candidate moves by taking as input all sets consisting of 3 mutually adjacent cells. The proposed CNN model is tested against an existing CNN model called “NeuroHex,” and the comparison results show that our CNN model is superior to NeuroHex on a \(13\,\times \,13\) board even though our CNN model is trained on an \(11\,\times \,11\) board. We also use the proposed model as an ordering function and test it against the world-champion Hex program “MoHex 2.0” on an \(11\,\times \,11\) board. The results show that the proposed model can be used as a better ordering function than the ordering function created by minimax tree optimization, and we obtained a win rate of 49.0% against MoHex 2.0 (30 s/move).