Developing Computer Hex Using Global and Local Evaluation Based on Board Network Characteristics

Abstract

The game of Hex was invented in the 1940s, and many studies have proposed ideas that led to the development of a computer Hex. One of the main approaches developing computer Hex is using an evaluation function of the electric circuit model. However, such a function evaluates the board states only from one perspective. Consequently, it is recently defeated by the Monte Carlo Tree Search approaches. In this paper, we therefore propose a novel evaluation function that uses network characteristics to capture features of the board states from two perspectives. Our proposed evaluation function separately evaluates the board network and the shortest path network using betweenness centrality, and combines the results of these evaluations. Furthermore, our proposed method involves changing the ratio between global and local evaluations through a support vector machine (SVM). So, it yields an improved strategy for Hex. Our method is called Ezo. It was tested against the world-champion Hex program MoHex. The results showed that our method was superior to the 2011 version of MoHex on an $11 \times 11$ board.