Deep Learning for Evaluation of Game and Puzzle Positions (Algebraic system, Logic, Language and Related Areas in Computer Sciences II)

Abstract

In recent years, it has been demonstrated that a powerful game AI can be implemented in Go with the combination of well-learned neural networks and traditional search algorithms. In this study, we made an equivalent to the policy network in the Go AI for, FreeCell, a solitaire card game, by deep learning. We also made an equivalent to the value network in the Go AI for Pentago, a two-player board game. The network structure, ResNet, was adopted for the above cases of deep learning, and supervised learning was performed with the evaluation values of moves in FreeCell and positions in Pentago as teacher data. It is well-known that ResNet is useful for deep learning with deep convolutional neural networks.