Study on Evaluation Function Design of Mahjong using Supervised Learning

抄録

The evaluation function for an imperfect information game is always hard to define but has a significant impact on the playing strength of a program. Deep learning has made great achievements in several recent years, and already exceeded the level of top human players in perfect information games such as AlphaGo. Predicting opponents moves and hidden states is important in imperfect information games. This paper describes a model on building a Mahjong artificial intelligence with deep learning method and supervised learning theory. Four deep neural network for discarding and predicting opponents' waiting, waiting tiles and point changes are combined into one model and performs good during games. With improved feature engineering, our accuracies on validation data of these networks reach higher than Dr. Mizukami and Professor Tsuruoka's network.