Behavior imitation of individual board game players

Chao-Fan Pan, Xue-Yang Min, Heng-Ru Zhang, Guojie Song, Fan Min

Abstract: Modeling and predicting player behavior is of the utmost importance in game development and matchmaking. A variety of methods have been proposed to build artificial intelligence (AI), human-like players. However, these human-like players have a limited ability to imitate the behavior of individual players. In this paper, we propose a player behavior imitation method using imitation learning under the framework of meta-learning. A generic behavior model of game players was learned from historical records using adversarial imitation learning. Then, we personalized the policy by imitating the behavior of each individual player. Convolutional neural networks were used to construct the feature extractor of game board states. The experiments were conducted using the Reversi game, and 18,000 game records of different players were used to train the generic behavior model. The behavior of each new player was learned using only hundreds of records. The results demonstrate that our method can be utilized to imitate individual behavior in terms of action similarity well.