Opponent Behavior Prediction in a Multi-Player Game with Imperfect Information

Taiwanese Mahjong (Tai-Mah) is a popular multiplayer tile-based game in Taiwan that presents imperfect information. Predicting the behavior of opponents is an essential aspect of Tai-Mah, and superior ability in doing so distinguishes master players from ordinary players. This paper presents a rule- based Tai-Mah AI using an opponent behavior prediction agent to optimize the AI. The agent can predict the probability of the Mahjong tiles that an opponent needs based on heuristic Tai-Mah principles, which is then verified by game records. It was observed that the AI exhibited better performance than the average human player even before the use of the prediction agent, with the winning percentage being 25.455% over 21.638%. After the implementation of the prediction agent, the performance of AI significantly improved compared to the previous version, with the winning percentage increasing to 28.830% over 17.953%. The proposed AI is easy to adjust, can reduce players' waiting time (since a Mahjong game requires four players), and improve players' gaming experience. Therefore, it is suitable for commercial use, and our AI is currently being used in the industry.