Taiwanese Mahjong AI — Dartrix
Improving opponent modeling in a DRL-based Taiwanese Mahjong AI with heuristics and a neural discard model. Best Paper @ TCGA 2025; Silver @ ICGA Computer Olympiad 2024.
Dartrix is a deep-reinforcement-learning agent for Taiwanese Mahjong — an imperfect-information game with hidden state and high randomness, where modeling opponent behavior is central to strong play.
Abstract
This study tries to improve opponent behavior simulation in a DRL-based Taiwanese Mahjong AI. We introduce heuristic techniques and a neural network-based discard model to simulate opponent hands and improve the realism of discard behavior. The improved version of our Taiwanese Mahjong AI program, named Dartrix, participated in the 2024 TCGA and the 2024 ICGA Computer Olympiad, achieving particularly commendable results in the ICGA Computer Olympiad. The results underscore the practical benefits of enhanced simulation accuracy in improving win rates and optimizing gameplay strategies, laying a solid foundation for future applications of deep reinforcement learning in Taiwanese Mahjong AI.
Highlights
- Improves opponent behavior simulation with heuristic techniques and a neural-network-based discard model that infers opponents’ hands and makes discard behavior more realistic.
- Combines learning with Monte Carlo simulation to handle imperfect information and high randomness.
- The improved agent, Dartrix, competed at 2024 TCGA and the 2024 ICGA Computer Olympiad (Silver Medal), building on the earlier Rowlet program.
- More accurate simulation translated into higher win rates and better strategy — a foundation for future DRL work in Taiwanese Mahjong.
Keywords: Mahjong · Heuristic Methods · Monte Carlo Simulation · Deep Reinforcement Learning
Publication
Chen, H. Y., & Huang, K. C. (2025). Improved Opponent Modeling in DRL-Based Taiwanese Mahjong AI. Taiwan Computer Game Association (TCGA 2025), Taiwan, May 2025. — Best Paper Award