Suphx: Mastering Mahjong with Deep Reinforcement Learning

• DOI: 10.48550/arxiv.2003.13590
• Publication Date: 2020-01-01
• Authors: Junjie Li; Sotetsu Koyamada; Qiwei Ye; Guoqing Liu; Chao Wang; Ruihan Yang; Zhao Li; Tao Qin; Tie‐Yan Liu; Hsiao-Wuen Hon
• Keywords: 0301 basic medicine; FOS: Computer and information sciences; 03 medical and health sciences; Artificial Intelligence (cs.AI); Computer Science - Artificial Intelligence

Artificial Intelligence (AI) has achieved great success in many domains, and game AI is widely regarded as its beachhead since the dawn of AI. In recent years, studies on have gradually evolved from relatively simple environments (e.g., perfect-information games such Go, chess, shogi or two-player imperfect-information heads-up Texas hold'em) to more complex ones multi-player hold'em StartCraft II). Mahjong a popular worldwide but very challenging for research due playing/scoring rules rich hidden information. We design an Mahjong, named Suphx, based deep reinforcement learning with some newly introduced techniques including global reward prediction, oracle guiding, run-time policy adaptation. Suphx demonstrated stronger performance than most top human players terms stable rank rated above 99.99% all officially ranked Tenhou platform. This first time that computer program outperforms Mahjong.