An Empirical Evaluation of Applying Deep Reinforcement Learning to Taiwanese Mahjong Programs

Abstract

We introduce deep reinforcement learning (DRL) into a Taiwanese mahjong program, and compare the pros and cons of traditional tree search methods and DRL in terms of effectiveness and efficiency. Our DRL-based program aims to learn good strategies comparable to the state-of-the-art Taiwanese mahjong program Verylongcat, and has demonstrated effective learning capability in the experiments. Moreover, the required computation time of our DRL-based program is significantly lower than the Verylongcat version, bringing great advantage in time-limited tournaments.