Applying Importance Sampling to MCTS for Mahjong

Abstract

Mahjong is a four-player stochastic imperfect-information game. In this article, we utilize importance sampling within Monte Carlo tree search (MCTS) to enhance the playing strength of our Mahjong program, MeowCaTS. First, we propose a tree structure called the merging solitary tile model, which facilitates the application of importance sampling. This model also reduces the branching factor of the search tree. Second, we apply importance sampling to MCTS and introduce the calculation of importance weights during the backpropagation stage. Finally, we design a multidepth transposition table to accumulate simulation results of similar positions in MCTS, further enhancing the strength of MeowCaTS. In the experiments, the performance of the proposed methods was analyzed, and the results showed a significant improvement. Notably, MeowCaTS won the first place in Computer Olympiad 2023.