Abstract: Monte Carlo tree search algorithm is a commonly used reinforcement learning algorithm, and the exponential growth of the dynamic space of the algorithm in the game process has become a factor that restricts the improvement of the algorithm learning efficiency. Based on the parallel approach to optimize the Monte Carlo tree search algorithm, a parallel Monte Carlo tree search algorithm based on the transfer of winning rate estimate is proposed. The improved parallel game search strategy framework consists of one main process and several sub-processes, in which the sub-processes are used for exploration, and the main process makes decisions according to the winning rate estimate data transmitted by the sub-processes. Combined with the multi-agent game platform Pommerman for experimental validation, the parallel Monte Carlo tree search algorithm can enhance the resource utilization rate, game-winning rate, and decision-making efficiency over the traditional Monte Carlo tree search algorithm.

Key words: multi-agent game, Pommerman, multi-process, parallel Monte Carlo tree search