Monte Calro Tree Search Parallelization by Specialized Process Ratio Control Depending on Game Progress

Abstract

Monte Carlo Tree Search(MCTS) is one of the key techniques that developed the recent game AI remarkably, and its parallelization is important to find the best action in a limited time. One popular parallelization is to assign each node in a game tree to one of processors, and various tasks for the node are processed by the assigned processor. In that parallelization, load balancing over processors is difficult due to the large computational-time differences among MCTS component tasks; tasks are piled up in the processors to which heavy tasks are assigned, and instead some processors become idle. In this paper, we propose a method to increase parallel efficiency of MP-MCTS, which is one of the parallel extensions of MCTS. In our method, we divide processors for simulations, which are heavy tasks, from processors for other tasks, and change the ratio of simulation-processors depending on game progress. We implement the proposed MP-MCTS, and let it play against the original MP-MCTS in Othello game. The proposed MP-MCTS outperforms the original MP-MCTS by increasing its parallelization effect.