Parallel Monte Carlo Tree Search Scalability Discussion

Abstract

In this paper we are discussing which factors affect the scalability of the parallel Monte Carlo Tree Search algorithm. We have run the algorithm on CPUs and GPUs in Reversi game and SameGame puzzle on the TSUBAME supercomputer. We are showing that the most likely cause of the scaling bottleneck is the problem size. Therefore we are showing that the MCTS is a weak-scaling algorithm. We are not focusing on the relative scaling when compared to a single-threaded MCTS, but rather on the absolute scaling of the parallel MCTS algorithm.