Scalability and parallelization of Monte-Carlo tree search

Abstract

Monte-Carlo Tree Search is now a well established algorithm, in games and beyond. We analyze its scalability, and in particular its limitations and the implications in terms of parallelization. We focus on our Go program MoGo and our Havannah program Shakti. We use multicore machines and message-passing machines. For both games and on both type of machines we achieve adequate efficiency for the parallel version. However, in spite of promising results in self-play there are situations for which increasing the time per move does not solve anything. Therefore parallelization is not a solution to all our problems. Nonetheless, for problems where the Monte-Carlo part is less biased than in the game of Go, parallelization should be quite efficient, even without shared memory. © 2011 Springer-Verlag.