Scalability and Parallelization of Monte-Carlo Tree Search

International audienceMonte-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, in particular for our program MoGo but also for our Havannah program Shakti. In particular, we get a good efficiency for the parallel versions, both for multicore machines and for message-passing machines, but in spite of promising results in self-play there are situations for which increasing the time per move does not solve anything, and therefore parallelization is not the solution either. Nonetheless, for problems on which the Monte-Carlo part is less biased than in Go, parallelization should be very efficient even without shared memory