Distributed Monte Carlo Tree Search: A Novel Technique and its Application to Computer Go

Abstract

Monte Carlo tree search (MCTS) has brought about great success regarding the evaluation of stochastic and deterministic games in recent years. We present and empirically analyze a data-driven parallelization approach for MCTS targeting large HPC clusters with Infiniband interconnect. Our implementation is based on OpenMPI and makes extensive use of its RDMA based asynchronous tiny message communication capabilities for effectively overlapping communication and computation. We integrate our parallel MCTS approach termed UCT-Treesplit in our state-of-the-art Go engine Gomorra and measure its strengths and limitations in a real-world setting. Our extensive experiments show that we can scale up to 128 compute nodes and 2048 cores in self-play experiments and, furthermore, give promising directions for additional improvement. The generality of our parallelization approach advocates its use to significantly improve the search quality of a huge number of current MCTS applications.