Parallel Monte-Carlo tree search for HPC systems and its application to computer go

Monte-Carlo Tree Search (MCTS) is a class of simulation-based search algorithms. It brought about great success in the past few years regarding the evaluation of deterministic two-player games such as the Asian board game Go. A breakthrough was achieved in 2006, when Rémi Coulom placed 1st at the ICGA Computer Go Olympiad in Turin with his MCTS based Go programm CrazyStone in the 9× 9 devision. Until today, MCTS highly dominates over traditional methods such as αβ search in the field of Computer Go. Over the years, MCTS found applications in several search domains. A recent survey of MCTS methods lists almost 250 MCTS related publications originating only from the last seven years, which demonstrates the popularity and importance of MCTS. It is currently emerging as a powerful tree search algorithm yielding promising results in many search domains such as connection games Hex and Havannah, combinatorial games Breakthrough and Amazons as well as General Game Playing and real-time games. Apart from games, MCTS finds applications in combinatorial optimization, constraint satisfaction, scheduling problems, sample-based planning and procedural content generation. In this thesis, we present a parallelization of the most popular MCTS variant for large HPC compute clusters that efficiently shares a single game tree representation in a distributed memory environment and scales up to 128 compute nodes and 2048 cores. It is hereby one of the most powerful MCTS parallelizations to date. We empirically confirmed its performance with extensive experiments and showed our parallelization’s power in numerous competitions with solutions of other research teams around the world. In order to measure the impact of our parallelization on the search quality and remain comparable to the most advanced MCTS implementations to date, we implemented it in a state-of-the-art Go engine Gomorra, making it competitive with the strongest Go programs in the world.