CUDA Implementation of Computer Go Game Tree Search

Abstract

Go is a fascinating game that has yet to be played well by a computer program due to its large board size and exponential time complexity. This paper presents a GPU implementation of PV-Split, a parallel implementation of a widely used game tree search algorithm for two-player zero-sum games. With many game trees, it often takes too much time to traverse the entire tree, but theoretically, the deeper the tree is traversed, the more accurate the best move found will be. By parallelizing the Go game tree search, we have successfully reduced the computation time, enabling deeper levels of the tree to be reached in smaller amounts of time. Results for the sequential and GPU implementations were compared, and the highest speedup achieved with the parallel algorithm was approximately 72x at 6 levels deep in the game tree. Although there has been related work with respect to game tree searches on the GPU, no exact best move search algorithms have been presented for Go, which uses significantly more memory due to its large board size. This paper also presents a technique for reducing the amount of required memory from previous game tree traversal methods while still allowing each processing element to play out games independently.