Convolutional Monte Carlo Rollouts for the Game of Go

Peter H Jin, Kurt Keutzer

Abstract: In this work, we present a Monte Carlo tree search-based program for playing Go which uses convolutional rollouts. Our method performs MCTS in batches, explores the Monte Carlo tree using Thompson sampling and a convolutional policy network, and evaluates convnet-based rollouts on the GPU. We achieve strong win rates against an open source Go program and attain competitive results against state of the art convolutional net-based Go-playing programs.

Conflicts: berkeley.edu, cs.berkeley.edu, eecs.berkeley.edu