PASCAL - Pattern Analysis, Statistical Modelling and Computational Learning

Time Management for Monte-Carlo Tree Search Applied to the Game of Go
Shih-Chieh Huang, Rémi Coulom and Shun-Shii Lin
In: International Conference on Technologies and Applications of Artificial Intelligence, 19 Nov - 20 Nov, Hsinchu City, Taiwan.

Abstract

Monte-Carlo tree search (MCTS) is a new technique that has produced a huge leap forward in the strength of Go-playing programs. An interesting aspect of MCTS that has been rarely studied in the past is the problem of time management. This paper presents the effect on playing strength of a variety of time-management heuristics for $19\times19$ Go. Results indicate that clever time management can have a very significant effect on playing strength. Experiments demonstrate that the most basic algorithm for sudden-death time controls (dividing the remaining time by a constant) produces a winning rate of 43.2$\pm$2.2\% against GNU Go 3.8 Level 2, whereas our most efficient time-allocation strategy can reach a winning rate of 60$\pm$2.2\% without pondering and 67.4$\pm$2.1\% with pondering.

EPrint Type:Conference or Workshop Item (Paper)
Project Keyword:Project Keyword UNSPECIFIED
Subjects:Learning/Statistics & Optimisation
Theory & Algorithms
ID Code:7422
Deposited By:Rémi Coulom
Deposited On:17 March 2011