Effect of potential model on Monte-Carlo Go: Pruning the igo game tree using potential and potential gradient

Abstract

In this study, we tackled the reduction of computational complexity by pruning the igo game tree using the potential model based on the knowledge expression of igo. The potential model considers go stones as potentials. Specific potential distributions on the go board result from each arrangement of the stones on the go board. Pruning, using the potential model, categorizes the legal moves into effective and ineffective moves in accordance with the threshold of the potential. In this experiment, 4 kinds of pruning strategies using the potential and 5 kinds of pruning strategies using the potential gradients were evaluated. The reduction rates differed according to how the potential and potential gradients were set. The best pruning strategy resulted in a 20% reduction of the computational complexity. In this research we have successfully demonstrated pruning using the potential model for reducing computational complexity of the go game. © 2013 Springer-Verlag.