Virtual global search: Application to 9×9 Go

Abstract

In games, Monte-Carlo simulations can be used as an evaluation function for Alpha-Beta search. Assuming w is the width of the search tree, d its depth, and g the number of simulations at each leaf, then the total number of simulations is at least g × (2 × wd/2). In games where moves permute, we propose to replace this algorithm by a new algorithm, Virtual Global Search, that only needs g × 2d simulations for a similar number of games per leaf. The algorithm is also applicable to games where moves often but not always permute, such as Go. We specify the application for 9×9 Go. © Springer-Verlag Berlin Heidelberg 2007.