Localizing Search in Monte-Carlo Go using Statistical Covariance

Abstract

The classical Asian game of Go presents a difficult challenge for artificial intelligence programmers because of its vast search space and the difficulty of static board evaluation. One of the techniques humans employ while playing Go is spatial localization – they direct their attention toward the important points on the board and ignore the unimportant areas. This note describes using statistical covariance as a tool to execute spatial localization in our Go program, OREGO.