Board state evaluation in the game of Go - Preliminary WPE report

The game of Go is very interesting from a machine learning point of view since it has been shown to be PSPACE hard to solve explicitly [9]. A major factor for the complexity is due to being played on a 19x19 board where moves can not be considered using only spacial locality. Heuristics are difficult to make for Go because much of the game is qualitative and it is hard to assign quantitative values to such situations. This paper extends GNU Go [6], a hybrid heuristic based Go program, in three ways: decomposition of patterns for influence propagation, adaptive strategy for move generation and risk identification and resolution for move generation. Pattern decomposition is analyzed and shown to be inferior to the original patterns for influence propagation. Adaptive strategy and risk identification and resolution are shown to improve the performance of GNU Go by a statistically significant amount.