Knowledge Representation and Its Refinement in Go Programs

Abstract

It is often said that, unlike chess, an exhaustive search of the game tree for deciding the next move for Go is impossible, because of the huge search space. Therefore, in Go search is limited to local fields only. For more global decisions, various strategic human knowledge should be applied. Almost all Go playing programs (for example, Zobrist 1970b; Ryder 1971; Sanechika et al. 1981) settle on an evaluation function to represent this knowledge. These programs decide on the next move by selecting candidates with maximum evaluation function values. However, Go programmers often judge that the program’s maximum-value moves are bad or unsuitable for a situation (a board state, or placement of stones on the board), so that the only way to get the proper result is to refine the evaluation function. But a temporary refinement for one situation may cause a contradiction for another.