A Small Go Board Study of Metric and Dimensional Evaluation Functions

Abstract

Abstract. The difficulty to write successful 19x19 go programs lies not only in the combinatorial complexity of go but also in the complexity of designing a good evaluation function containing a lot of knowledge. Leaving these obstacles aside, this paper defines very-little-knowledge evaluation functions used by programs playing on very small boards. The evaluation functions are based on two mathematical tools, distance and dimension, and not on domaindependent knowledge. After a qualitative assessment of each evaluation function, we built several programs playing on 4x4 boards by using tree search associated with these evaluation functions. We set up an experiment to select the best programs and identify the relevant features of these evaluation functions. Thanks to the results obtained by these very-little-knowledge-based programs, we can foresee the usefulness of each evaluation function.

THE MOVE-DECISION STRATEGY OF Indigo

“…Following the general trend provided by [Chen 1990[Chen , 2000 or [Mueller 2002], our paper provides the example of the move decision strategy of Indigo [Bouzy 1995a[Bouzy , 1995b[Bouzy , 1999[Bouzy , 2002a. Although [Chen 2001a[Chen , 2001b classifies Indigo into the "global selective search" category, (which was in keeping with our answer), this paper will show that Indigo could also have been classified into the "static analysis" category or into the "incentive/temperature approximation" category as well.…”

“…But it is hard to find the adequate corresponding EF which could be used by a tree search to verify the validity of moves: such is the problem we are faced with in developing Indigo. [Bouzy 2002b] addressed this problem. In summary, for some classes of moves, there are no corresponding EF, and Indigo has to use another method than TS to select the best urgent move.…”