Igo Hatsuyōron 120 and
Go-Playing Programs (2009 - 2011)
The Importance of Variable Komi
"Modern" Go-playing programs play at their strongest, when they are confronted (in their own estimation) with an approximately balanced overall situation. If they feel "too far" in front, there is the danger that they tend to confine themselves to very secure moves, not trying to extend their lead any further. For a lead of half a point at the end of the game is indeed fully sufficient for victory. However, if they feel "too far" behind, their moves begin to become the more irrational, the larger the supposedly unassailable lead of their opponent is. With humans, in such cases, we would speak of "desperation moves". By a suitable adjustment of the Komi, one can fool the program into assuming that the score is balanced. First, this means, that the analysis of a particular position must be performed several times, each time with different Komi. For the program's assessment of the score may differ from ours quite substantially, so we have to get an approximate idea of this assessment. Secondly, it is important to note that the Komi - given to assume a "balanced" position - will change in the further course of the move sequences. In general, the assessment of the program will become more adapted to the human one, the further the game progresses. If we want the program to make us show some viable candidate moves, it is advisable to make the program believe that it is behind to some small extent, but that it could probably still catch up. The required multiple analyses of identical positions, each with varying Komi, results in using an enormous amount of time. Therefore, the processing of the complete variation-tree of the tenuki paths after the guzumi always took me several months. | |