Re: AI opinions on "fairest" komi by board size
Posted: Mon Aug 14, 2023 6:35 am
2x2? To see if KataGo does worse than Bill Taylor's theorem (IIRC, Black's 1 point win under area scoring and superko)?
We can see KataGo's thought process for 2x2 on multiple rulesets!RobertJasiek wrote:2x2? To see if KataGo does worse than Bill Taylor's theorem (IIRC, Black's 1 point win under area scoring and superko)?
Also, an idea is instead of trying multiple komis, simply train only on no komi games and increase the weight of a win according to winning score, so when a net wins by 3 points, the effect of it's weightings on creating the new version are 3 times what it would have been if it won by one point, for example. But it means the winrate for a particular komi is just derived from the proportion of games won by black by that margin or higher in these no komi games. The total number of games between area rulesets Chinese, AGA, Ing, New Zealand, and territory rulesets Korean, Japanese, etc would be even with the runs of rulesets within those also distributed equally, however, there will often be games played using one ruleset that would be legal in another, and in those cases the results of the winning margin should included in the data for those rulesets too, or better yet, whenever a move is played that is legal in some rulesets but not others, split the run into multiple branches that comply with different rulesets. I had the thought of making dual area-territory rules the standard in lentitear, so the final score is half the difference in number of stones on the board plus the difference in territory, and since fair komi for territory is 6.5 and fair komi for area is 7, fair komi in dual area-territory scoring with each weighted at a half would be 6.75!