Page 1 of 1
Impact of komi differences
Posted: Sat Jan 11, 2020 6:44 am
by John Fairbairn
It seems that many people are taken by Katago's ability to give a score estimate (and I can see why). As far as I know, it's all based on a komi of 7.5. But the 0.5 is there just to avoid jigos. I'm curious to know how much Katago's score estimates might change if training komi were to be set at 7.000000000000000000000000000000000000000000000000000000000001. Would it simply be the "obvious" ~0.5, or are more complex issues involved?
Re: Impact of komi differences
Posted: Sat Jan 11, 2020 7:08 am
by lightvector
KataGo's code validates that komi*2 is an integer, and if not it rejects it as an invalid input, the same way it would reject it if you said the komi was "foobar" or some other non-numeric value.
Integers values are supported (because integer*2 is an integer). You can set it to exactly 7, if you like. Training occurs on a wide variety of komi values, both integer and half-integer, so KataGo already knows how to adjust its play and its evaluations for different komi. Within reason - maybe if you set a komi of 100 but give a position where black is winning by about that much on the board it won't behave as well in judging how well the komi balances the board, since that's going to be enormously outside of KataGo's training experience.