Gian-Carlo Pascutto: <[EMAIL PROTECTED]>: >Hideki Kato wrote: >> Gian-Carlo Pascutto: <[EMAIL PROTECTED]>: >>> Hideki Kato wrote: >>> >>>> delta_komi = 10^(K * (number_of_empty_points / 400 - 1)), >>>> where K is 1 if winnig and is 2 if loosing. Also, if expected >>>> winning rate is around 50%, Komi is unmodified. >>> I don't think the formula you posted is correct. >> >> It's just an experimental formla with no theoritical basis. I don't >> think we can discuss its "correctness." > >Even if it's a heuristic the outcome should be sensible. For what you >posted it doesn't seem to be at all. So I think that what you posted is >not what is in the program. That is what I mean by "correct". > >>> In the opening it gives delta_komi = 0.8 and in the endgame it gives >>> delta_komi = 0.1. >> >> I'm sorry but I don't understand what do you want to say. Too >> small? > >Yes. I don't understand how it can do anything if it adds _at most_ 0.8 >points difference. > >Should it have been 10* instead of 10^ or something?
I tried much bigger value first but it made my program very unstable. I decreased by a bit time by time and converged this. Of course, this value depends on when you start increasing or decreasing komi. Please remember this delta is added to or subtracted from komi on _each_ play. Adding 1.0 to komi on each move cuuld change the world dramatically in the middle of games on 19x19. IIRC, this small value, less than 1.0, is better to adjust komi _smoothly_. In other word, this formula is a conservative one. Sure, this value is sometime not enough. I had tried to find more effective formula but, in such cases, it was very hard to tune factors to get stable. Please note that I had to modify my code to 19x19 board, Japanese rules, etc, all in a few weeks so that I couldin't try many. I strongly believe you or others will be able to find much better formulae. -Hideki -- [EMAIL PROTECTED] (Kato) _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/