[computer-go] opponent strength differences for testing komi

[Aloril]

On Wed, 2006-05-03 at 08:20 -0400, Don Dailey wrote:
> I don't think strength matters for testing komi - when 2 programs
> play with significantly different strength, it's just harder to
> measure accurately - it takes more games.
> 
> For instance,  if you took a random program vs a very strong program,
> it might take billions of games to ascertain which komi is correct.
> But it could in principle be done.
> 
> The win ratio's of course would be much smaller.   If we image that
> 6.5 is the right komi  then 2 equal program might show black winning
> 55% but if you tested 2 programs of significantly different strength it
> might show up as 50.1 %
> 
> So the data from really weak players is almost useless when playing
> against strong opponents.  
> 
> In my own self-testing with botnoid I discovered that strength DOES
> seem to matter when both players are equal.    A small komi seems to
> work better for very weak programs and 7.5 seems to be the best komi
> for programs like Lazarus - one of the stronger ones on CGOS.
> 
> Has someone ever done this test with random players?   It would be
> interesting to see the results.

1.5 as komi seemed appropriate:
http://londerings.novalis.org/wlog/index.php?title=Even_games

-- 
Aloril <aloril at iki.fi>