Re: [gnugo-devel] Level 15

[<address@hidden>]

> Even better would be to
> force-feed existing games into the two versions, let them both choose a
> move, and have a human arbiter tell which move was better. Of course this is
> hardly practical.

I believe that there are corpora of professional and very-good-amateur games. 
It should be possible to measure the probability, for a random move in a random 
game a corpus that gnugo level Y made the same move as the player.

This evaluation measure fails, of course, if developers explicitly look at the 
games to see where gnugo can be improved, because that's testing on the 
training data.

stuart 



--    stuart yeates <address@hidden> aka `loam'
  "To err is human--but it feels divine."  --  Mae West
X-no-archive:yes