Author: David Dahlem
Date: 10:59:45 12/17/05
Go up one level in this thread
On December 17, 2005 at 12:55:30, Dagh Nielsen wrote: >On December 17, 2005 at 12:42:26, David Dahlem wrote: > >>>A comment about test position scores: >>> >>>Ideally, I would aim for test positions that: >>> >>>(1) Score close to 50%. >>>(2) Have a relatively low draw score (=interesting positions). >>> >>>In both instances, the point would be not to give "free" points to the weaker >>>engine. In light of (2), you may want to provide statistics for the 1-0/½-½/0-1 >>>scores of the different positions. >> >>I would like to compile my own custom opening testset. Is there a list somewhere >>with the win/loss/draw stats for the various openings only for computer players, >>not human players? If not, is there a software that will output these stats from >>a large database of computer games? >> >>Thanks >>Dave >> >> >>> >>>I see no problem if black scores slightly better, positions are played from both >>>sides anyway (so, in a sense, test positions are color neutral). >>> >>>Regards, >>>Dagh Nielsen > >Hi Dave > >Here's one approach: > >You can download a lot of engine games at the following address: > >http://www.michael-waesch.de/trade/trade.htm > >There are more than one million engine games there. > >Then you can put those games into a chessbase tree (the books/trees with *.ctg >extension). And then, at least in Fritz UCI, you can click opening book, right >click and choose properties, and tick "show statistics". This will give you the >win/draw/loss statistiscs for the position. > >Just one word of caution: Those stats are based on games where the engines play >with books usually containing lots of moves beyond the current position. So it >will not necessarily give a pointer about the expected "raw" stats that would >result from using such a position as a test position without further book moves. >Probably one would have to just conduct experiments, and if a position turns out >to give an even score, and a low draw rate, it can be included in a test set. > >Next thing is contemplating how one ensures that a test set is reasonably >"fair"; a good testset would include a set of interesting positions that is >expected to somehow represent "the full opening theory", ie., it should not be >biased towards specific pawn structures, or overly tactical positions, etc. The >whole point being not to unduly favor one engine skill over another. In other >words, it should include a bit of everything... Hope that makes sense :-) > >Regards, >Dagh Nielsen Hi Dagh Thanks for the excellent tips. I'll give it a try. Regards Dave
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.