Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: The superior Rybka chess knowledge

Author: Ross Boyd

Date: 12:30:03 01/18/06

Go up one level in this thread


On January 18, 2006 at 15:05:36, Dann Corbit wrote:

>On January 18, 2006 at 14:39:40, Fernando Villegas wrote:
>
>>Some days ago I talked of knowledege to know which knowledge is irrelevant as a
>>probable feature of Rybka. And almost nobody answered or commented my post.
>>Surely it is an advantage to be a chess programmer in a site like this.
>
>Probably the only one who can answer correctly is the author.
>
>But you can look at some interesting generalizations.
>
>I have found that some evaluation terms add nothing either to the EPD solution
>count or the contest point gathering effectiveness of an engine.
>
>An example is queen mobility.  I don't know of any logical reason why its
>computation should not help, but it doesn't.
>
>I have seen other chess engine authors come to the same conclusion.

Agreed. I do pseudo mobility only for bishops and knights. Queen mobility
improves nothing. However, I do piece-square so the queen knows about
centralisation.

>
>So there can be a case where you add this wonderful knowledge term that makes
>your engine smarter and have better chess insight.  But when you compute it and
>add it in to the overall score, the net effect is nothing except to slow the
>engine down.

If you do add knowledge it has to be relevant, non-redundant and preferably
augments the search. Mobility is a good eval function because of its general
usefulness in 95% of positions. Good candidates for chess knowledge are terms
that effectively increase the depth of the search... eg recognition of bad
bishop and rook pawn draws. Search alone is not enough since a good static eval
can detect this situation many plies earlier.

Ross



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.