Computer Chess Club Archives




Subject: Re: The superior Rybka chess knowledge

Author: enrico carrisco

Date: 15:18:44 01/19/06

Go up one level in this thread

On January 18, 2006 at 13:34:17, Chrilly Donninger wrote:

>After playing several engine matches against Rybka (chess programming is a
>rather boring job) I have come to the conclusion: There are a few special
>evaluation features of Rybka which are really unique. It is interesting that
>some seamingly relative unimportant feature appear regularily on the board. The
>opponent has no idea of this feature and does not prevent it. And the search
>always finds a way to reach the pattern. Rybka has e.g. some special passed pawn
>evaluation terms. I do not want to tell the details, but the game Zappa-Rybka,
>Paderborn 2005 is a prototype game for one of these special features.
>But the main chess-knowledge which sets Rybka appart from other engines is
>ignorance. The omission of features which other engines have incorporated.
>I have written a longer article for the German "Schachkalender 2006". The
>message of this article is: Most of the published chess knowledge is completly
>useless. Give your favorite chess-enemy your chess-books as a present. They will
>do some harm on his play.
>Rybka seems to be to prove of this hypothesis. If a feature is - in a given
>position - correct, it is of course an advantage if a programm has implemented
>it. But if its wrong, the programm hangs on an advantage which does no really
>not exist. Or even worse, it sacrificies another advantage to reach the pattern.
>I realized the principle: "It is sometimes more important to remove features
>than to add ones" several times in the Hydra project. E.g. Piece-Square Tables
>are generally considered as a "must have". Strong Chessplayers do not like them.
>It is very unnatural for them to evaluate a piece without considering the
>context of the other pieces. It took some time till GM Lutz convinced me to
>remove them in Hydra. And indead, the programm played considerably stronger with
>Insofar is the Rybka approach intelligent ignorance.


Let me restate what you have written and add some of my own points (some of
which I have stated here previously, in other Rybka threads...)

So you think Rybka is a fast searcher with little knowledge... :)  Well, I don't
dispute that (in fact I've claimed it many times in past threads...)

However, what it _does_have_ is a very well tuned evaluation which seems to be
correct 99.9% of the time.  Why?  It concentrates on what is known to be
successful in chess.  It speculates in evaluation on a few things:  passed pawns
and activity -- and this is obvious by the short (and sometimes long) pause,
often with no search information being updated (position recognition/bit board
access.)  Couple this with the fast search and it often makes these count.

Rybka uses an invisible Genius-like search to protect itself from tactical
mistakes and sooner or later, when you are searching very very fast and deep,
your opponent will go wrong.  The extra search (especially when not included in
its definition of a node) makes it appear like it has more knowledge and that is
clearly what sells chess programs (ex: low nps display...)

Clearly you have a point that wrong knowledge is much worse than no knowledge
because it means you trade one possible real advantage for a different but
unreal advantage.  Also, in computer chess, if you are outsearched it can look



>P.S.: The omission of Piece-Square-Tables is a feature of Hydra. I do not state,
>that Rybka as skipped this feature too.

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.