Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Rybka ... THE BAD (analysis)

Author: Dagh Nielsen

Date: 01:01:29 01/10/06

Go up one level in this thread


On January 10, 2006 at 03:32:03, Majd Al-Ansari wrote:

>Actually after further analysing this position I have come to the conclusion
>that it is indeed a DRAW after Kf5! instead of Kh5?  So the human intuition was
>correct ;}   Black can draw this quite easily by waiting for white to commit. If
>white commits white manages to gobble up the white pawns in time for a Queen
>just as white pawns.
>
>After Kf5!
>
>[D]8/5p2/8/pp3k2/3P3p/7K/PP4P1/8 w - - 0 2
>
>Now if Kxh4 then Ke4! is a draw.  This is confirmed by Fritz 9, Fruit 2.2.1 and
>Shredder 9.1 all in less than 5 seconds .... The only engine giving a heavy plus
>for white after reasonable amount of time is Rybka (I stopped Rybka analysis
>engine after 10 min with Rybka showing +2.56 for white).  Obviously Rybka needs
>some more endgame knowledge in these kind of simplified pawn endgames.

Hehe, I stand corrected. I couldn't believe it would be necessary to count :P

AFAIK, it is common knowledge on playchess that Rybka is quite blind in pawn
endings. It is a little mystifying for me, since it handles advanced pawns so
well in the middlegame, but I figure they are not the same.

Another thing I've noticed, referring to my second example, it seems like Rybka
is very good at finding "paths" exploiting weaknesses, both with king and with
knight (though I have not "tested" this statement with much observation). This
may "only" be a result of just a generally excellent search, but I'm also
wondering if maybe Vas has managed to do something like having weaknesses and
opponent weakness-attackers reinforce eachother, both in evaluation and in
search. Maybe something like this:

1. Identify potential weakness (potential only since no weakness is a weakness
without an exploit :P).

2. Check for potential exploits (piece maneuvres or something, maybe square
color for bishops).

3. Expand search on moves included in the maneuvre OR just up the evaluation
where condition 1 and 2 holds.

I admit I have absolutely no idea if it is possible to do this effectively (ie.,
avoiding all kind of stupid expansions), though I would think programs already
do something like this at least for endgame evaluation.

Condition 1 could maybe be modified instead for identifying "hot spots", for
example, squares ahead of a passed pawns, check for and expand supporting
maneuvres.

Dunno, it just impressed me a lot when Rybka found that kingwalk 2-3 plies
earlier than Deep Fritz 8, and also that it evaluated the position prior to that
as advantageous for white even though it was a knight vs. bishop endgame. (I
haven't checked with other engines, maybe it is just an instance of Deep Fritz 8
being stupid :P.)

Kind regards,
Dagh Nielsen



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.