Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: FINAL results Nunn test Hiarcs - CST

Author: Robert Hyatt

Date: 07:59:07 10/06/98

Go up one level in this thread


On October 05, 1998 at 03:56:45, Bruce Moreland wrote:

>
>On October 04, 1998 at 10:39:57, Moritz Berger wrote:
>
>>Performance on ICC was:
>>
>>Opponents avg. rating: 2393
>>CSTWin/BETA +19 =9 -38  total: 36%
>>ELO-Performance 2284
>
>>In the Nunn test, CSTAL/Black achieved an ELO performance of 2300. Quite
>>accurate, in my opinion.
>
>An ICC rating can't have any relation to reality other than that the numbers are
>something over 2200.
>
>Lots of those games were with Crafty, I think, and Bob had it set to use only
>one processor, so that must have had some effect.  Crafty is probably the wrong
>program to use if you want to see CST rack up a big score, because it is so
>paranoid about getting mated.
>
>CST plays funky stuff, which I like.  I think that is a good thing, it makes for
>interesting chess against the humans, and sometimes makes computer opponents
>look stupid.
>
>The problem is that it's not like this aggressive style is an enhancement to the
>strength of the underlying engine, it seems that instead that there is a
>downside to this style, if the goal is to maximize points.  Sometimes it can't
>win the positions it gets, it's like it tries to do one sacrifice too many or
>something.
>
>I think CST is a wonderful project but I don't buy arguments that it is only a
>little bit of fine tuning away from being champion.  It seems to have a hard
>time competing with the rest of the crowd, but I give it credit for not being a
>part of the crowd.
>
>bruce


I think you are right with respect to Crafty.  I used to have lots of
conversations with an IM that uses the handle "greg1".  He is a 2500+ USCF
player and *loves* kingside attacks.  He used to discuss his attacking style
with me all the time. and for a good while, one comment that kept my attention
was "Crafty is playing fine as far as positional and endgame chess goes, but
it is not as "quick" as WchessX is to detect kingside attacks, and, as a result,
is easier to defeat with a slow king-side buildup than wchessx.  He said the
same thing about Ferret (this being a year ago at least).  Bruce and I continued
to chat and to work on king safety, and the last 30 games he played with Crafty
saw 3 losses and the rest wins...  And he commented that it finally was becoming
acutely aware of kingside buildup and was reacting better, as well as having
better judgement on how to "crack" the kingside if it had to do "something" (IE
it seemed to understand files better after all the changes.)

So, in Bruce's words, "crafty is paranoid about king safety" is a pretty good
summation, and it does make it more difficult to simply move pieces to the king-
side and wait for something good to happen...  This used to work quite well, as
even WchessX would play like this, but as I slowly tuned king-safety, these
"attacks" worked less and less frequently.  Not that this is "solved" of course,
because in understanding attacks, it also understands how to attack, and there
is a very fine line between attacking your opponent and attacking yourself, in
some positions...

But in watching a lot of these games, I think a human would have more serious
problems with  the attacks than a cool piece of silicon with even rudimentary
king-safety...  but in the middle of the attacks, CSTal will often toss a pawn
to avoid serious retreating, and in many games, that pawn wins the game when the
attack doesn't lead to a mate or win of material.

This is a difficult problem, to be sure, because Crafty also sacs a pawn on
occasion, or an exchange on occasion, and it had better be right.  IE the
eval had beatter be measuring those things that make the position good, and
not those things that are present *because* the position is good.  IE it had
better be measuring causes and not effects...

But, in watching, the new CSTal does seem tactically stronger to me...  IE I
don't see wild piece sacrifices for nothing but "a hope and prayer" so that it
doesn't toss games away in the middlegame, which is good.  But if it has a
weakness, it has to be in the endgame, because it appears to be too slow there
and gets into deep trouble with deep searchers, because it still seems to be
using its "keep it complicated" evaluation even when the position is not, and
that costs time...  IE with RNPPP vs RNPPP, there is not going to be a serious
attack most likely, yet it seems to keep trying, and ignores a pawn that slowly
creeps forward, until it is too late...

But the skill of the program is moving in the right direction...  And a program
that is getting better has no upper bound on how far this can go...



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.