Author: Thorsten Czub
Date: 00:56:47 12/11/00
Go up one level in this thread
On December 10, 2000 at 23:06:56, Christophe Theron wrote:
>You are much too harsh against Fritz, Thorsten. Actually against Frans.
>I don't see drawing score in difficult positions as a charateristic of Fritz,
>even in your example.
but I see the drawing-score for years by now. and it happens very often.
i know this because i watch the games the autoplayer generates.
and i relate the scores manually with my own eyes.
what i do normally is, to use another program when the 2 have those different
score behaviour.
and i can tell you from experience that this "behaviour" is in fritz for many
versions.
i like this behaviour. it helps the other programs to win against fritz, and i
like it because it makes me laugh. i can be sure when fritz has these scores,
it is blind.
>So if the search does not say "it works", the programmer's choice is to give a
>rather low weight to the king attack.
well - the search is wrong not saying "it works". the search does not KNOW
if it works. it would maybe know if it would compute 2 hours on that move on a
1000 Mhz machine, but it cannot be SURE within 3 minutes.
therefore - if you are right here - the problem is even more stupid:
frans tells fritz, or the search tells fritz: evaluation := arround_sero
idea behind:
low scores help to stand king-attacks against speculative programs.
>And it's a fact that the side that has a strong attack can generally force a
>draw if it wants to.
why should it do so ? it has a strong attack.
would you fire the guy who makes the most money for you in your company ??
ì guess not.
> And generally, forcing the win takes much more plies, so
>only the draw can be foreseen in a reasonnable time by search.
if it is not seeing the attack as an attack, and if it believes the
opponent would make a draw, it should better try to play checkers instead of
chess. "4 wins" or other games. but not chess.
>That's all you can say about this behaviour. I don't think it's something
>written in stone in the Fritz engine. It's only because the programmer's choice
>is to rely more on search than on evaluation to detect strong king attacks.
this behaviour is very often. its in almost any game fritz loses. its in the
engine for versions. they never changed. and they changed much from version 4 to
5.32 to 6.
>Now you can disagree, as I do, with this choice. However I can understand it,
>because until version 12.0, Chess Tiger was doing the same. And not taking care
>of king attacks is much, much simpler in term of programming efforts.
right. but this way it is not plying chess.
chesstiger12 and even 11.2 played better quality of chess than fritz ever did.
thats the difference.
i don't care how you do it, as long as it produces good chess.
i care in the moment it produces weak shit chess (lost checker games).
>Now I have changed my mind, and the best proof is Gambit Tiger. But my task is
>huge. Correct evaluation of king attacks is extremely difficult (solve this and
>you are close to solving chess).
there is no correct evaluation in chess.
you have to create your own idea how much a king attack is worth for YOU
depending on your playing style.
what is a king attack worth ? a pawn ? 2 pawns ? 3 ? 4 if you have mate chances
? 5 if you make mate-net ? 6 if you get 6/7th rank ? and 9 is you
have mates ??
what is a connected passed pawn worth ?
which pieces can win against king in an attack, which not ?
this is all in cstal !
cstal knows which pieces, and it gives a value ! but these values are not
accurate. nobody knows what is accurate. they are inaccurate. it doen't matter,
the only reason cstal loses is not because the values are inaccurate
but because it gets outsearched and because chris has such a slow search.
("all chess programmers are lazy").
>Fritz maybe underestimates the king attacks.
right. it's unknown country for it.
>As Genius and many others do.
good for other programs. they get no title at championships without knowing.
they play checkers, other program play chess.
>That
>gives them an incredible steel-nerves style. When it works, they seems to play
>like ET. And when it fails it makes them look very stupid.
right.
>Gambit Tiger has another approach and takes care of king attacks. It gives it an
>interesting attacking style. When it works, it seems to play like a real GM. But
>when it fails, it looks as stupid as the programs of the other group.
right.
same with cstal.
only difference between cstal and gambit-tiger:
gambit tiger has a good search.
cstal knows more.
there are more bugs in cstal.
>I don't see the draw score behaviour like an intrinsic weakness of Fritz. It's a
>choice, a pragmatic choice, call it a bet, that has been working for years in
>computer chess. We will see if it will keep on working in the next years, but
>that's another story...
the bet loses. for years.
junior is like fritz, but goes different. it knows a little about king-safety
and also has low scores. and gets 50% against GT.
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.