Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Rybka and a different evaluation

Author: Stan Arts

Date: 07:22:29 02/14/06

Go up one level in this thread


On February 14, 2006 at 07:44:47, David B Weller wrote:

>Hello Chtistopher,
>
>You ask a very interesting question / make an interesting remark
>
>With regard to your Knight example:
>
>Certainly PST comes into play, as well as limited mobility, but I doubt most
>programs can ever justify nearly -600 points for for the sum of those [and
>purhaps other] defects.
>
>Fact is, IIRC, most versions of Crafty clamp the value of King Safety to 180!
>
>So theoretically, you could surround a king with 8 queens and the static eval
>would be -180 ! [for KS that is]
>
>So it is hard to identify/ define these facts without much, much code
>
>-David
>
>Look forward to more comments on this....



Giving a score of -300 to a knight on the rim will mean that the program will go
nuts about the knight and might sacrifice material (maybe 300 worth of
material..) to free that knight or whatever, and then you'll lose the game
because you threw away lots of material to have a mobile knight. But you can't
mate with a lonely knight.

Actually, the scores agressive programs (Gothmog, Zappa, etc.) give is already
high and about the maximum you can give to have a high overall playingstrength I
guess. Going further in that will sometimes be beneficial when the program
guesses right (it turns more into gambling then) , but when it guesses wrong it
loses because some attack or initiative fails and loses a lost endgame.

The same goes for king safety. Maybe you can go upto scoring it a rook or more,
as long as you have intelligent king safety, but that's the problem, and it
might take you years of experimenting to come up and tune the right patterns. If
you can't spend years on that, it's better to use lower scores, and "hint" the
program in the right direction.
So having low-ish scores of one or two pawns for something like kingsafety hints
the program in the right direction, without taking big risks with it's material,
and will play a good solid game overall. But sometimes it will miss a
speculative shot. So having lots of knowledge and high scores would be best, but
if your program isn't that smart, better not use very high scores, because it
doesn't know what it's doing.
Also, a balanced evaluation using low-ish scores for everything will still cause
a program to pay much attention to a knight on the rim etc. even if the program
might just score it 200. Also programs search deep and accurate, and often find
a way out for a trapped piece, so you can't easely say if a trapped piece is
really lost or not, etc.

Stan



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.