Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: bugs and strength of programs

Author: Roberto Waldteufel

Date: 02:48:04 09/20/98

Go up one level in this thread



On September 19, 1998 at 19:42:26, Inmann Werner wrote:

>On September 19, 1998 at 19:29:41, g.müller wrote:
>
>>I have the same problems with CAISSA98 (it plays with 0 points of six in mclanes
>>summer t.). In version 5.00 it gets more knownledge about pawn-structures, but
>>it plays really worse. In version 5.14 it plays a good endgame, in a test game
>>aginst mirage it made remis because of a very good endgamestyle. (in move 34 it
>>sees herself with -3.14 and in move 61 remis!) But a good endgame with a worse
>>middlegame is nothing!!!!!!!!!!
>>Parameters for a good middlegame i search sice years, but without any sucess.
>>Sometimes i think fastness is the only thing, but i do not want to accept this,
>>because i hate stupid tactical programs.
>
>I agree with you. I also hate stupidity.
>But lets have an example.
>In endgame, I tried to fix the problem of running pawns. I put in much code and
>knowledge, but in the end, the program was only slow, and did only seldom hit
>the point. I put everything out. Now I have about 10 lines of code, where I
>only "estimate" the race and in most cases it works really fine! It also solves
>some other "pawn problems". I will not change the lines, because i dont really
>understand, why it works, but never change a "winning team".
>What I want to say: Putting in knowledge must be a "Kompromiss" (whats the word
>in English?). You have to get balance between speed and knowledge.
>
>Werner

I think it is useful to consider the elements that constitute an advantage in
chess as viewed by strong human players. There are several different features,
and decideing which ones are the most important is one of the things that
requires most skill in high level chess. Now some of these, like material
balance, are easy in a program, and all programs take this into account. Others,
like space control, piece activity, king safety and pawn structure are more
difficult to program. I think the obvious pawn features like passed pawns, or
isolated, doubled and backward pawns, can be counted easily enough, but it is
very difficult to correctly asses the relative importance of all these features,
and if the weightings are out the program may play worse, or at least not pay
for the extra CPU time spent evaluating all the features (although this is not
very costly if you hash the pawn structure evaluation).

I have noticed that some strong (human) players place much less emphasis on pawn
structure defects than others. Often a damaged pawn formation comes with
enhanced piece activity, and this may sometimes more than compensate for the
weakness of the pawns, eg a doubled or isolated pawn implies an open file for
the rooks. The bishop pair is often very effective with broken pawn structures.

With this in mind, perhaps we have the reason why the pawn structure evaluation
fails to make really big improvements in playing standard of some programs.
Indeed, for this reason I tried to emphasize piece activity in Rabbit's
evaluation, and only consider pawn structure near the king in any detail.
Generally Rabbit prefers active pieces to good pawn structure.

Best wishes,
Roberto



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.