Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: M-Chess Pro7 : strength ??

Author: Robert Hyatt

Date: 10:26:36 12/28/97

Go up one level in this thread


On December 28, 1997 at 11:49:41, Jeroen Noomen wrote:

>On December 27, 1997 at 05:39:21, Thorsten Czub wrote:
>
>Thorsten,
>
>I completely agree with you on this subject. I cannot see why people
>try to defend 40-moves opening lines ending with +4 and an easy win.
>This is disgusting and not in the interest of CUSTOMERS who want to
>buy a chess program.
>
>Regards, Jeroen
>g

while I agree, I don't see how this can be rationalized as bad while
commercial programmers spend 24 hours a day 7 days a week auto-playing
against their closest competition to tune to maximize results against
specific opponents without much regard for how the changes affect
overall
play.

I've seen the following, and have simply given up on seeing real valid
SSDF (or any other) test results as a result:

1.  programmers tune for well-known problem sets.  the Kopec-Bratko was
the first such example, but now the BT series has also been attacked
quite
successfully by the same approach.

2.  programmers tune for their top competitors by playing hundreds of
games
and adjusting things to maximize results.

3.  programmers cook books directly by hand-coding winning lines, or by
automated auto-play learning.  IE take Crafty and play a few hundred
games against a top commercial program and it will learn which book
lines
to avoid as bad, and which to favor as winning, without any human inter-
vention at all..

I'm sure there are other cases.

I have (at times) had a book that went *very* deep.  I can recall one
game
against a GM where he resigned at move 47 yet Crafty was still in book.
I
don't keep such deep lines any more because they are *mucho dangerous*
as
there can be lots of strong replies along a narrow line that have not
been
played.  For every game where crafty came out +4, there were two or
three
where it came out -4.  Bruce and I used to see these all the time in our
matches against each other on ICC.  At times one of every 2 or 3 games
was decided before either program had to think on its own...

So deep lines don't necessarily mean Mchess cooked anything, they could
represent a game that was played and included.  Someone with Chessbase
or
something could likely verify this.  Of course, it could be automated
learning as well...

But the days of knowing which commercial engine is best are long gone.
They
are all good.  But that's all we know now...



>
>
>>>I think you misunderstand what we wrote. Of course the learner-option
>>
>>Right. Learning function that works is passive defense against active
>>cooked lines of opponents !
>>I have nothing against this.
>>But putting a mass of deep lines in a book that were produced at home
>>with autoplaying against the commercial enemy-versions is nothing
>>passive.
>>This is an attack and increases the tug-of-war.
>>
>>>is not the point, we were talking about cooked lines. In reaction on a
>>>posting by Thorsten, who found out in a game MCP versus Hiarcs that
>>>MCP played 37 moves of 'theory' and left the book with a +4 score.
>>>
>>>THAT is a development I dislike.
>>
>>Right. I know that the learning function of Mchess6 was not that
>>brilliant.
>>So what shall i have in marty programming a better one in mchess7.
>>But the cooked lines are the problem.
>>
>>
>>>
>>>Regards, Jeroen



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.