Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: learning evaluation weights (was Re: Genetic algorithms for chess?)

Author: Komputer Korner

Date: 22:38:24 05/24/98

Go up one level in this thread


On May 23, 1998 at 17:10:56, Robert Hyatt wrote:

>On May 23, 1998 at 14:18:04, Ren Wu wrote:
>
>>I think KK miss the point here.
>>
>>Don's paper has outlined a framework to apply TD methods to chess
>>evaluator. In fact, it is not just for chess, his framework can be
>>easily applied to most other games as well. And what it can learn is not
>>just piece values but can be any evaluaion terms.  For me, the method,
>>or the way to solve the problem, is a lot more important.
>>
>>I doubt that the search depth will effect the knight value that much,
>>but that is once again not the point, the point here is the way doing
>>things, not the experiment results.
>>
>>If you suspect that search depth will effect the knight's value, why
>>don't you repeat the experiment but with greater depth? And report the
>>result here. Even if you come up a vast different values, there is still
>>no flaw in this framework, because you are still in the same framework.
>>The framework has no search depth limitations.
>>
>>There is no major flaw in that research. ( And maybe there was  flaws in
>>your way to look things. :-)  )
>>
>>Ren (renw@iname.com)
>
>
>
>the biggest "flaw" was that Alan looked at "the result of the results"
>rather than at "the results".  The learning looked good to me.  Just
>because
>the piece values appear to be "a little different" is not cause for
>alarm
>IMHO.  First, there's nothing to say that 1,3,5,9 are right (I don't use
>those, for example.)  And there is also nothing to suggest that the
>values
>might be different depending on the "core" of the program being used to
>learn the values...  IE a program with strong pawn structure analysis
>might have a pawn value > 1, while a strong tactician might have
>knight/queen
>!= 3,9, and so forth...


The 1,3,3,5,9 ratio has stood the test of time with the only other
significant one that I have seen is the one by Larry Kaufmann who put it
at 3,10,10,15,29    3=pawn  29 = queen   with the 2 bishops =21
These point counts are important and I didn't question the experiment
itself. I simply questioned the lack of search depth for a chess
experiment that purports to come up with a different point count and
pass this point count off as maybe more accurate than the historical
one. It may be more accurate but the Knight value looks too low. There
have always been particular materiel equation exceptions to the
historical point count and I looked at the experiment in that light. I
was interested in it from the chess perspective only.  One particularly
gnawing exception to the historical point count is the 2R+N=2B+R
equation. This is supposed to be equal but of course it depends on
pawns. Of course there could be a different optimum point count for each
materiel equation on the board including pawns, and ultimately there is
an optimum different point count depending on the exact position. Of
course we are eventaully talking of the point count in general, because
point counts that depend on the position are useless and even point
counts that depend on materiel equation are impractical for OTB players.
Computers are a different story here. However for practical players, one
general point count or at least no more than one for each materiel
equation (not counting pawns) is a necessity for playing chess well. The
historical point count has held up fairly well. That is why I was
surprised that the experiment showed such a low value for the knight and
backed the accuracy up with another match based on search engines that
had positional algorithms as well. Don Beal's admission that the extra
match with historical point count against the experiment point count was
carried out with the same 4 ply depth limit, shows that the extra match
results are not meaningful for determining the accuracy of knights in
general. 4 plies is not enough to show the knights in their best light.
Rooks too can be undervalued at that limit and the results of the
experiment show that the rooks are undervalued against the historical
point count. 2 wrongs may make a right here but I would like to see
experiments of point count against point count at much deeper search
depths.



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.