Computer Chess Club Archives


Search

Terms

Messages

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

Author: Komputer Korner

Date: 01:38:19 05/23/98

Go up one level in this thread


Well Don, if your experiment had nothing to do with search depths, then
why not do a 1 ply search? As soon as you go above 1 ply, you skew the
results because of searching. Obviously a knight move is different from
the others and temporal differences are affected differently with a
knight because it is not a sliding piece. If the whole article wasn't
about piece values but of something else, then ignore what I have said
up to now. Learning methods are affected by the environment and you
conducted an environment that had a search of 4 plies as a way of
determining results of games. If the values learnt were not the result
of a search, then where did you get the values? The article stated that
the values learnt were learnt without any domain specific knowledge. If
this is so, then the only other possible way that you could obtain
results is by searching or else you would have to traverse the whole
chess tree or learn the values by  piece handicaps but you didn't do
this. In  your match results section 5.1 you state that you verified the
final values by playing matches beteen identical search engines. At what
ply lookahead did these matches take place?  Just because you randomized
all candidate moves, doesn't obviate the need for knights being treated
special when calculating values. In your experiment it is obvious to me
that knights would be underated and from your results it certainly
verifies this. You didn't specify whether the candidate moves were
randomized in the matches. I will assume that they were not randomized.
Of course your results from the special matches seem to back up the
accuracy of the "new" values but this is misleading for the following
reason. The accuracy of the other pieces has a very big bearing on the
results. The accuracy of the traditional value of 5 points for  the rook
in relation to the other pieces has always been in question. an example
of this is the materiel balance of 2R+N =2B+R. This has been known from
the days of Spielmann. Your new point results more accurately reflect
this balance. Also B+p vs R balance is more accurately reflected in your
new balance. The advantage of the bishop is when you have both and this
isn't reflected in your charts. It may be that since the knights are
undervalued in your new figures, this reflects nicely in the power of
the 2 bishops. So your new values by accident more accurately reflect
the real values of the game when only considering one piece against
another and not considering combinations of pieces. However your
experiment could only value individual pieces, but i contend that your
knights would't have died as many deaths if the search depth had been
longer. However in that case perhaps your values results would have been
less accurate as a whole when looking at all 5 piece values. it is
interesting that the relation of the queen to the rook in your new
values might be more realistic than the traditional values. all these
materiel balance reasons  could more than make up for your undervaluing
the knight . All experiments so far with large statistical studies of
bishop vs Knight suggest that their values are much much closer than the
values you come up with.


On May 22, 1998 at 15:33:00, Don Beal wrote:

>On May 19, 1998 at 12:45:59, Komputer Korner wrote:
>
>>There was an article on refining piece values in the ICCA journal last
>>year, but there was a major flaw in the research becuase they used
>>lookahead search of only 4 plies and that is not enough for knight
>>manoeuvers.
>
>The article was "Learning Piece Values Using Temporal Differences"
>ICCA Journal, Vol 20, No 3, Sept 1997.
>
>I was concerned to find the inaccurate comment "major flaw in
>the research".
>
>Whilst I don't think "inkompetent computers :-)" are likely to
>damage reputations too seriously, I'd just like to respond...
>
>The intention of the paper was to present the learning *method*,
>and its ability to start from no information whatsoever about piece
>values and obtain effective ones.  The values learnt were *not*
>the result of the research; the fact that the method obtained
>sensible values as opposed to ineffective values *was*.
>
>That said, however, the values were, of course, of interest, which
>is why we reported them.
>
>We made no claim that the values would be the optimum values
>for larger search depths, nor, as we commented in the paper,
>optimum if the evaluation included positional terms.  (Our
>"proof of principle" experiment used piece values as the only
>evaluation!)
>
>As regards the lookahead depth, we didn't try large depths, but
>over the range of depths we did try, we found that the depth of
>search had only a small effect on the values obtained.  This is
>perhaps less surprising when one realises that the learning method
>reacts to *positions* and the associated game outcome, not to move
>sequences that might, or might not, occur between the position and
>the end of the game.
>
>
>Other people are experimenting with TD learning - if any are
>reading CCC, perhaps they have values for comparison?
>
>Don Beal.



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.