Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: likelihood instead of pawnunits? + chess knowledge

Author: James Swafford

Date: 14:44:18 10/25/02

Go up one level in this thread


On October 25, 2002 at 14:49:24, Sune Fischer wrote:

>On October 25, 2002 at 14:17:49, James Swafford wrote:
>
>>On October 25, 2002 at 13:15:03, Sune Fischer wrote:
>>
>>>On October 25, 2002 at 12:39:38, Ingo Lindam wrote:
>>>
>>>>I would really like to see the computers measure a position rather in a
>>>>set of probabilities e.g. (P+,P=), where
>>>>
>>>>P+ = Probability in the position to evaluate white/player to move will
>>>>win and
>>>>P= = Probability that position will end in a draw
>>>>P- = Probability in the position to evaluate white/player to move will
>>>>lose
>>>>
>>>>with P+ + P= + P- = 1
>>>>
>>>>(also a confidation measure about the Probabilities might be useful)
>>>
>>>The pawn scale is equivalent so it doesn't matter, you can see the mapping in
>>>eg. the article on TD(lambda) for the KnightCap engine (I think Dann has it on
>>>his ftp).
>>>
>>
>>It is a hyperbolic tangent function, so the probabilities are in
>>the closed interval [-1..1].  So small changes in the "pawn units"
>>evaluation yield a more dramatic shift in the probablility if the
>>score is drawish (vice if you're already losing or winning by a queen).
>
>Yes but the point it you would need to go back and forth anyway. Your material
>score indicates you are down 1 pawn, now you need to do a table lookup to find
>what this corresponds to in probability of a win, it is just a futile
>translation at every node that servers no real purpose AFAICS. Sometimes the
>most natural choice isn't the optimal choice, I think this is one of those
>cases.

I don't disagree - I was just elaborating on the mapping you referred
to.  The mapping is necessary for learning (well, for TD learning at
least).  I'd rather "see" the pawn score, like you.

>
>>>>I expect that a CD (or DVD) full of positional chess patterns drawn out
>>>>of a suitable number and choice of chess games (out of a permanently
>>>>growing number) will have a much greater effect on the play and results
>>>>of a chess knowledge using chess engine than 4 or 5  pieces tablebases
>>>>have nowadays on the results of tablebases using chess engines.
>>>
>>>Patterns are used in chessprograms, connected passed pawns or rook on open file
>>>for instance. Consider that computers are too small and slow to be using a
>>>neural net with 100 billion neurons for the entire postion like humans. Besides
>>>how would you train it? You need scores on every position, which you do not
>>>have.
>>
>>Agreed about the size of the neural net, though that may not be true
>>forever.  I think you should be able to derive a score from game outcomes.
>
>Maybe or maybe not, speed / evaluation accuracy will always be a walk of
>balance, I'm not sure there will ever come a time where you can afford a 100
>fold slowdown go get better positional evaluations, well not in eng-eng matches
>;)

That depends on whether you believe there are diminishing returns.
See "Additional Returns for Additional Search in Chess",
Junghanns, Schaeffer, Brockington, Bjornsson, Marsland.  They
claim diminishing returns do exist in chess, but the effects are
hidden in shallow depths due to poor decision quality.


>
>>>For "probing" a CD you first need to design some clever index scheme and if you
>>>plan to use it at every node it's going to slow your program to a crawl (of
>>>course that wouldn't matter much if it was extremely accurate).
>>
>>True, but ... :)  computer scientists are often not interested in
>>hardware limitations.  Instead they concern themselves with what
>>is theoretically possible.
>
>I think sometimes things take a while to sink in, even with researchers.
>All these ideas may look good on paper, but there are some hefty pratical
>obstacles.

Practical obstacles should never get in the way of ideas.


Consider the resolution on such an evaluator for instance, let's say
>this NN is all the positional evaluation you have (and then some material bean
>counting). Now you search 500 kNps and do maybe 100 kNps full evaluations (just
>to throwing out some numbers). If you search for 10 secs that's 10^6
>evaluations, and you probably have not gone more than ~10-12 plies which means
>most of the positions (except those few heavily extended) are going to be very
>similar and extremely hard for a NN to distinguish, you really need an awsome NN
>and with that kind of power better tricks are probably at your disposal.

I agree, it's a hard problem. :)  I don't agree that it's not worth
some effort, though.

>
>-S.




This page took 0.01 seconds to execute

Last modified: Thu, 07 Jul 11 08:48:38 -0700

Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.