Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Evaluation Accuracy

Author: Ricardo Gibert

Date: 10:09:26 11/19/00

Go up one level in this thread


On November 19, 2000 at 12:31:20, Daniel Kang wrote:

>On November 19, 2000 at 02:30:21, Ricardo Gibert wrote:
>
>>On November 19, 2000 at 01:34:14, Uri Blass wrote:
>>
>>>On November 19, 2000 at 01:15:22, Ricardo Gibert wrote:
>>>
>>>>On November 18, 2000 at 22:13:35, Peter Kappler wrote:
>>>>
>>>>>On November 18, 2000 at 21:23:54, Ricardo Gibert wrote:
>>>>>
>>>>>>On November 18, 2000 at 12:37:20, Amir Ban wrote:
>>>>>>
>>>>>>>On November 18, 2000 at 06:03:39, Graham Laight wrote:
>>>>>>>
>>>>>>>>On November 17, 2000 at 19:24:23, Amir Ban wrote:
>>>>>>>>
>>>>>>>>>
>>>>>>>>>If your criterion of knowledge is based on accuracy of evaluation then I
>>>>>>>>>respectfully apply for membership in the exclusive "knowledge based" club (and
>>>>>>>>>IMO some members don't belong there).
>>>>>>>>>
>>>>>>>>>BTW, accuracy of evaluation is the best criterion for being knowledgable that
>>>>>>>>>I'm aware of. I've posted here in the past that, to start with, we don't have a
>>>>>>>>>real definition of what good evaluation means. This is the focus of my work with
>>>>>>>>>Junior for more than a year.
>>>>>>>>
>>>>>>>>IMHO, a truly accurate evaluation of a position would yield one of the following
>>>>>>>>3 ordinal values:
>>>>>>>>
>>>>>>>>Win
>>>>>>>>Draw
>>>>>>>>Lose
>>>>>>>>
>>>>>>>>-g
>>>>>>>>
>>>>>>>>>Amir
>>>>>>>
>>>>>>>I can easily fake evaluation that gives only those values. I suppose that you
>>>>>>>mean that the values should be true values. How do you propose to do that ? If I
>>>>>>>have an eval that gives absolutely correct values 60% of the time (and the rest
>>>>>>>wrong), do you expect my program to be weak or strong ? If I get 70% right, am I
>>>>>>>necessarily stronger ?
>>>>>>>
>>>>>>>The question is, given two evaluation functions, to decide which is more
>>>>>>>accurate.
>>>>>>>
>>>>>>>This is a good question. Your answer does not seem to lead anywhere.
>>>>>>>
>>>>>>>Amir
>>>>>>
>>>>>>With 100% correct evaluations of just win, lose or draw, can a program mate in K
>>>>>>+ R vs K? I think it will just wander around unless mate happens to fall within
>>>>>>the program search horizon. Yes?
>>>>>
>>>>>Yep, it would wander around until it lucked into a mate or until the "threat" of
>>>>>a draw by the 50-move rule forced it to play a mating line.
>>>>>
>>>>>--Peter
>>>>
>>>>The 50 move rule may or may not force it to play a mating line. Example:
>>>>
>>>>Lets say the program has played 40 moves without pawn move or capture and is
>>>>able to search only 20 ply. At that point, it may find that a draw due to the
>>>>move rule is a problem, but may not be able to anything about it, since the
>>>>position may actually require more than 10 moves (20 ply) to mate.
>>>
>>>If the program has accurate evaluation the 50 move rule is not relevant because
>>>it will never go to a position that is drawn by the 50 move rule because the
>>>evaluation will not let it to do it because it is going to tell it that it is a
>>>draw(the same position with different history of the game should be evaluated as
>>>a win but accurate evaluation should consider also the history of the game).
>>>
>>>If the program has accurate evaluation of draw,win,loss one ply search is enough
>>>to win won positions.
>>>
>>>Uri
>>
>>It would be a pretty amazing eval that detects 50-move draws in the eval rather
>>than in the search. I think the normal assumption is that it is detected in the
>>search. I think that is quite clear that is the operative assumption in this
>>thread.
>
>Right. But we are talking about a *perfect* evaluation function. If the
>evaluation function can only distinguish between the three ultimate outcomes,
>search (beyond one ply) becomes completely redundant. I guess it depends on how
>you see it but I'd expect a perfect eval to be able to tell if a position is
>drawn or not.
>
>Dan.

Well, the trouble is, what we are discussing here is a three-value-eval
consisting of win, lose, or draw. I don't see how you can do it in eval rather
than in search without at least *implicitly* using Mate-in-N information. This
would represent the use of a much finer eval than a three-value-eval.

The other problem is, to be consistent, you must also incorporate repetition
detection in the eval rather than in the search. This is really stretching
things.

Your interpretation is possible, but I don't see its "real world" value. My
interpretation is actually doable in a limited domain such as K+R vs K. I'm just
exploring the "real-world" consequences of using a three-value-eval and in that
light, I think I am making the more reasonable assumption. Yes? No? Maybe?



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.