Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Crafty Static Evals 2 questions

Author: Uri Blass

Date: 04:04:40 02/25/04

Go up one level in this thread


On February 25, 2004 at 05:56:16, martin fierz wrote:

>On February 24, 2004 at 11:25:23, Robert Hyatt wrote:
>
>>On February 24, 2004 at 11:17:32, martin fierz wrote:
>>
>>>On February 24, 2004 at 11:06:00, Robert Hyatt wrote:
>>>
>>>>On February 24, 2004 at 10:37:21, martin fierz wrote:
>>>>
>>>>>On February 24, 2004 at 10:19:51, Anthony Cozzie wrote:
>>>>>
>>>>>>On February 24, 2004 at 09:32:08, martin fierz wrote:
>>>>>>
>>>>>>>On February 23, 2004 at 23:05:58, Robert Hyatt wrote:
>>>>>>>
>>>>>>>>On February 23, 2004 at 18:52:36, Geoff Westwood wrote:
>>>>>>>>
>>>>>>>>>Hi
>>>>>>>>>
>>>>>>>>>I was perusing the latest table of results, Crafty's static eval of 2 of the
>>>>>>>>>passed pawn positions were interesting.
>>>>>>>>>
>>>>>>>>>Assuming I havent made a mistake in the cutting and pasting
>>>>>>>>>
>>>>>>>>>Position 1
>>>>>>>>>8/4k3/8/7P/1P6/3p4/4p3/4K3 b - -; id "PP-00004"
>>>>>>>>>
>>>>>>>>>[D]8/4k3/8/7P/1P6/3p4/4p3/4K3 b - -
>>>>>>>>>
>>>>>>>>>Crafty reckons this is +4.8 (good for white). This is rather clever as although
>>>>>>>>>the black king could catch either of the white passed pawns, it cannot stop
>>>>>>>>>both. Also blacks 2 advanced pawns cant do anything as the white king gobbles
>>>>>>>>>them up easily. Only Crafty and Tinker understand this position statically. Any
>>>>>>>>>tips on what the algorithm is to sort this one out ?
>>>>>>>>
>>>>>>>>
>>>>>>>>This is the idea I have reported here before, pointed out (demanded to be fixed
>>>>>>>>in fact) by a GM friend of mine.  The idea is that the two separated pawns are
>>>>>>>>better than the two connected passers.  The king stops the two connected passers
>>>>>>>>easily until the enemy king supports them, meanwhile the split passers walk on
>>>>>>>>in...
>>>>>>>
>>>>>>>i don't like the generality of your statemtent here, but - it is a small price
>>>>>>>to pay if it's right in most cases. which perhaps is the case. anyway, here's my
>>>>>>>question:
>>>>>>>
>>>>>>>what does your static eval say for the black king on e6/e5/e4/e3 ? i wouldn't be
>>>>>>>surprised if it got it wrong in some cases now...
>>>>>>>
>>>>>>>cheers
>>>>>>>  martin
>>>>>>
>>>>>>The question is always "what do you put in the search, what do you put in the
>>>>>>eval" <shrug>.
>>>>>
>>>>>sort of - for me the answer is clear. the point i wanted to make (not the first
>>>>>time, BTW) is that returning huge evaluations in positions like this may not be
>>>>>a good idea because they are *very* sensitive to details like king position.
>>>>>e.g. if i got it right, then it's a white win with the king on e6, but a black
>>>>>win with the king on e5. do you really want to allow your static eval to return
>>>>>a white win when it might be a black win?
>>>>>of course you can say that if you get it right 60% of the time, it is better
>>>>>than returning an equal eval in this kind of position. but wouldn't it be better
>>>>>then to return something like +- 1 so that you never blunder into this when you
>>>>>are e.g. a piece up and see this type of transition?
>>>>>i generally try to return huge evals only when i am very certain that they are
>>>>>correct.
>>>>>
>>>>>cheers
>>>>>  martin
>>>>
>>>>You are looking at this the wrong way.  If you want 100% accuracy, you will die
>>>>from it.  :)
>>>
>>>hehe, i never claimed i wanted 100%. i think your score should reflect the
>>>amount of certainty you have.
>>>
>>>>If you don't do what I do, you will like connected passers, and in 90% of the
>>>>games I will beat you when that comes up.
>>>
>>>nope. your rule "connected passers are strong when there are many pieces,
>>>separated passers when there are few pieces" is good in some cases. but it is
>>>not good in quite a lot of cases IMO, with no disrespect to your GM friend. if i
>>>make a more accurate version of that rule, i will beat you when it comes up :-)
>>>
>>
>>Perhaps you misunderstood my idea.  Connected passers are better when there are
>>any pieces on the board.  But split passers are better when there are none.  I
>>just have a smooth transition from one to the other to avoid yet another problem
>>called "an evaluation discontinuity"..
>
>yes, in this case i misunderstood, but i think my misunderstanding is better
>than your understanding :-)
>connected passers are better with many pieces, split passers are better against
>*some* pieces. and of course when there are no pieces left. split passers which
>are one file apart are worse than connected passers usually. it all depends on a
>lot more than just a one-line-rule...
>
>i don't know whether i should believe the eval discontinuity thing. i know
>somebody recently quoted a paper on this, but it's just a fact: exchanging any
>pieces necessarily changes the evaluation. sometimes not by very much. big
>changes are usually the exchange of the queen, the exchange of the last rook,
>the exchange of the last piece. these eval discontinuities are *real*. i don't
>believe in smoothing them out. perhaps if you write an eval with discontinuities
>it's harder to get it right that everything fits in with each other, and that's
>why it's supposed to be bad?!
>
>
>
>>
>>The king distance is an easy one to fix.  IE my passed pawn race code already
>>handles it correctly, in that a pawn can "run" if the king is close enough to
>>defend the queening square before the opponent's king can get there, as one
>>example of multiple special cases I handle...
>>
>>
>>
>>>> Do you want to be right 100% of the
>>>>cases you recognize, leaving 95% as "unclear and probably lost" or do you want
>>>>to be right in 90% of the total cases?
>>>>
>>>>I choose the latter...
>>>>
>>>>No doubt it can be made more accurate.  But no doubt that without it, it is even
>>>>less accurate...
>>>
>>>that wasn't my point. i assume you are returning a +5 for white even with the
>>>black king on e5 and you can lose games where you are completely winning because
>>>of this.
>>
>>Maybe.  But most likely not.  Wait for my hash collision paper to come out, for
>>some _eye-popping_ information about the overall tree search space and how
>>resistant to errors it really is.
>
>it won't pop *my* eyes. i once reduced hash key sizes in my checkers program
>beyond all sensible settings, because there was a discussion here about whether
>you really need 64-bit keys. in my checkers program, i have 64 bit keys, but
>effectively it's only using about 52 bits. i have about a 20 bit part which is
>used for the hashindex with %, and of the remaining 44 bits i store only 32 as a
>check. i reduced those 32 down to about 8 (!!) bits and in 100 test positions
>only saw one different move played IIRC. ridiculous, i must have lots of
>collisions there. unfortunately, i didn't count the collision number, or write
>down the results - but i know what you're talking about!
>
>> But that aside, remember that I do a decent
>>search as well, _before_ doing static evaluations.  If you put the king at e4,
>>it doesn't take much of a search to see that black wins, and the eval can't hide
>>that because I only evaluate positions reached in the q-search.
>
>right, if you are already in this position. but there was a position about 12-16
>ply before this, and there you just might have chosen the wrong move because of
>this. very, very unlikely, i know. but still...
>
>cheers
>  martin
>
>>> my theory is that you should try to recognize who is winning, but
>>>differentiate between being certain that it's a win, and just guessing it's a
>>>win. there are very many positions where you can be 99% certain, and for these
>>>returning a +5 is ok IMO. for those where it's not so clear, you should guess
>>>the result, but only return +1 or so.
>>>if you allow a +5 for such positions, you can get into this from a position
>>>where you were a piece up, because you simplify to this ending. if you want that
>>>to happen, fine :-)
>>>
>>
>>Note that there is an uncertainty in the evaluation.  Otherwise wouldn't you
>>expect a +9 which is a _real_ queen??? (+8 actually, since the pawn goes and the
>>queen appears).
>
>not necessarily. as a human, i never think in the context of +3 / +5 / +8. in
>such situations i think in the context of 99.9% winning probability or not. my
>static eval in my brain says "this is unclear" to the type of positions
>discussed. if i have a choice of playing this or playing with a piece up and 3
>pawns each for example, i will always choose the piece up position. no margin of
>error there...
>
>cheers
>  martin

I can add that I agree with you in this discussion.
Crafty is of course not perfect and can be improved.

I also do not like this evaluation of Crafty.
I have not complicated evaluation and the main problem is that writing
complicated evaluation without bugs is a problem but I have not big scores for
unstoppable passed pawns unless they are more advanced than all of the opponent
pawns.

Today I even do not evaluate unstoppable passed pawns that are not more advanced
than all of the opponent pawns.

Uri



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.