Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Positional scores in Eval()

Author: José Carlos

Date: 06:15:35 04/11/01

Go up one level in this thread


On April 10, 2001 at 13:57:33, Tony Werten wrote:

>On April 10, 2001 at 09:16:28, José Carlos wrote:
>
>>On April 10, 2001 at 06:58:26, Tony Werten wrote:
>>
>>>On April 09, 2001 at 22:25:14, Robert Hyatt wrote:
>>>
>>>>On April 09, 2001 at 21:17:59, Normand M. Blais wrote:
>>>>
>>>>>On April 09, 2001 at 18:20:07, Robert Hyatt wrote:
>>>>>
>>>>>>On April 09, 2001 at 16:49:21, Normand M. Blais wrote:
>>>>>>
>>>>>>>On April 09, 2001 at 16:21:56, Andrei Fortuna wrote:
>>>>>>>
>>>>>>>
>>>>>>>>positional score > 2 PAWN_VALUE. And that will hurt my quiescence and my
>>>>>>>>futility pruning if I assume that 2*PAWN_VALUE is max positional score. It all
>>>>>>>>boils down to the magnitude of the positional scores versus pawn value, I think
>>>>>>>>I have to choose either to keep big bonuses and turn futility off (or set a
>>>>>>>>bigger margin for futility but in that case it would make futility more
>>>>>>>>inefficient) or keep small bonuses and enjoy the reductions I get from futility
>>>>>>>>and quiescence.
>>>>>>>
>>>>>>>What if you multiply the value of the material by 10 (i.e. pawn = 1000 Knight =
>>>>>>>3000 Bishop = 3000 Rook = 5000 Queen = 10000 ) and keep the positional score as
>>>>>>>it is.
>>>>>>>
>>>>>>>N.M.B.
>>>>>>
>>>>>>
>>>>>>Then you become _very_ materialistic.  You will grab pawns whenever
>>>>>>possible, even if it wrecks your king position.  You will grab a pawn but
>>>>>>leave your opponent with an outside passer that wins.  Etc...
>>>>>
>>>>>I understand that the positional score is related to the material score (the
>>>>>pawn or tempi being the unit of measurement). I made a mistake by suggesting
>>>>>that the material value can be magnified without adjusting both the positional
>>>>>bonuses and penalties. But would it be a good idea to have one more digit to
>>>>>work with? A value of 1000, for instance, could represent 100.0.  And scores
>>>>>like 112.4 and 112.8 could be set apart? Not a sophisticated idea but anyhow.
>>>>>
>>>>>Regards,
>>>>>
>>>>>N.M.B.
>>>>
>>>>I don't personally think the idea of "millipawns" makes a lot of sense.  That
>>>>is _very_ fine resolution.  I used it in Cray Blitz, and in early versions of
>>>>Crafty, but it adds to the hash entry size for one thing...
>>>
>>>XiniX is using it. Not because it makes any sence but I like the idea that I can
>>>give small bonusses for some stupid positional thing that I like, without
>>>messing up the eval and still having the idea that if two positions score almost
>>>the same, it will prefer the one I like.
>>>
>>>Gives me a sort of personal tough feeling. ( As I said, it's not because it
>>>makes any sence )
>>>
>>>cheers,
>>>
>>>Tony
>>
>>  Too fine resolution makes more difficult beta cutoffs, IMO, so, even not
>>having tried it myself, I'd say your tree must be bigger that way than it would
>>be with smaller resolution.
>>  I even recall that, some time ago, I think it was John Dart that made a test
>>reducing the resolution of his eval to 1/4 pawn, but I actually can't remember
>>how did it work.
>
>You could even return -1,1,0 (lose,win or neither ). You would get a lot of beta
>cutoffs, but I don't think it will give you a strong engine. ( maybe a good
>matesolver )

  Sure. So what you need it something around the middle. A good compromise
between eval precision and beta cutoffs.

  José C.

>Tony
>
>>
>>  José C.



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.