Author: Uri Blass
Date: 15:03:36 12/29/01
Go up one level in this thread
On December 29, 2001 at 16:58:49, Tom Kerrigan wrote:
>On December 29, 2001 at 06:48:50, Uri Blass wrote:
>
>>I believe that even if the evaluation is only piece square table evaluation
>>there are tactical tricks that happen mainly in some situation(for example white
>>knight at a8 is often bad).
>
>Of course it's often bad. Everybody knows that already. That's why there's no
>program on Earth that has a knight piece/square table where a8 is higher than
>e5.
> What you are proposing is to decrease that value to account for major,
>short-term tactical issues that may or may not exist. I don't think this is a
>good idea, but you don't have to take my word for it. Download Crafty and try it
>out.
Positions with knight at a8 are not common enough to get a conclusion
by games but I guess that a knight at a8 should be evaluated as something like 1
pawn less than knight at e5 and not something like 0.2 pawns less than a knight
at e5.
Testing by games if Crafty earns 1 elo point from changing
the score for knights takes too much time to do and I also do not know if Crafty
does not punish knight at a8 by mobility considerations.
If I understand Crafty correctly white knight at e5 is considered as only
0.12 pawns more than a knight at a8 but I see no comments to tell me if the
array nval_w is really the piece square table for knights
I also guess that the first squares in the array are a8-h8 because I know that
programs like to hate b1 and g1 for the white knights in order to develop them
and if b1 and g1 are the worst they must be -16
Here is the relevant array from Data.c of Crafty
signed char nval_w[64] = { -8, -8, -8, -8, -8, -8, -8, -8,
-8, -8, 0, 0, 0, 0, -8, -8,
-8, 2, 2, 2, 2, 2, 2, -8,
-8, 2, 4, 4, 4, 4, 2, -8,
-8, 2, 6, 6, 6, 6, 2, -8,
-8, 2, 6, 6, 6, 6, 2, -8,
-16,-16, 4, 4, 4, 4,-16,-16,
-16,-16, -8, -8, -8, -8,-16,-16};
>
>>I also said that we should not use pawns to evaluate how much a guess is wrong
>>and there is no big difference if the guess is wrong by 5 pawns or by 6 pawns.
>
>I have no idea what you're talking about.
I meant to the following from this dicussion:
"You can do what the DT guys did--make a test suite from GM games, search the
positions deeply, and discard the ones where the deep score didn't agree with
the static eval by a certain margin (indicating that something tactical
happened). Optimizing against these deep scores has the same local maxima
problem that you were all worried about a few posts ago, though."
I understood from it that the DT guys tried to have an evaluation that
guess correctly the score after a deep search only when there is no big
difference in the score.
If the deep scores are scores of expected result and not score in pawns and
if there are enough positions for the relative weights that you think to change
then I believe that the test of DT may be more productive if you do not ignore
cases when the difference is big.
If there are not enough positions for some weights than it may be better not to
test changing them because you cannot get something that you can trust even if
there is no tactics.
>
>>Note that simple calculation say that you may need 100000 games to check if a
>>small change is really an improvement of 1 elo and analyzing 100000 positions
>>when every position analyzed for few minutes is clearly faster than playing
>>100000 games.
>
>But analyzing the positions will not necessarily tell you if you gained or lost
>an ELO point, that's the problem.
It may tell you if your program does more mistakes or less mistakes in these
position and you can hope that doing less mistakes is going to help you
to earn elo points.
Here I mean to the test of avoiding bad moves that is different from the test of
the DT guys.
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.