Piece/square table challenge

[Rebel]

I played a while with several posted PST's and especially the Adam Hair PST's are a lot of fun, lot's of instant solved tactical shots but in eng-eng it does bad like some of the others I tried. The latter is no surprise, piece values + PST's need to be in balance with eval.

[Uri Blass]

Results for today:

Code: Select all

    Program                          Elo    +   -   Games   Score   Av.Op.  Draws

  1 Rodent 1.3 (build adam)        : 2081   11  11  3642    64.9 %   1974   18.8 %
  2 Rodent 1.3 (build pawel)       : 2060   10  10  3641    61.4 %   1979   22.3 %
  3 Rodent 1.3 (build lyudmil)     : 2016   13  13  2060    46.9 %   2038   22.4 %
  4 Rodent 1.3 (build mikko+)      : 2006   12  12  2582    50.4 %   2003   21.4 %
  5 Rodent 1.3 (build ufo)         : 2000   10  10  3524    51.8 %   1987   24.5 %
  6 Rodent 1.3 (build lucas)       : 1965   13  13  2026    42.1 %   2021   24.7 %
  7 Rodent 1.3 (build ERROR)       : 1935   14  14  2036    36.9 %   2028   22.0 %
  8 Rodent 1.3 (build daniel)      : 1836   15  15  2063    23.7 %   2040   17.5 %

Lyudmil's set comes third (and the best among "general use" tables). It is worth nothing that it scored worse against Mikko's set, but relatively better against "doctored" sets).

More news tomorrow. Enjoy!

I totally dislike adam's set because it is not symmetric to files and it can be interesting if you can test a symmetric version of his set by simply replacing every 2 numbers by the average of them so if knight at b3 and knight at g3 have different values you replace the values simply by the average.

I wonder if this version cannot score even better because I doubt the idea of attacking the opponent king when you even do not know where it is can be good.

In most cases the king go to the king side with symmetric evaluation and still attacking in that side may change the decision of the opponent in part of the cases because the opponent can see for tactical reasons in some cases that it is better not to go to the king side.

I also think that attacking in the queen side can also be effective even the opponent castle short because there are other targets to attack in a chess game and not only the opponent king.

[Evert]

I totally dislike adam's set because it is not symmetric to files and it can be interesting if you can test a symmetric version of his set by simply replacing every 2 numbers by the average of them so if knight at b3 and knight at g3 have different values you replace the values simply by the average.

I wonder if this version cannot score even better because I doubt the idea of attacking the opponent king when you even do not know where it is can be good.

In most cases the king go to the king side with symmetric evaluation and still attacking in that side may change the decision of the opponent in part of the cases because the opponent can see for tactical reasons in some cases that it is better not to go to the king side.

I also think that attacking in the queen side can also be effective even the opponent castle short because there are other targets to attack in a chess game and not only the opponent king.

I agree. As a general PST it's probably not so useful.

What I posted yesterday (fitting the table with a linear combination of symmetric tables and an asymmetric "tropism" table) is my first attempt at making something a bit more general based on Adam's table. The fit isn't very good, but that may not be so important for a number of reasons (not knowing how much noise there is on each value is one of them). The main point is to replicate the idea with a few tables and some weights.

Eventually, I plan to try the following: I identify interesting "targets", like the enemy king, a weak pawn, a strong square. Each of these is assigned a weight (to be tuned) and the product of the weight and the tropism value gets added to the PST score. The weights could be game-phase dependent (attacking the king may be more important in the middle-game than the end-game, for instance). It'll be a while before I really get around to doing that though...

[Steve Maughan]

Hi Evert,

(...)
with coefficients -1.26987, 9.64933, -8.48597, 4.45139 and 0.0887156. (...)

I'm not sure negative coefficients make sense in this context. A negative coefficient is saying the opposite variation in score is appropriate - which I don't think makes sense.

May I ask:

- What were the t-stats on each coefficient?
- Did you have a constants in the model (which I think you need) or did you force through origin

Steve

[Evert]

I'm not sure negative coefficients make sense in this context. A negative coefficient is saying the opposite variation in score is appropriate - which I don't think makes sense.

It does, but it's a bit ugly.

Let's look at the third coefficient first, which multiplies the tropism table. The sign is due to the way the tropism table works: it records the number of steps required to reach a particular square. That means the value is high for squares that are far away from the target - but this is the opposite of what you want, because you really want squares that are "close" to give you the largest bonus. So really, the table should be transformed into (max_value_in_table - table), but then the first of these is a constant that can be taken out and you end up just subtracting the table. So in a sense, the only value that really makes sense is a negative value (otherwise a piece is better placed if it is far away from the king; could be of course, but if that's what we find we can scrap the idea).

The negative value for the first coefficient looks more suspicious (because at face-value it says that it's better for the knight to be way from the centre), but you have to remember that the fourth table (which is based on mobility scores) is also a centralisation table and it's really the combination of these two that matters. The relative difference between coefficient 1 and 4 more reflects what exactly the centralisation PST should look like than anything else.
If you look at the final table you'll see that it still favours centralisation.

- What were the t-stats on each coefficient?
- Did you have a constants in the model (which I think you need) or did you force through origin

I actually don't have any real data on the quality of the fit, since it was very trial-and-error and ad-hoc. I do have the (unreduced) chi-squared, which is something like 36000 (ie, poor). I don't think getting a very good fit is all that important, actually, because in the games that the empirical table is based on the king is not glued to g8, but gets to wander around and there may be other targets that affect where you prefer to place your pieces. The main thing I take away from it at this stage is the ratio of the asymmetric part to the symmetric part (~8.4/4.4) which tells me something about how important it is to get the piece close to the enemy king, relative to the centralisation. Actually, the advancement table also looks like it is more important than I thought (hoped) it would be, which annoys me a little.

[Steve Maughan]

Evert,

OK - certainly for tropism the negative coefficient is fine.

This should be a straightforward linear regression. If I have time I'll do it.

Steve

[PK]

Code: Select all

    Program                          Elo    +   -   Games   Score   Av.Op.  Draws

  1 Rodent 1.3 (build adam)        : 2082   11  11  3642    64.9 %   1975   18.8 %
  2 Rodent 1.3 (build pawel)       : 2060   10  10  3641    61.4 %   1979   22.3 %
  3 Rodent 1.3 (build lyudmil)     : 2021   12  12  2560    48.6 %   2031   21.8 %
  4 Rodent 1.3 (build mikko+)      : 2005   11  11  3082    50.1 %   2004   22.1 %
  5 Rodent 1.3 (build pio)         : 2001   15  15  1545    48.7 %   2010   23.9 %
  6 Rodent 1.3 (build ufo)         : 2000    9   9  4069    51.6 %   1990   24.9 %
  7 Rodent 1.3 (build lucas)       : 1966   13  13  2026    42.1 %   2021   24.7 %
  8 Rodent 1.3 (build ERROR)       : 1935   14  14  2036    36.9 %   2029   22.0 %
  9 Rodent 1.3 (build daniel)      : 1836   15  15  2063    23.7 %   2039   17.5 %

Today I had a busy day, so I give only some preliminary results for Pio's version - not bad at all.

[Sergei S. Markoff]

I totally dislike adam's set because it is not symmetric to files and it can be interesting if you can test a symmetric version of his set by simply replacing every 2 numbers by the average of them so if knight at b3 and knight at g3 have different values you replace the values simply by the average.

I wonder if this version cannot score even better because I doubt the idea of attacking the opponent king when you even do not know where it is can be good.

In most cases the king go to the king side with symmetric evaluation and still attacking in that side may change the decision of the opponent in part of the cases because the opponent can see for tactical reasons in some cases that it is better not to go to the king side.

I also think that attacking in the queen side can also be effective even the opponent castle short because there are other targets to attack in a chess game and not only the opponent king.

I think it brings up again an idea to use, at least, king sides flags as additional dimension. See: http://www.talkchess.com/forum/viewtopic.php?t=50294

[WinPooh]

Here are GreKo values. They are very simple. Doubt they can provide any sort of strong play, but nevertheless...

Pawn (mg and eg):

Code: Select all

  0,   0,   0,   0,   0,   0,   0,   0,
  0,   0,   0,   0,   0,   0,   0,   0,
  0,   0,   0,  10,  10,   0,   0,   0,
  0,   0,   5,  10,  10,   5,   0,   0,
  0,   0,   5,  10,  10,   5,   0,   0,
  0,   0,   5,   5,   5,   5,   0,   0,
  0,   0,   0,   0,   0,   0,   0,   0,
  0,   0,   0,   0,   0,   0,   0,   0

Knight (mg and eg):

Code: Select all

-10,  -6,  -3,   0,   0,  -3,  -6, -10,
 -6,  -3,   0,   3,   3,   0,  -3,  -6,
 -3,   0,   3,   6,   6,   3,   0,  -3,
  0,   3,   6,  10,  10,   6,   3,   0,
  0,   3,   6,  10,  10,   6,   3,   0,
 -3,   0,   3,   6,   6,   3,   0,  -3,
 -6,  -3,   0,   3,   3,   0,  -3,  -6,
-10,  -6,  -3,   0,   0,  -3,  -6, -10

Bishop (mg and eg):

Code: Select all

-10,  -6,  -3,   0,   0,  -3,  -6, -10,
 -6,  -3,   0,   3,   3,   0,  -3,  -6,
 -3,   0,   3,   6,   6,   3,   0,  -3,
  0,   3,   6,  10,  10,   6,   3,   0,
  0,   3,   6,  10,  10,   6,   3,   0,
 -3,   0,   3,   6,   6,   3,   0,  -3,
 -6,  -3,   0,   3,   3,   0,  -3,  -6,
-10,  -6,  -3,   0,   0,  -3,  -6, -10

Rook (mg and eg):

Code: Select all

  0,  0,  0,  0,  0,  0,  0,  0,
 20, 20, 20, 20, 20, 20, 20, 20,
  0,  0,  0,  0,  0,  0,  0,  0,
  0,  0,  0,  0,  0,  0,  0,  0,
  0,  0,  0,  0,  0,  0,  0,  0,
  0,  0,  0,  0,  0,  0,  0,  0,
  0,  0,  0,  0,  0,  0,  0,  0,
  0,  0,  0,  0,  0,  0,  0,  0

King (mg):

Code: Select all

-40, -40, -40, -40, -40, -40, -40, -40,
-40, -40, -40, -40, -40, -40, -40, -40,
-40, -40, -40, -40, -40, -40, -40, -40,
-40, -40, -40, -40, -40, -40, -40, -40,
-40, -40, -40, -40, -40, -40, -40, -40,
-40, -40, -40, -40, -40, -40, -40, -40,
-20, -20, -20, -20, -20, -20, -20, -20,
-10,   0,   0, -10,   0, -10,   0, -10

King (eg):

Code: Select all

-40, -26, -13,   0,   0, -13, -26, -40,
-26, -13,   0,  13,  13,   0, -13, -26,
-13,   0,  13,  26,  26,  13,   0, -13,
  0,  13,  26,  40,  40,  26,  13,   0,
  0,  13,  26,  40,  40,  26,  13,   0,
-13,   0,  13,  26,  26,  13,   0, -13,
-26, -13,   0,  13,  13,   0, -13, -26,
-40, -26, -13,   0,   0, -13, -26, -40

[shahil4242]

Can a chess engine with pesto evaluation beat 2000+ player?