PSQ tables depending on king sides, pawn patterns etc.

[Henk]

Why you don't know? Why not just test it on a huge set of games?

I don't want my notebook being busy all day.

[Sergei S. Markoff]

I've post answer here: http://www.talkchess.com/forum/viewtopi ... 6b21870ff7

[Daniel Shawul]

Sergei,
Thanks for taking the time to explain the procedure. I was not quite sure what you were trying to achieve with ANOVA because my use of it was for something different from yours. I see now that you are mining eval terms with the most benefit to add to your engine, so in a way programming eval automatically. I don't know if that will be more efficient than manual tuning with a human master, but it is indeed an interesting approach. In the past I have tried to automatically extract rules for endgames only using Quinlan's algorithm for the purpose of reducing sizes of TBs, but I wasn't much successful with it.
In any case, my thought was to use PSQT as a rough evaluation term (mostly mobility) without much thinking. So PSQ or not, you can still use your method to improve eval().
Good luck
Daniel

[Laszlo Gaspar]

Hi Sergei,

I did exactly the same! I constructed 6*6=36 different PSQT and other tables, each one is a case of the kings' relative position (queenside, center and can castle kingside, center and can castle queenside, center and can castle both sides, center and cannot castle, kingside - and the combinations for both king). Alltogether 73000 parameters, tuned automatically by playing 5000 games against a strong opponent. It might be luck but after a few trial I got a quite reasonable play. I'm still using my old data file.
I recently started thinking of the level of the contribution of my tables to the strength because some major terms are not part of the tables (piece values, passer and candidate pawn values) and without them I loose 200 Elo. This is quite high value considering that ProDeo for instance looses "only" 70 elo without PSQT (even if I have other than the PSQ parameters in it). So this idea has some merit it seems but I might overlook something and have to check my code once again to be sure that I didn't make a false analyzis.

[bob]

Robert, Pawel, it's not about game-time processing. It's about to have not just opening/endgame PSQs, but also some additional tables to correct PSQ scores, for example:

1. Tables for 1) white king kingside and black king kingside, 2) white king kingside, black king queenside, 3) white king queenside, black king kingside, 4) both kings queenside or maybe another way — tables to correct PSQs depending on own king at H-file and own king A-file with scaling depending on file and same tables to correct PSQ depending on opponent king file.
2. Tables with PSQ ajustments for every open files formation (8 bit pattern); you're really don't need 256 tables, you can join most similar tables using cluster analysis (ANOVA) and also you're able to join values inside tables using some polynomials — there are a lot of ways how to do it — but the general approach is to produce eval terms using statistical methods to reach not only near-to-optimal values of terms but also a near-to-optimal set of terms themselves.

A PST is a "general-purpose" or "general rule-of-thumb" method of scoring, dealing with things like "A knight on the rim is dim" and such rules that really apply to most (but not all) positions. I think that beyond general concepts such as centralization, PSTs don't apply, it is better to use game-specific code to recognize things that are not "generally applicable everywhere."

[Rebel]

Hi Sergei,

I did exactly the same! I constructed 6*6=36 different PSQT and other tables, each one is a case of the kings' relative position (queenside, center and can castle kingside, center and can castle queenside, center and can castle both sides, center and cannot castle, kingside - and the combinations for both king). Alltogether 73000 parameters, tuned automatically by playing 5000 games against a strong opponent. It might be luck but after a few trial I got a quite reasonable play. I'm still using my old data file.
I recently started thinking of the level of the contribution of my tables to the strength because some major terms are not part of the tables (piece values, passer and candidate pawn values) and without them I loose 200 Elo. This is quite high value considering that ProDeo for instance looses "only" 70 elo without PSQT (even if I have other than the PSQ parameters in it). So this idea has some merit it seems but I might overlook something and have to check my code once again to be sure that I didn't make a false analyzis.

It's better to have no PST's at all than faulty PST's.