Piece/square table challenge

[Steve Maughan]

Interesting!

(...)Adam's set is asymmetric, which raises important point: is vertical symmetry of piece/square tables a requirement?(...)

They are very asymmetrical! A pawn on white pawn on g7 is -16 and b7 is +121!

Steve

[PK]

not only interesting, but also strong! Adam's set scored 59,4% I wonder how much can be gained from further normalization of his values. Anyhow, right now I'm running his set against my set, the next match being initial template against Mikko's set, and then filling the gaps, so that we have a nice little round-robin. of course, new entrants are welcome!

[Steve Maughan]

This is +65 ELO!!

[Lyudmil Tsvetkov]

not only interesting, but also strong! Adam's set scored 59,4% I wonder how much can be gained from further normalization of his values. Anyhow, right now I'm running his set against my set, the next match being initial template against Mikko's set, and then filling the gaps, so that we have a nice little round-robin. of course, new entrants are welcome!

Hi Pawel.

I would like if you could run also my Stockfish values. I hope the format is the same, only thing is in Stockfish middlegame and endgame values are specified in the same place. I do not know how to reformat them, but the 5 tables (without king) are here:

http://www.talkchess.com/forum/viewtopi ... 34&t=50778

From top to bottom are the values for pawn, knight, bishop, rook and queen.

For king table you can use your king table or Adam's king table, for example, I do not mind. The important thing would be to test the piece parameters. Also, Stockfish tables are adjusted to piece values that are 2-2.5 times bigger than normal, I hope this does not have some bad effect.

Would be glad if you could run the test.

[Adam Hair]

I was looking over an email to Miguel last year describing how I derived these tables, and I saw that I forgot how the endgame tables were created. As it turns out, I had noticed that the mined endgame values lacked much of the bilateral asymmetry that the midgame tables have. So, I averaged the two sides and used the averages as new pst values. Then, I found models that best fit the average values for each of the endgame tables, and used those equations to find the final piece/square endgame values.

As for the midgame tables, I tried something similar but the results were not as good.

[PK]

@Adam,

I had a look at Your tables, and it is possible that some of their success can be attributed to betting that opponent has castled short. I am unable to describe all the differences in the mathematical terms, even though I see some patterns. Bishop table is especially funny, as it gives bonuses for queenside squares from which kingside is accessible.

I thought that averaging file scores for knight tables will give some information, but all I see is mess: all I kan say is that Your knight table prefers kingside, but I am totally surprised by immense preference for F anf H files. Overall, it seems that normalizing for assumed king distance would be much harder than I anticipated.

Code: Select all

A file: -42,125  H file: -7,875 (slant:  34,250)
B file:   -4,875  G file:  0,5    (slant:   4,375)
C file:    5,875  F file: 34,635 (slant: 28,760)
D file:  15,75    E file: 23,125 (slant:  7,375)

[PK]

@Lyudmil,

I'd like to include Your tables in the competition, but please type Your values in the template I provided and fill in the gaps (piece values, king tables). For the record: exact divisor for Stockfish pst values is 2,56.

[Steve Maughan]

Hi Adam,

Do you have a description of how you derived the tables? I'd love to understand more.

Steve

[Adam Hair]

Hi Adam,

Do you have a description of how you derived the tables? I'd love to understand more.

Steve

Hi Steve,

There is a description in my first post in this thread. I can give a more detailed and slightly corrected description this evening (I believe that you and I live in the same time zone).

Adam

[Adam Hair]

@Adam,

I had a look at Your tables, and it is possible that some of their success can be attributed to betting that opponent has castled short. I am unable to describe all the differences in the mathematical terms, even though I see some patterns. Bishop table is especially funny, as it gives bonuses for queenside squares from which kingside is accessible.

I thought that averaging file scores for knight tables will give some information, but all I see is mess: all I kan say is that Your knight table prefers kingside, but I am totally surprised by immense preference for F anf H files. Overall, it seems that normalizing for assumed king distance would be much harder than I anticipated.

Code: Select all

A file: -42,125  H file: -7,875 (slant:  34,250)
B file:   -4,875  G file:  0,5    (slant:   4,375)
C file:    5,875  F file: 34,635 (slant: 28,760)
D file:  15,75    E file: 23,125 (slant:  7,375)

Hi Pawel,

I see patterns also, but I am hampered by being a weak chess player. Perhaps I can work with the data to extract additional information.

By the way, how did you create your tables?

Adam