TalkChess.com

Are there any engines above, say, 2700 that have eschewed the use of piece-square tables?

Edit: Excluding neural nets.

I don't know how necessary they are. In Winter I use them for the positions of the King and Knight, but no other pieces. Mine are completely symmetric in terms of rotations and mirroring, which means mine probably encode less information than most PSTs.

I have full 8x8 psqt for every piece (well, 8x6 for pawn). Most times when I introduce a new evaluation feature, the psqt of the corresponding piece(s) needs to be retuned to make the new feature work. So there seems a lot of "value" in the psqt.

./Andreas

zenpawn wrote: ↑Sat Feb 16, 2019 3:39 pm Are there any engines above, say, 2700 that have eschewed the use of piece-square tables?

Edit: Excluding neural nets.

If Stockfish removed them, they will probably need to add some positional parameters and retune everything to compensate. I suppose that it will become a little stronger, as dynamic stuff is better than static one, as it can be adapted to the current position. A lot of work in any case.

cdani wrote: ↑Mon Feb 18, 2019 8:41 pm

zenpawn wrote: ↑Sat Feb 16, 2019 3:39 pm Are there any engines above, say, 2700 that have eschewed the use of piece-square tables?

Edit: Excluding neural nets.

If Stockfish removed them, they will probably need to add some positional parameters and retune everything to compensate. I suppose that it will become a little stronger, as dynamic stuff is better than static one, as it can be adapted to the current position. A lot of work in any case.

I see no reason for the assumption that dynamic stuff is better than combination of static stuff and dynamic stuff that stockfish has.

Uri Blass wrote: ↑Tue Feb 19, 2019 9:23 am I see no reason for the assumption that dynamic stuff is better than combination of static stuff and dynamic stuff that stockfish has.

Personal experience tuning everything on hand.

Hi Erin! I made an experiment on my engine, removing the PSQT for each piece on 5k games (This is the strength of my engine psqt):

https://github.com/ratosh/pirarucu/projects/14

What i noticed from my experiment:

- Knight psqt seems weak, i need to improve it;
- Pawn psqt seems too strong, i need to transfer it to pawn structure evaluation;

You can also see a similar experiment on stockfish https://onedrive.live.com/view.aspx?res ... 9qAPiXN5z8