TalkChess.com

That TCEC will not even allow data created by Stockfish to be used to train another net is the height of ignorance, and I would say even offensive.
Suppose that data is created by independent testers like CCRL and CEGT? Does that data still belong exclusively to the Stockfish team?
And if you don't want someone to use your data, then don't publish it.

I don't like the direction of the arrow that computer chess is pointed in.

Dann Corbit wrote: ↑Sat Oct 10, 2020 6:38 pm That TCEC will not even allow data created by Stockfish to be used to train another net is the height of ignorance, and I would say even offensive.
Suppose that data is created by independent testers like CCRL and CEGT? Does that data still belong exclusively to the Stockfish team?
And if you don't want someone to use your data, then don't publish it.

I don't like the direction of the arrow that computer chess is pointed in.

What I do not understand is why TCEC still allows using data created by Leela to train a net (AllieStein) but doesn't allow data created by Stockfish to train a net (NNUE). Either allow both or forbid both. Double standards by TCEC, and unfair for NNUE based engines.

Also, what makes training nets different from tuning? Should TCEC ban engines like Ethereal that once used Stockfish data in the past to tune the evaluation? Does training with TCEC data count as Stockfish data, seeing as how Fire and Houdini are both Stockfish clones, and there are Stockfish - Fire games and Stockfish - Houdini games in TCEC?

Not releated to TCEC rules but we should not forget regular AB engines with hand-crafted evaluations too. They have been tuning their evals from datasets scored by stockfish such as Zurichess's set. Even using raw CCRL/CEGT data for training is basically learning from other engines. How about learning from millionbase (human) games which I assume is Ok ... I guess you have to learn from weak humans or weak engines (whole CCRL set unfiltered) for it to be Ok. This bogles my mind.

If Stockfish self-played a game that also happened to be played by two grandmasters or correspondence chess players, does that game count as Stockfish data or not?

Keeping "personalities" of our engines through their standard evaluations used as input for building nets seems important (to me at least).
Of course many parts of our engine are inspired, sometimes nearly copy/pasted from other, and often from Stockfish, but at the end they all play different and that counts. And Texel-tuned engines are often optimized to game outcome (whatever engine played the game), not the score itself.

By the way, I tried already to build a net "mixing" engines evaluation (from CCRL, FGRL and CEGT games) but this doesn't seem to work very well.

xr_a_y wrote: ↑Sat Oct 10, 2020 7:50 pm And Texel-tuned engines are often optimized to game outcome (whatever engine played the game), not the score itself.

This doesn't matter much. Infact training against game outcome (Z) is better than score (Q) because Z can potentially incorporate
long-term effects. A0 used Z entirely but you can also mix in say 10% of Q to Z for faster learning.

Madeleine Birchfield wrote: ↑Sat Oct 10, 2020 7:26 pm If Stockfish self-played a game that also happened to be played by two grandmasters or correspondence chess players, does that game count as Stockfish data or not?

Computer chess continually fantasising up nonsense constructs. There is no Stockfish data, data is free for everybody, chess engines do not own their output, they doubly do not own it because non-human entities can’t own anything. Triply nobody owns chess game data. Zero sum game data is not cooperatively generated and copyright can’t apply.

chrisw wrote: ↑Sat Oct 10, 2020 7:59 pm Computer chess continually fantasising up nonsense constructs. There is no Stockfish data, data is free for everybody, chess engines do not own their output, they doubly do not own it because non-human entities can’t own anything. Triply nobody owns chess game data. Zero sum game data is not cooperatively generated and copyright can’t apply.

This is primarily a construct of TCEC.

Madeleine Birchfield wrote: ↑Sat Oct 10, 2020 8:18 pm

chrisw wrote: ↑Sat Oct 10, 2020 7:59 pm Computer chess continually fantasising up nonsense constructs. There is no Stockfish data, data is free for everybody, chess engines do not own their output, they doubly do not own it because non-human entities can’t own anything. Triply nobody owns chess game data. Zero sum game data is not cooperatively generated and copyright can’t apply.

This is primarily a construct of TCEC.

Secondarily.

Primarily it’s a construct of factions with the National Union of chess engine programmers and some others who threaten TCEC and then TCEC bending over to accommodate them.

Daniel Shawul wrote: ↑Sat Oct 10, 2020 7:56 pm

xr_a_y wrote: ↑Sat Oct 10, 2020 7:50 pm And Texel-tuned engines are often optimized to game outcome (whatever engine played the game), not the score itself.

This doesn't matter much. Infact training against game outcome (Z) is better than score (Q) because Z can potentially incorporate
long-term effects. A0 used Z entirely but you can also mix in say 10% of Q to Z for faster learning.

Yes that was my point. No need for SF score in tuning data for Texel tuning method.
In fact quiet-labeled.edp and violent.edp from Zurichess doesn't have SF score inside ... just game outcome...
And those games were played with Zurichess ... not Stockfish (http://talkchess.com/forum3/viewtopic.p ... dp#p686204)
So I don't get your statement

Daniel Shawul wrote: ↑Sat Oct 10, 2020 7:19 pm They have been tuning their evals from datasets scored by stockfish such as Zurichess's set.