Komodo 14.1 Release and Dragon Announcement

[Madeleine Birchfield]

Good side : Leela is going to get kicked out of the next TCEC SuFi.
Bad side : training on Komodo data won't prevent Komodo Dragon from losing a lot of K's originality, for the most part becoming "Stockfish but worse".

On the other hand Stockfish's NNUE training code and architecture are suboptimal, according to our resident Andrew Grant, so Komodo Dragon could end up becoming "Stockfish but better".

[AndrewGrant]

Good side : Leela is going to get kicked out of the next TCEC SuFi.
Bad side : training on Komodo data won't prevent Komodo Dragon from losing a lot of K's originality, for the most part becoming "Stockfish but worse".

On the other hand Stockfish's NNUE training code and architecture are suboptimal, according to our resident Andrew Grant, so Komodo Dragon could end up becoming "Stockfish but better".

Well, suboptimal in the sense that the mirroring is incorrect for chess. But the fact that its not been fixed/replaced yet, implies its of little consequence. But as a whole I'm not impressed by the SF Trainer, aside from the Loss computations.

EDIT: Or maybe it has been replaced? I have no clue, I don't follow along. I just know some of the CFish code from when I hacked NNUE into Ethereal to prove a point a few weeks back.

[Ozymandias]

Good side : Leela is going to get kicked out of the next TCEC SuFi.
Bad side : training on Komodo data won't prevent Komodo Dragon from losing a lot of K's originality, for the most part becoming "Stockfish but worse".

Considering that "The search is Komodo search (with some parameters tuned), and the nets we use are all trained on Komodo games and Komodo evals.", I don't see why that would be so.

[Madeleine Birchfield]

Well, suboptimal in the sense that the mirroring is incorrect for chess. But the fact that its not been fixed/replaced yet, implies its of little consequence. But as a whole I'm not impressed by the SF Trainer, aside from the Loss computations.

Or it could mean that replacing rotate with mirror in Stockfish's NNUE means that they would have to train a completely different net from scratch, and their terrible neural network trainer and the departure of Sergio Vieri means that none of them were able to train a net that is better than the current SV based nets. The Stockfish developers have resorted to SPSA tuning a small part of their current net because they couldn't figure out how to improve their trainer.

[Madeleine Birchfield]

Someone told me the other day that you said in TCEC's Twitch chat something along the lines of "There is no NNUE training code that does not use SF's qsearch". Is this something you said, or am I hearing it second hand incorrectly? Can you elaborate. Are the networks you trained using SF's NNUE Trainer, and thus SF's qsearch?

If Twitch chat is correct, I think dkappe made that statement before Seer and Halogen 7 came out, because I highly doubt either engine used Stockfish's q search in their training process.

[AndrewGrant]

Someone told me the other day that you said in TCEC's Twitch chat something along the lines of "There is no NNUE training code that does not use SF's qsearch". Is this something you said, or am I hearing it second hand incorrectly? Can you elaborate. Are the networks you trained using SF's NNUE Trainer, and thus SF's qsearch?

If Twitch chat is correct, I think dkappe made that statement before Seer and Halogen 7 came out, because I highly doubt either engine used Stockfish's q search in their training process.

Well, Halogen's original networks are trained in Pytorch. Halogen's current network is trained by a private NN/NNUE trainer I wrote; but Halogen is NOT using the NNUE King-Piece, Shogi Style, whatever you want to call it, framework. All Halogen networks are trained on game results I believe, not evaluations, so there is no SF involved at any step.

I am not familiar with what Seer is doing, although I have talked to the author a fair amount.

[Madeleine Birchfield]

EDIT: Or maybe it has been replaced? I have no clue, I don't follow along. I just know some of the CFish code from when I hacked NNUE into Ethereal to prove a point a few weeks back.

This existing open issue in the official Stockfish github repository would indicate no, not yet.

https://github.com/official-stockfish/S ... ssues/3021

[dkappe]

I think Minic maybe swapped in it’s own qsearch, otherwise I’m not aware of any net author other than myself that has replaced the qsearch. I can’t speak to dragon, but I replaced the qsearch for training night nurse with a very simple one. You’re not looking for the best qsearch for playing, just quiet positions for training.

[AndrewGrant]

I think Minic maybe swapped in it’s own qsearch, otherwise I’m not aware of any net author other than myself that has replaced the qsearch. I can’t speak to dragon, but I replaced the qsearch for training night nurse with a very simple one. You’re not looking for the best qsearch for playing, just quiet positions for training.

So, it is the same training code then, except you've cut out the SF bits?

[dkappe]

I think Minic maybe swapped in it’s own qsearch, otherwise I’m not aware of any net author other than myself that has replaced the qsearch. I can’t speak to dragon, but I replaced the qsearch for training night nurse with a very simple one. You’re not looking for the best qsearch for playing, just quiet positions for training.

So, it is the same training code then, except you've cut out the SF bits?

For night nurse, yes.