Request for someone to train an NNUE for Ethereal

[AndrewGrant]

I have the code to run NNUEs in Ethereal. I'll probably clean it up and merge it if I had a network I could call "mine". Ethereal with an NNUE will very easily smash all engines in TCEC aside from Stockfish, and _maybe_ Leela. I'm willing to add whoever makes an original Net, one without running Stockfish evals or qsearches, to the authors output. I don't even care if its trained on Ethereal data or games -- just enough to pass TCEC's pathetic criteria.

I can provide something like 5 million Ethereal self-play games to help. But no one has been opposed to using Stockfish games (for some reason?) so there are ~50 million of those freely available, and they are quite easy to extract and strip to a small size.

If no one is immediately interested, I'd be willing to talk about $. I can run the training session on a 32-core CPU if desired. I'll be buying RTX3080s whenever they are again available.

[Madeleine Birchfield]

Why is TCEC and beating other engines so important to you that you feel the need to abandon your principles of not using NNUE in Ethereal?

[towforce]

Why is TCEC and beating other engines so important to you that you feel the need to abandon your principles of not using NNUE in Ethereal?

It's difficult enough to understand small neural networks: the human brain has around 10^11 neurons, with each neuron connecting to about 10,000 others, making a total of 10^15 synapses. It's wrong to assume that anyone is able to understand their own behaviour or motivations: they don't.

Having said that, whether I'm right or wrong, I now feel I have a satisfying explanation for Andrew's retirement messages: he wants to place high in TCEC, and felt that without having the latest weapon, he wouldn't be able to. He has now decided to acquire the new weapon, so retirement is postponed - at least temporarily...

[Raphexon]

I think my data generation should be just about complete.
When I create a net with it, it will be a flip net though.
Look at Noob's changes what to change to make it work: (not a lot btw)
https://github.com/noobpwnftw/Stockfish/tree/flip

Unlikely it will pass Fishtest so you can use it.
I'll try a few settings and see if it can be more different from SF eval than Ls15 is from T60.

[AndrewGrant]

I think my data generation should be just about complete.
When I create a net with it, it will be a flip net though.
Look at Noob's changes what to change to make it work: (not a lot btw)
https://github.com/noobpwnftw/Stockfish/tree/flip

Unlikely it will pass Fishtest so you can use it.
I'll try a few settings and see if it can be more different from SF eval than Ls15 is from T60.

Is this a network that meets TCEC's criteria -- not trained on SF evals and searches? If so, do you want to talk more?

We can talk here, through pms, Discord (Andrews#9072), or Email (andrew@grantnet.us)

[Raphexon]

I think my data generation should be just about complete.
When I create a net with it, it will be a flip net though.
Look at Noob's changes what to change to make it work: (not a lot btw)
https://github.com/noobpwnftw/Stockfish/tree/flip

Unlikely it will pass Fishtest so you can use it.
I'll try a few settings and see if it can be more different from SF eval than Ls15 is from T60.

Is this a network that meets TCEC's criteria -- not trained on SF evals and searches? If so, do you want to talk more?

We can talk here, through pms, Discord (Andrews#9072), or Email (andrew@grantnet.us)

Seems like the TCEC requirements (for NNUE) have become more strict, I won't be able to abide to that.

Feels a little arbitrary though. If you train purely on results you're not using SF eval as target, and since the Stockfish team can't claim ownership of individual games played it shouldn't be a problem. But I guess it is.

[AndrewGrant]

Seems like the TCEC requirements (for NNUE) have become more strict, I won't be able to abide to that.

Yup! Mission Accomplished. This thread can be ignored now safely.

[syzygy]

Are those TCEC rules written up somewhere?

[AndrewGrant]

Are those TCEC rules written up somewhere?

TCEC refuses to take a stance on anything, not even Houdini.
They have "Guidelines", found here : https://wiki.chessdom.org/TCEC_Cup_7_rules#Guidelines
Note that they change every 24 hours. People not as jaded as me, will complain just as much.

[syzygy]

Are those TCEC rules written up somewhere?

TCEC refuses to take a stance on anything, not even Houdini.
They have "Guidelines", found here : https://wiki.chessdom.org/TCEC_Cup_7_rules#Guidelines
Note that they change every 24 hours. People not as jaded as me, will complain just as much.

Thanks.

Do the NNUE guidelines apply outside NNUE technology?
No, of course not.

Why not!?
Why would you be allowed to tune your non-NNUE eval by playing games against Stockfish if you're not allowed to tune your NNUE eval with games played by (or rather: positions generated by) Stockfish.

3. All NNUE training data should be generated by the unique engine's own search and/or eval code.

So only self-play games are allowed (when tuning NNUE weights but not when tuning other weights)?