Fat Fritz 2

[JohnWoe]

It is not 'just a NNUE player'. But NNUE player is definitely one of its functions. Which does not use the built-in HCE. FF2 only uses SF in its capacity of NNUE player; it doesn't use the HCE.

Oh I see. I stand corrected!

[Raphexon]

It is not 'just a NNUE player'. But NNUE player is definitely one of its functions. Which does not use the built-in HCE. FF2 only uses SF in its capacity of NNUE player; it doesn't use the HCE.

It does use HCE....

Fat Fritz 2 is literally Stockfish, including the hybrid eval. So it literally does use the HCE.
You can also turn off NNUE in the settings and it will be (again literally) identical to SFdev with NNUE turned off.

https://github.com/DanielUranga/Stockfish/tree/ff2

Why do you make such statements when you could have looked at the source code before you talked?

[hgm]

Well, I cannot read the source code. It is C++.

I was under the impression that Stockfish was pure NNUE, and that I read somewhere that this in general gives the best result. If it is hybrid, then I stand corrected.

Of course you can turn the NNUE off. But as you say, then it isn't Fat Fritz anymore, and just a regular Stockfish (for which you payed $99). You are not expected to do that.

[noobpwnftw]

Take the FF2 code, revert weak edits to the original state(or grab one from abrok, commit "Fix static_assert."), compile and run, boom you just made it 10 elo stronger than what you were sold for.

This is backed up by rating lists today, in all time controls:
http://ccrl.chessdom.com/ccrl/404/
http://ccrl.chessdom.com/ccrl/4040/

Not sure if this is even relevant to hybrid or anything else, I think turning NNUE off in FF2 can save it from being exposed that the "commercial net" actually made it much worse.

[Ckappe]

Well, I cannot read the source code. It is C++.

I was under the impression that Stockfish was pure NNUE, and that I read somewhere that this in general gives the best result. If it is hybrid, then I stand corrected.

Of course you can turn the NNUE off. But as you say, then it isn't Fat Fritz anymore, and just a regular Stockfish (for which you payed $99). You are not expected to do that.

But CB does currently not provide an option to turn hybrid off. So they only support hybrid mode currently- no pure NNUE !

[Dann Corbit]

Well, I cannot read the source code. It is C++.

I was under the impression that Stockfish was pure NNUE, and that I read somewhere that this in general gives the best result. If it is hybrid, then I stand corrected.

Of course you can turn the NNUE off. But as you say, then it isn't Fat Fritz anymore, and just a regular Stockfish (for which you payed $99). You are not expected to do that.

It was true for a while that NNUE replaced the standard eval. But they discovered that highly polar positions (where one side has a big advantage) are handled better by the traditional Eval. So they have a test that decides whether to use NNUE or standard evaluation. This hybrid evaluation is stronger than either one by itself.

[Dann Corbit]

Everyone seems to be hung up on "It is not very different from SF"

It does not have to be.

It is OK to not like what Chessbase has done, but that does not make it illegal.

It was clever to try doubling the size of the NNUE net. Nobody else did that. Yes, that change was trivial, but it was also clever.
You can make the same change in about a minute, but you would also have had to think of it and try it

I know that Albert has some interesting theories about training nets. i do not know if the theories are correct or not, but it is worthwhile to try them.

Are people getting their money's worth? Not sure. I did not buy it but if testing shows it has interesting properties I might buy it.
Or I might just build my own. My only problem there is that I am way too lazy to build my own net from scratch.
I do have lots of compute power, but I use it for things that are more interesting to me than building a net

[hgm]

It was true for a while that NNUE replaced the standard eval. But they discovered that highly polar positions (where one side has a big advantage) are handled better by the traditional Eval. So they have a test that decides whether to use NNUE or standard evaluation. This hybrid evaluation is stronger than either one by itself.

That sounds like you would not really need a clever evaluation, to finish off games where you already have a huge advantage. But probably they did not train the NNUE to recognize such polar positions.

It was clever to try doubling the size of the NNUE net. Nobody else did that. Yes, that change was trivial, but it was also clever.

You apply a very low standard. I would say it was pretty obvious. The 'width' of the first layer obviously is a parameter of the net architecture that can be changed, and it is custom for top-engine developers to tune any such parameter for an optimum. Just using 256 without trying 128 or 512 would be stupid. How would you know 8 wouldn't already be enough? (With an enormous speedup as a consequence.)

[dkappe]

That sounds like you would not really need a clever evaluation, to finish off games where you already have a huge advantage. But probably they did not train the NNUE to recognize such polar positions.

At fixed depth, NNUE wins. When time is a factor, you don’t need the overhead of NNUE in a branch where you’re a rook up.

[hgm]

Sure. But you probably also would not need the Stockfish HCE. A simple PST eval probably could already do it.

Of course I understand that the easiest solution is to use the evaluation you already happen to have, but saying that the strength of the hybrid depends for a large part on the quality of the Stockfish eval seems a bit of a stretch.