Fat Fritz 2

[dkappe]

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.

I’m not sure this is true. The quality of the eval and how closely the HCE corresponds to the NNUE influences the shape of the search tree. Still, it would be worth an experiment.

Right now I have a naive material eval for the ab search in a0lite when it’s beyond abs(450) cp, otherwise it uses night nurse. I’ll be plugging in a psqt eval in a few weeks. Will report back my findings.

[hgm]

My reasoning would be that as long as the root score still makes it a contest, any leaf that evaluates near +/-450 cP would have been a refutation of something no matter how accurate it is. And when the root score gets close enough to +/-450cP, so that the difference between NNUE and HCE can start having an effect... Well, how strong would that effect have to be in order to bungle a +450cP advantage? (Especially when it has the Stockfish search behind it.)

[noobpwnftw]

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.

False.

1. People did try many of those variants.
2. Not sure if it is anything clever if the result is actually worse.

[dkappe]

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.

False.

1. People did try many of those variants.
2. Not sure if it is anything clever if the result is actually worse.

I think you are being uncharitable. If someone in the SF “fold” had trained a net of that size that was inside the error margin to sfdev, you would be jumping up and down with joy. At least it has inspired you to suggest spsa tuning the 512x16x16 attempt in the discord.

[jshriver]

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.

That is a great attitude and exactly what I'm doing. I applied the patches to nodechips codebase and currently working on training a 512 net.

[noobpwnftw]

I think you are being uncharitable. If someone in the SF “fold” had trained a net of that size that was inside the error margin to sfdev, you would be jumping up and down with joy. At least it has inspired you to suggest spsa tuning the 512x16x16 attempt in the discord.

Which means I at least know all too well that it isn't good enough to start making wild claims.

Nobody is against trying something different, it may or may not bear fruit, but claiming things you didn't "invent" and lying about its performance is not something people usually do in the "SF fold". Is that something very common around you?

[dkappe]

I think you are being uncharitable. If someone in the SF “fold” had trained a net of that size that was inside the error margin to sfdev, you would be jumping up and down with joy. At least it has inspired you to suggest spsa tuning the 512x16x16 attempt in the discord.

Which means I at least know all too well that it isn't good enough to start making wild claims.

Nobody is against trying something different, it may or may not bear fruit, but claiming things you didn't "invent" and lying about its performance is not something people usually do in the "SF fold". Is that something very common around you?

Since I don’t sell chess software for money, I don’t have to make wild marketing claims. I have been overconfident with my predictions at times (as have you in the early days of your trainer rewrite ). For instance, I predicted that Dragon would be 200 elo stronger at knight odds than SF12. I and another tester found it was only 50 elo stronger. Of course those are predictions that can be tested. Many marketing claims cannot.

[Ckappe]

Let's get back to some basic facts in this thread.

a) What A. Silver has done is not anything unique nor revolutionary.. larger nets have been tried before. training from other engines pgn/fens as well (and also failed then as well)

b) A.Silvers network has never passed any objective quality controls. And has not in any way been proven trough fishtest or other trusted testing-frameworks to improve at all over current default-net.

c) Marketing a re-packaging of Stockfish with a not-so-tested network as "the strongest chess entity" in the world is a misleading and an outright marketing lie. Just because it is "marketing", and that you can "get away with it" in a court, does not make it less dishonest.

d) Calling Stockfish (or Leela) something else **-engine just for marketing & profit reasons, may be legal, but certainly not very honorable or respectable, to consumers and the chess community in large. Most chessplayers are no idiots and feels very strongly when they feel some company or persons try to scam them.

e) The Stockfish-compile + network sold by CB uses partly SF eval and partly the modified weight files included by CB/A.Silver.

f) CB first distributed the network as part of the binary but then released a new version of code to break it out when they realized that the network is part of the GPLv3 when embedded.

g) CB also tries to muddy the waters even more by releasing undocumented "fake" networks on GitHub (even calls these fakes the same - FatFritz2_v1.bin) along with the minor code changes they made to read the net they distribute.

[dkappe]

Let's get back to some basic facts in this thread.

e) The Stockfish-compile + network sold by CB uses partly SF eval and partly the modified weight files included by CB/A.Silver.

Moderator edit: condescending tone removed.
For concision, I’ve edited out all the claims I don't consider factual but simply opinion.

[gaard]

Let's get back to some basic facts in this thread.

a) What A. Silver has done is not anything unique nor revolutionary.. larger nets have been tried before. training from other engines pgn/fens as well (and also failed then as well)

b) A.Silvers network has never passed any objective quality controls. And has not in any way been proven trough fishtest or other trusted testing-frameworks to improve at all over current default-net.

c) Marketing a re-packaging of Stockfish with a not-so-tested network as "the strongest chess entity" in the world is a misleading and an outright marketing lie. Just because it is "marketing", and that you can "get away with it" in a court, does not make it less dishonest.

d) Calling Stockfish (or Leela) something else **-engine just for marketing & profit reasons, may be legal, but certainly not very honorable or respectable, to consumers and the chess community in large. Most chessplayers are no idiots and feels very strongly when they feel some company or persons try to scam them.

e) The Stockfish-compile + network sold by CB uses partly SF eval and partly the modified weight files included by CB/A.Silver.

f) CB first distributed the network as part of the binary but then released a new version of code to break it out when they realized that the network is part of the GPLv3 when embedded.

g) CB also tries to muddy the waters even more by releasing undocumented "fake" networks on GitHub (even calls these fakes the same - FatFritz2_v1.bin) along with the minor code changes they made to read the net they distribute.

Well put.

Don't feed the trolls.