Stockfish - Update architecture to "SFNNv4

[Sopel]

Is it really true this latest build / network is based on Leela data? This is what I see on the notes.

“From scratch” simply means that the starting net is a random collection of weights. Last time I checked, they were using leela data for that first go round, then a bunch of dodgy position data to clean up flaws in the second go round.

it's sf data first, then lc0

[Rebel]

Regarding the Leela data, are they using annotated PGN's (with scores) or PGN's without.

It’s from leela training data which has Z (result), Q (score), best move, position, and a bunch of other stuff. I was converting the Bad Gyal data like this to nnue data back in early 2020 to train Night Nurse.

I think they’ve got maybe 50b positions from leela converted (after rescoring with TB). That leaves a lot of room for filtering and massaging.

That's indeed very powerful, I consider Leela the real deal when it's about neural nets.

[dkappe]

Regarding the Leela data, are they using annotated PGN's (with scores) or PGN's without.

It’s from leela training data which has Z (result), Q (score), best move, position, and a bunch of other stuff. I was converting the Bad Gyal data like this to nnue data back in early 2020 to train Night Nurse.

I think they’ve got maybe 50b positions from leela converted (after rescoring with TB). That leaves a lot of room for filtering and massaging.

That's indeed very powerful, I consider Leela the real deal when it's about neural nets.

Well, leela and dragon are the only major engines doing intensive reenforcement learning (RL). Dragon is still quite young as RL networks go.

[Cornfed]

I downloaded last night and put it thru some paces as I normally do with new versions.

I pull up a particular large pgn file in Chessbase which is opening prep and use whatever my default engine for analysis is - currently 1/19/22 (if I remember right...), set each with the same parameters and often with 2 pv and 2 cores.

I then jump around to various positions to compare analysis - which engine gets to a particular choice of move...how fast...does it jump back and forth a lot...etc, etc.

Frankly, I'm sticking with 1/19/22 as my default for now...

In the short term you could probably get this kind of gain with the old architecture. Maybe there’s more headroom for future gains.

The challenge is that SF is essentially massaging large amounts of leela data and trying to get more out of it. Barring some way of generating useful data of its own (other than dodgy position data), moving to new architectures and tweaking training hyperparameters is the only way forward.

Still, SF Dev has made impressive advances. Over at CCC blitz the performance is trending toward +70 uho elo vs dragon which works out to +35 CCRL elo. Not bad.

Yes, all very understandable. Oh, I'll keep checking each iteration for sure.

My 'testing' in this sense is almost entirely subjective. But when I see a newer version generally taking a tad longer to decide on the top or top 2 moves...and those move being what the older version suggests, I keep the 'old' as my default for analysis. It would not surprise me at all if I jettison this for a new 'default' soon!

[bmp1974]

The new updated arch SFNNv4 & network is now able to solve some difficult puzzles with lower depths & lesser time.
Some examples:
1. 2b1rk2/5p2/p1P5/2p2P2/2p5/7B/P7/2KR4 w - - 0 1
2. 6br/1KNp1n1r/2p2p2/P1ppRP2/1kP3pP/3PBB2/PN1P4/8 w - - 0 1
3. r5k1/p1pb2bp/3p3r/P1pPp1p1/2B1Pq2/1R2QPp1/1P4PP/5RBK b - - 0 1
4. 4q1kr/p6p/1prQPppB/4n3/4P3/2P5/PP2B2P/R5K1 w - - 0 1
5. 8/8/4kpp1/3p1b2/p6P/2B5/6P1/6K1 b - - 0 47 (Shirov's brilliant Bh3 move position)

[Vernon Crawford]

https://abrok.eu/stockfish/

Author: Tomasz Sobczyk
Date: Thu Feb 10 19:54:31 2022 +0100
Timestamp: 1644519271

Update architecture to "SFNNv4". Update network to nn-6877cd24400e.nnue.

Impressively good. Congratulations!

FishTitz !

[Eduard]

The new updated arch SFNNv4 & network is now able to solve some difficult puzzles with lower depths & lesser time.
Some examples:
1. 2b1rk2/5p2/p1P5/2p2P2/2p5/7B/P7/2KR4 w - - 0 1
2. 6br/1KNp1n1r/2p2p2/P1ppRP2/1kP3pP/3PBB2/PN1P4/8 w - - 0 1
3. r5k1/p1pb2bp/3p3r/P1pPp1p1/2B1Pq2/1R2QPp1/1P4PP/5RBK b - - 0 1
4. 4q1kr/p6p/1prQPppB/4n3/4P3/2P5/PP2B2P/R5K1 w - - 0 1
5. 8/8/4kpp1/3p1b2/p6P/2B5/6P1/6K1 b - - 0 47 (Shirov's brilliant Bh3 move position)

In my own testset with interesting positions from server practice, this new Stockfish is the best.

[Chessqueen]

The new updated arch SFNNv4 & network is now able to solve some difficult puzzles with lower depths & lesser time.
Some examples:
1. 2b1rk2/5p2/p1P5/2p2P2/2p5/7B/P7/2KR4 w - - 0 1
2. 6br/1KNp1n1r/2p2p2/P1ppRP2/1kP3pP/3PBB2/PN1P4/8 w - - 0 1
3. r5k1/p1pb2bp/3p3r/P1pPp1p1/2B1Pq2/1R2QPp1/1P4PP/5RBK b - - 0 1
4. 4q1kr/p6p/1prQPppB/4n3/4P3/2P5/PP2B2P/R5K1 w - - 0 1
5. 8/8/4kpp1/3p1b2/p6P/2B5/6P1/6K1 b - - 0 47 (Shirov's brilliant Bh3 move position)

In my own testset with interesting positions from server practice, this new Stockfish is the best.

Where can I download the newer Stockfish with the SFNNv4 version ?

[jpqy]

On abrok https://abrok.eu/stockfish/

[Cornfed]

Taking about a half hour on Sundays, "Big search tuning (version 2)"...

The 'point eval' may indeed be more realistic early in the search. Certainly the kN/s dwarf my current default (think it's 1/19/22) for what that is worth...

So, an improvement...not sure yet if it will replace my default just yet.