Stockfish - Update architecture to "SFNNv4

[Eduard]

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!

[Jouni]

+2,81 ELO for 60000 @ 60+0.6 th 1.

[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...

[AdminX]

[dkappe]

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.

[Rebel]

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.

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

[adnoh]

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

https://github.com/official-stockfish/S ... /pull/3927

First session:

The first session was training a network from scratch (random initialization). The exact trainer used was slightly different (older) from the one used in the second session, but it should not have a measurable effect. The purpose of this session is to establish a strong network base for the second session. Small deviations in strength do not harm the learnability in the second session.

Second session:

The second training session was done starting from the best network (as determined by strength testing) from the first session.

[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.

[dkappe]

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.

[Sopel]

get yourself a tshirt in celebration