StockfishNNUE! Probably the future of Stockfish!

If you look at computer chess competitions, like TCEC and chess.com computer championship, you will see testing of this new engine.

1. It is a Neural Network evaluation of Stockfish rather than traditional hand crafted evaluation

2. It differs from Leela, A0 and other NN cos it doesnt require Graphic Cards.

3. It is only 50% slower than Stockfish in search but better evaluation make this stronger than Stockfish. In comparision Leela is approx 1/1000 to 1/10000 speed of Stockfish, depending on your hardwares.

4. It is approx +30 elo stronger than current Stockfish ( since after the initial release in 2 months ago). Ongoing training and testing.

5. Here is original post from talkchess and current best network.

6. SF NNUE, Neural Network size is still massively smaller than Leela and Alpha Zero.( So, in term of 1 node search per evaluation, Leela will be massively stronger , but Leela is very slow)

Best Leela NN = 150 MB

Alpha Zero= 35 MB

SFNNUE= 30 MB

https://workupload.com/file/ggEUrvNVgmH

who?

It is a difficult subject to understand. Was Stockfish's playing style positional like Leela's? How was this neural network trained?

@barretoff, they let the engine run in million games for depth 8 or depth 16 of Stockfish. Then evaluate the position based on the evaluation of those depth, rather than the original static evaluation.

This is static evaluation of original Stockfish, at depth 0. ( put any position and see the base Stockfish evaluation at 0 search)

https://hxim.github.io/Stockfish-Evaluation-Guide/

Originally, if Stockfish can search 1 million positions, every position is evaluated by above table and Stockfish will choose the most favourable position for her from those 1 million position.

Where are the tests that show StockfishNNUE is 30 elo stronger than Stockfish?

Many tests posted in talkchess.

Here are a few of them.

http://talkchess.com/forum3/viewtopic.php?f=2&t=74484

Hmmm. Leela is beating StockfishNNUE now. But StockfishNNUE is supposed to be stronger than Stockfish, which beat Leela quite recently. Something seems off.

Probably the admin of cccc, the real greco, messed up the setting of SFNNUE .

But that Leela was one, this one is another. Leela now uses a different network, different from the TCEC 18 final.

This site (CCRL) is not very accurate in relation to Leela and Leelestein. In this list there the networks and versions used are no longer the strongest and they have not updated this to date.

oh I didn't know that

CCRL is not updated rating list but they are testing the newest version vs their standarized pool to get certain rating comparision. In their list, latest Stockfish developmental version is not included which is +25 elo to SF 11 and also T60 is not included which may be +25 elo to their T40.

In term of large scale 4X GPU, you can see reliable performance of SF NNUE in Tcec.

I saw it, lost to Alie and won 3 games for a weaker engine.

For me, this site here is much better than the CCRL, because it is always updated with the new engine versions: https://www.sp-cc.de/

I agreed! He just finished testing a while ago.

Why don't they have SF NNUE in the brackets at https://tcec-chess.com/live.html ?

Stockfish and Alliestein is coming up, should be interesting to see these games, but after that it's just LCZero and the winner of the current match between Fire and Xiphos. I don't see SF NNUE listed anywhere.

Stockfish NNUE is technically Stockfish fork or clone.

Stockfish team must choose Stockfish or SFNNUE . (they cant submit both) 

(In fact, there are more than a dozen Stockfish forks , eg sugar,crystal, Corfish etc)

Who else noticed that stockfish has a higher rating that Stockfish NNUE in the CCCC?

I know, as I mentioned above, it is alleged that the admin of cccc messed up set up SFNNUE.

However , SFNNUE is doing well in  TCEC , as majority of people behind TCEC team are highly skilled professionals. 

I have a game that SF NNUE beat SF 11! But SF Dev is still better than SF NNUE. This is the game.