Future of NNUE is Dimension 3072 network!

[Draude]

Hello Talkchess, this is the latest work to making Stockfish NNUE much stronger:

A version with changed parameters for the small network (Big-SF 150324). This version beats SF dev by a large margin with my own variants test book running on a fast server (Core i7-3770K, 3 cores for each SF, ponder ON). I prefer to use this book because it is clearly better for analysis, credit to Eduard for creating it!

Yes, Big-SF 150324 use Dimension 3072, this made it learn more and play better in many positions I tested.

SF programmers are finally learning from the 3072 network previously posted in this forum, A new 3072 network is currently being tested on Fishtest today! And they are training a new one soon. I hope, it will not make the engine unusable like last time

tests.stockfishchess.org/nns
Network : nn-fe11b1ade500.nnue
Username : linrock

I have been using the large network for a long time because it plays better in balanced openings and especially in complex closed positions. Previously I use the corchess engine which was the best SF clone available. In my local version I included the network in the developers' latest engine update from yesterday. In analysis, it is clearly the best Stockfish dev I have tested after Big-SF 110324 .


NNUE 3072 is the future of chess engine! In fact, I have already started developing Next Dimension 4096 networks. They are not ready yet, but already showing great improvement in positions like this one:

Qxf6!
[d]1B1r4/rp2npkp/2b1pbp1/1qp5/nPN1R3/1P1P1QP1/2P2PBP/5R1K w - - 0 1

Analysis by Big-SF 17Mar24-4096-avx2:

1.Qxf6+ Kxf6 2.Be5+ Kg5 3.Bg7 Bxe4 4.f4+ Kh5 5.Bxe4 g5 6.Ne5 Qc6 7.Bf6 h6 8.Bxe7 Qd5 9.Bxd8 Qxe5 10.fxe5 Nc3 11.bxc5 Nxe4 12.dxe4
Depth: 24/55 00:00:06 56MN

I hope SF developers will hurry up, add Dimension 3072 and Dimension 4096 network to their engine! After all, they are such greedy for Elo. Otherwise, I can confidently say they will be losing a lot to Lc0 at SSDF!

[towforce]

It looks as though this goes to the question of how big the NN should be. People who are familiar with my position know that I believe in a "God algorithm" for chess, but I'm going to take a moment to discuss this interesting question:

1. Having too little knowledge (implying a smaller NN) is fatal. We know that hand-coded evaluations are uncompetitive, because they lack important chess knowledge - even though they can get ridiculously high NPS (nodes per second)

2. Having too few NPS per second is fatal. Even the best NNs clearly have some missing knowledge, because to get all that knowledge, they have to be bigger, which leads to lower NPS. Therefore we can say that at the current time, even the biggest NNs lack essential knowledge that only a deep search (hence high NPS) can provide

3. This "3072 network" thread appears to be saying that Stockfish would perform better if it had a bigger network. A bigger network implies a lower NPS. Therefore, the OP seems to be saying that this new network contains enough extra high-value knowledge to make up for the loss of NPS

So the question becomes: why is this new, larger network able to more than make up for the loss of NPS that a larger network implies?

[Draude]

So the question becomes: why is this new, larger network able to more than make up for the loss of NPS that a larger network implies?

Hello, thank you for your question! I always believe Dimension 4096 network has just enough knowledge to provide the God algorithm to SF-Big engine! Much more than "official" Stockfish!

Example? Fat Fritz 2 when released was a truly great step for SF. It had big Dimension network. After it succeeded, SF network dimension increased!

Luckily I have Fat Fritz 2 engine locally and I can test on the following position:

Qxf6!
[d]1B1r4/rp2npkp/2b1pbp1/1qp5/nPN1R3/1P1P1QP1/2P2PBP/5R1K w - - 0 1

Analysis by [d]Fat Fritz 2:
1.Qxf6+ Kxf6 2.Be5+ Kg5 3.Bg7 Bxe4 4.f4+ Kh5 5.Bxe4 g5 6.Ne5 Qc6 7.Bf6 h6 8.Bxe7 Qd5 9.Bxd8 Qxe5 10.fxe5 Nc3 11.bxc5 Nxe4 12.dxe4 Ra2 13.c6 bxc6 14.Rf2 g4

It found the best move in 15 seconds. SF 16.1 cannot find it after 1 minute. Impressive for the engine released in 2021!

[Werewolf]

So the question becomes: why is this new, larger network able to more than make up for the loss of NPS that a larger network implies?

Hello, thank you for your question! I always believe Dimension 4096 network has just enough knowledge to provide the God algorithm to SF-Big engine! Much more than "official" Stockfish!

Example? Fat Fritz 2 when released was a truly great step for SF. It had big Dimension network. After it succeeded, SF network dimension increased!

Luckily I have Fat Fritz 2 engine locally and I can test on the following position:

Qxf6!
[d]1B1r4/rp2npkp/2b1pbp1/1qp5/nPN1R3/1P1P1QP1/2P2PBP/5R1K w - - 0 1

Analysis by [d]Fat Fritz 2:
1.Qxf6+ Kxf6 2.Be5+ Kg5 3.Bg7 Bxe4 4.f4+ Kh5 5.Bxe4 g5 6.Ne5 Qc6 7.Bf6 h6 8.Bxe7 Qd5 9.Bxd8 Qxe5 10.fxe5 Nc3 11.bxc5 Nxe4 12.dxe4 Ra2 13.c6 bxc6 14.Rf2 g4

It found the best move in 15 seconds. SF 16.1 cannot find it after 1 minute. Impressive for the engine released in 2021!

I tested this on a high thread count and SF 16.1 got it in 5 seconds. It's search luck.
Pruning allows a deeper search, but at a cost.
And we know in a match SF 16.1 would beat FF2. Maybe the day will come and we can have the best of both worlds.

[gordonr]

For this test position, can anyone post Stockfish analysis that does *not* show a winning move within a few seconds or so? Qxf6 and bxa4 and Bxa7 all win. I understand that faster wins are "nicer" but they all score 1-0. Surely if we're trying to show that one engine is better than another in some situations, it would be a lot more convincing to highlight examples of "finds win" versus "only draws or loses".

[Draude]

This is precisely why I make Stockfish clones! The developers are greedy about only ELO and cannot understand how to play good chess. In the position I sent only Qxf6 is the best move! The other moves are engine garbage. It shows the engine does not understand the position.

My dimension 3072 networks can understand the position with its training knowledge and, therefore, play better chess! It is simple.

[Eduard]

Keep up the good work and keep fooling me. Where can I find your compiles? You are an arrogant fake news poster. That's why I will now only share all my versions privately with my friends and delete all downloads with the parameters I changed.

[Eduard]

Everything deleted. I wish the forum a lot of fun with Draude and his new great clones. Keep it up.
(I'll say it again: My engine creation was just for my own fun and I wanted to share it. If that doesn't work, then I won't share anything anymore. Also good, no problem)

[Viren]

One last gift before you go: https://cdn.discordapp.com/attachments/ ... 37a51e601&

[Eelco de Groot]

This is a heavily moderated forum and they hire some two year olds to do it.

Look at it from the bright side: This is Stockfish in optima forma Everbody can join and be part of the winning team, even when you have just broken your ankle in your last 1000 m. race. It does not matter, everybody wins! Go Suzanne!