Where did you download Stockfish NNUE?
The playing style of StockfishNNUE
Sort:
neveraskmeforadraw wrote:
Where did you download Stockfish NNUE?
The first one I downloaded was from https://f9.rapiduservers.net/download/86d5df2b03e22eb2011cdc2fe88e1510c3001565/Stockfish-NNUE%202020-07-06.7z.
That one is actually outdated, but it's the one featured in the games I've been posting (except for the one in Sam Copeland's video).
Here is the one I just switched to, on advice from someone who knows a lot more than I do --
https://github.com/nodchip/Stockfish.git
It did require about an hour to figure out how to get the thing running. These programmers, they don't follow the same logic us peasants do, hah. I may be able to help if you have the same hardware I have (AMD Ryzen 9 3900X, RTX 2060) but if not the Stockfish Discord knows a lot more than me.
They're really helpful!
Reset my own testing, since I've been using the wrong version of SF NNUE the entire time.
Game 1, Stockfish 11 -- specific version 20011801, no openings book, vs SF NNUE -- specific version 20200722-2141. (Numbers that make zero sense to me but apparently distinguish between different Stockfish versions)
Once the Queens were traded, Black was relatively comfortable, electing to enter a bishop + 3 pawns vs rook imbalance. User adjudicated when it became clear there was no way either side would make progress.
We continue to see history being made. Over in chess.com's Computer Chess Championships, a 200-game match between LC0 and Stockfish NNUE was finished. Stockfish NNUE ... lost. Convincingly. Somewhere around 110-80 if I remember correctly.
But then in true programmer's fashion, Stockfish's dev team complained that the configuration for Stockfish NNUE was flawed, and requested a rematch. The CCC organizers complied, and this time switched to a 600-game match between LC0, Stockfish 11 and Stockfish NNUE.
The result is quite possibly the best chess, to date, the world has ever seen, excluding perhaps the best correspondence matches. The current score has LC0 and NNfish tied at 57.5, but with LC0 at 113 games finished vs NNfish's 112.
The loser is Stockfish 11 at 54 points. It may not seem like much of a difference, but it is a major one, for the following reason:
Each matchup has a manually selected opening that all three players must play, both as Black and as White. Some openings are almost certain to score a given result (it is hard to imagine an engine winning the Latvian as Black, for example), so this is a very good test of each engine's ability to score where it matters.
And, so far, Stockfish NNUE has struck a few victories vs Stockfish 11 in openings SF11 was unable to return the favor. That speaks a ton.
As of now, Stockfish NNUE is trashing SF11 30.5-25.5. However! LC0 is also edging out SFNNUE head to head, 29 to 27. This is very intriguing if it continues, as LC0 vs Stockfish 11 is currently tied at 28.5 points each.
Here are a few openings where Stockfish 11, vs NNUE, was unable to match a lost point with reversed colors:
[edit]-- I was unable to find the Grob match in the archive, though I could have sworn there was one. Oh well. But I did see that NNUE defeated Stockfish 11 in the King's Gambit Accepted, and I had failed to spot it-- 1. e4 e5 2. f4 exf4 3. Bc4 Qh4+ 4. Kf1 d6 5. Nc3 Nf6 6. d4 Be7 7. Qe1 Qxe1+
8. Kxe1 *, NNUE won as Black, drew as White
The French, 1. e4 e6 2. d4 d5 3. e5 c5 4. Qg4 -- NNUE won with White, drew with Black
The Benoni, 1. d4 Nf6 2. c4 c5 3. d5 e5 -- NNUE won with White, drew with Black
The French again, 1. e4 e6 2. d4 d5 3. Nc3 Bb4 4. e5 c5 -- NNUE won with White, drew with Black
And the above game in the King's Indian, 1. d4 Nf6 2. c4 g6 3. Nc3 Bg7 4. e4 d6 5. Nf3 O-O 6. Be2 e5 7. O-O Nc6 8. -- NNUE won with White, drew with Black.
For a lot of these obscure openings, these engines are making up completely new opening lines that may be of interest, despite the time control only being 3|2. It's truly a fascinating matchup to watch.
d5 Ne7 *
Jul 23, 2020
0
#26
<iframe id="6891178" allowtransparency="true" frameborder="0" style="width:100%;border:none;" src="//www.chess.com/emboard?id=6891178"></iframe><script>window.addEventListener("message",e=>{e['data']&&"6891178"===e['data']['id']&&document.getElementById(`${e['data']['id']}`)&&(document.getElementById(`${e['data']['id']}`).style.height=`${e['data']['frameHeight']+30}px`)});</script>
King123thegreat wrote:
6. Be6?? was apparently a planned opening by the CCC organizers. No idea why they tossed that one in. Now NNUE will have to defend it in the next game.
Stockfish NNUE has taken the lead, andbis now ahead of both S11 and Lc0.
A new era begins.
Time will tell, but it looks promising.
The pattern continues into today: NNUE is simply shredding Stockfish 11, 39 points to 35. However, it's actually lagging behind LC0 head to head, 39 to 38.
In terms of ELO gain, NNUE will certainly gain plenty of points over Stockfish 11, making it a clear #1 on the CCRL board. But this does beg the question: if it struggles against LC0 1v1, how is it going to win the next TCEC superfinals?
We're entering completely uncharted territory here, but for the Stockfish team, pivoting into the neural net territory will plunge them into an arms race they are at least a year behind. LC0 and its sister, Allie, have an enormous head start here.
There's also another interesting question here:
How will NNUE's newfound inspiration translate into its scores against the other CCRL ranks?
That's a tricky one actually. The CCRL has tested Stockfish 11 vs the rest of the top 20, and the absolute truth is it's hard to improve on Stockfish 11's results against anyone in the top 20. That's right--in current testing, Stockfish has lost only three games out of 482. Two of them were to LC0 ranked third, and one to SugaRNN ranked second, which is apparently a second derivative neural net based on Stockfish (I'm unsure if or how StockfishNNUE is related). Against the rest of the top 20, Stockfish 11 simply does not lose in 40|15|. When you get past the top 6 it gets even wilder: Stockfish 11 scores over 76% against everyone below Ethereal 11.75.
The only legitimate competition to Stockfish 11 has been from neural nets and the Komodo/Houdini team. It would be quite amusing if Stockfish NNUE gains a surge in rating simply by dominating other Stockfish iterations, but without any change in score against other competition, which is possible. Further testing needed there!
Testing continues.
Currently in a round robin tournament with Stockfish NNUE, Stockfish 11, Allie and LC0, one minute per move. One duplicate of each engine using a custom openings book from my own data.
Hardware is Ryzen 9 3900X & RTX 2060.
The field is intense. No one's giving an inch so far! 
How exactly is it a stupid time control? It fits my needs: it's long enough to produce decent-quality moves, especially since the net LC0 is using does poorly in blitz, and it's short enough that I can still see a few games finished per day.
I have no interest in seeing an engine lose because of time pressure in the endgame.
pfren wrote:
Dyslexic_Goat έγραψε:
How exactly is it a stupid time control? It fits my needs: it's long enough to produce decent-quality moves, especially since the net LC0 is using does poorly in blitz, and it's short enough that I can still see a few games finished per day.
I have no interest in seeing an engine lose because of time pressure in the endgame.
Which means that you force the engines to play some bad games, and you get deluded that this is the best chess ever played.
I don't say I'm sorry for your delusion- the computer is yours, I do not pay the electricity bill myself, so feel free to go on like this- I could not care less.
Maybe your definition of "bad" is the delusional side here. If these engines were making blunders there'd be a lot more decisive results. Any mistakes made here have so far been too small for their top competitors to exploit.
I get it, life's given you the role of "condescending old IM man" to play, and you normally play that role pretty well, but it may help to parse through your statements before posting them.
Jul 24, 2020
0
#37
Dyslexic_Goat wrote:
pfren wrote:
Dyslexic_Goat έγραψε:
How exactly is it a stupid time control? It fits my needs: it's long enough to produce decent-quality moves, especially since the net LC0 is using does poorly in blitz, and it's short enough that I can still see a few games finished per day.
I have no interest in seeing an engine lose because of time pressure in the endgame.
Which means that you force the engines to play some bad games, and you get deluded that this is the best chess ever played.
I don't say I'm sorry for your delusion- the computer is yours, I do not pay the electricity bill myself, so feel free to go on like this- I could not care less.
Maybe your definition of "bad" is the delusional side here. If these engines were making blunders there'd be a lot more decisive results. Any mistakes made here have so far been too small for their top competitors to exploit.
I get it, life's given you the role of "condescending old IM man" to play, and you normally play that role pretty well, but it may help to parse through your statements before posting them.
Dyselxic_Goat: 1
pfren: 0
Don't get me wrong. I have a lot of respect for Pfren's ability as a chess player.
But I have to give credit where it's due. The most important calculations for an engine happen in the first few seconds of analysis. There comes a point in search depth -- for "conventional" engines usually around the time they switch on A/B pruning -- where, if the engine hasn't solved its blind spots already, it will continue to have them virtually indefinitely. This is why an engine will commit to a move at search depth x and continue recommending that move at search depth x +20 unless something miraculous is discovered in another line.
So far, a one-minute per move time control has eliminated the idiotic mistakes that an engine will make under time pressure. At one minute per move, the engine will recommend the same move it would recommend at one day per move 99% of the time--though again, I do suspect neural nets may handle it a bit more efficiently, so, credit where it's due.
On this rig, Stockfish 11 hits an average search depth of 40+ and 1.8 billion positions analyzed in 60 seconds, averaging 30 million per second. StockfishNNUE is around 20-30% slower, but has proven to be more efficient in its use of that analysis.
Conventional engine wisdom suggests LC0 and Allie, who analyze at ~5k-20k per second on my rig, should struggle the most if time were an issue. So far they have not.
The first decisive match of this round robin. Allie 0.6, playing White with no openings book, offered its refutation of the Albin countergambit with itself using an openings book playing Black.
Over at the CCC, results have become more problematic for Stockfish NNUE.
It still dominates Stockfish 11, 62.5-59.5, but against LC0, Leela is leading 63.5-57.5. Stockfish 11 is still tied vs LC0 at 61 points each.
This confirms what I had suspected, but hoped wasn't the case here: Stockfish NNUE's improvements from fighting itself has led to more self-victories, but it may not be translating to better play against its opponents.
Well, shiver me timbers! I stand corrected!
Everything I thought I knew has just been thrown out the window, as it turns out the Stockfish NNUE I downloaded was actually outdated by an entire month. Go figure. An observer was able to help find the most recent version, and an important lesson has been reminded today: When using Stockfish & other fast-updated programs, it is important to notate the version history.
More to come soon. Forgive my blunder, things just got a whole lot more interesting...
[edit]
And on a side note, I also updated my rig's LC0 net too.