SF-NNUE going forward...

Discussion of anything and everything relating to chess playing software and machines.

Moderators: Harvey Williamson, Dann Corbit, hgm

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
User avatar
Ozymandias
Posts: 1243
Joined: Sun Oct 25, 2009 12:30 am

Re: SF-NNUE going forward...

Post by Ozymandias » Sat Sep 05, 2020 3:17 pm

Ovyron wrote:
Tue Jul 28, 2020 10:32 pm
Raphexon wrote:
Tue Jul 28, 2020 9:03 pm
Has NNUE refuted a few lines yet?
Yes, there's a little opening revolution as we speak, people that haven't updated their lines have lost +70 elo points. I'm currently outperforming people in much faster hardware than mine and I'm not even using NNUE, just holding up with drawish openings that don't lose while others go down in flames.

This is very recent, since Eman Enigma 1.00 was released an started winning tourneys and people started jumping towards it, refuting opening theory still used by Stockfish dev users (mainly, out of book positions where, say, Stockfish dev depth 38 still claims 0.00, while NNUE at just depth 20 already shows 1.00 white advantage, then the white side goes on to win positions that the book claimed were fine for black, but weren't.)

NNUE is not outsearching Stockfish dev, but out-evaluating it, I used to play those 0.00 lines and get draws, now the NNUEs are killing me, I think anybody picking a great book from last week will be destroyed today when they play black (some white lines have fallen, but white still holds so NNUE isn't strong enough to break white Stockfish dev.)

Someone went and played with NNUE without book and got better performance than Stockfish dev with book, just because of those poor book likes that are being uncovered. I can't really imagine how out of opening eval will look like in 2021, but it's possible 0.00 positions of today will be 2.00 by then :shock:
I never cared much for evaluations, so I don't know what out of book positions will be evaluated as in 2021, neither by SF nor by whatever is found to be best at that time.

What I do care about are winning chances, and I don't see them going up because of NNUE, but rather that the draw ratio is increasing. For the two months, previous to NNUE surpassing SF, I have an 82.7%; for June I have an 83.5%, and for the last two months I see a whooping 85.2% draw ratio. This last figure for the July-August period, approximately the one you mention as with players "going down in flames".

It looks to me like fire extinguishers were readily available (i.e. just use the same engine that beat yours a few times and update the book).

chrisw
Posts: 3851
Joined: Tue Apr 03, 2012 2:28 pm

Re: SF-NNUE going forward...

Post by chrisw » Sat Sep 05, 2020 4:18 pm

Dann Corbit wrote:
Thu Aug 06, 2020 2:55 pm
Zenmastur wrote:
Thu Aug 06, 2020 3:36 am
Dann Corbit wrote:
Thu Aug 06, 2020 3:18 am
It seems to me that the size of the net that is optimal is likely a function of the power of the computer.
A one core Pentium would need a tiny net, and a 128 core dual Epyc would be able to use a much larger net.

I don't have any proof, just another thought experiment
It seems to me that if you are running 128 threads and each thread is handling part of the search then it will also have to handle the evaluation of the leaf nodes, which means it has to run the evaluation function. This mean you could have 128 threads each evaluation a different leaf node. Which means each thread must be able to run the evaluation function efficiently. So, I fail to see how having more cores will allow the use of larger nets WITHOUT degrading each cores NPS.
Even at that, the new architecture cores run at higher speed, and with lots ore cores you will also have a lot more hash hits. Considering that the eval is expensive, a zero ply eval hash might be a good idea
Chunks of the eval generally get hashed already, pawns alone, or pawns and kings. They tend not to use too much memory, just relying on the most recent batch of positions being kind of similar and overwriting older data.
There’s two problems with eval hash. First, to have the same “efficiency” as normal hash, you’ld need a lot more of it. Normal hash basically stores up to the leaf minus one ply, and not all programs store anything for qsearch. There’s a lot more eval nodes that there are tree nodes. and hence a much larger memory requirement (for equivalent hit rate). Second, main hash is unitary, it allows for communication across threads, but then there’s collision danger, race danger and so on to attend to. If your idea of a massive thread common eval hash were implemented, you’ld get these problems in spades (many more calls to save/probe hash due to larger numbers of leaf nodes than tree nodes). I’ld guess this is why threaded programs with hashed pawn eval have separate hash for each thread.

User avatar
Ovyron
Posts: 4405
Joined: Tue Jul 03, 2007 2:30 am

Re: SF-NNUE going forward...

Post by Ovyron » Sat Sep 05, 2020 11:27 pm

Ozymandias wrote:
Sat Sep 05, 2020 3:17 pm
What I do care about are winning chances
Some people at the top have been playing perfect chess for a while. The winning chances against them is 0%, no matter what one does. NNUE or no NNUE there's nothing to do, this is a flaw of chess itself, the draw margin is too wide.

So it only makes sense to talk about people that are already beatable. There's some way to beat these guys (I'm one of those guys, currently losing a Spanish game, anybody could have defeated me by playing my opponent's moves against me.) The draw ratio is going up because more and more people reach perfection, but these new tools available will make beating the beatable easier, specially if they continue to be engine slaves that just download Stockfish 12 and relay its moves to you in some way.

If you want winning chances, create a new account at chess-server.net, they'll pair you up against people that play without engine assistance, and winning chances will be 100%. As your rating goes up you'll face stronger opposition, and the wall will be hit as perfect opponents will be faced, but those that aren't perfect should be easier to defeat, even if they are fewer and fewer as they go over the perfection threshold.

User avatar
Ozymandias
Posts: 1243
Joined: Sun Oct 25, 2009 12:30 am

Re: SF-NNUE going forward...

Post by Ozymandias » Sun Sep 06, 2020 6:29 am

Ovyron wrote:
Sat Sep 05, 2020 11:27 pm
Some people at the top have been playing perfect chess for a while. The winning chances against them is 0%, no matter what one does. NNUE or no NNUE there's nothing to do, this is a flaw of chess itself, the draw margin is too wide.

So it only makes sense to talk about people that are already beatable. There's some way to beat these guys (I'm one of those guys, currently losing a Spanish game, anybody could have defeated me by playing my opponent's moves against me.) The draw ratio is going up because more and more people reach perfection, but these new tools available will make beating the beatable easier, specially if they continue to be engine slaves that just download Stockfish 12 and relay its moves to you in some way.

If you want winning chances, create a new account at chess-server.net, they'll pair you up against people that play without engine assistance, and winning chances will be 100%. As your rating goes up you'll face stronger opposition, and the wall will be hit as perfect opponents will be faced, but those that aren't perfect should be easier to defeat, even if they are fewer and fewer as they go over the perfection threshold.
Preaching to the choir, mate.

I'd say you're talking corr chess now. It gave me the impression you were discussing an ongoing situation in the Playchess Engine Room, hence my stats.

I already tried chess-server.net one or two years ago, but beating someone who plays by himself isn't my idea of winning.

User avatar
Ovyron
Posts: 4405
Joined: Tue Jul 03, 2007 2:30 am

Re: SF-NNUE going forward...

Post by Ovyron » Sun Sep 06, 2020 7:47 pm

Ozymandias wrote:
Sun Sep 06, 2020 6:29 am
I'd say you're talking corr chess now. It gave me the impression you were discussing an ongoing situation in the Playchess Engine Room, hence my stats.
Ah, I didn't expect the draw rate to be that high on Playchess.

People have been bragging more and more about longer and longer streaks of games without a single loss, it seems 5 0 chess is steadily reaching perfection. But I can beat people in much stronger hardware that are using NNUE by just using a net that happens to understand some position better, that wasn't possible before NNUE, it'd be interesting if someone can create a net specifically designed to defeat the most popular net, I could see a rock-paper-scissors effect where net A beats net B which beats net C, which beats net A.

Or maybe the opening holes will be filled out and we'll have draw death like in corr chess.

User avatar
Ozymandias
Posts: 1243
Joined: Sun Oct 25, 2009 12:30 am

Re: SF-NNUE going forward...

Post by Ozymandias » Sun Sep 06, 2020 8:27 pm

Ovyron wrote:
Sun Sep 06, 2020 7:47 pm
Or maybe the opening holes will be filled out and we'll have draw death like in corr chess.
That I would easily believe, just look how long it's taking to organize the 45th Engine Masters, after surpassing the 90% mark in the 44th.

User avatar
Ovyron
Posts: 4405
Joined: Tue Jul 03, 2007 2:30 am

Re: SF-NNUE going forward...

Post by Ovyron » Sun Sep 06, 2020 10:33 pm

Ozymandias wrote:
Sun Sep 06, 2020 8:27 pm
That I would easily believe, just look how long it's taking to organize the 45th Engine Masters, after surpassing the 90% mark in the 44th.
So, reduce the time control until the draw rate goes down to more manageable levels. In the future bullet 1 0 chess is going to be played at a level higher than in the 44th. If that level is acceptable in the 44th, why it doesn't become acceptable when a higher level is available?

1 0 chess is still far enough from draw death that it may be the future, organizers just have to realize current time controls are too long for the level displayed (and reduce it - I'm not supporting 1 0, but approaching it, perhaps try 11 +1?), otherwise the winner will be some random guy that luckily beat a fool somewhere in the tourney, and drew the rest of the games, and it becomes meaningless.

Cornfed
Posts: 114
Joined: Sun Apr 26, 2020 9:40 pm
Full name: Brian D. Smith

Re: SF-NNUE going forward...

Post by Cornfed » Mon Sep 07, 2020 1:25 am

Ozymandias wrote:
Sat Sep 05, 2020 3:17 pm
Ovyron wrote:
Tue Jul 28, 2020 10:32 pm
Raphexon wrote:
Tue Jul 28, 2020 9:03 pm
Has NNUE refuted a few lines yet?

What I do care about are winning chances,...
Might I suggest 'human chess'? It's the 'cats meow'...the real deal...you will get your winning chances if you are any good at chess.

User avatar
Ozymandias
Posts: 1243
Joined: Sun Oct 25, 2009 12:30 am

Re: SF-NNUE going forward...

Post by Ozymandias » Mon Sep 07, 2020 5:38 am

Ovyron wrote:
Sun Sep 06, 2020 10:33 pm
Ozymandias wrote:
Sun Sep 06, 2020 8:27 pm
That I would easily believe, just look how long it's taking to organize the 45th Engine Masters, after surpassing the 90% mark in the 44th.
So, reduce the time control until the draw rate goes down to more manageable levels. In the future bullet 1 0 chess is going to be played at a level higher than in the 44th. If that level is acceptable in the 44th, why it doesn't become acceptable when a higher level is available?
The problem seems to be that they don't deem that draw ratio acceptable. 1 0 might solve the problem, but it would most likely make the winner the guy with the best internet connection, as many of the loses would come from lag; not to mention the GUI, which never reached CB responsiveness.

They're certainly in a pickle, because even 3 0 is only 3% below rapid, in the draw ratio arena.

User avatar
Ovyron
Posts: 4405
Joined: Tue Jul 03, 2007 2:30 am

Re: SF-NNUE going forward...

Post by Ovyron » Tue Sep 08, 2020 5:38 am

Ozymandias wrote:
Mon Sep 07, 2020 5:38 am
They're certainly in a pickle, because even 3 0 is only 3% below rapid, in the draw ratio arena.
Wow! I still thought we had years left. I guess the progress was even more dramatic than I thought.

I wonder if 1 0 chess will reach >90% draw ration in a year of less, it seems very soon it'll become completely pointless to use a chess engine to play it (unassisted chess seems more healthy than ever, specially with online tournaments that allow people to stay home and win championships.)

Post Reply