Leela Zero( A Neural Network engine similar to Alpha Zero)

Sort:
AIM-AceMove
Elroch wrote:

I was wondering when that was going to happen.

A great milestone. Chess engines are doomed!

[What was the time control?]

Happy for Leela improving. Not sure about chess engines, but humans definetely not. I was shocked and didnt belived it. A guy jose sylva had won and drawn many games in 10 min plus 5 , playleela account. And he is FM around skills? If it was C then long ago to be banned. And he was outplaying her on 10 min plus 5!! Also dogbody is drawing her in closed positions in blitz.

I mean how can we talk leelas 2850 fide and 3000 rating if this hapens? Sure different hardware and time control and stuff, but definetely not 2850 / 3000 for blitz.  Did she figured out how to checkmate with knight and bishop yet?!

A new chess match is needed vs a GM that knows what Leelas openings and weaknesses, and exploit them with smart play, not someone completely in the dark.

Then i will be pretty happy and cheer for her.

I believe GM Hammertime Magnus 2nd has interest playing her, he did once in his streams

 

AIM-AceMove

ID 235 checkmated her in blitz,  kings gambit accepted trap. 

https://lichess.org/epBSkUwI/white

Godeka

@AIM-AceMove:

Classic chess engines are weak in playing positional, which is exactly where LCZ has its strength, so one can expect LCZ gets better results against other engines. Humans are better in exploiting LCZ’s weakness.

I think LCZ has a lot of blind spots and will learn about it in the far future. Even Leela (Go) played at professional level without knowing how to play ladders – and we knew that AGZ learned them very late too.

LGZ makes a nice progress now, has still a smaller NN compared to AZ, and it had some severe bugs which are resolved. Currently everything looks fine.

drmrboss
Elroch wrote:

Of course: my main point was that AlphaZero is a bit nearer than it might appear!

Yes, I guess. Some people estimate 8x1080ti cards can reproduce the same amount of A0 nodes. Once there is bug free version of cuDNN, people can test exactly the same condition A0 vs stockfish  by doing 8 mins per move on Leela on 1080Ti vs 1 min per move on stockfish.

drmrboss

They already have exactly the same A0 blocks x filters (20x256)  version of Leela. Some people might be testing A0 replica in those buggy versions!

drmrboss

https://drive.google.com/file/d/11ueg-mdxYbe7_PmurO8a8ym2Vh3fPjw6/view

20x256 Leela, Replica of A0.😍

drmrboss
Virtually we already got the exact replica of A0.
 
Today I tested the replica of Alpha Zero, 20x256 blocks and filters of leela vs full strength stockfish, 2 games played , 2 games draw vs stockfish is amazing. :D 

Please test this leela. 
https://drive.google.com/uc?id=11ueg-md ... t=download
This Leela, 20x256, seemed quite strong. 
I played 3+2 mins blitz cuDNN 21st May (3rd version) Leela vs full strength latest vestion 4 cores i5 3.0GHZ stockfish on my 1060, for two games and both games draw. Opening book, perfect chess, 2017, 8 moves.


I will test more games later.

Nay

P.s average speed of Leela is around 2-2.5 Kn/s and average speed of stockfish is around 5 Mn/s.
https://lichess.org/UQytNWdL

https://lichess.org/IlhlNsul
 
AIM-AceMove
Godeka wrote:

@AIM-AceMove:

Classic chess engines are weak in playing positional, which is exactly where LCZ has its strength, so one can expect LCZ gets better results against other engines. Humans are better in exploiting LCZ’s weakness.

I think LCZ has a lot of blind spots and will learn about it in the far future. Even Leela (Go) played at professional level without knowing how to play ladders – and we knew that AGZ learned them very late too.

LGZ makes a nice progress now, has still a smaller NN compared to AZ, and it had some severe bugs which are resolved. Currently everything looks fine.

Nice explanation i see. Still difference gap is gigantic.

AIM-AceMove
drmrboss wrote:
Elroch wrote:

Of course: my main point was that AlphaZero is a bit nearer than it might appear!

Yes, I guess. Some people estimate 8x1080ti cards can reproduce the same amount of A0 nodes. Once there is bug free version of cuDNN, people can test exactly the same condition A0 vs stockfish  by doing 8 mins per move on Leela on 1080Ti vs 1 min per move on stockfish.

Cant they buy TItan?? Also gtx1180 is coming in couple months acording to speculations with new GDDR6

drmrboss

Yes, a few people have $3000 Titan V cards already. 

AIM-AceMove

Have leela beated komodo 1 yet? Whats current rank list

drmrboss

https://docs.google.com/spreadsheets/d/1XSJiCcQpCLv0fNwrUn7jXjdkZFU63YFEWpdXv6dSSg0/edit#gid=0

drmrboss

How to run Leela CUDA (should be around 4x faster, than Leela GPU(!)) in Fritz/Shredder GUI:

Important: Leela CUDA needs 3 .dll-files from the NVIDIA-developer-website, otherwise it can not run!
And - of course - you need a NVIDIA CUDA GPU in your PC...

1) Update your Nvidia driver for your GPU to the latest version here: http://www.nvidia.de/Download/index.aspx
After that: reboot your PC.
2) Get LC0.exe for the cuDNN version from: https://crem.xyz/lc0/ and unpack it and put it in your Leela CUDA folder.
3) Install CUDA v9.2 from https://developer.nvidia.com/cuda-92-download-archive (you need an account) and get cublas64_92.dll and cudart64_92.dll from the bin directory of CUDA v9.2 and copy it to your Leela CUDA-folder (Alternatively if you don't want to install CUDA, you can just unpack the installer and search in the unpacked folder and files for the .dlls with <Strg>+F in the Windows-Explorer, that was, what I did). Put the 2 files in your Leela CUDA-folder.
4) Get cuDNN for CUDA 9.2 from https://developer.nvidia.com/rdp/cudnn-download (you also need an account for this) and get cudnn64_7.dll from the bin directory of the zip and extract it and put it in your Leela CUDA-folder.
5) Download the latest network from http://lczero.org/networks and unpack it (Leela CUDA will autodetect it in any form either as weights.txt or however it is named) and put it in your Leela CUDA-folder.
6) Create the UCI-Engine in Fritz/Shredder.

Done.

7) I strongly recommend to change the UCI-settings of Leela CUDA to this:

Scale thinking time = 2.500000
FPU reduction = -0.068300
Cpuct MCTS = 3.168000

Why these values? Because FPU reduction and Cpuct were optimized with the CLOP-tool to these values.
And raising Scale thinking time to 2.5 makes Leela moving slower, which means a better use of the game-time.
In my testing, Leela plays definitly stronger with this settings.

drmrboss
Stockfish 10 dev version 4 cores, vs Leela 20x256 on 1060 GTX, 3+2 blitz. This draw game was exciting, whereas stockfish exchanged 1 minor vs 3 pawns, there was big disagreement between two engines since move 10, whereas stockfish saw he had big advantage from his side ,but leela saw big advantage from her side.
 
 
Armstrong-s-a
drmrboss wrote:
Stockfish 10 dev version 4 cores, vs Leela 20x256 on 1060 GTX, 3+2 blitz. This draw game was exciting, whereas stockfish exchanged 1 minor vs 3 pawns, there was big disagreement between two engines since move 10, whereas stockfish saw he had big advantage from his side ,but leela saw big advantage from her side.
 
 

Great game with a very low avg diff!

Leela was white in this game?

If Leela has beaten Stockfish 8 or 9 so far?

Armstrong-s-a

Leela Zero beats SF8 in 2 min TC!

Surprisingly, SF9 thinks that SF8 played better than Leela Zero, though Leela wins the game!

Indication that Leela Zero on-the-way to become world beater!

 

Armstrong-s-a

Sorry. There was an error in the preceding post. SF9 saw 4 inaccuracies and 2 mistakes in SF8 play.

DmitryReviews

YOU HAVE TO SEE THIS. I decide to fight Leela against other engines, like Rybka 2.2 with book, and finally agains't Stockfish with few resources and no book, and i spected Stockfish to win, but i NEVER spected this, even a 8 years old kid could see this:

 

drmrboss

 If you dont get 3000 + rating skill for Leela, there is something wrong with your setting, or low power graphic card. 

Btw, Lczero project will be discarded in next week, due to bugs.

There is new Leela or Lc0 , cuDNN version of Leela .

http://testserver.lczero.org/

DiogenesDue
thougtdancer wrote:

Shouldn't we be getting paid for giving them data (games we play against it}

Did someone coerce you to play Leela?