You gotta love Leela crazy aggressive attack! (Lc0 vs Komodo 9)
Have a look!
Leela Zero( A Neural Network engine similar to Alpha Zero)
Sort:
Leela is fantastic!
This was one of the game on SF vs Leela , 100 games match, (SF on 20 thread vs Leela on 2x1080Ti, 5+5 blitz. Total result, SF won in total, +56 elo above Leela.)
Leela played out of opening principles and played like a beginner. In fact she is not, she created new opening Novelty. SF was dead with Leela's crazy King side attack.
Leela's first clash with top 1 engine Stockfish in chess.com, computer chess championhip ended as draw.
Future of Leela!
Initial goals of Leela is almost complete!
Leela is very close to A0 level (-50 elo, she would probably be exactly the same level of A0 in new bug free nets) in simulation tests. She got very similar results vs SF8 in simulation tests, 80knps of Leela vs 70 Mnps of SF8.
Those will soon be possible evolution of leela. ( Likely stronger than A0)
Poll voting from Leela Discord as of today!
As we are slowly approaching strength of alphazero, and there are many ideas what to do better than what is described from alphazero paper, we are considering implementing support of experiments so that all crazy ideas could be easily checked and incorporated if they prove to be useful (with the speed things happen in LCZero, likely not within a month though). What do you think about that, where LCZero should go? 
Allow all changes, including ones which don't follow zero principle or possibly make style less "interesting", as long as it improves Elo. (E.g. mixing A/B with handwritten eval and MCTS+NN) 
Allow sidestepping from AlphaZero paper, as long as it doesn't violate "zero" principle. E.g. different NN in the beginning and endgame, MiniMax instead of MCTS, different board representation, etc. 
It's too early to sidestep, stick to AlphaZero paper for now.
I think default visit (search position) is 200k or 400k. If you want to do infinite analysis, you can do
" -v100000000" in engine parameter.
P.S. Tests showed that there is almost no measaurable elo gain after searching 400k visits( positions).
Also please tell me your
1. Graphic card
2. Your Leela Network ID number
3. Engine version number
I might figure out the best for you.
(Diagram, i put "-v 100000000000000 " in engine pararmeter
You are doing the right thing! (Open CL , Leela Zero with 15x192 network with your GPU) .
Unfortunately your graphic card doent support Lc0 ( cuDNN version ) which is about 4x-6x faster than Leela zero.
Regarding analysis, I checked again and saw that you have been analysing for 3 hours and you already have 1400 KN, that is more than enough for Leela.
Unless you are playing correspondence game, 400 k nodes for Leela and 1 billion nodes for SF is generally enough for causal analysis. ( it usually takes 3 mins for Leela and 5 mins for stockfish in my GTX 1060 card with 4 cores i5 desktop ). Beyond those nodes, engines strength increase is significantly tapered cos engines already searched important positions and their principal variation do not change very much!
tests for 20x256 showed that doubling 200 kn to 400 kn for leela gains 0 elo only.(400 kn for 20x256 is roughtly equal to 800 kn for 15x196 )
doubling 400 mnodes to 800 mnodes for Stockfish gains +17 elo only.
World chess championship live with Leela
http://viralcast.io/sports/world-chess-championship-2018-live-with-a-neural-network/
Sep 23, 2018
0
#372
drmrboss wrote:
(Diagram, i put "-v 100000000000000 " in engine pararmeter
You are doing the right thing! (Open CL , Leela Zero with 15x192 network with your GPU) .
Unfortunately your graphic card doent support Lc0 ( cuDNN version ) which is about 4x-6x faster than Leela zero.
Regarding analysis, I checked again and saw that you have been analysing for 3 hours and you already have 1400 KN, that is more than enough for Leela.
Unless you are playing correspondence game, 400 k nodes for Leela and 1 billion nodes for SF is generally enough for causal analysis. ( it usually takes 3 mins for Leela and 5 mins for stockfish in my GTX 1060 card with 4 cores i5 desktop ). Beyond those nodes, engines strength increase is significantly tapered cos engines already searched important positions and their principal variation do not change very much!
tests for 20x256 showed that doubling 200 kn to 400 kn for leela gains 0 elo only.(400 kn for 20x256 is roughtly equal to 800 kn for 15x196 )
doubling 400 mnodes to 800 mnodes for Stockfish gains +17 elo only.
What was the uncertainty on those increases? It is important to make sure it is small relative to the difference in Elo you are trying to measure. Larger increases in processing power / compute time would help to increase the accuracy of the estimation.
It would be surprising if a doubling in compute would have such a small effect even for Stockfish. Generally speaking you would expect a bigger increase for Leela from a doubling in compute (from the same starting Elo).
What HW is needed to get good Kns ? I only get about 7k/nps using a 1080Titan and Dual Xeon 44-core CPUs. Same config gets about 100Mns SF/ASMF. 80Kns mentioned in thread? 4x2080ti???
7kn/s for 1080ti is a bit slow, it should be 9kn/s.
Version 18 engine is even a bit faster, I guess it should be 11kn/s in version 18. (my estimate, i have 1060 gtx and version 18 speed up my card about 20%)
Try nvidia card setting for best performance like that. And also remove you windows bloatwares.
https://lc0bench.netlify.com/bench-result.html
Thanx for info :-) Tried optimizing the Nvidia settings, seems to increase slightly. However when running engine GPU hist 15% max on the task mgr. and CPU howvers about 1-3%.. Is that to be expected? Tried different number of workin threads.. Does the multithreading really work in 18rc2? seems only to go slower when increasing threads beyond 2?
In what language was leela coded in?
NichtGut wrote:
In what language was leela coded in?
https://github.com/LeelaChessZero/lc0
Source is here!
C++, beautiful. God bless denmark.
Leela is giving lessons to Komodo now.
There's a huge event coming up on Chess.com with Leela showing her force