LCZero update

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

Moderators: bob, hgm, Harvey Williamson

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
Locked
mkchan
Posts: 86
Joined: Thu Oct 06, 2016 7:17 pm
Location: India
Contact:

Re: LCZero update

Post by mkchan » Fri Mar 23, 2018 2:59 pm

My game with Leela doesn't even deserve a graphical view yet:
1. e4 Nf6
2. e5 Ng8
3. d4 f6
4. Bd3 fxe5
5. Qh5+ g6
6. Bxg6+ hxg6
7. Qxg6#

Hopefully will make some progress soon! :lol:

MonteCarlo
Posts: 111
Joined: Sun Dec 25, 2016 3:59 pm

Re: LCZero update

Post by MonteCarlo » Fri Mar 23, 2018 3:25 pm

I played a few games against it, and yeah, it's especially uneven at such fast time controls.

When I steer the game into closed maneuvering, its play stays fairly reasonable, but if you just go after it with a bunch of threats it just collapses.

I'm not sure how much of that is the super-fast time control and how much of that is just the inherent tactical issues with MCTS.

Right now it reminds me a lot of how the last iteration of spawk.fish played before that project died.

At the current pace, it's only a matter of time before it gets quite reasonable indeed! Now if only I had some GPUs... :(

EDIT: I suppose I should make it clear that this is from playing it at play.lczero.org.
Last edited by MonteCarlo on Fri Mar 23, 2018 3:34 pm, edited 1 time in total.

Nay Lin Tun
Posts: 628
Joined: Mon Jan 16, 2012 5:34 am

Re: LCZero update

Post by Nay Lin Tun » Fri Mar 23, 2018 3:28 pm

mkchan wrote:My game with Leela doesn't even deserve a graphical view yet:
1. e4 Nf6
2. e5 Ng8
3. d4 f6
4. Bd3 fxe5
5. Qh5+ g6
6. Bxg6+ hxg6
7. Qxg6#

Hopefully will make some progress soon! :lol:
What version of Leela?
[pgn][Event "Computer chess game"]
[Site "DESKTOP-U9VVTOA"]
[Date "2018.03.23"]
[Round "?"]
[White "Nay Lin Tun"]
[Black "Play [001]"]
[Result "1-0"]
[BlackElo "2000"]
[ECO "A07"]
[Opening "Reti Opening"]
[Time "23:16:36"]
[Variation "KIA, 2.g3"]
[WhiteElo "2400"]
[TimeControl "180+2"]
[Termination "normal"]
[PlyCount "111"]
[WhiteType "human"]
[BlackType "program"]

1. Nf3 d5 {(d7-d5 e2-e3 Ng8-f6 d2-d4 a7-a5 Nf3-e5 c7-c6) -0.03/22 5} 2. g3
e5 {(e7-e5 Nf3xe5 f7-f6 Ne5-f3 Nb8-c6 h2-h3 h7-h5 d2-d4) +0.17/21 5} 3.
Nxe5 f6 {(f7-f6 Ne5-f3 Ng8-e7 d2-d4 Nb8-c6 c2-c3 a7-a5 a2-a4) -0.01/22 5}
4. Nf3 Ne7 {(Ng8-e7 Bf1-g2 Nb8-c6 d2-d4 a7-a5 c2-c3 Ra8-b8) -0.19/22 5} 5.
Bg2 c6 {(c7-c6 O-O a7-a5 c2-c4 d5xc4 Nb1-c3 Nb8-a6 e2-e4) -0.39/22 5} 6. d4
g5 {(g7-g5 h2-h3 Bf8-g7 Nb1-c3 O-O e2-e4 d5xe4 Nc3xe4) 0.00/22 5} 7. e3 g4
{(g5-g4 Nf3-g1 Bf8-g7 h2-h3 f6-f5 h3xg4 f5xg4 Nb1-c3 O-O e3-e4 d5xe4
Nc3xe4) +0.46/21 5} 8. Nh4 Bg7 {(Bf8-g7 f2-f3 f6-f5 O-O O-O f3xg4 f5xg4
Rf1xf8+ Kg8xf8 Nb1-c3) +0.66/22 5} 9. Nc3 O-O {(O-O O-O f6-f5 Nc3-e2 a7-a5
Ne2-f4 Nb8-a6 c2-c3 Bg7-f6) +0.72/22 5} 10. Ne2 f5 {(f6-f5 O-O Ne7-g6
Nh4xg6 h7xg6 Ne2-f4 g6-g5 Nf4-d3 a7-a5) +0.62/22 5} 11. Nf4 h5 {(h7-h5
Nh4-g6 Ne7xg6 Nf4xg6 Rf8-f7 h2-h3 g4xh3 Bg2xh3 Nb8-a6 Qd1xh5) +0.22/22 5}
12. Nxh5 Rf7 {(Rf8-f7 Nh5xg7 Rf7xg7 h2-h3 g4xh3 Bg2xh3 a7-a5 Nh4-f3 Nb8-a6
Nf3-e5) -0.15/22 5} 13. Qd3 c5 {(c6-c5 Nh5xg7 Kg8xg7 d4xc5 Nb8-c6 Qd3-d1
Ra8-b8 h2-h3 g4xh3) +0.89/22 5} 14. dxc5 Nbc6 {(Nb8-c6 Nh5xg7 Kg8xg7 O-O
a7-a5 Qd3-d1 Ra8-b8 c2-c4 d5xc4) +0.35/22 5} 15. Nxg7 Kxg7 {(Kg8xg7 O-O
a7-a5 c2-c4 d5xc4 Qd3xc4 Nc6-e5 Qc4-c3 Rf7-f8 Qc3xe5+) -0.91/22 5} 16. Bd2
a5 {(a7-a5 Bd2-c3+ Kg7-g8 h2-h3 g4xh3 Bg2xh3 Nc6-b4 Qd3-d1 Nb4-c6 Qd1-h5)
-0.42/22 5} 17. Bc3+ Kg8 {(Kg7-g8 Qd3-d1 Nc6-b4 h2-h3 Nb4-c6 h3xg4 f5xg4
a2-a4 d5-d4 e3xd4 Ra8-b8) -1.51/22 5} 18. O-O-O Nb4 {(Nc6-b4 Qd3-d2 Nb4xa2+
Kc1-b1 Na2xc3+ Qd2xc3 d5-d4 e3xd4 Ra8-b8 Rh1-e1 Ne7-d5) -1.03/22 5} 19.
Bxb4 axb4 {(a5xb4 Qd3-b3 Qd8-e8 Rh1-e1 Kg8-h7 Qb3xb4 Ne7-c6 Qb4-c3 Bc8-e6
Bg2xd5 Be6xd5) -1.72/22 5} 20. Bxd5 Nxd5 {(Ne7xd5 Qd3xd5 Qd8-e8 Nh4-g6
Bc8-e6 Qd5-e5 Be6xa2 Qe5xe8+ Ra8xe8 h2-h4 b4-b3 c2xb3 Ba2xb3 Rd1-d6 Bb3-c4
h4-h5) -1.45/22 5} 21. Qxd5 Qe8 {(Qd8-e8 Nh4-g6 Bc8-e6 Qd5-d4 Be6xa2 Ng6-e5
Rf7-e7 f2-f4 b4-b3 Rh1-e1 b3xc2) -1.90/23 6} 22. Rd4 Be6 {(Bc8-e6 Qd5-e5
Ra8xa2 Rh1-e1 Qe8-c8 Qe5-d6 Ra2-a6 Qd6-e5 Be6-d7 Qe5-d6 Ra6xd6) -1.94/22 6}
23. Qd6 Ra6 {(Ra8-a6 Qd6-e5 Qe8-c8 Rh1-d1 b7-b6 c5xb6 Be6xa2 Nh4xf5 Qc8xf5
Qe5xf5 Rf7xf5 b6-b7) -1.82/22 6} 24. Qf4 Qc8 {(Qe8-c8 Rh1-d1 Qc8xc5 Qf4-g5+
Kg8-h7 Nh4xf5 Qc5xf5 Qg5xf5+ Be6xf5 e3-e4 Bf5xe4 Rd4xe4 Ra6xa2) -1.41/22 6}
25. Rhd1 Qxc5 {(Qc8xc5 Qf4-g5+ Kg8-h7 Qg5-g6+ Kh7-h8 Rd4-d8+ Rf7-f8 a2-a3
Be6-f7 Rd8xf8+ Qc5xf8 Qg6xf5 b4xa3 b2xa3 Qf8-g7 Rd1-d7 Ra6xa3) -2.26/22 6}
26. Rd8+ Kh7 {(Kg8-h7 Qf4-g5 Qc5-c7 Rd1-d4 Ra6xa2 Rd8-e8 Qc7xc2+ Kc1xc2
b4-b3+ Kc2-b1) -1.99/22 6} 27. Qg5 Rf8 {(Rf7-f8 Rd8xf8 Qc5xf8 Rd1-d8 Qf8-g7
Qg5xg7+ Kh7xg7 b2-b3 Ra6xa2 Kc1-b1 Ra2-a7 Rd8-b8 b7-b5 Kb1-b2 Be6-f7
Nh4xf5+ Kg7-h7) -3.29/22 6} 28. Rxf8 Qxf8 {(Qc5xf8 Rd1-d8 Qf8-g7 Qg5xg7+
Kh7xg7 b2-b3 Ra6xa2 Kc1-b1 Ra2-a7 Rd8-d4 b7-b5 Rd4xb4 Ra7-a5 c2-c4 b5xc4
b3xc4 Kg7-f6 Kb1-b2 Ra5-a6 c4-c5) -3.33/22 6} 29. Rd8 Qg7 {(Qf8-g7 Qg5xg7+
Kh7xg7 b2-b3 Ra6xa2 Kc1-b1 Ra2-a6 Kb1-b2 Ra6-b6 f2-f3 g4xf3 Nh4xf3 Kg7-f6
Nf3-d4 Be6-f7 Rd8-c8 Bf7-g6) -3.36/22 6} 30. Qh5+ Qh6 {(Qg7-h6 Rd8-h8+
Kh7xh8 Qh5xh6+ Kh8-g8 Qh6-g6+ Kg8-f8 Nh4xf5 Be6-f7 Qg6xg4 Ra6xa2 Nf5-d6
Bf7-d5 Qg4-d1 Bd5-e6 Nd6xb7 Be6-c4) -10.69/21 5} 31. Rh8+ Kxh8 {(Kh7xh8
Qh5xh6+ Kh8-g8 Qh6-g6+ Kg8-f8 Nh4xf5 Be6-f7 Qg6xg4 Ra6xa2 Kc1-b1 Ra2-a6
Nf5-d4 Ra6-a8 Qg4-f3 Kf8-e7) -12.16/22 5} 32. Qxh6+ Kg8 {(Kh8-g8 Qh6-g6+
Kg8-f8 Nh4xf5 Be6-f7 Qg6xg4 Ra6xa2 Qg4xb4+ Kf8-g8 Qb4xb7 Ra2-a1+ Kc1-d2
Ra1-a2 g3-g4 Ra2-a1 h2-h4) -12.47/23 5} 33. Qg6+ Kf8 {(Kg8-f8 Nh4xf5 Be6-f7
Qg6xg4 b7-b6 Qg4xb4+ Kf8-g8 c2-c4 Ra6xa2 Nf5-d6 Bf7-g6 Qb4xb6 Ra2-a1+
Kc1-d2) -13.16/23 5} 34. Nxf5 Bf7 {(Be6-f7 Qg6xg4 b7-b6 Qg4xb4+ Kf8-g8
c2-c4 Ra6xa2 Nf5-d6 Bf7-g6 Qb4xb6 Ra2-a1+ Kc1-d2 Ra1-a2) -13.49/23 5 Black
resigns} 35. Qg7+ Ke8 {(Kf8-e8 Kc1-b1 Bf7xa2+ Kb1-c1 Ke8-d8 Qg7xb7 Ra6-f6
e3-e4 b4-b3 c2xb3 Ba2xb3 Qb7xb3) -11.89/21 5 Black resigns} 36. Qe5+ Re6
{(Ra6-e6 Qe5-c7 Ke8-f8 Qc7xb7 Re6-f6 Qb7xb4+ Kf8-g8 Nf5-e7+ Kg8-h7 Ne7-d5)
-13.33/21 5 Black resigns} 37. Ng7+ Ke7 {(Ke8-e7 Ng7xe6 Bf7xe6 b2-b3 Ke7-d7
Kc1-b2 Be6-g8 e3-e4 Kd7-c6) -15.28/21 5 Black resigns} 38. Nxe6 Bxe6
{(Bf7xe6 b2-b3 Ke7-d7 Kc1-b2 Be6-g8 Qe5-g7+ Kd7-c6 Qg7xg8 b7-b6) -16.60/21
5 Black resigns} 39. Qc5+ Kd8 {(Ke7-d8 b2-b3 Be6-d7 Qc5xb4 Bd7-c6 e3-e4
Kd8-c7 Qb4-c4) -15.88/21 4 Black resigns} 40. Qxb4 Bc8 {(Be6-c8 b2-b3
Kd8-c7 Kc1-b2 Bc8-d7 c2-c3 Bd7-c6) -15.39/21 4 Black resigns} 41. Kd2 Bd7
{(Bc8-d7 Qb4xb7 Bd7-c8 Qb7-b6+ Kd8-d7 b2-b4 Kd7-e7 b4-b5) -15.93/20 4 Black
resigns} 42. e4 Bc8 {(Bd7-c8 c2-c3 Bc8-d7 Qb4xb7 Bd7-e8 e4-e5 Be8-c6)
-16.36/20 4 Black resigns} 43. e5 Bd7 {(Bc8-d7 Qb4xb7 Bd7-c8 Qb7xc8+ Kd8xc8
e5-e6 Kc8-d8 b2-b4 Kd8-e7) -17.16/20 4 Black resigns} 44. Qd6 Kc8 {(Kd8-c8
c2-c4 Bd7-c6 e5-e6 Bc6-f3 e6-e7 b7-b5 e7-e8Q+ Kc8-b7) -17.67/20 4 Black
resigns} 45. e6 Bxe6 {(Bd7xe6 Qd6xe6+ Kc8-c7 Qe6xg4 Kc7-d6 h2-h4 Kd6-d5
h4-h5 Kd5-e5) -18.24/20 3 Black resigns} 46. Qxe6+ Kc7 {(Kc8-c7 Qe6xg4
Kc7-d6 h2-h4 Kd6-e5 Qg4-d1 Ke5-f5 Qd1-h5+) -19.47/21 3 Black resigns} 47.
c4 Kd8 {(Kc7-d8 Qe6xg4 Kd8-c7 Kd2-c3 Kc7-d6 h2-h4 Kd6-e5) -19.60/20 3 Black
resigns} 48. Qxg4 Kc7 {(Kd8-c7 h2-h4 Kc7-d6 h4-h5 Kd6-e5 h5-h6 Ke5-f6)
-21.07/20 3 Black resigns} 49. h4 Kd6 {(Kc7-d6 h4-h5 Kd6-e5 h5-h6 Ke5-f6
h6-h7 Kf6-f7 h7-h8Q) -22.03/20 3 Black resigns} 50. h5 Ke5 {(Kd6-e5 h5-h6
Ke5-f6 h6-h7 Kf6-f7 h7-h8Q Kf7-e7 Qg4-g7+ Ke7-e6) -22.74/20 3 Black
resigns} 51. h6 Kf6 {(Ke5-f6 h6-h7 Kf6-f7 h7-h8Q Kf7-e7 Kd2-c3 Ke7-f7
Qh8-h7+) -23.76/20 3 Black resigns} 52. h7 Kf7 {(Kf6-f7 h7-h8Q Kf7-e7
Qg4-g7+ Ke7-e6 Qg7xb7 Ke6-d6) -24.72/20 3 Black resigns} 53. h8=Q Ke7
{(Kf7-e7 Qg4-g7+ Ke7-e6 Kd2-c3 Ke6-f5 Qg7xb7 Kf5-g6 Qh8-g8+) -25.54/19 2
Black resigns} 54. Qh6 Kf7 {(Ke7-f7 Qg4-e6+) -25.63/20 2 Black resigns} 55.
Qgg7+ Ke8 {(Kf7-e8 Qh6-h8+) -37.84/20 2 Black resigns} 56. Qhh8# 1-0
[/pgn]
Not easy in playing 3'+2" min blitz on my geforce gtx 1060 3GB card. I lose several games on time.

koedem
Posts: 91
Joined: Fri Mar 18, 2016 9:45 pm

Re: LCZero update

Post by koedem » Fri Mar 23, 2018 3:39 pm

I just played against someone who ran it on decently fast hardware (he said on average 10k playouts per move) and it actually played some nice chess. https://lichess.org/G7brl6KQ/white#0
Not only was it quite an effort to win, also the game looked very human like. Very impressive. :)
(but yeah, you need a decent number of playouts for it to play good chess)

Nay Lin Tun
Posts: 628
Joined: Mon Jan 16, 2012 5:34 am

Re: LCZero update

Post by Nay Lin Tun » Fri Mar 23, 2018 3:52 pm

How do they play Leela on lichess?

MonteCarlo
Posts: 111
Joined: Sun Dec 25, 2016 3:59 pm

Re: LCZero update

Post by MonteCarlo » Fri Mar 23, 2018 3:53 pm

Yeah, it really needs a decent amount of time to avoid getting completely confused.

I decided to put up with the delays of slow mode at play.lczero.org (there is candidates to watch during the delay :) ), and this was the game:

https://lichess.org/8idDxNek

I didn't even attempt to play correct chess...just blitzed out some crazy nonsense and it got confused pretty quickly.

It'd be interesting to see how its strength scales with extra time. I'd imagine that's a pretty steep slope right now.

koedem
Posts: 91
Joined: Fri Mar 18, 2016 9:45 pm

Re: LCZero update

Post by koedem » Fri Mar 23, 2018 3:59 pm

Nay Lin Tun wrote:How do they play Leela on lichess?
You have someone with strong hardware who manually transmits the moves. :)

Nay Lin Tun
Posts: 628
Joined: Mon Jan 16, 2012 5:34 am

Re: LCZero update

Post by Nay Lin Tun » Fri Mar 23, 2018 4:03 pm

Jesse Gersenson wrote:To get the current versions of this project, which is changing often, you will want to go directly to the source. The sources are listed below.

CLIENT
The official client, including source, is here:
https://github.com/glinscott/leela-chess/releases

ENGINE
A windows compile of the engine is built automatically and is available here (details below):
https://ci.appveyor.com/project/glinscott/leela-chess

NN 'network' FILE
The current network file is available here:
http://lczero.org/networks

Details of how to get engine:
1. go to https://ci.appveyor.com/project/glinscott/leela-chess
2. click a 'JOB NAME', for example "Environment: APPVEYOR_BUILD_WORKER_IMAGE=Visual Studio 2017"
3. click 'ARTIFACTS'
4. download engine

You likely want to get the network file with the highest ID number. Some users are confussing others and using this ID number together with the word 'generation'. For example, ID number 8 they call '8th generation'. More accurate would be to use the hex value given under NETWORK, as in, "A new network file, 03d94904, is available."
Thank you, I got it!

Uri Blass
Posts: 8790
Joined: Wed Mar 08, 2006 11:37 pm
Location: Tel-Aviv Israel

Re: LCZero update

Post by Uri Blass » Fri Mar 23, 2018 5:13 pm

koedem wrote:
It's a pity that it uses so little time however the guy who runs it said he considers writing a slow mode. Where the engine can think a lot longer. (though the user might have to wait a certain time between moves)
Right now it still blunders pieces left and right.
There is already a slow mode but it does not do what it is supposed to do.

It only prevents me from playing for 20 seconds but does not use 2 seconds per move but reply immediately.

It even can miss a mate in 1 based on my experience.

Note that I lost one game against it without slow mode when I played fast
(at 0.2 seconds per move I have no chance against it) but after it I won against it 3 times(of course I used more timethan the opponent in the games that I won) when the faster win was in slow mode when it only prevented me to play fast but even did not use the 2 seconds it is supposed to use.

Uri Blass
Posts: 8790
Joined: Wed Mar 08, 2006 11:37 pm
Location: Tel-Aviv Israel

Re: LCZero update

Post by Uri Blass » Fri Mar 23, 2018 5:23 pm

mkchan wrote:My game with Leela doesn't even deserve a graphical view yet:
1. e4 Nf6
2. e5 Ng8
3. d4 f6
4. Bd3 fxe5
5. Qh5+ g6
6. Bxg6+ hxg6
7. Qxg6#

Hopefully will make some progress soon! :lol:
I think that you were lucky and usually it does not play so weak in games based on my experience.

Locked