LCZero update

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

Moderators: hgm, Harvey Williamson, bob

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
Werewolf
Posts: 1174
Joined: Thu Sep 18, 2008 8:24 pm

Re: LCZero update

Post by Werewolf » Sat Mar 24, 2018 3:24 pm

[Date "2018.03.24"]
[White "Master "]
[Black "LCZ"]
[Result "1-0"]

New match!
LCZ vs a common dedicated machine: "The Master" made by Mephisto / Saitek and around 2100 elo. Note LCZ turned down a draw.

{Master played at G20. LCZ in slow (2 sec / move) mode.}
[pgn] 1. e4 e5 2. Nf3 Nf6 3.Nxe5 d6 4. Nf3 Nxe4 5. d4 d5 6. Bd3 Be7 7. O-O O-O 8. c4 Nf6 9. cxd5 Nxd5 10.Nc3 Nf6 11. Re1 h6 12. Qb3 Nc6 13. Be3 Be6 14. Qxb7 Nb4 15. Bb1 Rb8 16. Qxa7 Ra8 17. Qb7 Rb8 18. Qa7 Ra8 19. Qb7 Bd6 20. Ne5 Nfd5 21. Qb5 Nxe3 22. Rxe3 f6 23. Nc6 Qc8 24. Nxb4 Rb8 25. Qd3 Rxb4 26. Qh7+ Kf7 27. Bg6+ Ke7 28. Rae1 Kd7 29. Qxg7+ Kc6 30. Rxe6 Rg8 31. Qxf6 Qf8 32. Bf7 Rg7 33. Rxd6+ Qxd6 34. Qxg7 Qxd4 35. Bd5+ Kb6 36. Re6+ Ka5 37. Qxc7+ Rb6 38. Qa7+ Kb4 39. Qa3# [/pgn] 1-0

Werewolf
Posts: 1174
Joined: Thu Sep 18, 2008 8:24 pm

Re: LCZero update

Post by Werewolf » Sat Mar 24, 2018 3:56 pm

Second match game between dedicated machine: The Master, and LCZ ended in another win for the Master.

This was interesting though. LCZ gave away (or maybe blundered, not sure) 2 pawns in the opening for serious compensation, but yet again tactics was its undoing.

[Event "Match Game 2"]
[Date "2018.03.24"]
[White "LCZ"]
[Black "Master"]
[Result "0-1"]

{Master played at G20, LCZ in Slow Mode (2 sec / move)} [pgn] 1. e4 e5 2. Nf3 Nc6 3. Bc4 Bc5 4. c3 Nf6 5. d4 exd4 6. cxd4 Bb4+ 7. Nbd2 Nxe4 8. d5 Na5 9. Be2 c6 10. d6 Nxd6 11. O-O O-O 12. a3 Bxd2 13. Bxd2 Re8 14. Bxa5 Qxa5 15. Qxd6 Rxe2 16. Rae1 Re6 17. Rxe6 dxe6 18. Rd1 h6 19. h3 Kh7 20. Qe7 f6 21. b4 Qxa3 22. Rd8 Qc1+ 23. Kh2 Qf4+ 24. Kg1 c5 25. bxc5 Qc1+ 26. Kh2 b6 27. cxb6 Qf4+ 28. Kg1 axb6 29. Rd4 Ra1+ 30. Rd1 Rxd1+ 31. Ne1 Rxe1# [/pgn] 0-1

Score:
Master = 2
LCZ = 0

Werewolf
Posts: 1174
Joined: Thu Sep 18, 2008 8:24 pm

Re: LCZero update

Post by Werewolf » Sat Mar 24, 2018 5:18 pm

The 3rd match game was a very interesting opening, which I felt LCZ played intelligently. The idea with ...a5 followed by ...Ne4 was noteworthy.

Unfortunately the 20 MHz Frans Morsch machine is rather good when it gets an attack, which it did. LCZ lost its way in the tactics and it finished quickly.

[Event "3rd Match Game"]
[Date "2018.03.24"]
[White "Master"]
[Black "LCZ"]

{Master played at G20, LCZ in "slow mode" (2 sec / move)}
[pgn] 1. e4 e5 2. Nc3 Be7 3. Nf3 d6 4. Bc4 Nf6 5. d3 O-O 6. O-O c6 7. Bb3 Nbd7 8. Bd2 a5 9. d4 a4 10. Bxa4 exd4 11. Nxd4 Nxe4 12. Nxe4 Rxa4 13. c3 Ra8 14. Nf5 d5 15. Qg4 g6 16. Ned6 Bf6 17. Bf4 Qb6 18. Nxc8 Raxc8 19. Nh6+ Kg7 20. Qxd7 Ra8 21. Ng4 Rfd8 22. Bh6+ Kh8 23. Qxf7 Rf8 24. Bxf8 Bd4 25. Nf6 Bxf6 26. Qxf6+ Kg8 27. Qg7# [/pgn]
1-0

Match Score:
Master = 3
LCZ = 0

gladius
Posts: 537
Joined: Tue Dec 12, 2006 9:10 am

Re: LCZero update

Post by gladius » Sat Mar 24, 2018 5:50 pm

JJJ wrote:Maybe I should try the slow mode after all
Slow mode definitely spots the mate and avoids it :). The policy network hasn't quite figured out mates in one yet. Will be interesting to see when it starts getting that (if it does!).

User avatar
CMCanavessi
Posts: 832
Joined: Thu Dec 28, 2017 3:06 pm
Location: Argentina

Re: LCZero update

Post by CMCanavessi » Sat Mar 24, 2018 6:03 pm

Broadcasting LCZero Gen 20 gauntlet right now, we'll see how much "real life" improvement there was in the past 4 generations.
Follow my tournament and some Leela gauntlets live at http://twitch.tv/ccls

Werewolf
Posts: 1174
Joined: Thu Sep 18, 2008 8:24 pm

Re: LCZero update

Post by Werewolf » Sat Mar 24, 2018 6:24 pm

On the play LCZ website, please could you show which version of LCZ it is?

It would be really fun to see which version it is in general, but also for the match with the Saitek Master.

Damir
Posts: 2023
Joined: Mon Feb 11, 2008 2:53 pm

Re: LCZero update

Post by Damir » Sat Mar 24, 2018 7:00 pm

Here is my game agaisnt LCZero:

History:
1. e4 e5
2. Nc3 Nf6
3. d3 d6
4. f4 exf4
5. Bxf4 Be7
6. Nf3 O-O
7. Be2 Qe8
8. Qd2 Nbd7
9. O-O-O h6
10. h3 Nb6
11. g4 a5
12. Rdg1 a4
13. g5 Nh5
14. gxh6 g6
15. Be3 Bf6
16. Ng5 Qc6
17. Bxh5 gxh5
18. Ne6+ Kh7
19. Nxf8+ Kh8
20. Qg2 a3
21. Qg8#

I checkmated it in 21 moves.

:) :) :D

noobpwnftw
Posts: 291
Joined: Sun Nov 08, 2015 10:10 pm

Re: LCZero update

Post by noobpwnftw » Sat Mar 24, 2018 7:01 pm

Is the ceiling of 6x64 still quite far away after 3 days?

CheckersGuy
Posts: 273
Joined: Wed Aug 24, 2016 7:49 pm

Re: LCZero update

Post by CheckersGuy » Sat Mar 24, 2018 7:21 pm

noobpwnftw wrote:Is the ceiling of 6x64 still quite far away after 3 days?
I would assume so. This net would need to outperform the NN that was trained with human games. Until then I dont think the net is saturated at all.

jkiliani
Posts: 143
Joined: Wed Jan 17, 2018 12:26 pm

Re: LCZero update

Post by jkiliani » Sat Mar 24, 2018 7:21 pm

noobpwnftw wrote:Is the ceiling of 6x64 still quite far away after 3 days?
I think it's still quite a way to the ceiling. So far it hasn't surpassed a supervised net of the same architecture yet, and it's proven that reinforcement learning nets will outperform SL nets of the same size by a significant margin.

Locked