Progress seems slighty faster now. I remember when LCzero was way behind in term of elo and many guys crying about progress stalling. Lack of optimism here.
Anway, it's good to see Lczero becoming that strong.
LCZero: Progress and Scaling. Relation to CCRL Elo
Moderators: hgm, Rebel, chrisw
-
- Posts: 1346
- Joined: Sat Apr 19, 2014 1:47 pm
-
- Posts: 10948
- Joined: Wed Jul 26, 2006 10:21 pm
- Full name: Kai Laskos
Re: LCZero: Progress and Scaling. Relation to CCRL Elo
I am guessing that on GTX 1060 current LC0 with v0.8 binary might be close to 3000 CCRL 40/4'.JJJ wrote:Progress seems slighty faster now. I remember when LCzero was way behind in term of elo and many guys crying about progress stalling. Lack of optimism here.
Anway, it's good to see Lczero becoming that strong.
I adjusted time controls according to my CPU to CCRL 40/4' rating list, with LC0 ID245 on simulated GTX 1060 (TC on CPU 4 cores: 15m + 15s) and Naum 4.2 (TC on CPU 1 core: 2m + 2s, equivalent to 40/4' on CCRL), with CCRL 40/4' rating of Naum 4.2 on 1 core of 2987.
The result was equal in 4 games: +1 -1 =2
[pgn][Event "?"]
[Site "?"]
[Date "2018.05.05"]
[Round "1"]
[White "lczero_15_15"]
[Black "Naum_42_2_2"]
[Result "1/2-1/2"]
[BlackTimeControl "120+2"]
[FEN "rnbqk1nr/pp1p1ppp/4p3/2p5/1bPP4/2N5/PP2PPPP/R1BQKBNR w KQkq - 0 1"]
[PlyCount "154"]
[SetUp "1"]
[WhiteTimeControl "900+15"]
1. dxc5 {+0.12/19 24s} Nf6 {0.00/15 3.1s} 2. g3 {+0.12/20 24s}
Nc6 {-0.04/15 7.6s} 3. Bg2 {+0.21/20 21s} Qa5 {+0.03/15 3.0s}
4. Bd2 {+0.18/19 19s} Qxc5 {-0.38/13 4.5s} 5. a3 {+0.24/20 23s}
Bxc3 {-0.17/15 3.0s} 6. Bxc3 {+0.33/20 5.9s} Qxc4 {-0.19/16 4.7s}
7. Rc1 {+0.27/20 11s} Qg4 {-0.19/17 5.9s} 8. Nh3 {+0.22/20 28s}
d5 {-0.15/17 11s} 9. Bxf6 {+0.25/20 22s} gxf6 {+0.06/6 0.012s}
10. Nf4 {+0.30/20 8.8s} Bd7 {+0.03/15 6.4s} 11. O-O {+0.57/20 30s}
Rc8 {-0.04/15 4.4s} 12. Qb3 {+0.66/19 21s} b6 {+0.10/14 2.3s}
13. Rc3 {+0.60/20 32s} O-O {-0.08/15 5.7s} 14. Qa4 {+0.59/20 20s}
Rc7 {0.00/16 3.1s} 15. Rfc1 {+0.69/19 24s} Ne5 {-0.02/16 5.4s}
16. Qb4 {+0.86/19 25s} Rc4 {-0.07/16 2.6s} 17. Rxc4 {+0.74/19 27s}
dxc4 {-0.12/17 3.4s} 18. h4 {+0.64/20 24s} Qf5 {-0.04/15 3.5s}
19. b3 {+0.75/20 24s} cxb3 {-0.27/15 3.2s} 20. Bh3 {+1.14/19 17s}
Nc6 {-0.20/16 1.5s} 21. Qxf8+ {+1.07/19 27s} Kxf8 {+0.05/3 0.011s}
22. Bxf5 {+1.08/19 14s} exf5 {+0.20/16 1.9s} 23. Kf1 {+1.06/20 18s}
Nd4 {0.00/16 2.5s} 24. Ke1 {+1.21/18 28s} Nc2+ {+0.54/20 2.5s}
25. Kd2 {+2.59/18 25s} b2 {+0.50/21 3.0s} 26. Kxc2 {-0.40/20 27s}
bxc1=Q+ {+0.43/23 3.4s} 27. Kxc1 {-0.31/20 18s} Ke7 {+0.42/24 3.4s}
28. Kd2 {-0.36/21 28s} Kd6 {+0.42/24 3.3s} 29. Kc3 {-0.38/21 25s}
Ke5 {+0.59/21 3.3s} 30. Kc4 {-0.47/21 52s} a6 {+0.62/22 5.2s}
31. Kd3 {-0.62/21 51s} Bb5+ {+0.62/23 2.9s} 32. Kc2 {-0.72/21 55s}
a5 {+0.62/22 1.8s} 33. Kc3 {-1.02/21 45s} Ba6 {+0.62/23 3.7s}
34. Kd2 {-1.33/21 50s} Bb7 {+0.62/23 3.6s} 35. Nh5 {-0.63/20 44s}
Bc8 {+0.30/23 14s} 36. Nf4 {-0.89/21 44s} Bd7 {+0.30/22 8.3s}
37. Nh5 {-0.69/21 50s} Ke6 {+0.30/23 3.3s} 38. Kc3 {-0.41/20 34s}
Bb5 {+0.30/22 2.4s} 39. Kd4 {+0.74/18 28s} Ke7 {+0.30/22 1.6s}
40. Nf4 {+0.34/18 27s} Bc6 {+0.29/22 1.8s} 41. Nd5+ {+0.19/19 16s}
Bxd5 {+0.29/22 1.4s} 42. Kxd5 {+0.04/20 22s} Kd7 {+0.29/23 3.4s}
43. h5 {-0.01/22 44s} a4 {+0.38/21 2.0s} 44. h6 {+2.19/17 26s}
Kc7 {+0.38/23 1.6s} 45. Kc4 {+2.34/18 21s} Kc6 {+0.38/25 2.2s}
46. Kb4 {+2.32/19 17s} b5 {+0.45/25 2.4s} 47. f3 {-0.13/19 37s}
Kb6 {+0.45/25 2.4s} 48. Kc3 {-3.02/21 25s} Kc5 {+0.45/25 1.9s}
49. e3 {-2.65/24 22s} Kc6 {+0.45/25 5.3s} 50. Kc2 {-2.69/24 17s}
Kd6 {+0.45/24 3.9s} 51. Kd2 {-2.80/24 24s} Ke6 {+0.45/25 1.3s}
52. Kc2 {-2.74/25 29s} Kd7 {+0.45/24 1.3s} 53. Kd1 {-2.79/25 29s}
Kc7 {+0.45/25 2.9s} 54. Kc2 {-2.74/24 25s} Kb6 {+0.45/26 4.3s}
55. Kd3 {-2.81/24 21s} Ka6 {+0.45/26 2.5s} 56. Kd4 {-2.78/24 25s}
b4 {0.00/23 3.7s} 57. axb4 {+1.96/19 22s} f4 {0.00/25 3.1s}
58. gxf4 {-0.22/18 23s} Kb5 {0.00/25 1.1s} 59. Kc3 {-0.25/18 1.0s}
f5 {0.00/26 1.5s} 60. e4 {-0.30/18 1.0s} a3 {0.00/26 1.6s}
61. exf5 {-1.55/19 26s} a2 {0.00/28 1.2s} 62. Kb2 {-0.44/19 19s}
Kxb4 {0.00/30 1.8s} 63. Kxa2 {-0.24/20 15s} Kc4 {0.00/32 1.2s}
64. Ka3 {-0.08/20 19s} Kd4 {0.00/33 1.4s} 65. Kb4 {+0.04/20 21s}
f6 {0.00/34 2.2s} 66. Kb5 {+0.03/20 16s} Ke3 {0.00/34 1.9s}
67. Kc4 {+0.02/20 8.0s} Kxf4 {0.00/32 1.8s} 68. Kd5 {+0.21/19 22s}
Kxf5 {0.00/33 1.3s} 69. Kd6 {+0.14/21 19s} Kg5 {0.00/33 1.5s}
70. Ke6 {+0.06/23 33s} f5 {0.00/32 1.4s} 71. Ke5 {+0.03/23 15s}
f4 {0.00/33 1.6s} 72. Ke4 {+0.02/23 24s} Kxh6 {0.00/33 1.7s}
73. Kxf4 {+0.01/23 19s} Kg6 {0.00/33 2.3s} 74. Kg4 {0.00/23 36s}
h6 {0.00/34 3.4s} 75. Kh4 {0.00/22 42s} h5 {0.00/35 1.2s} 76. f4 {0.00/23 35s}
Kf5 {0.00/39 1.4s} 77. Kxh5 {0.00/21 24s}
Kxf4 {0.00/60 0.40s, Draw by insufficient mating material} 1/2-1/2
[Event "?"]
[Site "?"]
[Date "2018.05.05"]
[Round "2"]
[White "Naum_42_2_2"]
[Black "lczero_15_15"]
[Result "0-1"]
[BlackTimeControl "900+15"]
[FEN "rnbqk1nr/pp1p1ppp/4p3/2p5/1bPP4/2N5/PP2PPPP/R1BQKBNR w KQkq - 0 1"]
[PlyCount "124"]
[SetUp "1"]
[WhiteTimeControl "120+2"]
1. e3 {+0.27/15 3.8s} Nf6 {-0.14/19 23s} 2. Nf3 {+0.33/16 5.7s}
O-O {-0.09/20 22s} 3. a3 {+0.28/16 4.2s} Bxc3+ {-0.01/19 24s}
4. bxc3 {+0.20/6 0.011s} Qa5 {-0.02/20 14s} 5. Bd2 {+0.26/15 4.7s}
d6 {+0.03/19 19s} 6. Bd3 {+0.29/16 18s} e5 {0.00/20 22s} 7. dxe5 {+0.32/16 5.3s}
dxe5 {+0.40/19 18s} 8. Nxe5 {+0.20/17 3.5s} Re8 {+0.47/21 25s}
9. f4 {+0.53/17 3.8s} Nbd7 {+0.38/21 26s} 10. Nf3 {+0.47/16 5.5s}
Qc7 {+0.34/21 16s} 11. O-O {+0.68/15 3.3s} b6 {+0.37/20 13s}
12. Qe2 {+0.62/15 3.9s} Bb7 {+0.34/21 33s} 13. Rfd1 {+0.38/16 10s}
Re7 {+0.85/19 24s} 14. h3 {+0.63/14 4.0s} Nh5 {+0.98/20 36s}
15. Ng5 {+0.69/16 5.4s} Nhf6 {+1.00/20 22s} 16. Ra2 {+0.61/16 15s}
h6 {+1.16/20 31s} 17. Nf3 {+0.39/16 3.2s} Nh5 {+1.19/20 24s}
18. Qf2 {+0.47/16 3.4s} Rae8 {+1.19/21 22s} 19. Nh4 {+0.47/15 4.6s}
Ndf6 {+1.29/19 22s} 20. g4 {+0.27/14 3.6s} Qd7 {+1.07/20 31s}
21. Bc2 {-0.18/15 4.1s} Ne4 {+1.00/20 18s} 22. Qh2 {+0.09/16 5.7s}
Nhg3 {+1.20/20 28s} 23. Be1 {-0.18/15 11s} Qc6 {+1.40/20 8.3s}
24. Raa1 {-0.18/15 11s} Nxc3 {+2.78/19 26s} 25. Bxg3 {-1.18/14 1.6s}
Nxd1 {+2.54/20 39s} 26. Rxd1 {-2.17/16 10s} Rxe3 {+2.68/20 14s}
27. Rf1 {-2.17/16 5.4s} Qe6 {+3.17/20 38s} 28. Bd1 {-1.66/13 1.7s}
Qxc4 {+3.80/20 26s} 29. Nf5 {-3.17/15 5.7s} Rxa3 {+4.30/20 25s}
30. Qb2 {-2.62/14 2.4s} Qc3 {+4.62/19 17s} 31. Qxc3 {-2.76/16 3.6s}
Rxc3 {+4.93/20 13s} 32. Bh4 {-2.74/16 1.7s} Rxh3 {+5.02/19 33s}
33. Nd6 {-2.77/16 0.83s} Rh1+ {+4.82/19 20s} 34. Kf2 {-2.77/3 0.011s}
Rh2+ {+5.30/21 53s} 35. Kg1 {-2.76/17 0.96s} Rg2+ {+5.60/20 21s}
36. Kh1 {-2.76/3 0.011s} Rxg4+ {+5.96/20 3.0s} 37. Nxb7 {-2.77/18 1.0s}
Rxh4+ {+6.65/19 25s} 38. Kg2 {-2.77/19 6.8s} Re6 {+6.70/20 22s}
39. Kg3 {-2.77/17 5.2s} g5 {+7.97/19 22s} 40. fxg5 {-3.36/17 1.7s}
hxg5 {+7.88/19 16s} 41. Rf5 {-3.36/17 4.1s} Rd4 {+8.19/20 23s}
42. Bf3 {-3.65/17 1.7s} Kg7 {+9.46/19 30s} 43. Bd5 {-3.71/16 3.4s}
Rf6 {+10.36/20 51s} 44. Rxg5+ {-4.99/15 1.7s} Rg6 {+13.05/19 26s}
45. Rxg6+ {-4.99/17 0.90s} Kxg6 {+14.14/20 36s} 46. Ba2 {-5.14/18 3.4s}
Kg5 {+14.26/20 56s} 47. Bxf7 {-4.23/19 1.9s} Rd7 {+15.62/19 28s}
48. Nxc5 {-4.23/17 1.8s} bxc5 {+15.86/19 29s} 49. Be6 {-4.06/21 2.8s}
Rd3+ {+18.65/19 31s} 50. Kf2 {-4.06/20 2.0s} Kf4 {+21.17/19 23s}
51. Bc4 {-4.06/18 2.0s} Rd2+ {+22.93/19 34s} 52. Kf1 {-8.04/17 2.0s}
Ke3 {+24.07/20 38s} 53. Be6 {-8.82/18 1.8s} a5 {+22.85/20 51s}
54. Ke1 {-M22/19 0.96s} a4 {+30.00/19 34s} 55. Bf7 {-M20/3 0.011s}
a3 {+36.08/19 36s} 56. Be6 {-M18/3 0.011s} a2 {+41.13/19 35s}
57. Bxa2 {-M12/14 0.018s} Rxa2 {+44.63/19 22s} 58. Kd1 {-M10/3 0.011s}
Rh2 {+52.15/21 45s} 59. Kc1 {-M10/4 0.011s} Kd3 {+71.42/20 22s}
60. Kb1 {-M6/5 0.011s} Kc3 {+92.17/20 20s} 61. Ka1 {-M4/3 0.011s}
Kb3 {+116.59/21 27s} 62. Kb1 {-M2/3 0.012s} Rh1# {+127.12/19 22s, Black mates}
0-1
[Event "?"]
[Site "?"]
[Date "2018.05.05"]
[Round "3"]
[White "lczero_15_15"]
[Black "Naum_42_2_2"]
[Result "0-1"]
[BlackTimeControl "120+2"]
[FEN "rnbqkb1r/pp1p1ppp/4pn2/2p5/2P5/1P3N2/P2PPPPP/RNBQKB1R w KQkq - 0 1"]
[PlyCount "102"]
[SetUp "1"]
[WhiteTimeControl "900+15"]
1. g3 {0.00/20 25s} Nc6 {+0.17/17 4.7s} 2. Bg2 {0.00/20 19s} d5 {+0.20/17 3.7s}
3. cxd5 {0.00/20 13s} exd5 {+0.18/17 7.4s} 4. O-O {+0.03/20 14s}
g6 {+0.13/16 12s} 5. Bb2 {+0.37/19 19s} Bg7 {+0.06/16 8.1s}
6. Qc1 {+0.42/20 18s} O-O {-0.19/15 16s} 7. Qxc5 {+0.38/19 13s}
Re8 {-0.11/16 5.6s} 8. e3 {+0.29/20 24s} Bg4 {-0.22/15 3.9s}
9. Qb5 {+0.35/20 16s} Qd7 {-0.03/14 5.4s} 10. Bxf6 {+0.46/20 16s}
a6 {0.00/16 5.4s} 11. Qb6 {+0.56/19 18s} Bxf6 {+0.16/16 3.3s}
12. Nc3 {+0.51/20 9.2s} Re6 {+0.32/16 1.7s} 13. d4 {+0.57/19 25s}
Nxd4 {+2.40/16 2.7s} 14. Qxd4 {-2.85/19 22s} Bxd4 {+2.40/6 0.011s}
15. Nxd4 {-2.72/20 12s} Re5 {+2.40/17 2.7s} 16. Nce2 {-2.09/20 42s}
Bxe2 {+2.70/16 1.6s} 17. Nxe2 {-2.02/19 17s} Qb5 {+2.70/17 1.7s}
18. Bf3 {-2.11/20 40s} g5 {+2.70/18 3.0s} 19. Nd4 {-1.68/20 39s}
Qa5 {+2.70/19 4.1s} 20. Rfd1 {-1.73/20 41s} h6 {+2.70/18 2.9s}
21. h3 {-1.76/20 41s} Rf8 {+2.70/17 4.1s} 22. Kg2 {-1.62/20 44s}
Kg7 {+2.70/15 4.3s} 23. Rac1 {-1.31/20 46s} Kh8 {+2.70/14 3.2s}
24. g4 {-1.24/20 41s} Qxa2 {+2.70/17 14s} 25. Rc7 {-0.82/19 26s}
b5 {+2.69/15 12s} 26. Rdc1 {-0.77/20 48s} a5 {+1.82/12 2.6s}
27. Rb7 {-0.87/20 36s} a4 {+1.59/13 2.3s} 28. Rxb5 {-0.95/19 28s}
Qb2 {+1.38/15 3.3s} 29. Rc2 {-0.56/19 27s} a3 {+1.40/16 2.2s}
30. Ra5 {-0.96/20 18s} Rfe8 {+1.30/16 4.5s} 31. Ra7 {-1.02/19 28s}
R5e7 {+1.21/15 2.9s} 32. Ra6 {-1.09/20 23s} Kg8 {+1.21/14 4.0s}
33. Bxd5 {-1.53/19 29s} Rd7 {+2.55/14 1.9s} 34. Rc5 {-5.10/20 46s}
a2 {+4.48/13 2.6s} 35. Bc6 {-5.49/20 28s} Rxd4 {+4.93/14 1.2s}
36. Bxe8 {-5.70/20 14s} Rd1 {+4.94/16 1.6s} 37. Rf5 {-6.36/19 43s}
a1=Q {+5.02/17 1.9s} 38. Bxf7+ {-7.92/20 41s} Kg7 {+5.10/16 1.5s}
39. Bc4 {-9.42/20 33s} Rg1+ {+18.13/10 2.6s} 40. Kf3 {-11.31/20 18s}
Qd1+ {+22.12/10 2.6s} 41. Ke4 {-13.41/20 17s} Rxg4+ {+20.29/15 2.8s}
42. hxg4 {-15.20/20 16s} Qxg4+ {+20.29/6 0.011s} 43. Rf4 {-15.88/20 13s}
gxf4 {+39.05/15 3.0s} 44. Ra7+ {-16.05/20 12s} Kg6 {+M21/15 0.27s}
45. Ra6+ {-19.22/19 26s} Kg5 {+M19/11 0.015s} 46. Ra5+ {-23.49/20 23s}
Kh4 {+M17/5 0.011s} 47. Rf5 {-26.00/20 33s} Qc2+ {+M15/4 0.011s}
48. Bd3 {-26.59/19 31s} Qc6+ {+M7/3 0.011s} 49. Kd4 {-43.01/20 51s}
Qg7+ {+M5/3 0.011s} 50. Re5 {-62.18/20 19s} Qa7+ {+M3/3 0.011s}
51. Rc5 {-90.87/12 0.078s} Qaxc5# {+M1/3 0.011s, Black mates} 0-1
[Event "?"]
[Site "?"]
[Date "2018.05.05"]
[Round "4"]
[White "Naum_42_2_2"]
[Black "lczero_15_15"]
[Result "1/2-1/2"]
[BlackTimeControl "900+15"]
[FEN "rnbqkb1r/pp1p1ppp/4pn2/2p5/2P5/1P3N2/P2PPPPP/RNBQKB1R w KQkq - 0 1"]
[PlyCount "118"]
[SetUp "1"]
[WhiteTimeControl "120+2"]
1. g3 {-0.11/16 4.7s} d5 {-0.02/20 29s} 2. Bg2 {-0.24/16 3.8s} d4 {+0.15/20 19s}
3. O-O {-0.28/16 5.3s} Nc6 {+0.24/20 14s} 4. e3 {-0.21/16 4.0s}
e5 {+0.27/20 13s} 5. exd4 {-0.06/16 4.7s} exd4 {+0.30/21 29s}
6. Na3 {-0.04/17 4.3s} Be7 {+0.38/20 18s} 7. d3 {-0.31/16 4.6s}
O-O {+0.36/20 19s} 8. Re1 {-0.23/16 3.4s} h6 {+0.33/21 30s}
9. Bf4 {0.00/16 2.9s} Nh5 {+0.32/21 30s} 10. Bc1 {-0.02/16 5.1s}
Nf6 {+0.30/21 20s} 11. Bf4 {0.00/18 2.8s} Bd6 {+0.32/20 7.0s}
12. Qd2 {-0.09/16 4.0s} Bf5 {+0.42/20 15s} 13. Ne5 {+0.08/15 3.3s}
Bxe5 {+0.57/20 10s} 14. Bxe5 {-0.11/17 7.6s} Re8 {+0.72/20 9.0s}
15. Bxf6 {-0.21/16 5.6s} Qxf6 {+0.83/20 8.2s} 16. Nb5 {-0.13/17 3.0s}
Re7 {+0.84/20 13s} 17. Rxe7 {-0.20/16 1.8s} Qxe7 {+1.09/19 17s}
18. Bxc6 {-0.19/17 2.3s} bxc6 {+0.97/20 23s} 19. Re1 {-0.23/18 2.1s}
Qd7 {+0.96/20 7.9s} 20. Na3 {-0.30/18 4.8s} a5 {+0.95/20 9.4s}
21. Nc2 {-0.23/17 3.6s} a4 {+0.98/21 45s} 22. b4 {-0.21/17 12s}
cxb4 {+1.08/20 11s} 23. Nxb4 {-0.27/18 5.4s} Re8 {+1.11/20 16s}
24. Rxe8+ {-0.14/16 9.9s} Qxe8 {+1.20/20 8.4s} 25. h4 {-0.16/17 3.7s}
c5 {+1.10/20 33s} 26. Nd5 {-0.04/16 1.6s} a3 {+1.09/21 34s}
27. Qd1 {0.00/17 2.2s} Qe6 {+1.08/20 32s} 28. Kh1 {0.00/16 3.8s}
g5 {+1.46/19 33s} 29. g4 {-0.07/17 2.1s} Bh7 {+0.74/20 43s}
30. h5 {0.00/18 2.7s} Kg7 {+0.67/22 47s} 31. Kg2 {0.00/18 1.3s}
Qc6 {+0.73/22 57s} 32. Qd2 {0.00/18 3.2s} Qe6 {+0.78/21 41s}
33. Qd1 {0.00/20 1.4s} Qe5 {+0.60/22 34s} 34. Kf1 {0.00/19 2.5s}
Qh2 {+0.47/22 39s} 35. Qf3 {0.00/21 2.2s} Qd6 {+0.44/24 57s}
36. Kg2 {0.00/21 2.3s} Qe5 {+0.41/24 27s} 37. Qd1 {0.00/20 1.2s}
Qb8 {+0.42/24 21s} 38. Qe2 {0.00/20 3.6s} Qd6 {+0.40/22 29s}
39. Qf3 {0.00/21 2.7s} Qe5 {+0.32/22 21s} 40. Qd1 {0.00/22 2.5s}
Qe6 {+0.33/23 19s} 41. Qf3 {0.00/20 1.6s} Qd6 {+0.51/21 19s}
42. Kf1 {0.00/22 5.1s} Qe5 {+0.20/23 51s} 43. Qe2 {0.00/22 2.8s}
Qd6 {+0.18/23 57s} 44. Qf3 {0.00/22 2.3s} Qe5 {+0.11/22 43s}
45. Qe2 {0.00/23 2.7s} f6 {+0.18/20 41s} 46. f3 {0.00/19 2.0s}
Qd6 {+0.29/20 6.1s} 47. Kg2 {-0.15/19 5.0s} f5 {+0.60/21 28s}
48. Ne7 {-0.07/19 2.6s} Kf6 {+0.56/22 33s} 49. Nxf5 {0.00/20 3.0s}
Qe6 {+0.44/20 39s} 50. Qe4 {0.00/20 3.0s} Bxf5 {+0.05/21 22s}
51. gxf5 {0.00/20 1.4s} Qxf5 {+0.04/22 21s} 52. Qc6+ {0.00/19 1.4s}
Kg7 {0.00/24 46s} 53. Qc7+ {0.00/19 1.5s} Kf6 {0.00/24 25s}
54. Qb6+ {0.00/20 3.3s} Qe6 {0.00/20 25s} 55. Qxc5 {0.00/21 2.2s}
Qe2+ {0.00/19 1.1s} 56. Kg3 {0.00/21 2.8s} Qe1+ {0.00/19 1.1s}
57. Kg2 {0.00/23 1.5s} Qe2+ {0.00/19 1.1s} 58. Kg3 {0.00/22 5.3s}
Qe1+ {0.00/18 2.3s} 59. Kg2 {0.00/24 11s}
Qe2+ {0.00/18 25s, Draw by 3-fold repetition} 1/2-1/2[/pgn]
LC0 lost as White the game 3 from an equal position by a tactical blunder:
[D]r5k1/1p1q1p1p/pQn1rbp1/3p4/6b1/1PN1PNP1/P2P1PBP/R4RK1 w - - 0 13[/D]
Here LC0 played 13.d4, losing a piece. But generally LC0 is not hideously weak tactically in these conditions. Losing 0.5 points by a tactical blunder in 4 games is about 100 Elo points. Sure, only 4 games, but it was interesting for me to see if it can stand a 3000 Elo engine.
-
- Posts: 228
- Joined: Sun Mar 12, 2006 3:11 pm
Re: LCZero: Progress and Scaling. Relation to CCRL Elo
Despite the obvious improvements in general strength, tactics seem not to be proportionately better - if at all. I still see lczero falling for simple 1-3 ply combos and, as in your position, 1 ply discovered attacks - although it is a few moves to their resolution of course.
No idea if this will be an enduring feature of the NN or whether it magically improves dramatically. Would be interesting to know how good A0 was at tactics. Clearly it did not suffer in an obvious way from the evidence of the published SF games, but it would be nice to know its performance on standard test suites, for example (history plane problems or not).
No idea if this will be an enduring feature of the NN or whether it magically improves dramatically. Would be interesting to know how good A0 was at tactics. Clearly it did not suffer in an obvious way from the evidence of the published SF games, but it would be nice to know its performance on standard test suites, for example (history plane problems or not).
-
- Posts: 1346
- Joined: Sat Apr 19, 2014 1:47 pm
Re: LCZero: Progress and Scaling. Relation to CCRL Elo
The interesting thing is to get a 3000 elo engines "weak" in tactics. That's means being 3000 elo is still "weak" in chess compared to top strenght.
And that could mean Lczero will improve a lot without problem for a while.
And that could mean Lczero will improve a lot without problem for a while.
-
- Posts: 10948
- Joined: Wed Jul 26, 2006 10:21 pm
- Full name: Kai Laskos
Re: LCZero: Progress and Scaling. Relation to CCRL Elo
Yes, tactics is just a bit better than much earlier nets on tactical testsuites. I tested a day or two ago on ECM200 easy tactical suite. It still performs (in conditions similar to the shown match) at 2000 Elo level standard AB engine. I should have said "LC0 is not hideously weak in tactics occurring often in games". The ECM200 easy tactical positions occur say in 1 out of 4 games (as it happened in the above match). Fruit 2.1 (2700) already solves 90%+ of the positions, and SF9 (3500) practically all. So, decisive tactical shots Fruit 2.1 doesn't solve and SF9 solves, are very rarely occurring in real games, maybe in 1 out of 50 or 100 games. Therefore, A0 might have been similar in tactical strength to Fruit 2.1, never showing its tactical weakness compared to SF in 100 games.frankp wrote:Despite the obvious improvements in general strength, tactics seem not to be proportionately better - if at all. I still see lczero falling for simple 1-3 ply combos and, as in your position, 1 ply discovered attacks - although it is a few moves to their resolution of course.
No idea if this will be an enduring feature of the NN or whether it magically improves dramatically. Would be interesting to know how good A0 was at tactics. Clearly it did not suffer in an obvious way from the evidence of the published SF games, but it would be nice to know its performance on standard test suites, for example (history plane problems or not).
-
- Posts: 2567
- Joined: Fri Nov 26, 2010 2:00 pm
- Location: Czech Republic
- Full name: Martin Sedlak
Re: LCZero: Progress and Scaling. Relation to CCRL Elo
Interesting, I just finished a 100 game match 40 moves 1 minute between Leela 0.8 net 245 (4 threads + 980Ti) vs Cheng 4.39 4 threads and Leela scored only 36%, which is a bit less than I expected (could be noise as 100 games is not enough).
It's obvious the GPU version didn't seem to gain anything going from 2 to 4 CPU threads; last time I checked the speed difference between 1 CPU thread and 1 CPU+GPU was ~23x.
Judging from what I saw, many games Leela lost were actually due to tactical blunders. Always the same story, Leela got a slight advantage initially, then blundered and lost the games. This happened in "won/drawn" games.
(moving into mate in 4, losing a piece and so on; on the bright side I didn't see any blunders due to trivial forks)
So tactics seems to be the Achilles heel of Leela, at least at this fast TC.
It's obvious the GPU version didn't seem to gain anything going from 2 to 4 CPU threads; last time I checked the speed difference between 1 CPU thread and 1 CPU+GPU was ~23x.
Judging from what I saw, many games Leela lost were actually due to tactical blunders. Always the same story, Leela got a slight advantage initially, then blundered and lost the games. This happened in "won/drawn" games.
(moving into mate in 4, losing a piece and so on; on the bright side I didn't see any blunders due to trivial forks)
So tactics seems to be the Achilles heel of Leela, at least at this fast TC.
-
- Posts: 1470
- Joined: Mon Apr 23, 2018 7:54 am
Re: LCZero: Progress and Scaling. Relation to CCRL Elo
jp wrote:What does the graph of net size vs. number of playouts (fixed time) look like?Laskos wrote: Nah, this is self-play with fixed number of playouts. But the nets became slower. On GPU, with the introduction of ID227 larger 15x192 net, maybe there was some improvement, 20-30 Elo points or so, no more, because the nets were much slower.
Are you saying going from 10x128 to 15x192 was just a 3-4% slowdown in the number of playouts? Doesn't sound like much.Laskos wrote: Taking into account a 3-4% slowdown of the net, I guess that the improvement since ID227 is in the range of 25 Elo points, maybe even a bit more. Have yet to test at fixed time control.
-
- Posts: 10948
- Joined: Wed Jul 26, 2006 10:21 pm
- Full name: Kai Laskos
Re: LCZero: Progress and Scaling. Relation to CCRL Elo
Not, from same 15x192 nets, ID227 to later, but the speed seems to have been stable later. From smaller to bigger net (10x128 to 15x192), there is a factor of maybe 2 in speed for GPU, and maybe 3 for CPU.jp wrote:jp wrote:What does the graph of net size vs. number of playouts (fixed time) look like?Laskos wrote: Nah, this is self-play with fixed number of playouts. But the nets became slower. On GPU, with the introduction of ID227 larger 15x192 net, maybe there was some improvement, 20-30 Elo points or so, no more, because the nets were much slower.Are you saying going from 10x128 to 15x192 was just a 3-4% slowdown in the number of playouts? Doesn't sound like much.Laskos wrote: Taking into account a 3-4% slowdown of the net, I guess that the improvement since ID227 is in the range of 25 Elo points, maybe even a bit more. Have yet to test at fixed time control.
-
- Posts: 251
- Joined: Sat Dec 02, 2006 10:47 pm
- Location: Toronto
- Full name: Peter Kasinski
Re: LCZero: Progress and Scaling. Relation to CCRL Elo
The interesting change is that in going from v 0.7 to 0.8 the 1080 ti went from 836.9 GFLOPs to 1,134.3 GFLOPs after tuning (both in silent mode of the GPU I am using).
It would appear that the new network size allows Leela to take better advantage of this GPU.
PK
It would appear that the new network size allows Leela to take better advantage of this GPU.
PK
-
- Posts: 10948
- Joined: Wed Jul 26, 2006 10:21 pm
- Full name: Kai Laskos
Re: LCZero: Progress and Scaling. Relation to CCRL Elo
Still consistent with about 3000 CCRL 40/4' rating list. Add to your result some 30 Elo points in better scaling from your 1' to CCRL 2' (or 4' with old core).mar wrote:Interesting, I just finished a 100 game match 40 moves 1 minute between Leela 0.8 net 245 (4 threads + 980Ti) vs Cheng 4.39 4 threads and Leela scored only 36%, which is a bit less than I expected (could be noise as 100 games is not enough).
Interesting. I saw that in 1 out of 4 (lol) games at twice the time control, but you have anyway worse score, so this weakness should be more apparent. Yes, tactics seems to not improve a lot, in fact it is now (ID245 or so) not much better than ID104 or so on decisive tactical blunders tests like ECM200. I will check today how it scales with time control on ECM200, my past result (long ago) was not that good again.
It's obvious the GPU version didn't seem to gain anything going from 2 to 4 CPU threads; last time I checked the speed difference between 1 CPU thread and 1 CPU+GPU was ~23x.
Judging from what I saw, many games Leela lost were actually due to tactical blunders. Always the same story, Leela got a slight advantage initially, then blundered and lost the games. This happened in "won/drawn" games.
(moving into mate in 4, losing a piece and so on; on the bright side I didn't see any blunders due to trivial forks)
So tactics seems to be the Achilles heel of Leela, at least at this fast TC.