Lc0 DAG-tt Engine

[Eduard]

Lc0 DAG-tt Engine

In another forum you can read that the new Lc0 DAG-tt engine plays great. Dragon was defeated in the blitz tournament on Chess.com. Can someone explain what's new and special about this new Lc0 engine?

On PlayChess.com Lc0 DAG-tt won the blitz tournament today with a T78 net and on a laptop with RTX 3060m.



PGN with RTX 3060m:

[pgn][Event "Friday 1 July Blitz Tour 5+0, 15 Rounds"]
[Site "Machines A"]
[Date "2022.07.01"]
[Round "6"]
[White "Python, Lc0 v0.30.0-dev+git"]
[Black "Hisoka798, Stockfish 200622"]
[Result "1-0"]
[WhiteElo "2819"]
[BlackElo "2824"]
[PlyCount "151"]
[EventDate "2022.07.01"]
[EventType "swiss (blitz)"]

1. d4 {0.00/0 0} d5 {0} 2. c4 {0.00/0 0} e6 {0} 3. cxd5 {0.00/0 0} exd5 {0} 4.
Nc3 {0.00/0 0} c6 {0} 5. Bf4 {0.00/0 0} Bd6 {0} 6. Bxd6 {0.00/0 0} Qxd6 {0} 7.
e3 {0.00/0 0} Bf5 {0} 8. g4 {0.00/0 0} Be6 {0} 9. h4 {0.00/0 0} Nd7 {14} 10.
Nh3 {0.00/0 0} O-O-O {4} 11. Nf4 {0.00/0 0} Kb8 {4} 12. Qe2 {0.08/10 14} h6 {
0 (g6)} 13. f3 {0.13/13 8} f5 {3} 14. Qh2 {0.19/21 4} Re8 {17} 15. Ng6 {
0.23/27 2} Qxh2 {3} 16. Rxh2 {0.22/24 2} Rh7 {3} 17. Bd3 {0.20/22 22} fxg4 {6}
18. h5 {0.22/22 2} Ngf6 {27} 19. Rh4 {0.27/21 2} gxf3 {20} 20. Kf2 {0.29/19 2}
Kc7 {2 (Sb6)} 21. Rg1 {0.31/13 19} Kd6 {4} 22. Nb1 {0.33/26 2} c5 {39 (Lf7)}
23. Nd2 {0.52/13 10} c4 {19 (cxd4)} 24. Bc2 {0.62/13 7} Rc8 {0 (b5)} 25. Nxf3 {
0.64/14 6} Rc6 {5 (b5)} 26. Nfe5 {0.83/13 8} Ra6 {8} 27. Ra1 {0.90/13 2} b5 {
6 (Sb8)} 28. a4 {1.05/16 8} b4 {3} 29. a5 {1.12/17 3} b3 {26} 30. Bb1 {1.21/21
2} Nb6 {20 (Kc7)} 31. Ke2 {1.42/16 14} Nc8 {1} 32. Kd2 {1.38/15 2} Kc7 {5} 33.
Kc3 {1.48/16 2} Bg8 {1 (Td6)} 34. Rh1 {1.82/15 12} Rd6 {4} 35. Nf4 {1.89/16 4}
Ne7 {0} 36. Bxh7 {2.04/15 7} Bxh7 {3} 37. Kb4 {2.25/14 16} Bc2 {1 (Sf5)} 38.
Rhg1 {2.68/13 6} Nf5 {0} 39. Rg6 {2.78/13 3} Kb8 {12 (Se8)} 40. Rag1 {4.36/12 8
} Be4 {1 (Kc8)} 41. Ng4 {6.94/11 8} Kb7 {3 (Se8)} 42. Nxf6 {6.27/13 6} gxf6 {5}
43. Rg8 {6.78/13 2} Rc6 {3} 44. Ne2 {7.16/14 5} c3 {1} 45. Nxc3 {7.34/14 3}
Nxe3 {0} 46. Nxe4 {7.69/14 9} Nc2+ {1} 47. Kxb3 {6.77/15 12} Nxd4+ {0} 48. Ka2
{7.53/13 5} dxe4 {0} 49. Re8 {8.80/11 2} Rc5 {1 (f5)} 50. Rxe4 {17.16/11 4}
Rxa5+ {0} 51. Kb1 {16.24/1 0} Nc6 {0} 52. Rh4 {17.27/9 4} Rf5 {0 (Se5)} 53. Rg6
{17.54/9 4} Rf1+ {0} 54. Kc2 {16.96/9 5} Ne7 {0 (Tf2+)} 55. Rxh6 {20.07/8 8}
Nf5 {1} 56. Rh7+ {23.42/9 2} Kb6 {0 (Kc6)} 57. Rh3 {26.18/8 2} Nd4+ {0} 58. Kd2
{22.46/7 2} Rf5 {2 (Se6)} 59. Rh8 {31.67/6 2} Nf3+ {0} 60. Kc2 {27.70/6 1} Ng5
{1} 61. Rh1 {30.35/6 1} Rc5+ {2} 62. Kb1 {34.02/6 1} Nf3 {2 (Se4)} 63. b4 {
43.22/6 1} Rf5 {1 (Tc4)} 64. h6 {39.06/5 1} Ng5 {0} 65. h7 {32.23/6 1} Nf7 {
2 (Sxh7)} 66. Rf8 {92.26/5 1} Nd6 {1 (Sh8)} 67. h8=Q {118.18/4 0} Rf4 {0} 68.
Rb8+ {112.31/4 0} Kc6 {0} 69. Rd1 {116.12/4 0} Rf1 {0 (Txb4+)} 70. Rxf1 {
124.13/5 0} Kd5 {0} 71. Qxf6 {124.33/4 0} a6 {0 (Se4)} 72. Rf4 {#107/4 0} Ne4 {
0} 73. Rxe4 {#106/1 0} Kxe4 {0} 74. Kc2 {#105/2 0} Kd5 {0 (a5)} 75. Kd3 {
#103/1 0} a5 {0} 76. Rb5# {#1/1 0} 1-0[/pgn]

[dkappe]

Dragon was defeated in the blitz tournament on Chess.com.

Only 210 of 480 games played and the elo gap continues to shrink.

I addressed this new search in my post, A Leela improvement that won't help Stockfish?

The MCTS algorithms build a search tree and backup data to the root node. The problem is that if a position appears several times in a tree, it doesn’t benefit from the search information of those duplicates. ClassicAra had already used a Monte Carlo Graph Search algorithm which allowed the engine to deal with transpositions. Leela has now implemented something similar.

For more technical detail and info on what a Directed Acyclic Graph (DAG) is, see https://ojs.aaai.org/index.php/ICAPS/article/view/15952

[Eduard]

Dragon has no chance on PlayChess (with normal hardware), Lc0 can always win tournaments, even with weak hardware.

[Krzysztof Grzelak]

Please see here (I also play dag-tt) - forum3/viewtopic.php?f=6&t=77071&start=30

[AdminX]

I had not updated or used Leela in a while, but this post and the one by Dkappe got me interested to take a look. So I downloaded the latest dev build of it along with a 78 net and played some games on Lichess using my bot account (NeuralAbyss). Rating did start to climb. I enjoyed the idea leela played in the mid game here.

https://lichess.org/fNAwP8Nn#91

[pgn]
[Event "Rated Blitz game"]
[Site "lichess.org"]
[Date "2022.07.02"]
[Round "?"]
[White "NeuralAbyss"]
[Black "beckapplebot"]
[Result "1/2-1/2"]
[ECO "E06"]
[WhiteElo "2506"]
[BlackElo "2610"]
[Annotator "lichess.org"]
[PlyCount "122"]
[EventDate "2022.??.??"]
[EventType "blitz"]
[TimeControl "300+3"]

1. d4 {6} Nf6 {6} 2. c4 {7} e6 {0} 3. g3 {12} d5 {0} 4. Bg2 {0} Be7 {0} 5. Nf3
{0 E06 Catalan Opening: Closed} O-O {1} 6. O-O {6} dxc4 {0} 7. Qc2 {5} b5 {0}
8. a4 {0} b4 {0} 9. Nfd2 {1} c6 {0} 10. Nxc4 {0} Qxd4 {0} 11. Rd1 {9} Qc5 {0}
12. Be3 {5} Qh5 {0} 13. Nbd2 {7} Ng4 {0} 14. Nf3 {14} Nxe3 {0} 15. Nxe3 {1} a5
{0} 16. Nd4 {28} Ba6 {1} 17. Rac1 {1} Ra7 {0} 18. Qe4 {18} Bb7 {0} 19. g4 {3}
Qg5 {22} 20. h4 {11} Qxh4 {1} 21. Nef5 {1} Qf6 {39} 22. Qe3 {1} Bc8 {10} 23.
Nb3 {0} Rc7 {15} 24. Nxe7+ {1} Qxe7 {9} 25. g5 {39} Ba6 {0} 26. f4 {16} Rd7 {1}
27. Rxd7 {12} Nxd7 {2} 28. Kf2 {11} c5 {1} 29. Nxc5 {5} Nxc5 {7} 30. Rxc5 {2}
Bb7 {9} 31. Rxa5 {7} Qd8 {17} 32. Qc5 {11} Bxg2 {10} 33. Kxg2 {0} Qd2 {10} 34.
Qe5 {6} Qc2 {10} 35. Kf2 {15} Qb3 {0} 36. Qd4 {29} g6 {1} 37. Kg2 {12} Re8 {10}
38. Qe5 {13} Qc2 {16} 39. Kf2 {2} Qb3 {7} 40. Qd4 {19} Rf8 {1} 41. Ra6 {14} Qh3
{13} 42. Rd6 {1} e5 {11} 43. Qxe5 {1} Qh2+ {12} 44. Ke3 {15} Qg1+ {0} 45. Kd2 {
7} Rc8 {4} 46. Qe7 {4} Qc1+ {14} 47. Kd3 {1} Qd1+ {10} 48. Ke3 {0} Qc1+ {10}
49. Rd2 {33} Qg1+ {10} 50. Kf3 {3} Qf1+ {10} 51. Ke3 {7} Qh3+ {8} 52. Ke4 {10}
Rc4+ {10} 53. Rd4 {1} Qf5+ {14} 54. Ke3 {1} Qh3+ {7} 55. Ke4 {3} Qf5+ {6} 56.
Ke3 {2} Qh3+ {6} 57. Kf2 {7} Qh2+ {9} 58. Kf3 {0} Qh1+ {8} 59. Kf2 {1} Qh2+ {9}
60. Kf3 {3} Qh3+ {5} 61. Kf2 {13} Qh2+ {5 Normal The game is a draw.} 1/2-1/2
[/pgn]

[dkappe]

From a game pair win vs DAG-tt at CCC. Double exchange sac and three white queens at the end. The reverse hold might have been even better.

[pgn] [Event "CCC 17 Blitz: Challenger (5|5)"]
[Site "?"]
[Date "2022.07.02"]
[Round "1"]
[White "Dragon"]
[Black "Lc0"]
[Result "1-0"]
[ECO "A80"]
[GameDuration "00:24:33"]
[GameEndTime "2022-07-02T11:50:06.321 PDT"]
[GameStartTime "2022-07-02T11:25:32.452 PDT"]
[Opening "Dutch"]
[PlyCount "211"]
[TimeControl "300+5"]

1. d4 f5 2. Nc3 d5 3. Bf4 c6 4. e3 Nd7 5. Bd3 e6 6. Nf3 Be7 7. O-O Ngf6 8. Ne2
O-O 9. c4 Nh5 10. h3 Nxf4 11. exf4 Nf6 12. c5 a5 13. a3 Bd7 14. b4 Be8 15. Qc2
Qc7 16. Rfe1 Bh5 17. Ng5 Bf7 18. b5 Ne4 19. Nxf7 Rxf7 20. b6 Qb8 21. Rf1 Qf8 22.
Rac1 g6 23. f3 Nf6 24. h4 Nd7 25. g3 Bf6 26. Kf2 Qg7 27. Qc3 Rff8 28. Rh1 h5 29.
Bc2 Rfe8 30. Kg2 Be7 31. Qe3 Bf8 32. Rhe1 Qh6 33. Ng1 Bg7 34. Rb1 Kf7 35. Nh3
Qh8 36. Red1 Bh6 37. Ng5+ Kg8 38. Ba4 Bg7 39. Nh3 Bh6 40. Qf2 Kf7 41. Qb2 Kg8
42. Qc3 Bf8 43. Rd3 Qf6 44. Ng1 Bg7 45. Qd2 Bh6 46. Qc3 Bf8 47. Bc2 Qh8 48. Rbd1
Qg7 49. Re3 Qf6 50. a4 Qe7 51. Rde1 Bg7 52. Ne2 Bh6 53. Nc1 e5 54. Rxe5 Nxe5 55.
dxe5 Red8 56. Nb3 Kh8 57. Qd4 Qe6 58. Bd3 Bf8 59. Rc1 Qe7 60. Rc2 Rd7 61. Rb2
Rdd8 62. Re2 Qe6 63. Rc2 Qg8 64. Be2 Be7 65. Rc1 Kg7 66. Bd3 Kh8 67. g4 Rf8 68.
gxf5 gxf5+ 69. Kh3 Qe6 70. Rg1 Rg8 71. Rg5 Bxg5 72. hxg5 Rad8 73. Bc2 Rgf8 74.
Nxa5 Qe7 75. Kh4 Kh7 76. Nb3 Kg6 77. a5 Qd7 78. a6 bxa6 79. Na5 Rc8 80. Ba4 Kh7
81. Nc4 a5 82. Nd6 Rcd8 83. Qd3 Kg6 84. b7 d4 85. Qa6 d3 86. Bxc6 Qe6 87. Bd5
Rxd6 88. cxd6 d2 89. Qd3 Qd7 90. Qxd2 Qb5 91. e6 Qb1 92. e7 Qh1+ 93. Kg3 h4+ 94.
Kf2 Qh2+ 95. Ke3 Qg1+ 96. Kd3 Rb8 97. d7 Qf1+ 98. Kc2 Qb5 99. d8=Q Qc5+ 100. Qc3
Qf2+ 101. Kb3 a4+ 102. Kb4 Rxb7+ 103. Bxb7 Qb2+ 104. Qxb2 Kh7 105. Qbh8+ Kg6
106. e8=Q# 1-0[/pgn]

[Eduard]

In such tournaments and also in the TCEC, not only the best variants are played, many curious positions arise. That's on purpose. But I see that Lc0 still has a 10-point lead.

[cc2150dx]

Has anyone tested/experimented with Lc0 dag-bord2 ?

https://github.com/Etcaqab/lc0/tree/dag-bord2

[AdminX]

Has anyone tested/experimented with Lc0 dag-bord2 ?

https://github.com/Etcaqab/lc0/tree/dag-bord2

I pulled down the last one about one hour ago (see link), It works did not play allot of games with it, just one or two on lichess.

https://ci.appveyor.com/project/Etcaqab/lc0

[cc2150dx]

Thanks Ted. I tried it myself and it seems to perform fairly well.