Allie 0.1

[jmartus]

Allie vs Gull, TCEC Division 3:

[d]1r4k1/r1b2pp1/1pQp3p/pR2p2P/P3PqP1/1BP2P2/KP6/3R4 w - - 64 77
AllieStein played nice Rxd6! and won.

Analysis by Cfish 12-03-19:
77. Qd7 Rf8 78. Rf1 Kh8 79. Rd5 Bb8 80. Qf5 Qg5 81. Rdd1 Re7 (+1.60)

g4-g5 also wins

FEN: 1r4k1/r1b2pp1/1pQp3p/pR2p2P/P3PqP1/1BP2P2/KP6/3R4 w - - 0 1

SugaR Tactical_R

32/44 01:04 420,458k 6,552k +3.02 Rb5-d5 Kg8-h8 Qc6-b5 Rb8-f8 Rd1-f1 f7-f6 Rd5-d1 Ra7-a8 Rd1-e1 Ra8-d8 Qb5-c6 Bc7-b8 Bb3-e6 Rf8-e8 Be6-f5 Re8-e7 Qc6-b5 Bb8-c7 Re1-e2 Rd8-b8 Qb5-d5 Rb8-d8 Ka2-b1 Qf4-g3 Qd5-b5 Rd8-b8 Re2-d2 Qg3-f4 Qb5-e2 b6-b5 a4xb5 Rb8xb5
32/51 01:04 420,458k 6,552k +3.71 g4-g5 Qf4xg5 f3-f4 Qg5-f6 Rb5-d5 Rb8-f8 Qc6-b5 Bc7-b8 Qb5xb6 Kg8-h7 Rd5xd6 Bb8xd6 f4xe5 Bd6-c5 Qb6xa7 Bc5xa7 e5xf6 g7xf6 Rd1-d7 Ba7-c5 Bb3xf7 Kh7-h8 Rd7-c7 Bc5-d6 Rc7-a7 Bd6-b8 Ra7-a8 Rf8xf7 Ra8xb8+ Kh8-g7 Rb8-b5 Rf7-a7 Rb5-f5 Ra7-a8 Ka2-a3 Kg7-f7 b2-b4 a5xb4+ c3xb4 Kf7-e6 a4-a5

[Eduard]

Thanks. Can you post a analysis after g5 hxg5?

[jdart]

AllieStein played nice Rxd6! and won.

That is an interesting move. I gave Houdini 6 a long think on 6 cores and it likes Rf1 or Rg1 (score +1.45 at depth 42). But if I force Rxd6 it starts to get a higher score. The mainline looks to be 77. Rxd6 Bxd6 78. Qxd6 Rc8 79. Rxb6, with Black giving back a pawn. It is still not seeing a win.

--Jon

[Eduard]

[Event "TCEC Season 15 - Division 3"]
[Site "http://tcec.chessdom.com"]
[Date "2019.03.16"]
[Round "2.1"]
[White "AllieStein v0.2-n4"]
[Black "Gull 3"]
[Result "1-0"]
[TimeControl "1800+5"]
[BlackElo "3300"]
[ECO "B56"]
[GameDuration "01:16:35"]
[GameEndTime "2019-03-17T01:06:27.941 W. Europe Standard Time"]
[GameStartTime "2019-03-16T23:49:52.126 W. Europe Standard Time"]
[Opening "Sicilian"]
[PlyCount "208"]
[Termination "adjudication"]
[TerminationDetails "TCEC win rule"]
[WhiteElo "3233"]

1. e4 c5 2. Nf3 d6 3. d4 cxd4 4. Nxd4 Nf6 5. Nc3 Nc6 6. f3 e5 7. Nb3 Be7 8. Be3
O-O 9. Qd2 a5 10. Bb5 Be6 11. O-O-O Na7 12. Bxa7 Rxa7 13. Kb1 Qb6 14. Qe2 Raa8
15. a4 Rfc8 16. Rd3 Rc7 17. g4 Nd7 18. Nd5 Bxd5 19. Rxd5 Nf8 20. Nd2 Ne6 21.
Nc4 Qa7 22. Qd1 Nd4 23. Na3 Rac8 24. c3 Ne6 25. Rd2 Bg5 26. Rd3 Be7 27. Rd2 Bg5
28. Rd5 Be7 29. h4 Nf4 30. Rd2 h6 31. Qb3 Qe3 32. Qd1 Qa7 33. Ka2 Kh8 34. Nc2
Qb6 35. Ne1 Qe3 36. Nd3 Nxd3 37. Rxd3 Qf4 38. Qg1 Rc5 39. Qd1 R5c7 40. Qb3 Kg8
41. h5 Qg3 42. Rf1 Qh2 43. Kb1 Qf4 44. Rd5 b6 45. Bd3 Qe3 46. Ka2 Ra8 47. Bc4
Rac8 48. Bb5 Ra8 49. Bd3 Rd8 50. Ba6 Ra8 51. Bb5 Rb8 52. Ba6 Ra8 53. Bb5 Rb8
54. Rd3 Qf4 55. Bc4 Rf8 56. Qb5 Rc5 57. Qa6 Rc7 58. Bb3 Rb8 59. Rd5 Qe3 60. Qb5
Qf4 61. Qe2 Rd8 62. Qd3 Rdc8 63. Qa6 Rb8 64. Qb5 Kh8 65. Qd3 Kg8 66. Qa6 Qe3
67. Qb5 Qf4 68. Qd3 Rd8 69. Rb1 Re8 70. Rd1 Rb8 71. Qa6 Rcb7 72. Qc4 Bf8 73.
Rb5 Be7 74. Rf1 Bd8 75. Qc6 Bc7 76. Rd1 Ra7 77. Rxd6 Bxd6 78. Qxd6 Rbb7 79.
Rxe5 Kh7 80. Re6 Qxd6 81. Rxd6 Rb8 82. Bd5 Rc7 83. c4 g5 84. Kb3 Re7 85. Kc3
Kg7 86. Kd4 f6 87. Kc3 Ree8 88. Rd7+ Kh8 89. b4 Red8 90. Rf7 axb4+ 91. Kxb4 Rd6
92. Rc7 Rbd8 93. Kb5 Re8 94. Bc6 Rb8 95. Bd5 Re8 96. Rc6 Rxc6 97. Kxc6 Kg7 98.
Kxb6 Rc8 99. a5 Rf8 100. a6 Rb8+ 101. Bb7 Rd8 102. a7 Rg8 103. c5 f5 104. c6
Kf6 1-0

[crem]

It is using mcts search in tcec. The alphabeta version has not been released yet. And the engine playing in TCEC is officially “Allie+Stein” as it is using Mark’s excellent SL “Stein” network. Without that, Allie could not play chess.

Just want to note, that it's not correct to say that "Stein" is a neural network. It's a set of neural network weights, but neural network is not just weights, it's also at least neural network structure, and probably also the implementation of evaluation, storage and training.

From the "Allie+Stein" name it may feel that there's no Lc0 at all in this project, but in fact there are lots of parts of Lc0 between Stein and Allie:

• The structure of Neural Network (Lc0 project does many experiments with that, e.g. SE-networks, policy head experiments, WDL-head, and any improvements there go into Allie+Stein for free)
• The NN backends, especially CuDNN one. GPU programming is a pretty rare skill, and writing optimized backend is not an easy task. Ankan's contribution to the project (writing CuDNN backend) is one of the most important reasons of Lc0 being so strong now. Also, all changes to NN structure "automatically" get into the backends. Lc0 NN backends is not just a library to eval generic neural network, it's an implementation of particular neural network structure which is changed with every experiment.
• The weights file format. That one is not that important compared to other items though.
• Training script. That would probably be also not that hard to rewrite, as it's pretty standard tensorflow script, telling just for completeness.

Also note that Lc0 NN alone without any search (just play moves with highest prior) has strength of 2400 Elo (by some scale). In my opinion, that 2400 really comes from Lc0 project (not even from "Stein" project), and search code builds on top of that.

[dkappe]

It is using mcts search in tcec. The alphabeta version has not been released yet. And the engine playing in TCEC is officially “Allie+Stein” as it is using Mark’s excellent SL “Stein” network. Without that, Allie could not play chess.

Just want to note, that it's not correct to say that "Stein" is a neural network. It's a set of neural network weights, but neural network is not just weights, it's also at least neural network structure, and probably also the implementation of evaluation, storage and training.

From the "Allie+Stein" name it may feel that there's no Lc0 at all in this project, but in fact there are lots of parts of Lc0 between Stein and Allie:

• The structure of Neural Network (Lc0 project does many experiments with that, e.g. SE-networks, policy head experiments, WDL-head, and any improvements there go into Allie+Stein for free)
• The NN backends, especially CuDNN one. GPU programming is a pretty rare skill, and writing optimized backend is not an easy task. Ankan's contribution to the project (writing CuDNN backend) is one of the most important reasons of Lc0 being so strong now. Also, all changes to NN structure "automatically" get into the backends. Lc0 NN backends is not just a library to eval generic neural network, it's an implementation of particular neural network structure which is changed with every experiment.
• The weights file format. That one is not that important compared to other items though.
• Training script. That would probably be also not that hard to rewrite, as it's pretty standard tensorflow script, telling just for completeness.

Also note that Lc0 NN alone without any search (just play moves with highest prior) has strength of 2400 Elo (by some scale). In my opinion, that 2400 really comes from Lc0 project (not even from "Stein" project), and search code builds on top of that.

So you’re saying AllieStein is not unique and should not be allowed in TCEC? Your grapes seem to be sour.

[jp]

AllieStein v0.2-n4 - Arasan TCEC15 0-1 Game #45 C47 S15 - Division 3

was not so good for Allie.

1. e4 e5 2. Nf3 Nc6 3. Nc3 Nf6 4. d4 exd4 5. Nxd4 Bb4 6. Nxc6 bxc6 7. Qd4 Qe7 8. f3 Bc5 9. Qd3 O-O 10. Bg5 Rb8 11. O-O-O d5 12. exd5 h6 13. Bh4 Ba6 14. Bxf6 Be3+ 15. Kb1 Rxb2+ 16. Ka1 gxf6 17. Qxa6 Rfb8 18. Bd3 Qb4 19. Ne2 Rxa2+ 20. Kxa2 Qb2# 0-1

[mwyoung]

It is using mcts search in tcec. The alphabeta version has not been released yet. And the engine playing in TCEC is officially “Allie+Stein” as it is using Mark’s excellent SL “Stein” network. Without that, Allie could not play chess.

Just want to note, that it's not correct to say that "Stein" is a neural network. It's a set of neural network weights, but neural network is not just weights, it's also at least neural network structure, and probably also the implementation of evaluation, storage and training.

From the "Allie+Stein" name it may feel that there's no Lc0 at all in this project, but in fact there are lots of parts of Lc0 between Stein and Allie:

• The structure of Neural Network (Lc0 project does many experiments with that, e.g. SE-networks, policy head experiments, WDL-head, and any improvements there go into Allie+Stein for free)
• The NN backends, especially CuDNN one. GPU programming is a pretty rare skill, and writing optimized backend is not an easy task. Ankan's contribution to the project (writing CuDNN backend) is one of the most important reasons of Lc0 being so strong now. Also, all changes to NN structure "automatically" get into the backends. Lc0 NN backends is not just a library to eval generic neural network, it's an implementation of particular neural network structure which is changed with every experiment.
• The weights file format. That one is not that important compared to other items though.
• Training script. That would probably be also not that hard to rewrite, as it's pretty standard tensorflow script, telling just for completeness.

Also note that Lc0 NN alone without any search (just play moves with highest prior) has strength of 2400 Elo (by some scale). In my opinion, that 2400 really comes from Lc0 project (not even from "Stein" project), and search code builds on top of that.

So you’re saying AllieStein is not unique and should not be allowed in TCEC? Your grapes seem to be sour.

Why is this sour grapes. Who cares what TCEC allows. They are a private chess engine exhibition. And answer only to themselves.

It will be funny to watch 100 Lc0 based NN engines playing next year...