A Gilden rematch

[lkaufman]

This was the first win by a human master in the last twelve pre-announced knight odds rapid games on chess.com. Komodo failed to obtain any initiative or compensation for the piece, Gilden methodically simplified and won cleanly.

[pgn][Event "Live Chess - Knight Odds"] [Site "Chess.com"] [Date "2019.17.dec"] [Round "4"] [White "PlayKomodo"] [Black "FM lgilden"] [Result "0-1"] [SetUp "1"] [FEN "rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/R1BQKBNR w KQkq - 0 1"] [TimeControl "1800+10"] 1. f4 d5 2. Nf3 g6 3. e3 Bg7 4. g3 c5 5. a3 Nc6 6. Rb1 Nf6 7. Bg2 a5 8. O-O O-O 9. Qe2 d4 10. d3 dxe3 11. Bxe3 Nd4 12. Nxd4 cxd4 13. Bf2 Ng4 14. Kh1 Nxf2+ 15. Qxf2 Rb8 16. Qd2 Be6 17. Rf2 Bd5 18. h4 Bxg2+ 19. Rxg2 Qd5 20. Kg1 Qa2 21. Qc1 Qd5 22. Qf1 b5 23. Ra1 b4 24. axb4 axb4 25. b3 Rfc8 26. Re1 Rb7 27. Ree2 Rbc7 28. Qe1 Qd6 29. h5 e6 30. hxg6 hxg6 31. Rgf2 Bf6 32. Qf1 Ra7 33. Re1 Qd5 34. Qg2 Qxg2+ 35. Kxg2 Rca8 36. Re4 Ra1 37. g4 Rc1 38. Kf3 Raa1 39. Rg2 Rf1+ 40. Rf2 Kf8 41. Rxf1 Rxf1+ 42. Kg3 Rc1 43. Re2 g5 44. fxg5 Bxg5 45. Re5 Be7 46. Kf3 Rxc2 47. Rb5 Rb2 48. Ke4 Rxb3 49. Kxd4 Rb1 50. g5 Rg1 51. Ke3 Rxg5 52. Rb7 Re5+ 53. Kd2 Rd5 54. Ra7 f6 55. Ra6 Kf7 56. Rb6 f5 57. Rb8 Kf6 58. Rh8 Ke5 59. Rh7 Bg5+ 60. Kc2 b3+ 61. Kxb3 Rxd3+ 62. Kc4 Rd4+ 63. Kb5 Ke4 64. Rh1 Be3 65. Rh7 e5 66. Rh8 f4 67. Rh4 Kd3 68. Rh6 e4 69. Rh1 f3 70. Kc5 Rc4+ 71. Kd6 Rc1 72. Ke6 Rxh1 73. Kf5 f2 74. Kg6 Rg1+ 75. Kh5 f1=Q 76. Kh4 Qf6+ 77. Kh5 Qh8# 0-1[/pgn]

[lkaufman]

White got a strong attack right out of the opening (The Scandinavian 1...d5 is ideal at b1 odds but not so good at g1 odds) and won quite quickly. Final score 3.5 to 1.5 (3 wins for Komodo, 1 win for Gilden, 1 draw). Komodo won all three at g1 odds, but lost 1.5 to 0.5 at b1 odds. Perhaps 1.f4 isn't so good at knight (b1) odds, but with 1.e4 dubious at that handicap due to 1...d5 the move 1.f4 was often chosen by 19th century masters when they were worried about 1.e4 d5; it's not clear what other White move offers chances of a complicated enough game for knight odds.

[pgn][Event "Live Chess - Knight Odds"] [Site "Chess.com"] [Date "2019.17.dec"] [Round "5"] [White "PlayKomodo"] [Black "FM lgilden"] [Result "1-0"] [SetUp "1"] [FEN "rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKB1R w KQkq - 0 1"] [TimeControl "1800+10"] 1. e4 d5 2. exd5 Qxd5 3. Nc3 Qa5 4. d4 Nf6 5. Bd2 c6 6. Bc4 Bg4 7. f3 Bf5 8. g4 Bg6 9. Qe2 Qd8 10. O-O-O Nbd7 11. f4 b5 12. Bb3 a5 13. f5 a4 14. fxg6 axb3 15. gxf7+ Kxf7 16. cxb3 h6 17. Rhe1 Ke8 18. Qf3 Nd5 19. Nxb5 cxb5 20. Qxd5 Rxa2 21. Qh5+ 1-0[/pgn]

[dannyb]

thanks for the games and congrats to FM lgilden for winning a game against Komodo! and of course to Team Komodo for winning the match.

will this Komodo NN be released at some point?

[lkaufman]

thanks for the games and congrats to FM lgilden for winning a game against Komodo! and of course to Team Komodo for winning the match.

will this Komodo NN be released at some point?

Our intent is for this Komodo NN to be a new Komodo product. It is already quite a bit stronger than Komodo MCTS without NN (assuming you have a good GPU), but we think it needs to be stronger than normal Komodo to be a viable product, although it is already my engine for choice for analysis, since it is already stronger than either normal or MCTS Komodo in multiPV mode. Whether that will take a week or a year is impossible to predict; if we just improve it incrementally it may take a while, but if we hit on one good idea it could happen quickly.
I hope we'll be able to do a knight odds match in the near future with an IM or even a GM at a time limit suitable based on the opponent. One thing I'll have to think about based on this match is whether we should continue to alternate the knight or just do g1 odds. With the g1 knight missing White can play normal openings generally, but with b1 missing he has to resort to unusual openings.

[dannyb]

Our intent is for this Komodo NN to be a new Komodo product. It is already quite a bit stronger than Komodo MCTS without NN (assuming you have a good GPU), but we think it needs to be stronger than normal Komodo to be a viable product, although it is already my engine for choice for analysis, since it is already stronger than either normal or MCTS Komodo in multiPV mode. Whether that will take a week or a year is impossible to predict; if we just improve it incrementally it may take a while, but if we hit on one good idea it could happen quickly.

good luck to you and Mark then. Looking forward to this new version.

[Nordlandia]

The announcement Komodo NN can be something similar to a christmas gift from team Komodo.

Maybe opt for GPU donation for Komodo NN improvement. Stockfish and Leela relies on voluntary CPU / GPU donation. Maybe this is something to consider.

[lkaufman]

The announcement Komodo NN can be something similar to a christmas gift from team Komodo.

Maybe opt for GPU donation for Komodo NN improvement. Stockfish and Leela relies on voluntary CPU / GPU donation. Maybe this is something to consider.

Right now the focus is on how to best combine Komodo with any existing NN, rather than making our own NN. Once we have found a sufficiently good way to do that, we will likely turn our attention to the NN itself. But for the way we are using the NN now, it wouldn't be a sensible use of resources to try to make a better net, it's much more important to find the best use of existing nets.

[dannyb]

Right now the focus is on how to best combine Komodo with any existing NN, rather than making our own NN. Once we have found a sufficiently good way to do that, we will likely turn our attention to the NN itself. But for the way we are using the NN now, it wouldn't be a sensible use of resources to try to make a better net, it's much more important to find the best use of existing nets.

is Komodo NN based on the MCTS version or on the AB one?

[Laskos]

Right now the focus is on how to best combine Komodo with any existing NN, rather than making our own NN. Once we have found a sufficiently good way to do that, we will likely turn our attention to the NN itself. But for the way we are using the NN now, it wouldn't be a sensible use of resources to try to make a better net, it's much more important to find the best use of existing nets.

is Komodo NN based on the MCTS version or on the AB one?

Must be MCTS, NNs are convoluted and slow, AB for now is undoable with them. But Mark and Larry have a peculiar MCTS, let's see how it will fare with existing nets when they will iron out the engine.

[lkaufman]

Right now the focus is on how to best combine Komodo with any existing NN, rather than making our own NN. Once we have found a sufficiently good way to do that, we will likely turn our attention to the NN itself. But for the way we are using the NN now, it wouldn't be a sensible use of resources to try to make a better net, it's much more important to find the best use of existing nets.

is Komodo NN based on the MCTS version or on the AB one?

This NN version is Komodo MCTS with a little help from NN. We also have a version of normal K that gets help from NN, but the benefit is too small to be interesting so far in that version.