Why is Stockfish dominating the TCEC so far?

[Leo]

Is it because of the new contempt? Did Houdart set his contempt to high? I am curious what you think.

[Ras]

I guess it's because SF makes better moves.

[Norm Pollock]

The domination is even stronger that it appears. The lone Houdini victory used the same opening with which Stockfish also won, with colors reversed. Stockfish won in 36 moves while it took Houdini 76 moves.

Code: Select all

[Event "TCEC Season 11 - Superfinal"]
[Site "http://tcec.chessdom.com"]
[Date "2018.04.03"]
[Round "31"]
[White "Stockfish 260318"]
[Black "Houdini 6.03"]
[Result "1-0"]
[BlackElo "3489"]
[ECO "C05"]
[EventDate "2018.03.29"]
[GameDuration "02:47:56"]
[Opening "French: Tarrasch, Closed, 7.Ngf3 Be7"]
[PlyCount "72"]
[Termination "adjudication"]
[TerminationDetails "TCEC win rule"]
[TimeControl "7200+15"]
[WhiteElo "3546"]

1. e4 e6 2. d4 d5 3. Nd2 Be7 4. Ngf3 Nf6 5. e5 Nfd7 6. Bd3 c5 7. c3 Nc6 
8. O-O a5 9. a4 O-O 10. Re1 b6 11. Bb1 Ba6 12. Nb3 Re8 13. g3 Qc7 
14. Bf4 Rac8 15. h4 Nf8 16. h5 h6 17. Qc1 Qd7 18. Qe3 Bc4 19. Nbd2 cxd4 
20. Nxd4 Nxd4 21. cxd4 Ba6 22. Nf3 b5 23. Bxh6 gxh6 24. Qxh6 Qd8 
25. g4 bxa4 26. g5 Bb4 27. Re3 Rc1+ 28. Kg2 Bf1+ 29. Kh2 Rc7 30. Kg3 a3 
31. bxa3 Bc3 32. Nh2 Bxa1 33. Ng4 f5 34. Nf6+ Qxf6 35. exf6 f4+ 
36. Kxf4 Bxd4 1-0 

[Event "TCEC Season 11 - Superfinal"]
[Site "http://tcec.chessdom.com"]
[Date "2018.04.03"]
[Round "32"]
[White "Houdini 6.03"]
[Black "Stockfish 260318"]
[Result "1-0"]
[BlackElo "3546"]
[ECO "C05"]
[EventDate "2018.03.29"]
[GameDuration "04:34:25"]
[Opening "French: Tarrasch, Closed, 7.Ngf3 Be7"]
[PlyCount "151"]
[Termination "adjudication"]
[TerminationDetails "TCEC win rule"]
[TimeControl "7200+15"]
[WhiteElo "3489"]

1. e4 e6 2. d4 d5 3. Nd2 Be7 4. Ngf3 Nf6 5. e5 Nfd7 6. Bd3 c5 7. c3 Nc6 
8. O-O a5 9. Bc2 cxd4 10. cxd4 b5 11. Re1 a4 12. a3 Nb6 13. Nf1 Nc4 
14. Ne3 Bb7 15. Rb1 N4a5 16. Nf1 Nc4 17. h3 Rc8 18. Ng3 h5 19. Bd3 Rb8 
20. Ne2 Bc8 21. Nd2 Bb7 22. Bc2 Ba8 23. Nxc4 dxc4 24. Be3 Na5 25. Ng3 Nb3 
26. Nxh5 g6 27. Ng3 Rb7 28. Ne2 Rd7 29. Bxb3 axb3 30. Nc3 Bc6 31. Qe2 Kf8 
32. Rbd1 Qa8 33. f3 Kg8 34. Bf2 Qd8 35. d5 exd5 36. Bd4 Bf8 37. f4 Rh5 
38. Qe3 Qe7 39. Ne2 Rd8 40. Bc3 Rh7 41. Bb4 Qc7 42. Bxf8 Rxf8 43. Rd2 Rd8 
44. Nc3 f6 45. Ne2 Rf8 46. e6 Rb8 47. Nd4 Be8 48. f5 Rh4 49. g4 g5 
50. Kg2 Rh7 51. Nf3 d4 52. Nxd4 Qb7+ 53. Kf2 Qa7 54. Qc3 Rd8 55. Re3 Qb8 
56. Kf1 Qb7 57. Kg1 Qb6 58. Kg2 Qc7 59. Ne2 Kg7 60. Rxd8 Qxd8 61. Kf2 Qd1 
62. Qd4 Qxd4 63. Nxd4 Kf8 64. Ke1 Rc7 65. Kd2 Ke7 66. Kc3 Rc5 67. Re2 Rc7 
68. Kb4 Rc8 69. Re1 Rb8 70. Rb1 Kf8 71. Ne2 Bc6 72. Nc3 Bf3 73. Kc5 Rd8 
74. Nxb5 Rc8+ 75. Kb4 Ke7 76. Rc1 1-0 

[Leo]

Interesting point.

[Nordlandia]

Lowering contempt to zero may yield more draws than otherwise lost games. Even negative contempt may be appropriate.

Code: Select all

Level of contempt to avoid draws in game play.
Default 2, min -10, max 10.

The notion of "contempt" implies that Houdini will try to avoid draws by evaluating its
own position slightly too optimistically. Houdini will value its own pieces higher than the
opponent pieces and will only exchange them if there’s a clear positional advantage in
doing so.
The Contempt level can be chosen between -10 and 10. A contempt value of 2 will
typically reduce the draw rate in a match by about 4%. The impact on the evaluation
(the "optimism" of the engine, if you like) will be about 10 centipawn for each level of
contempt.
The objectively highest level of play is achieved without contempt, i.e. Contempt
value 0.

[Vizvezdenec]

stockfish has record-breaking progress since sf 9 (like 18 elo in 2 months - it's faster than in progressed after sf7 or sf8 releases, thx to noobpwnftw who increased fishtest amount of cores by 800) so now it's stronger by 50-60 elo.
SF was on par with H6 by the time of H6 release, but since then it gained this 50-60 elo (actually last TCEC SF caught really bad timing since premier div was played exactly when SF had 0 progress for 2 months and after it started SF team found a lot of elo gainers... But it was too late) - that's the difference you see in TCEC now.

[Nordlandia]

Michael Chaly:It's a mystery why RH choosed to use default contempt value of +2 against SF that is like you're implying (50 +/-) elo stronger now than when H6 was released.

[Vizvezdenec]

+2 contempt makes almost 0 difference in terms of elo according to multiple tests, maybe like 1 or 2 which sure wouldn't help you if you are lacking 50+

[Mike S.]

I do not think that it is a matter of contempt between these two giants. Houdini 6.03 is strong, but Stockfish is just (much) stronger. I do not know what else to say about this situation. I saw a game where Stockfish understood very much earlier that it was winning, many moves earlier than Houdini. But I am not sure if it is a matter of effective search depth alone.

[Dann Corbit]

Is it because of the new contempt? Did Houdart set his contempt to high? I am curious what you think.

Contempt against lesser opponents is worth maybe 15 Elo.
Contempt against peers has almost no value.
An eval tweak is likely more important than a good contempt setting.
And a search improvement is ten times more important.

SF is winning more because H held still and SF moved ahead.