TCEC Season 7 - Main Thread

[M ANSARI]

Komodo seemed to play really well in that format. I was shocked at how many wins Gull managed against SF and Houdini! Komodo seemed to have the most balanced play between attack and defense and felt like it had the best evaluation or idea of what was going on. But SF seemed to be playing way below strength. Had it not been for the heavy toll Komodo had on Gull, it could have very well been in the final.

[Jouni]

I calculated, that we have new (old) champion just before 2015! Around 28.-29.12., if games are not much different from earlier stages.

[lkaufman]

I would just like to comment on the rating system used by TCEC. Just one look at the ratings shows that they do not reflect the results of the latest cycle to any meaningful degree; rather they reflect the total results over an extended period. This is of course the proper way to do ratings when the engines involved are unchanged, but when they are updated after each round it doesn't make much sense. There is no logic to weighting results of old versions equally to current results; it simply results in the underrating of those engines that do keep updating.
The simplest solution would be just to use the FIDE rating system, where ratings would be updated treating each stage as a complete tournament. This automatically results in the latest stages getting more weight than the old ones. The situation is really the same as for human players, who also constantly "update" themselves. In both cases it is only proper that older results get aged out reasonably quickly, so that ratings reflect current strength as much as is practical.
Thanks again for a great job running TCEC!

Best regards,
Larry

[Martin Thoresen]

I would just like to comment on the rating system used by TCEC. Just one look at the ratings shows that they do not reflect the results of the latest cycle to any meaningful degree; rather they reflect the total results over an extended period. This is of course the proper way to do ratings when the engines involved are unchanged, but when they are updated after each round it doesn't make much sense. There is no logic to weighting results of old versions equally to current results; it simply results in the underrating of those engines that do keep updating.
The simplest solution would be just to use the FIDE rating system, where ratings would be updated treating each stage as a complete tournament. This automatically results in the latest stages getting more weight than the old ones. The situation is really the same as for human players, who also constantly "update" themselves. In both cases it is only proper that older results get aged out reasonably quickly, so that ratings reflect current strength as much as is practical.
Thanks again for a great job running TCEC!

Best regards,
Larry

Larry, the rating list is based on all games since Season 4. No games involving crashes or disconnects etc. are counted.

You are right that a FIDE-like system is more suitable, but the solution for that isn't there right now. It would be hopelessly time consuming for me to do it manually.

Best,
Martin

[whereagles]

Komodo seemed to play really well in that format. I was shocked at how many wins Gull managed against SF and Houdini! Komodo seemed to have the most balanced play between attack and defense and felt like it had the best evaluation or idea of what was going on. But SF seemed to be playing way below strength. Had it not been for the heavy toll Komodo had on Gull, it could have very well been in the final.

Agree. I'm quite surprised as well. SF seems like underperforming. Could it be the lack of tablebases?

[Jouni]

Obviously engines are approaching perfection and most games are draws now! Match is in shedule and will end as I predicted around 28.12.-29.12.

[Jouni]

Additionally some observations about games. Komodo is eager to play h4 in many white games. Stockfish is not caring much about one or two pawns for temporary sacrifice. But most common evaluation is now 0,00 vs 0,00 .

[M ANSARI]

Komodo won a game yesterday that was absolutely remarkable and shows what tremendous progress computer chess engines have made in the last few years. A position that seemed like totally a fortress position was expertly handled by Komodo, and I really think it is the only engine that would have managed to win that endgame. What made it more impressive is the fact that no EGTB's were used. While it is still way too early in the match and things could quickly change, the TCEC Komodo seems incredibly strong at the TC and hardware used.

1. d4 Nf6 2. c4 g6 3. g3 Bg7 4. Bg2 d5 5. cxd5 Nxd5 6. e4 Nb6 7. Ne2 c5 8. d5 e6 9. O-O O-O 10. Nbc3 exd5 11. exd5 Bf5 12. b3 Re8 13. Be3 Na6 14. Qd2 Qd7 15. Rad1 Rad8 16. Bh6 Be5 17. Bg5 Rc8 18. a3 c4 19. b4 Nc7 20. Rfe1 Bg4 21. d6 Nb5 22. f3 Nxc3 23. Nxc3 Bf5 24. Be7 Bg7 25. f4 Bg4 26. Rc1 Be6 27. Re2 Bg4 28. Re3 h6 29. Rce1 Be6 30. Qd1 Bg4 31. Qc1 Bd4 32. Qd2 Bg7 33. h4 Be6 34. Qc1 Bh3 35. Bh1 Bg4 36. Kh2 h5 37. Bg2 Bf5 38. b5 Rc5 39. Bf6 Rcc8 40. Be5 f6 41. Bd4 Red8 42. Bxb6 axb6 43. Nd5 Qxd6 44. Ne7+ Kh7 45. Nxc8 Rxc8 46. Re7 Rc7 47. Rxc7 Qxc7 48. Bf1 Qc5 49. Qe3 Qxb5 50. Qd4 Bf8 51. Qxf6 Bxa3 52. Qf7+ Kh6 53. Qxc4 Qxc4 54. Bxc4 Bc5 55. Be6 Bd3 56. Bc8 Ba6 57. Re5 Kg7 58. Be6 Bb5 59. Bd5 Kf6 60. Bxb7 Bf1 61. Bc8 Bc4 62. Bd7 Ba3 63. Bb5 Bf7 64. Bd3 Bc5 65. Kg2 Be7 66. Kf2 Bd6 67. Rg5 Bc5+ 68. Kg2 Be8 69. Re5 Bf7 70. Re1 Ba2 71. Be4 Bc4 72. Bc6 Bd6 73. Rd1 Bc5 74. Bd5 Be2 75. Re1 Bb5 76. Ra1 Bd7 77. Ra7 Be8 78. Rh7 Bd4 79. Rb7 Bc5 80. Kf3 Bg1 81. Rb8 Ke7 82. Rc8 Bc5 83. Be4 Bf7 84. Rc6 Be8 85. Rxg6 Bxg6 86. Bxg6 b5 87. Bxh5 Bb4 88. Bg6 Kf8 89. h5 Kg8 1-0

[michiguel]

I would just like to comment on the rating system used by TCEC. Just one look at the ratings shows that they do not reflect the results of the latest cycle to any meaningful degree; rather they reflect the total results over an extended period. This is of course the proper way to do ratings when the engines involved are unchanged, but when they are updated after each round it doesn't make much sense. There is no logic to weighting results of old versions equally to current results; it simply results in the underrating of those engines that do keep updating.
The simplest solution would be just to use the FIDE rating system, where ratings would be updated treating each stage as a complete tournament. This automatically results in the latest stages getting more weight than the old ones. The situation is really the same as for human players, who also constantly "update" themselves. In both cases it is only proper that older results get aged out reasonably quickly, so that ratings reflect current strength as much as is practical.
Thanks again for a great job running TCEC!

Best regards,
Larry

The simplest solution would be to use Ordo with its features, loose and relative anchors. We tested last year and it gave very reasonable results.

Miguel

[IGarcia]

Komodo won a game yesterday that was absolutely remarkable and shows what tremendous progress computer chess engines have made in the last few years. A position that seemed like totally a fortress position was expertly handled by Komodo, and I really think it is the only engine that would have managed to win that endgame. What made it more impressive is the fact that no EGTB's were used. While it is still way too early in the match and things could quickly change, the TCEC Komodo seems incredibly strong at the TC and hardware used.

1. d4 Nf6 2. c4 g6 3. g3 Bg7 4. Bg2 d5 5. cxd5 Nxd5 6. e4 Nb6 7. Ne2 c5 8. d5 e6 9. O-O O-O 10. Nbc3 exd5 11. exd5 Bf5 12. b3 Re8 13. Be3 Na6 14. Qd2 Qd7 15. Rad1 Rad8 16. Bh6 Be5 17. Bg5 Rc8 18. a3 c4 19. b4 Nc7 20. Rfe1 Bg4 21. d6 Nb5 22. f3 Nxc3 23. Nxc3 Bf5 24. Be7 Bg7 25. f4 Bg4 26. Rc1 Be6 27. Re2 Bg4 28. Re3 h6 29. Rce1 Be6 30. Qd1 Bg4 31. Qc1 Bd4 32. Qd2 Bg7 33. h4 Be6 34. Qc1 Bh3 35. Bh1 Bg4 36. Kh2 h5 37. Bg2 Bf5 38. b5 Rc5 39. Bf6 Rcc8 40. Be5 f6 41. Bd4 Red8 42. Bxb6 axb6 43. Nd5 Qxd6 44. Ne7+ Kh7 45. Nxc8 Rxc8 46. Re7 Rc7 47. Rxc7 Qxc7 48. Bf1 Qc5 49. Qe3 Qxb5 50. Qd4 Bf8 51. Qxf6 Bxa3 52. Qf7+ Kh6 53. Qxc4 Qxc4 54. Bxc4 Bc5 55. Be6 Bd3 56. Bc8 Ba6 57. Re5 Kg7 58. Be6 Bb5 59. Bd5 Kf6 60. Bxb7 Bf1 61. Bc8 Bc4 62. Bd7 Ba3 63. Bb5 Bf7 64. Bd3 Bc5 65. Kg2 Be7 66. Kf2 Bd6 67. Rg5 Bc5+ 68. Kg2 Be8 69. Re5 Bf7 70. Re1 Ba2 71. Be4 Bc4 72. Bc6 Bd6 73. Rd1 Bc5 74. Bd5 Be2 75. Re1 Bb5 76. Ra1 Bd7 77. Ra7 Be8 78. Rh7 Bd4 79. Rb7 Bc5 80. Kf3 Bg1 81. Rb8 Ke7 82. Rc8 Bc5 83. Be4 Bf7 84. Rc6 Be8 85. Rxg6 Bxg6 86. Bxg6 b5 87. Bxh5 Bb4 88. Bg6 Kf8 89. h5 Kg8 1-0

Besides the low amount of games (for a statistic POV) Komodo is showing great play, and seems it has coded a much better understanding of chess, something always Don was looking for.

However I noticed on all games Komodo took advantage (games 14, 16 and 40) SF was in time trouble. Not to discredit Komodo achievement, but asking myself what if SF has a little better time management the results will be the same. This repeats in almost all games, Komodo having time advantage.

So far, great games and tournament!
Thaks for all Efforts to Martin.