Komodo Dragon 2.6.1 released

[lkaufman]

KomodoChess has released Dragon 2.6.1 at KomodoChess.com. It is primarily a bug-fix version, needed because in 2.6 setting Auto-Skill disabled the Elo settings, making Auto-Skill useless. Also, in response to many requests, we changed the UCI options for Elo and LimitStrength to use the underscore, so that they may be used more easily on some GUIs. Additionally, some of the internal settings for the Elo settings have been revised based on new information; below 1650 they were made a bit weaker, above 1650 a bit stronger. There is no change in the net used; some tiny speedups were included and one small internal parameter change; we estimate the net effect is a boost of about two elo. Obviously there is no need for the testers to retest for this. This version is free for anyone who received version 2.6 from us, just download it, and any new purchasers will get this version.

[Chessqueen]

KomodoChess has released Dragon 2.6.1 at KomodoChess.com. It is primarily a bug-fix version, needed because in 2.6 setting Auto-Skill disabled the Elo settings, making Auto-Skill useless. Also, in response to many requests, we changed the UCI options for Elo and LimitStrength to use the underscore, so that they may be used more easily on some GUIs. Additionally, some of the internal settings for the Elo settings have been revised based on new information; below 1650 they were made a bit weaker, above 1650 a bit stronger. There is no change in the net used; some tiny speedups were included and one small internal parameter change; we estimate the net effect is a boost of about two elo. Obviously there is no need for the testers to retest for this. This version is free for anyone who received version 2.6 from us, just download it, and any new purchasers will get this version.

I have tested a few of CCRL engines against Komodo Dragon 2.6.1 and you have done an excellent job in equalizing their strength, but the question is if the rating of CCRL engines are close to Human Fide at T/C of 10 Minutes ? Here is an interesting games out of the 100 games that I matched against Snowy and setting the UCI Elo to 2000. I decided to test it under 10 Minutes to get a feel of it, but in Chess.com most humans prefer to challenge engines either at T/C of 3 minutes or 3/2 Blitz, anything greater than those T/C will give some humans a chance to cheat using other engines and after reaching 30 to 40 moves play on their own when they have an advantage. Let say they play Blitz in 5/5 minutes and use Stockfish after reaching 38 moves they get a great advantage, they can easily play the rest of the game against their own Elo Strength and beat Komodo Dragon.

[pgn][Event "Computer chess game"]
[Site "MININT-UB2PIMJ"]
[Date "2022.01.10"]
[Round "?"]
[White "Dragon-2.6.1-64bit-avx2"]
[Black "Snowy_0_2_x64"]
[Result "1/2-1/2"]
[BlackElo "2000"]
[ECO "D45"]
[Opening "Semi-Slav"]
[Time "14:04:36"]
[Variation "6.Qc2 Bd6 7.Be2"]
[WhiteElo "2000"]
[TimeControl "600"]
[Termination "normal"]
[PlyCount "212"]
[WhiteType "program"]
[BlackType "program"]

1. d4 Nf6 2. c4 c6 3. Nf3 d5 4. Nc3 e6 5. e3 Nbd7 6. Qc2 Bd6 7. Be2 dxc4 8.
a4 Nd5 9. O-O b5 10. Rd1 O-O 11. b3 cxb3 12. Qxb3 bxa4 13. Nxa4 Rb8 14. Qc4
Bb7 15. Nc5 Bxc5 16. dxc5 a5 17. e4 N5f6 18. Bf4 Ra8 19. Bd6 Ba6 20. Qc2
Bxe2 21. Qxe2 Re8 22. Nd2 a4 23. h3 Ra7 24. Nc4 Qa8 25. Qc2 a3 26. Kh2 Ra4
27. f3 Qa7 28. Nb6 Ra5 29. Nc4 Ra6 30. Qd2 a2 31. Qf2 Ra8 32. Rd2 Ne8 33.
Be7 Ra4 34. Nb2 Ra5 35. Nd3 Ndf6 36. Bxf6 Nxf6 37. Rb2 Rd8 38. Rbxa2 Qc7+
39. e5 Rxa2 40. Rxa2 Rxd3 41. Ra8+ Rd8 42. Rxd8+ Qxd8 43. exf6 Qxf6 44. Qg3
e5 45. Qe1 Qf4+ 46. Kh1 g5 47. Qa5 Qc1+ 48. Kh2 Kg7 49. Qb6 Qf4+ 50. Kh1 e4
51. Qxc6 exf3 52. gxf3 Qg3 53. Qd5 Qxh3+ 54. Kg1 Qg3+ 55. Kf1 h5 56. c6 Kg6
57. Qd3+ f5 58. Qc3 Qc7 59. Qc4 g4 60. Qe6+ Kg5 61. Qd7 Qf4 62. Qd8+ Kg6
63. c7 Qxf3+ 64. Ke1 Qg3+ 65. Kd2 Qf2+ 66. Kd3 Qf3+ 67. Kd4 Qf4+ 68. Kc5
Qc1+ 69. Kd5 Qd1+ 70. Ke6 Qe2+ 71. Kd7 Qb5+ 72. Ke6 Qb3+ 73. Qd5 Qe3+ 74.
Qe5 Qb3+ 75. Kd6 Qa3+ 76. Kd7 Qa7 77. Ke7 f4 78. Qxf4 Qb7 79. Qf7+ Kh6 80.
Qf6+ Kh7 81. Qd6 g3 82. Qd3+ Kh6 83. Qxg3 Qe4+ 84. Kf6 Qc6+ 85. Kf5 Qc2+
86. Ke6 Qc4+ 87. Kd7 Qa4+ 88. Kd8 Qd4+ 89. Kc8 h4 90. Qg8 Qe5 91. Kd7 Qb5+
92. Kd8 Qd3+ 93. Ke7 Qa3+ 94. Kd7 Qd3+ 95. Kc6 Qc3+ 96. Kb7 Qb2+ 97. Kc8
Qe5 98. Qg4 Qe8+ 99. Kb7 Qb5+ 100. Ka8 Qd5+ 101. Ka7 Qc5+ 102. Kb7 Qb5+
103. Ka8 Qd5+ 104. Ka7 Qc5+ 105. Kb8 Qb5+ 106. Ka7 Qc5+ {3-fold repetition}
1/2-1/2[/pgn]

[lkaufman]

KomodoChess has released Dragon 2.6.1 at KomodoChess.com. It is primarily a bug-fix version, needed because in 2.6 setting Auto-Skill disabled the Elo settings, making Auto-Skill useless. Also, in response to many requests, we changed the UCI options for Elo and LimitStrength to use the underscore, so that they may be used more easily on some GUIs. Additionally, some of the internal settings for the Elo settings have been revised based on new information; below 1650 they were made a bit weaker, above 1650 a bit stronger. There is no change in the net used; some tiny speedups were included and one small internal parameter change; we estimate the net effect is a boost of about two elo. Obviously there is no need for the testers to retest for this. This version is free for anyone who received version 2.6 from us, just download it, and any new purchasers will get this version.

I have tested a few of CCRL engines against Komodo Dragon 2.6.1 and you have done an excellent job in equalizing their strength, but the question is if the rating of CCRL engines are close to Human Fide at T/C of 10 Minutes ? Here is an interesting games out of the 100 games that I matched against Snowy and setting the UCI Elo to 2000. I decided to test it under 10 Minutes to get a feel of it, but in Chess.com most humans prefer to challenge engines either at T/C of 3 minutes or 3/2 Blitz, anything greater than those T/C will give some humans a chance to cheat using other engines and after reaching 50 to 60 moves play on their own when they have an advantage.

[pgn][Event "Computer chess game"]
[Site "MININT-UB2PIMJ"]
[Date "2022.01.10"]
[Round "?"]
[White "Dragon-2.6.1-64bit-avx2"]
[Black "Snowy_0_2_x64"]
[Result "1/2-1/2"]
[BlackElo "2000"]
[ECO "D45"]
[Opening "Semi-Slav"]
[Time "14:04:36"]
[Variation "6.Qc2 Bd6 7.Be2"]
[WhiteElo "2000"]
[TimeControl "600"]
[Termination "normal"]
[PlyCount "212"]
[WhiteType "program"]
[BlackType "program"]

1. d4 Nf6 2. c4 c6 3. Nf3 d5 4. Nc3 e6 5. e3 Nbd7 6. Qc2 Bd6 7. Be2 dxc4 8.
a4 Nd5 9. O-O b5 10. Rd1 O-O 11. b3 cxb3 12. Qxb3 bxa4 13. Nxa4 Rb8 14. Qc4
Bb7 15. Nc5 Bxc5 16. dxc5 a5 17. e4 N5f6 18. Bf4 Ra8 19. Bd6 Ba6 20. Qc2
Bxe2 21. Qxe2 Re8 22. Nd2 a4 23. h3 Ra7 24. Nc4 Qa8 25. Qc2 a3 26. Kh2 Ra4
27. f3 Qa7 28. Nb6 Ra5 29. Nc4 Ra6 30. Qd2 a2 31. Qf2 Ra8 32. Rd2 Ne8 33.
Be7 Ra4 34. Nb2 Ra5 35. Nd3 Ndf6 36. Bxf6 Nxf6 37. Rb2 Rd8 38. Rbxa2 Qc7+
39. e5 Rxa2 40. Rxa2 Rxd3 41. Ra8+ Rd8 42. Rxd8+ Qxd8 43. exf6 Qxf6 44. Qg3
e5 45. Qe1 Qf4+ 46. Kh1 g5 47. Qa5 Qc1+ 48. Kh2 Kg7 49. Qb6 Qf4+ 50. Kh1 e4
51. Qxc6 exf3 52. gxf3 Qg3 53. Qd5 Qxh3+ 54. Kg1 Qg3+ 55. Kf1 h5 56. c6 Kg6
57. Qd3+ f5 58. Qc3 Qc7 59. Qc4 g4 60. Qe6+ Kg5 61. Qd7 Qf4 62. Qd8+ Kg6
63. c7 Qxf3+ 64. Ke1 Qg3+ 65. Kd2 Qf2+ 66. Kd3 Qf3+ 67. Kd4 Qf4+ 68. Kc5
Qc1+ 69. Kd5 Qd1+ 70. Ke6 Qe2+ 71. Kd7 Qb5+ 72. Ke6 Qb3+ 73. Qd5 Qe3+ 74.
Qe5 Qb3+ 75. Kd6 Qa3+ 76. Kd7 Qa7 77. Ke7 f4 78. Qxf4 Qb7 79. Qf7+ Kh6 80.
Qf6+ Kh7 81. Qd6 g3 82. Qd3+ Kh6 83. Qxg3 Qe4+ 84. Kf6 Qc6+ 85. Kf5 Qc2+
86. Ke6 Qc4+ 87. Kd7 Qa4+ 88. Kd8 Qd4+ 89. Kc8 h4 90. Qg8 Qe5 91. Kd7 Qb5+
92. Kd8 Qd3+ 93. Ke7 Qa3+ 94. Kd7 Qd3+ 95. Kc6 Qc3+ 96. Kb7 Qb2+ 97. Kc8
Qe5 98. Qg4 Qe8+ 99. Kb7 Qb5+ 100. Ka8 Qd5+ 101. Ka7 Qc5+ 102. Kb7 Qb5+
103. Ka8 Qd5+ 104. Ka7 Qc5+ 105. Kb8 Qb5+ 106. Ka7 Qc5+ {3-fold repetition}
1/2-1/2[/pgn]

The elo settings on Dragon are for people to use on their own computers, so I don't need to be concerned about cheating, this isn't for online bots. The chess.com bots are Komodo-based but are pre-NNUE and scaled differently. The ratings given on Dragon 2.6.1 are intended to match humans at 15' + 10"; we could always change that to some faster tc in the future if enough people request that. Probably matching dragon levels vs. CCRL engines in the 2000 ballpark is reasonable in blitz; ccrl engines rated 2000 maybe play blitz about like humans rated 2000 play rapid. So far we've had four serious test games between dragon 2.6.1 Elo settings and titled human players (2 with me, 2 with my son Ray) at 15' + 10" with elo set to match human FIDE rating; result is two wins for Dragon, one win for human (me), and one draw. One of the two Dragon wins was closely fought and should have ended in a draw. So it does appear that the elo settings are at least not way off. But if we tried to play blitz against those same elo levels we would be crushed. I think I have underestimated how much better humans play Rapid than they play blitz.

[Chessqueen]

Komodo Dragon 2.6.1 at T/C of 30 seconds total for the entire game become unbeatable, but at 2 minutes and above Cicada is too stoong for the UCI_Elo set at 1600 for Komodo Dragon 2.6.1., but against any human rated around 1600 Fide at T/C 5/5 it would be just perfect.

[pgn][Event "Computer chess game"]
[Site "MININT-UB2PIMJ"]
[Date "2022.01.10"]
[Round "?"]
[White "Dragon-2.6.1-64bit-avx2"]
[Black "Cicada"]
[Result "1-0"]
[BlackElo "1550"]
[ECO "D30"]
[Opening "QGD"]
[Time "20:10:55"]
[Variation "3.Nf3 Nf6 4.e3 c6 5.Nbd2"]
[WhiteElo "1600"]
[TimeControl "30+0"]
[Termination "normal"]
[PlyCount "138"]
[WhiteType "program"]
[BlackType "program"]

1. Nf3 d5 2. d4 Nf6 3. c4 e6 4. e3 c6 5. Nbd2 c5 6. cxd5 exd5 7. b3 Nc6 8.
Bb2 cxd4 9. Nxd4 Bd7 10. Be2 Bd6 11. Nb5 Bb4 12. O-O Ke7 13. Rc1 Qa5 14.
Nf3 a6 15. Nbd4 Qxa2 16. Nxc6+ Bxc6 17. Qc2 Qa5 18. Nd4 Rac8 19. Qf5 Bd7
20. Qe5+ Kd8 21. Qg5 Rxc1 22. Rxc1 Bf8 23. Qe5 Bc5 24. Bxa6 Bxd4 25. exd4
Re8 26. Qb8+ Ke7 27. Qxb7 Qd2 28. Ba3+ Ke6 29. Rf1 Qxd4 30. Rc1 Qd2 31. h3
Ke5 32. g4 Be6 33. Bd6+ Kd4 34. Qc6 Ke4 35. Bg3 Bd7 36. Rc4+ Kf3 37. Rc3+
Re3 38. Rxe3+ Qxe3 39. Qc2 Qe1+ 40. Bf1 Qxf1+ 41. Kxf1 Bb5+ 42. Kg1 Ba6 43.
Qd2 d4 44. Qxd4 Ke2 45. Qe5+ Kd1 46. Qa1+ Kc2 47. Qa2+ Kc3 48. Be5+ Kb4 49.
Qxa6 Kxb3 50. Qb6+ Kc4 51. Qd4+ Kb5 52. Qd3+ Kc5 53. Qd4+ Kb5 54. f3 h5 55.
Qd3+ Kb4 56. Bd6+ Ka5 57. gxh5 Ne8 58. h6 gxh6 59. Bc5 Ng7 60. Kf2 Ne6 61.
Bd6 Kb6 62. f4 Kb7 63. Qb5+ Ka7 64. f5 Ng5 65. Bc5+ Ka8 66. Ke3 Ne6 67.
Qa6+ Kb8 68. fxe6 Kc7 69. Bd6+ Kd8 {Black resigns} *[/pgn]

[Cornfed]

Just an observation. I don't know anything about Cicada, but 12...Ke7?? - no human would (and no computer should) ever think to make that kind of move (or likely several others in the game, but that's the obvious one), even at 1550 or 1600 elo...or 1200 ( 4...c6. 5...c5. maybe...). But again, this must be the programmers burden when trying to simulate a rating level for 'bad human play'. How the heck does one realistically it? Missing some pretty obvious, sure, poor positional play, sure...but moves like 12...Ke7? I'm not even sure one should test against an engine that does that.

[lkaufman]

Just an observation. I don't know anything about Cicada, but 12...Ke7?? - no human would (and no computer should) ever think to make that kind of move (or likely several others in the game, but that's the obvious one), even at 1550 or 1600 elo...or 1200 ( 4...c6. 5...c5. maybe...). But again, this must be the programmers burden when trying to simulate a rating level for 'bad human play'. How the heck does one realistically it? Missing some pretty obvious, sure, poor positional play, sure...but moves like 12...Ke7? I'm not even sure one should test against an engine that does that.

Note that this "Cicada" engine is not a crippled engine trying to simulate bad human play, but the full strength engine; it is rated 1537 on CCRL Rapid.

[Cornfed]

Just an observation. I don't know anything about Cicada, but 12...Ke7?? - no human would (and no computer should) ever think to make that kind of move (or likely several others in the game, but that's the obvious one), even at 1550 or 1600 elo...or 1200 ( 4...c6. 5...c5. maybe...). But again, this must be the programmers burden when trying to simulate a rating level for 'bad human play'. How the heck does one realistically it? Missing some pretty obvious, sure, poor positional play, sure...but moves like 12...Ke7? I'm not even sure one should test against an engine that does that.

Note that this "Cicada" engine is not a crippled engine trying to simulate bad human play, but the full strength engine; it is rated 1537 on CCRL Rapid.

A really horrible one if among the many moves available to it, it plays an unprovoked 12...Ke7.

With it making moves like that, do you think results from such an engine as this is in any meaningful way valuable in trying to correlate Dragon's play/elo to a human with a similar rating? It strikes me that the overwhelming majority of errors by a human with that rating would be of a tactical nature or poor positional understanding...not totally crazy king moves which hands the opponent a nice edge on a silver platter.

I mean, I guess one could make the argument that it doesn't matter and 'whatever the error' Dragon still has to perform against it and if it can only score 50% then 'set whatever tweaks you have made' to be 'that level elo'. The approach just give me pause...but maybe I'm not seeing a bigger picture.

[lkaufman]

Just an observation. I don't know anything about Cicada, but 12...Ke7?? - no human would (and no computer should) ever think to make that kind of move (or likely several others in the game, but that's the obvious one), even at 1550 or 1600 elo...or 1200 ( 4...c6. 5...c5. maybe...). But again, this must be the programmers burden when trying to simulate a rating level for 'bad human play'. How the heck does one realistically it? Missing some pretty obvious, sure, poor positional play, sure...but moves like 12...Ke7? I'm not even sure one should test against an engine that does that.

Note that this "Cicada" engine is not a crippled engine trying to simulate bad human play, but the full strength engine; it is rated 1537 on CCRL Rapid.

A really horrible one if among the many moves available to it, it plays an unprovoked 12...Ke7.

With it making moves like that, do you think results from such an engine as this is in any meaningful way valuable in trying to correlate Dragon's play/elo to a human with a similar rating? It strikes me that the overwhelming majority of errors by a human with that rating would be of a tactical nature or poor positional understanding...not totally crazy king moves which hands the opponent a nice edge on a silver platter.

I mean, I guess one could make the argument that it doesn't matter and 'whatever the error' Dragon still has to perform against it and if it can only score 50% then 'set whatever tweaks you have made' to be 'that level elo'. The approach just give me pause...but maybe I'm not seeing a bigger picture.

Of course it is much better to test Dragon elo settings against humans, and we are gradually accumulating data, although most people don't test it under the specified 15' +10" serious game format. Some weaker engines are at least reasonably human-like in their play, others are not. In any case engines with CCRL ratings below master level will absolutely clobber humans of the same FIDE rating in blitz or even Rapid games, that's quite clear now. I know it's hard to believe that an engine that would play ...Ke7 there would destroy a 1550 FIDE human in Rapid, but it would. Tactics are too dominant in chess. "Cicada" playing blitz might be even with a 1550 human playing Rapid, but we don't know that. The only engine in that human level ballpark for which I have a lot of human data (from LiChess) is Safrad 2.2, which performed at the equivalent of FIDE 1527 against humans at Rapid (10' + 5" to 15' + 10" or equiv.); FIDE 1723 at slow blitz (about 5' + 3"), and at FIDE 1922 at 3' + 0". But it is only rated 1006 CCRL blitz. I think this is a good engine to use for such tests of Dragon elo settings, since we know how it performs vs. humans and since it at least is not outrageously silly in its play.

[Chessqueen]

Just an observation. I don't know anything about Cicada, but 12...Ke7?? - no human would (and no computer should) ever think to make that kind of move (or likely several others in the game, but that's the obvious one), even at 1550 or 1600 elo...or 1200 ( 4...c6. 5...c5. maybe...). But again, this must be the programmers burden when trying to simulate a rating level for 'bad human play'. How the heck does one realistically it? Missing some pretty obvious, sure, poor positional play, sure...but moves like 12...Ke7? I'm not even sure one should test against an engine that does that.

You forgot that Cicada only made that horrible move on T/C of 30 seconds for the entire game, if you allow Cicada at least 2 minutes for the entire game it would NOT make that horrible move and it would constantly beat Komodo Dragon 2.6.1 set at CI Elo of 1600. Try it yourself, and match them.

[Chessqueen]

Cicada still play weak moves at T/C game in 1 minute but look at the result. Cicada which is rated only 1537 by CCRL can beat 75% of human Fide rated below 1600 at T/C of 1 minute per game.

[pgn][Event "Computer chess game"]
[Site "MININT-UB2PIMJ"]
[Date "2022.01.11"]
[Round "?"]
[White "Dragon-2.6.1-64bit-avx2"]
[Black "Cicada"]
[Result "0-1"]
[BlackElo "1537"]
[ECO "D85"]
[Opening "Grünfeld"]
[Time "18:43:13"]
[Variation "Modern Exchange, 8.Bb5+"]
[WhiteElo "1600"]
[TimeControl "60+0"]
[Termination "normal"]
[PlyCount "70"]
[WhiteType "program"]
[BlackType "program"]



1. d4 Nf6 2. c4 g6 3. Nf3 Bg7 4. Nc3 d5 5. cxd5 Nxd5 6. e4 Nxc3 7. bxc3 c5
8. Bb5+ Bd7 9. Bxd7+ Nxd7 10. O-O e5 11. d5 f5 12. Ng5 Bf6 13. Ne6 Qa5 14.
Rb1 Rb8 15. exf5 Qxa2 16. Be3 gxf5 17. Qh5+ Ke7 18. Qxf5 Qxd5 19. Nc7 Qc6
20. Nb5 a6 21. Na7 Qc7 22. Rfd1 b6 23. Bg5 Bxg5 24. Qxg5+ Kf7 25. Qh5+ Kf8
26. Qf5+ Ke7 27. Qg4 Rbg8 28. Qh4+ Nf6 29. f4 exf4 30. c4 f3 31. Re1+ Kf7
32. g3 Rg4 33. Qh6 Rxg3+ 34. hxg3 Qxg3+ 35. Kf1 Qg2# 0-1[/pgn]