Dragon by Komodo Chess

[lkaufman]

Very early chess960 results here are incredibly impressive. Massive jump in performance.

Yes, in a direct match Dragon beat Komodo 14.1 in chess960 (FRC) by about 200 elo in a long match on one thread at 2' + 1", and also beat SF12 in a long FRC match. We trained partly on 960 games, so this is not just a fluke.

[Steve Maughan]

<snip>We trained partly on 960 games, so this is not just a fluke.

It's just a gut feel, but I suspect this may be the secret sauce needed to create a super-aggressive evaluation function.

--- Steve

[JohnS]

The net used is embedded, so no need for the user to do anything differently than with previous Komodo versions, although there is an option (Use Regular Eval) to turn off the use of the net.

Is Dragon with Use Regular Eval the same as Komodo 14.1, thanks.

[lkaufman]

The net used is embedded, so no need for the user to do anything differently than with previous Komodo versions, although there is an option (Use Regular Eval) to turn off the use of the net.

Is Dragon with Use Regular Eval the same as Komodo 14.1, thanks.

Well, not exactly the same, a couple UCI options are set differently, and the program runs a bit slower in Dragon using regular eval than when running Komodo 14.1. Probably Komodo 14.1 is about ten elo stronger, which can be reduced to maybe five elo by duplicating the UCI options from Komodo 14.1. Eval should be the same, as well as time management and options.

[MikeB]

Customers who purchased Komodo 13 or later can get a 20% discount;

Thanks for komodo Dragon and for the discount for 13+ customers too, whilst at sales- site only 14+ are mentioned only so far.
Looking forward to buy my shiny new Dragon by Komodo Chess soon now regards

Update: The discount for K13 owners is now there. Thanks very much to Larry , Mark and team. A job well done!

[JohnS]

The net used is embedded, so no need for the user to do anything differently than with previous Komodo versions, although there is an option (Use Regular Eval) to turn off the use of the net.

Is Dragon with Use Regular Eval the same as Komodo 14.1, thanks.

Well, not exactly the same, a couple UCI options are set differently, and the program runs a bit slower in Dragon using regular eval than when running Komodo 14.1. Probably Komodo 14.1 is about ten elo stronger, which can be reduced to maybe five elo by duplicating the UCI options from Komodo 14.1. Eval should be the same, as well as time management and options.

Thanks Larry for clarifying.

[peter]

Customers who purchased Komodo 13 or later can get a 20% discount;

Thanks for komodo Dragon and for the discount for 13+ customers too, whilst at sales- site only 14+ are mentioned only so far.
Looking forward to buy my shiny new Dragon by Komodo Chess soon now regards

Update: The discount for K13 owners is now there. Thanks very much to Larry , Mark and team. A job well done!

Got it already too.
More thanks and happy regards

[BrendanJNorman]

Very, very impressed.

The play reminds me of Alpha Zero.

[pgn][Event "Elites, Blitz 2.0min+1.0sec"]
[Site "?"]
[Date "2020.11.11"]
[Round "1.2"]
[White "Dragon by Komodo Chess, 64-bit"]
[Black "Stockfish 12"]
[Result "1-0"]
[ECO "C78"]
[Annotator "0.55;0.38"]
[PlyCount "111"]
[EventDate "2020.11.11"]
[EventType "tourn"]
[SourceTitle "Fritz Engine Tournament"]

{Intel(R) Core(TM) i7-8565U CPU @ 1.80GHz 1992 MHz W=23.8 plies; 3,987kN/s
B=26.7 plies; 4,240kN/s} 1. e4 e5 2. Nf3 Nc6 3. Bb5 a6 4. Ba4 Nf6 {0.38/25 9}
5. O-O {0.55/19 4} b5 {0.38/24 6} 6. Bb3 {0.52/21 3} Bc5 {0.49/20 1} 7. a4 {
0.43/22 12} Bb7 {0.49/24 2} 8. c3 {0.46/23 6 (d3)} d6 {0.24/20 2} 9. d4 {
0.57/24 13} Bb6 {0.27/23 2} 10. Re1 {0.51/22 4 (Bg5)} O-O {0.19/23 2 (h6)} 11.
Bg5 {0.45/22 3} h6 {0.00/25 5} 12. Bh4 {0.58/22 2} g5 {0.15/24 5 (exd4)} 13.
Nxg5 {1.08/20 6 (Bg3)} hxg5 {0.00/21 1} 14. Bxg5 {1.11/20 4} exd4 {0.00/22 2}
15. Bd5 {1.13/19 3} dxc3 {0.18/25 9 (d3)} 16. Nxc3 {0.65/20 5} b4 {0.64/24 4}
17. a5 {1.88/22 5} Ba7 {0.92/28 5} 18. Bxc6 {2.23/23 3} Bxc6 {2.04/28 6} 19.
Nd5 {1.75/25 4} Bxd5 {2.08/25 2} 20. exd5 {1.78/23 2} Kg7 {1.77/27 2} 21. Qf3 {
1.80/25 3} Rh8 {1.66/26 2} 22. Rac1 {2.16/24 3} Rh5 {2.00/28 5} 23. h4 {
1.99/24 2} Rxg5 {1.64/31 4} 24. hxg5 {0.66/24 5} Nh7 {1.18/26 2} 25. Re4 {
0.92/24 6 (Rc4)} Nxg5 {0.52/24 4} 26. Qg3 {0.70/23 10} f6 {0.18/24 3 (Kg6)} 27.
Rce1 {1.20/22 10} Qg8 {2.97/26 10} 28. Re7+ {1.21/19 2} Kh6 {3.65/27 5} 29. Qf4
{0.57/21 4 (Kh2)} Kg6 {1.75/25 3} 30. Qc4 {0.68/22 2} Rf8 {1.53/22 1} 31. Qxa6
{1.29/21 4} Bxf2+ {0.94/25 3} 32. Kxf2 {1.58/20 2} Qxd5 {0.84/25 2} 33. Qe2 {
1.54/21 2} Qxa5 {0.96/25 4 (f5)} 34. Qc2+ {1.75/18 1} f5 {0.62/25 2} 35. Qxc7 {
1.78/22 3} Qc5+ {0.89/26 2 (Ne4+)} 36. Qxc5 {1.47/21 2} dxc5 {1.43/28 4} 37.
Re8 {2.24/23 1} Rxe8 {1.15/28 3} 38. Rxe8 {1.30/26 1} Kf7 {2.79/31 6} 39. Rc8 {
2.50/27 3} Ke6 {2.85/31 2} 40. Ke3 {2.48/26 1} Kd5 {2.10/35 2} 41. Rf8 {
2.72/27 2} Ke5 {3.24/34 10} 42. Kd3 {2.69/27 2} Ne6 {3.25/31 1} 43. Rf7 {
3.41/27 3} Ng5 {3.48/32 2} 44. Rd7 {3.99/27 2} Ne4 {3.57/31 4 (Ke6)} 45. Kc4 {
4.30/28 1} Ke6 {3.91/31 2} 46. Rd5 {4.57/26 1 (Rg7)} Kf6 {4.27/30 5} 47. Rd3 {
4.96/28 2 (Rd8)} Ke5 {4.24/32 3} 48. Re3 {4.90/27 1 (Rh3)} Kd6 {4.26/31 1 (Kf4)
} 49. g4 {5.67/24 1 (Kb5)} Kc6 {6.32/26 3} 50. Kb3 {6.07/24 1} Nd2+ {6.26/26 1}
51. Kc2 {6.44/24 1} f4 {5.55/28 1} 52. Re6+ {6.88/26 1} Kd5 {6.09/29 1} 53. Rf6
{6.93/30 1} Nf3 {6.69/25 2} 54. Rxf4 {6.94/30 1} Nd4+ {7.27/23 2 (Ne5)} 55.
Rxd4+ {8.35/28 1 (Kb1)} cxd4 {60.61/33 2 (Kxd4)} 56. g5 {9.17/34 4} 1-0[/pgn]



Several times in this game SF has no idea what was coming and Dragon's eval was steadily rising.




And to think people are talking about SF being "within 200 Elo of perfect chess/32 man TBs" lol

[Redshift]

Hi Larry

Slightly puzzled that the macOS and Windows versions, using Dragon NNUE eval and MCTS, output different evaluations when single threaded.

The Mac is running a Xeon W and the Windows machine an AMD 2700X. I used the non-AVX2 binary for both the below printouts but it makes no difference - same results with the AVX2 binaries, and on repeat runs.

Mac:

Code: Select all

uciok
setoption name Use MCTS value true
position fen 1r3r2/4bpkp/1qb1p1p1/3pP1P1/p1pP1Q2/PpP2N1R/1Pn1B2P/3RB2K w - - 0 1
go infinite
info depth 9 seldepth 1 time 2000 score cp -125 nodes 74 nps 37 pv e1d2
info depth 11 seldepth 3 time 4001 score cp -122 nodes 191 nps 47 pv h3h6 c2a3 h2h4 a3c2
info depth 12 seldepth 6 time 6003 score cp -68 nodes 301 nps 50 pv h3h6 c2a3 h2h4 a3c2 f3h2 a4a3 h2g4
info depth 12 seldepth 7 time 8003 score cp -41 nodes 378 nps 47 pv h3h6 c2a3 h2h4 a3c2 f3h2 a4a3 h2g4 c2e1
info depth 13 seldepth 7 time 10005 score cp -58 nodes 455 nps 45 pv h3h6 c2a3 h2h4 a3c2 f3h2 a4a3 h2g4 c2e1
info depth 13 seldepth 7 time 12006 score cp -59 nodes 571 nps 47 pv h3h6 b6a7 e1d2 c2a3 h2h4 a3c2 h4h5 a4a3
info depth 13 seldepth 9 time 14006 score cp -56 nodes 650 nps 46 pv h3h6 b6a7 e1d2 c2a3 h2h4 a3c2 h4h5 a4a3 f4h4
info depth 14 seldepth 11 time 16007 score cp -54 nodes 786 nps 49 pv h3h6 b6a7 e1d2 c2a3 h2h4 a3c2 h4h5 a4a3 h1g2 a3a2 h5g6 f7g6
info depth 14 seldepth 11 time 18009 score cp -46 nodes 859 nps 47 pv h3h6 b6a7 e1d2 c2a3 h2h4 a3c2 h4h5 a4a3 h1g2 f8h8 h5g6
info depth 14 seldepth 11 time 20010 score cp -49 nodes 909 nps 45 pv h3h6 b6a7 e1d2 c2a3 h2h4 a3c2 h4h5 a4a3 h1g2 f8h8 h5g6
info depth 14 seldepth 11 time 22010 score cp -56 nodes 991 nps 45 pv h3h6 b6a7 e1d2 c2a3 h2h4 a3c2 h4h5 a4a3 h1g2 f8h8 d1h1 a3b2
info depth 14 seldepth 11 time 24012 score cp -52 nodes 1099 nps 45 pv h3h6 b6a7 e1d2 c2a3 h2h4 a3c2 h4h5 a4a3 h1g2 f8h8 d1h1 a3b2
info depth 15 seldepth 11 time 26013 score cp -46 nodes 1173 nps 45 pv h3h6 b6a7 e1d2 c2a3 h2h4 a3c2 h4h5 a4a3 h1g2 f8h8 d1h1 a3b2
info depth 15 seldepth 11 time 28013 score cp -45 nodes 1244 nps 44 pv h3h6 b6a7 e1d2 c2a3 h2h4 a3c2 h4h5 a4a3 h1g2 f8h8 d1h1 a3b2
info depth 15 seldepth 11 time 30014 score cp -42 nodes 1314 nps 43 pv h3h6 b6a7 e1d2 c2a3 h2h4 a3c2 h4h5 a4a3 h1g2 f8h8 d1h1 a3b2
info depth 15 seldepth 11 time 32015 score cp -38 nodes 1467 nps 45 pv h3h6 b6a7 e1d2 c2a3 h2h4 a3c2 h4h5 a4a3 h1g2 f8h8 d1h1 a3b2
info depth 15 seldepth 11 time 34016 score cp -40 nodes 1501 nps 44 pv h3h6 b6a7 e1d2 c2a3 h2h4 a3c2 h4h5 a4a3 h1g2 f8h8 d1h1 a3b2
info depth 15 seldepth 11 time 36018 score cp -43 nodes 1577 nps 43 pv h3h6 b6a7 e1d2 c2a3 h2h4 a3c2 h4h5 a4a3 h1g2 f8h8 d1h1 a3b2
info depth 15 seldepth 11 time 38019 score cp -40 nodes 1686 nps 44 pv h3h6 b6a7 e1d2 c2a3 h2h4 a3c2 h4h5 a4a3 h1g2 f8h8 d1h1 a3b2
info depth 15 seldepth 11 time 40020 score cp -44 nodes 1754 nps 43 pv h3h6 b6a7 e1d2 c2a3 h2h4 a3c2 h4h5 a4a3 h1g2 f8h8 d1h1 a3b2
info depth 16 seldepth 11 time 42020 score cp -44 nodes 1890 nps 44 pv h3h6 b6a7 e1d2 c2a3 h2h4 a3c2 h4h5 a4a3 h1g2 f8h8 d1h1 a3b2
info depth 16 seldepth 11 time 44021 score cp -47 nodes 1919 nps 43 pv h3h6 b6a7 e1d2 c2a3 h2h4 a3c2 h4h5 a4a3 h1g2 f8h8 d1h1 a3b2
info depth 16 seldepth 11 time 46022 score cp -48 nodes 1993 nps 43 pv h3h6 b6a7 e1d2 c2a3 h2h4 a3c2 h4h5 a4a3 h1g2 f8h8 d1h1 a3b2
info depth 16 seldepth 11 time 48023 score cp -51 nodes 2063 nps 42 pv h3h6 b6a7 e1d2 c2a3 h2h4 a3c2 h4h5 a4a3 h1g2 f8h8 d1h1 a3b2
info depth 16 seldepth 12 time 50025 score cp -54 nodes 2142 nps 42 pv h3h6 b6a7 e1d2 c2a3 h2h4 a3c2 h4h5 a4a3 h1g2 f8h8 d1h1 a3b2
info depth 16 seldepth 12 time 52026 score cp -55 nodes 2246 nps 43 pv h3h6 b6a7 e1d2 c2a3 h2h4 a3c2 h4h5 a4a3 h1g2 f8h8 d1h1 a3b2
info depth 16 seldepth 12 time 54027 score cp -60 nodes 2324 nps 43 pv h3h6 b6a7 e1d2 c2a3 h2h4 a3c2 h4h5 a4a3 h1g2 f8h8 d1h1 a3b2
info depth 16 seldepth 12 time 56027 score cp -60 nodes 2431 nps 43 pv h3h6 b8b7 e1d2 e7a3 f4f6 g7g8 f6f4 a3b2 d1g1 a4a3 f4h4 f8e8 g1g4
info depth 16 seldepth 12 time 58028 score cp -61 nodes 2501 nps 43 pv h3h6 b8b7 e1d2 e7a3 f4f6 g7g8 f6f4 a3b2 d1g1 a4a3 f4h4 f8e8 g1g4
info depth 16 seldepth 12 time 60029 score cp -60 nodes 2575 nps 42 pv h3h6 b8b7 e1d2 e7a3 f4f6 g7g8 f6f4 a3b2 d1g1 a4a3 f4h4 f8e8 g1g4
info depth 16 seldepth 12 time 62030 score cp -57 nodes 2723 nps 43 pv h3h6 b8b7 e1d2 e7a3 f4f6 g7g8 f6f4 a3b2 d1g1 a4a3 f4h4 f8e8 g1g4
info depth 16 seldepth 12 time 64031 score cp -58 nodes 2790 nps 43 pv h3h6 b8b7 e1d2 e7a3 f4f6 g7g8 f6f4 a3b2 d1g1 a4a3 f4h4 f8e8 g1g4
info depth 16 seldepth 12 time 66032 score cp -56 nodes 2867 nps 43 pv h3h6 b8b7 h2h4 e7a3 h4h5 a3b2 f3h2 b6d8 e1d2 c2d4 c3d4
info depth 17 seldepth 12 time 68033 score cp -56 nodes 2977 nps 43 pv h3h6 b8b7 h2h4 e7a3 h4h5 a3b2 f3h2 b6d8 e1d2 c2d4 c3d4 b2d4 f4d4
info depth 17 seldepth 13 time 70035 score cp -57 nodes 3047 nps 43 pv h3h6 b8b7 h2h4 e7a3 h4h5 a3b2 f3h2 b6d8 e1d2 c2d4 f4d4 a4a3 h2g4
info depth 17 seldepth 13 time 72036 score cp -59 nodes 3124 nps 43 pv h3h6 b8b7 h2h4 e7a3 h4h5 a3b2 f3h2 b6d8 e1d2 c2d4 c3d4 b2d4 f4d4 b3b2
info depth 17 seldepth 14 time 74037 score cp -60 nodes 3201 nps 43 pv h3h6 b8b7 h2h4 e7a3 h4h5 a3b2 f3h2 b6d8 e1d2 c2d4 c3d4 b2d4 f4d4 b3b2 h2g4
info depth 17 seldepth 14 time 76038 score cp -61 nodes 3240 nps 42 pv h3h6 b8b7 h2h4 e7a3 h4h5 a3b2 f3h2 b6d8 e1d2 c2d4 c3d4 b2d4 f4d4 b3b2 h5g6
info depth 17 seldepth 14 time 78039 score cp -61 nodes 3320 nps 42 pv h3h6 b8b7 h2h4 e7a3 h4h5 a3b2 f3h2 b6d8 e1d2 c2d4 c3d4 b2d4 f4d4 b3b2 h5g6
info depth 17 seldepth 16 time 80040 score cp -58 nodes 3407 nps 42 pv h3h6 b8b7 h2h4 e7a3 h4h5 a3b2 f3h2 b6d8 e1d2 c2d4 c3d4 b2d4 h2g4 b3b2 h5g6 h7g6
stop
info nodes 3483
info depth 17 seldepth 16 time 81843 nodes 3483 nps 42 score cp -57 tbhits 0 hashfull 1000 pv h3h6 b8b7 h2h4 e7a3 h4h5 a3b2 f3h2 b6d8 e1d2 c2d4 c3d4 b2d4 h2g4 b3b2 h5g6 f7g6 f4d4
bestmove h3h6 ponder b8b7

Windows:

Code: Select all

uciok
setoption name Use MCTS value true
position fen 1r3r2/4bpkp/1qb1p1p1/3pP1P1/p1pP1Q2/PpP2N1R/1Pn1B2P/3RB2K w - - 0 1
go infinite
info depth 9 seldepth 1 time 2000 score cp -39 nodes 76 nps 38 pv f3d2
info depth 9 seldepth 2 time 4000 score cp -117 nodes 107 nps 26 pv d1b1 b6b7 h3h6
info depth 11 seldepth 2 time 6000 score cp -105 nodes 181 nps 30 pv d1b1 b6b7 h3h6
info depth 12 seldepth 4 time 8001 score cp -118 nodes 295 nps 36 pv e1d2 b6b7 d2e3 c2a3 b2a3
info depth 12 seldepth 5 time 10001 score cp -128 nodes 343 nps 34 pv e1d2 b6b7 d2e3 c2a3 b2a3 e7a3
info depth 12 seldepth 6 time 12001 score cp -139 nodes 390 nps 32 pv e1d2 b6b7 d2e3 c2a3 b2a3 e7a3 f4f6
info depth 13 seldepth 6 time 14001 score cp -108 nodes 499 nps 35 pv e1d2 b6b7 d2e3 c2a3 b2a3 e7a3 f4f6
info depth 13 seldepth 8 time 16001 score cp -102 nodes 611 nps 38 pv e1d2 b6b7 h1g2 c2a3 h3h6 a3c2 h2h4 a4a3 h4h5
info depth 13 seldepth 10 time 18002 score cp -101 nodes 687 nps 38 pv e1d2 b6b7 h1g2 c2a3 h3h6 a3c2 h2h4 a4a3 h4h5 a3b2 d1b1
info depth 14 seldepth 12 time 20002 score cp -99 nodes 796 nps 39 pv e1d2 b6b7 h1g2 c2a3 h3h6 a3c2 h2h4 a4a3 h4h5 a3b2 h5g6 f7g6 f4h2
info depth 14 seldepth 14 time 22002 score cp -95 nodes 902 nps 40 pv e1d2 b6b7 h1g2 c2a3 h3h6 a3c2 h2h4 a4a3 h4h5 a3b2 h5g6 f7g6 d1h1 f8h8 f4h2
info depth 14 seldepth 17 time 24003 score cp -90 nodes 1027 nps 42 pv e1d2 b6b7 h1g2 c2a3 h3h6 a3c2 h2h4 a4a3 h4h5 a3b2 h5g6 f7g6 d1h1 f8h8 f4h2 g7f7 f3h4 b7a8
info depth 14 seldepth 18 time 26003 score cp -90 nodes 1061 nps 40 pv e1d2 b6b7 h1g2 c2a3 h3h6 a3c2 h2h4 a4a3 h4h5 a3b2 h5g6 f7g6 d1h1 f8h8 f4h2 g7f7 f3h4 b7a8 h1f1
info depth 14 seldepth 19 time 28003 score cp -91 nodes 1065 nps 38 pv e1d2 b6b7 h1g2 c2a3 h3h6 a3c2 h2h4 a4a3 h4h5 a3b2 h5g6 f7g6 d1h1 f8h8 f4h2 g7f7 f3h4 b7a8 h1f1 f7e8
info depth 15 seldepth 22 time 30003 score cp -88 nodes 1153 nps 38 pv e1d2 b6b7 h1g2 c2a3 h3h6 a3c2 h2h4 a4a3 h4h5 a3b2 h5g6 f7g6 d1h1 f8h8 f4h2 g7f7 f3h4 b7a8 h1f1 f7e8 h2h3 c2e1 f1e1
info depth 15 seldepth 24 time 32003 score cp -80 nodes 1228 nps 38 pv e1d2 b6b7 h1g2 c2a3 h3h6 a3c2 h2h4 a4a3 h4h5 a3b2 h5g6 f7g6 d1h1 f8h8 f4h2 g7f7 f3h4 b7a8 h1f1 f7e8 h2h3 c2e1 f1e1 e8d7 e2g4
info depth 15 seldepth 28 time 34004 score cp -77 nodes 1385 nps 40 pv e1d2 b6b7 h1g2 c2a3 h3h6 a3c2 h2h4 a4a3 h4h5 a3b2 h5g6 f7g6 d1h1 f8h8 f4h2 g7f7 f3h4 b7a8 h1f1 f7e8 h2h3 c2e1 f1e1 e8d7 e2g4 d7c7 h4f3 a8a1 h6h7
info depth 15 seldepth 32 time 36004 score cp -68 nodes 1513 nps 42 pv e1d2 b6b7 h1g2 c2a3 h3h6 a3c2 h2h4 a4a3 h4h5 a3b2 h5g6 f7g6 d1h1 f8h8 f4h2 g7f7 f3h4 b7a8 h1f1 f7e8 h2h3 c2e1 f1e1 e8d7 e2g4 d7c7 h4f3 a8a1 h6h7 a1e1 d2e1 b2b1q h7h8
info depth 15 seldepth 34 time 38004 score cp -63 nodes 1559 nps 41 pv e1d2 b6b7 h1g2 c2a3 h3h6 a3c2 h2h4 a4a3 h4h5 a3b2 h5g6 f7g6 d1h1 f8h8 f4h2 g7f7 f3h4 b7a8 h1f1 f7e8 h2h3 c2e1 f1e1 e8d7 e2g4 d7c7 h4f3 a8a1 h6h7 a1e1 d2e1 b2b1q h7h8 b1c2 e1f2
info depth 15 seldepth 35 time 40004 score cp -58 nodes 1683 nps 42 pv e1d2 b6b7 h1g2 c2a3 h3h6 a3c2 h2h4 a4a3 h4h5 a3b2 h5g6 f7g6 d1h1 f8h8 f4h2 g7f7 f3h4 b7a8 h1f1 f7e8 h2h3 c2e1 f1e1 e8d7 e2g4 d7c7 h4f3 a8a1 h6h7 a1e1 d2e1 b2b1q h7h8 b1c2 e1f2 b8h8
info depth 15 seldepth 37 time 42005 score cp -53 nodes 1734 nps 41 pv e1d2 b6b7 h1g2 c2a3 h3h6 a3c2 h2h4 a4a3 h4h5 a3b2 h5g6 f7g6 d1h1 f8h8 f4h2 g7f7 f3h4 b7a8 h1f1 f7e8 h2h3 c2e1 f1e1 e8d7 e2g4 d7c7 h4f3 a8a1 h6h7 a1e1 d2e1 b2b1q h7h8 b1c2 e1f2 b8h8 h3h8 b3b2
info depth 15 seldepth 39 time 44005 score cp -50 nodes 1797 nps 40 pv e1d2 b6b7 h1g2 c2a3 h3h6 a3c2 h2h4 a4a3 h4h5 a3b2 h5g6 f7g6 d1h1 f8h8 f4h2 g7f7 f3h4 b7a8 h1f1 f7e8 h2h3 c2e1 f1e1 e8d7 e2g4 d7c7 h4f3 a8a1 h6h7 a1e1 d2e1 b2b1q h7h8 b1c2 e1f2 b8h8 h3h8 b3b2 g4e6 b2b1q
info depth 16 seldepth 42 time 46005 score cp -47 nodes 1857 nps 40 pv e1d2 b6b7 h1g2 c2a3 h3h6 a3c2 h2h4 a4a3 h4h5 a3b2 h5g6 f7g6 d1h1 f8h8 f4h2 g7f7 f3h4 b7a8 h1f1 f7e8 h2h3 c2e1 f1e1 e8d7 e2g4 d7c7 h4f3 a8a1 h6h7 a1e1 d2e1 b2b1q h7h8 b1c2 e1f2 b8h8 h3h8 b3b2 g4e6 b2b1q h8c8 c7b6 c8b8
info depth 16 seldepth 43 time 48005 score cp -47 nodes 1860 nps 38 pv e1d2 b6b7 h1g2 c2a3 h3h6 a3c2 h2h4 a4a3 h4h5 a3b2 h5g6 f7g6 d1h1 f8h8 f4h2 g7f7 f3h4 b7a8 h1f1 f7e8 h2h3 c2e1 f1e1 e8d7 e2g4 d7c7 h4f3 a8a1 h6h7 a1e1 d2e1 b2b1q h7h8 b1c2 e1f2 b8h8 h3h8 b3b2 g4e6 b2b1q h8c8 c7b6 c8b8 b6a5
info depth 16 seldepth 44 time 50005 score cp -46 nodes 1893 nps 37 pv e1d2 b6b7 h1g2 c2a3 h3h6 a3c2 h2h4 a4a3 h4h5 a3b2 h5g6 f7g6 d1h1 f8h8 f4h2 g7f7 f3h4 b7a8 h1f1 f7e8 h2h3 c2e1 f1e1 e8d7 e2g4 d7c7 h4f3 a8a1 h6h7 a1e1 d2e1 b2b1q h7h8 b1c2 e1f2 b8h8 h3h8 b3b2 g4e6 b2b1q h8c8 c7b6 c8b8 b6a5 b8a7
info depth 16 seldepth 45 time 52006 score cp -45 nodes 1894 nps 36 pv e1d2 b6b7 h1g2 c2a3 h3h6 a3c2 h2h4 a4a3 h4h5 a3b2 h5g6 f7g6 d1h1 f8h8 f4h2 g7f7 f3h4 b7a8 h1f1 f7e8 h2h3 c2e1 f1e1 e8d7 e2g4 d7c7 h4f3 a8a1 h6h7 a1e1 d2e1 b2b1q h7h8 b1c2 e1f2 b8h8 h3h8 b3b2 g4e6 b2b1q h8c8 c7b6 c8b8 b6a5 b8a7 a5b5
info depth 16 seldepth 47 time 54006 score cp -43 nodes 1958 nps 36 pv e1d2 b6b7 h1g2 c2a3 h3h6 a3c2 h2h4 a4a3 h4h5 a3b2 h5g6 f7g6 d1h1 f8h8 f4h2 g7f7 f3h4 b7a8 h1f1 f7e8 h2h3 c2e1 f1e1 e8d7 e2g4 d7c7 h4f3 a8a1 h6h7 a1e1 d2e1 b2b1q h7h8 b1c2 e1f2 b8h8 h3h8 b3b2 g4e6 b2b1q h8c8 c7b6 c8b8 b6a5 b8a7 a5b5 f3d2 b1b2
info depth 16 seldepth 51 time 56006 score cp -26 nodes 2189 nps 39 pv e1d2 b6b7 h1g2 c2a3 h3h6 a3c2 h2h4 a4a3 h4h5 a3b2 h5g6 f7g6 d1h1 f8h8 f4h2 g7f7 f3h4 b7a8 h1f1 f7e8 h2h3 c2e1 f1e1 e8d7 e2g4 d7c7 h4f3 a8a1 h6h7 a1e1 d2e1 b2b1q h7h8 b1c2 e1f2 b8h8 h3h8 b3b2 g4e6 b2b1q h8c8 c7b6 c8b8 b6a5 b8a7 a5b5 f3d2 b1b2 a7b8 b5a4 b8a7 a4b5
info depth 16 seldepth 51 time 58006 score cp -28 nodes 2262 nps 38 pv e1d2 b6b7 h1g2 c2a3 h3h6 a3c2 h2h4 a4a3 h4h5 a3b2 h5g6 f7g6 d1h1 f8h8 f4h2 g7f7 f3h4 b7a8 h1f1 f7e8 h2h3 c2e1 f1e1 e8d7 e2g4 d7c7 h4f3 a8a1 h6h7 a1e1 d2e1 b2b1q h7h8 b1c2 e1f2 b8h8 h3h8 b3b2 g4e6 b2b1q h8c8 c7b6 c8b8 b6a5 b8a7 a5b5 f3d2 b1b2 a7b8 b5a4 b8a7 a4b5
info depth 16 seldepth 51 time 60006 score cp -30 nodes 2339 nps 38 pv e1d2 b6b7 h1g2 c2a3 h3h6 a3c2 h2h4 a4a3 h4h5 a3b2 h5g6 f7g6 d1h1 f8h8 f4h2 g7f7 f3h4 b7a8 h1f1 f7e8 h2h3 c2e1 f1e1 e8d7 e2g4 d7c7 h4f3 a8a1 h6h7 a1e1 d2e1 b2b1q h7h8 b1c2 e1f2 b8h8 h3h8 b3b2 g4e6 b2b1q h8c8 c7b6 c8b8 b6a5 b8a7 a5b5 f3d2 b1b2 a7b8 b5a4 b8a7 a4b5
info depth 16 seldepth 51 time 62007 score cp -31 nodes 2445 nps 39 pv e1d2 b6b7 h1g2 c2a3 h3h6 a3c2 h2h4 a4a3 h4h5 a3b2 h5g6 f7g6 d1h1 f8h8 f4h2 g7f7 f3h4 b7a8 h1f1 f7e8 h2h3 c2e1 f1e1 e8d7 e2g4 d7c7 h4f3 a8a1 h6h7 a1e1 d2e1 b2b1q h7h8 b1c2 e1f2 b8h8 h3h8 b3b2 g4e6 b2b1q h8c8 c7b6 c8b8 b6a5 b8a7 a5b5 f3d2 b1b2 a7b8 b5a4 b8a7 a4b5
info depth 16 seldepth 51 time 64007 score cp -33 nodes 2549 nps 39 pv e1d2 b6b7 h1g2 c2a3 h3h6 a3c2 h2h4 a4a3 h4h5 a3b2 h5g6 f7g6 d1h1 f8h8 f4h2 g7f7 f3h4 b7a8 h1f1 f7e8 h2h3 c2e1 f1e1 e8d7 e2g4 d7c7 h4f3 a8a1 h6h7 a1e1 d2e1 b2b1q h7h8 b1c2 e1f2 b8h8 h3h8 b3b2 g4e6 b2b1q h8c8 c7b6 c8b8 b6a5 b8a7 a5b5 f3d2 b1b2 a7b8 b5a4 b8a7 a4b5
info depth 16 seldepth 51 time 66007 score cp -34 nodes 2585 nps 39 pv e1d2 b6b7 h1g2 c2a3 h3h6 a3c2 h2h4 a4a3 h4h5 a3b2 h5g6 f7g6 d1h1 f8h8 f4h2 g7f7 f3h4 b7a8 h1f1 f7e8 h2h3 c2e1 f1e1 e8d7 e2g4 d7c7 h4f3 a8a1 h6h7 a1e1 d2e1 b2b1q h7h8 b1c2 e1f2 b8h8 h3h8 b3b2 g4e6 b2b1q h8c8 c7b6 c8b8 b6a5 b8a7 a5b5 f3d2 b1b2 a7b8 b5a4 b8a7 a4b5
info depth 16 seldepth 51 time 68007 score cp -35 nodes 2630 nps 38 pv e1d2 b6b7 h1g2 c2a3 h3h6 a3c2 h2h4 a4a3 h4h5 a3b2 h5g6 f7g6 d1h1 f8h8 f4h2 g7f7 f3h4 b7a8 h1f1 f7e8 h2h3 c2e1 f1e1 e8d7 e2g4 d7c7 h4f3 a8a1 h6h7 a1e1 d2e1 b2b1q h7h8 b1c2 e1f2 b8h8 h3h8 b3b2 g4e6 b2b1q h8c8 c7b6 c8b8 b6a5 b8a7 a5b5 f3d2 b1b2 a7b8 b5a4 b8a7 a4b5
info depth 16 seldepth 51 time 70007 score cp -37 nodes 2729 nps 38 pv e1d2 b6b7 h1g2 c2a3 h3h6 a3c2 h2h4 a4a3 h4h5 a3b2 h5g6 f7g6 d1h1 f8h8 f4h2 g7f7 f3h4 b7a8 h1f1 f7e8 h2h3 c2e1 f1e1 e8d7 e2g4 d7c7 h4f3 a8a1 h6h7 a1e1 d2e1 b2b1q h7h8 b1c2 e1f2 b8h8 h3h8 b3b2 g4e6 b2b1q h8c8 c7b6 c8b8 b6a5 b8a7 a5b5 f3d2 b1b2 a7b8 b5a4 b8a7 a4b5
info depth 16 seldepth 51 time 72008 score cp -38 nodes 2842 nps 39 pv e1d2 b6b7 h1g2 c2a3 h3h6 a3c2 h2h4 a4a3 h4h5 a3b2 h5g6 f7g6 d1h1 f8h8 f4h2 g7f7 f3h4 b7a8 h1f1 f7e8 h2h3 c2e1 f1e1 e8d7 e2g4 d7c7 h4f3 a8a1 h6h7 a1e1 d2e1 b2b1q h7h8 b1c2 e1f2 b8h8 h3h8 b3b2 g4e6 b2b1q h8c8 c7b6 c8b8 b6a5 b8a7 a5b5 f3d2 b1b2 a7b8 b5a4 b8a7 a4b5
info depth 16 seldepth 51 time 74008 score cp -39 nodes 2907 nps 39 pv e1d2 b6b7 h1g2 c2a3 h3h6 a3c2 h2h4 a4a3 h4h5 a3b2 h5g6 f7g6 d1h1 f8h8 f4h2 g7f7 f3h4 b7a8 h1f1 f7e8 h2h3 c2e1 f1e1 e8d7 e2g4 d7c7 h4f3 a8a1 h6h7 a1e1 d2e1 b2b1q h7h8 b1c2 e1f2 b8h8 h3h8 b3b2 g4e6 b2b1q h8c8 c7b6 c8b8 b6a5 b8a7 a5b5 f3d2 b1b2 a7b8 b5a4 b8a7 a4b5
info depth 17 seldepth 51 time 76008 score cp -40 nodes 3012 nps 39 pv e1d2 b6b7 h1g2 c2a3 h3h6 a3c2 h2h4 a4a3 h4h5 a3b2 h5g6 f7g6 d1h1 f8h8 f4h2 g7f7 f3h4 b7a8 h1f1 f7e8 h2h3 c2e1 f1e1 e8d7 e2g4 d7c7 h4f3 a8a1 h6h7 a1e1 d2e1 b2b1q h7h8 b1c2 e1f2 b8h8 h3h8 b3b2 g4e6 b2b1q h8c8 c7b6 c8b8 b6a5 b8a7 a5b5 f3d2 b1b2 a7b8 b5a4 b8a7 a4b5
info depth 17 seldepth 51 time 78009 score cp -41 nodes 3110 nps 39 pv e1d2 b6b7 h1g2 c2a3 h3h6 a3c2 h2h4 a4a3 h4h5 a3b2 h5g6 f7g6 d1h1 f8h8 f4h2 g7f7 f3h4 b7a8 h1f1 f7e8 h2h3 c2e1 f1e1 e8d7 e2g4 d7c7 h4f3 a8a1 h6h7 a1e1 d2e1 b2b1q h7h8 b1c2 e1f2 b8h8 h3h8 b3b2 g4e6 b2b1q h8c8 c7b6 c8b8 b6a5 b8a7 a5b5 f3d2 b1b2 a7b8 b5a4 b8a7 a4b5
info depth 17 seldepth 51 time 80009 score cp -44 nodes 3288 nps 41 pv e1d2 b6b7 h1g2 c2a3 h3h6 a3c2 h2h4 a4a3 h4h5 a3b2 h5g6 f7g6 d1h1 f8h8 f4h2 g7f7 f3h4 b7a8 h1f1 f7e8 h2h3 c2e1 f1e1 e8d7 e2g4 d7c7 h4f3 a8a1 h6h7 a1e1 d2e1 b2b1q h7h8 b1c2 e1f2 b8h8 h3h8 b3b2 g4e6 b2b1q h8c8 c7b6 c8b8 b6a5 b8a7 a5b5 f3d2 b1b2 a7b8 b5a4 b8a7 a4b5
info depth 17 seldepth 51 time 82009 score cp -47 nodes 3360 nps 40 pv e1d2 b6b7 h1g2 c2a3 h3h6 a3c2 h2h4 a4a3 h4h5 a3b2 h5g6 f7g6 d1h1 f8h8 f4h2 g7f7 f3h4 b7a8 h1f1 f7e8 h2h3 c2e1 f1e1 e8d7 e2g4 d7c7 h4f3 a8a1 h6h7 a1e1 d2e1 b2b1q h7h8 b1c2 e1f2 b8h8 h3h8 b3b2 g4e6 b2b1q h8c8 c7b6 c8b8 b6a5 b8a7 a5b5 f3d2 b1b2 a7b8 b5a4 b8a7 a4b5
stoinfo depth 17 seldepth 51 time 84009 score cp -48 nodes 3375 nps 40 pv e1d2 b6b7 h1g2 c2a3 h3h6 a3c2 h2h4 a4a3 h4h5 a3b2 h5g6 f7g6 d1h1 f8h8 f4h2 g7f7 f3h4 b7a8 h1f1 f7e8 h2h3 c2e1 f1e1 e8d7 e2g4 d7c7 h4f3 a8a1 h6h7 a1e1 d2e1 b2b1q h7h8 b1pc2 e1f2 b8h8 h3h8 b3b2 g4e6 b2b1q h8c8 c7b6 c8b8 b6a5 b8a7 a5b5 f3d2 b1b2 a7b8 b5a4 b8a7 a4b5

info nodes 3393
info depth 17 seldepth 51 time 84424 nodes 3393 nps 40 score cp -49 tbhits 0 hashfull 1000 pv e1d2 b6b7 h1g2 c2a3 h3h6 a3c2 h2h4 a4a3 h4h5 a3b2 h5g6 f7g6 d1h1 f8h8 f4h2 g7f7 f3h4 b7a8 h1f1 f7e8 h2h3 c2e1 f1e1 e8d7 e2g4 d7c7 h4f3 a8a1 h6h7 a1e1 d2e1 b2b1q h7h8 b1c2 e1f2 b8h8 h3h8 b3b2 g4e6 b2b1q h8c8 c7b6 c8b8 b6a5 b8a7 a5b5 f3d2 b1b2 a7b8 b5a4 b8a7 a4b5
bestmove e1d2 ponder c2a3

[AndrewGrant]

<snip>We trained partly on 960 games, so this is not just a fluke.

It's just a gut feel, but I suspect this may be the secret sauce needed to create a super-aggressive evaluation function.

--- Steve

So I tried to keep the lid on this after having talked to Mark and finding out that we arrived at the same idea for two different contexts. As of a few months ago, Ethereal is trained with a 50/50 mix of FRC and Standard games. The difference in training quality and diversity was large. I did not say much about it since it was a novel idea at the time. I did quietly add it to my paper though a few weeks back.

SF's NNUE style is not trained with castling rights. I assume that therefore Komodo's NNUE is not trained with castling rights. FRC increases natural king placement diversity, which appears to be the greatest shortcoming in training NNUE's in my estimation. The field as a whole will benefit from a shift to FRC play -- after all, many engines like Demolito and Ethereal are actually FRC engines, that just happen to play Standard chess.