New engine: Caissa

[Rubinus]

15m+10s, both cpu=6, GUI DR3/F11-CB, book C30-39-LC0.ctg, depth 8.

Caissa 0.4 -Fritz 18 1:3=8
Caissa 0.4 -Pedone 3.1 2:3=1 Significant improvement over 0.3.
Caissa 0.4 - Chiron 5 1:4=7 It's a little better here too, but Chiron doesn't seem to be a popular opponent.
Caissa 0.4 - Revenge 1.0 1:6=5
Caissa 0.4 - Eman 7.70 1:6=5 I'm not sure if it's really Eman and not some fake-engine like Hydra-Fruit used to be. I've read such preferences that I'd expect 0:12, similar to how I'd expect to see Stockfish on Caisse.
Caissa 0.4 - Arasan 23.0.1 1:1=10 It's a lot better here than I would have expected based on the CCRL rating.
so far
against Wasp 5.00 0:1=1
all my Caissa games

[Rubinus]

Ugh! I guess it would be good to have a different rating when playing according to Syzygy, instead of 0.00... Maybe it was evaluated as a draw offer, which the opponent in a losing position naturally accepted. Maybe you could look at your database of trial games at 1+1 tempo to see how many such weird draws you have there.

[pgn][Event "15m+10"]
[Site "?"]
[Date "2021.11.25"]
[Round "6"]
[White "Deep Junior Yokohama"]
[Black "Caissa 0.4"]
[Result "1/2-1/2"]
[ECO "C32"]
[Annotator "0.11;-0.55"]
[PlyCount "145"]
[TimeControl "900+10"]

1. e4 {B 0} e5 {B 0} 2. f4 {B 0} Nf6 {B 0} 3. Nc3 {B 0} d5 {B 0} 4. exd5 {B 0}
e4 {B 0} 5. d3 {B 0} Bb4 {B 0} 6. Nge2 {B 0 Oba mimo knihovnu} Qe7 {-0.55/18 33
} 7. h3 {0.11/18 22 (a3)} O-O {-1.32/18 35} 8. a3 {-0.01/21 47} Ba5 {-1.32/19
31 (Sxc3+)} 9. Kf2 {0.11/20 31} exd3 {-1.11/19 72} 10. Qxd3 {0.04/20 17} Na6 {
-0.83/18 25 (Vd8)} 11. b4 {0.18/19 31 (Kg3)} Bb6+ {-1.34/16 5} 12. Be3 {0.11/
19 25 (Kg3)} c5 {-0.64/19 59 (Vd8)} 13. dxc6 {-0.11/20 93} bxc6 {-1.36/19 35}
14. Bxb6 {-0.11/20 23} axb6 {-0.75/18 10} 15. Re1 {-0.11/20 17} c5 {-0.71/20 36
} 16. Ng3 {-0.22/19 20} Qc7 {-0.51/20 27} 17. Re5 {-0.26/19 16 (Jce4)} cxb4 {
-2.39/20 80} 18. Nb5 {-0.26/20 33} Qd7 {-1.58/21 77} 19. Be2 {-0.13/20 34
(axb4)} Qxd3 {-0.82/19 38} 20. Bxd3 {-0.26/21 38} bxa3 {-1.66/21 22} 21. Ra1 {
-0.23/21 24} Nb4 {-1.22/22 33 (Jc5)} 22. Rxa3 {-0.38/22 25} Rxa3 {-1.21/23 49}
23. Nxa3 {-0.52/24 65} Nxd3+ {-1.32/22 26} 24. cxd3 {-0.48/24 45} Ba6 {-1.00/
23 32} 25. Nf5 {-0.48/23 7 (Je4)} Bxd3 {-2.28/22 25 (Vd8)} 26. Ne7+ {-0.60/22
15} Kh8 {0.01/0 0} 27. g4 {-0.63/22 62} h6 {-1.58/22 57 (Jd7)} 28. Re3 {-0.54/
22 39} Ne4+ {-1.52/23 24} 29. Kf3 {-0.46/22 28} Nc5 {-1.59/25 38} 30. Nd5 {-0.
54/24 70} Rb8 {-1.55/25 46} 31. Nb4 {-0.55/25 64} Bf1 {-1.49/25 43 (Vd8)} 32.
Kf2 {-0.40/21 11} Ba6 {-1.42/24 7} 33. Nxa6 {-0.45/21 12} Nxa6 {-1.43/23 6} 34.
Nb5 {-0.48/22 18} Kg8 {-1.51/24 49 (Jc5)} 35. Kf3 {-0.48/20 7 (g5)} Nc5 {-1.39/
25 34} 36. f5 {-0.57/22 36 (h4)} Rd8 {-1.48/24 27} 37. Re5 {-0.63/21 10 (Kf4)}
Kf8 {-1.60/23 24 (f6)} 38. Nc7 {-0.56/22 15 (h4)} Nd7 {-2.29/24 20 (f6)} 39.
Re3 {-0.46/21 6} Nf6 {-2.22/25 27 (Vc8)} 40. Rb3 {-0.50/21 8} Rd6 {-2.20/24 42
(Vc8)} 41. Ke2 {-0.45/20 8 (Jb5)} Rd7 {-2.69/24 19 (Vc6)} 42. Nb5 {-0.49/22 14}
Ke7 {-2.69/24 22 (h5)} 43. Ke3 {-0.43/19 6 (Jc3)} Kd8 {-2.58/23 21 (g6)} 44.
Nd4 {-0.57/20 31 (Va3)} Kc7 {-3.21/22 20} 45. Kd3 {-0.63/22 14} Rd5 {-3.44/23
21} 46. Rb1 {-0.71/21 13} Rc5 {-3.52/22 15} 47. Nb5+ {-0.73/21 21} Kd7 {-3.23/
22 18 (Kc6)} 48. Rb2 {-0.73/22 10 (Kd4)} Rc1 {-3.75/22 14 (Kc6)} 49. Rb3 {-0.
71/19 9 (Kd4)} Ra1 {-4.79/22 17 (Vd1+)} 50. Na3 {-0.91/20 18} Kc6 {-5.21/22 22
(Jd5)} 51. Nc2 {-0.91/21 13 (Vc3+)} Rd1+ {-5.38/21 12} 52. Kc4 {-0.91/22 15}
Nd5 {-5.49/22 12 (Vc1)} 53. Nb4+ {-0.90/21 6 (Jd4+)} Nxb4 {-5.45/19 2} 54. Kxb4
{-0.88/21 10} Rd4+ {-5.03/22 13} 55. Ka3 {-0.88/23 11 (Kc3)} g5 {-5.53/22 17
(f6)} 56. fxg6 {-0.72/22 8} fxg6 {-5.57/22 2} 57. Rc3+ {-0.71/22 11} Kd5 {-5.
57/26 21 (Kb5)} 58. Rb3 {-0.71/25 9} g5 {-5.66/27 11 (Kc5)} 59. Rxb6 {-0.35/24
5 (Kb2)} Rd3+ {-12.61/22 2} 60. Kb2 {-0.35/24 6} Rxh3 {-13.31/22 2} 61. Rb4 {
-0.35/26 5} Rf3 {-15.29/25 3 (Ke5)} 62. Kc2 {-0.35/25 7 (Kc1)} Rf4 {-15.56/26 2
} 63. Rb6 {-0.36/25 5 (Vb3)} Ke4 {-16.30/25 2 (Vxg4)} 64. Rxh6 {-0.37/25 6
(Vb3)} Rxg4 {#999/35 36} 65. Kd2 {-0.42/25 11} Kf3 {0.01/0 0} 66. Rh3+ {-0.50/
25 10} Rg3 {0.01/0 0 (Kf2)} 67. Rh6 {-0.59/26 15} g4 {0.01/0 0 (Vg2+)} 68. Rf6+
{-0.62/27 11} Kg2 {0.01/0 0} 69. Rg6 {-1.05/29 9} Kf2 {0.01/0 0} 70. Rg8 {-1.
22/28 7} Rg1 {0.01/0 0} 71. Rf8+ {-1.22/25 4} Kg2 {0.01/0 0} 72. Rc8 {-2.53/27
30} g3 {0.01/0 0} 73. Rg8 {-2.58/26 12 Přijímám remis} 1/2-1/2
[/pgn]

[Rubinus]

And the previous game is weird, too. A position that according to the 50 moves rule cannot be won and - the engine has exceeded the time. I guess he quit thinking it was a draw and stopped playing, he had a little over a minute left. Bug?

[pgn][Event "15m+10"]
[Site "?"]
[Date "2021.11.25"]
[Round "5"]
[White "Caissa 0.4"]
[Black "Deep Junior Yokohama"]
[Result "0-1"]
[ECO "C32"]
[Annotator "-0.35;-0.04"]
[PlyCount "118"]
[TimeControl "900+10"]

{Intel(R) Core(TM) i7-8750H CPU @ 2.20GHz W=21.9 plies; 616kN/s; C30-39-LC0.
ctg B=21.4 plies; 14 805kN/s; C30-39-LC0.ctg} 1. e4 {B 0} e5 {B 0} 2. f4 {B 0}
Nf6 {B 0} 3. Nc3 {B 0} d5 {B 0} 4. exd5 {B 0} e4 {B 0} 5. d3 {B 0} Bb4 {B 0} 6.
Nge2 {B 0 Oba mimo knihovnu} O-O {-0.04/19 24} 7. a3 {-0.35/20 32 (h3)} Ba5 {
0.00/21 48 (Sxc3+)} 8. b4 {-0.85/19 27} Bb6 {0.05/22 146} 9. dxe4 {-0.84/20 46
(Jxe4)} Ng4 {-0.09/20 19} 10. Nd4 {-1.48/19 25} Nc6 {-0.31/20 40} 11. Nf3 {-2.
16/19 30} Qf6 {-0.48/20 19} 12. Qd3 {-2.34/19 31 (Sd2)} Nf2 {-0.59/20 14} 13.
Qc4 {-1.91/19 30} Nxh1 {-0.62/21 39} 14. e5 {-2.42/19 40} Qg6 {-0.62/21 20} 15.
dxc6 {-3.02/18 26} Bf2+ {-0.54/20 35} 16. Kd1 {-2.93/19 33} bxc6 {-0.70/20 39}
17. Bd2 {-2.28/19 38} a5 {-0.56/19 14} 18. Kc1 {-2.38/20 43} axb4 {-0.62/19 19}
19. Qxb4 {-2.18/20 39} Bg4 {-0.62/19 28 (Vd8)} 20. Qe4 {-1.02/19 28} Qxe4 {-0.
52/18 18 (Sb6)} 21. Nxe4 {-0.40/20 7} Ra4 {-0.35/18 15 (Sa7)} 22. Nc3 {-1.20/
20 34 (Sd3)} Ra7 {-0.49/20 15} 23. a4 {-1.25/21 45 (Je4)} Bb6 {-0.43/20 15} 24.
e6 {-0.90/21 33 (Je4)} Bxf3 {-0.48/20 17} 25. gxf3 {-0.98/23 55} fxe6 {-0.43/
20 13} 26. Bc4 {-0.80/22 28} Nf2 {-0.40/21 25 (Kf7)} 27. f5 {-0.82/22 57
(Sxe6+)} Rfa8 {-0.47/20 13 (Kf7)} 28. Bxe6+ {-1.46/21 26} Kf8 {-0.57/19 11} 29.
Bd7 {-1.40/22 26 (Vb1)} Rd8 {-0.18/21 19} 30. Bxc6 {-0.57/23 29} Rd6 {-0.18/21
24} 31. Bb5 {-1.22/23 24} c6 {-0.19/21 23} 32. Bc4 {-0.72/23 33} Rad7 {-0.28/
22 27} 33. a5 {-1.30/23 34} Ba7 {-0.07/21 19} 34. Nb1 {-1.01/23 32} Rd4 {-0.07/
21 17} 35. Ra4 {-1.11/23 44} Rh4 {-0.12/21 18} 36. a6 {-1.36/22 31 (Se6)} h5 {
-0.36/20 17 (Vxh2)} 37. Rb4 {-1.12/21 35 (Sb4+)} c5 {-0.40/23 28} 38. Be6 {-2.
47/22 31} cxb4 {-0.47/24 22} 39. Bxd7 {-2.26/21 6} Kf7 {-0.48/23 20 (Kg8)} 40.
f4 {-2.50/21 26} Rxh2 {-0.48/23 6} 41. Bxb4 {-1.80/23 27} h4 {-0.48/22 31} 42.
Bc6 {-1.71/23 35 (Se6+)} Ng4 {-0.83/20 13} 43. Na3 {-1.91/22 18} Ne3 {-1.04/21
11} 44. Bd2 {-1.86/22 21} h3 {-1.05/22 11} 45. Nb5 {-1.50/22 10} Rg2 {-1.28/23
23} 46. Nxa7 {-1.40/21 4} h2 {-0.89/24 13} 47. Bxe3 {-1.47/20 4} h1=Q+ {-0.90/
24 28} 48. Kb2 {-1.30/21 8} Qf1 {-0.90/24 11} 49. Bd5+ {-1.39/22 37 (Sxg2)} Kf6
{-1.35/23 35} 50. Bd4+ {-1.24/23 28} Kxf5 {-1.10/23 19} 51. Bxg2 {-1.34/23 33}
Qxg2 {-1.10/21 7} 52. Nb5 {-1.04/24 20} Qc6 {-0.98/22 4} 53. c4 {-1.14/23 3}
Qxa6 {-0.92/23 10} 54. Bxg7 {-1.30/26 12} Kxf4 {-0.92/23 6} 55. Bd4 {-1.40/30
18 (Kb3)} Ke4 {-0.92/25 11 (Dc6)} 56. Kb1 {-1.23/31 18 (Kb3)} Kd3 {-3.99/25 4
(Kf5)} 57. Bb2 {-1.09/35 18} Kxc4 {-3.99/28 8 (Dc6)} 58. Nd4 {0.01/0 0} Qg6+ {
-3.99/25 7} 59. Ka2 {0.01/0 0} Qg2 {-3.99/27 9 čas} 0-1
[/pgn]

[Rubinus]

15m+10s, both cpu=6, GUI DR3/F11-CB, book C30-39-LC0.ctg, depth 8.

Caissa 0.4 - Wasp 5.00 0:1=11
Caissa 0.4 - Deep Shredder 12x64 7:1=4
Caissa 0.4 - Deep Junior Yokohama 6:1=5
Caissa 0.4 - K-Dragon 2.5.1 64bit position 0:2=10
Caissa 0.4 - K-Dragon 2.5.1 64bit aggresive 1:4=7
Caissa 0.4 - K-Dragon 2.5.1 64bit defensive 0:4=8
Caissa 0.4 - K-Dragon 2.5.1 64bit active 3:6=3
Caissa 0.4 - K-Dragon 2.5.1 64bit endgame 0:6=6
Caissa 0.4 - K-Dragon 2.5.1 64bit beginner 0:0=12
Caissa 0.4 - K-Dragon 2.5.1 64bit human 1:7=4
Caissa 0.4 - K-Dragon 2.5.1 64bit 0:7=5
----------------------------------------------------------------------
Caissa 0.4 - K-Dragon 2.5.1 64bit skill34 0:9=3 Worse than against a full level?
Caissa 0.4 - K-Dragon 2.5.1 64bit skill33 0:4=8
Caissa 0.4 - K-Dragon 2.5.1 64bit skill32 1:2=9
Even against skill31 is already a sure defeat, I would expect it to be even with 32 or so, but I guess we will have to play against 30 and 29 before the advantage is clear.
all my Caissa games

[Rubinus]

15m+10s, both cpu=6, GUI DR3/F11-CB, book C30-39-LC0.ctg, depth 8.

Caissa 0.4 - K-Dragon 2.5.1 64bit skill31 4:6=2
Caissa 0.4 - K-Dragon 2.5.1 64bit skill30 5:2=5
Caissa 0.4 - K-Dragon 2.5.1 64bit skill29 7:2=3
It's a small lot. But Caissa seems to be approaching Dragon with level 31... With that, apparently levels are limited by the depth setting, so the power probably doesn't move in a completely linear fashion. And if I estimate Caissa at CCRL level 3230 (6 cpu), then Dragon level 31 is probably 10-20 ELO better off.
all my Caissa games
I'm done testing version 0.4 for now.

[Wolfgang]

Short question: is the actual NNUE from Caissa still "from Stockfish"?

[Witek]

Short question: is the actual NNUE from Caissa still "from Stockfish"?

Yes. I'm planning to get rid of it before doing a proper "release" and replace it it with my own neural network implementation.

[Witek]

Caissa 0.5 has been released!
https://github.com/Witek902/Caissa/releases/tag/0.5

It contains huge amount of improvements and bugfixes, almost 200 Elo points gain compared to version 0.4 when using the same evaluation net.

Unfortunately, it's still using SF network, I'm in the middle of process of implementing my own NNUE and trainer for it.

[Frank Quisinsky]

Hi Michal,

I saw that in the morning, added the update information on my site.

Caissa 0.4 NNSf is still running in my tourney.
200 Elo is a lot with the same NNSf file.

v0.5 NNSf will be test with my next run-4.

Keep up your good work and nice to read with your own NN file soon.

Go Caissa go!

Best
Frank

[Sylwy]