New engine: Caissa

[Witek]

Hi, my name is Michal and this is my first post on this forum.

I'd like to present you a new UCI chess engine - Caissa. I was developing it completely from scratch since March 2021 with zero chess programming knowledge before. At this moment it's around 2800 Elo (10+1), but I need to run some more games with different engines.
You will find sources and Windows release on Github: https://github.com/Witek902/Caissa.

It's a classic alpha-beta search engine with features like PVS, aspiration windows, quiescence search, shared transposition table, MultiPV, multithreading, late move reductions, futility prunning, SEE, MVV/LVA, killer/history heuristics, magic bitboards, Syzygy support, etc. For now I was mostly focusing on search and didn't pay much attention to evaluation - that's why I "borrowed" NNUE from Stockfish. I implemented my own neural network and trainer for it, but it's still in progress.

I was mainly inspired by this forum and Chess Programming Wiki, I took some ideas from Stockfish and Ethereal engines.

Looking forward to hearing some opinions from you

Cheers

[algerbrex]

Hi Michal, welcome to TalkChess, and congrats on making your engine public! Hopefully, someone here will look into testing your engine and getting it up on the CCRL, CGET, or something similar. And I look forward to playing around with it.

I'm curious how much Elo you estimate using Stockfish's NNUE added to your engine?

[amanjpro]

Hi, my name is Michal and this is my first post on this forum.

I'd like to present you a new UCI chess engine - Caissa. I was developing it completely from scratch since March 2021 with zero chess programming knowledge before. At this moment it's around 2800 Elo (10+1), but I need to run some more games with different engines.
You will find sources and Windows release on Github: https://github.com/Witek902/Caissa.

It's a classic alpha-beta search engine with features like PVS, aspiration windows, quiescence search, shared transposition table, MultiPV, multithreading, late move reductions, futility prunning, SEE, MVV/LVA, killer/history heuristics, magic bitboards, Syzygy support, etc. For now I was mostly focusing on search and didn't pay much attention to evaluation - that's why I "borrowed" NNUE from Stockfish. I implemented my own neural network and trainer for it, but it's still in progress.

I was mainly inspired by this forum and Chess Programming Wiki, I took some ideas from Stockfish and Ethereal engines.

Looking forward to hearing some opinions from you

Cheers

Congrats on the engine...

I believe most of the major rating list testers (CEGT, CCRL and others) are not interested in SF NNUE unless the net is trained solely on the engine's own games

[Gabor Szots]

I believe most of the major rating list testers (CEGT, CCRL and others) are not interested in SF NNUE unless the net is trained solely on the engine's own games

That's true.

[Witek]

I'm curious how much Elo you estimate using Stockfish's NNUE added to your engine?

It is ~500 Elo gain in 10+1 TC, and ~400 Elo in 1+0.1. But I know, that my classic evaluation function is really bad...

I believe most of the major rating list testers (CEGT, CCRL and others) are not interested in SF NNUE unless the net is trained solely on the engine's own games

Yes, I'm aware of that. That's why I want to get rid of Stockfish NNUE and integrate my own neural network.

[Rubinus]

2800? No.
Caissa 0.1 - Smarthink 1.98 0:12 ... (Smarthink 3025 CCRL)
Another network, the default one I didn't know where to download. nn-04f258f90427.nnue
Book - My experimental with the C30-C39 theme. 8 moves or less. 15+10.
The first few moves are fine, but then gradually the poor pricing becomes apparent. Or even tactical blindness example 1, 11.Qxb7??
I can place the parts somewhere for download on request.

example:
[pgn][Event "15m+10s"]
[Date "2021.10.26"]
[Round "7"]
[White "Caissa 0.1"]
[Black "SmarThink 1.98"]
[Result "0-1"]
[ECO "C32"]
[Annotator "0.00;-0.12"]
[PlyCount "76"]
[TimeControl "900+10"]

{Intel(R) Core(TM) i7-8750H CPU @ 2.20GHz W=18.2 plies; 1 084kN/s; C30-39-LC0.
ctg B=21.9 plies; 1 408kN/s; C30-39-LC0.ctg} 1. e4 {B 0} e5 {B 0} 2. f4 {B 0}
d5 {B 0} 3. exd5 {B 0} e4 {B 0} 4. d4 {B 0} Qxd5 {B 0} 5. c4 {B 0 Oba mimo
knihovnu} Bb4+ {-0.12/22 31} 6. Nc3 {0.00/18 27} Bxc3+ {-0.04/23 17} 7. bxc3 {
-0.01/19 51} Qa5 {-0.12/24 40} 8. Qb3 {0.00/19 61 (Je2)} Nf6 {-0.04/23 36} 9.
Ba3 {0.10/18 60 (Je2)} Bd7 {-0.50/23 26 (Jg4)} 10. Ne2 {0.06/17 50} Na6 {-0.38/
23 30} 11. Qxb7 {0.03/17 30 (Sc1)} Rb8 {-5.77/20 11} 12. Bb4 {-0.81/20 23
(Dxa7)} Qa4 {-5.64/20 8 (Dxb4)} 13. Qxa7 {-3.21/18 70} Rxb4 {-6.41/21 22} 14.
Rd1 {-3.27/17 31 (a3)} Rb2 {-7.13/21 45} 15. Qa8+ {-3.59/17 26 (f5)} Rb8 {-7.
18/22 17} 16. Qa7 {-4.28/18 24} Qxc4 {-7.34/23 23} 17. d5 {-4.48/18 44} Rb2 {
-7.36/23 50 (Jxd5)} 18. Qd4 {-4.04/17 26} Qxd4 {-7.43/23 12} 19. Rxd4 {-4.39/
19 54 (Jxd4)} Nc5 {-7.56/22 9} 20. Ng3 {-4.59/18 30 (Jc1)} Rxa2 {-7.79/21 12
(0-0)} 21. Be2 {-4.37/19 30} Ke7 {-7.93/22 13 (Vc2)} 22. O-O {-4.20/17 30} Rb8
{-8.21/22 16} 23. Re1 {-4.44/18 42 (Vfd1)} Rbb2 {-8.65/22 18} 24. Bf1 {-4.56/
17 21} h5 {-8.74/22 17} 25. h4 {-4.79/17 21} g6 {-9.29/23 49 (Kd6)} 26. Rc4 {
-4.83/17 34 (Je2)} Kd6 {-9.29/22 7} 27. Rb4 {-5.24/17 25 (Vd4)} Rf2 {-9.68/19
8 (Vc2)} 28. Be2 {-5.32/17 34 (Vbb1)} Rxf4 {-9.76/20 13} 29. c4 {-5.44/17 24
(Jf1)} Rxh4 {-10.55/23 27} 30. Rf1 {-5.66/18 30 (Vbb1)} Ke5 {-10.74/22 13 (Ke7)
} 31. Re1 {-6.33/18 23 (Vfb1)} Ng4 {-11.17/21 12 (Vf4)} 32. Rbb1 {-6.07/18 31
(Veb1)} Ne3 {-11.42/21 7} 33. Ra1 {-6.20/20 38} Rxa1 {-11.80/22 12} 34. Rxa1 {
-6.56/22 34} f5 {-12.05/23 8 (Kd4)} 35. Ra7 {-5.79/19 23 (Va5)} f4 {-12.54/22 8
} 36. Nf1 {-6.71/20 22} Nxf1 {-12.93/23 9 (Jxg2)} 37. Bxf1 {-6.79/20 18} e3 {
-13.75/23 55 (Kd6)} 38. Ra8 {-8.60/20 23 (Vxc7)} Ne4 {-20.54/20 11} 0-1
[/pgn]

[pgn][Event "15m+10s"]
[Date "2021.10.26"]
[Round "8"]
[White "SmarThink 1.98"]
[Black "Caissa 0.1"]
[Result "1-0"]
[ECO "C32"]
[Annotator "-0.02;0.04"]
[PlyCount "83"]
[TimeControl "900+10"]

{Intel(R) Core(TM) i7-8750H CPU @ 2.20GHz W=22.7 plies; 1 553kN/s; C30-39-LC0.
ctg B=19.9 plies; 1 454kN/s; C30-39-LC0.ctg} 1. e4 {B 0} e5 {B 0} 2. f4 {B 0}
d5 {B 0} 3. exd5 {B 0} e4 {B 0} 4. d4 {B 0} Qxd5 {B 0} 5. c4 {B 0 Oba mimo
knihovnu} Bb4+ {0.04/18 31} 6. Nc3 {-0.02/23 26} Bxc3+ {-0.01/19 26} 7. bxc3 {
-0.01/23 15} Qa5 {0.09/19 41} 8. Qb3 {-0.07/24 53} Nh6 {0.08/17 26 (Jf6)} 9.
Ne2 {0.33/23 37 (Sa3)} O-O {-0.01/18 42} 10. Ng3 {0.32/23 24} Nc6 {-0.14/18 42
(Ve8)} 11. Qa3 {0.22/22 13} Re8 {-0.22/19 23} 12. h3 {0.50/24 37 (d5)} Qxa3 {
-0.30/19 34 (b6)} 13. Bxa3 {0.45/24 14} Nf5 {-0.31/19 23 (b6)} 14. Nxf5 {0.94/
23 10} Bxf5 {-0.19/21 44} 15. d5 {0.92/24 23 (g4)} Na5 {-0.25/21 50 (Jd8)} 16.
g4 {1.44/25 36 (Se2)} Bd7 {-0.25/20 23} 17. Bb4 {1.61/26 13 (Se2)} b6 {-0.51/
20 31} 18. Bxa5 {1.61/26 15 (Se2)} bxa5 {-0.52/20 35} 19. Be2 {1.61/25 57 (c5)}
Rab8 {-0.61/19 30 (e3)} 20. Kf2 {1.72/23 12 (0-0-0)} Ba4 {-0.34/18 44 (g6)} 21.
Rhb1 {2.27/25 68} e3+ {-0.08/20 40} 22. Kg3 {2.39/26 38} Bc2 {0.12/21 28} 23.
Rxb8 {2.41/25 48} Rxb8 {0.67/22 54} 24. c5 {2.58/24 24} h5 {0.74/20 27 (f5)}
25. Kf3 {3.27/19 7} Rd8 {0.85/20 36 (h4)} 26. Rc1 {4.04/20 8 (Sc4)} Bh7 {1.42/
20 35 (Sa4)} 27. c4 {3.85/21 8} hxg4+ {1.64/19 46 (Ve8)} 28. hxg4 {4.54/21 10}
Re8 {1.76/21 31 (Kf8)} 29. d6 {4.77/22 9 (Vc3)} cxd6 {1.41/20 25} 30. cxd6 {5.
16/22 7} Be4+ {1.51/21 24} 31. Kxe3 {5.58/23 10} Bc6+ {1.59/21 21} 32. Kd2 {5.
98/23 13 (Kf2)} Re4 {1.10/19 27 (Kf8)} 33. c5 {6.25/23 13 (Vd1)} Rd4+ {1.39/18
25 (Vxf4)} 34. Kc3 {8.11/23 37 (Ke1)} Re4 {1.93/21 41} 35. Bd3 {8.34/23 12} Re6
{2.11/20 29 (Ve3)} 36. Rb1 {11.01/22 23 (a4)} Re8 {2.86/20 38 (Ve3)} 37. Bb5 {
13.65/22 13 (a4)} Bxb5 {5.15/22 33 (Vb8)} 38. Rxb5 {13.66/20 14} Rd8 {5.98/21
34 (Kf8)} 39. Rxa5 {15.95/22 61} Kf8 {6.79/22 20 (a6)} 40. Rxa7 {16.57/19 24}
g6 {9.25/23 28} 41. Kc4 {22.12/20 56 (Kd4)} Kg8 {11.52/21 19 (g5)} 42. Kd5 {31.
38/18 42 (d7)} 1-0
[/pgn]

[Rubinus]

Caissa 0.1 - Nimzo 7.32 so far 3:2=1. Same conditions as before against SmarThink.

[Witek]

2800? No.
Caissa 0.1 - Smarthink 1.98 0:12 ... (Smarthink 3025 CCRL)
Another network, the default one I didn't know where to download. nn-04f258f90427.nnue
Book - My experimental with the C30-C39 theme. 8 moves or less. 15+10.
The first few moves are fine, but then gradually the poor pricing becomes apparent. Or even tactical blindness example 1, 11.Qxb7??
I can place the parts somewhere for download on request.

Hey, huge thanks for testing!

Yeah, I forgot to mention which network should be used. I was mostly using nn-04cf2b4ed1da.nnue - the engine only supports older NNUE (the one from SF12, maybe you tried to load too new network).

That's quite possible that the Elo is much lower than 2800 in 15+10 TC. I was mostly testing in much faster, e.g. 30+1, which is probably my bad... I will for sure do some more testing in slow-paced TCs.

You can PM me PGNs of your games, that would be very helpful.

Cheers

[Rubinus]

I'd use your network if I knew where to download it. I just couldn't find it through Google. I got to Stockfish's site where I downloaded one with a similar name. If you post a link to your net here, I can download it and test with it. It's handy for me, I'm building a library for a king gambit and need different engines to verify it. So I'm happy to run a few dozen more games.
games (Caissa-Nimzo 7.32 5:6=1, Nimco 8 CCRL 2503, 7.32 is cca -20 ELO) ----> Caissa cca 2450.

I made a makeshift logo from this image, but maybe someone has some rights to this image...

Well, I'm sure someone will create a logo for you, like Graham Banks.

[Rubinus]

I already found the net, very old, from 12.10.2020?