New engine: Caissa

[Witek902]

Hi,

My name is Michal. Over last year I was developing a new chess engine from scratch. Version 0.7 finally qualifies for CCRL testing, as I dropped any external dependencies (except from TB probing code) and implemented my own neural network for evaluation. Would it be possible to run some CCRL tournaments so it gets assigned "official" rating?

Source code and Windows executables are available on Github: https://github.com/Witek902/Caissa

Estimated Elo: 3100

More details can be found on TalkChess forum: http://talkchess.com/forum3/viewtopic.p ... faff3f98c9

[Gabor Szots]

Hi,

My name is Michal. Over last year I was developing a new chess engine from scratch. Version 0.7 finally qualifies for CCRL testing, as I dropped any external dependencies (except from TB probing code) and implemented my own neural network for evaluation. Would it be possible to run some CCRL tournaments so it gets assigned "official" rating?

Source code and Windows executables are available on Github: https://github.com/Witek902/Caissa

Estimated Elo: 3100

More details can be found on TalkChess forum: http://talkchess.com/forum3/viewtopic.p ... faff3f98c9

Hi Michal,

I'm starting a test gauntlet tomorrow.

[Witek902]

FYI, Caissa 0.8 has been released: https://github.com/Witek902/Caissa/releases/tag/0.8

Estimated Blitz Elo: 3240

[Graham Banks]

In ChessGUI, no parameters are showing up, but the engine runs fine with whatever the default parameters are.
I am using 256mb hash plus 1 core, but for 4CPU testing (assuming Caissa is SMP), I'll need to use 1024mb hash and set cores to 4.

[bastiball]

FYI, Caissa 0.8 has been released: https://github.com/Witek902/Caissa/releases/tag/0.8

Estimated Blitz Elo: 3240

I'm currently testing 8CPU so Gabor will test this version I think.

[Gabor Szots]

I'm currently testing 8CPU so Gabor will test this version I think.

Maybe later. My hands are full at the moment.

[Witek902]

In ChessGUI, no parameters are showing up, but the engine runs fine with whatever the default parameters are.
I am using 256mb hash plus 1 core, but for 4CPU testing (assuming Caissa is SMP), I'll need to use 1024mb hash and set cores to 4.

Thanks for letting know. I will check that and try to fix in next release. Default hash table size is 256MB. "Hash" and "Threads" UCI options are supported by the engine. My guess something is wrong with reporting list of available options.

[bastiball]

I forgot to tell that if it's okay if I invite this engine as well in upcoming Mini-TCEC Season 2?

[Witek902]

Caissa 0.9 is out now: https://github.com/Witek902/Caissa/releases/tag/0.9

~70 Elo improvement compared to version 0.8

[Witek902]

Caissa 1.0 is out now: https://github.com/Witek902/Caissa/releases/tag/1.0

Code: Select all

Elo difference: 73.5 +/- 20.0, LOS: 100.0 %, DrawRatio: 32.9 %

[bastiball]

Caissa 1.0 is out now: https://github.com/Witek902/Caissa/releases/tag/1.0

Code: Select all

Elo difference: 73.5 +/- 20.0, LOS: 100.0 %, DrawRatio: 32.9 %

Congrats! Any chances to implement FRC?

[Witek902]

Caissa 1.0 is out now: https://github.com/Witek902/Caissa/releases/tag/1.0

Code: Select all

Elo difference: 73.5 +/- 20.0, LOS: 100.0 %, DrawRatio: 32.9 %

Congrats! Any chances to implement FRC?

Yes, I'm planning to add FRC support soon.

[bastiball]

Caissa 1.0 is out now: https://github.com/Witek902/Caissa/releases/tag/1.0

Code: Select all

Elo difference: 73.5 +/- 20.0, LOS: 100.0 %, DrawRatio: 32.9 %

Congrats! Any chances to implement FRC?

Yes, I'm planning to add FRC support soon.

Can't wait to test it in FRC

[Witek902]

Any chances to implement FRC?

FRC support implemented in version 1.1: https://github.com/Witek902/Caissa/releases/tag/1.1

Estimated CCRL 40/2 FRC rating: 3270

I only tested under cutechess-cli so there can be bugs.

[bastiball]

Any chances to implement FRC?

FRC support implemented in version 1.1: https://github.com/Witek902/Caissa/releases/tag/1.1

Estimated CCRL 40/2 FRC rating: 3270

I only tested under cutechess-cli so there can be bugs.

I'll put it on a queue

[Graham Banks]

Going to replace Caissa 1.0 in here - https://tlcv.net/16094

[Witek902]

Going to replace Caissa 1.0 in here - https://tlcv.net/16094

Makes sense. I don't expect significant Elo change compared to 1.0. The net is also the same.

[Graham Banks]

Caissa games don't seem to update after each move in the broadcast. I'll attach a couple of debug files for you.

Let me know when you've got them so that I can delete them.

[Witek902]

Caissa games don't seem to update after each move in the broadcast. I'll attach a couple of debug files for you.

Let me know when you've got them so that I can delete them.

I got them. My suspicion is castling moves in PV line. After adding Chess960 support, they are now reported as "e1h1" for instance, instead of "e1g1". I have to fix it. I wonder what happens if casting is reported as best move...

[Graham Banks]

Caissa games don't seem to update after each move in the broadcast. I'll attach a couple of debug files for you.

Let me know when you've got them so that I can delete them.

I got them. My suspicion is castling moves in PV line. After adding Chess960 support, they are now reported as "e1h1" for instance, instead of "e1g1". I have to fix it. I wonder what happens if casting is reported as best move...

Not sure, but Caissa 1.1 has the same issue.

[bastiball]

Caissa playing well in FRC

[bastiball]

https://www.computerchess.org.uk/ccrl/4 ... Caissa_1_1

[Witek902]

Version 1.2 released: https://github.com/Witek902/Caissa/releases/tag/1.2

~60 Elo gain in regular chess and ~75 Elo gain in FRC.

Issue with PV printing should be now fixed.

[Witek902]

Caissa 1.3 released: https://github.com/Witek902/Caissa/releases/tag/1.3

Most gain visible in ultra-fast time controls and "X moves in Y time" time controls. Also some small gain in FRC.

For instance, in 1sec per whole game there's ~50 Elo gain compared to 1.2 - that's only ~200 Elo less than Stockfish 15

[Witek902]

Caissa 1.4 released: https://github.com/Witek902/Caissa/releases/tag/1.4

Fixed pondering bug. Quite significant Elo gain in FRC.