Discussion of anything and everything relating to chess playing software and machines.
Moderator: Ras
Ciekce
Posts: 192Joined: Sun Oct 30, 2022 5:26 pmFull name: Conor Anstey
Post by Ciekce Thu Jan 02, 2025 11:09 am
the engine programming list is not a rating list per se, just a list of engines with authors in the engine programming server
the ratings listed are from CCRL, as the column headers indicate
LazySMP
Post by LazySMP Thu Jan 02, 2025 7:13 pm
jefk wrote: ↑ Thu Jan 02, 2025 4:17 am
So your idea of 'success' is being on the Ccrl list ? Get real. Learning programming/improving programming skills is a fine goal, but making a chess engine elo 2500 imo isn't a success; maybe it would be for some AI, not for humans anymore.
My engine has already been tested in CCRL Blitz by Mr. Gabor.
https://computerchess.org.uk/ccrl/404/c ... 5.0 64-bit
Gabor Szots wrote: ↑ Tue Nov 19, 2024 8:09 am
I have downloaded v5 and now I have free CPU time as well so I could start testing it immediately.
Also, if I can implement FRC, it will be tested by Mr. Basti as well.
bastiball wrote: ↑ Mon Nov 25, 2024 11:46 pm
I only test FRC and Chess324 as of the moment. I hope you implement FRC so I can test it
My engine is only untested on CCRL 40/15. Graham had said that if my engine reached 2400, it would be tested.
Graham Banks wrote: ↑ Sat Nov 16, 2024 2:50 am
If you can eventually reach 2400, I'll be happy to include it in my Amateur Series.
When my engine reached a rating of 2500, he didn't test it and even said:
Graham Banks wrote: ↑ Wed Jan 01, 2025 7:15 pm
You're like that one small piece of dog shit that one just can't shake off their walking shoes.
This is not only unprofessional, but also simply disrespectful. As I said in previous comments, I would never want my engine tested by Graham.
LazySMP
Post by LazySMP Thu Jan 02, 2025 10:12 pm
[pgn]AlphaZero' s developers. Please note that Giraffe supports Transposition Table but LazySMP does not.
LazySMP
Post by LazySMP Fri Jan 03, 2025 4:09 am
Ras wrote: ↑ Fri Nov 15, 2024 12:03 pm LazySMP wrote: ↑ Fri Nov 15, 2024 7:45 am
What is your engine rating in CCRL?
Around 2700.
Hi Rasmus! Are you sure CT800 rating is around 2700?
[pgn][Event "Blitz 4.0min+2.0sec"]
[Date "2025.01.03"]
[Round "1"]
[White "LazySMP Version 9.0"]
[Black "CT800 V1.46 64 bit"]
[Result "1-0"]
[Variant "Standard"]
[TimeControl "240+2"]
[ECO "B02"]
[Opening "Alekhine Defense: Normal Variation"]
1. e4 Nf6 2. e5 Nd5 { B02 Alekhine Defense: Normal Variation } 3. Nf3 d6 4. Nc3 dxe5 5. Nxe5 Nxc3 6. bxc3 Nd7 7. Qh5 g6 8. Bc4 e6 9. Qe2 Be7 10. d4 Nxe5 11. dxe5 c6 12. O-O b5 13. Bd3 O-O 14. Bh6 Re8 15. a4 bxa4 16. Rxa4 Bb7 17. h4 Qc7 18. h5 c5 19. Rg4 Kh8 20. Qe3 Bc6 21. Rf4 Bd8 22. hxg6 hxg6 23. Rg4 Kg8 24. Bxg6 fxg6 25. Rxg6+ Kh7 26. Qg3 Be4 27. Rg7+ Kxh6 28. Rxc7 Bxc7 29. Qe3+ Kg7 30. Qxe4 Rab8 31. Qg4+ Kf7 32. Qh5+ Kg7 33. Qg5+ Kf8 34. Qf6+ Kg8 35. Ra1 { Black resigns. } 1-0[/pgn]
LazySMP 9.0 beat CT800 V1.46 in 30 moves
Modern Times
Posts: 3699Joined: Thu Jun 07, 2012 11:02 pm
Post by Modern Times Fri Jan 03, 2025 4:25 am
LazySMP wrote: ↑ Fri Jan 03, 2025 4:09 am
Are you sure CT800 rating is around 2700?
You could look at the CCRL lists yourself and check.
Ras
Posts: 2694Joined: Tue Aug 30, 2016 8:19 pmFull name: Rasmus Althoff
Post by Ras Fri Jan 03, 2025 7:37 am
LazySMP wrote: ↑ Fri Jan 03, 2025 4:09 am Hi Rasmus! Are you sure CT800 rating is around 2700?
Yes.
LazySMP 9.0 beat CT800 V1.46 in 30 moves
King safety has always been an engine weakness, so losses like this can happen, and if they do, the game is shorter naturally because it doesn't go into an endgame. However, it's not about how quickly the engine loses if it does, but how often, with varying openings of course. If you say LazySMP 9.0 should be around 2500, that's 200 Elo difference, so LazySMP should get about 2500 points in a 10000 games match - including quite some wins, obviously.
I usually test at 10s per game for each side, no increment. With an eight core CPU plus hyperthreading, running 14 games in parallel isn't an issue with a proper engine driver; I'm using c-chess-cli under Linux. CPU turbo boost should be switched off to get a more constant CPU speed. 10k games take something like 3 hours or so.
Daniel Anulliero
Posts: 770Joined: Fri Jan 04, 2013 4:55 pmLocation: Nice
Post by Daniel Anulliero Fri Jan 03, 2025 11:35 am
LazySMP wrote: ↑ Fri Jan 03, 2025 4:09 am Ras wrote: ↑ Fri Nov 15, 2024 12:03 pm LazySMP wrote: ↑ Fri Nov 15, 2024 7:45 am
What is your engine rating in CCRL?
Around 2700.
Hi Rasmus! Are you sure CT800 rating is around 2700?
LazySMP 9.0 beat CT800 V1.46 in 30 moves
You play ONE game vs ct800 and you claim your "thing" is better than ct800 ? How "funny" guy you are .
Remember another guy who test with 1-2 games and never follow experts advices , but I can't remember his name exactly
Welll ..........
LazySMP
Post by LazySMP Fri Jan 03, 2025 8:17 pm
Ras wrote: ↑ Fri Jan 03, 2025 7:37 am
King safety has always been an engine weakness, so losses like this can happen, and if they do, the game is shorter naturally because it doesn't go into an endgame.
You're right. BTW, king safety is the most important part of the evaluation. I am not familiar with your engine code, otherwise I would have helped you to fix its weakness. I think that you don't spend as much time on improving the evaluation function as you do on improving the search function.
Daniel Anulliero wrote: ↑ Fri Jan 03, 2025 11:35 am
You play ONE game vs ct800 and you claim your "thing" is better than ct800 ?
No, my point is CCRL rating may be wrong. Please see LazySMP 5.0 rating which is 2070.
Ras
Posts: 2694Joined: Tue Aug 30, 2016 8:19 pmFull name: Rasmus Althoff
Post by Ras Fri Jan 03, 2025 8:28 pm
LazySMP wrote: ↑ Fri Jan 03, 2025 8:17 pm You're right. BTW, king safety is the most important part of the evaluation.
There is ofc king safety code in there, which is why in the game, the CT800 castles properly and sets up a standard pawn structure. What it doesn't see is king attacks coming in. It's not that I havn't tried tons of stuff, with danger zones, scoring attackers and whatnot - but it never worked. The software architecture is also limited by hardware constraints - the UCI version is only the testbed, the actual project is the microcontroller version with 1MB ROM for code/data and 192kB RAM for dynamic data (variables, stack, hash tables) with only 16kB RAM remaining as future reserve.
No, my point is CCRL rating may be wrong.
It's consistent with my own tests, usually within 20 Elo. However, I test with six different engines, 10000 games each, plus 50000 games against the previous CT800 version, and an 80000 opening position book to avoid duplicate games. Unless you even start to ramp up your testing, all you see is error margins.
LazySMP
Post by LazySMP Sat Jan 04, 2025 3:37 am
Ras wrote: ↑ Fri Jan 03, 2025 8:28 pm
There is ofc king safety code in there, which is why in the game, the CT800 castles properly and sets up a standard pawn structure. What it doesn't see is king attacks coming in. It's not that I haven't tried tons of stuff, with danger zones, scoring attackers and whatnot - but it never worked.
You're right. It's too hard! I remember many years ago when Joerg Oster tuned values for piece check and attack unit factors, king safety went down.
Please see this:
https://github.com/official-stockfish/S ... dd82d0fadf
Joerg Oster wrote:
A middle ground patch of two successful tuning patches, one at STC, the other at LTC, which now passed both.
Code: Select all
const int QueenCheck = 52; const int QueenCheck = 62;
const int RookCheck = 45; const int RookCheck = 57;
const int BishopCheck = 5; const int BishopCheck = 48;
const int KnightCheck = 17; const int KnightCheck = 78;
