Page 2 of 3
Re: Torch 2 is here
Posted: Mon Mar 25, 2024 8:58 am
by ThatsIt
CEGT 40/20
http://www.cegt.net/40_40%20Rating%20Li ... liste.html
Code: Select all
1. Stockfish 16.1 x64 1CPU 3606 +-14 1712 games
2. Torch 2 x64 1CPU 3590 +-13 1619 games
3. Komodo Dragon 3.3 x64 1CPU 3581 +-11 2687 games
...
Re: Torch 2 is here
Posted: Mon Mar 25, 2024 9:31 am
by AndrewGrant
The +- error bars are quite strong here.
Torch is much closer to SF16.1 than Dragon 3.3, at least that is what is seen in UHO books.
Hopefully more games will nudge Torch up.
Re: Torch 2 is here
Posted: Mon Mar 25, 2024 9:51 am
by Graham Banks
AndrewGrant wrote: ↑Mon Mar 25, 2024 9:31 am
The +- error bars are quite strong here.
Torch is much closer to SF16.1 than Dragon 3.3, at least that is what is seen in UHO books.
Hopefully more games will nudge Torch up.
If you gave closely rated humans UHO openings to play, their ratings would be artificially skewed as well.
Re: Torch 2 is here
Posted: Mon Mar 25, 2024 9:55 am
by AndrewGrant
Graham Banks wrote: ↑Mon Mar 25, 2024 9:51 am
AndrewGrant wrote: ↑Mon Mar 25, 2024 9:31 am
The +- error bars are quite strong here.
Torch is much closer to SF16.1 than Dragon 3.3, at least that is what is seen in UHO books.
Hopefully more games will nudge Torch up.
If you gave closely rated humans UHO openings to play, their ratings would be artificially skewed as well.
Absolutely; no contest here.
Its just difficult to elicit an elo difference outside error bars with these books, without tens of thousands of games.
As the underdog, there is value in the current book. With some luck, Torch 2 could have come out ahead of Stockfish 16.1. But if you did the same test using highly unbalanced books, the probability of that upset goes from a couple percent, to a few thousandths of a percent.
This is why I like Fischer. Its got some inherent imbalance, and an inherent understanding that it must be played as game-pairs, even for humans. So you can still see elo differences in engines without massive hardware... but are not crushing the spirit of chess so much.
Re: Torch 2 is here
Posted: Mon Mar 25, 2024 2:49 pm
by Modern Times
AndrewGrant wrote: ↑Mon Mar 25, 2024 9:55 am
This is why I like Fischer. Its got some inherent imbalance, and an inherent understanding that it must be played as game-pairs, even for humans. So you can still see elo differences in engines without massive hardware... but are not crushing the spirit of chess so much.
Yes, I've been a big fan of this for several years.
Re: Torch 2 is here
Posted: Tue Apr 02, 2024 1:32 pm
by AlexChess
I don't know Chess.com as playground for chess engines. Is it possible to let engines play against a Torch 2 instance?
(Like on Lichess.org I mean, preferably using APIs)
Thanks
PS: Maybe I could play against it manually, times ago I have played against MITTENS and other bots probably running Dragon, but they were only avatars ordered by strength.
Re: Torch 2 is here
Posted: Tue Apr 02, 2024 1:36 pm
by AndrewGrant
AlexChess wrote: ↑Tue Apr 02, 2024 1:32 pm
I don't know Chess.com as playground for chess engines. Is it possible to let engines play against a Torch 2 instance?
(Like on Lichess.org I mean, preferably using APIs)
Thanks
PS: Maybe I could play against it manually, times ago I have played against MITTENS and other bots probably running Dragon, but they were only avatars ordered by strength.
There is no interface for it. Also, the version in the browser is NOT the same as "proper Torch". You won't be able to make a meaningful comparison using the browser version, even if you try your best to account for things like the browser being slower.
Re: Torch 2 is here
Posted: Tue Apr 02, 2024 6:17 pm
by AlexChess
AndrewGrant wrote: ↑Tue Apr 02, 2024 1:36 pm
AlexChess wrote: ↑Tue Apr 02, 2024 1:32 pm
I don't know Chess.com as playground for chess engines. Is it possible to let engines play against a Torch 2 instance?
(Like on Lichess.org I mean, preferably using APIs)
Thanks
PS: Maybe I could play against it manually, times ago I have played against MITTENS and other bots probably running Dragon, but they were only avatars ordered by strength.
There is no interface for it. Also, the version in the browser is NOT the same as "proper Torch". You won't be able to make a meaningful comparison using the browser version, even if you try your best to account for things like the browser being slower.
Thank you very much Andrew for these useful infos and congratulations for your new engine!
Re: Torch 2 is here
Posted: Wed Apr 03, 2024 1:06 pm
by Jouni
Huge CCRL progress
Torch v1 64-bit 8CPU 3798
Torch v2 64-bit 8CPU 3798
Torch v1 64-bit 3752
Torch v2 64-bit 3751
Re: Torch 2 is here
Posted: Wed Apr 03, 2024 1:17 pm
by Uri Blass
Jouni wrote: ↑Wed Apr 03, 2024 1:06 pm
Huge CCRL progress
Torch v1 64-bit 8CPU 3798
Torch v2 64-bit 8CPU 3798
Torch v1 64-bit 3752
Torch v2 64-bit 3751
I see some small FRC progress in CCRL
Torch v2 4036 +13 −13 68.0% −119.9 58.2% 2219
Torch v1 4021 +11 −11 65.1% −97.0 62.0% 3388
Berserk progressed faster in FRC that was enough to have rating above komodoDragon but still did not capture Torch
Berserk 13 3998 +17 −17 55.5% −33.4 70.1% 1092
Dragon 3.3 by Komodo 3982 +11 −11 59.1% −59.0 65.1% 3558
Berserk 12.1 3966 +12 −12 55.9% −36.6 65.1% 2464
Berserk 12 3942 +15 −15 57.6% −55.9 55.7% 1380