Stockfish 15.1 is ready

[Uri Blass]

Progress for SF 15.1 over SF 15.0 :

Each test : 60'000 games, Time Control = 60s+0.6s, 1 thread.

+17.04 : Book UHO (unbalanced-human-openings)
+53.51 : Double Fischer Random Chess
+29.45 : Fischer Random Chess
+7.46 : Standard Chess, Standard Book


From : https://tests.stockfishchess.org/tests/ ... ltc_only=1

I finally see some progress in CCRL 40/15 rating list(not with 15.1 but with something newer than 15).
https://ccrl.chessdom.com/ccrl/4040/rat ... t_all.html


Of course we need more games to be sure that it is not luck but the results seem promising and after a long time we may have something that is more than 20 elo stronger than stockfish13 against non stockfish engines in the CCRL rating list

Note that I have no idea why CCRL decided to test 20.11.2022 version
and I have no idea why stockfish decided to release 15.1 not a long time later.

It could be more natural in case the CCRL test 15.1 instead of some developement version.

Stockfish 20221120 64-bit 4CPU 3552 +21 −21 64.7% −80.7 70.6% 596
Stockfish 14 64-bit 4CPU 3530 +15 −15 65.1% −87.8 68.7% 1316
Stockfish 15 64-bit 4CPU 3530 +14 −14 66.7% −97.7 66.4% 1512
ShashChess 26.0 64-bit 4CPU 3528 +23 −22 61.8% −65.3 75.5% 498
Stockfish 13 64-bit 4CPU 3526 +14 −14 71.6% −137.1 56.6% 1544

[Modern Times]

+27 Elo (bayeselo) on CCRL 4015 single thread after 400 games.

[Ajedrecista]

Hello Eduard:

But there are also tournaments where one side has a good advantage! TCEC for example. Or now on InfinityChess similar to TCEC. I don't want to despise +8 Elo!

To put in in perspective, knowing that E Elo difference is by definition:

Code: Select all

E := Elo difference (measured in logistic Elo).

W := win ratio.
L := lose ratio.

E = 400 × log10{ [ 1/2 + (W-L)/2 ] / [ 1/2 - (W-L)/2 ] }
[...]
W-L = [ 10^(E/400) - 1 ] / [ 10^(E/400) + 1 ]

That is, E ~ +7.46 Elo (Fishtest match here) means W-L ~ 0.0215 = 2.15% more less. [Correct: (3430 - 2142)/60000 ~ 0.0215].

Long time ago, I calculated an approximation for small Elo differences (Taylor series) that 1% difference in W-L translates to 8/ln(10) Elo ~ 3.47 Elo. (Less accurate as Elo differences become larger). Nothing new at all: there is an old rule of the thumb of circa 7 Elo difference for each 1% over 50% score (that percentage is half of W-L) [with my approach: 2 × 8/ln(10) ~ +6.95 Elo].

Regards from Spain.

Ajedrecista.

[Modern Times]

After 568 games on CCRL 4015 one thread, it is +28 elo (bayeselo) and +37 Ordo elo. I'd say those are good gains for what was already an incredible engine.

[Dariusz]

After 568 games on CCRL 4015 one thread, it is +28 elo (bayeselo) and +37 Ordo elo. I'd say those are good gains for what was already an incredible engine.

https://www.chessengeria.com/post/stoc ... 1-released

Stockfish 15.1 compiled for Macs is noticeably stronger than its predecessor.

[pocpit]

Any opinion about this new evaluation calibration? (or however this should be called):

"An evaluation of +1 is now no longer tied to the value of one pawn, but to the likelihood of winning the game. With a +1 evaluation, Stockfish has now a 50% chance of winning the game against an equally strong opponent."

[Werewolf]

After 568 games on CCRL 4015 one thread, it is +28 elo (bayeselo) and +37 Ordo elo. I'd say those are good gains for what was already an incredible engine.

https://www.chessengeria.com/post/stoc ... 1-released

Stockfish 15.1 compiled for Macs is noticeably stronger than its predecessor.

Great news. This may add weight to Magnum's view that previous compiles weren't that good.

One way the SF Dev team could show they care about Apple Silicon is by releasing compiles as standard for it, as they already do for Windows and Linux

[criko]

my version for Mac M1

https://www.mediafire.com/file/ggx10fdq ... 1.zip/file

[Jouni]

Good thing is, that downhill of nps is stopped. 6 cores bench

Code: Select all

SF14     14,4Mnps
SF14.1   13,1
SF15     11,4
SF15.1   11,8

[Wolfgang]

After 568 games on CCRL 4015 one thread, it is +28 elo (bayeselo) and +37 Ordo elo. I'd say those are good gains for what was already an incredible engine.

We have a regression of approx. 20 ELO after 800 games
(only!) with shorter TC
https://cegt.forumieren.com/t1830-testi ... ish-15-1nn