Stockfish 070324 a Disaster

Discussion of anything and everything relating to chess playing software and machines.

Moderators: hgm, chrisw, Rebel

Eduard
Posts: 1439
Joined: Sat Oct 27, 2018 12:58 am
Location: Germany
Full name: N.N.

Re: Stockfish 070324 a Disaster

Post by Eduard »

I am still pleased that the new NNUE with Dimension 3072 has defeated the Master Network with Dimension 2560:

Fishtest:
SPRT nn-8747f5ce199f.nnue vs master diff
LLR: -0.01 (-2.94,2.94) <0.00,2.00>
Total: 25474 W: 6357 L: 6325 D: 12792
Ptnml(0-2): 1, 2515, 7674, 2545, 2

And that without tuning! Now it is being optimized.
Vast
Posts: 2
Joined: Sun Oct 15, 2023 6:31 pm
Full name: Joseph Pasfield

Re: Stockfish 070324 a Disaster

Post by Vast »

The second line of that result states that the Log Likeliness Ratio of that SPRT (LLR) is -0.01. This is indicative that the test has barely progressed and is slightly negative. I don't think that means the 3072x net has beaten the 2560x one. It means there is much more testing to be done.
Eduard
Posts: 1439
Joined: Sat Oct 27, 2018 12:58 am
Location: Germany
Full name: N.N.

Re: Stockfish 070324 a Disaster

Post by Eduard »

It's on par with 2560 in the Bullet! In my tests, Dimension 3072 has been ahead of Dimension 2560 for a long time. I know that Dimension 3072 is better for analysis. I don't care if 2560 remains the master network. All the better for me! :-)
Eduard
Posts: 1439
Joined: Sat Oct 27, 2018 12:58 am
Location: Germany
Full name: N.N.

Re: Stockfish 070324 a Disaster

Post by Eduard »

Something else:

Sudden Death – Improve TM
Author: Lemmy
Date: Thu Mar 14 22:12:53 2024 +0100
Timestamp: 1710450773

Due to the estimated turn horizon of 50, a sudden death game occurred once
1 second would cause the time division for the optimal time to go negative
and SF would start instamoving.

To counteract this, the movement horizon is shifted as soon as limits.time is below 1 second
begins to decrease at a rate of 1 movement per 20 ms. That was exactly the right thing
seemed like a reasonable decay rate.
--

I've known that this is the case for a long time! Such games often took place on the online server and Stockfish lost the game. It took the developers a lot of time (despite billions of test games) to finally realize this themselves! :roll:
(Server LAG has even more impact on the rest of the time)