Last green Net nn-62ef826d1a6d.nnue

[Paloma]

Why is the last green NNUE already 2 months old ?
After that, about 100 new networks have already appeared, but none is recommended (i.e. green).
Are they all worse than nn-62ef826d1a6d.nnue ?

[dkappe]

Why is the last green NNUE already 2 months old ?
After that, about 100 new networks have already appeared, but none is recommended (i.e. green).
Are they all worse than nn-62ef826d1a6d.nnue ?

The series of “green” nets were trained via RL (reinforcement learning) and possibly are very close to the maximum strength possible for the net architecture and size. What’s surprising isn’t that it topped out, but that the SF team got there so quickly, but given that Sergio Viera was a leela net training veteran, perhaps it wasn’t so surprising.

The SF team is working on new training software and architectures and Sergio has returned after some time off. Things are looking promising.

[pohl4711]

Why is the last green NNUE already 2 months old ?
After that, about 100 new networks have already appeared, but none is recommended (i.e. green).
Are they all worse than nn-62ef826d1a6d.nnue ?

The series of “green” nets were trained via RL (reinforcement learning) and possibly are very close to the maximum strength possible for the net architecture and size. What’s surprising isn’t that it topped out, but that the SF team got there so quickly, but given that Sergio Viera was a leela net training veteran, perhaps it wasn’t so surprising.

The SF team is working on new training software and architectures and Sergio has returned after some time off. Things are looking promising.

What about this net:
https://tests.stockfishchess.org/tests/ ... 97de3ef2d0

STC Framework test was green...But no LTC?!?

[Jouni]

Difficult to follow all test now. One of most extreme was:

Code: Select all

ELO: 271.32 +-2.6 (95%) LOS: 100.0%
Total: 60000 W: 41331 L: 2137 D: 16532
Ptnml(0-2): 40, 504, 3572, 11990, 13894 

This was 60000 @ 60+0.6 th 1 test .