|
Post by matejst on Dec 26, 2022 16:57:13 GMT
I downloaded this last dev. version almost entirely based on Lc0 data. The problem with SF was Is it still? From what I read, it seems that most of the data is Leela's. It is here:
"Author: Linmiao Xu Date: Wed Dec 21 07:14:58 2022 +0100 Timestamp: 1671603298
Update default net to nn-335a9b2d8a80.nnue
Created by retraining the master net with a combination of:
the previous best dataset (Leela-dfrc_n5000.binpack), with about half the dataset filtered using depth6 multipv2 search to throw away positions where either of the 2 best moves are captures Leela T80 Oct and Nov training data rescored with best moves, adding ~9.5 billion positions
Trained effectively the same way as the previous master net: [...]"
|
|
|
Post by Ozymandias on Dec 26, 2022 17:46:05 GMT
But is it still "an optimisation of known ideas" or have things changed, now that it uses Lc0 data?
|
|
|
Post by matejst on Dec 26, 2022 19:53:18 GMT
But is it still "an optimisation of known ideas" or have things changed, now that it uses Lc0 data? I thought initially that you asked this question, but then, somehow changed my mind. Yes, more than ever. Basically, SF remains Rybka + Komodo + Giraffe + Shogi. Nihil novum sub sole. And they were late at everything -- I still have Texel GI where Giraffe's net was implemented and Peter Osterlund proved the concept was working, and the SF team missed the idea. They did not even try like the Leela authors to develop a net from zero, nor from TBs like the author of Seer, they did not manage to make multiple nets work like Jonathan Kreuzer (or a few else, last example is Caissa), nor did they try with radically different net structures like Chris or Michael Borgstadt, etc.
They are too afraid to lose Elo, so wait for others to prove and mature a concept before incorporating it in SF. SF is a compilation, an engineering feat, but it is a desert of new idea. Small tweaks. Nothing revolutionary.
|
|