SPCC: Testrun of Lc0 610889 finished

[pohl4711]

NN-testrun of Lc0 0.28.2 610889 finished.


https://www.sp-cc.de

(Perhaps you have to clear your browsercache or reload the website)

[dkappe]

So 3660, a little ways back from 3667 of 610062. And all with error margins of 17 or 18.

[mehmet123]

Lc0 0.28.2 610889 is -1 elo behind of Lc0 0.27.0 68695.
Still no progress after 2194 net and more than 70 millions self play games.

[pohl4711]

Lc0 0.28.2 610889 is -1 elo behind of Lc0 0.27.0 68695.
Still no progress after 2194 net and more than 70 millions self play games.

Yes, I doubt, that T60 training-run will make progress in the future. Generally, IMO Lc0-project is not very promising anymore, since Stockfish, Komodo etc. are using nnue neural nets, which improve their positional understanding a lot. Of course, Lc0 has still a better positional play, but not so much better, that it compensates the much higher nps and brutal optimized search of Stockfish and Komodo.
Nevertheless, the Alphazero/Lc0 project was groundbreaking and extremly important for the computerchess in general. But now, it seems, it reached its maximum and the AB-engines with nnue neural nets have taken over.
This is how evolution works...So, no need to be disappointed of Lc0.
But I am thinking about abandoning my NN-testruns in the near future. At the moment they are a waste of energy (which is extremly expensive here in Germany!) and resources...

[Ozymandias]

Not just in Germany, but all over the EU. I guess it only makes sense to test really new nets, like 40b, once they're above 30b. Or if they're able to improve the executable, something Ceres has been a promise of for quite some time now. If they decide to scrap 15b in favor of 20b (which hasn't been developed for over a year) that could also be interesting, but in general it's a waste of time and money.