SPCC: Testrun of Lc0 610594 finished

[pohl4711]

NN-testrun of Lc0 0.28.0 610594 finished.

https://www.sp-cc.de

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

[mehmet123]

Lc0 0.28.0 610594 (30x384) is behind of Lc0 0.27.0 68695 (30x384) at the rating list.
After 1899 net and ~60 millions self play game there is no regression. Unless there is a radical structural change it is not right to expect any progress from Lc0 anymore.

[Ozymandias]

No regression, no progress, I guess the question is whether the people donating their HW and money, are also in this for the theoretical research into NN, or they were actually expecting some ELO gain.

Has the number of active users and games been declining?

[mehmet123]

After 1899 net and ~60 millions self play game there is no regression.

AlphaZero has trained by 44 millions self play games. It's not good if 60 millions self play games aren't improving Lc0. We need a new NN engine with new technical codes.

[pohl4711]

After 1899 net and ~60 millions self play game there is no regression.

AlphaZero has trained by 44 millions self play games. It's not good if 60 millions self play games aren't improving Lc0. We need a new NN engine with new technical codes.

Perhaps the problem is the tactics... Lc0 is running around 800x slower (in nps), than the nnue-engines. Because of this, it is clear for me, that Lc0 can not (and will never be) tactically as strong as its opponents. So, the question for me is, if Lc0 reached the Elo-ceiling, because the weak tactics lead to so much lost games, that even the highest progress of positional understanding (=net-learning and improvement) can not compensate this.
Lc0 is a superhuman, with the same strengths (positional understanding) and weaknesses (tactics) as humans. And like the humans, it can not beat the best engines anymore, because Lc0 and humans are "outcalculated" by the nodecrunchers. Lc0 could have an unsolveable problem here, because it runs so slow.

Just my 2cents