SPCC: Testrun of Lc0 T75_2400k finished

[pohl4711]

NN-testrun of Lc0 0.28.2 T75_2400k finished.


https://www.sp-cc.de

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

[brianr]

Thank you for testing this net.
By request, I'm currently running some self-play RL "finishing-off" training that might add another 10ish Elo.
It will take at least another week.

This net and the prior 15b and 10b nets were not intended to be my strongest nets.
They are to compare net sizes with the exact same training as much as possible.
The 15b strength improvement over the 10b was a much larger improvement than the 20b v 15b.

Code: Select all

# PLAYER             :  RATING  ERROR  POINTS  PLAYED   (%)    W     D    L  D(%)  CFS(%)
1 256x20-T75-32MGames:      47      9   457.5     880    52   57   801   22    91      99
2 192x15-T75-32MGames:      39      7   921.0    1750    53  149  1544   57    88     100
3 128x10-T75-32MGames:       0   ----   501.5    1130    44   21   961  148    85     ---

I was hoping that 32M games and 2.4M steps would be enough for the larger nets to pull further ahead of the smaller ones.
There are significantly diminishing returns with the 20b, which took 37 days to train and another week to run the match.
I do not have the patience with my hardware (single 2080ti) to try a 25b, nevermind a 30b.

My best 10b net is within about 30 Elo of the Lc0 project full RL training runs.
https://drive.google.com/file/d/1CNTaih ... sp=sharing
That last Elo gap is proving to be exceedingly difficult to close with an SL/RL combination.

There is quite a lot of promising work currently being experimented with in the main project by trying significantly different net architectures and sizes (more filters/fewer blocks, attention nets, policy changes, and even reducing SE [aka squeeze excitation], and hopefully some search changes too.

Thanks again,
Brian