NM youtubing under Chessnetwork is on a bit of a Leela kick of his own. He’s got two interesting
wins by two different ID’s, the first a very strategic Indian that Houdini 1.5a misunderstands. The second
against Stockfish 5 is a gambit in the Scotch.
I would like to see a few more wins on the black side, to be honest, but we’ll take what we can get.
We have started training a 256x20 net in the test pipeline. Even with the speed of lc0, the generation of self play games is much slower, which points out how critical lc0 will be to the next phase of training. There are some elo fluctuations we’re trying to iron out. Stay tuned.
The 192x15 main server net has crashed through the 6000 self play elo barrier!
Resign is working. Fingers crossed.
Even compressed, the 256x20 weights file is 130MB. That’s a burden on clients, servers and chews up a lot of time. The last thing before rollout of lc0 to the main pipeline, is shrinking the weights file, either by reducing precision to fp16 (there’s also a 8 bit experimental effort), or by moving to a specialized compression library. This is the major item to resolve before launch.
The blas backend computed the NN incorrectly, and this was fixed today (7/2/2018)
lc0 now supports fp16 computation! WARNING: this is for high end cards like the Titan V; GTX cards not supported.
New pipeline won’t be next week, and probably not week after next, so more cooking on the existing main net. On the other hand, it could happen next week. ;-)
[Editor’s note: Leela certainly wouldn’t play the Dunst opening on her own. But even in openings far from her preferred lines, she can find some surprising plans, like the theoretical novelty (TN) in this win vs Gull 3.
Tuning of hyperparameters in the test pipeline continues, along with parameters to lc0. An update on what has been learned in a future post.
Tweaking resignation continues. The idea is that once situation is hopeless for long enough in self play, the losing side resigns. Keeping the rate of resignations down to a low percentage and the rate of “false positives” — where a resignation happens in a drawn or winning position — to an even lower percentage.
There’s new time management in the works. The initial results in terms of elo gain are encouraging.
The rollout of lc0 and the new test pipeline will not happen this week or next.