TalkChess.com

https://lczero.org/

I imagine someone will say it does not really mean that.
But that sure is what the graph looks like.

I wonder if someone made a naughty patch.

If you test the individual nets against a non-LCZ engine you'll probably notice no or little elo gain the past few days despite what the self elo says. Unless it starts gaining some real elo soon I think they might have to do the second LR drop. I'm no expert in NN nets but I imagine when you see a sharp increase in self-play elo all it probably means is that LCZ has found a winning opening against the older nets, and then when the self-play elo sharply decreases it probably means that a weakness has been found in the winning opening. I might be wrong though...

That is an interesting theory. So they don't use opening books at all in testing?
Seems to me like it might be a good idea to force lots of oddball stuff like long and short castled sides, and things that might be puzzling like fianchetto bishops found in the Indian openings.

Temperature decay is the key. Temperature (aka "randomness") starts very high in self-play games and is decreased as the game evolves. That makes an opening book not necessary, as you will see a lot of variation in the beginning of the game.

Dann Corbit wrote: ↑Wed Nov 14, 2018 11:15 pm https://lczero.org/

I imagine someone will say it does not really mean that.
But that sure is what the graph looks like.

I wonder if someone made a naughty patch.

Well, your graph just got updated !

I wonder if anybody can say something sensible about this data...

Looks more and more Lc0 is a hoax... that works !!!

Well, that hoax played pretty well in TCEC.

LCO needs a little help. Maybe a pill.

He could never lift that pill out of there. He always had to get Polly the poodle to do it for him.
He should have just given her the ring.

Dann Corbit wrote: ↑Fri Nov 16, 2018 3:25 am He could never lift that pill out of there. He always had to get Polly the poodle to do it for him.
He should have just given her the ring.

lol

Dann Corbit wrote: ↑Fri Nov 16, 2018 2:03 am Well, that hoax played pretty well in TCEC.

That's what I said... it works !

Not sure anybody has a deep understanding about how it progresses...