Leela Zero as a computer opponent

#10 "Playing against LCZ doesn't make her stronger. The playable version is frozen. And even if it wasn't it would slow down learning as explained in the AGZ paper."

I agree, yet I think years from now people will want a strong AI which emulates human behavior. I'm not suggesting that Leela does that, because her definition of "best outcome" is a win. (But again, Leela is a costly opponent to run, and currently Lichess lacks a provision for sanctioned AI opponents.)

Training on human games is one interesting mode for a future fork of LZ/LCZ. One interesting point is that most Go players think that AGZ plays more humanlike then the previous Alphago, even though the old Alphago was mostly trained to imitate human games.

Other interesting forks would be one for chess960, if it learn even classical chess faster then the conventional trained one.
Or a version that values draws at a different point then half of a win.

FYI kontrachess.com/ just released 3 Leelas into the wild! Good luck, and I for one hail our robot overlords.

@Allonautilus is right. See github.com/ornicar/lila/issues/4228#issuecomment-379804993

Please fix my numbers and address my concerns.

@thibault How about to limit the number of people who can simultaneously play Leela? Like when many subscribers want to play a streamer, they throw challenges and then the streamer picks them one by one - the similar queue can be organized for Leela. As this probably can be abused by using multiaccounts and the queue will quickly become very large, this can be solved by turning the queue into a priority queue by some meaningful parameter. For example, holding Daily Leela Arenas with different controls during a week, and using the total sum of Leela tournament points for the last week as the priority of an account in the Leela queue.

Of course I could also propose that people could play Leela for a donate, but we all know that's not the Lichess way. Limiting of play by certain tournament points seems to be not any bit more unfair than limiting the access to Lichess Titled Arena.

@thibault

Leela is still half-decent even when limited to just 1 playout (which I'm guessing is under 100 ms of CPU). Well, she's a challenge for me as a 1300 player anyway, maybe not so much for stronger players. And that will surely change as Leela improves.

If i understand correctly, Leela would not benefit from being used on lichess. While lichess would eventually take a serious hit by offering such a cooperation at unduly times.
... makes no sense.
I suggest, reconsidering, if Leela ever gets somewhere meaningful (and there are reasons to doubt that. Too many misunderstandings at the basis of that project. Let them sort this out eventually)

Fun fact: Leela is weak tactically.

I tried using it against different openings. Here's my attempt at the poisoned pawn variation in the Sicilian. ( I used hard mode by the way.) I let Leela play on its own from move 15:



It's interesting that it thought it had over 50% chance after move 16, which is actually a forced mate in 10 or so.

In another game it hang mate in 1, albeit a losing position (about +6 by SF).
So all in all it makes an interesting player if you can't or don't want to play vs. humans.

Pardon my ignorance, but couldnt Lichess store all the games played against Leela and every certain time send them to http://lczero.org/ ?

@MateoReq Leela only trains by playing against itself, not humans. So there would be no use in sending the games to lczero.org.