1. his quality varied massivly during the last 2 weeks
(the creator changed it at one point to elo 1200 and then back to 2000)
2. it is not self learning!!!
I would expect a self learning bot (like the term/name Leela suggests) not to blunder 2 times in the same way.
I did find out that its quite easy to beat him continously. Just beat him once like I did a few minutes ago (or look at his last games where he was beaten with black or white pieces), memorize your moves and his moves. and just play the same moves. Quite boring, I know...
I think that if it was self-learning, it would get stronger and no longer be "tame". It needs to be "frozen" at a certain strength to manage this.
nevertheless it should have some variation in his playing style.
stockfish level 4 or 5 always play different moves.
As u can see in my first post, leela nice just blundered the same way.
I wasnt even playing very well.
I see, but I can't imagine if it is possible for machine learning to "learn how to blunder differently" or whatever, it's the opposite of the current "self-learning to improve" model. Maybe it is possible, but I don't see it being a focus of LCZ development.
First, Leela only learns from self-play games generated by the client executable you can get from the site. She doesn't doesn't do "live" online learning while playing other opponents.
On the variety thing, yeah, it's tricky because of the way LeelaNice's strength is limited.
The way to get variety with Leela is to use "temperature", which basically tells it to pick moves in proportion to how much they're searched (the way its search works, it ends up searching moves it thinks are better substantially more often than moves it thinks are worse, so it still mostly plays the moves it thinks are best, but occasionally plays suboptimal moves).
The problem is that LeelaNice (unless this has changed recently), importantly relies on doing a single node "search" for limiting strength.
The problem is that temperature doesn't do anything then; since you've only searched a single move, you're going to play that no matter what.
On the other hand, if you increase the number of nodes Leela visits, you can get some variety, but at the "cost" of her becoming much stronger.
There probably is some middle ground that would work decently, like tweaking FPU reduction and pUCT (settings that basically control how much Leela focuses on the moves it thinks are best compared to how much it searches moves it thinks are inferior) and turning on temperature, but that would require a decent bit of tweaking to get right.
Probably worthwhile to investigate, though. I might mention it in the Discord if I remember :)
And what's the point of this rant?
According to the description you are playing a crippled engine
"Like
@LeelaChess, this bot run lczero. But LeelaNice has been tweaked to play more like a 1800 rated lichess player. It accepts all.."
If you don't enjoy playing this engine, what's forcing you?
thx
@a_pleasant_illusion it sounds interesting, I read a little bit about the leela project, and also played a few games on the lczero site. But I just observe the progress right now.
@BlackSalt I am sorry if my choice of words did provoque any irritation. I just think that it is not fair to call a bot leela if there are some grave limitations and the bot doesnt resemble the real leela enough. E.g. u wouldn't and shouldn't call your bot Google, Alphabet or DeepMind if its not a google engine/project.
Just my 2 pence.