Hello,
It would be incredibly cool if, in analysis mode, Leela suggestions were available alongside SF9 ones. Having such a set up would be incredibly useful when looking for new perspectives in late-opening and middlegame positions ...
Are they any plans to do something like this?
Thanks!
I'm not sure, but as far as I understand it is barely possible. Leela and Alpha Zero work completely different than our standard engines. Believe me, all of us (including Top GMs) would love to ask these engines to play/evaluate a certain position.
But, while SF/Komodo... works with, this is the position, just calculate accoriding to the algorithms, stop after a time, evaluate the position,... They absolutely don't care about time, input, position, etc. They jsut calculate and evaluate the position by their standards (e.g. pawn is +1, bad king might be -0.5; in positions where they top calculating further).
Well, Leela (isn't it Leelo?) does not work like that. She does not have any human evaluations, so she cannot just tell you that 34.Ng2 is +0.5 while 34.Kg2 is -0.2. She does not know any of this. All she does know is her own experience. Like she would tell you that she would play (or did already played agaisnt herself) Ng2 with better success. Which comes to the next thing- she can play against you. She even might tell that she considers e.g. French to be a bad opening and tell you why, by destroying you. But she won't tell you how to play better, she just simply says 'French is bad, everything is losing'. She can prove it by playing, but she won't help you play french better. She just simply proves her point. She is about beeing (aka) playing right. The whole engine hates 'wrongness'.
So if you put your position, which was played with a lot of mistakes from her perspective, she cannot help you. She will says- that is hopeless won/lost in the best case. Other than this she will say she never had this (bad of a) starting position. She will need experience to play it against herself to get to the truth. he will need to calculate everything from the scratch. And if you force her immediatly to move her first suggestions might be the worst you ever face.
I love the engine, and her playstyle, but her playstyle i her own and based only on her own games against herself. Her experience lies in the critical lines, the ones where it is hard to get an advantage for white (or hard to keep it as black). These are the positions she is familiar with, she spends time in. If she could speak she would say something like
Leela: 'your move 15.Ne5 is a mistake, Nd2 is better'
you: 'Yes, but I played 15.Ne5 and what about 20. xy?'
Leela: '15.Ne5? Rxe5!! and black would have won on the spot.'
you: 'But he played differently and I asked you about position at move 20'
Leela: 'The position at move 20 i invalid for my algorithm, if you would play 15.Nd2 then I can help. Everything else does not bother me'
This is as far as I understand the engine, as a mathematician working on self learning algorithms. So sadly it will be beneficial only for the opening theory, since after the first mistake Leela cannot help you any longer
Thanks ToLazy ... points taken but I'm not 100% that's how Leela works ... Jon Hammer did a vid a while ago
www.youtube.com/watch?v=TxiNUPKv-Es (with a very young version of leela) here, where she was offering evaluations (presumably based on her % chance of winning estimations). Also, she was forced to play a lot of bad lines in the CCCC and fared pretty well – surely that wasn't purely based on her exhaustive explorations of those positions in her millions of self play games?
Anyway, what do I know. I just thought it would be really super cool if Leela was on hand to offer her positional insights on things (even if she misses the odd basic tactic ;) )
