Yeah the 250 strength reviews are not as reliable unfortunately, I have a few ideas on how to make that better but for now I took the liberty of rerunning the analysis at full strength and it had much better results I think.
I noticed the big blunders were actually a couple moves back - that seems to be a common pattern when there’s not enough playouts, it’s a couple moves lagged before the ai “sees” it.
thx for that! i really hope this endeavor turns out the way everyone wants it to!
i’ve been away from the forums for a bit and wasnt aware the analysis ran on different strengths… ill read up on it .
EDIT: I read up on the review-strength-system on the supporter page and found an unrelated minor error. while the english page correctly states that i am supporting ogs with a donation of 25€ a year, the german version says i donate 25€ a month, which, as much as i like it here, is not the case (i checked xD) .
Is this the right place to mention that if you do a full review and then go back to the top three moves, if you tap the full ai button to send to do the full review again from scratch. I was expecting to just flip to the already completed AI screen. Seems unnecessary to be able to trouble the server with another review when one already exists.
This is intentional so the reviews can be upgraded. There are different levels of reviews and it would be unfortunate if a low level review locked out a higher one. But there could probably be some kind of check in place to see if the requested review is better than existing ones.
Actually I was looking back at the game I was linking before to do with the leela variations being displayed in a confusing manner. I can’t remember if it was always like this but at the end of the game it thinks white has an 80% chance of winning even though none of the suggested moves change the score. I’m imagining leela’s neural network isn’t taking into account that it’s a handicap game, and so it could think white gets 6.5 or 7.5 komi, whatever the usual is supposed to be for it, and since black is winning by 5.5 it thinks white is actually winning.
Not only plausible, but known. Currently the bot assumes 19x19 and komi (i forget which but probably 7.5) so for handicap games the result is either going to be inaccurate or plain wrong.
The developers aim to one day be able to support different board sizes and komi values.
Before then, I’m not sure what the optimal way to address this issue is. Close games will be buggy like you observed, but there could still be some worthy benefit in more conclusive victories regarding where losses were made (specifically when black loses).
Oh yeah I don’t doubt the value of having the variations (except in some cases where leela knows it’s behind and looks at terrible variations), just I was thinking regarding another topic:
It would seem at the moment there’s an issue to use the ai to decide correspondence handicap games that timeout if they’re estimated to be closer in score than the komi at least. If the win rate is based on non-existent komi though it might even be hard to use the ai at all to decide correspondence games in the case of a serial timeout. (I know this is probably minor given the lack of handicap games anyway.)
I might link this to that thread if it’s worth having there.
Is it because the initial blunder count based off the 1 playout analysis - Leela’s ‘intuition’ - where there is one move with more than 10% swing?
The full analysis which is run after (should be more accurate) and shows no moves above the blunder threshold
Yeah - you would think that after full analysis has been run, it would no longer be telling us about the poor analysis blunder analysis. In fact, it appears there is no way to get the blunder analysis of the full run?
This review seems to have been stuck on procession for a few days. https://online-go.com/game/18564761
I think I actually got a message saying it had finished, but when I checked it was still in queue.
Wondering what I can do about this.