( There are numerous examples of top pros losing games after the bot assessed their winrates over 90%... so even a 9.5% drop could be "noise" for humans... But, at least entertaining to know. )
Thanks!
Re: dutchie's journal
Posted: Sun Aug 26, 2018 10:03 am
by Tryss
At this point of the game, a -6% error is not noise, it's significant. Not a huge mistake, but still noticeable...
And yes, LZ want to hane at your (a) instead of the descent
( There are numerous examples of top pros losing games after the bot assessed their winrates over 90%... so even a 9.5% drop could be "noise" for humans... But, at least entertaining to know. )
I mean, if it was the true winrate, pro would lose 1/10th of the games that were assessed win over 90% in their favour. 1/10th is not a small number. Roll two dices, and they'll lose the game if you roll 11 or 12 points
Re: dutchie's journal
Posted: Sun Aug 26, 2018 11:29 am
by Bill Spight
Elf generates more extreme winrates than Leela Zero, but yes, an estimated 6% difference is hardly noise, even for Elf. For Leela Zero my guess is that it indicates a serious error.
Re: dutchie's journal
Posted: Sun Aug 26, 2018 12:14 pm
by Tryss
My current scale with LZ (#153) :
-0.5% to -2.5% : inaccuracy
-2.5% to -5% : error
-5% to -10% : significant error
over -10% : critical error (not necessarily game losing).
Of course, that depend on the situation. Toward the end of a very close game, yose inaccuracy can drop your winrate by 50%
But maybe we should move this conversation elsewhere
Posted: Sun Aug 26, 2018 2:31 pm
by EdLee
Hi Tryss, Bill
Hi Tryss, thanks. There's an existing thread or even threads about winrates. Anyway, my understanding ( or rather, my gut feeling ) about super-human bots' winrates is they are for bots. Human brains are different; that's why the bots are at super-human levels.
I mean, if it was the true winrate, pro would lose 1/10th of the games that were assessed win over 90% in their favour.
And who has done the research on this ? I wouldn't be surprised if this is true. To be clear: the 90% ( or say, over 75% ) winrate is for a particular moment in the game. Obviously the moment matters. If the 90% is assessed at move 349 out of a 350-move game, that's very different from at move 70 out of a 350-move game. In this example, it's .
Thanks, Tryss! I have to get one of them bots...
Live example: Iyama v. Cho, post 282: human winrates all over the place (80%+ and ups and downs...)
Re: dutchie's journal
Posted: Mon Feb 03, 2020 12:49 pm
by dutchie
I'm back! I've been playing off and on while I haven't been posting, and I have somehow managed to attain 7k on OGS. I think that's mostly because I'm higher ranked at correspondence than live, but I'll take it.
However, I just got destroyed by a 5k. A lot of my groups ended up dying, and there were definitely a few brain fades in there. My main question is how should I deal with W approaching my three stone wall at move 46? The pincer I played seemed obvious, but in the end my group died. I'm not sure if the misread net was the problem, or if I should just run (with P13 or something?).
Also, the OGS AI's win percentage dropped a lot after I played P6. I'm not a site supporter, but that seemed a very natural shape move. Should I be defending the cut at Q5 more directly there, and allowing W to hane?
SGF with general comments attached.
Re: dutchie's journal
Posted: Mon Feb 03, 2020 1:53 pm
by jlt
I explored a bit with Katago. Concerning your wall, your mistake was at move 24. A general principle is to play urgent moves before big moves. You played at K3, this is big but doesn't threaten anything so is gote. On the other hand, after White played the marked stone, the corner group becomes strong so your 3-stone wall becomes weak in comparison, and needs an extension at one of the points a-d.
Concerning move 46, Katago wants you to play elsewhere (at H4) but if you want to respond locally, here is a possible sequence with the same idea of pincering, but which secures your group:
About move 34: I think that the AI doesn't like your move because of the cut at R4 which forces Black to capture the cutting stone at some point. For instance, if White plays at Q2 and Black ignores, this can be very annoying:
$$Wc
$$ . . . . . . . . . . . . |
$$ . . . . . . . X . X . . |
$$ . . . . . . O X 2 . . . |
$$ . . , . . . . O X 1 3 . |
$$ . . X . . . . O O X . . |
$$ . . . . . . . . W X . . |
$$ . . . . . . . . . . . . |
$$ -----------------------
[go]$$Wc
$$ . . . . . . . . . . . . |
$$ . . . . . . . X . X . . |
$$ . . . . . . O X 2 . . . |
$$ . . , . . . . O X 1 3 . |
$$ . . X . . . . O O X . . |
$$ . . . . . . . . W X . . |
$$ . . . . . . . . . . . . |
$$ -----------------------[/go]
Re: dutchie's journal
Posted: Mon Feb 03, 2020 6:17 pm
by dutchie
jlt wrote:I explored a bit with Katago. Concerning you wall, your mistake was at move 24. A general principle is to play urgent moves before big moves. You played at K3, this a big but doesn't threaten anything so is gote. On the other hand, after White played the marked stone, the corner group becomes strong so your 3-stone wall becomes weak in comparison, and needs an extension at one of the points a-d.
Ah, I see. So if White approaches at a-d without playing the marked stone, then I should attack the corner somehow? I'm not used to the post-AI 3-3 invasion joseki yet.
Re: dutchie's journal
Posted: Tue Feb 04, 2020 1:05 am
by jlt
Yes, if White attacks the wall there are several continuations to the corner joseki, you can check josekipedia.com for suggestions. According to Katago, here is a possible sequence:
So the three Q17 stones are almost dead, but they still have aji, Black gets sente, has taken part of the corner territory and gets potential on the right side.
If you don't like the idea of losing your three stones, you also have some forcing moves to strengthen your wall:
(a) hane better ?
.
at K3, this is big but doesn't threaten anything so is gote. On the other hand, after White played the marked stone, the corner group becomes strong so your 3-stone wall becomes weak in comparison, and needs an extension at one of the points a-d.