Life In 19x19
http://lifein19x19.com/

dutchie's journal
http://lifein19x19.com/viewtopic.php?f=48&t=13928
Page 4 of 5

Author:  Tryss [ Wed Jun 13, 2018 1:09 pm ]
Post subject:  Re: dutchie's journal

LZ with over 50k playout consider localy :

E4 : 21.3% with 40k playouts :

Attachment:
File comment: E4 continuation by LZ
E4-LZContinuation.jpg
E4-LZContinuation.jpg [ 365.76 KiB | Viewed 12059 times ]


F3 : 21.6% with 16k playouts :

Attachment:
File comment: F3 continuation by LZ
F3-LZContinuation.jpg
F3-LZContinuation.jpg [ 346.45 KiB | Viewed 12059 times ]


B3 : 20.7% with 4.7k playouts :

Attachment:
File comment: B3 continuation by LZ
B3-LZContinuation.jpg
B3-LZContinuation.jpg [ 349.91 KiB | Viewed 12059 times ]


G3 and E3 get less than 20 playouts (so, not really seriously considered), with sub 20% winrate each

Other options are to tenuki, but with kinda low playouts (lower than B3).

Edit : added the "best" continuation that LZ plan for each of these moves

Author:  Uberdude [ Wed Jun 13, 2018 1:21 pm ]
Post subject:  Re: dutchie's journal

Nice to see LZ approves of my f3 in response to white b3 :)

Author:  Tryss [ Wed Jun 13, 2018 1:31 pm ]
Post subject:  Re: dutchie's journal

It's even the only move it consider (over 65k playout for it, while the others are all under 500)

Author:  EdLee [ Wed Jun 13, 2018 2:37 pm ]
Post subject: 

Hi Tryss,

Thank you.

Noob question: is there a significance between the 40k (E4) v. 16k (F3) playouts ? Their winrates are very close ( 21.3% & 21.6% ),
but the playouts are quite different ?

Is it Lizzie ? Very pretty interface. :tmbup:

Just curious: what's LZ's winrate for :w1: , given 2-stones ?
Does :white:'s winrate start from ~0% and climb up to ~20% by :b18: ?
Or is it the opposite ? :white:'s winrate starts out higher than 20% at :w1: ( but under 50%, obviously ), and drops to ~20% by :b18: ?

Author:  Tryss [ Wed Jun 13, 2018 3:15 pm ]
Post subject:  Re: dutchie's journal

Before :w1:, LZ gives white around 21% winrate, so there is no real winrate change by :b18:. Note that this evaluation is with 7.5 komi for white .


For the playouts, I'm pretty sure the difference is noise (due to the randomness of the tree exploration). I just did another run, and the results were :

F3 : 21.4% winrate with 42k playouts

E4 : 21.0% winrate with 33k playouts

B3 : 20.5% winrate with 8k playouts

So while last time LZ would have played E4, this time it would have played F3.

(I spend a little more playouts here, but the result would have been the same if I stopped a little earlier : F3 took the lead early this time)


And yes, this is lizzie

Author:  EdLee [ Wed Jun 13, 2018 4:29 pm ]
Post subject: 

Tryss, Thank you.

Author:  dutchie [ Thu Jun 14, 2018 5:45 am ]
Post subject:  Re: dutchie's journal

Huh, interesting analysis. I'm surprised W is down to only ~20% winrate after less that 20 moves!

Author:  Bill Spight [ Thu Jun 14, 2018 5:58 am ]
Post subject:  Re: dutchie's journal

dutchie wrote:
Huh, interesting analysis. I'm surprised W is down to only ~20% winrate after less that 20 moves!

EdLee wrote:
Just curious: what's LZ's winrate for , given 2-stones ?

Tryss wrote:
Before , LZ gives white around 21% winrate, so there is no real winrate change by . Note that this evaluation is with 7.5 komi for white .

Author:  dutchie [ Wed Aug 22, 2018 10:39 am ]
Post subject:  Re: dutchie's journal

Again it's been too long. I've played (and lost) a couple tonight. One was a fun close fight that I ended up losing by 22.5. I'll attach it in case anyone is interested and wants to open up the SGF themselves for any comments, but I reviewed it with my opponent and I'm much more curious about the second game, where I ended up with a wall across the centre and then messed up attacks (kills?) on two invading groups. I went through and added some thoughts of my own, but the opening was a little weird as well as messing up the fights in the midgame.



Attachments:
20180822-Neosilver.sgf [11.22 KiB]
Downloaded 496 times
20180822-Aralsa.sgf [5.81 KiB]
Downloaded 606 times

Author:  EdLee [ Thu Aug 23, 2018 3:16 am ]
Post subject: 

Hi dutchie,

:w8: human.

:w10: slow.

:w14: human:

:w16: slow, small.
( The game was B to lose after here. )

:b19:, :b21: how about kick-pincer combo ?

:b21: Does LZ like o17 area ?

:b23: :scratch: What does LZ think of this ?

:b27: What does LZ think about :b23: @ F9 directly ?
( foundation shapes study... :study: )

( :b21:+ :b23: ), ( :b27:+ :b31: ) Foundation shapes study :study: .
Instead of these 4 moves, imagine just 1 move around o17 area.
( Variation in SGF here still not sensing global perspective. )

Recent thread about direction.

:b37: note in SGF, "...make shape in the middle" --
Still missing global perspective, what mistakes led B to this position,
how to handle these stones.

( :b61:+ :b63: ) Hmm. Note in SGF, "No mindless atari" is good;
but insufficient here. ( emphasis on mindless, not atari. )
Problem here is knowledge gaps in foundation shapes.

( :b21:+ :b23: ), ( :b61:+ :b63: ) See Toothpaste and fractured shapes.

:b91: P16. Foundation shapes study :study:.

Author:  Tryss [ Thu Aug 23, 2018 4:38 am ]
Post subject:  Re:

EdLee wrote:
:w16: slow, small.
( The game was B to lose after here. )
After :w16:, LZ gives white around 30% winrate

Quote:
:b19:, :b21: how about kick-pincer combo ?
LZ choice. Interesting note : her favorite pincer after the kick is D11. This move is never on my radar, so I learned something new here. (but other pincers seems playable)


Quote:
:b21: Does LZ like o17 area ?
LZ choice too

Quote:
:b23: :scratch: What does LZ think of this ?
-11% winrate, black advantage is mostly gone (black has now 53% winrate)

Quote:
:b27: What does LZ think about :b23: @ L9 directly ?
( foundation shapes study... :study: )
You mean D9? (L9 is in the middle of the board). If so, less bad, but LZ favorite move is the D8 kosumi

The elephant eye at E11 feels like a serious mistake. I mean, LZ can manage to get an ok result from it, but the knight move at E10 or the 1 space jump at E9 feels much less error prone.


After 37, black winrate is down to 22%, a loss of over 40% winrate compared to the situation after :w20:

Author:  EdLee [ Thu Aug 23, 2018 4:55 am ]
Post subject: 

Hi Tryss,

Thanks.
Quote:
favorite pincer after the kick is D11. This move is never on my radar
Yea, D11 quite natural. :) Human reasoning: D12 too close to
W's iron pillar, so pull back to D11 ( but I don't know bot reasoning :) ).
Quote:
You mean D9? but LZ favorite move is the D8 kosumi
Thanks; typo: I meant F9 knight's. D8 kosumi, thanks.
Quote:
:b21: [cap] at E11 feels like a serious mistake.
@dutchie: LZ's 'hint' to you: what was your plan for :b21: at E11 ? You're not building a huge moyo. Just take a big point ( O17 area ) and keep the lead.
Quote:
After :b37: , black winrate is down to 22%, a loss of over 40% winrate compared to the situation after :w20:
@dutchie, LZ's lesson for you (thanks to Tryss): how to lose the game in 8 moves...

Author:  dutchie [ Thu Aug 23, 2018 6:29 am ]
Post subject:  Re: dutchie's journal

Thank you folks, a lot to think about. I honestly completely forgot that kick/pincer moves even existed, Direction of play and whole board thinking are so important!

Author:  EdLee [ Sat Aug 25, 2018 10:49 pm ]
Post subject: 

Hi Tryss,
Speaking of pincer...
No connection to dutchie's game.
But in case you're still reading this thread, a curious question:
Click Here To Show Diagram Code
[go]$$B
$$ -----------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . 7 8 . 5 . 6 . . . . . . . . . . . |
$$ | . . 9 4 . . . . . , . . . . . 1 . . . |
$$ | . . 0 . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . 2 . . . . . , . . . . . 3 . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ----------------------------------------[/go]
Does LZ have a significant change in winrate for W on :w10: ?
Thanks. :)

Author:  Tryss [ Sun Aug 26, 2018 2:43 am ]
Post subject:  Re:

EdLee wrote:
Hi Tryss,
Speaking of pincer...
No connection to dutchie's game.
But in case you're still reading this thread, a curious question:
Click Here To Show Diagram Code
[go]$$B
$$ -----------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . 7 8 . 5 . 6 . . . . . . . . . . . |
$$ | . . 9 4 . . . . . , . . . . . 1 . . . |
$$ | . . 0 . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . 2 . . . . . , . . . . . 3 . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ----------------------------------------[/go]
Does LZ have a significant change in winrate for W on :w10: ?
Thanks. :)


Yes, LZ (#153) consider this a -6% error. What's surprising is that the "punishment" at C18 is considered a -9.5% mistake !

Click Here To Show Diagram Code
[go]$$Bc
$$ -----------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . 1 3 . . . . . . . . . . . . . . . |
$$ | . . X O . X . O . . . . . . . . . . . |
$$ | . . X O . 4 . . . , . . . . . X . . . |
$$ | . . O . . . . . . . . . . . . . . . . |
$$ | . . . 2 . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . . . . . , . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ----------------------------------------[/go]


I'll explore this with stronger networks to see if they agree

Author:  EdLee [ Sun Aug 26, 2018 8:46 am ]
Post subject: 

Hi Tryss,
Quote:
Yes, LZ (#153) consider this a -6% error.
Thanks! So almost "noise"... :)
Quote:
...C18 is considered a -9.5% mistake !
Click Here To Show Diagram Code
[go]$$Bc
$$ -----------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . a . . . . . . . . . . . . . . . |
$$ | . . X O . X . O . . . . . . . . . . . |
$$ | . . X O . . . . . , . . . . . X . . . |
$$ | . . O . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |[/go]
Does LZ like :black: (a) hane better ?

( 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!

Author:  Tryss [ Sun Aug 26, 2018 10:03 am ]
Post subject:  Re: dutchie's journal

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

Quote:
( 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 :mrgreen:

Author:  Bill Spight [ Sun Aug 26, 2018 11:29 am ]
Post subject:  Re: dutchie's journal

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.

Author:  Tryss [ Sun Aug 26, 2018 12:14 pm ]
Post subject:  Re: dutchie's journal

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% :lol:


But maybe we should move this conversation elsewhere :oops:

Author:  EdLee [ Sun Aug 26, 2018 2:31 pm ]
Post subject: 

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.
Quote:
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 :w10: .

Thanks, Tryss! I have to get one of them bots... :blackeye:

Live example: Iyama v. Cho, post 282: human winrates all over the place (80%+ and ups and downs...)

Page 4 of 5 All times are UTC - 8 hours [ DST ]
Powered by phpBB © 2000, 2002, 2005, 2007 phpBB Group
http://www.phpbb.com/