It is currently Tue Mar 19, 2024 2:13 am

All times are UTC - 8 hours [ DST ]




Post new topic Reply to topic  [ 1056 posts ]  Go to page Previous  1 ... 31, 32, 33, 34, 35, 36, 37 ... 53  Next
Author Message
Offline
 Post subject: Re: This 'n' that
Post #661 Posted: Fri Oct 11, 2019 2:31 pm 
Gosei
User avatar

Posts: 1747
Liked others: 176
Was liked: 488
:b5: is not on AlphaGo Teach either.

AlphaGo thinks that "a" has 46.8% winrate.


This post by jlt was liked by: Bill Spight
Top
 Profile  
 
Offline
 Post subject: Re: This 'n' that
Post #662 Posted: Sat Oct 12, 2019 12:54 am 
Lives in gote

Posts: 502
Liked others: 1
Was liked: 153
Rank: KGS 2k
GD Posts: 100
KGS: Tryss
Elf is bad for reviewing games.

With 30k+ playouts (but the result is the same with much less) Katago think :b5: is a -0.4% winrate drop compared to a, (or a loss of 0.2pts), meaning it's purely stylistic

Top
 Profile  
 
Offline
 Post subject: Re: This 'n' that
Post #663 Posted: Sat Oct 12, 2019 2:34 am 
Honinbo

Posts: 10905
Liked others: 3651
Was liked: 3374
Tryss wrote:
Elf is bad for reviewing games.

With 30k+ playouts (but the result is the same with much less) Katago think :b5: is a -0.4% winrate drop compared to a, (or a loss of 0.2pts), meaning it's purely stylistic


Maybe so. But the problems with Elf are the same as with other bots, if perhaps more extreme. Elf has strong preferences and we are more certain that its preference reflects its style that that it indicates an error. But that is so for every top bot I know of. It's just that their preferences may not be so strong.

As jlt has noted, AlphaGoTeach also prefers a as its top choice, and does not report :b5: as an alternative. Let's take a look.

Click Here To Show Diagram Code
[go]$$Bc AlphaGoTeach's preferences
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . 2 . . . . . . . . . . d . . . . |
$$ | . . . , . . . . . , . . . . c , 1 . . |
$$ | . . . b . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . a . . |
$$ | . . 3 , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . 4 . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]


It shows four possibilities, none including :b5:, in order of preference a - d. It winrate estimates are a - 46.8%, b - 45.5%, c - 45.3%, d - 45.0%. If we believe that these differences are a matter of style, then we want a winrate difference (near 50%) of more than 1.8% to think that it indicates an error.

But is it just a matter of style? Each top bot has its own developmental path dependency. How many of them pick a as its top choice? Does KataGo have a different top choice? If not, then that's three so far. :)

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

Visualize whirled peas.

Everything with love. Stay safe.

Top
 Profile  
 
Offline
 Post subject: Re: This 'n' that
Post #664 Posted: Sat Oct 12, 2019 3:36 am 
Gosei

Posts: 1733
Location: Earth
Liked others: 621
Was liked: 310
KataGo:
Attachment:
Kata1.JPG
Kata1.JPG [ 157.12 KiB | Viewed 9416 times ]


Last edited by Gomoto on Sat Oct 12, 2019 7:24 am, edited 1 time in total.
Top
 Profile  
 
Offline
 Post subject: Re: This 'n' that
Post #665 Posted: Sat Oct 12, 2019 4:29 am 
Gosei

Posts: 1733
Location: Earth
Liked others: 621
Was liked: 310
I forgot to close Lizzie in the background :-)

Attachment:
kata2.JPG
kata2.JPG [ 163.03 KiB | Viewed 9406 times ]


This post by Gomoto was liked by: Bill Spight
Top
 Profile  
 
Offline
 Post subject: Re: This 'n' that
Post #666 Posted: Sat Oct 12, 2019 1:30 pm 
Lives with ko

Posts: 218
Liked others: 23
Was liked: 14
Rank: IGS 3k
KGS: Bki
IGS: mlbki
What's interesting here is that all the bots seems to favor some kind of approach moves. If the difference between the evaluation was due to noise, you would expect that one bot might rank the enclosures (any of them) higher. But it doesn't, which suggest that while it can't be called a mistake, enclosing a corner might indeed be inferior in this position.

Does elf refusal to even consider the large knight move enclosure a matter of "style"? Or is this because in such a position this move is dominated by another move (i.e. at best equal) and so is never worth exploring?


This post by Bki was liked by: Bill Spight
Top
 Profile  
 
Offline
 Post subject: Re: This 'n' that
Post #667 Posted: Sat Oct 12, 2019 2:31 pm 
Gosei

Posts: 1733
Location: Earth
Liked others: 621
Was liked: 310
Oh, KataGo for example is considering lots of moves. All the red spots are considered by the bot. (I currently have not set up an ELF network.)

Top
 Profile  
 
Offline
 Post subject: Re: This 'n' that
Post #668 Posted: Thu Oct 17, 2019 4:31 pm 
Honinbo

Posts: 10905
Liked others: 3651
Was liked: 3374
My meeting with Takemiya

In the 1970s I spent one summer in Japan and made the cusp of 3 dan. For a few weeks I won every game as 2 dan but lost every game as 3 dan. ;) The last weekend before coming back to the US I went to a go weekend at a resort. The highlight on the last day was a game between Takagawa and Takemiya. Ishida Yoshio gave the running commentary. There was time to kill while the players got several moves ahead. Ishida warmed up the room. He was very good at it. (If you don't know, he is also a professional jazz guitarist. He used to perform solo at jazz coffee shops and kissaten in Tokyo.) At one point he asked who had come the farthest to be there. I raised my hand, which got a big laugh from the audience. "So it seems," said Ishida, and then said, "No, really. Who came the farthest to be here?" I was not amused.

Afterwards a doctor who was a 5 dan invited a couple of us to come to his rooms (rich guy!) that evening after his private teaching game with Takemiya. The doctor's wife met the two of us at the door. She was a very smiley and gracious roly-poly woman. After the four men were seated around a small table she retreated to what I supposed was the bedroom. Takemiya was very friendly. The conversation covered a lot of topics, including the practice of rich Turkish women in Istanbul cruising the streets looking for attractive young men to invite into their limosines. It turned out that Takemiya had arrived the night before and spent the evening in the resort's game room, which had a billiards table, but nobody else played billiards. If I had only known about that room, I might have played billiards with Takemiya! (Not that I am any good at billiards.) After almost two hours there was a lull in the conversation. At that point the adjoining door slid open and the doctor's wife appeared with a tray of fresh, hot tea and munchies. She earned my admiration. Whatever you may think about the paternalistic Japanese culture of the time and her apparent subservience, she was obviously on the ball. :bow: :bow: :bow:

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

Visualize whirled peas.

Everything with love. Stay safe.


This post by Bill Spight was liked by 3 people: Applebaps, Gomoto, SoDesuNe
Top
 Profile  
 
Offline
 Post subject:
Post #669 Posted: Fri Oct 18, 2019 5:34 am 
Honinbo
User avatar

Posts: 8859
Location: Santa Barbara, CA
Liked others: 349
Was liked: 2076
GD Posts: 312
Hi Bill,

Do you happen to know how's Mrs. Takemiya these days?

Top
 Profile  
 
Offline
 Post subject: Re:
Post #670 Posted: Fri Oct 18, 2019 5:36 am 
Honinbo

Posts: 10905
Liked others: 3651
Was liked: 3374
EdLee wrote:
Hi Bill,

Do you happen to know how's Mrs. Takemiya these days?


No idea.

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

Visualize whirled peas.

Everything with love. Stay safe.

Top
 Profile  
 
Offline
 Post subject: Re: This 'n' that
Post #671 Posted: Thu Nov 07, 2019 10:59 am 
Honinbo

Posts: 10905
Liked others: 3651
Was liked: 3374
A discussion of the opening of a game between Hattori Inshuku (W) and Yamamoto Genkichi on March 29, 1801 (GoGoD 1801-03-29a), based upon ELf's commentary. OC, it was a no komi game by territory scoring, so Elf's evaluation will be off for the actual conditions.

Hattori is the author of the handicap go classic, Okigo Jizai. :)

Click Here To Show Diagram Code
[go]$$Bc Popular start with minor errors
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . 7 . . . . . . . . . . . . . . . |
$$ | . . 1 6 5 9 . . . , . . . . . , . . . |
$$ | . . . 8 2 . . 3 . . . . . . . 4 . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]


:w2: and :b3: had been popular for over a century, but are minor errors, according to Elf. Either may be OK sans komi, OC. :b5: and :w6: are also minor mistakes, as is :b9:.

Click Here To Show Diagram Code
[go]$$Wcm6 Elf variation for :w6:
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . 1 . . . |
$$ | . . . 2 . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . 7 . . . . . . . . . . . . . |
$$ | . . X . X 3 . . . , . . . . . , . . . |
$$ | . . . . O 4 5 X . . . . . . . O . . . |
$$ | . . . . . 6 8 . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]


:w8: - :b11: are joseki. :w12: seems natural, but does not appear in Waltheri.

Click Here To Show Diagram Code
[go]$$Wcm14 Elf variation for :w6:, II
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . O . . . |
$$ | . . . X . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . 8 . 6 . . . . . . . . . . . . . . . |
$$ | . . . . 4 . . . . . . . . . . . . . . |
$$ | . . 7 3 2 O . . . . . . . . . . . . . |
$$ | . . X 1 X O . . . , . . . . . , . . . |
$$ | . . . 5 O X O X . . . . . . . O . . . |
$$ | . . . . . X X . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]


:w14: and :w16: push through. Then :b19: and :b21: confine White to the corner.

Click Here To Show Diagram Code
[go]$$Wcm22 Elf variation for :w6:, III
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . O . . . |
$$ | . . . X . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . 1 . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . X . X . . . . . . . . . . . . . . . |
$$ | . . . . X . . . . . . . . . . . . . . |
$$ | . . O O X O . . . . . . . . . 4 2 . . |
$$ | . . X O X O . . . , . . . . . , 3 . . |
$$ | . . . O O X O X . . . . . 5 . O . . . |
$$ | . . . . . X X . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]


:w22: encloses the top right corner. Then :b23: - :w26: play a joseki that is popular with the bots.

Click Here To Show Diagram Code
[go]$$Bcm9 Elf variation for :b9:
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . 9 . . . . . , . . . . . 8 . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . 7 . . . . . . . . . . . . . . . |
$$ | . . . X 6 . . . . . . . . . . . . . . |
$$ | . . X O X 0 . 5 . , . . . . . , . . . |
$$ | . . 1 O O . 4 B . . . . . . . O . . . |
$$ | . 3 2 . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]


:b9: at 18 is joseki, but Elf regards it as a minor error and prefers the 3-3, claiming the corner and sandwiching White between the corner and the :bc: stone. :b11: does not appear in Waltheri.

Click Here To Show Diagram Code
[go]$$Bcm19 Elf variation for :b9:, II
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . 1 . . . . . |
$$ | . . . X . . . . . , . . . . . O . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . 2 . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . X . 3 . . . . . . . . . . . . . |
$$ | . . . X O . 4 . . . . . . . . . . . . |
$$ | . . X O . O . X . , . . . . . , . . . |
$$ | . . X O O . O X . . . 5 . . . O . . . |
$$ | . X O . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]


Interesting how the play pingpongs back and forth between the bottom left corner and the rest of the board. :b21: - :b23: is a nice flow of the stones.

More later. :)

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

Visualize whirled peas.

Everything with love. Stay safe.

Top
 Profile  
 
Offline
 Post subject: Re: This 'n' that
Post #672 Posted: Sun Dec 15, 2019 5:28 am 
Honinbo

Posts: 10905
Liked others: 3651
Was liked: 3374
Sorry to have kept you waiting. Here is the next diagram.

Click Here To Show Diagram Code
[go]$$Wcm10 Joseki, but . . .
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . 2 6 X . . . . . . . . . . . . . . . |
$$ | . 5 X O X X 4 . . , . . . . . , . . . |
$$ | . . 1 O O 3 . X . . . . . . . O . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]


This was, and has been, joseki, but Elf sees a couple of mistakes. It says that :w12: is a serious error. (I am classifying Elf's errors as minor errors, serious errors, and blunders, mainly by guesswork.) It loses 11½% to par, in Elf's estimation, because it does not play the sente, :w14:.

Click Here To Show Diagram Code
[go]$$Wcm12 Mainline variation for :w12:
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . 4 . . . . . , . . . . . 3 . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . X 2 X . . . . . . . . . . . . . . . |
$$ | d 1 X O X X . . . , . . . . . , . . . |
$$ | . b O O O 5 a X . . . . . . . O . . . |
$$ | e c . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]


White plays the sente at :w12:. Then :w14: and :b15: occupy the open corners. OC, at the time the 4-4 plays were out of favor, not to make a comeback until late in the 19th century. Then White comes back to the push, :w16:. The point is that if Black later plays at a and then b, White c - e makes life in the corner.

Elf says that :b13: in the game loses 21% to par, placing it in blunder territory, I think. Black should take advantage of White's error.

Click Here To Show Diagram Code
[go]$$Bcm13 Mainline variation for :b13:
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . 7 . . . . . . . . . . . . |
$$ | . X . X . . 5 6 8 . . . . . . . . . . |
$$ | . . X O X X 4 3 . , . . . . . , . . . |
$$ | . 1 O O O O 2 X . . . . . . . O . . . |
$$ | . . . . . . 9 0 . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]


As White does not have a forcing play against :b13:, he has to scramble a bit. After :w14: Black puts up a stout resistance. :w22: is tesuji.

Click Here To Show Diagram Code
[go]$$Bcm23 Mainline variation for :b13:, continued
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . X . . . . . . . . . . . . |
$$ | . X . X . . X O O 5 . . . . . . . . . |
$$ | . . X O X X O X . 3 . . . . . , . . . |
$$ | . X O O O O O X a . . . . . . O . . . |
$$ | . . . . . 4 X W 1 . . . . . . . . . . |
$$ | . . . . . . . 2 . . . . . . . . . . . |
$$ ---------------------------------------[/go]


OC, if Black plays atari on :wc: and :w24:, White will atari at a. This is a good result for Black. :)

More to come.

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

Visualize whirled peas.

Everything with love. Stay safe.

Top
 Profile  
 
Offline
 Post subject: Re: This 'n' that
Post #673 Posted: Sun Dec 15, 2019 12:16 pm 
Oza

Posts: 3644
Liked others: 20
Was liked: 4620
Bill

A bit of clarification, please. You say the nobi W12 loses 11.5% under Dobby the Elf. Princess Leela, however, says the win rate between that and atari only falls 3.5% (16k playouts).

I think you've said before that Dobby's numbers are consistently on the high side.

So do we assume all Dobby's numbers are three to four times higher than Leela's? Or do we have to say "apples and pears" - Dobby is so much stronger than Leela that a comparison is a bit futile?

Top
 Profile  
 
Offline
 Post subject: Re: This 'n' that
Post #674 Posted: Sun Dec 15, 2019 1:35 pm 
Honinbo

Posts: 10905
Liked others: 3651
Was liked: 3374
John Fairbairn wrote:
Bill

A bit of clarification, please. You say the nobi W12 loses 11.5% under Dobby the Elf. Princess Leela, however, says the win rate between that and atari only falls 3.5% (16k playouts).

I think you've said before that Dobby's numbers are consistently on the high side.

So do we assume all Dobby's numbers are three to four times higher than Leela's? Or do we have to say "apples and pears" - Dobby is so much stronger than Leela that a comparison is a bit futile?


Are you making both plays by hand? :w12: as an atari and :w12: as a nobi? If not, doing so may give you a larger difference with LZ. I can't say anything about the comparative strengths of the programs. My guess is that in this range Elf may give a difference twice as large as LZ. I'm happy if they prefer the same play, when the Elf differenece is more than 5%. ;)

Edit: If you are not used to doing comparisons this way, what you should see is the winrate estimate for :b13: in each case.

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

Visualize whirled peas.

Everything with love. Stay safe.

Top
 Profile  
 
Offline
 Post subject: Re: This 'n' that
Post #675 Posted: Sun Dec 15, 2019 9:16 pm 
Lives in sente

Posts: 757
Liked others: 114
Was liked: 916
Rank: maybe 2d
John Fairbairn wrote:
Bill

A bit of clarification, please. You say the nobi W12 loses 11.5% under Dobby the Elf. Princess Leela, however, says the win rate between that and atari only falls 3.5% (16k playouts).

I think you've said before that Dobby's numbers are consistently on the high side.

So do we assume all Dobby's numbers are three to four times higher than Leela's? Or do we have to say "apples and pears" - Dobby is so much stronger than Leela that a comparison is a bit futile?

(I'm not quite sure of the connotation you're trying to communicate by calling them weird names :scratch: so maybe I'm misunderstanding something but... )

Leela Zero (with latest 40-block networks) is quite a bit stronger than ELF. Definitely stronger at equal visits, probably stronger at equal time, unless you're using very small numbers of playouts or something is unusual about your hardware. And this is even while being less opinionated about opening moves, in the sense of giving less extreme winrates for them.

Top
 Profile  
 
Offline
 Post subject: Re: This 'n' that
Post #676 Posted: Sun Dec 15, 2019 9:44 pm 
Honinbo

Posts: 10905
Liked others: 3651
Was liked: 3374
lightvector wrote:
John Fairbairn wrote:
Bill

A bit of clarification, please. You say the nobi W12 loses 11.5% under Dobby the Elf. Princess Leela, however, says the win rate between that and atari only falls 3.5% (16k playouts).

I think you've said before that Dobby's numbers are consistently on the high side.

So do we assume all Dobby's numbers are three to four times higher than Leela's? Or do we have to say "apples and pears" - Dobby is so much stronger than Leela that a comparison is a bit futile?

(I'm not quite sure of the connotation you're trying to communicate by calling them weird names :scratch: so maybe I'm misunderstanding something but... )


JF and I both like word play, so no problem as far as I am concerned. IIUC, Dobby is an elf in the Harry Potter books.

Quote:
Leela Zero (with latest 40-block networks) is quite a bit stronger than ELF. Definitely stronger at equal visits, probably stronger at equal time, unless you're using very small numbers of playouts or something is unusual about your hardware. And this is even while being less opinionated about opening moves, in the sense of giving less extreme winrates for them.


Thanks for the info. The Elf commentaries were mostly created in October of 2018, ancient history in bot time. :) But they use a large number of playouts, so there is that.

JF's reference to a fall in winrates suggests to me that he is using game analysis mode, or some such, which focuses on winrate changes between consecutive plays, which may be the drop he is referring to. And when comparing options for the same play, the less preferred option will often have many, many fewer playouts than the top choice, which means that its winrate estimate is much less reliable that that of the top choice. That's why I recommend yoyoma's suggestion to enter the less preferred play by hand. :)

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

Visualize whirled peas.

Everything with love. Stay safe.

Top
 Profile  
 
Offline
 Post subject: Re: This 'n' that
Post #677 Posted: Sun Dec 22, 2019 6:59 am 
Oza
User avatar

Posts: 2401
Location: Tokyo, Japan
Liked others: 2337
Was liked: 1332
Rank: Jp 6 dan
KGS: ez4u
Let's look at this in a little more detail: some historical and some bot. My sources are GoGoD Winter 2019 edition and LZ #254.

This fuseki with the immediate approach answered by the 2-space low pincer has a long history but it is not correct that it was popular at the time of this game. If we search GoGoD we find a total of 85 games stretching from 1647 to 1933. It was, however, mainly a tool of the Yasui family and the main source of 17th century games are biased toward the more famous Honinbos. In the 85 games, 50 featured a player as Black with "Yasui" in their name and 53 featured a "Honinbo" as White. The "Page from Go History" article in Go World #11 is about Dosaku and the superiority of his fuseki theory over that of his rivals of the time. The featured game starts with this 1-2-3. All of these games occurred in the 16th and early 17th centuries.

The last game with a Yasui player as Black appeared in 1722. GoGoD has only 8 games later than 1722. Interestingly, four of those games featured the Yamamoto Genkichi that we see here, two as Black in 1801 (including this one) and two as White in 1809 and 1813.

During the battles of the 17th century, the 2-space low pincer was replaced by the 3-space low pincer where Black responds immediately to White 2, probably due to Dosaku's influence. This is interesting because LZ calculates a nearly identical winrate for Black (at lower visits the 2-space is preferred but the calculated difference falls to only 0.1% better beyond ~100K visits).

Also interesting is what happens when Black responds to White 2 by playing elsewhere. The human preferred play was the 5-3 point at P3 in the lower right, facing the lower left corner. LZ calculates Blue as the upper left star point (D16) instead. However, there is a problem with this analysis. P3 in the lower right does not appear in LZ's policy net. As a result, even with 200K visits LZ had only 9 visits at P3. So should we conclude that D16 is "better"?

Of course the analysis here is more about the joseki rather than the fuseki and indeed the joseki appears far more often than the specific fuseki in this game. If we search GoGoD on an otherwise empty 10x10 corner, we have 167 games with the 1-2-3-"tenuki"-5 of this analysis. White more often replied to the attachment than not, but playing elsewhere was a significant choice at least temporarily. In terms of local responses the two main variations are what is discussed in this analysis: the wedge at D4 played in this game and the outside hane at F4 that Elf "preferred". Both plays first appear in the 1660's in GoGoD.

The outside hane goes back to 1660. The intent when it was played by humans was to trade the outside for the corner. It anticipates the cut by Black and then a series of ataris that ends with White capturing Black's original 3-4 stone.

Click Here To Show Diagram Code
[go]$$W Joseki part 1
$$
$$| . . . , . . . . . ,
$$| . . . . . . . . . .
$$| . . . . . . . . . .
$$| . . . . . . . . . .
$$| . . . . 7 . . . . .
$$| . . . 5 4 8 . . . .
$$| . . X 3 X O . . . ,
$$| . . 9 6 O X 1 X . .
$$| . . . . . 2 . . . .
$$| . . . . . . . . . .
$$+ - - - - - - - - - -[/go]


Play continues to :w6: below and then Black plays elsewhere
Click Here To Show Diagram Code
[go]$$B Joseki part 2
$$
$$| . . . , . . . . . ,
$$| . . . . . . . . . .
$$| . . . . . . . . . .
$$| . 6 . 5 . . . . . .
$$| . . . 3 O . . . . .
$$| . . 2 O X X . . . .
$$| . 4 X O X O . . . ,
$$| . . O X O X O X . .
$$| . . . . 1 X . . . .
$$| . . . . . . . . . .
$$+ - - - - - - - - - -[/go]


Elf chooses a very different variation. The stand that it uses after the first atari does not appear in GoGoD. When we try in LZ, it calculates Blue as the same stand at F5 as Elf. OK let's write that up - the humans are wrong again. Alternatively let's just... wait a while. Two choices dominate LZ's policy net, the stand at F5 and the atari at D4. Its initial calculations do not cover anything else. However, at around 50K visits it has a first peek at the strange-looking alternative of hane at H4. Then at around 130K LZ calculates H4 in more detail. If you have a really fast machine (or in my case if I go and have dinner and then watch a little TV), LZ calculates Blue as shown below given enough visits.

Attachment:
New move cropped.jpg
New move cropped.jpg [ 360.64 KiB | Viewed 8152 times ]


With a million visits LZ calculates a main line as shown below. There are two interesting things here IMHO. First, this variation uses the same corner maneuver by White as the human joseki to take the lower left corner. The difference is that interpolating H4 and the other moves near it leaves more potential on the bottom side. Second, watching the calculated variations rolling out in Lizzie, the bottom left corner (up to White 15 shown in the picture) is stable relatively quickly (by ~160K visits). Thereafter, out to 800K is almost entirely about the rest of the board.
Attachment:
New move continuation cropped.jpg
New move continuation cropped.jpg [ 373.3 KiB | Viewed 8152 times ]

_________________
Dave Sigaty
"Short-lived are both the praiser and the praised, and rememberer and the remembered..."
- Marcus Aurelius; Meditations, VIII 21


This post by ez4u was liked by 2 people: Bill Spight, Uberdude
Top
 Profile  
 
Offline
 Post subject: Re: This 'n' that
Post #678 Posted: Tue Feb 18, 2020 8:32 am 
Honinbo

Posts: 10905
Liked others: 3651
Was liked: 3374
Just a brief note. :) It's been a busy new year for me so far.

I have finally received my 2080 Ti graphics card, along with the rest of the parts for building a computer. I can get it built locally at a reasonable price, and they will stress test it, as well. I should have it up and running by March. :)

As those who have been reading along will know, I have had a bee in my bonnet about the margin of error of the bots' winrate estimates. Surely, I thought, the developers must know what it is, but nobody is telling. Well, I just stumbled across a guy who talks about measures of uncertainty in deep learning, of which margin of error is one. :) He is Yarin Gal, whose doctoral dissertation was entitled, Uncertainty in Deep Learning (2016). He is now an associate professor at Oxford and leads the Oxford Applied and Theoretical Machine Learning (OATML) group. I may not learn enough to apply his methods to my own research, but I am certainly going to study his work. :) If you are interested, here is a link to one of his blog posts, on the science of deep learning. http://www.cs.ox.ac.uk/people/yarin.gal ... _5058.html

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

Visualize whirled peas.

Everything with love. Stay safe.


This post by Bill Spight was liked by 2 people: justtysen, xela
Top
 Profile  
 
Offline
 Post subject: Re: This 'n' that
Post #679 Posted: Tue Feb 18, 2020 3:49 pm 
Lives in sente

Posts: 757
Liked others: 114
Was liked: 916
Rank: maybe 2d
Bill Spight wrote:
As those who have been reading along will know, I have had a bee in my bonnet about the margin of error of the bots' winrate estimates. Surely, I thought, the developers must know what it is, but nobody is telling.


I hope you've realized by now that the developers don't know either. :)

And part of the reason is the issue that you've earlier claimed is "not your problem" - that there's a mathematical issue defining what "margin of error" means in a quantitative way that also corresponds to being useful for Go players. Regardless of whose problem it is, it's not surprising that if you find it tricky yourself to define what you want, then the bot developers - who are on average much, much worse at Go than you, and therefore are much less familiar with what would constitute a success - would fare no better.

And the curse of the machine learning age is that the machine learns the same anyways - the developers don't need to dig into such issues, so they haven't. :)


This post by lightvector was liked by: ez4u
Top
 Profile  
 
Offline
 Post subject: Re: This 'n' that
Post #680 Posted: Tue Feb 18, 2020 4:23 pm 
Honinbo

Posts: 10905
Liked others: 3651
Was liked: 3374
lightvector wrote:
Bill Spight wrote:
As those who have been reading along will know, I have had a bee in my bonnet about the margin of error of the bots' winrate estimates. Surely, I thought, the developers must know what it is, but nobody is telling.


I hope you've realized by now that the developers don't know either. :)


Yes, thanks to the kind remarks of yourself, yoyoma, dfan, jlt, jann, and others, I realized that the learning programs do not produce such error estimates.

Quote:
And part of the reason is the issue that you've earlier claimed is "not your problem" - that there's a mathematical issue defining what "margin of error" means in a quantitative way that also corresponds to being useful for Go players. Regardless of whose problem it is, it's not surprising that if you find it tricky yourself to define what you want, then the bot developers - who are on average much, much worse at Go than you, and therefore are much less familiar with what would constitute a success - would fare no better.

And the curse of the machine learning age is that the machine learns the same anyways - the developers don't need to dig into such issues, so they haven't. :)


Well, I was glad to find someone else who was bothered by that, and who seems to be pretty sharp. Associate prof at Oxford within 4 years of getting his Ph.D., which was on this topic. He has found a way to measure uncertainty in recurrent neural networks by approximating a Bayesian approach, and even to distinguish between noise and what he calls epistemic uncertainty. I'm a bit skeptical about the last, and this may be over my head, anyway, but it looks interesting. ;)

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

Visualize whirled peas.

Everything with love. Stay safe.

Top
 Profile  
 
Display posts from previous:  Sort by  
Post new topic Reply to topic  [ 1056 posts ]  Go to page Previous  1 ... 31, 32, 33, 34, 35, 36, 37 ... 53  Next

All times are UTC - 8 hours [ DST ]


Who is online

Users browsing this forum: No registered users and 1 guest


You cannot post new topics in this forum
You cannot reply to topics in this forum
You cannot edit your posts in this forum
You cannot delete your posts in this forum
You cannot post attachments in this forum

Search for:
Jump to:  
Powered by phpBB © 2000, 2002, 2005, 2007 phpBB Group