It is currently Thu Mar 28, 2024 3:16 pm

All times are UTC - 8 hours [ DST ]




Post new topic Reply to topic  [ 1056 posts ]  Go to page Previous  1 ... 34, 35, 36, 37, 38, 39, 40 ... 53  Next
Author Message
Offline
 Post subject: Re: This 'n' that
Post #721 Posted: Thu May 21, 2020 4:41 pm 
Judan

Posts: 6725
Location: Cambridge, UK
Liked others: 436
Was liked: 3719
Rank: UK 4 dan
KGS: Uberdude 4d
OGS: Uberdude 7d
jlt, hopefully this makes Bill's point clearer. Which of these boards is better for black (under usual endgame assumption that outside stones alive and rest of board dame, so I'm not asking about value of the big walls)?
Click Here To Show Diagram Code
[go]$$B
$$ -------------
$$ . X O O . O .
$$ . X X X X O .
$$ . . . . . O .
$$ . X O O , O .
$$ . X X X X O .
$$ . . . . . O .
$$ . X O O . O .
$$ . X X X X O .
$$ . . . . . O .
$$ . X O O , O .
$$ . X X X X O .
$$ . . . . . . .[/go]


Click Here To Show Diagram Code
[go]$$B
$$ -------------
$$ . O X X . X .
$$ . O O O O X .
$$ . . . . . X .
$$ . O X X , X .
$$ . O O O O X .
$$ . . . . . X .
$$ . O X X . X .
$$ . O O O O X .
$$ . . . . . X .
$$ . O X X , X .
$$ . O O O O X .
$$ . . . . . . .[/go]


This post by Uberdude was liked by: jlt
Top
 Profile  
 
Offline
 Post subject: Re: This 'n' that
Post #722 Posted: Sun May 24, 2020 9:11 pm 
Honinbo

Posts: 10905
Liked others: 3651
Was liked: 3374
As I guess you all know by now, I am still playing around with the Elf GoGoD commentaries, despite having better programs available now. But the Elf commentaries have a large number of rollouts, and the bulk of the work is already done. :) And as far as significant human mistakes are concerned, I doubt if it's opinion will differ much from today's top bots. OC, the winrate estimates will rarely agree. ;)

Anyway, here is a game where I have worked out my best estimate, given the commentary, for winrate gains and losses to par (according to Elf). I have marked plays with more than a 12% estimated loss as Bad Moves. and I will mark those with more than an estimated 18% loss as Blunders. Since Elf has labeled a correct play in the endgame as a 6% loss, I have enlarged what I think of as Elf's margin of error. :) OC, this is a no komi game from the 19th century, so you have to take the commentary with a large grain of salt.



Enjoy!

This is an unfinished game. I don't know if a complete, annotated game would upload successfully.

_________________
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 #723 Posted: Mon May 25, 2020 3:04 pm 
Honinbo

Posts: 10905
Liked others: 3651
Was liked: 3374
Memorial Day. Be well. :)

Another annotated game. Short, like the last. Same format.

BTW, the Elf commentaries call every choice by Elf a Tesuji. In case you were wondering. ;)



In case you're wondering, Tamura is the future Honinbo Shusai Meijin.

Also in case you're wondering, White's kosumi, :w10:, got a tesuji tag because it got a better winrate estimate than Elf's top choice. OC, there's a lot of noise in the winrate estimates, even with a lot of rollouts.

_________________
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 #724 Posted: Mon May 25, 2020 8:54 pm 
Honinbo

Posts: 10905
Liked others: 3651
Was liked: 3374
I wrote:

Bill Spight wrote:
Here is a simple endgame corner position. Per convention, the outside White stones are alive. Territory scoring, counting territory in seki.

Click Here To Show Diagram Code
[go]$$Bc Endgame corner
$$ -------------
$$ | . . . X O . .
$$ | O X X X O . .
$$ | . . X O O . .
$$ | X . X O . . .
$$ | O X X O . . .
$$ | O O O O . . .
$$ | . . . . . . .[/go]


Easy questions.

1) What is the usual territorial value of the corner?

2) How much does a gote or reverse sente gain?


jlt answered these questions. No point in hiding the answers now. :)

1) The usual territorial value of the corner is 3 pts, (for Black, by convention).

2) A gote or reverse sente gains 4 pts,

Next question:

Quote:
3) When Black is komaster, what is the territorial value of the corner?


This question is ambiguous, as the territorial value depends upon whether Black makes the ko or not. My apologies for a poorly worded question.

jlt also answered this question, as well. it is 7 pts. if Black makes and wins the ko, and 3 pts. if Black does not make the ko.

Note, however, that if Black makes the ko, White gets points elsewhere in compensation for losing the ko, depending on when the ko is fought. That compensation is called the ko exchange.

That leaves three questions.

Quote:
4) How much does each play in the ko gain, on average?

5) How much does Black gain by winning the ko versus the territorial value of the corner?

6) How should White play this corner?


4) is easy. 5) is ambiguous, as well. 6) is implied in the previous discussion between jlt and me, and the answer to 4).

4) How much does each play in the ko gain, on average?

This question is easy to answer for area scoring, as we just have to count the area covered by the corner. It is 16 pts. So winning or losing the ko makes a difference of 32 pts., and each move in the ko gains on average 32/3 = 10⅔ pts. To get the average gain under territory scoring which counts points in seki, we subtract 1 pt. and get 9⅔ pts.


5) How much does Black gain by winning the ko versus the territorial value of the corner?

Well, assuming correct play, Black will win the ko when she should and then the value of the corner is 7 pts., so she gains nothing by winning the ko. Not a good question, I am afraid, as it is a question of definition. My apologies.

6) How should White play this corner?

An important question, the answer to which may be counterintuitive.

In theory we assume that there are plays elsewhere about which we know nothing except how much the largest play or plays gain, which we call the ambient temperature. If we know more, the answer can get complicated. I know, because this is a question the likes of which I worked on for years. Berlekamp's komaster theory makes giving an approximate answer tractable. :) As we know more, the theory may be expanded to cover more complicated situations.

Since Black can win the ko Black does not need to waste a move to prevent the ko. Black can wait for White to atari the single stone, and then Black can make and win the ko. If White never plays that atari, then Black gets 7 pts. in the corner without trying. The question for White, then, is when to play the atari.

If White waits to play the atari until the ambient temperature, which is how much White can gain in the ko exchange, is 3 pts., then Black gets 7 pts. in the corner and White gets 3 pts., for a net result of 4 pts., on average, which is 1 pt. better than the usual average value of the corner. So White should play the atari, sooner, if possible.

How soon can White effectively make the atari?

1) Suppose that Black makes and wins the ko. Then White gains the ambient temperature, T, in exchange for 7 pts. in the corner. The net result is 7 - T, on average.

2) Suppose that Black ignores the atari and lets White kill the corner. The net result is 2T - 22. This is better for Black when T > 9⅔. So White can play the atari when T ≤ 9⅔. I.e., when the ambient temperature is low enough to fight the ko.

Edit: This is true so far as it goes, but as jlt pointed out, since by connecting Black returns to a position worth 3 pts., White can play the atari when the temperature is as high as 12½. :)

3) Suppose that Black simply connects without making seki. If T > 4, then play will continue elsewhere (as a rule) and the average value of the corner will be 3 pts. This is better for Black, on average, than making the ko, as 3 > 7 - T.

Putting all this together means that when Black is komaster, if White waits until T < 4 to play the atari, then Black will make and win the ko for a net result of 7 - T. However, when 9⅔ > T > 4, White will normally be able to play the atari with sente, and Black will connect for an average result of 3 pts. in the corner.

Typically such corner positions arise in the middle game, when the ambient temperature is high. So White will normally have many opportunities to play the atari with sente, and should do so, as a rule.

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

Visualize whirled peas.

Everything with love. Stay safe.


Last edited by Bill Spight on Tue May 26, 2020 6:57 am, edited 1 time in total.

This post by Bill Spight was liked by: jlt
Top
 Profile  
 
Offline
 Post subject: Re: This 'n' that
Post #725 Posted: Tue May 26, 2020 2:00 am 
Gosei
User avatar

Posts: 1753
Liked others: 177
Was liked: 491
I still don't understand what happens if 9⅔ < T < 12½.

Since 2T-22 < 3, if :w1: ataris it is more interesting for Black to connect than to play elsewhere and let White kill the corner.

But how do you determine if :w1: is better at A17 (atari) or elsewhere?

Top
 Profile  
 
Offline
 Post subject: Re: This 'n' that
Post #726 Posted: Tue May 26, 2020 2:44 am 
Honinbo

Posts: 10905
Liked others: 3651
Was liked: 3374
jlt wrote:
I still don't understand what happens if 9⅔ < T < 12½.

Since 2T-22 < 3, if :w1: ataris it is more interesting for Black to connect than to play elsewhere and let White kill the corner.


You're right. :) White can atari with sente up to T = 12½.

Quote:
But how do you determine if :w1: is better at A17 (atari) or elsewhere?


In general, White will not take a loss by waiting unless the temperature drops below 4. But the closer the temperature gets to 4, the more likely that conditions will arise so that Black can afford to make the ko.

_________________
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 #727 Posted: Tue May 26, 2020 2:49 pm 
Honinbo

Posts: 10905
Liked others: 3651
Was liked: 3374
Welcome news! :D

The book about my wife's great-great-grandmother, Henrietta Wood, has won the Pulitzer Prize in History. :) I got an email from the author, Caleb McDaniel, today.

See this note: https://lifein19x19.com/viewtopic.php?p=248096#p248096

And here is a picture that Caleb took of Winona and me when he visited a few years ago. (Repost.)


Attachments:
Winona Bill LR.jpg
Winona Bill LR.jpg [ 155.23 KiB | Viewed 8850 times ]

_________________
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 7 people: EdLee, ez4u, gowan, Kirby, SoDesuNe, sorin, Uberdude
Top
 Profile  
 
Offline
 Post subject: Re: This 'n' that
Post #728 Posted: Sat May 30, 2020 2:32 pm 
Honinbo

Posts: 10905
Liked others: 3651
Was liked: 3374
We don't need no stinking ladders

Here is a game of the early modern, i.e., post 1500, era, between Nakamura Doseki, Meijin, and Yasui Santetsu. Of some interest is White 102, which serves as a ladder maker. In its commentary, however, Elf prefers to make the atari for the non-ladder. :) OC, ladders are a weak spot for more than one top bot. You may wonder why I chose to show this variation. Then again, you may not. ;)


_________________
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: Farodin, Knotwilg
Top
 Profile  
 
Offline
 Post subject: Re: This 'n' that
Post #729 Posted: Sat Jun 06, 2020 4:50 am 
Honinbo

Posts: 10905
Liked others: 3651
Was liked: 3374
I have been thinking lately about how to advance as an SDK, and I remembered a player who was frustrated by being stuck for a long time at a certain level, who rose to the occasion, remade his game, and made solid progress. :)

So here's a shoutout to our friend, Fedya, an inspiration to us all. :D

_________________
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 #730 Posted: Sat Jun 06, 2020 6:11 am 
Honinbo

Posts: 9545
Liked others: 1600
Was liked: 1711
KGS: Kirby
Tygem: 커비라고해
Congratulations about the book. And thanks for sharing your photo, again. I’ve seen that picture before, but it’s the only way I’ve ever been able to put a face to your name.

Regarding Fedya, what steps might have been involved to “remake his game”? I suppose it means to play better, or maybe with a different style. Anything else?

_________________
be immersed

Top
 Profile  
 
Offline
 Post subject: Re: This 'n' that
Post #731 Posted: Sat Jun 06, 2020 7:07 am 
Honinbo

Posts: 10905
Liked others: 3651
Was liked: 3374
Kirby wrote:
Congratulations about the book. And thanks for sharing your photo, again. I’ve seen that picture before, but it’s the only way I’ve ever been able to put a face to your name.

Regarding Fedya, what steps might have been involved to “remake his game”? I suppose it means to play better, or maybe with a different style. Anything else?


I know that Fedya worked with Leela 11, and adopted a more aggressive style. His overall attitude improved, as well. :) A number of us recognized the transformation in a game he posted in April of last year. https://www.lifein19x19.com/viewtopic.php?f=4&t=16561

_________________
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 #732 Posted: Sat Jun 06, 2020 1:27 pm 
Honinbo

Posts: 10905
Liked others: 3651
Was liked: 3374
Think like a Grand Robot.

Bots do make unusual plays. I just ran across one. :)

Click Here To Show Diagram Code
[go]$$Bcm9 Up in the air, flying high
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . . . . . . . . . . X . . . . |
$$ | . . . , . . . . . , . . . . . , X . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . O . . . . . . . . . . . . . . . . |
$$ | . . . c . . . . . . . . . . . . . . . |
$$ | . . b . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , a . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . 7 . . . . . |
$$ | . . X . . . . . . . . . . . . . . . . |
$$ | . . . , 3 . . . . 6 . . . 1 . , O . . |
$$ | . . . X 2 4 . . . . . . . . O . . . . |
$$ | . . . 5 . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]

The four enclosure position is from a 19th century game, but could arise today. Black has fallen behind, because of the two tight enclosures.

In the game Black played :b9: at a, the traditional extension between facing enclosures, but the shoulder play is modern AI style. This diagram shows Elf's main variation for :b9:. The White response, :w10: - :w14:, is also the modern AI style. The jump. :b15:, is a new one on me, however. ;) Elf gives it 17.9k rollouts. It is 1% better than b and c, which get 5.6k and 3.1k rollouts, respectively.

Is this jump on any other bot's radar? If not, how does it evaluate it? Thanks.

_________________
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: ez4u, Uberdude
Top
 Profile  
 
Offline
 Post subject: Re: This 'n' that
Post #733 Posted: Sat Jun 06, 2020 2:18 pm 
Lives in sente

Posts: 757
Liked others: 114
Was liked: 916
Rank: maybe 2d
Yes.

KataGo recent 40 blocks, large number of playouts, analysisWideRootNoise = 0.05, so as to force exploration and evaluation of a much wider range of moves.

Attachment:
kata40bwide005.png
kata40bwide005.png [ 1.15 MiB | Viewed 8602 times ]


And here's the raw policy, showing that the move also is fairly prominent in the bot's "instinct", suggesting that this "kind" of one point jump into the center is something that it would consider a plausible move in situations "like" this one, based on whatever generalizations the net is making from its training experience. Rather than something the bot itself considers unusual.

Attachment:
kata40brawpolicy.png
kata40brawpolicy.png [ 1.06 MiB | Viewed 8602 times ]


This post by lightvector was liked by 3 people: Bill Spight, ez4u, Uberdude
Top
 Profile  
 
Offline
 Post subject: Re: This 'n' that
Post #734 Posted: Sat Jun 06, 2020 3:32 pm 
Oza
User avatar

Posts: 2401
Location: Tokyo, Japan
Liked others: 2338
Was liked: 1332
Rank: Jp 6 dan
KGS: ez4u
If Black does not play locally next (e.g. extends at the top), does White have a specific local continuation to exploit Black's failure to play another move at the bottom? In other words, can the bot give us a hint why the local play is rated so high?

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

Top
 Profile  
 
Offline
 Post subject: Re: This 'n' that
Post #735 Posted: Sat Jun 06, 2020 3:47 pm 
Honinbo

Posts: 10905
Liked others: 3651
Was liked: 3374
Many thanks, lightvector. Verrrry interesting. :)

Elf, at any rate, does not particularly like either the New Fuseki or Cosmic Go. However, in many situations it does like jumps toward the center and other center tending moves. :) Interesting that in this specific case KataGo values this jump, as well.

_________________
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 #736 Posted: Thu Jun 11, 2020 11:42 pm 
Judan

Posts: 6725
Location: Cambridge, UK
Liked others: 436
Was liked: 3719
Rank: UK 4 dan
KGS: Uberdude 4d
OGS: Uberdude 7d
ez4u wrote:
If Black does not play locally next (e.g. extends at the top), does White have a specific local continuation to exploit Black's failure to play another move at the bottom? In other words, can the bot give us a hint why the local play is rated so high?


Tenuki to top side is only 0.5% or 0.1 points worse, so I wouldn't say local play is so much better :). Normal p4 block is a white continuation from my phone bot, but it seems to like my human idea of cap when I try it so maybe that's a reason for urgency. Also it likes q7, if black jumps after that White can reasonably cap again.

Top
 Profile  
 
Offline
 Post subject: Re: This 'n' that
Post #737 Posted: Fri Jun 12, 2020 3:52 am 
Oza

Posts: 3647
Liked others: 20
Was liked: 4626
If I may switch to another game, I think there are some insights to be gleaned on the above from a commentary by Ohashi Hirofumi.



In his book "Encyclopaedia of Go AI", Ohashi looks in some depth at a game in the 1st Nihon Ki-in International Championship between Iyama Yuta (White) and DeepZen (2017-03-23b if you want to look it up in GoGoD). In the position above, where the last two moves were the triangled ones, he called the one just played by Iyama an Iyama-ism. It looks rather AI-ish, doesn't it? But Leela for one doesn't like it. More to the point, Ohashi discusses this particular position under the heading "The difference between humans and go AI".

He doesn't explain that in the depth I'm sure we would all like, so I am going to try to pull together the little that he does say there with what he and other pros say elsewhere. Do keep that "I" in mind. Have a cellarful of salt handy. But Ohashi is human, too, so even for him a grain of salt may have to be reserved.

As a background point first, one of the first and most important features of AI play that Japanese pros thought they had spotted and deigned to share with us was a strong AI emphasis on inducing overconcentration in the opponent's positions very early on. In Bill's example above, this would be exemplified by moves such as White 10. Another feature was the use of very early probes. These two features often come together, and so an attempt at overconcentrating may be left off part way through.

If you look at many pro games even today, you can make sense of them to a large degree by not viewing them as fights over territory or thickness or influence but as fights over overconcentrating each other. For me the best analogy is a push-and-shove sumo bout, which of course may also start with a slap and tickle probing stage. Now, in that kind of sumo bout, within reason it's not really about who's bigger or heavier or faster. It's about who ends up first in bad shape, or in overconcentrated shape. He is the one who can be made to topple over or step out of the ring first. In go terms, bad shape punishes itself.

Obviously good sumo wrestlers (who also throw a lot of salt) learn good shape. The circled move by Iyama above can be seen as an example of that. Ohashi makes the point that this is the sort of shape that is very attractive to pros. It appears to demolish aji in the corner, and White can look forward to a large corner territory. Which is true. At least that's how it worked out in the game. But it appears that Iyama overlooked one salient point. And that blond spot may be what all pros had been missing until AI showed them the error of their ways.

Before I get on to that specific point, have a look again at the position as a whole and try to form some judgements about how it will all end up. Thoughts on what Iyama ought to have played for his last centre move and how Black now should respond would not go amiss.

What Ohashi himself notes is that Black has first occupation of two corners below, and with two stones on the fourth lines at the top, he has an advantage in speed.

What I will add is based on the final position. 235 moves were played. Iyama resigned when about 15 points behind on the board. Nothing was killed. He was just crushed. That is, overconcentrated.

White ended up with a huge upper-left corner territory - about 35 points. He also ended up making a very large investment there - about 45 stones. But in the rest of the board, Black has the upper hand (i.e. prior first occupation) and so whenever White played there, White would have to expect to take a bit of battering - which is another way of saying he would end up overconcentrated. In fact he made three extra groups, all live but at a significant cost - a total of close to 60 stones for a total of about 20 points of territory (and he also lost some stones as prisoners). That really was overconcentration. Black, meanwhile, with the upper hand, can control the development of the game and so is much more likely to end up in good shape. He is like the sumo wrestler who is genuinely heavier of faster or heavier than his opponent. That doesn't guarantee victory, but it surely helps. And one can't but help but note that it was Go Seigen's speed of foot that made his go so potent - and AI like!

Now to the specifics. Ohashi's point first: he noted that DeepZen punished White's play by playing at C. Ohashi said that this is the kind of play that characterises the strength of not just DeepZen but all AI bots. What is going on is sabaki.

If you are hamstrung by thinking of sabaki as making light, flexible shape you won't get very far in understanding what Ohashi saw. The point is, sabaki just means "coping". Ending up with light or flexible shapes may be counted as common attributes of sabaki. But so is ending up with a heavy but solid shape. That is too often overlooked. What is apparently even more overlooked, even by pros, is that another useful attribute of good sabaki is that it overconcentrates the opponent. It's pretty obvious as soon as you think about it, but that doesn't stop it being overlooked. And it worked a treat that way for Black in this game.

Keeping that in mind will perhaps allay possible surprise at the move Leela recommended for White instead of Iyama's flying bedstead. It was A. At first sight this may seem like self-inflicted overconcentration. But that's also what a honte is, is it not? The point there is that it accepts a small dose of overconcentration now to avoid a bad case of it later. In topical terms, it's a sort a vaccine. And, note well, because it eliminates sabaki for Black, it automatically makes Black weak in this area.

Moving on now to another favoured choice of Leela's: if White really does have to play in the centre, Leela suggests B. We can now see some similarity with Bill's game and Katago's "thinking".

Katago shows that it considers lots of moves in the centre, and especially in the lower-right part of the centre. But why does it plump for 15, and why did Leela prefer B?

I used to think it was all down to the bots' ability to see further ahead, and to gauge how fights in each corner would eventually coalesce at some region in the centre - invisible to us but visible to bots. I now feel ashamed of such naïve thinking.

Based partly on what Ohashi says (in connection with other games rather than the one above), I am now inclined to believe that the bots are just making shape moves in the centre, and the choice of move is simply the one that gives the shape that is least likely to end up overconcentrated.

It seems that pros have already twigged that - or something like that. The problems they are having with such high moves are down rather to understanding the timing and (maybe more important if I interpret Ohashi correctly) the order of such moves. In other words, these moves too, like the pure overconcentration-inducing moves, also have a probe element. If we extend the pugilistic analogy, these moves are like timing the changeover tag in tag wrestling. There's a lot more to it than most fans realise.


This post by John Fairbairn was liked by 2 people: Bill Spight, gowan
Top
 Profile  
 
Offline
 Post subject: Re: This 'n' that
Post #738 Posted: Fri Jun 12, 2020 8:31 am 
Honinbo

Posts: 10905
Liked others: 3651
Was liked: 3374
John, thank you for your thorough and interesting analysis. :)

John Fairbairn wrote:
If I may switch to another game, I think there are some insights to be gleaned on the above from a commentary by Ohashi Hirofumi.



In his book "Encyclopaedia of Go AI", Ohashi looks in some depth at a game in the 1st Nihon Ki-in International Championship between Iyama Yuta (White) and DeepZen (2017-03-23b if you want to look it up in GoGoD). In the position above, where the last two moves were the triangled ones, he called the one just played by Iyama an Iyama-ism. It looks rather AI-ish, doesn't it? But Leela for one doesn't like it.

FWIW, neither does Elf. ;)

John Fairbairn wrote:
More to the point, Ohashi discusses this particular position under the heading "The difference between humans and go AI".

He doesn't explain that in the depth I'm sure we would all like, so I am going to try to pull together the little that he does say there with what he and other pros say elsewhere. Do keep that "I" in mind. Have a cellarful of salt handy. But Ohashi is human, too, so even for him a grain of salt may have to be reserved.

As a background point first, one of the first and most important features of AI play that Japanese pros thought they had spotted and deigned to share with us was a strong AI emphasis on inducing overconcentration in the opponent's positions very early on.

I noted early on that AlphaGo pincers about half as often as humans. Maybe I was the first? ;)

Anyway, I think that this may exemplify starting from where we are. The pros thought about overconcentration, I thought about pincers.

John Fairbairn wrote:
In Bill's example above, this would be exemplified by moves such as White 10. Another feature was the use of very early probes. These two features often come together, and so an attempt at overconcentrating may be left off part way through.

:w10:, the side attachment against the enclosure, goes back in human play to at least the 19th century, if not in this context.

I definitely agree about early probes. However, I have some doubts about the overconcentration hypothesis. If the purpose of the play is to induce overconcentration, why does the opponent reply? The bots are happy with both the initial play and the local response. Why accept overconcentration? Why not tenuki? We know the bots tenuki a lot. The fact that they don't indicates that the local reply does not result in overconcentration. In addition, the bots at times willingly accept what appears at first blush to be overconcentration with no prompting at all. My working hypothesis is that the bots' attitude (to anthropomorphize) to overconcentration is more like that of top players in the 19th century than that of top players in the 20th century or today. They care, but not so much.

If that's the case, then what is the purpose of plays like :w10:, aside from possibly being probes? My working hypothesis is that the purpose is to build up some strength or influence radiating in the direction of future play. It's like the leaning part of a leaning attack without the attack. ;)

John Fairbairn wrote:
If you look at many pro games even today, you can make sense of them to a large degree by not viewing them as fights over territory or thickness or influence but as fights over overconcentrating each other. For me the best analogy is a push-and-shove sumo bout, which of course may also start with a slap and tickle probing stage. Now, in that kind of sumo bout, within reason it's not really about who's bigger or heavier or faster. It's about who ends up first in bad shape, or in overconcentrated shape. He is the one who can be made to topple over or step out of the ring first. In go terms, bad shape punishes itself.

Emphasis mine. :)

John Fairbairn wrote:
Obviously good sumo wrestlers (who also throw a lot of salt) learn good shape. The circled move by Iyama above can be seen as an example of that. Ohashi makes the point that this is the sort of shape that is very attractive to pros. It appears to demolish aji in the corner, and White can look forward to a large corner territory. Which is true. At least that's how it worked out in the game.

It also prevents the Black tachi on the same point, which would bolster Black's pincer and start to make some territory.

John Fairbairn wrote:
But it appears that Iyama overlooked one salient point. And that blond spot may be what all pros had been missing until AI showed them the error of their ways.

That's a big maybe, good buddy. :)

John Fairbairn wrote:
What Ohashi himself notes is that Black has first occupation of two corners below, and with two stones on the fourth lines at the top, he has an advantage in speed.

Right. The problem, in my view, and in Elf's, is the high approach to the top right corner with :w6: instead of occupying the last open corner. Elf docks :w6: 6½% by comparison. :w6: violates my proposed last move principle for the opening: Occupy the last open corner.

Click Here To Show Diagram Code
[go]$$Wcm6 Elf's recommended variation for :w6:
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . X . . . . . . . . . . . . . |
$$ | . . . O . . . . . , . . . 2 . , X . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . O . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . 1 . . . . . , . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . 3 . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]

Moi, I do not find this development appealing for White. The top right enclosure looks so good, but what can I say? ;)

OTOH, I do like DeepZen's play in response to the high approach. :)

Click Here To Show Diagram Code
[go]$$Wcm6 Actual game
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . X . . . . . . . 3 2 4 . . . |
$$ | . . . O . . . . . , . . . 5 1 , X . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . O . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . 6 , . . . . . , . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]

DeepZen secures the top right corner with sente and then occupies the last open corner. What's not to love? :)

John Fairbairn wrote:
Now to the specifics. Ohashi's point first: he noted that DeepZen punished White's play by playing at C. Ohashi said that this is the kind of play that characterises the strength of not just DeepZen but all AI bots. What is going on is sabaki.

Well, not all AI bots. Elf does not like the hane at C. In fact, it docks it 6½% by comparison with its top choice in the next diagram.

Click Here To Show Diagram Code
[go]$$Bcm17 One space jump from the side
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . O X . . . . . . . O X X . . . |
$$ | . . . O . O . . . X . . . O O , X . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . O . . . . . . . O . . 1 . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . X , . . . . . , . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]

This is a kind of play that I am coming more and more to appreciate, a strong play near what appears to be the hot spot or focus of interest. Instead of simply strengthening Black on the top side, which is what I would do, :b17: strengthens Black on the right side while threatening White and bolstering Black on the top side indirectly. I have to admire the play. :)

Click Here To Show Diagram Code
[go]$$Bcm17 Elf's variation for :b17:
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . 5 . . . . . . . |
$$ | . . . . O X 4 . . . . 3 . O X X . . . |
$$ | . . . O . O . . . X . . . O O , X . . |
$$ | . . . . . . . . . . . 2 . . . . . . . |
$$ | . . O . . . . . . . W . . 1 . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . X , . . . . . , . . . . . X . . . |
$$ | . . . . 6 . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]

:w18: breaks the sector line and then :b19: takes away White's potential base while making a small base for Black. :w20: is big. Then :b21: builds strength and threatens to connect underneath. Next, :w22: approaches the bottom left corner. :b17: - :b21: refutes the high flying :w16: ( :wc:).

John Fairbairn wrote:
Keeping that in mind will perhaps allay possible surprise at the move Leela recommended for White instead of Iyama's flying bedstead. It was A. At first sight this may seem like self-inflicted overconcentration.

That's Elf's play, as well. And yes, at first glance it does seem overconcentrated. White has played 5 of his first 8 moves (4 net) in the top left corner, while his other 3 stones are under attack. But again, I am reminded of pre-20th century play by top players who knew about overconcentration.

John Fairbairn wrote:
But that's also what a honte is, is it not? The point there is that it accepts a small dose of overconcentration now to avoid a bad case of it later. In topical terms, it's a sort a vaccine. And, note well, because it eliminates sabaki for Black, it automatically makes Black weak in this area.

Moving on now to another favoured choice of Leela's: if White really does have to play in the centre, Leela suggests B. We can now see some similarity with Bill's game and Katago's "thinking".

Also AlphaGo, who taught us that the three White stones in the top right do not need an extension, but can jump into the center if pincered.

John Fairbairn wrote:
Katago shows that it considers lots of moves in the centre, and especially in the lower-right part of the centre. But why does it plump for 15, and why did Leela prefer B?

I used to think it was all down to the bots' ability to see further ahead, and to gauge how fights in each corner would eventually coalesce at some region in the centre - invisible to us but visible to bots. I now feel ashamed of such naïve thinking.

Based partly on what Ohashi says (in connection with other games rather than the one above), I am now inclined to believe that the bots are just making shape moves in the centre, and the choice of move is simply the one that gives the shape that is least likely to end up overconcentrated.

I think there is something there. :) Bots read the whole board, OC, which is why they sometimes make local mistakes, but I have come to think that their "whole board thinking" is basically different from that of 20th century and 21st century pros. That is why, perhaps, we do not see examples of direction of play as understood in the 20th century, while we do see examples like those from the 19th century. Indeed, if the view of the bots of the center is more local than that of humans, then making shape in the center would be important, just is it is important in any other local region of the board. :)

_________________
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 #739 Posted: Fri Jun 12, 2020 8:46 am 
Oza
User avatar

Posts: 2408
Location: Ghent, Belgium
Liked others: 359
Was liked: 1019
Rank: KGS 2d OGS 1d Fox 4d
KGS: Artevelde
OGS: Knotwilg
Online playing schedule: UTC 18:00 - 22:00
Bill, I find the argument you offer against the hypothesis of overconcentration flawed. Whatever the concept, if the opponent bot answers it's because they judge that not answering locally is even worse. That doesn't mean they are not forced into some degree of overconcentration.

That's the whole point of sente, regardless of the underlying concept: you play a move because you think the opponent should answer locally and you think it gives you some benefit. When the opponent answers, that can still be his best move. And that in turn doesn't mean the sente was wrong.

Many concepts would lose their meaning, like "shape", if refusing to be coerced into it, like "bad shape", were always better than allowing for it.

Go concepts exist because the game isn't solved. When the game is solved, there is only the concept of "best move".

Top
 Profile  
 
Offline
 Post subject: Re: This 'n' that
Post #740 Posted: Fri Jun 12, 2020 8:57 am 
Judan

Posts: 6087
Liked others: 0
Was liked: 786
A solved game need not lose concepts, which might classify correct plays of same kinds.

Top
 Profile  
 
Display posts from previous:  Sort by  
Post new topic Reply to topic  [ 1056 posts ]  Go to page Previous  1 ... 34, 35, 36, 37, 38, 39, 40 ... 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