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

armpit approach to komoku
http://lifein19x19.com/viewtopic.php?f=48&t=16615		 Page 1 of 1
Author:  MikeKyle [ Fri May 10, 2019 1:28 pm ]
Post subject:  armpit approach to komoku
How do people feel about this nonsense?:

Click Here To Show Diagram Code
[go]$$W
$$ --------------------+
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |
$$ . . . . . . 1 . . . |
$$ , . . . . . , X . . |
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |
$$ . . . . . . , . . . |
$$ . . . . . . . . . . |[/go]


Elfv2 seems to have some slightly confused feelings about it (if we credit it with some feelings of course.)
At 100k+ playouts, Elf v2 wants to start with hoshi for 44.0% black win rate. komoku is only slightly worse at 43.7%. At a reasonable number of playouts, white responds to the polite komoku (top right as shown above) with the natural komoku in the upper left (facing the empty corner). If given enough playouts (~50k) white 2 is changed to a 3-3 in that corner (maybe because of the armpit hit?) But then for black 3, Elf soon finds the armpit hit, and at 100k+ playouts Elf thinks it get's black a 46.5% win rate.
ie. Black plays a very slightly suboptimal opening move, but if white plays a very natural next move then Elf thinks by playing something very strange looking black gets a better possition than the empty board!
It's not really clear to me that this isn't just a quirk in the network.

What does it mean?

The position has symetry across the diagonal which feels slightly creepy to me. whoever plays next in this corner will want to take the 3-3 point or 4-4 point.

if black takes 3-3 then white seems to play away, content with this exchange.
Click Here To Show Diagram Code
[go]$$W Early exchange
$$ --------------+
$$ . . . . . . . |
$$ . . d . c . . |
$$ . a . Q Y . . |
$$ . b . , X . . |
$$ . . . . . . . |
$$ . . . . e . . |[/go]


Now if black plays a or b as he might otherwise to enclose the corner, then the exchange of the triangled stones is interesting. From my (limited) understanding, from a big enclosure, black would usually play 4-4 in response to the 3-4 invasion early in the game. Later in the game, if there's more support, then black might respond at 3-3 like this exchange.

If white wants to follow up (white has kind of prevented herself from a "normal" approach) white can consider c or d but the most frequent seems to be playing e, to which black responds at a.

If black plays 4-4 in response to the armpit hit then white can take 3-3 to default to all that lovely 3-3 invasions stuff. White also has the option to jump or extend along the side, so maybe this is better for white than an early 3-3 invasion.

My pro database hasn't got any occurances of this early in the game, but it's not right up to date.
Is anyone rushing out to try this in their games?
Any feelings?
Author:  Joaz Banbeck [ Fri May 10, 2019 2:23 pm ]
Post subject:  Re: armpit approach to komoku
The two things that bother me about it are:
1) Black can tenuki easily. ( It is a symmetrical position. If white does not mind black having sente here, then black can't mind white having sente either, so it costs him very little to tenuki )
2) If black does decide to play locally, he has great flexibility. He can go high with the 4-4, or low with the 3-3.

The two of these together make the game way too flexible for black. Any natural move works.
I like to put my opponent in a position where he is hard pressed to find a good move. Here, short of blatant stupidity, it is hard to find a bad move.
Author:  atarihuana [ Fri May 10, 2019 4:47 pm ]
Post subject:  Re: armpit approach to komoku
Joaz Banbeck wrote:
The two things that bother me about it are:
1) Black can tenuki easily. ( It is a symmetrical position. If white does not mind black having sente here, then black can't mind white having sente either, so it costs him very little to tenuki )


what? no. yes it is symetrical, but it also is a "contact" play. do not tenuki, the area is "hot"
Author:  Joaz Banbeck [ Fri May 10, 2019 6:22 pm ]
Post subject:  Re: armpit approach to komoku
atarihuana wrote:
Joaz Banbeck wrote:
The two things that bother me about it are:
1) Black can tenuki easily. ( It is a symmetrical position. If white does not mind black having sente here, then black can't mind white having sente either, so it costs him very little to tenuki )


what? no. yes it is symmetrical, but it also is a "contact" play. do not tenuki, the area is "hot"


If it is a bad play for black to tenuki, thus giving white the next move in the position, it was therefore a bad move by white to create the position in which black has first move.

I'm not going to claim to know how valuable the first move is in this position. But if it is a bad move for black to let white have the first move, it is also a bad move for white to let black have the first move.
Author:  Kirby [ Fri May 10, 2019 6:23 pm ]
Post subject:  Re: armpit approach to komoku
atarihuana wrote:
Joaz Banbeck wrote:
The two things that bother me about it are:
1) Black can tenuki easily. ( It is a symmetrical position. If white does not mind black having sente here, then black can't mind white having sente either, so it costs him very little to tenuki )


what? no. yes it is symetrical, but it also is a "contact" play. do not tenuki, the area is "hot"


If white plays here:

Click Here To Show Diagram Code
[go]$$W black tenuki
$$ --------------------+
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |
$$ . . . . . . 1 . . . |
$$ , . . . . . 3 X . . |
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |
$$ . . . . . . , . . . |
$$ . . . . . . . . . . |[/go]


black could play 3-3, and it'd be the same as a 3-3 invasion:
Click Here To Show Diagram Code
[go]$$W
$$ --------------------+
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |
$$ . . . . . . 1 4 . . |
$$ , . . . . . 3 X . . |
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |
$$ . . . . . . , . . . |
$$ . . . . . . . . . . |[/go]


same as:
Click Here To Show Diagram Code
[go]$$W
$$ --------------------+
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |
$$ . . . . . . 3 2 . . |
$$ , . . . . . 1 4 . . |
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |
$$ . . . . . . , . . . |
$$ . . . . . . . . . . |[/go]


If white plays here:
Click Here To Show Diagram Code
[go]$$W black tenuki
$$ --------------------+
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |
$$ . . . . . . 1 3 . . |
$$ , . . . . . , X . . |
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |
$$ . . . . . . , . . . |
$$ . . . . . . . . . . |[/go]


It's a little more painful for black, I guess. Then it's as if black's the one that did your armit thing :-)

Click Here To Show Diagram Code
[go]$$B
$$ --------------------+
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |
$$ . . . . . . O 2 . . |
$$ , . . . . . , 1 . . |
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |
$$ . . . . . . , . . . |
$$ . . . . . . . . . . |[/go]


Edit: Joaz said things more elegantly.
Author:  Uberdude [ Fri May 10, 2019 11:20 pm ]
Post subject:  Re: armpit approach to komoku
See also https://forums.online-go.com/t/elf-open ... ss/21101/1
There Elf not only plays kosumi approach but also kosumi enclosure!
Author:  Bill Spight [ Sat May 11, 2019 2:21 am ]
Post subject:  Re: armpit approach to komoku
I have been expecting to see different zero bots develop different styles of play, particularly in the opening. This may be an example. :)

Top bots play at superhuman strength, yet the opening is so far from the end of the game that there are usually a number of moves that, against best play, will produce the same result. This is especially so if you are only interested in win or loss. But in their quest to win the bots will will not deviate far from what they perceive as best. And if a refutation exists, which may well not be the case in the opening, in self play your opponent is unlikely to find it. So we can expect that zero bots' choices will have an element of path dependence. IOW, they will develop their own style.

Looking at the Elf1 commentaries, it has seemed to me that Elf likes to make strong corners with secure territory early in the game, with the exceptions of starting on the 4-4 and playing the press from the 5-3 point against the 3-4 point, which it shares with other top bots. That does not mean that it plays so territorially a little later in the opening, however.

So seeing Elf2 like the 4-3 approach to the 3-4 and the double komoku enclosure is not too surprising. Maybe other strong bots will follow suit in the next year or two, but maybe not. :) But at this point I doubt if other plays can be considered mistakes, unless Elf estimates them as losing at least 5%.
Author:  Uberdude [ Sat May 11, 2019 2:28 am ]
Post subject:  Re: armpit approach to komoku
MikeKyle wrote:
Elfv2 seems to have some slightly confused feelings about it (if we credit it with some feelings of course.)
At 100k+ playouts, Elf v2 wants to start with hoshi for 44.0% black win rate. komoku is only slightly worse at 43.7%.

Elfv2 is weird/blindspotty in the opening. At low playouts it thinks black 4-4 is 48% and only imagines white 4-4s in reply. If you then manually play an adjacent good (so approach isn't on same side as 4-4) 3-4 the black winrate drops by ~5%, what an unusual genius move the refute the 4-4 it never considered! It does find it for itself after about 50k playouts. I haven't checked if this is a foible of the LZ engine running Elf and if the native Elf engine is any better.
Author:  Bill Spight [ Sat May 11, 2019 2:29 am ]
Post subject:  Re: armpit approach to komoku
atarihuana wrote:
Joaz Banbeck wrote:
The two things that bother me about it are:
1) Black can tenuki easily. ( It is a symmetrical position. If white does not mind black having sente here, then black can't mind white having sente either, so it costs him very little to tenuki )


what? no. yes it is symetrical, but it also is a "contact" play. do not tenuki, the area is "hot"


Probably not. In theory, non-random plays close to other stones usually reduce the temperature. We may have the feeling that they raise it, but today's top bots do not make that assumption. :)
Author:  Bill Spight [ Sat May 11, 2019 2:36 am ]
Post subject:  Re: armpit approach to komoku
Uberdude wrote:
MikeKyle wrote:
Elfv2 seems to have some slightly confused feelings about it (if we credit it with some feelings of course.)
At 100k+ playouts, Elf v2 wants to start with hoshi for 44.0% black win rate. komoku is only slightly worse at 43.7%.

Elfv2 is weird/blindspotty in the opening. At low playouts it thinks black 4-4 is 48% and only imagines white 4-4s in reply. If you then manually play an adjacent good (so approach isn't on same side as 4-4) 3-4 the black winrate drops by ~5%, what an unusual genius move the refute the 4-4 it never considered! It does find it for itself after about 50k playouts. I haven't checked if this is a foible of the LZ engine running Elf and if the native Elf engine is any better.


I would not trust Elf, or any other bot, for analysis in the opening with fewer than 100k playouts. And Elf has a higher margin of error for winrates than other bots. (Which does not mean that its choices are worse. As a bot's choices get better, its winrate estimates should approach 100% or 0% more closely, which will raise its margin of error. ;)) A 5% drop for Leela 11 probably indicates an error, a 5% drop for Elf2 may be the equivalent of a raised eyebrow. {shrug}

Edit: Analysis is different from play. A strong player may have an idiosyncratic style, but for that reason may underestimate plays in a different style. In that case its analysis may be misleading. That's why my preference for an analyst is one trained against a variety of opponents, not by self play. But self play continues to produce stronger bots, so I may have to wait awhile. ;)
Author:  Uberdude [ Sat May 11, 2019 3:25 am ]
Post subject:  Re: armpit approach to komoku
Bill Spight wrote:
Edit: Analysis is different from play. A strong player may have an idiosyncratic style, but for that reason my underestimate plays in a different style. In that case its analysis may be misleading. That's why my preference for an analyst is one trained against a variety of opponents

When doing thorough analysis in Lizzie I will switch between a variety of networks (15b-40btrained, latest LZ, Elfv2, MiniGo, Leela Master) to get a variety of opinions. (Pro tip, Ctrl+number plus engine-command-list setting on config file). When all the bots agree on something it's more likely they are right.
Author:  mhlepore [ Sat May 11, 2019 6:18 am ]
Post subject:  Re: armpit approach to komoku
MikeKyle wrote:
...
My pro database hasn't got any occurances of this early in the game, but it's not right up to date.
Is anyone rushing out to try this in their games?
Any feelings?


Waltheri's pattern search (http://ps.waltheri.net/) reveals 40 games where this approach occurs. Several occur before move 30, and one occurs at move 6 (Mine Yasuhiro vs.Kunisawa Hiroto (2017_07_24)).

And no, I am not rushing to try it out in my games. :-)
Author:  Uberdude [ Sat May 11, 2019 6:30 am ]
Post subject:  Re: armpit approach to komoku
Kosumi approach to 3-4 on such an open board is unusual I think even for bots. What is more common is to see it when Black already has some extra stones in the area e.g Chinese opening so you are more in invade moyo and want maximum wyespace even if strengthen opponent mode. mhlepore, I'm guessing most of those hits had something else in the quadrant?
Author:  mhlepore [ Sat May 11, 2019 6:34 am ]
Post subject:  Re: armpit approach to komoku
Yup Uberdude - that is the case the vast majority of the time.
Author:  MikeKyle [ Sat May 11, 2019 3:49 pm ]
Post subject:  Re: armpit approach to komoku
I can report that I (curious 2 kyu) tried this out in a tournament today.

Unfortunately although it gave me a clearly advantageous opening, my 3 Dan opponent found a nice sequence throughout the middle game which overcame the deficit.
Author:  MikeKyle [ Sat May 11, 2019 4:17 pm ]
Post subject:  Re: armpit approach to komoku
Bill Spight wrote:
atarihuana wrote:
Joaz Banbeck wrote:
The two things that bother me about it are:
1) Black can tenuki easily. ( It is a symmetrical position. If white does not mind black having sente here, then black can't mind white having sente either, so it costs him very little to tenuki )


what? no. yes it is symetrical, but it also is a "contact" play. do not tenuki, the area is "hot"


Probably not. In theory, non-random plays close to other stones usually reduce the temperature. We may have the feeling that they raise it, but today's top bots do not make that assumption. :)


I'm with atarihuana here.

What about the normal approaches? Eg. Knights move:
Click Here To Show Diagram Code
[go]$$W
$$ --------------------+
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |
$$ . . . . . 1 . . . . |
$$ , . . . . . , X . . |
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |
$$ . . . . . . , . . . |[/go]


This kind of approach is very tenuki-able. If black ignores White's approach then it's just as if white played an unpopular opening move (5-3) and black played a very reasonable approach.
White gets a result as if she played the first move in the corner, but that move has to be the unpopular 5-3 move and it seems a little less likely that the 5-3 stone was played with good direction because white may have expected a response.

With the armpit hit whoever plays there next gets a result as if they played the first 3-4 stone in that corner.

When we approach a corner I think we usually accept that if the other player
-played first in the corner
-takes gote
Then we expect that they will have the noticeably better position. We just hope that we will have a reasonably good position considering these factors.
Page 1 of 1 All times are UTC - 8 hours [ DST ]
Powered by phpBB © 2000, 2002, 2005, 2007 phpBB Group
http://www.phpbb.com/