Locally, I think hane is just better:

Empty board, attachment to tengen: hane is 0.35 pt better than extend according to Katago
Attachment to corner stones, I guess something similar (0.2-0.4 according to some simple tests.)

But usually opponents attach when they try to make a sabaki. Extend is sometimes suggested in the literature, to give less forcing moves.
But reviewing my games with the help of silicon, I realised that I extend too often.

My question is: when is extending appropriate? Do you have some heuristic of some clear example(s)?

In my opinion first instincts should be as follows:

1) attachment --> hane
2a) cross cut --> extend
2b) two-step-hane --> defend your cutting point (broadly speaking, could just be another extension in some cases)

There are a lot of patterns especially in corner positions (probe against an enclosure).

Aside from these instincts, you have to read because depending on the surrounding position, you can actually play anything as a response. Even a third hane after the two-step-hane.

If you get stronger, you might want to calculate which sequence ends in sente for you. So a small local loss may be better globally, if you have the initiative.

My advice: Experiment, experiment, experiment- see how things unfold on your level and review what you could have done better : )

_________________
My "guide" to become stronger in Go

I have already experimented a bit, now is a good time to try to investigate this detail a bit more (my level is around KGS 1 dan).
I would hope for some more abstract heuristic. Maybe I will try to come up with one.

But for now, I will thy to search for professional games with attach-extend, and check if the bot agree, to try to find some cases where extending is appropriate.

Example 1:
B played extend, but it is 1.2pt worse than L13 and 0.7pt worse than J13. So I guess it is not easy even for pros.

Are those, board points or percentage points? If they are percentage points, the small differences tell me that all three plays are playable.

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

Visualize whirled peas.

Everything with love. Stay safe.

The normal theory I suppose is to extend rather than hane when your opponent is trying to make sabaki in your strong area, because you are trying to reduce the number of subsequent sente moves he gets to help make shape (counter hanes, crosscuts). When you are weaker or around equal strength in the area then hane is the default better move.


My expanded version of the oft quoted to beginners proverb to not atari but extend from a crosscut is:

When crosscut is a bad move: extend (is normally best).
When crosscut is a good move: atari (is often better).

BadukDoctor had a nice video on such a situation recently:

I think there's a fundamental thinking error here. It's a bit like when you plot a course from a map. If you set off a millimetre at the wrong angle, you very soon end up a mile from where you want to be.

The millimetre error here is assuming it's hane or extend. In fact the second option is hiki not nobi, and that is part of a very different dynamic. In this case the hiki is part of a combination shape made by the follow-up keima at H12. The terms are differentiated for a reason.

There's also a deeper thinking flaw, I feel. We can see this if we tackle the problem, as the poster wishes, of coming up with a general heuristic. There is already a good one: 上手真っすぐ、下手コスム. This means the strong player plays straight (nobi, hiki), the weak player plays diagonally. The main diagonal play is kosumi, of course, and tends to be highlighted because of the phrase ヘボのコスミ - the duffer's kosumi. But hane is a kind of diagonal play.

One reason the strong player plays straight is to avoid leaving weaknesses. If you judge that you will have to come back to this local area and make a repair move at some point later, you should probably be playing straight - otherwise diagonal is usually more severe. One of the best examples of the nice judgement required is when making the fukure swelling shape (tiger's mouth upside down) on the 3rd/4th lines. It feels good, it looks good, but it often bites you in the bum.

But I think there's a deeper reason for playing straight. It cuts down the number of variations. Even pros get lost in the forest of variations. So, where there is no over-riding reason to play a severe diagonal kind of move, they often take the safe option.

In my view, that still makes sense even if a bot tells you a hane/kosumi/whatever is better. The bots can justify that severe move with their much deeper and more reliable tree searches (and I am really referring to global play rather than local play). If you (even be you a pro) can't justify the apparent severity, you should play straight and safe so as to keep a modicum of control. Play to your human strengths and weaknesses, not the bots'.

This lesson applies in cricket, too. Play with a straight bat.

Of course when they attach for sabaki, tenuki is also an option; a common shape is this tenuki of side attach to a strong 3rd line stone to take the side extension immediately, e.g. from http://ps.waltheri.net/database/game/43319/, here it's also making a base for the white group to the right.

Click Here To Show Diagram Code

[go]$$Bc
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . O . . . |
$$ | . . . . . . . . . . O . . O . . X . . |
$$ | . . O O . . O . . , . . . . . X . . . |
$$ | . . . . . . . . . . . . . O . . . . . |
$$ | . . X . O . O . . X . . . . . X . . . |
$$ | . . . X X X X X . . . . . X . . . . . |
$$ | . . O . . . . . . X . . . . . . . . . |
$$ | . . . O . O . O . . . . . . . . . . . |
$$ | . . X , . . . . . X . . . . . , . . . |
$$ | . X . . X . O . . . . . . . . . X . . |
$$ | . X O O . . . . . . . . . . . . . . . |
$$ | . . X . . O . . . . . . . . . . X . . |
$$ | . . X . . . . . . . . . O . . . . . . |
$$ | . . O . . . . . . . . . . . X O . . . |
$$ | . . . O . . . . . , . . O . X , . . . |
$$ | . . . . . O 1 . 2 . . . . O . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]

A lot of interesting discussions, thanks

@Bill board points (correspond to few % changes)

@Uberdude

This is precisely what I was getting at. I would like to find an example, where the bot agrees with this heuristics. I notice that the bots disagree with it based on my games, now I am searching for pro games where pros play in this way(extend). So far I found two examples, in both cases Katago disagrees with the pros and prefers the hane. (I attached example 2, move 68, where the hane is 0.8pt better, but best move is tenuki at R10)

@John
Thanks for the post, I will try to understand it better latter. If I understand you, the point is to play safer, even if it is suboptimal? Or it is easier to play? Then there surely must be some example where hane is an overplay. Can you provide it?

I agree and I should think about tenuki more often. But in this example Katago again strongly prefers hane on top (1.6 pt, 10% better than the game move; 0.5pt better than extend).

We would need to know at least the playouts for your presented statistics but I guess here comes the point John Fairbairn made into play.

The bot with his superhuman abilities favours the hane - based on his neural net of self-play experiences. The human player argues (I guess) that the hane could be stronger but that it gets murky down the variation tree. And before he takes his opponent into the woods, where possibly only of them surfaces again... Why not play the checking extension?

The idea is that this checking extension is easier to calculate and easier to judge positionally. And if you can calculate and judge it to be enough, why push further and maybe offset your risk-reward-balance?

White won that game.

_________________
My "guide" to become stronger in Go

Now that I've looked up hiki (it's not a Japanese term I use), John's comments match my feeling that this example was not a good one if you are trying to find a game setting where a "hane vs extend against attachment" proverb might apply. I wouldn't even think of that proverb here because it is a dynamic position with moving groups, it's all about haengma in relation to the existing m14 and n13 stones and how black stops white from moving into the left area; can he cut?. I would also describe j14 as "pulling back" which is what hiki seems to mean. In this case the shape principle that would favour pull-back over hane would be the idea that after pull-back white has a cut at L13 he needs to worry about, and if he fixes (simply by playing l13 himself) that then black has sente, whereas if you hane, white pulls-back at l13, then black has a cut to fix at j14 but that's a sad direction because now white marches on into the left side with k11 jump, or maybe even j12 hane. There is some reading of variations involved, e.g. can white keep pushing at j13 himself or is that an overplay when black cuts at l13 (looks like the latter to me), or does white have a more efficient way to fix l13 than simply playing there, ideas like k11. But the idea against playing pull back is that it might be slack, can black reasonably cut right away? This looks like a case of atari from crosscut potentially being a good technique, though there's still quite some reading to do if it actually works, is this what KataGo recommends? If so it's rather disappointing that some 9ps didn't think enough to play something better a random 4d amateur did lying on a swing seat in his garden.

Click Here To Show Diagram Code

[go]$$W
$$ +---------------------------------------+
$$ | . . . . . . . . . . . . . O . . . . . |
$$ | . . . . . . . . . . . . O . O X . . . |
$$ | . . . O O O . . . . . . O . O O X . . |
$$ | . . O , X . . . . , . . O O X O X . . |
$$ | . . O X . . X . . . . . . X X X . . . |
$$ | . . X . . . . . . X 3 O . . . . X . . |
$$ | . . . X . . . . 4 1 2 . O . . O X . . |
$$ | . . . . . . . . . 5 6 . . . . O O X . |
$$ | . . . . . . . . . . . . . . . X O X . |
$$ | . . . , . . . . . , X . X X . X X O . |
$$ | . . X . . . . . . . . X . . X O O O . |
$$ | . . . . . . . . . . . X O O O . . . . |
$$ | . . . . . . . . X O O O X X O O . . . |
$$ | . . . . . . . . . X X X X . O X X . . |
$$ | . . X . . . . . . . . . . O . X O . . |
$$ | . . . , . . . . . O . O O X . X . . . |
$$ | . . . O . . O . . . . . . O . X . . . |
$$ | . . . . . . . . . . . . . O . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ +---------------------------------------+[/go]

That doesn't surprise me so much: one of the lessons I've picked up from bots is how much they dislike letting the opponent make gains in local efficiency of shapes in sente*, even if there is a directional benefit to it. And tenuki-ing from an attachment means they get to play first in a 1-vs-1 local sitaution so get a nice sente exchange, eg here hane on top and white pull back, and that's already a benefit without further continuation. The human pro probably thought that local loss was justified by the global gain of getting the lower side extension, but bots seem to think that to endure the local loss the global gain needs to be bigger than humans have tended to think.

* e.g. peep the cutting point and get a connect in reply against the wall of 3-3 invasion of 4-4 if you hane and extend.

I usually play the move and wait for at least 10k playouts on Katago (Zbaduk).

Katago's sequence:

Click Here To Show Diagram Code

[go]$$W
$$ +---------------------------------------+
$$ | . . . . . . . . . . . . . O . . . . . |
$$ | . . . . . . . . . . . . O . O X . . . |
$$ | . . . O O O . . . . . . O . O O X . . |
$$ | . . O , X . . . . , . . O O X O X . . |
$$ | . . O X . . X . . . . . . X X X . . . |
$$ | . . X . . . . . 4 X 3 O . . . . X . . |
$$ | . . . X . . . . 6 1 2 . O . . O X . . |
$$ | . . . . . . . . . 7 5 . . . . O O X . |
$$ | . . . . . . . . . . . . . . . X O X . |
$$ | . . . , . . . . . , X . X X . X X O . |
$$ | . . X . . . . . . . . X . . X O O O . |
$$ | . . . . . . . . . . . X O O O . . . . |
$$ | . . . . . . . . X O O O X X O O . . . |
$$ | . . . . . . . . . X X X X . O X X . . |
$$ | . . X . . . . . . . . . . O . X O . . |
$$ | . . . , . . . . . O . O O X . X . . . |
$$ | . . . O . . O . . . . . . O . X . . . |
$$ | . . . . . . . . . . . . . O . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ +---------------------------------------+[/go]

I think that I already understand this local sequence better.

My summary would be:
- attachment locally loses a bit, and playing hane is the way to exploit that
- hane is very rarely a bad move
- extend can simplify a game and does not lose much
- attach extend shape usually arises in different order: after a shoulder hit
- this might mean that shoulder hit is better locally than attach (but it is less "hot")
- this is all in vacuum and the differences are small, should think about it in context of haengma

Just a kyu observation.
Extend; a. when the attachment is from a weak stone. You do not want to leave some aji for the opponent to make sabaki; b. when you want sente. if the cut against the opponent attachment stone is severe, the opponent would have to go back and defend, and you have sente; c. when you have to go back and fix the cutting point after you hane, and, from a tewari perspective, your hane move ends up as a small move.
Hane; everything else