I posted a while ago about some komoku, high approach, low pincer patterns. I thought I'd do another deep dive and share it here again for a feeling of accountability.

Again, I'm aiming at:

• Having fun trying to understand go better
• Improve my instinct in the opening
• Perhaps pick up some good habits (shape, fighting, positional judgement)

I'm investigating and writing primarily for my own benefit, but I'd be thrilled if anyone could offer suggestions, corrections, thoughts, or even if people just read and take something from it.
I'm going to look at the fairly common position:

Click Here To Show Diagram Code

[go]$$B
$$ | . . . . . . . . . .
$$ | . . . . . . . . . .
$$ | . . . . . . . . . .
$$ | . . . 1 . . . . . .
$$ | . . . . . 2 . 3 . .
$$ | . . . . . . . . . .
$$ | . . . . . . . . . .
$$ --------------------[/go]

I will stick to trying to consider some of these questions:

• Black 3
  
  • When do professionals play black 3
  • When do AI (mainly ELFv1 unless i find a new toy) play black 3
  • Meaning in relation to other pincers
  • Meaning in relation to other choices
• What happens next
  
  • What happens next in joseki books
  • what happens next in pro play
  • What happens next in AI play
  • What happens in special cases
  • What happens next in weak play
  • Can/should deviations be punished

But I've been looking at batch analysis of patterns arising in pro games, so maybe it will be centered around this.

I did some automated analysis with Elfv1. I think the whole board position is important, so I wanted to look at the patterns in the context in which they happen.

I sketched out the joseki I knew and a few more from books allong with a couple of new patterns and a couple of things i was curious about in an sgf. I created some scripts that do a pattern search to find board positions where these patterns get played and run them through Elf. All games used are 9p vs 9p games. It then taggs the sgf with some summary stats about what Elf thinks about the move accross the range of board positions.



I classified things as a mistake only if Elf thinks the win rate drops by more than 3%. iqr is the inter-quartile range ie. the central 50% of the data. Since I filtered down to 9p vs 9p games only, some of the patterns don't have so much data.
I was planning to use this as a starting point and have a closser look at the patterns where Elf has strong opinions.

There's one modern trend here that is very stark and certain and I don't think I really feel i properly understand. Maybe a stronger player could help me see?

In this shape:

Click Here To Show Diagram Code

[go]$$B
$$ | . . . . . . . . . .
$$ | . . . . . . . . . .
$$ | . . . 7 . . . . . .
$$ | . . 6 1 . . . . . .
$$ | . . 4 5 a 2 . 3 . .
$$ | . . 0 8 9 b . . . .
$$ | . . . . . . . . . .
$$ --------------------[/go]

a has disappeared and b has replaced it.
It seems that Alphago did this in one of the master online games, and I remember it being in the teaching tool, and it's taken off since then. I still expected that it might depend on the whole board and that a might still be appropriate, but it looks like Elf1 thinks that a is always a mistake of at least around 5% (usually about 9%). b is apparently never a mistake in my sample of positions. Pros seem to agree - in my filtered down 9p vs 9p search the pattern with b appears only 7 times before 2017 (starting in 1961 with Go Seigen of course) but then appears 25 times in 2018 alone, whilst the pattern with a is down to only 3 times in the year.

I understand that playing at a leaves white with the move at b to start an attack on the wall, but is that really so bad when the pincer is tight? if it's dangerous then black can connect on top right? looking at the Elfs projected sequences, the AI seems to cash in on the aji fairly soon by just playing a move on the lower side, but surely you concede a lot of forcing moves on the lower side with b too? It seems that both choices leave behind aji/forcing moves, and in any choice like this there is going to be a trade off, but does anyone feel they have a good handle on why humans used to unanimously think that a offered the better deal and now think b does?

It always struck me that playing at 'a' looked like particularly bad shape when 'b' was available. It might be as simple as not making the empty triangle being a more efficient way to do the same thing (maintaining/increasing liberties for example).

My slightly more in depth undertanding of it is this, presuming some other stuff happens elsewhere on the board white might have the chance to play like this, black has to play 3 to separate the white groups and then comes under attack:

Click Here To Show Diagram Code

[go]$$B
$$ | . . O . . . . . .
$$ | . . . X . . . . .
$$ | . . O X . 4 . . .
$$ | . . O X 1 O . X .
$$ | . . O O X 2 . . .
$$ | . . . . 3 . . . .
$$ +------------------[/go]

By comparison, if white tries to cut black in this variation, black's outside shape remains solid and white has to have significantly more support to attempt the cut. That's not to say it can't happen but it's harder for white to attack:

Click Here To Show Diagram Code

[go]$$B
$$ | . . O . . . . . .
$$ | . . . X . . . . .
$$ | . . O X . . . . .
$$ | . . O X 2 O 3 X .
$$ | . . O O X 1 . . .
$$ | . . . . . . . . .
$$ +------------------[/go]

In addition, if black tenukis before connecting at all black can even sacrifice on the 2nd line to maintain good shape for relatively little gain on white's part:

Click Here To Show Diagram Code

[go]$$W
$$ | . . O . . . . . .
$$ | . . . X . . 8 . .
$$ | . . O X . 6 . . .
$$ | . . O X 1 O 4 X .
$$ | . . O O X 2 3 . .
$$ | . . . . 7 5 . . .
$$ +------------------[/go]

Alternatively white could descend looking for fewer points but keeping holes in black's shape, but black could easily defend at a, b or c.

Click Here To Show Diagram Code

[go]$$W
$$ | . . O . . . . . .
$$ | . . . X . a . . .
$$ | . . O X . b c . .
$$ | . . O X . O . X .
$$ | . . O O X 1 . . .
$$ | . . . . . . . . .
$$ +------------------[/go]

At this point if white playing to cut off the black stone on the 2nd line has become that good a move, either you have reached the early endgame (obviously it's still a huge move), or the black stones on the outside have already been neglected too much.

I think just looking at the shapes that arise from white attempting to cut through black in either variation it should be relatively obvious which is better locally. The only hesitation I have is that if black has come under heavy attack already it may be easier to defend like this if black had already connected solidly instead, starting to make eyeshape and getting out at a/b/c/d/e:

Click Here To Show Diagram Code

[go]$$W
$$ | . . O . . e . . .
$$ | . . . X . . . d .
$$ | . . O X a 2 c . .
$$ | . . O X X O b X .
$$ | . . O O X 1 . . .
$$ | . . . . 3 . . . .
$$ +------------------[/go]

There may also be less aji in the corner for white when black doesn't connect solidly and likewise the 1st line endgame may end up being better for white, later on moves at 'a' or 'b' might be sente for white:

Click Here To Show Diagram Code

[go]$$W
$$ | . . O . . . . . .
$$ | . . . X . . . . .
$$ | . . O X . . . . .
$$ | . . O X . O . X .
$$ | . . O O X X . . .
$$ | . . . a b . . . .
$$ +------------------[/go]

That's my two cents anyway! I think it's probably one of this instances where connecting underneath is locally better but it can also be highly situational

Even more surprisingly, when this shape is played without black's pincer present on the board, b is preferred to a nowadays, I have seen something like this:

Click Here To Show Diagram Code

[go]$$B
$$ | . . . . . . . . . .
$$ | . . . . . . . . . .
$$ | . . . X . . . . . .
$$ | . . O X . . . . . .
$$ | . . O X 5 O 2 . 3 .
$$ | . . O O X 1 4 . . .
$$ | . . . . . . . . . .
$$ --------------------[/go]

Presumably, the idea is to make white heavier before connecting.

(Most recently, this occurred in the Sumire (soon to be 1p) game vs Choi 9p.

_________________
Sorin - 361points.com

The main thing I notice when comparing these diagrams (besides the fact that White has one more move coming to them in the second one) is that White has six liberties in the first and two in the second.

That's why black would not descend in the first diagram quoted above, but would play hane with 3 in the following diagram, at least that would be my default answer locally:

Click Here To Show Diagram Code

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

Of course, the descend is a viable move too, depending on the overall position.

_________________
Sorin - 361points.com

Thanks for answers struttnoddy, sorin, dfan.
I always forget about endgame considerations - I guess if we assume that there is territory on both sides (certainly doesn't seem a given for black) then I think b is two points better in swing value (the point at a + going from sente endgame for white under a, to double gotte under b. My technical endgame is terrible, so please correct me if this is wrong)

When thinking about the aji left if black plays a I wasn't really thinking about playing b directly because I thought white would rather keep good endgame and can cash in on the aji less directly. However I looked at some of the follow ups played by Elfv2 and it looks as though playing b is often the right move. (the hane underneath does seem to be better than the descent in all the games i analysed)

I took the batch of games that were analysed at the start, I cropped the sgfs at the choice between a and b and i used twogtp to get elfv2 to play out the rest of the games under each choice to see how the aji was used. I captured some of the interesting or recurring sequences in this sgf. It suffers badly from lacking the whole board context that guides the sequences!


a lot of the same sequences seem to be played either way, but maybe black can be a little bit bolder when connected underneath - if nothing else then it's still possible to sacrifice something.

I thought a fairly big difference is that this move seems much less intimidating if you're connected underneath:

Click Here To Show Diagram Code

[go]$$W
$$ | . . d . c . . . .
$$ | . . O 1 . . . . .
$$ | . . . X . . . . .
$$ | . . O X . . . . .
$$ | . . O X . O . X .
$$ | . . O O X X . . .
$$ | . . . . . . . . .
$$ +------------------[/go]

In the games where black chose the other move, black would often find time to play 'c', which is a nice move to play as it has nice follow up at d (if white is feeling slow, white might even protect at d) but perhaps it's more of a protection from white taking the nice forcing sequence on that side.

I would definitely agree the hane seems best in this situation, but I think the point stands that if white has any strength nearby black is likely to come under attack after white descends, hane or not. Black probably does have more eye shape and liberties with the hane, but the situation could have been avoided entirely by playing the other variation, albeit opening the door to a whole different set of possibilities!

Additionally if black hanes it removes something of the corner aji by giving white the decent at 'a' in sente later, which if black has sealed the white corner group in could be very useful:

Click Here To Show Diagram Code

[go]$$W
$$ | . . . . . . . . .
$$ | . . Y . . . . . .
$$ | . . Y . Y . . . .
$$ | . Y O . . . . . .
$$ | . Y O X . . . . .
$$ | . O O X . . . . .
$$ | . . O X X O . X .
$$ | . . O O X 1 . . .
$$ | . . . a . 2 . . .
$$ +------------------[/go]

This might not be unrealistic if the black group had to run up the board somewhat but would of course be a minor loss when compared to saving the outside black group, far be it from me to disagree with Elf!

Elf doesn't often like to block the other way (when 9 dan pros play this, it's a mistake of over 3% about 1/3 of the time) but when you do block the other way Elf thinks that the sequence to play is almost always:

Click Here To Show Diagram Code

[go]$$B
$$ +------------------
$$ . . . . . . . . . |
$$ . . . . . . 4 . . |
$$ . . . . . 1 O . . |
$$ . . . . . X 2 . . |
$$ . . . . . 3 . . . |
$$ . . . . . 5 O a . |
$$ . . . . . 7 6 . . |
$$ . . . . . . X 8 . |
$$ . . . . . . . 9 . |
$$ . . . . . . . . . |[/go]

I think humans had realised that playing 4 at a was probably not good, but 4 at 5 seemed to be popular. humans also used to play 6 at a. I think master played this 6 and it seems that Elf found it too.

if you do play 4 at 5, then Elf likes this sequence:

Click Here To Show Diagram Code

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

Which has only happened very rarely in pro games. I always thought that one of the advantages of 1 was that if black plays the more usual sequence (4 at a) then white is completely alive in the corner (with no aji) and so white can play more freely on the top side. This 4 seems to keep the whole white group under pressure.

http://www.alphago-games.com/view/event ... 31/move/18

P.S. The second sequence crops up in the AlphaGo teaching tool too.

Humans had already found . I like it, but I did not find it myself. However, I would generally play at "a" or the solid connection, because I do not like giving Black the two step hane.



Yes, very nice. Thanks.

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

Visualize whirled peas.

Everything with love. Stay safe.

I'd be interested as to why pros seem to dislike black 4 above, a quick search of josekipedia gives these result as (presumably locally) "good for white":

Click Here To Show Diagram Code

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

Click Here To Show Diagram Code

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

My guess is that black has made white strong on the outside for relatively little profit in the corner, needing 3 moves to capture two stones. Given these sequences I'm not overly sure that the hane at 4 is designed to keep pressure on the white group, there's plenty of aji but it's no where close to death locally. Of course ELF may think that actually the outer white group is very attackable but I can't (locally) see how with my measly 1d logic

What are ELF's follow ups to the black hane at 4?

I should say that this is a little bit dependant on the board position and sometimes Elf does prefer the human cut. Also the idea that white is kept under pressure is my 2kyu attempt to interpret what's going on.

I did some twogtp play-outs for a range of board positions and Elf's follow up is:

Click Here To Show Diagram Code

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

but then white seems never to play at 'a'. 'b' and 'c' are the moves that appear more than once, but it gets quite board specific. Maybe Elf thinks white benefits from strengthening with one more move and maybe considers 'a' a bit too much.
Although I'd agree that white isn't weak exactly, from my playing around with AI I feel like they place large value on having groups completely settled without aji, and white isn't there yet so black can feel a little happier for that?

Elf really dislikes jumping out a lot of the time. It's a mistake (>3%) on 58% of occasions when it was played by 9d pros. Elfs response is a close call between 'a' and 'b':

Click Here To Show Diagram Code

[go]$$W
$$ --------------------+
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |
$$ , . . b a . X . . . |
$$ . . . . . . . . . . |
$$ . . . . . 1 . O . . |
$$ . . . . . . . . . . |
$$ . . . . . . . X . . |
$$ . . . . . . . . . . |
$$ . . . . . . , . . . |
$$ . . . . . . . . . . |[/go]

'a' is the common human move. It was probably among the first 'joseki' moves I learned. Elf broadly seems to think that it's fine - it's rarely much of a mistake at all.
'b' is one that personally I wasn't aware of at all before AI (I don't have my game database to hand to check if go seigen played it 80 years ago).

When I've briefly looked at AI playouts or pro games where it features, I haven't really got a flavour of the different features of 'b' compared to 'a'. Do any stronger players have any ideas of what the positives and negatives are with this choice?

I don't pretend to have an answer to your original question (besides the intuition that bots may prefer the looser extension instead of the human one space jump for efficiency reasons, just like they prefer looser corner enclosures while humans used to play tighter ones).

However I find this position very interesting for the following reason: humans tend to follow up with "a" in the following diagram (giving black the choice of either peacefully living on the right side while letting white develop thickness in the center, or fight by pushing on the 4th line and cut), while AlphaGo and LeelaZero strongly prefer white "b" instead, forcing a fight. Maybe bots prefer that since it allows fewer options to the opponent?

This seems to be another blind spot in human positional evaluation.

Click Here To Show Diagram Code

[go]$$W
$$ --------------------+
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |
$$ , . . . X . X . . . |
$$ . . . . . . . . . . |
$$ . . . . . O . O . . |
$$ . . . . . . . . . . |
$$ . . . . . . . X . . |
$$ . . . . . . a . . . |
$$ . . . . . . , b . . |
$$ . . . . . . . . . . |[/go]

_________________
Sorin - 361points.com

Human blind spot? Do top bots have a decided preference one way or another? I.e., > 3% for LZ, >5% for Elf?

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

Visualize whirled peas.

Everything with love. Stay safe.

By "human blind spot" I meant pros only seem to play the counter-attack at "b" in special circumstances (in the pre-2016 games), and they seem to think "a" is the natural move; while bots seems to think "b" is the natural move.

The win-rate deltas are pretty small, see details embedded in the screenshots below; for LeelaZero the number of visits for "b" (which is what decides which candidate move is chosen) is overwhelming, compared to "a", in both cases I looked at.
AlphaZero gives "b" 2% more compared to "a", but it doesn't publish the number of visits; and it only has one of the two positions I looked at in their online tool database.

_________________
Sorin - 361points.com

Thanks. I wonder if both human and bot preferences are the result of path dependency, of different sorts: historical for humans, computational for bots. IIRC, in chess Emmanuel Lasker said if you find a good move, look for a better one. MCTS bots, it seems, don't do that so much.

Playing around with Deep Leela, when I compare plays I try to make the number of playouts more equal by forcing the program to make each play, rather than comparing the original winrate estimates. I played out the main variations on AlphaGo Teach until tenuki (a 3-3 corner invasion in both cases) and had Deep Leela make the same plays. Both programs preferred the pincer, Deep Leela by 2%, AlphaGo by 1% (down from an original 2% preference). Even if the bots started out trying both plays fairly equally, the results could favor the pincer to a moderate depth, even if only by a small amount, and it would make sense not to waste time searching the variations starting with "a".

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

Visualize whirled peas.

Everything with love. Stay safe.

Thanks Sorin, Bill Spight for your thoughts.


Are you referring to the original alphago teaching tool? I was under the impression that the teaching tool was based on AlphaGo Master ie. somewhere around the version that beat Ke Jie. I'd love to be proven wrong, but I didn't think that we had any resources based on AlphaGo Zero or AlphaZero except for the bot vs bot games that they published? (and of course the papers, leading to all these brilliant bots we now have!)

If we are comparing AG master vs Elf/LZ then it's an interesting comparison. Elf certainly seems to be more opinionated than Master. I'm perfectly happy to be challenged on this, but I think that the zero method, with it's emphasis on building a good network rather than mcts, means that I trust it more in the opening. I think that while Master on Google's hardware would have no problem beating Elf/lz on sensible pc hardware, I might actually trust the zero bots more on opening theory. I would trust Master more on reading, ladders/blindspots, life and death, and I accept that the lines between these are always blurred. Elf/LZ also offer the benefit of being able to probe their choices, which helps to investigate, as Bill points out, and make sure the bot isn't missing something important further down the tree. This might be another discussion though.



My preferred metric is to set the non-trivial mistake threshold at 3% and look at a range of pro board positions and see how frequently the move is a mistake (above this threshold.)

Click Here To Show Diagram Code

[go]$$W
$$ --------------------+
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |
$$ , . . . X . X . . . |
$$ . . . . . . . . . . |
$$ . . . . . O . O . . |
$$ . . . . . . . . . . |
$$ . . . . . . . X . . |
$$ . . . . . . a . . . |
$$ . . . . . . , b . . |
$$ . . . . . . . . . . |[/go]

From the sgf at the top of the thread (according to Elfv1):
'a':
52 out of 108 times this was a mistake(48.0%) (median delta is -2.88, half of deltas in the range -4.68 to -1.44)
'b':
there is less data - fewer 9d vs 9d games include this move
0 out of 10 times this was a mistake(0.0%) (median delta is -1.04, half of deltas in the range -1.39 to -0.49)



It's tough to be conclusive with only 10 board positions for b, but it looks to me like Elfv1 has a systematic preference for the counter pincer.