Life In 19x19

While reviewing one of my games I came to the conclusion this move is not joseki:

Click Here To Show Diagram Code

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

As always consider tenuki, but if you look for the best local move this is much better usually:

Click Here To Show Diagram Code

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

The bad move is played a lot by pros. As always I challenge you to show me games were the critisized move is indeed the better option. Have fun!

Checking Waltheri, Takemiya and Seo Pong-su have each played in this position.

Click Here To Show Diagram Code

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

Hi Bill/Gomoto,

How about Go Seigen ?

EdLee wrote:Hi Bill/Gomoto,

How about Go Seigen ?

No record, as far as I know.

Funny, I don't fancy myself better than pros, but the "right" move seems like the obvious one.

Gomoto: Agreed that the low crawl is normal among pros, but your statement that it is 200:1 invites query. That implies you have at least 201 pro games with this pattern. Do you really? GoGoD has only 61 (of which 3 are the bend on top).

Ed: Yes, GSG played the low crawl (1933-02-21, vs Kitani).

Hi John,

Thanks. The low crawl felt like Kitani-style.
Too bad Go Seigen didn't foresee the engine move in this case.

(Thanks John, I made a mistake about the 200:1 ratio.)

Picking up where I left off (I had to go out this afternoon):

Click Here To Show Diagram Code

[go]$$W
$$ --------------------+
$$ . . . . . . . . . . |
$$ . . . . . . O . O . |
$$ . . j i O . X O . . |
$$ , . . h . . X . . . |
$$ . . . . . . . X . . |
$$ . . . . . X X O X . |
$$ . . . . . a O O b . |
$$ . . . . . . . . . . |
$$ . . . . . . d f . . |
$$ , . . . . . e g . . |[/go]

I found the Go Seigen connection interesting, and so ran LZ for 15 minutes on the relevant position. LZ pondered NINE moves in this area (and three elsewhere). 'a' did indeed come out on top but actually with a lower score than other moves with marginally fewer visits. In fact, FIVE moves showed a higher score and for a long time 'd' (the highest scorer) was actually "best". Further, Go's move 'b' was one of the candidates and was bottom, but within 2.5 percentage points of the "best" move. I believe someone (?Bill) said on this forum that any difference up to about 3 percentage points can be disregarded, but in this case we must also remember that it was a no-komi game and Go was Black, so a base-emphasising or territory-emphasising move can be more easily justified.

So, as far as I'm concerned Go himself has not been shot down by this exercise, but actually I'm not trying to make any such points: this is just the groundwork for some questions.

1. Is the 3% range sensible - at least to the extent that it's presumably better to play a move you understand rather than a higher scoring but incomprehensible love?

2. There were times when LZ was showing the best (blue) move 'a' with a lower score AND a lower number of visits than 'd'. Is that a possible bug?

3. I have noticed often before that LZ likes to choose what we might call common shape moves such as 'a' as its first candidates (just as we humans might do, of course). Being so common these shape moves are presumably encountered an uncommon number of times when LZ is training itself. It therefore must take extra account of them, at every stage in the tree. Is it possible that this process introduces blind spots in LZ, in a way similar to what happens in humans?

Edit: diagram corrected

While I think the position Bill and John showed is very interesting, I proposed this starting position:

Click Here To Show Diagram Code

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

In this case we are talking about >10% difference in favour for a.

John Fairbairn wrote:Picking up where I left off (I had to go out this afternoon):

$$W
$$ --------------------+
$$ . . . . . . . . . . |
$$ . . . . . . O . . . |
$$ . . j i O . X O . . |
$$ , . . h . . X . . . |
$$ . . . . . . . X . . |
$$ . . . . . X X O b . |
$$ . . . . . a O O . . |
$$ . . . . . . . . . . |
$$ . . . . . . d f . . |
$$ , . . . . . e g . . |

Click Here To Show Diagram Code

[go]$$W
$$ --------------------+
$$ . . . . . . . . . . |
$$ . . . . . . O . . . |
$$ . . j i O . X O . . |
$$ , . . h . . X . . . |
$$ . . . . . . . X . . |
$$ . . . . . X X O b . |
$$ . . . . . a O O . . |
$$ . . . . . . . . . . |
$$ . . . . . . d f . . |
$$ , . . . . . e g . . |[/go]

I found the Go Seigen connection interesting, and so ran LZ for 15 minutes on the relevant position. LZ pondered NINE moves in this area (and three elsewhere). 'a' did indeed come out on top but actually with a lower score than other moves with marginally fewer visits. In fact, FIVE moves showed a higher score and for a long time 'd' (the highest scorer) was actually "best". Further, Go's move 'b' was one of the candidates and was bottom, but within 2.5 percentage points of the "best" move. I believe someone (?Bill) said on this forum that any difference up to about 3 percentage points can be disregarded, but in this case we must also remember that it was a no-komi game and Go was Black, so a base-emphasising or territory-emphasising move can be more easily justified.

So, as far as I'm concerned Go himself has not been shot down by this exercise, but actually I'm not trying to make any such points: this is just the groundwork for some questions.

1. Is the 3% range sensible - at least to the extent that it's presumably better to play a move you understand rather than a higher scoring but incomprehensible love?

Anything I say about this must be taken as tentative. There is not a lot of research about the margin of error of bots' winrate estimates. But I have compared the estimates of Leela Zero at the 200k setting versus the estimates at the 100k setting and found more than occasional differences of 3% or more. A good study would require looking at a few hundred games at least, and, as Uberdude suggested, looking at the variability of estimates at the same setting, as well. (IOW, I ain't gonna do it. )

3. I have noticed often before that LZ likes to choose what we might call common shape moves such as 'a' as its first candidates (just as we humans might do, of course). Being so common these shape moves are presumably encountered an uncommon number of times when LZ is training itself. It therefore must take extra account of them, at every stage in the tree. Is it possible that this process introduces blind spots in LZ, in a way similar to what happens in humans?

I don't think so. You can't ignore the obvious or the defaults. And training refines these judgements. To take a human example (myself) hane at the head of two stones is the default to consider, but I have learned to recognize some situations where my stones are too weak or the opponent's stones are too strong for that to be effective, and so I do not spend any time consciously reading the hane out. OC, I will sometimes be mistaken, but I can spend time exploring other alternatives.

Here is one way I think that blind spots may arise. IIUC, LZ leans toward assuming that ladders work. In self play, it may then avoid a variation that leads to a ladder, even though the ladder does not work. In that case, it never discovers that that ladder does not work, and so continues to assume that ladders work. Or, playing one side, it creates a ladder that actually does not work, but playing the other side, it also assumes that the ladder works, and so does not try to escape. The shared assumption is never tested.

Gomoto: I copied Bill's diagram without noticing it was different, but the rest of my comments stand (and the GSG game was with your variation). Sorry.

A related shape you may find interesting to study and compare with: Kim Jiseok recently played a pincer (not so common these days!) and following double approach white had the choice to play hane from 3-3 stone but played hane from the 3 outside forcing black to make the rather clumsy empty triangle that I thought people generally didn't like so much (3-3 still has aji and black's shape can end up short of liberties, particularly if white pushes on the outside and black ignores so white can hane). I've seen this new variation in a few pro games recently.

I like white's result in that local variation, Uberdude.

I guess Kim Jiseok made things work, but how much of that is due to the joseki, and how much of that is due to being Kim Jiseok?