Some interesting discussion on Iyama's s12 peep, a common move from humans in this new joseki before sacrificing the 3 stones, which some versions of LeelaZero don't like
over on facebook.
Here's my attempts at figuring out why LeelaZero #198 doesn't like it. First its suggested line for both (move 1 with 35k playouts, 10k at end). Some comments:
- 2 is one choice, cut is another but -3% now
- 4 is a nice tesuji, white is reluctant to atari at m18 because then after defending at L16 n13 works*
- 6 after making white make an ugly connection of 5 black's got some stones in sente which will be useful in making a mess on the top side later (compare to game where Iyama got m18 and made a big territory there)
- 7 gets good value out of losing the 3 stones, white can grow the middle later if that becomes big
White starts and ends this sequence with 59% (white then approaches at q5, in mainline white makes a group there, black then makes a group on top side, using k18 sente to help).
- Click Here To Show Diagram Code
[go]$$Wc Dia 1
$$ +---------------------------------------+
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . O . . . |
$$ | . . . . . . . . . 4 1 X . X . O X O . |
$$ | . . . O . . . . . , 5 O 2 . X X O O . |
$$ | . . . . . . . . . . . 3 9 O 0 O X X . |
$$ | . . . . . . . . . . . . 8 O 6 O O X . |
$$ | . . . . . . . . . . . 7 X X . X X . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ +---------------------------------------+[/go]
* this is why n15 doesn't work without black tricking white into making black get a stone at n17:
If white peeps, the idea is white is going to sacrifice the 3 stones and then it's a good exchange as that one white stone could be helpful later in doing things on the right side (and it's not sente once 3 stones captured). However LZ 198 says it loses 10%, let's see what happens next.
- LZ says for 2 black won't n16 now (-7%) but will l16 crosscut instead
- white 3 at o16 is also good, but makes comparison harder to ignore that choice for now
- with 4 black trades, sacrificing the top for a ponnuki and sente
- the question is, is the marked exchange now good or bad?
White starts and ends this sequence with 49%.
- Click Here To Show Diagram Code
[go]$$Wc Dia 2
$$ +---------------------------------------+
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . 5 9 . . O . . . |
$$ | . . . . . . . . . . 1 8 3 X . O X O . |
$$ | . . . O . . . . . , 2 O 4 7 X X O O . |
$$ | . . . . . . . . . . . 6 . O . O X X . |
$$ | . . . . . . . . . . . . . O . O O X . |
$$ | . . . . . . . . . . . . X X . X X B . |
$$ | . . . . . . . . . . . . . . . . . W . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ +---------------------------------------+[/go]
Black could have chosen the crosscut if white didn't peep, but LZ thinks that's not so good. 2 lost 3%, so white ends this with 62%. The only difference between this diagram and the one above is the s12 s13 exchange.
- Click Here To Show Diagram Code
[go]$$Wc Dia 3
$$ +---------------------------------------+
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . 5 9 . . O . . . |
$$ | . . . . . . . . . . 1 8 3 X . O X O . |
$$ | . . . O . . . . . , 2 O 4 7 X X O O . |
$$ | . . . . . . . . . . . 6 . O . O X X a |
$$ | . . . . . . . . . . . . . O . O O X . |
$$ | . . . . . . . . . . . . X X . X X . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ +---------------------------------------+[/go]
Is it really so bad? What's the difference?
- If the right side becomes black territory white lost 2 points vs playing endgame hane at
a- If white invades right side later then maybe she would peep at s12 anyway, but maybe not. You'd rather not pre-lose the points.
- Adds more liberties to q13 stones. It's kinda unlikely as terrible inside for outside exchanges, but say white wedged at p13 to create 2 cutting points, s12 actually means black can more easily answer on the outside with 4 and safely answer 5 at 6 as white can't s13 cut to capture the s14 stones (but white gets endgame squeeze, so black might actually atari under anyawy, depends on what's down the side, I'm just focusing on the local connectivity), without s12 maybe 4 would be at q11 and then there's a cutting point at o12 (probably used to peep as part of a reduction rather than to cut).
- Click Here To Show Diagram Code
[go]$$Wc Dia 4
$$ +---------------------------------------+
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . O O . . O . . . |
$$ | . . . . . . . . . . O X O X . O X O . |
$$ | . . . O . . . . . , X . X O X X O O . |
$$ | . . . . . . . . . . . X . O . O X X . |
$$ | . . . . . . . . . . . . . O 3 O O X . |
$$ | . . . . . . . . . . . . X X 1 X X X . |
$$ | . . . . . . . . . . . . . . 2 5 . O . |
$$ | . . . . . . . . . . . . . 4 . 6 . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ +---------------------------------------+[/go]
LZ 198 says yes, its -13% bad exchange, quite significant. You can even make the s12 exchange after the crosscut and ponnuki trade variation and see the same winrate drop. I guess it's mostly the chance of it ending up as mochikomi that causes the dislike. I like how LZ changed its answer to white l17 hane to turn s12 into a bad move instead of a good move, a common technique seen in high level probing play.
Edit: to further isolate the mochikomi aspect of the s12 s13 exchange I added some outside exchanges to make it more likely that right area will become a black territory so that the white peep won't be useful in a future invasion there (note this part of analysis is with <1k playouts). Starting from diagram 3 (black crosccuts for ponnuki, white hasn't peeped) it is black's move and he is at 38%. If I pass for black and allow white to reduce at 1 and have black answer at 2 and then pass for white to make it black to move again we now have black at 30%. So LZ thinks this is a good exchange for white (I thought it probably would be as black's answer seems soft, but I wasn't sure before checking with LZ as there is a downside of solidifying black). Note that neither of these moves are the best if they had a free choice, LZ 198 says approach (for 1) or enclose (for 2) at q5 is best. The purpose is to create a somewhat realistic whole board position in which LZ will think the right side is black territory (and also means p13 no longer creates 2 cutting points). o11 to q8 is a bad shape so I add another reducing exchange of 3-4 that's also slightly helping white: black is now down to 27-28% (LZ thinks black's best next move is either q5: very normal, or interestingly n5: weird shape to be 5th line, n4 is what I would expect if trying to be more ambitious than solid shimari). Then to try to get black's winrate back up closer to the 38% original I make the peep-connect exchange of 5-6 in reverse, in hypothetical play probably peep of 6 would come before defending with 4, though really if black attempted to peep there first LZ says white should just ignore and q5 approach. This makes white heavier and gets black an outside stone so now his winrate is up to 31% (with suggested next move at q5 again).
- Click Here To Show Diagram Code
[go]$$Wc Dia 5
$$ +---------------------------------------+
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . O O . . O . . . |
$$ | . . . . . . . . . . O X O X . O X O . |
$$ | . . . O . . . . . , X . X O X X O O . |
$$ | . . . . . . . . . . . X . O . O X X . |
$$ | . . . . . . . . . . . . . O . O O X . |
$$ | . . . . . . . . . . . . X X . X X . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . 1 . 2 . . . . . |
$$ | . . . , . . . . . , 6 5 . . . , . . . |
$$ | . . . . . . . . . . . 3 . 4 . . . . . |
$$ | . . . . . . . . . . . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ +---------------------------------------+[/go]
So with the setting prepared, what difference to winrate does it make if we make the s12 s13 exchange? Answer: black is up to 53% (again q5 next move), so 22% mistake now instead of 13% when it's more even likely to be mochikomi. Having said that, LZ as white still wants to tease black's right side: after q5 enclose it wants to r7 invade and then just leave it for aji.
- Click Here To Show Diagram Code
[go]$$Wc Dia 6 Black's move and ends at 53%
$$ +---------------------------------------+
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . O O . . O . . . |
$$ | . . . . . . . . . . O X O X . O X O . |
$$ | . . . O . . . . . , X . X O X X O O . |
$$ | . . . . . . . . . . . X . O . O X X . |
$$ | . . . . . . . . . . . . . O . O O X . |
$$ | . . . . . . . . . . . . X X . X X 2 . |
$$ | . . . . . . . . . . . . . . . . . 1 . |
$$ | . . . . . . . . . . . O . X . . . . . |
$$ | . . . , . . . . . , X O . . . , . . . |
$$ | . . . . . . . . . . . O . X . . . . . |
$$ | . . . . . . . . . . . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ +---------------------------------------+[/go]
For completeness, what if white makes the endgame exchange? This is premature (losing a ko threat and the peep aji if an invasion did still happen) so loses some winrate for white, but isn't as bad as giving black 2 points and black ends at 42%, so white lost 11%. (Black's next move either q5 or n4 or n5 or m6 or m5 (#1 after a few k playouts)).
- Click Here To Show Diagram Code
[go]$$Wc Dia 7 Black's move and ends at 42%
$$ +---------------------------------------+
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . O O . . O . . . |
$$ | . . . . . . . . . . O X O X . O X O . |
$$ | . . . O . . . . . , X . X O X X O O . |
$$ | . . . . . . . . . . . X . O . O X X 1 |
$$ | . . . . . . . . . . . . . O . O O X . |
$$ | . . . . . . . . . . . . X X . X X 2 . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . O . X . . . . . |
$$ | . . . , . . . . . , X O . . . , . . . |
$$ | . . . . . . . . . . . O . X . . . . . |
$$ | . . . . . . . . . . . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ +---------------------------------------+[/go]