Interesting top policy move

lightvector · #1

I found a position where KataGo's favorite move is an interesting move, and one that I find particularly interesting because this move is also its top raw policy move by a pretty wide margin. (Raw policy = no explicit search, the move it would choose "intuitively" with only 1 playout).

White to play.

Click Here To Show Diagram Code: [go]$$Wc White to play $$ --------------------------------------- $$ | . . . . . . . . . . . . O . . . . . . | $$ | . . O . . . . . . . . O . O . . . . . | $$ | . X X O O . . . . . . O O X O O O . . | $$ | . . . X . . . . . , X O X X X X X . . | $$ | . . . X O O . . . . . X O . . . . . . | $$ | . . . . X O . . . . . . O . X . . . . | $$ | . . . . X O . . . . O . . O X . . . . | $$ | . . . X . . . . . . . . . . X . . . . | $$ | . . . . . . . . . . . . . X . O . . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . X . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . X . X . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . O O . . . . . . . . . . . . . | $$ | . O X . O X . O . . . . O . . O O . . | $$ | . X O O X O . . . , . . . . . , X . . | $$ | . X X X X O . . . . . . . X . X . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

Edited to add the answer, which various people have figured out at this point:

Shaddy · #2

Uberdude · #3

My policy network:

Gomoto · #4

ez4u · #5

Uberdude wrote:

My policy network:

I guessed (not reasoned) the wedge thinking, "What could be wacky enough to brighten lightvector's eyes?" I saw something similar on a Japanese TV show some years ago. The idea of probing with a placement in the middle of a one-space jump goes back to the 18th century at least. The pattern that I saw on TV first appears in GoGoD here.

Click Here To Show Diagram Code: [go]$$Wc Inoue Shunseki (W) - Honinbo Satsugen (1767-04-25a.sgf) $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . O . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , X . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . X X X . . | $$ | . . . . . . . . . . . . . O X O . . . | $$ | . . . , . . . . . , . . . O . , . . . | $$ | . . . . . . . . . . . . . O X X O . . | $$ | . . . . . . . . . . . . O X . X . . . | $$ | . . 2 . . . . . . . . . . O . . . . . | $$ | . . . 1 . . . . . . . . . O . X . . . | $$ | . . X O . 6 . 4 . 5 . . . . . . X . . | $$ | . . X 7 . . 9 . . 0 . . . . O , . . . | $$ | . . 8 . O . . X . 3 . . . . . O . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

Interestingly enough it occurred after White uses the press in response to the 2-space low pincer - AI Style

Black probes White in response to White's peep. This probe is top in katago's policy in this game as well.

Attachment:

Playing into the one space jump policy.jpg [ 152.71 KiB | Viewed 8490 times ]

Don't start thinking this is the solution to all your indecisive moments however. Katago was much less enthusiastic about Go Seigen's choice against Takagawa in 1955 and even less so about Yang Kaiwen's choice against Chen Xian last year in the A League.

Attachment:

Takagawa - Go 1955.jpg [ 131.08 KiB | Viewed 8490 times ]

Attachment:

Yang Kaiwen - Chen Xian 2019.jpg [ 126.57 KiB | Viewed 8490 times ]

Bill Spight · #6

Gomoto wrote:

Uberdude · #7

Whilst Dave's first example from the 18th century is a nice other example of wedging a one-point jump as a probe, it has quite a different flavour to the one here. The shape nugget is essentially as below, I've drawn it near the edge as that's where it often it but it could apply further into the centre too, or even (rarely) not perpendicular to the edge of the board. The logic is white doesn't want to simply connect at a in gote, but wedges because then if black ataris from b side THEN you connect at a and have created a big followup at c for later to cut that one-point jump if black continues with whatever operations the peep were about to the left (which black wanted to be sente). And if black ataris from the right with c then white pulls back to b and doesn't need to inefficiently connect at a, also giving black 2 cutting points. For the tesuji of :w2:

to be valid you also need to have read a good outcome if black simply ignores it and pushes at a.

Click Here To Show Diagram Code: [go]$$B $$ . . . . . $$ . O . X . $$ 1 a b 2 c $$ . O . X . $$ . . . . . $$ . . . . . $$ ---------[/go]

This example from lightvector, and the Lee-Gu game, have a different flavour: their shape nugget is simply a wedge without friendly stones on the 2 adjacent lines but further away and you have the flexibility to not mind which one gets stronger and which weaker, in fact in both cases here the friendly stones are a 2 point jump on either side of the wedge. Does this shape deserve a name? How about "proton torpedo"?

Click Here To Show Diagram Code: [go]$$W $$ ? . . X . . ? $$ O . , 1 . . O $$ ? . . X . . ? $$ ? . . . . . ? $$ ? . . . . . ? $$ -------------[/go]

I remember a commentary, probably in Go World, about a pro (O Meien/Rissue iirc) wedging a 3-4 5-4 shimari winning some "pro move of the year" award. That didn't have the friendly stone on one side, but had the same flavour of creating miai: either black atarid from inside and white created 2 cuts so the wedge was better than the a peep (and it helped with some outside fight), or black ataris from outside and white can more easily live in the corner than simply peeping at b. It that case the edge of the board is sort of functioning as the 'friendly group' on that side, being an easy place to build eyes.

Click Here To Show Diagram Code: [go]$$W Pro game wedge $$ . . . . . . | $$ . . X . . . | $$ . a 1 b . . | $$ . . X . . . | $$ . . . . . . | $$ . . . . . . | $$ ------------+[/go]

Edit: found it, O Meien in 1984, move 58: http://ps.waltheri.net/database/game/13495/

lightvector · #8

Original post edited with the answer. I really like Uberdude's conceptualization of when this is a move worth considering.

Uberdude · #9

I wonder how often this "proton torpedo" shape crops up, so did a pattern search on waltheri (only supports rectangular area, don't have kombilo avaiable atm) for that 7x3 pattern: http://ps.waltheri.net/database/game/13495/. Then apply my human reasoning of if I think wedge is appropriate, i.e. at the 2 side stones both about the same strength / lightness? Then check with KataGo. I found 18 examples, here's a few:

First one in early opening Rin Kaiho vs Yamabe Toshiro in 1976 http://ps.waltheri.net/database/game/16256/ . By my reasoning wedge is not appropriate here because the q8 stone is weak and black doesn't want to sacrifice it because then the white one-point jump gets strong, rather jumping out is natural. Rin indeed jumped at a which KataGo thinks is a good move (and top policy with 30%), though knight jump to b best (14% policy). It does give a few playouts to the wedge (0.3%) policy and interestingly it's only 1% winrate worse but as expected white would atari from below and although black builds the top q8 suffers.

Click Here To Show Diagram Code: [go]$$Bc $$ +---------------------------------------+ $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . X . . . . X . . . . . . . . | $$ | . . . O . . . . . , . . . . . , X . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . O . . . . . . . . . . . . X . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . O M O . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . b . . . . . | $$ | . . . . . . . . . . . . . a . X . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . O . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . X . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . . O . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ +---------------------------------------+[/go]

Second game looks closer to the sort of situation wedge could apply, with single stones either side. Liu Xing (white) vs Kono Rin in 2002 http://ps.waltheri.net/database/game/10331/ . There is an asymmetry in that white would probably rather save the d12 stones, because then you are attacking black d14, but also if black gobbles up d6 that's pretty big, so that stone settling with a wedge and sacrificing d12 for now (can always sacrifice on small scale with white c16 likely sente later) looks good too. So I'd rate this is a fairly strong candidate for the wedge. In the game Liu jumped out to g12, presumably with his eye on d14 and also the weakish black group on the top side. KataGo gives the wedge as top policy with 17%! f12 one space jump is 2nd at 15%, 2 space jump a distant 1.6% of policy. And letting KG read to 30k playouts the wedge remains a strong blue preference. Pleasing that my pattern match and a little mid-dan amateur logic can find better moves than top pros, which would likely get a "wow! so strong AI move that human wouldn't think of" in a commentary.

Click Here To Show Diagram Code: [go]$$Bc $$ +---------------------------------------+ $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . . . . . . O . . . . | $$ | . . . , . O . . . O . X . . . , X . . | $$ | . . . . . . . . . . . . . . . O X . . | $$ | . . . X . . . . . . . . . . O . X . . | $$ | . . . . . . . . . X . X . . . . . . . | $$ | . . . O . . a . . . . . . . O . . . . | $$ | . . . . . . . . . . . O . . . . X . . | $$ | . . . , . . . . . , . . . . O , . . . | $$ | . . X M X . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . X . . . , . . . . . O . . . | $$ | . . . X . . . . . . X . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ +---------------------------------------+[/go]

P.S. If the wedge is the best move here, it begs the question: was the one point jump a mistake and wedge is punishment, or is it a fair result? If e9 was the last move here KG strongly agrees with it with 66% policy, and strong blue through to 50k playouts, though then tenuki to q3 is looking promising too. So no jump isn't bad, just jump and wedge is AI joseki

Uberdude · **#10**

Example 3, another early one at move 12, white is Fujisawa Shuko 9p vs Hashimoto Shoji 9p http://ps.waltheri.net/database/game/37720/ . To the right we again have the 2-space high approach to a 3-4 with knight answer, to left a 4-4 pincer. There's similarities to example 2: the 2-space approach stone could be light and sacrificed via 3-3 attachment in corner later or saved if atari the other way; whilst the h16 stone is looking at attacking the f17 stone. So I'd say looking ripe for a wedge as whichever direction black ataris from white seems content. In the actual game Shuko continued with 2-space jumps with a-b-c-d. KataGo's top policy is j17 with 23%, but with only a few playouts the wedge shines blue, Shuko's actual move loses 3%, and continues shining blue on some the subsequent moves. Leela Zero also likes the wedge.

Click Here To Show Diagram Code: [go]$$B $$ +---------------------------------------+ $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . X . . . . X . . . . . . . . | $$ | . . . O . . . O . , M . . O . , X . . | $$ | . . . . . b . . . . X . . . . . . . . | $$ | . . . O . . . . . . . . . . . X . . . | $$ | . . . . . d . c . . . . . a . . . . . | $$ | . . . e . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . O . . . | $$ | . . X . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ +---------------------------------------+[/go]

Example 4 is Kitani Minoru 8p vs Iwamoto Kaoru 8p http://ps.waltheri.net/database/game/5954/ . Here I'd say the wedge is not applicable, because although it matched the simple 3x7 pattern, the black group on the left is a lot bigger and heavier than the single stone on the right, so white has an obvious better direction to reply with atari from left and g3 peep covers the cut in sente to boot. KataGo agrees, no policy on the wedge, some on peeps either side but most on d8 press. Blue is p3 attachment. If played it loses 13%. So far that's 2 out of 2 cases where I've correctly identified the wedge isn't good, and 2 out of 2 where I have correctly identified it is good (and the human pros didn't). Nice.

Click Here To Show Diagram Code: [go]$$Bc Kitani (white) vs Iwamoto 1951 $$ +---------------------------------------+ $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . X . . . . . . . . . . . . . . | $$ | . . X , . . . . . , . . . . X , X . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . O . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . O . . . . . . . . . . . . . . . . | $$ | . . . . X . . . . . . . . . . . . . . | $$ | . . . . X . . . O . . . . . . . O . . | $$ | . . O O O X . a M b . X . . . , . . . | $$ | . . . O X d c . O . . . . . . O . . . | $$ | . . O X . X . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ +---------------------------------------+[/go]

Example 5, a more modern and developed middlegame one Choi Cheolhan 9p white vs Gu Li 9p in 2011, http://ps.waltheri.net/database/game/51971/ . This one's a bit trickier. The q6 white stone isn't as light as the high approaches of before because black r4 (the last move) takes the corner so there's no option to sacrifice it on a small scale and go for a corner life. And is q12 strong or weak? It's reasonably light in that if it dies and q6 lives nicely in exchange that's probably ok, as the top right black group is already alive with r18. But what's going on with n14 group, that's not totally settled either. But if wedge black atari from below looks better to gobble up q6 and let q12, which wasn't as weak anyway, get stronger. So I think wedge is probably an ok move here, but maybe not the best. In the game Choi played p4 to pull out the q6 stone but, in combo with black's attack on q12 (which did live), it ended up dying in a ko. Kata Go top policy is r5 at 32% with about 15% each on jumps at o5 and o6, with the wedge 4th with 7%. Blue starts out at L10 (I didn't really consider the unsettled-ness of black's top group), but then flits to the wedge and after 50k we have L11 as blue with 10.1% winrate and 7.2 points black lead, wedge is a close 2nd choice at 9.8% and 7.0 points and r5 / o5 / o6 / o7 bringing up the rear. p4 loses 1%, it's not so bad a sente move. So I think can again count as a success for my new shape heuristic.

Click Here To Show Diagram Code: [go]$$Bc $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . X X O . O . X X O . . . . O X . . | $$ | . . X O O X . X O X . O . O . . X . . | $$ | . . X O . O X . O X . O . . . X . . . | $$ | . . . O . . O O O X X . . . . . . . . | $$ | . . X O X . X O X X . O O . . X . . . | $$ | . . X O . . O X O X O X . . . . . . . | $$ | . . O O . . O X . X . . . . . O . . . | $$ | . . . . . . . X . . . O . . . . . . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . X M X . . | $$ | . . . O . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . O . . . | $$ | . . O . . . . O . . . . . . . . . . . | $$ | . X X O . X . O . O . . X . a , B . . | $$ | . . . X . X O X X O . . . . . X . . . | $$ | . . . . . . O . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

Interesting top policy move

Who is online