Opening problems for AI: Problem 8

Bill Spight · #1

Sakata Eio was one of those players who went his own way, and rather successfully, being one of the winningest players of all time. In the diagrammed position, from GoGoD 1953-03-25a, versus Takagawa, there are a number of standard plays. Sakata picked one that Elf reckons lost 22½% to par.

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

hyperpape · #2

Bill Spight · #3

Click Here To Show Diagram Code: [go]$$Wcm20 Actual game $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O O X . . . X . . . . . X O . . | $$ | . . O , X X . . . , . . . . . X O . . | $$ | . . . O . . . . . 1 . . . . . X . . . | $$ | . . . . . 2 . . . . . . . . . . O O . | $$ | . . . . . . . . . . . . . . X . . . . | $$ | . . . . . . . . . 3 . . . . . . X . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . 4 , . . . . . , . . . . . X . . . | $$ | . . . . O . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

The boshi, :w20:

, strikes at the heart of Black's incipient moyo. But Elf much prefers enclosing the bottom left corner. Takagawa jumped out with :b21:

, a move that Elf also criticizes, and after Sakata played a two space jump, Black played in the bottom left corner.

Click Here To Show Diagram Code: [go]$$Wcm20 Elf's response $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O O X . . . X . . . . . X O . . | $$ | . . O , X X . . . , . . . . . X O . . | $$ | . . . O . . . . . 1 . . . . . X . . . | $$ | . . . . . . . . . . . . . . . . O O . | $$ | . . . . . . . . . . . . . . X . . . . | $$ | . . . . . . . . . . . . . . . . X . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . 0 . . . . . . . . . . . . . . . . . | $$ | . . . 9 . . . . . . . . . . . . . . . | $$ | . . 8 7 . . . . . . . . . . . . . . . | $$ | . . 6 3 . . . . . . . . . . . . . . . | $$ | . . . 4 5 . . . . , . . . . . X . . . | $$ | . . 2 . O . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

Elf chooses the 3-3 invasion of the bottom left corner, leaving the top side as is. Next, White invades the bottom right corner.

Click Here To Show Diagram Code: [go]$$Wcm20 ELf's mainline $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O O X . 6 8 X . . . . . X O . . | $$ | . . O , X X . 5 . 7 . . . . . X O . . | $$ | . . . O . . . . . . . . . . . X . . . | $$ | . . . . . 3 . . . . . . . . . . O O . | $$ | . . . . . . . . . . . . . . X . . . . | $$ | . . . . . . . . . . . . . . . . X . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . 1 , . 4 . . . , . . . . . X . . . | $$ | . . . . O . . . . . . . . 2 . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

simply encloses the bottom left corner. OC, many players would choose this play. What is shocking is how bad the boshi seems to be by comparison. :b21:

encloses the bottom right corner, and then :w22:

plays keima against the Black moyo. :b23:

plays a shoulder hit against the bottom left enclosure, in the modern style. Then :w24:

and

reduce Black's moyo.

John Fairbairn · #4

Quote:

What is shocking is how bad the boshi seems to be by comparison.

I feel shocked, too, but an important caveat for this crop of positions is that they are from an era with 4.5 komi (or sometimes 0). Another point is that there were rather few players in those days and they met each a lot. Sakata and Takagawa played each other over 100 times. They were very familiar with each other's styles. It was therefore common to play the man as much as the board. Kitani seemed specially addicted to this, but he was far from alone. I don't know how much Sakata indulged in it, but playing a boshi (a cap) against Boshi no Takagawa has the hallmarks.

It may of course be that, by adjusting for komi size, the win rate would change by exactly the same extent for both the cap and the 3-3 move. But I'd be interested to hear whether it's possible that the change would vary in each case, In other words, could changing the komi alter, say, the balance between the corners and the centre?

Uberdude · #5

John Fairbairn wrote:

Quote:

What is shocking is how bad the boshi seems to be by comparison.

It may of course be that, by adjusting for komi size, the win rate would change by exactly the same extent for both the cap and the 3-3 move. But I'd be interested to hear whether it's possible that the change would vary in each case, In other words, could changing the komi alter, say, the balance between the corners and the centre?

KataGo to the rescue!

Using 7.5 komi (and Chinese rules, I don't know how to change to Japanese), KG thinks (at 50k playouts) the shimari is 54.0% for white or a lead of 0.6 points. A very close second (53.8%) is just one space away from the bad cap, the classic bot attachment at k16, a move that's actually sente. After the cap and black's punishment invasion at bottom left 3-3, white's next move (KG agrees with Elf d5, I keep same player to play) is down to 41.8% and black leads 1.0 points. So a loss of 12.2% or 1.6 points.

Using 4.5 komi, KG still wants to shimari, with 33.4% and black leads 2.2 points. After the cap best play is still the same, black 3-3, white knight press at d5 is now 22.8% or 4.2 points lead for black. So a loss of 10.6% or 2.0 points. So pretty similar, though strangely more points loss with a smaller winrate loss. Maybe the Chinese rules lack of sharpness vs Japanese effect, or that measuring changes in percentage points rather than fractional change as you approach 0 distorts relative changes (2% to 1% is a halving of winrate whereas 50% to 49% is not much)?

P.S. Another human mistake that could have been avoided by following the Leela Zero Opening Gospel

lightvector · #6

@Uberdude - change to Japanese rules by pressing "E" in Lizzie to open the GTP console and type "kata-set-rules japanese" into it and hit enter. Ideally the GUI would send such commands automatically based on the rules in an SGF file... but unfortunately the GTP protocol never built in any support for rules, so despite KataGo having supported this GTP extension for a few months now, it seems many GUI programs haven't implemented it, because - granted - it is an extension and not part of the "official" protocol.

Alternatively, in your gtp.cfg or gtp_example.cfg or default_gtp.cfg, edit it to be "rules=japanese":
https://github.com/lightvector/KataGo/b ... le.cfg#L88

It will use whatever rules you put here each time it starts up. You can still of course dynamically change it by typing in the GTP console too.

Bill Spight · #7

John Fairbairn wrote:

It may of course be that, by adjusting for komi size, the win rate would change by exactly the same extent for both the cap and the 3-3 move. But I'd be interested to hear whether it's possible that the change would vary in each case, In other words, could changing the komi alter, say, the balance between the corners and the centre?

I wish to emphasize, if I haven't done so enough already, that my purpose is not to show how bad the players were, but how opening theory may change in the AI era. Yes, it is a good idea to see how komi matters and to judge plays accordingly.

In doing the research for these problems, I have found out a couple of interesting things. First, opening theory of the 20th century, aside from considering new opening patterns, was largely based upon best play in the 19th century, even though those games were played without komi. Second, looking at those ancient plays in the Elf commentaries revealed that Elf liked many of those moves in the 19th century games that it didn't like in the 20th century games. Even though they were no komi games.

ez4u · #8

As Uberdude wrote above, katago calculates that attaching on top of the extension in the original position is almost equivalent to playing in the lower left. We should also be aware of what comes before that.

In the position below, katago calculates that playing :b1:

instead of "a" in the bottom left is a bigger mistake than playing the later attachment (instead of the cap) in the center of the top! needless to say it also calculates that White is better off to tenuki at :w6:

in order to enclose the bottom left. (A year later Takagawa reached the position after 6 against Sugiuchi Masao. Instead of the low :b7:

, he switched to the high extension at K16 instead. Sugiuchi did not approach the top right, splitting the right side at R12 instead. What did Elf think of that?)

Step by step from :w6:

on, katago calculates that both Black and White would be better off to turn to the bottom left. An interesting exception comes after White invades the corner with :w10:

below. Katago calculates that Black should answer locally but by blocking at "c" rather than "d" as played in the game. Apparently there is more value to the right side formation than the top. The difference is not so large and might narrow or reverse with a deeper analysis but at 35K that is how things stand. Katago anticipates something like this a the followup (komi = 0 and Japanese rules as in the original game).

Click Here To Show Diagram Code: [go]$$Bc Earlier that same day $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . 4 2 3 . . . 7 . . . . . d 0 . . | $$ | . . O , 1 5 . . . , . . . . . X c . . | $$ | . . . 6 . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . b 8 . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . 9 . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . a , . . . . . , . . . . . X . . . | $$ | . . . . O . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

[/quote]

Bill Spight · #9

ez4u wrote:

A year later Takagawa reached the position after 6 against Sugiuchi Masao. Instead of the low :b7:

, he switched to the high extension at K16 instead. Sugiuchi did not approach the top right, splitting the right side at R12 instead. What did Elf think of that?

Good question!

So good, in fact, that it is the next problem.

Uberdude · **#10**

ez4u wrote:

A year later Takagawa reached the position after 6 against Sugiuchi Masao. Instead of the low :b7:

, he switched to the high extension at K16 instead.

Makes me wonder if Takagawa felt uncomfortable after Sakata's cap and so wanted to play high instead to prevent it. So even if Sakata's move was a mistake according to Elf and other bots, playing the man not the board it may have been an excellent one.

Bill Spight · **#11**

Uberdude wrote:

ez4u wrote:

A year later Takagawa reached the position after 6 against Sugiuchi Masao. Instead of the low :b7:

, he switched to the high extension at K16 instead.

Makes me wonder if Takagawa felt uncomfortable after Sakata's cap and so wanted to play high instead to prevent it. So even if Sakata's move was a mistake according to Elf and other bots, playing the man not the board it may have been an excellent one.

Sakata may or may not have been turning the tables on Boshi no Takagawa, but the boshi strikes me as Sakata's leave'em bleeding style, anyway.

dhu163 · **#12**

Opening problems for AI: Problem 8

Who is online