Who says AI is territorial? (Joseki reevaluation)

xela · #1

So I'm asking what KataGo thinks of some old joseki, and why it doesn't like pincers. (It doesn't use words, but that doesn't have to stop us having a conversation!) And it shows me this:

Click Here To Show Diagram Code: [go]$$Bc Start with a one-space high pincer... $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . , . . . . . X . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . 3 . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . 5 W . 4 . . . . . . . . . . . . . | $$ | . . 7 6 . . . . . . . . . . . . . . . | $$ | . . B 8 1 . . . . . . . . . . . . . . | $$ | . . 0 9 . 2 . . . , . . . . . X . . . | $$ | . . . W . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

This diagram is in good old Ishida, and we're told that :b9:

is "unreasonable", and :w10:

is the refutation. But KataGo seems to like this position for black. So I'm following along, looking for KataGo to show me the good move later in the sequence that people didn't notice back then.

Click Here To Show Diagram Code: [go]$$Bc KataGo keeps following the book; connects $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . , . . . . . X . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . X . . . . . . . . . . . . . . . | $$ | . . a . . . . . . . . . . . . . . . . | $$ | . . X O . O . . . . . . . . . . . . . | $$ | . . X O 8 6 . . . . . . . . . . . . . | $$ | . . X O X 3 4 b . . . . . . . . . . . | $$ | . . O X 1 O 7 . . , . . . . . X . . . | $$ | . . 0 O 2 5 . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

So this very position appears in Ishida. KataGo thinks :w4:

is a small mistake, but likes all the black moves. Next, Ishida tells us that a and b are miai, and black "has almost collapsed". But KataGo thinks black is comfortably ahead here! It hasn't found a new move for black -- it just disagrees with the evaluation.

Carrying on, KataGo says b is the better of the "miai" points, and this should follow. The next diagram is recommended play for both sides according to KataGo, and leaves black comfortably ahead.

Click Here To Show Diagram Code: [go]$$Bc White gets a big corner. $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . , . . . . . X . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . 7 . , . . . . . , . . . . . , . . . | $$ | 9 4 5 X . . . . . . . . . . . . . . . | $$ | . 3 2 6 . . . . . . . . . . . . . . . | $$ | . 8 X O . O . . . . . . . . . . . . . | $$ | . . X O O O . . . . . . . . . . . . . | $$ | . 0 X O X X O 1 . . . . . . . . . . . | $$ | . . O X X O X . . , . . . . . X . . . | $$ | . . O O O X . . . . . . . a . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

Next, black a. KataGo says to Ishida: there you go, you can have 20 points of solid territory, and I'll take a vague moyo in gote. Now who's the territorial one?

John Fairbairn · #2

Seems like apples and pears to me. Katago NEVER evaluates josekis (or tesujis etc) per se. It evaluates fusekis or other whole-board positions. The rest of the board changes the evaluation. In this case, the right-hand side of the board seems to make the miai at 'b' a no-brainer, but that could easily change with a different fuseki.

In contrast, when professionals talk about josekis, they assume the word 'LOCALLY' is implicit in their evaluations, and quite often they make that explicit also.

RobertJasiek · #3

Might be interesting but to give your study meaning, you need to state the number of visits of the top move, or better the second-top move, of the created position! Besides komi, statistical values and stability of the values.

Knotwilg · #4

Click Here To Show Diagram Code: [go]$$Bc Colors reversed $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . a . . . . . . . . . . . . . | $$ | . . . B . . . . . , . . . . . X . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . c . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . X . , . . . . . , . . . . . , . . . | $$ | X . X X . . . . . . . . . . . . . . . | $$ | . X O O . . . . . . . . . . . . . . . | $$ | . O X O . O . . . . . . . . . . . . . | $$ | . . X O O O . . . . . . . . . . . . . | $$ | . O X O X X O X . . . . . . . . . . . | $$ | . . O X X X X . . , . . . . . W . . . | $$ | . . O O O X . . . . . . . 1 . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

I reversed the colors of the marked stones and ran the whole sequence again to see if KataGo would evaluate this differently due to the difference in fuseki. It doesn't seem to do so, indicating we might see this indeed as a difference in evaluation of the joseki in general than how Ishida evaluated it. After "completing" the lower left, :b1:

is natural and endorsed by KataGo, with A being the blue move by a small margin. KataGo gives B+4 at this point.

I admit I would be more comfortable with those 20 points of territory and a strong group sticking its head out into the center. On the other hand I can easily buy into KataGo's evaluation, that at least Black is not behind and more "active".

Let's acknowledge that besides Black's clumsy shape at the time of "collapsing", Black has captured two White stones and is about to capture another one, which would compensate for the 3 captured Black stones. White too has a very inefficient double empty triangle, and should he connect the stone in atari, there's a 3rd one.

KataGo is known to put a high value on "efficiency". It also values sente.

Knotwilg · #5

Alternatives

Click Here To Show Diagram Code: [go]$$Bc Start with a one-space high pincer... $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . , . . . . . X . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . a . . . . . , . . . . . , . . . | $$ | . . . 3 . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . 5 W . 4 . . . . . . . . . . . . . | $$ | . . 7 6 . . . . . . . . . . . . . . . | $$ | . . X 8 1 b . . . . . . . . . . . . . | $$ | . . 0 9 c 2 . . . , . . . . . X . . . | $$ | . . . W . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

- KG has a small preference for A i/o :b3:

- likewise B i/o :b5:

- and for

KG prefers to play atari from the other side at C, creating a strong group on the outside, then take sente

Click Here To Show Diagram Code: [go]$$Bc KataGo keeps following the book; connects $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . X . . . . . , . . . . . X . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . X . . . . . . . . . . . . . . . | $$ | . . a . . . . . . . . . . . . . . . . | $$ | . . X O . O . . . . . . . . . . . . . | $$ | . . X O 8 6 . . . . . . . . . . . . . | $$ | . . X O X 3 4 b . . . . . . . . . . . | $$ | . . O X 1 O 7 . . , . . . . . O . . . | $$ | . . 0 O 2 5 . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

The real "mistake" leading into the 4 point (dis)advantage, is :w4:

here. According to KataGo it loses 3 points, which is a mistake I allow myself but is substantial at the pro level.

Instead ...

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

White should reinforce at :w1:

here, then ignore :b2:

for a moment, building up strength in the center, then come back to connect in sente

Click Here To Show Diagram Code: [go]$$Wc White's better sequence $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . X . . . . . , . . . . . X . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . 4 . 3 . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . X X 1 . . . , . . . . . , . . . | $$ | . . . X O O . . . . . . . . . . . . . | $$ | . . X O . . . . . . . . . . . . . . . | $$ | . . X O . O . . . . . . . . . . . . . | $$ | . . X O . . X . . . . . . . . . . . . | $$ | . . X O X X . . . . . . . . . . . . . | $$ | . . O X X O X . . , a . . . . O . . . | $$ | . O . O O O 2 . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

... reinforce again and eventually annoy black again at the bottom.

When reversing the corner stones as per xela's original opening, the suggestions remain approximately the same.

Knotwilg · #6

Two more contexts

1) In isolation

Click Here To Show Diagram Code: [go]$$Bc Isolation $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . 1 . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . X . , . . . . . , . . . . . , . . . | $$ | X . X X . . . . . . . . . . . . . . . | $$ | . X O O . . . . . . . . . . . . . . . | $$ | . O X O . O . . . . . . . . . . . . . | $$ | . . X O O O . . . . . . . . . . . . . | $$ | . O X O X X O X . . . . . . . . . . . | $$ | . . O X X X X . . , . . . . . , . . . | $$ | . . O O O X . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

In isolation, KataGo raises the advantage to 5 points.

2) occupied corners with 2 shimaris

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

When putting a few shimaris in, the evaluation holds its ground and remains around 4.5 better for Black

I checked the joseki at Josekipedia, where the KataGo analysis is mostly replicated, but the source credited is professional go player Jiang Ming jiu

See https://senseis.xmp.net/?JiangMingjiu

kvasir · #7

xela wrote:

So I'm asking what KataGo thinks of some old joseki, and why it doesn't like pincers. (It doesn't use words, but that doesn't have to stop us having a conversation!)[...]

I wonder where the myth about pincers came from. KataGo likes pincer when they are good and usually when it likes a pincer more than all other moves it is when we are close to the middle game, the third or fourth corner in human-human games come to mind. In komi games at least it tends to like narrow pincers more than wider pincers. Pros have been very similar, a few pro's appeared to play a pincer in every corner, but judging from game databases they are usually not the preferred move in the first and second corner.

Doesn't this make sense? If you don't have a special situation there is no reason why a pincer would be better than defending. The most common move has always been defending after an approach move.

Things change a bit if we consider specific openings. There are openings with pincers. Each case is different, why humans play pincers is more interesting than why computers do or don't.

Let me explain. The computer always does the same thing and one move has by the nature of the thing to be evaluated as better in each case. This usually only means marginally better in the early game. The computer also plays that move with no other considerations. A human may know which move is evaluated best according to some criteria but they can have different criteria. He knows many criteria, he knows human criteria. That is what makes it interesting, there are opinions to be expressed and exchanged, there is no idea of one shoe fits everyone or that there is one truth that everyone must accept.

As for this pincer. I came to the conclusion, when I analyzed it in positions that I'm likely to play, that it is often close to whatever KataGo thinks is the best move. The thing about the best move is that it can be hard to guess, this pincer is not hard to guess. It is also an interesting pincer, it was played by Go Seigen and there are many logical variations that follow, that allow for fairly easy to understand tenuki. And finally, lot of the analysis for booklovers was wrong, this was known long before AI could play Go. In particular it is wrong which was quoted from Ishida Joseki Dictionary, I think it is partially corrected in Takao's Basic Josekis but you can also just find pro games were this analysis is shown to be inaccurate.

So it is a fun pincer, more fun if you have a flexible attitude toward Go, maybe like Go Seigen.

RobertJasiek · #8

With ca. 5 million playouts for the top move, I checked quite a few early opening pincers. Typically, AI avoids them for losing ca. 0.1 points of empirical scores. 3-space pincers lose more.

xela · #9

John Fairbairn wrote:

...Katago NEVER evaluates josekis (or tesujis etc) per se...assume the word 'LOCALLY' is implicit in their [human pro] evaluations...

No need to shout. You've seen me round here before: you know I wasn't born yesterday. Give me a bit of credit here.

I did check the same position in a few other contexts. I decided to keep my post short and focussed, rather than adding extra diagrams and numbers, so that people don't get lost in the details and miss the point. You've managed to miss the point anyway ;-) Knotwilg has done a fine job of adding the extra context.

If you're willing to spend a bit of time learning the language of all the decimals, instead of dismissing them out of hand, you'll find that when the books say "this is joseki" or "this is locally fair to both sides", KataGo might say B+0.7 points on this full board, W+1.3 on that board, possibly as much as a 2 point advantage if the joseki really is inappropriate for the full board context. We're usually looking at fairly small swings here. But when it starts saying B+5 in a position with just star points in three corners, for a corner position that a pro evaluates as not just "better for white" but "black collapses", then something fairly unusual is going on. Without treating the numbers as gospel truth, they can still be useful signposts for things we could learn from. I thought this one was worth a conversation.

The point I'm trying to make is that a bunch of cliches are springing up about "AI style", and I think they deserve closer examination. The one I'm noticing most is "AI values the corners more highly". So here's a position where I think most humans would lean towards white because of the corner territory, but at least one AI gives a contrary assessment. And now kvasir has called out the cliche about "no pincers" (which also gets a chapter in the Fuseki Revolution book, promoting said cliche).

RobertJasiek · **#10**

xela wrote:

rather than adding extra diagrams and numbers, so that people don't get lost in the details and miss the point.

The point is that I do not trust claimed AI evaluations without any evidence. Current AI provides numbers as empirical evidence. Making a strong claim, such as "joseki reevaluation", requires evidence.

Knotwilg · **#11**

RobertJasiek wrote:

xela wrote:

rather than adding extra diagrams and numbers, so that people don't get lost in the details and miss the point.

The point is that I do not trust claimed AI evaluations without any evidence. Current AI provides numbers as empirical evidence. Making a strong claim, such as "joseki reevaluation", requires evidence.

The fact that xela, myself and possibly Jiang Mingjiu came to the same conclusion should be some sort of evidence. The nice thing about science is that you can verify an experiment for yourself, in case you have access to a lab. And you do too!

But to humour your request - I hope - I ran the position in isolation again, for 24K playouts. Komi being 7.5 on my version of KG and unable to reset it to 6.5, White starts out with ~1 point advantage.

Around 11K the advantage (for White in this case) drops a decimal from W+6.0 to W+5.9 and stays there until the end of my evaluation. With the original 1 subtracted that gives a 5 point loss for the side taking the corner territory (for Black in this case).

This 5 point loss comes from a couple of "mistakes", the bigger one being :b5:

in the next diagram.

Click Here To Show Diagram Code: [go]$$Bc $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . , . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . O . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . O X . X . . . . . . . . . . . . . | $$ | . . O X . . . . . . . . . . . . . . . | $$ | . . O X O 4 5 . . . . . . . . . . . . | $$ | . . 1 O 2 X . . . , . . . . . O . . . | $$ | . a . X 3 . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

KataGo wants Black to play :b5:

at A.
I let KG run again before and after :b5:

, for 24K runs each. Result: 6.1 - 1.8 = 4.3
The other, smaller, mistake contributing to the 5 point loss is :b1:

which KG prefers as an atari on the other side, at :w2:

, in order to keep sente.
Again, with 24K playouts, KG puts B+0.1 after that "good atari". 0.1 - (-1.8) = 1.9

As one may point out, 4.3 + 1.9 = 6.2 > 5. That's because in the moves before, White lost about 1 point too.

Lizzie version 0.7.4.

SGF evidence:

RobertJasiek · **#12**

Knotwilg wrote:

The fact that xela, myself and possibly Jiang Mingjiu came to the same conclusion should be some sort of evidence.

I am as unimpressed as by the pre-AI history of all professional joseki knowledge. AI shows us huge joseki branches overlooked before.

Quote:

24K playouts. Komi being 7.5

This is an improvement but my experience with KataGo joseki study has taught me not to trust anything below 100k per move, and often more is needed. The most difficult joseki decisions require more than 5 million playouts per candidate. Way beyond that, changes (other than reversions) are very rare (I occasionally do up to ca. 110 million when 64GB RAM are 99.8 ~ 99.9% filled but it depends on the position; sometimes ca. 50 million is the limit when there are several close candidates).

RobertJasiek · **#13**

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

a = 13.7 million playouts 46.1% -0.4 correct
b = 4.1k playouts 62.7% 1.5 mistake

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

a = 202.6 million playouts 61.9% 1.3 correct
b = 26.2 million playouts 63.7% 1.6 mistake

99.9% of 64 GB RAM and 100% swap file have been used.

Click Here To Show Diagram Code: [go]$$W variation 1, correct $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . d . . | $$ | . . . . . a . . . . . . . . . 4 3 . . | $$ | . . . X . . . . . , . . . . . X 5 . . | $$ | . . . . . . . . . . . . . . . 6 7 . . | $$ | . . . . . . . . . . . . . . . c 8 9 . | $$ | . . . . . . . . . . . . . . . . b . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . X . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . X O . O . . . . . . . . . . . . . | $$ | . . X O . . . . . . . . . . . . . . . | $$ | . . X O X . . . . . . . . . . . . . . | $$ | . . 2 X 1 O . . . , . . . . . O . . . | $$ | . . . O . . . . . . . . . . . . e . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

Click Here To Show Diagram Code: [go]$$W variation 2, correct continuation $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . X . . . . . , . . . . . X . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . X . . . . . . . . . . . . . . . | $$ | . . 5 . . . . . . . . . . . . . . . . | $$ | . . X O . O . . . . . . . . . . . . . | $$ | . . X O . . . . . . . . . . . . . . . | $$ | . 6 X O X 3 . . . . . . . . . . . . . | $$ | . 8 O X X O . . . , . . . . . O . . . | $$ | . . 1 O 2 7 . e . . . . . . . . . . . | $$ | . 9 b a 4 c d . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

Click Here To Show Diagram Code: [go]$$W variation 3, correct continuation from move 2 $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . X . . . . . , . . . . . X . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . d . , . . . . . , . . . . . , . . . | $$ | . . . X . . . . . . . . . . . . . . . | $$ | . . . 4 . . . 9 . . . . . . . . . . . | $$ | . . X O . O . . . . . . . . . . . . . | $$ | . c X O . 5 . 7 . . . . . . . . . . . | $$ | e b X O X 2 6 . . . . . . . . . . . . | $$ | . . O X X O 8 . . , . . . . . O . . . | $$ | . 3 . O 1 a . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

Gérard TAILLE · **#14**

RobertJasiek wrote:

xela wrote:

rather than adding extra diagrams and numbers, so that people don't get lost in the details and miss the point.

The point is that I do not trust claimed AI evaluations without any evidence. Current AI provides numbers as empirical evidence. Making a strong claim, such as "joseki reevaluation", requires evidence.

I agree with you. AI is a very useful tool indeed, but we have to be careful not to make Katago say what it doesn't say.

RobertJasiek wrote:

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

a = 13.7 million playouts 46.1% -0.4 correct
b = 4.1k playouts 62.7% 1.5 mistake

Here Katago says only
a = 13.7 million playouts 46.1% -0.4
b = 4.1k playouts 62.7% 1.5
and Katago do not say that "a" is correct and "b" is a mistake.
On contrary Katago says that after white "a" black may win (46.1%) while after white "b" black might lose (37.3%). In that case "b" might be the correct move.
Note that if Katago were God then Katago will always find a value equal to 0.5, 1.5, 2.5 ... or -0.5, -1.5, -2.5 ... and the pourcentage will be always 0% or 100%.
Unless in the late endgame Katago proposes explicitly a "estimate" value.

When you say "a" is correct and "b" is a mistake it is only "your own" interpretation of katago analysis.
I just say Katago believes "a" is better.

RobertJasiek wrote:

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

a = 202.6 million playouts 61.9% 1.3 correct
b = 26.2 million playouts 63.7% 1.6 mistake

With such near values you can be convinced that "a" is correct and "b" is a mistake but it is only your own interpretation. The result proposed by katago is really very uncertain.

Anyway, let me repeat katago is for me a very useful tool. No doubt with that.

xela · **#15**

OK, here's another one. (By the way, now I'm wishing I hadn't used the word "joseki". It's getting a disporportionate amount of attention. For me, this is about the common perception that "AI is territorial", and about exploring some positions where the AI assessment appears to go the other way.)

In Fuseki Revolution, theme 27, Shibano recommends this sequence as good for black. (KataGo mildly endorses this.)

Click Here To Show Diagram Code: [go]$$Bc19 $$ +---------------------------------------+ $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . . . . . X . . X . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . a . O . . | $$ | . . . . . . . . . . . . . . 8 1 2 . . | $$ | . . . . . . . . . . . . . 9 5 4 6 . . | $$ | . . . . . . . . . . . . . . . 3 O . . | $$ | . . . . . . . . . . . . . . . 7 . . . | $$ | . . . . . . . . . . . . . . . . X . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . . . . . . . X . . . | $$ | . . . . . . . . . b . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ +---------------------------------------+[/go]

Next, he recommends white a, and says that if white b instead:

Click Here To Show Diagram Code: [go]$$Bc19 $$ +---------------------------------------+ $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . 5 . . . . . . . . . . . . . | $$ | . . . O . . . . 7 . . . . X . . X . . | $$ | . . . . . . . . . . . . . . . . . 4 . | $$ | . . 6 . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . 3 1 2 O . . | $$ | . . . . . . . . . . . . . . O X O . . | $$ | . . . . . . . . . . . . . X X O O . . | $$ | . . . . . . . . . . . . . . . X O . . | $$ | . . . . . . . . . . . . . . . X . . . | $$ | . . . . . . . . . . . . . . . . X . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . . . . . . . X . . . | $$ | . . . . . . . . . W . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ +---------------------------------------+[/go]

This is where KataGo differs, offering this variation instead:

Click Here To Show Diagram Code: [go]$$Bc19 $$ +---------------------------------------+ $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . . . . . X . . X . . | $$ | . . . . . . . . . . . . . . . . . 5 . | $$ | . . . . . . . . . . . . . . . . 7 . . | $$ | . . . . . . . . . . . . . . 1 3 W . . | $$ | . . . . . . . . . . . . 6 2 O X W . . | $$ | . . . . . . . . . . . . 4 X X W W . . | $$ | . . . . . . . . . . . . . . . X W . . | $$ | . . . . . . . . . . . . 8 9 . X . . . | $$ | . . . . . . . . . . . . 0 . . . X . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . b . . . . . . . . . . . . a . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . . . . . . . X . . . | $$ | . . . . . . . . . O . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ +---------------------------------------+[/go]

Next, black a, white tenuki (possibly b but there are several almost equally good options). KataGo thinks that :b1:

was a small mistake, and :w2:

in this diagram is significantly better than :w2:

at

.

The five white stones on the side are not totally dead, but definitely unhealthy. It seems again that, in the right circumstances, KataGo is happy to give away large amounts of territory in order to build a moyo.

RobertJasiek · **#16**

Gérard TAILLE wrote:

When you say "a" is correct and "b" is a mistake it is only "your own" interpretation of katago analysis. [...]
"a" is correct and "b" is a mistake but it is only your own interpretation.

Yes.

An interpretation that has helped me extraordinarily well while studying josekis and openings for half a year with KataGo.

Quote:

The result proposed by katago is really very uncertain.

No. During my aforementioned study, my impression has been and is that, for josekis and openings, also small relative differences in KataGo's values are a very good indicator, provided that there are enough playouts and values have reached stability. Of course, one must analyse carefully and find out whether early values might be wrong by also exploring subsequent branches with KataGo. (I have not studied the middle game with KataGo carefully yet so cannot say much about its values then.)

ez4u · **#17**

xela wrote:

OK, here's another one. (By the way, now I'm wishing I hadn't used the word "joseki". It's getting a disporportionate amount of attention. For me, this is about the common perception that "AI is territorial", and about exploring some positions where the AI assessment appears to go the other way.)

In Fuseki Revolution, theme 27, Shibano recommends this sequence as good for black. (KataGo mildly endorses this.)

Click Here To Show Diagram Code: [go]$$Bc19 $$ +---------------------------------------+ $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . . . . . X . . X . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . a . O . . | $$ | . . . . . . . . . . . . . . 8 1 2 . . | $$ | . . . . . . . . . . . . . 9 5 4 6 . . | $$ | . . . . . . . . . . . . . . . 3 O . . | $$ | . . . . . . . . . . . . . . . 7 . . . | $$ | . . . . . . . . . . . . . . . . X . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . . . . . . . X . . . | $$ | . . . . . . . . . b . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ +---------------------------------------+[/go]

Next, he recommends white a, and says that if white b instead:

One of the things that strikes me about studying with Katago is that we can never really justify starting in the middle of things. We must go step by step. In this situation from Shibano's book, Katago considers the invasion of :w1:

below a poor choice. It would be better as an invasion at 'a' instead. If you do invade anyway and Black answers at :b2:

, do not play the rather poor extension of :b3:

. Instead just play 'b' in the upper left. If you insist on answering locally, the counter invasion at 'c' seems to be preferable. Finally, if you do extend and Black plays :b4:

(another "less than the best" according to Katago), of course you should not answer with :w5:

. The attachment in the corner with 'd' is the way to mix things up.

Click Here To Show Diagram Code: [go]$$Wc19 $$ +---------------------------------------+ $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . b . . . . . . . . . . . . . | $$ | . . . O . . . . , . . . . X . d X . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . 3 . . | $$ | . . . . . . . . . . . . . . . 4 5 . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , 1 . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . 2 . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . c . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . , . . . . . X . . . | $$ | . . . . . . . . . . . . . . . . a . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ +---------------------------------------+[/go]

The last idea mentioned of attaching at 'd' in the corner seems to take advantage of Black's extra wide shimari. It raises the question of whether Katago's evaluation is based on this particular situation. Hmmm....

If we set up the board with the original position but also the three alternative of 'a' through 'c' shown below, we see the following from Katago:

Click Here To Show Diagram Code: [go]$$Wc19 $$ +---------------------------------------+ $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . a b . . . . | $$ | . . . O . . . . , . . . . X c , X . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . e 3 . . | $$ | . . . . . . . . . . . . . . . 4 5 . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , 1 . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . 2 . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . , . . . . . X . . . | $$ | . . . . . . . . . . . . . . . . d . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ +---------------------------------------+[/go]

All in all, it does not seem that the book contains much work with real programs.
:scratch:

kvasir · **#18**

ez4u wrote:

One of the things that strikes me about studying with Katago is that we can never really justify starting in the middle of things.

I think you have really hit the nail on the head here. All too often discussion about computer variations devolves into hypotheticals — KataGo won't play it and any given pair of humans wouldn't be likely to either :-?

Knotwilg · **#19**

kvasir wrote:

ez4u wrote:

One of the things that strikes me about studying with Katago is that we can never really justify starting in the middle of things.

I think you have really hit the nail on the head here. All too often discussion about computer variations devolves into hypotheticals — KataGo won't play it and any given pair of humans wouldn't be likely to either :-?

Most of the time though it leads to very interesting discussions.

Also, what's most certainly is real, is that modern professionals use AI to help them understand Go better and find better moves. Of course professionals start from a different level of play and analysis so they are better equipped to interpret the sequences AI comes up with. But the practice of investigating variations (from books, from games) with AI in itself is probably a very good type of study.

ez4u · **#20**

Knotwilg wrote:

kvasir wrote:

ez4u wrote:

One of the things that strikes me about studying with Katago is that we can never really justify starting in the middle of things.

I think you have really hit the nail on the head here. All too often discussion about computer variations devolves into hypotheticals — KataGo won't play it and any given pair of humans wouldn't be likely to either :-?

Most of the time though it leads to very interesting discussions.

Also, what's most certainly is real, is that modern professionals use AI to help them understand Go better and find better moves. Of course professionals start from a different level of play and analysis so they are better equipped to interpret the sequences AI comes up with. But the practice of investigating variations (from books, from games) with AI in itself is probably a very good type of study.

Sorry if I created the wrong impression. These days I study using Katago very often. Mainly my own games, but I also run Katago while watching the NHK Tournament on TV here in Tokyo. I do it exactly so that I can explore other variations that are not mentioned by the commentators. My original point above is that it is no longer possible to study things like "joseki", at least not using AI. As JF and Xela back and forthed on earlier (ultimately agreeing I think), there is no local in AI assessments. The specifics of the whole board overwhelm the local variations in the corners, sides, etc. We might be interested in a particular fight, but we cannot really believe what the programs tell us as long as the next time we use a sequence the board position might be different.

Who says AI is territorial? (Joseki reevaluation)

Who is online