Who says AI is territorial? (Joseki reevaluation)

[xela]

[re Shibano's book] All in all, it does not seem that the book contains much work with real programs.

To be fair, Shibano's book is from 2021, and much of the research as probably done earlier. It's a long time ago in AI-years. I already noted in 2019 that different AIs can give different recommendations in the same position. And KataGo is not the only AI out there.

On the other hand, the book has an awful lot of "if white follows up this mistake with five more mistakes in a row, then black gets the advantage..." And apart from not much work with real programs, there's not much evidence of work with databases either (references to "noone plays this any more" or "AI invented this sequence" often don't line up with what I can see in GoGoD + Go4Go).

It's certainly interesting to get Shibano's opinions on the new trends, as someone on the front line of playing with or against these new patterns at a high level. But I would have preferred a smaller number of themes and some more in depth discussion, including some middlegame positions to show how the fuseki choices lead to different outcomes.

[xela]

...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.

There's a huge middle ground between "Just blindly put down the moves, you'll be OK, it's a joseki" versus "Everything depends on context and we understand nothing so forget trying to think about joseki". It doesn't need to be all-or-nothing here.

The fact is that if you look at a bunch of AI self-play games, you'll see that certain paterns pop up more than you'd expect by chance. And many of those patterns, if they don't work in every full board context, at least work well more often than not. You might easily conclude that these patterns are significant and should be given a special name... Since AlphaGo Zero, Leela Zero and KataGo have come to similar conclusions here, I think it's proven that joseki are an intrinsic part of go. Yes, they need to be handled with caution. We already knew that.

[xela]

...The first time I saw this position I confess I was impressed by the 20 white points of territory but here comes the help of katago which considers black position is quite good. After that I reconsidered the position and I said to me : you are a stupid boy, black position is OC far better...

Is it possible that this is simply an error in Ishida's joseki dictionary? The most common type of error that I encounter in Go material is swapping black and white.

My guess is that - to a pro of Ishida's generation -being squeezed as black is in that joseki sequence seems so unbearable and 'obviously wrong' that it affected the evaluation.

No, it's not a simple misprint in Ishida. There are multiple diagrams and descriptive explanatory text, so it's very clear that to him, the side being squeezed is the side that comes off worse. It's not a simple case of swapping two words or getting the colours mixed up.

I'm reminded of the many anecdotes about "expert judgement" in Thinking, slow and patronising. Although the book is overconfident on some points, there's a lot of good cautionary tales about how a group of experts can all move together in the wrong direction. The insight of "here comes the help of <new point of view> ... I said to me : you are a stupid boy ..." if you want to make it sound profound instead of comical, you can call it a paradigm shift. In primary school mathematics, we learn a lot of "obvious" things that took centuries to see for the first time. (But I did enjoy Gérard's description!)

See also: https://lifein19x19.com/viewtopic.php?f=12&t=19358

[RobertJasiek]

"AI invented this sequence" often don't line up with what I can see in GoGoD + Go4Go).

AI invents (or is used to invent) very much faster than professional players adopt so databases of their games report! Roughly 80% of what I find with AI in josekis I have never seen before but, on closer study with AI and my own subsequent study, is very convincing. Therefore, WRT josekis, databases of pro games are often hopelessly outdated for the sake of good joseki choice. These databases remain very useful but, in particular, not for thorough, general joseki study.

[Knotwilg]

The fact is that if you look at a bunch of AI self-play games, you'll see that certain patterns pop up more than you'd expect by chance. And many of those patterns, if they don't work in every full board context, at least work well more often than not. You might easily conclude that these patterns are significant and should be given a special name... Since AlphaGo Zero, Leela Zero and KataGo have come to similar conclusions here, I think it's proven that joseki are an intrinsic part of go. Yes, they need to be handled with caution. We already knew that.

I totally agree with xela here. Which doesn't mean I disagree with others per se, in some cases we might misunderstand each other. In my own schematic words:

1. pre-AI there were lots of joseki, established corner patterns, which give locally best play or acceptable variations, which need caution to apply in the context of the whole board
2. every decade or so, these joseki would get updates with the latest professional insights, some joseki going out of fashion, others coming back in, still others being new; the caution still applied
3. the past decade, there has been a major revision due to the advent of AI; some patterns have been discarded, others restored, many new have emerged; the caution still applies;
4. the fundamental difference pre- and post-AI is that we can now (statistically) quantify the results and can likewise quantify their value in a whole board context
5. contrary to a widespread misconception about the current state of AI - which I'm not accusing anyone in particular of - the evaluation remains largely statistical and does not portray one narrow path to victory, i.e. AI is not (yet) deterministic; the whole board mattes, but not (much) more than before AI; the difference in evaluation rarely goes beyond the margin of error and when it does, it is often easy to explain (like a ladder breaker); this is why, IMO, joseki still exist today

[RobertJasiek]

Points 1 to 3: ok.

Point 4: One must be careful not to misinterpret statistics! Typical mistakes: too few playouts, reversion statistics misunderstood, overlooking contradictions in statistics only dissolved by much deeper search.

Point 5: Far beyond statistics, go theory can be taken into account with related effort.

[xela]

"AI invented this sequence" often don't line up with what I can see in GoGoD + Go4Go).

AI invents (or is used to invent) very much faster than professional players adopt so databases of their games report!..

I'm referring to the opposite: several times Shibano shows a diagram and says "AI invented this move" (or words to that effect), yet I can see multiple examples in human games from before 2015.

[Gérard TAILLE]

The fact is that if you look at a bunch of AI self-play games, you'll see that certain patterns pop up more than you'd expect by chance. And many of those patterns, if they don't work in every full board context, at least work well more often than not. You might easily conclude that these patterns are significant and should be given a special name... Since AlphaGo Zero, Leela Zero and KataGo have come to similar conclusions here, I think it's proven that joseki are an intrinsic part of go. Yes, they need to be handled with caution. We already knew that.

I totally agree with xela here. Which doesn't mean I disagree with others per se, in some cases we might misunderstand each other. In my own schematic words:

1. pre-AI there were lots of joseki, established corner patterns, which give locally best play or acceptable variations, which need caution to apply in the context of the whole board
2. every decade or so, these joseki would get updates with the latest professional insights, some joseki going out of fashion, others coming back in, still others being new; the caution still applied
3. the past decade, there has been a major revision due to the advent of AI; some patterns have been discarded, others restored, many new have emerged; the caution still applies;
4. the fundamental difference pre- and post-AI is that we can now (statistically) quantify the results and can likewise quantify their value in a whole board context
5. contrary to a widespread misconception about the current state of AI - which I'm not accusing anyone in particular of - the evaluation remains largely statistical and does not portray one narrow path to victory, i.e. AI is not (yet) deterministic; the whole board mattes, but not (much) more than before AI; the difference in evaluation rarely goes beyond the margin of error and when it does, it is often easy to explain (like a ladder breaker); this is why, IMO, joseki still exist today

Knotwilg says that joseki exists sill today. I agree with that but I fear that we do not have all the same definition of joseki. OK we can simply say it is a standard sequence of moves in the corner (see https://senseis.xmp.net/?Joseki) but what means "standard" and what means "seqeunce". In particular how do you handle tenuki moves.

Let's take an example:

$$W
$$ | . . . . . . . . . . .
$$ | . . . , . . . . . , .
$$ | . . . . . . . . . . .
$$ | . . . . . . . . . . .
$$ | . . . 3 . . . . . . .
$$ | . . . . . . . . . . .
$$ | . . 2 . 4 . . . . . .
$$ | . . . , . 5 . . . , .
$$ | . . . 1 . . . . . . .
$$ | . . . . . . . . . . .
$$ | . . . . . . . . . . .
$$ +----------------------

Click Here To Show Diagram Code

[go]$$W
$$ | . . . . . . . . . . .
$$ | . . . , . . . . . , .
$$ | . . . . . . . . . . .
$$ | . . . . . . . . . . .
$$ | . . . 3 . . . . . . .
$$ | . . . . . . . . . . .
$$ | . . 2 . 4 . . . . . .
$$ | . . . , . 5 . . . , .
$$ | . . . 1 . . . . . . .
$$ | . . . . . . . . . . .
$$ | . . . . . . . . . . .
$$ +----------------------[/go]

After this seqeunce of 5 moves I guess black can play tenuki. If it is true then is this five moves sequence a joseki?

Later in the game black may followed by :

$$W
$$ | . . . . . . . . . . .
$$ | . . . , . . . . . , .
$$ | . . . 6 . . . . . . .
$$ | . . . . . . . . . . .
$$ | . . . 3 . . . . . . .
$$ | . . . . . . . . . . .
$$ | . . 2 . 4 . . . . . .
$$ | . . . , . 5 . . . , .
$$ | . . . 1 . . . . . . .
$$ | . . . . . . . . . . .
$$ | . . . . . . . . . . .
$$ +----------------------

Click Here To Show Diagram Code

[go]$$W
$$ | . . . . . . . . . . .
$$ | . . . , . . . . . , .
$$ | . . . 6 . . . . . . .
$$ | . . . . . . . . . . .
$$ | . . . 3 . . . . . . .
$$ | . . . . . . . . . . .
$$ | . . 2 . 4 . . . . . .
$$ | . . . , . 5 . . . , .
$$ | . . . 1 . . . . . . .
$$ | . . . . . . . . . . .
$$ | . . . . . . . . . . .
$$ +----------------------[/go]

Do you consider to being another joseki or do you consider the move alone as joseki in such situation?

$$W
$$ | . . . . . . . . . . .
$$ | . . . , . . . . . , .
$$ | . . . 6 . . . . . . .
$$ | . . . . . . . . . . .
$$ | . . 8 3 . 7 . . . . .
$$ | . . 0 9 . . . . . . .
$$ | . . 2 . 4 . . . . . .
$$ | . . . , . 5 . . . , .
$$ | . . . 1 . . . . . . .
$$ | . . . . . . . . . . .
$$ | . . . . . . . . . . .
$$ +----------------------

Click Here To Show Diagram Code

[go]$$W
$$ | . . . . . . . . . . .
$$ | . . . , . . . . . , .
$$ | . . . 6 . . . . . . .
$$ | . . . . . . . . . . .
$$ | . . 8 3 . 7 . . . . .
$$ | . . 0 9 . . . . . . .
$$ | . . 2 . 4 . . . . . .
$$ | . . . , . 5 . . . , .
$$ | . . . 1 . . . . . . .
$$ | . . . . . . . . . . .
$$ | . . . . . . . . . . .
$$ +----------------------[/go]

Seeing the sequence above it may well happen that tenuki moves can be played after 5, after 6, after 7 ...
What do you call joseki?
What about a sequence of 20 or 30 moves?

In addition remenber that "joseki" could also give a local result which is not even but which can be justied by the environment like the joseki beginning by

$$W
$$ | . . . . . . . . . . .
$$ | . . . . . . . . . . .
$$ | . . 2 1 . . . . . , .
$$ | . . 3 . . . . . . . .
$$ | . . . . . . . . . . .
$$ | . . . . . . . . . . .
$$ +----------------------

Click Here To Show Diagram Code

[go]$$W
$$ | . . . . . . . . . . .
$$ | . . . . . . . . . . .
$$ | . . 2 1 . . . . . , .
$$ | . . 3 . . . . . . . .
$$ | . . . . . . . . . . .
$$ | . . . . . . . . . . .
$$ +----------------------[/go]

[RobertJasiek]

'Joseki' can be used with different meanings incl.
- standard sequence
- standard sequence that is correct in at least some positions
- standard sequence that is established by professional players' games and consensus of equality (with the implicit interpretation that different numbers of played stones are taken into account)
etc.

No problem. Just every medium should clarify its intended meaning.

What is your problem with plays elsewhere?! Just annotate them as such in the otherwise local sequence!

[xela]

Point 4: One must be careful not to misinterpret statistics! Typical mistakes: too few playouts, reversion statistics misunderstood, overlooking contradictions in statistics only dissolved by much deeper search.

Yes, one must be careful! The percentages and winrates shown by KataGo are not statistics in either the Bayesian or frequentist sense (*). At this stage of our knowledge, they are not much more than a heuristic. How many playouts are too few? You can not name an exact number! Whatever number of playouts you do, there could always be more. However deep the search is, there may still be a surprise lurking at the next level. There is no certainty to be found in this direction.

I don't wish to discourage you from exploring. I'm just cautioning you that interpretation of the metrics (not statistics) produced by AI is still some mixture of art and science. When I convey (my interpretation of) KataGo's results in approximate terms, this is not due to ignorance on my part. Quite the opposite: it is a deliberate strategy.

Have you heard the saying "Don't let the perfect be the enemy of the good"?

(*) One can in fact calculate a frequentist confidence interval around the winrate, based on the number of playouts. This is after all what guides the choice of nodes for each playout. So there is an underlying statistic -- but one that's not displayed on the screen in most GUIs, and in any case you need to be very careful indeed about what it's measuring. Discussions in this forum about "margin of error" are generally using the term in a colloquial sense.

[RobertJasiek]

As to how many playouts are enough, people with weak hardware might deceive themselves by misinterpreting early stability as sufficient condidence. If some almost correct move is good enough for a purpose, this can be fine but also result in wild AI play. If one wants correct moves (such as for finding josekis), forget about anything below 100,000 playouts because AI changes its preference too often until then and early ghosts frequently turn out to be real alternatives! So now, for joseki study, I presume 100,000+ per top candidate. At least for the top-most but hetter for each candidate seriously competing for the top.

One can be confident to have enough playouts if stability in top candidate selection has been reached for quite some time of increasing numbers of playouts. The more difficult the position and the closer the next candidates the longer search must proceed! Furthermore, the most difficult josekis have their best move appearing only after ca. 5 million playouts per top move! So for such josekis, use at least as many playouts per move!

There are rare cases in the most difficult josekis. A few such cases within half a year of full time AI study. Cases for which AI changes its mind after 10, 50 (so far once) or 90 (once) million moves for the top candidate. Or, we can imagine, after an eternity. One must be pragmatic but also cautious whenever it is difficult and the candidates are rather close in values. I was motivated to wait for the RAM to fill when previous values contradicted the decision-making. Luckily, my 64GB RAM has always been large enough to dissolve contradictions. 32GB would have been too small.

[Knotwilg]

The fact is that if you look at a bunch of AI self-play games, you'll see that certain patterns pop up more than you'd expect by chance. And many of those patterns, if they don't work in every full board context, at least work well more often than not. You might easily conclude that these patterns are significant and should be given a special name... Since AlphaGo Zero, Leela Zero and KataGo have come to similar conclusions here, I think it's proven that joseki are an intrinsic part of go. Yes, they need to be handled with caution. We already knew that.

I totally agree with xela here. Which doesn't mean I disagree with others per se, in some cases we might misunderstand each other. In my own schematic words:

1. pre-AI there were lots of joseki, established corner patterns, which give locally best play or acceptable variations, which need caution to apply in the context of the whole board
2. every decade or so, these joseki would get updates with the latest professional insights, some joseki going out of fashion, others coming back in, still others being new; the caution still applied
3. the past decade, there has been a major revision due to the advent of AI; some patterns have been discarded, others restored, many new have emerged; the caution still applies;
4. the fundamental difference pre- and post-AI is that we can now (statistically) quantify the results and can likewise quantify their value in a whole board context
5. contrary to a widespread misconception about the current state of AI - which I'm not accusing anyone in particular of - the evaluation remains largely statistical and does not portray one narrow path to victory, i.e. AI is not (yet) deterministic; the whole board mattes, but not (much) more than before AI; the difference in evaluation rarely goes beyond the margin of error and when it does, it is often easy to explain (like a ladder breaker); this is why, IMO, joseki still exist today

Knotwilg says that joseki exists sill today. I agree with that but I fear that we do not have all the same definition of joseki. OK we can simply say it is a standard sequence of moves in the corner (see https://senseis.xmp.net/?Joseki) but what means "standard" and what means "seqeunce". In particular how do you handle tenuki moves.

Let's take an example:

$$W
$$ | . . . . . . . . . . .
$$ | . . . , . . . . . , .
$$ | . . . . . . . . . . .
$$ | . . . . . . . . . . .
$$ | . . . 3 . . . . . . .
$$ | . . . . . . . . . . .
$$ | . . 2 . 4 . . . . . .
$$ | . . . , . 5 . . . , .
$$ | . . . 1 . . . . . . .
$$ | . . . . . . . . . . .
$$ | . . . . . . . . . . .
$$ +----------------------

Click Here To Show Diagram Code

[go]$$W
$$ | . . . . . . . . . . .
$$ | . . . , . . . . . , .
$$ | . . . . . . . . . . .
$$ | . . . . . . . . . . .
$$ | . . . 3 . . . . . . .
$$ | . . . . . . . . . . .
$$ | . . 2 . 4 . . . . . .
$$ | . . . , . 5 . . . , .
$$ | . . . 1 . . . . . . .
$$ | . . . . . . . . . . .
$$ | . . . . . . . . . . .
$$ +----------------------[/go]

After this seqeunce of 5 moves I guess black can play tenuki. If it is true then is this five moves sequence a joseki?

Later in the game black may followed by :

$$W
$$ | . . . . . . . . . . .
$$ | . . . , . . . . . , .
$$ | . . . 6 . . . . . . .
$$ | . . . . . . . . . . .
$$ | . . . 3 . . . . . . .
$$ | . . . . . . . . . . .
$$ | . . 2 . 4 . . . . . .
$$ | . . . , . 5 . . . , .
$$ | . . . 1 . . . . . . .
$$ | . . . . . . . . . . .
$$ | . . . . . . . . . . .
$$ +----------------------

Click Here To Show Diagram Code

[go]$$W
$$ | . . . . . . . . . . .
$$ | . . . , . . . . . , .
$$ | . . . 6 . . . . . . .
$$ | . . . . . . . . . . .
$$ | . . . 3 . . . . . . .
$$ | . . . . . . . . . . .
$$ | . . 2 . 4 . . . . . .
$$ | . . . , . 5 . . . , .
$$ | . . . 1 . . . . . . .
$$ | . . . . . . . . . . .
$$ | . . . . . . . . . . .
$$ +----------------------[/go]

Do you consider to being another joseki or do you consider the move alone as joseki in such situation?

$$W
$$ | . . . . . . . . . . .
$$ | . . . , . . . . . , .
$$ | . . . 6 . . . . . . .
$$ | . . . . . . . . . . .
$$ | . . 8 3 . 7 . . . . .
$$ | . . 0 9 . . . . . . .
$$ | . . 2 . 4 . . . . . .
$$ | . . . , . 5 . . . , .
$$ | . . . 1 . . . . . . .
$$ | . . . . . . . . . . .
$$ | . . . . . . . . . . .
$$ +----------------------

Click Here To Show Diagram Code

[go]$$W
$$ | . . . . . . . . . . .
$$ | . . . , . . . . . , .
$$ | . . . 6 . . . . . . .
$$ | . . . . . . . . . . .
$$ | . . 8 3 . 7 . . . . .
$$ | . . 0 9 . . . . . . .
$$ | . . 2 . 4 . . . . . .
$$ | . . . , . 5 . . . , .
$$ | . . . 1 . . . . . . .
$$ | . . . . . . . . . . .
$$ | . . . . . . . . . . .
$$ +----------------------[/go]

Seeing the sequence above it may well happen that tenuki moves can be played after 5, after 6, after 7 ...
What do you call joseki?
What about a sequence of 20 or 30 moves?

In addition remenber that "joseki" could also give a local result which is not even but which can be justied by the environment like the joseki beginning by

$$W
$$ | . . . . . . . . . . .
$$ | . . . . . . . . . . .
$$ | . . 2 1 . . . . . , .
$$ | . . 3 . . . . . . . .
$$ | . . . . . . . . . . .
$$ | . . . . . . . . . . .
$$ +----------------------

Click Here To Show Diagram Code

[go]$$W
$$ | . . . . . . . . . . .
$$ | . . . . . . . . . . .
$$ | . . 2 1 . . . . . , .
$$ | . . 3 . . . . . . . .
$$ | . . . . . . . . . . .
$$ | . . . . . . . . . . .
$$ +----------------------[/go]

It's an interesting question.

My answer will be too vague for some, but I consider "joseki" to be a set of variations on a certain starting pattern, where one or more are considered to be "best for both" assuming alternate play, others are probably still acceptable and other variations are considered mistakes. Aspects for evaluation include: territory, influence, "strength" (of that influence), aji and sente. Tenuki variations are a little different: if one player has invested two stones more than the opponent, the result is expected to be lopsided but exactly how lopsided is not easy to judge. Since one player decided to play elsewhere the whole board is probably more relevant to the evaluation.

There is a notion of a significant drop in the local temperature by the end of what is considered joseki, so playing elsewhere at that point is more likely. We usually speak of "follow-ups" from that point onwards. With AI we have seen that tenuki can happen very early (depending ...). Since local play can resume later, often as if the play elsewhere was only incidental, I consider that to be one and the same joseki (variation). It is useful to mark the moves where tenuki is more likely / less harmful.

In some cases tenuki results in the position reverting to another joseki. A trivial example is 3-5, 4-3 tenuki = 3-4, 5-3 continued.

[kvasir]

Knotwilg says that joseki exists sill today. I agree with that but I fear that we do not have all the same definition of joseki. OK we can simply say it is a standard sequence of moves in the corner (see https://senseis.xmp.net/?Joseki) but what means "standard" and what means "seqeunce". In particular how do you handle tenuki moves.

It can be unclear but the word "joseki" at least is more specific than "standard sequence". It is also one less syllable and lot fewer stop sounds (when airflow in the mouth is completely obstructed), it is therefore easier to say. It also allows for "non-standard joseki", "joseki sequence" and "non-standard joseki sequence", were as "non-standard standard sequence sequence" would be unusable.

Maybe joseki only means the kind of diagrams you'd find in a joseki book?

That said, you are right about the tenuki. The tenuki at , and are all common. I believe pros have played away at most of the time since this joseki became common, there are other moves for white here. Tenuki at is now common. I believe tenuki at some of the other moves are rare in pro games, for the reason that they tend to anticipate it and time their play to make it more awkward to tenuki. Pros are also averse to aji-keshi. Then is one place where it is possible to tenuki, white has two continuations here, which are less urgent than the previous moves. I'd expect pros to try to time their move here while considering the possibility that is tenuki.

$$Bm6 Simple but lot of tenuki in the way of reaching this diagram.
$$ | . . . . . . . . . . .
$$ | . . . , . . . . . , .
$$ | . . . 1 . . . . . . .
$$ | . . . . . . . . . . .
$$ | . . 3 O . 2 . . . . .
$$ | . . 5 4 . . . . . . .
$$ | . . X 6 X . . . . . .
$$ | . . 7 8 . O . . . , .
$$ | . . . O . . . . . . .
$$ | . . . . . . . . . . .
$$ | . . . . . . . . . . .
$$ +----------------------

Click Here To Show Diagram Code

[go]$$Bm6 Simple but lot of tenuki in the way of reaching this diagram.
$$ | . . . . . . . . . . .
$$ | . . . , . . . . . , .
$$ | . . . 1 . . . . . . .
$$ | . . . . . . . . . . .
$$ | . . 3 O . 2 . . . . .
$$ | . . 5 4 . . . . . . .
$$ | . . X 6 X . . . . . .
$$ | . . 7 8 . O . . . , .
$$ | . . . O . . . . . . .
$$ | . . . . . . . . . . .
$$ | . . . . . . . . . . .
$$ +----------------------[/go]

Variations like the one in the OP would seem farfetched for some players and reasonable to others, dependent on if they are likely to follow the moves up to that point. It is a lot of moves in this case but it is the kind of thing I'd expect to find in multi-volume joseki books.

[Gérard TAILLE]

There is a notion of a significant drop in the local temperature by the end of what is considered joseki, so playing elsewhere at that point is more likely. We usually speak of "follow-ups" from that point onwards. With AI we have seen that tenuki can happen very early (depending ...). Since local play can resume later, often as if the play elsewhere was only incidental, I consider that to be one and the same joseki (variation). It is useful to mark the moves where tenuki is more likely / less harmful.

In some cases tenuki results in the position reverting to another joseki. A trivial example is 3-5, 4-3 tenuki = 3-4, 5-3 continued.

Interesting.
Assume a joseki begin at temperature T on the board. Does it make sense to say:
1) after the last move of the joseki the local temperature is not higher than T
2) after each other moves of the joseki the local temperature cannot be smaller than T
BTW can we consider that in the first phase of the game the temperature of the all board is always higher than 12?

[RobertJasiek]

Not always. Mistakes can raise the temperature.