Life In 19x19

As someone who grew up with the elementary series, I've long wanted to review the ideas we got there, with the new insights provided by the bots. In particular I wanted to review what is probably my favorite book of all Go books: Attack and Defense, by Akira Ishida and James Davies.

One of the major insights we got from the bots is that, when we attack, we should always have territory in mind. This idea is not new: "attack for profit" has always been the guideline. The difference between then and now is that "profit" in the old days could also be central influence. Attacking from above, capping the group under pressure, was often deemed a good idea. The bots prefer attacking sideways by and large, as this gives immediate profit in the form of territory, rather than future profit in the form of influence. This goes along with the overall shift the bots made in the balance between influence and territory, i.e. more territory oriented. The concept of "territory" is not only about surrounding points, but also and perhaps more importantly about making groups safe. We often see in the analysis that a variation which forces an opponent group into safety, is also considered worse than we may have thought before.

I believe I got my instincts to attack from above largely from the book "Attack & Defense". So let's review some of the diagrams in there. In order not to violate the intellectual property, at least not in spirit, I'll take material that is available as sample pages on Amazon.

Analysis 1 - attacking wholesale or peacemeal

Analysis 2 - How to attack for profit?

Analysis 3 - Attack or defend? (when to defend)

Analysis 4 - sacrifice tactics / driving / what if you're not happy with standard sequences?

I believe I got my instincts to attack from above largely from the book "Attack & Defense".

I share your instincts, acquired not from that book specifically but from being surrounded by western players. But I long ago came to believe that those instincts were wrong in many respects, and also that the source of the problem was other western players (including, perhaps, as in A&D, a western author), and more specifically western languages. My belief has been the source of constantly harping on about the meaning of thickness

I have recently become even more convinced that this mismatch in language is the root of slow progress in the west.

It's a subtle thing and I haven't got either the time or the inclination to make my case rigorously, but anecdotally, I can easily explain why my conviction has grown.

I introduced what I called Go Wisdom into my book Genjo-Chitoku. This was an appendix covering all aspects of go theory so designed as to enable the reader to think about the game commentaries in a more expansive and self-controlled way. Although the text discussing the theory topics may seem to be the main element, I found that the indexed references to each topic in the actual (pro-based) commentaries were in fact the most significant aspect for me. I was rather shocked to see how often certain words came up in the Japanese-based sources that were words that exist in an English form but are used by us much less - and vice versa!

To give two examples that relate specifically to your post: (1) The concept of bases (and, as part of the same nexus, settling) is much, much more common in Japanese. Tsume/checks (which are, after all, in part base-making or moyo-making moves are very, very common in Japanese, but rarely highlighted by westerners. (2) Attack is very, very common in English, but in Japanese is less so, and when it does occur it is in a much nuanced form. As I have mentioned before on L19, the Japanese tend rather to refer to specific kinds of fighting: torikake, seriai, karami, motare, semedori, ijime, etc - each with its own set of often vastly different parameters. 'Invasion' is likewise 'too' common in English, and the Japanese meaning is different anyway.

I first noticed this in Genjo-Chitoku, which had about 90 commented games, but I have now noticed it with knobs on in Games of Shuei (now being proof-read) which adds 133 commented games to the databank. And also more games in Go vs Iwamoto in the forthcoming book on their eleven-game match.

It will, therefore, presumably come as no surprise that I also find that the way the new insights from AI are being talked about differs strongly according to whether you are a Japanese pro or a western amateur, although I'm sure both are useful. FWIW things like insights about e.g. attacking sideways don't seem to come up in what I've read in Japanese books/mags on AI. There the focus seems to be predominantly on overconcentration and efficiency. Obviously it helps if you already have a high standard of efficiency, as pros do, and so western amateurs probably do have to start somewhere else. But I feel sure deeper consideration of the languages of go should be part of the package.

Incidentally, in your second example, where Lizzie recommends a magari, to someone who has seen Shuei's predilection for L shapes that came as no surprise.

John Fairbairn wrote:(...)
I have recently become even more convinced that this mismatch in language is the root of slow progress in the west.

(...)

It will, therefore, presumably come as no surprise that I also find that the way the new insights from AI are being talked about differs strongly according to whether you are a Japanese pro or a western amateur

A disclaimer first: despite being around for a long time (probably also because of it) I'm still a weak amateur by most standards. So whatever I draw from analyzing pro examples with Lizzie, is bound to be limited by my understanding of the game.

From earlier posts, I know I don't share your confidence in the power of language - if I may even quote you like that. I absolutely love language, for debate, to write essays, songs, ... but in the acquisition of expert knowledge it can become a hindrance rather than a device. Maybe the Japanese Go language has been developed more and better to grasp the expert knowledge than the Western language, perhaps even better equipped to do so altogether - but witnessing pros analyzing post mortem, seeing how little they talk and instead show sequences all the time; and witnessing LZ who doesn't talk but through sequences, I'm less convinced than you are that it matters which language you use to convey the ideas.

The concept of bases (and, as part of the same nexus, settling) is much, much more common in Japanese. Tsume/checks (which are, after all, in part base-making or moyo-making moves are very, very common in Japanese, but rarely highlighted by westerners.

I can't judge the difference but I'd say that the concept of making a base and the importance of doing so, is rather abundant in English go commentaries.

Attack is very, very common in English, but in Japanese is less so, and when it does occur it is in a much nuanced form. As I have mentioned before on L19, the Japanese tend rather to refer to specific kinds of fighting: torikake, seriai, karami, motare, semedori, ijime, etc - each with its own set of often vastly different parameters. 'Invasion' is likewise 'too' common in English, and the Japanese meaning is different anyway.

Here I agree that "attack" is both abundant and not very nuanced in English commentaries. And our bewilderment at the bots' preference for sideways attacks may point at underdeveloped conceptualization in English go theory of "attack".

FWIW things like insights about e.g. attacking sideways don't seem to come up in what I've read in Japanese books/mags on AI. There the focus seems to be predominantly on overconcentration and efficiency. Obviously it helps if you already have a high standard of efficiency, as pros do, and so western amateurs probably do have to start somewhere else. But I feel sure deeper consideration of the languages of go should be part of the package.

I do sense that notion of efficiency in many of the analyses done using bots, like Bill's thread, but I'm not confident writing about it.

Incidentally, in your second example, where Lizzie recommends a magari, to someone who has seen Shuei's predilection for L shapes that came as no surprise.

It did surprise me, because it was a response to a capping move, which was right where I'd want to play 'after' the magari. LZ can leave shapes behind which she thinks are resilient and perhaps Shuei saw things in the same manner, i.e. that no plausible sequences are effective against this kind of shape.

Analysis 5 - Problem 1 at the end of the book

Analysis 6 - Problem 20 of the book

John Fairbairn wrote:I introduced what I called Go Wisdom into my book Genjo-Chitoku. This was an appendix covering all aspects of go theory so designed as to enable the reader to think about the game commentaries in a more expansive and self-controlled way.

I confess that I have not read Genjo-Chitoku, but I am sure that the book and the Go Wisdom appendix are very valuable resources. Too many Western amateurs are almost totally unaware of go theory.

John Fairbairn wrote:To give two examples that relate specifically to your post: (1) The concept of bases (and, as part of the same nexus, settling) is much, much more common in Japanese. Tsume/checks (which are, after all, in part base-making or moyo-making moves are very, very common in Japanese, but rarely highlighted by westerners.

Ancient game records indicate that the concept of a base is one of the oldest ideas in go. In the AI era, however, it appears that that is a concept we are going to have to reconsider. For instance, in the 3-4, high approach joseki with an underneath attachment, after the approacher makes a solid connection and the 3-4 player extends one space on the side, the bots typically omit making a base, i.e., an extension from the connected stones. Before AlphaGo, however, I have seen that omission, except perhaps after a sente exchange, only once in game records of top level play, by Shusaku, IIRC. Segoe pointed out in the 20th century that that joseki was already suspect. But the remedy of human players was either to delay making the high approach, to accept a small loss, perhaps, or to omit the connection. It did not include making the solid connection and then omitting an extension. (There are other possible examples, but this one is obvious and well known.) In the AI era humans can devalue the idea of a base or modify the concept.

It will, therefore, presumably come as no surprise that I also find that the way the new insights from AI are being talked about differs strongly according to whether you are a Japanese pro or a western amateur, although I'm sure both are useful. FWIW things like insights about e.g. attacking sideways don't seem to come up in what I've read in Japanese books/mags on AI. There the focus seems to be predominantly on overconcentration and efficiency. Obviously it helps if you already have a high standard of efficiency, as pros do, and so western amateurs probably do have to start somewhere else. But I feel sure deeper consideration of the languages of go should be part of the package.

The sideways attack, by which I mean approaching the group being attacked from the side, but not as a pincer and not from the side and slightly ahead, is something I picked up from Okigo Jizai, which is about 200 years old. Hattori did not talk about a sideways attack in so many words, but such attacks appear often in his examples. In fact, I don't know of any go author who talks about sideways attacks.

As for how Western amateurs talk about AI play, I can only speak for myself. Western amateurs may have scientific and mathematical skills that can deepen our understanding of go. For instance, 25 years ago Howard Landman published "Eyespace values in go", which extended the concept of fractional eyes from the half eye, which was already well known among Eastern pros, to other values, such as the ¾ eye. Howard provided many examples from Maeda's tsumego problems. In particular, ¾ eye + ¾ eye + ½ eye = 2 eyes and life. Since then, has any pro talked about the ¾ eye? The same year saw the publication of Berlekamp and Wolfe's Mathematical Go, which deepened our understanding of the last play (tedomari). The Japanese translation sold out in a day or two. Since then, has any pro talked about Ups or Downs? I know when I submitted my article about getting the last play in a well known ancient game, showing a line of play that had apparently been overlooked by pro commentators for centuries, to the new magazine Myosu a few years ago, I met opposition from one of their editors, a strong Western amateur, who thought I was committing heresy. Also in 1994, Berlekamp introduced his idea of komaster. I talked about it on rec.games.go and Sensei's Library. I gather from your writings that pros do not use the idea of komaster to calculate the values of approach kos, 10,000 years kos, and the like, and consequently often get those values wrong. I spoke about it at ICOB 4, where we met. That was in 2006, right? Has any pro taken up the idea of komaster since then?

As for AI play, I believe that I was one of the first to point out that AlphaGo pincered about half as often as his human opponents. Having noticed that tendency in AlphaGo, it was easy for me to compile the statistics to show it more precisely. I also observed that, with one exception, AlphaGo's pincers were supported by a third stone on the other side. That limited the ability of the pincered stone to jump out and threaten to attack either side. Not a deep insight, but a start, I believe. These days, several people have pointed out that the human concept of direction of play has little, if any, effect on AI play. I think that there is something there, but have not done enough research to demonstrate anything. Western amateurs have also noticed that AI devalues the sides. In his 21st Century Go series, Go Seigen had started to do that, at least in the case of the splitting play, but was ambivalent about the idea. There is a traditional go concept of the last play of the opening (fuseki no tedomari). I have not noticed that in AI play (but I haven't looked very hard). However, in the Elf commentaries on GoGoD games I believe I have discovered a different tedomari, occupying the last empty corner. OC, that is not a problem in many parallel openings, but in komi games top level human players often fail to occupy the last empty corner when they have the chance. Again, not a deep insight, but a start.

Knotwilg wrote:Analysis 5 - Problem 1 at the end of the book

Do we know what game this is? I suspect it is one that has a pro commentary. Elf may have something to say about it.

Hmmm. I think the book was written before Wilcox came up with the idea of cutting sector lines, which Lizzie's play does.

Bill Spight wrote:
Hmmm. I think the book was written before Wilcox came up with the idea of cutting sector lines, which Lizzie's play does.

On the contrary, the book discussEs junction lines early on. I was surprised to see that Wilcox’s idea is already in there

Bill Spight wrote:Do we know what game this is? I suspect it is one that has a pro commentary. Elf may have something to say about it.

It's in GoGoD: Ishida Akira 7d - Kanno Kiyonori 6d, 20th Prime Minister's Cup, Quarter-final, 1976-07-15

KataGo wants to jump out at O10, or alternatively play a shoulder hit against the top left enclosure, or play one of several moves against the lower right. It does not consider the "right answer" and evaluates it as maybe 2.5% worse when it is put on the board.

bernds wrote:

Bill Spight wrote:Do we know what game this is? I suspect it is one that has a pro commentary. Elf may have something to say about it.

It's in GoGoD: Ishida Akira 7d - Kanno Kiyonori 6d, 20th Prime Minister's Cup, Quarter-final, 1976-07-15

KataGo wants to jump out at O10, or alternatively play a shoulder hit against the top left enclosure, or play one of several moves against the lower right. It does not consider the "right answer" and evaluates it as maybe 2.5% worse when it is put on the board.

Thanks.

Click Here To Show Diagram Code

[go]$$Bcm25 Game play, Black 25, losing around 2% according to Elf
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . . . . . O . . . X . . X . . |
$$ | . . . , . . . . . , . . . . . X . O . |
$$ | . . O . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . O . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . O . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . 2 . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . O . . . |
$$ | . . . X . . . . . . 1 . O . O . O . . |
$$ | . . . . . . . . . . . . . . X O X X . |
$$ | . . O , X . . . . , . X . . X , . . . |
$$ | . . . . a . . . . . . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]

was barely on Elf's radar, with 0.1k playouts, so it inherited its winrate estimate from Elf's choice for , as shown, with 23k playouts. White's actual play was at a.

Click Here To Show Diagram Code

[go]$$Bcm25 Elf's second choice
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . . . . . O . . . X . . X . . |
$$ | . . . , . . . . . , . . . . . X . O . |
$$ | . . O . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . O . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . O . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . O . . . |
$$ | . . . X . . . . . . . . O . O . O . . |
$$ | . 1 . . . . . . . . . . . . X O X X . |
$$ | . . O , X . . . . , . X . . X , . . . |
$$ | . . . . . . . . . . . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]

Elf's second choice for had 4.1k playouts and lost nothing by comparison with Elf's top choice. I expect that lesser choices, except for the game play, were not shown because they each got fewer than 1.5k playouts.

Click Here To Show Diagram Code

[go]$$Bcm25 Elf's top choice
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . . . . . O . . . X . . X . . |
$$ | . . . , . . . . . , . . . . . X . O . |
$$ | . . O . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . O . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . O . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . 1 . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . O . . . |
$$ | . . . X . . . . . . . . O . O . O . . |
$$ | . . . . . . . . . . . . . . X O X X . |
$$ | . . O , X . . . . , . X . . X , . . . |
$$ | . . . . . . . . . . . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]

Elf's top choice, the one space jump, breaking sector lines, had 27.2k playouts.

Knotwilg wrote:Analysis 4 - sacrifice tactics / driving / what if you're not happy with standard sequences?

This is the game: http://ps.waltheri.net/database/game/1444/
It is interesting that this is in a book by Ishida Akira because the idea was refuted by his opponent. White looks silly because he played away from a shoulder hit, three times!
Is there no mention of this in the book?

It looks like a good example of attacking because the attacking move makes a strong group weak.

Click Here To Show Diagram Code

[go]$$W The actual game
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . X . . . |
$$ | . . . . . . . . . . . . . . X O O O . |
$$ | . . O O X . . . X . . . . . . X O . O |
$$ | . . . X X . . . . , . . . X . X X O . |
$$ | . O . . . . . . . . . . . . . . . X . |
$$ | . . . O . . . . . . . . . . . . . . . |
$$ | . 6 . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . 4 X . . . . . , . . . . . , O . . |
$$ | . . O . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . O . . |
$$ | . . . 1 . 5 . . . . . . . . . . . . . |
$$ | . . X . 3 X . . . . . . . . . . . . . |
$$ | . . . X X O O . . , . . . . . , X . . |
$$ | . X X O O . . O . . . . . . X . . . . |
$$ | . 2 O . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]

It would be curious to see what the AI think of the counter-attack.