After a fair bit of time I am pleased to report that I have worked my way through John's Games of Shuei. Each of Shuei's magical moves have been lovingly placed on a physical go board and duly considered. Not sure if it counts as effortful practice, but there was certainly some effort involved.

Some thoughts and observations:

1. Shuei was strong. A trite observation, to be sure, but Shuei gave handicaps against established professionals and made the wins look effortless. It's worth studying those games just to try and glimpse the technique of playing white in ahandicap game against strong players. Anyone who drives the entire world to a handicap is worthy of a second look.

2. Eight points is enough to live. Not exactly, but Michael Redmond mentioned this rule of thumb in a video I saw recently. The general idea is that if a group has eight empty intersections under its control, it's enough to allow tenuki, even if the boundaries aren't completely firm. It did a nice job of explaining a number of tenuki plays that weren't quite obvious to me at first.

3. Pay attention to the middle. There are corners, edges and the centre, but also a middle area between them, maybe the fifth line to the seventh or eighth line. Interesting stuff happens there related to the strength of groups and there can be a surprising number of points that strong players accumulate in those areas. This also seems to be true with AlphaGo's games in its 60-0 online series--it's actually possible to make a reasonable guess which is the computer and which is the human just by looking at the final positions in those middle regions. AlphaGo seemed to pick up a surprising amount of territory in those regions.

4. There is strength in gote. We all know sente is important, but it's interesting to watch when the pros are willing to take gote, usually to play building moves. I remember a number of Chinese players years ago using the phrase "building a house". At the time I appreciated the idea, but I have come to appreciate the philosophy on a deeper note recently.

5. I still don't understand amashi. I know the textbook definition, but I don't really understand it even after seeing examples from the Master.

Next project is to look through AlphaGo's games. I am particularly looking forward to the AlphaGo vs the world series.

Nice!

What you call the middle, I can relate to "Go Seigen groups", a concept introduced by John (again), due to Go Seigen writing & playing in this sense.

Although Go Seigen groups do "relate" to Phil's concept of the "middle" they have their base on the centre part of the sides and so are really distinct, I think. I would suggest that Go got this locus from his study of old Chinese games, where there was much theoretical debate about the 6-3 then the 7-3 then 8-3 and finally the 9-3 points. Much of this is still valid today, despite group tax, because of Go's insights but also because it sheds light on AI moves.

Glimmerings of AI thinking can also be seen in old Chinese games in the area Phil is specifically mentioning, namely the 5th to 8th lines. In particular you will notice that old Chinese commentaries talk much about 收 which (like yose) gets wrongly mixed up with the endgame - it rarely even gets mentioned in the final stages of the game. It refers to boundary plays, and its constant mention in commentaries can be attributed to the old masters stressing the need to seal off boundaries in the "middle area" above all, and early - just like AI. It is not a mere matter of counting or just remembering to think about such moves - it is more (like AI) about strategic ways to seal off the boundaries. In my large corpus of old Chinese go commentaries, 收 appears in 50th place for frequency, WAY above words like territory, ko, life, eyes, attack, invasion.

I also like to recall the story by Go Seigen's doctor that whenever he visited him, Go had a go board out with rather few stones on it - and they always seemed to be in the middle of the board, he said.

A related concept that here is almost entirely overlooked in modern Japanese go is kakoi (surrounding moves). They occur usually in this "middle" area.

As a PS, I feel I should point out that old Chinese games also play in this "middle" area for another reason - lighthouse moves (zhaoying 照应). These also occur in commentaries with very high frequency - the same frequency as moyos as it happens. They are probably more important in old Chinese go than in modern go because they are related to group tax and the need to make long-range connections. But long-range connections are vital in AI go (just as they were in the GO Seigen version of Shin Fuseki). So it would foolish to overlook old Chinese games just because of group tax. Maybe we should be talking about Huang Longshi groups now!

Regardless of the meaning of this kanji (I cannot read it), a higher frequency for a particular term than territory, ko, life, eyes, attack, invasion does not imply their relatively lower importance. It might as well be that the old Chinese sources consider talking about more basic concepts less interesting. Having studied with AI for several hours every day for a year, these basics are, however, more important than I expected. E.g., AI always tries to assess life status correctly and very many of its moves are (also) chosen to emphasise the territory balance.



What do mean by "long-range connections"? Connections, haengma or something else? Depending on your clarification, I might agree or disagree to "are vital in AI go".

To make such a statement on what is vital for AI, on how much of your own study with AI do you rely it and on what confidence measured by numbers of playouts (or visits) for the second-top-most move candidates do you rely?

I'm not entirely sure the concept is a Go Seigen group, although I'd take your word for it if it was. The 'middle' or perhaps 'in-between' areas seem to figure differently into the games of strong players. I have sensed it with most pros, but it is more acute with the ultra-strong players like Shuei and AlphaGo.

I noticed it when I was flipping through printed copies of AlphaGo's 60-game match against online players. My copies are printed in only a single diagram, so you get a sense of the full board position at the end of the game. While flipping through the positions it struck me that the game records looked different than ones I'm used to seeing. When I started looking more, I realized that the in-between areas were qualitatively different, and that it was possible to make a good guess whether AlphaGo was playing black or white largely by looking at those regions. Here's a few examples:

Click Here To Show Diagram Code

[go]$$ End of game. Observe the top right 'in-between'.
$$ ----------------
$$ | . . O O . . . . . . . . . . . . . . . |
$$ | . O . . O . . . . . . O . . X . . . . |
$$ | . O . O X . O . . O O O X X . O O . . |
$$ | . O . O . . . . . , X X . . X X . O . |
$$ | . X O O X . . . O . . . . . O X X O . |
$$ | X . X X . X . . . O X . . . . X O O . |
$$ | . X . . . . . . . O X . . . X O X O . |
$$ | . X O . X . X . . O O X . . X O X . . |
$$ | . X O X . X X O . X X . . . . O O O . |
$$ | . O . O X O X . . , . . . . X , . . . |
$$ | . . . O O O . . O . X . . X . . O X . |
$$ | . O . O X X . . . . . . X . X . X . . |
$$ | . O X X O X . . O . . X . X O X . . . |
$$ | . O O X O X . . . O X X X O . . . . . |
$$ | . X O O O X . . . . O X O O . . O . . |
$$ | X . X O . X O . O O . O X X O O X X . |
$$ | . X X . O X . . . . O O O X X X . . . |
$$ | . . . X . X . O . . . O X X . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ -------------------[/go]

Click Here To Show Diagram Code

[go]$$ End of game. Observe the top left 'in-between'
$$ ---------------
$$ | . . . . . . . . . . . . O . O . O . . |
$$ | . . . . . . . . . . . . O . O O X O . |
$$ | . . . O O O . O O X . . O . O X X X . |
$$ | . . . , X X X . . X . O . O X X . . . |
$$ | . . X . . . . . X . . . O X . . . . . |
$$ | . . . . . . . . X O . . . O . . X . . |
$$ | . . . . X . . . . . . . . . O . X O . |
$$ | . . . . O X . . X . . . . . . X O O . |
$$ | . . O O X X . . . X . . . . . . . . . |
$$ | . . O , O . O . . , . O . . . , . . . |
$$ | . X X O O . O . O O . . . . . . O . . |
$$ | . . X X . X . X . X O O . . . O . . . |
$$ | . O O X . X . X . X X . . . X X O . . |
$$ | . . O O O O O . . X O O O . . . . . . |
$$ | . X O O . . O . O X X . . . . X . O . |
$$ | . O X O O O X . . X . . X . . X O . . |
$$ | . X X X X O X . . . . . . . . X . . . |
$$ | . . . X O X X . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ -------------[/go]

Click Here To Show Diagram Code

[go]$$ End of game. Observe the top left 'in-between'
$$ -----------------------------
$$ | . . . . O . . O . . . . X O O O X . . |
$$ | . . O O X X X O O . . . . O O X X . . |
$$ | . . O X . . X X O O . . O O X X . . . |
$$ | . . O X . . . . X O O . X X O , X . . |
$$ | . O O O X X . . X O X O O O . O X . . |
$$ | O O X X O . . . X O X O . O X O X X X |
$$ | O X . . X . . . . X X O O X . X X O O |
$$ | X . X X O X . . . X X . O X X X X X O |
$$ | . X O O O X . X X X X O X . . O X O O |
$$ | . X X O O O X X O X O . X X X O X O O |
$$ | . X O O O X X O O O O O . O X X O O . |
$$ | . X O X O O O O . O X O O O X O X O . |
$$ | . . X X X O . O X X X X X O X O . . . |
$$ | . X X X X X X O X . . X O O O . O . . |
$$ | X X O X O O O X . . . . X X O . . . . |
$$ | X O O O . O X . X X . . X O O O . . . |
$$ | O O . . . O X X X . . . X X O X O O . |
$$ | . . . . O O O X . . . . X . X X X O . |
$$ | . . . . . O X X . . . . . X X O O O . |
$$ -------------------------------[/go]

Click Here To Show Diagram Code

[go]$$ End of game. Observe the bottom right 'in-between'
$$ --------------------------------------
$$ | . . . . . . . O . X . . X O . O X . . |
$$ | . . . X O O O . O O X X X X O X X . . |
$$ | . . . X X . O . O X X O X O O . . . . |
$$ | . . . , . . O O X O O O O X . X . . . |
$$ | . . X X . . O X X . X O . . . . . . . |
$$ | . . . O . O . . . . X X O . X X X . . |
$$ | . X . . . . . . . . . O O O X O . X . |
$$ | . . . X . . X . X X . O X O O . X O . |
$$ | . X . . O . . . . O . O X O . X X O . |
$$ | . O O O X X . X X O . O X X X X O X . |
$$ | . . O X . . . X O . . . O O X O O X O |
$$ | . O X X . X . . O . . . . X O X O O . |
$$ | . . O O X . . . . . . . . . O X . . . |
$$ | . . . . O X . O O . . . . . O . O . . |
$$ | . . . . . . O . X O O . . . . . . O . |
$$ | . . O O . X . . O X O . O . . O O . O |
$$ | . . O . O O X X X X X X O . . . X O . |
$$ | . . . O X X . . . . X O . O X X . X X |
$$ | . . . . O . X . . . . . O . X . . . . |
$$ -------------------------------------[/go]

Compare to a random human game:

Click Here To Show Diagram Code

[go]$$ End of game.
$$ ---------------------------------------
$$ | . . . . . O . . . . . . . . . . . . . |
$$ | . . . X O X O . . . . . . X . X O X . |
$$ | . . X X X X O . O O . O O O O X . X . |
$$ | . . . O . X X O O , O X X . O O X . . |
$$ | . O O X . X . X O O X . . . O X . . . |
$$ | . X X O O X X O X . X . X O O X . . . |
$$ | . X O . O O O O X . X . . O X X . X . |
$$ | . . X O . O X X X O . . . O X O O O . |
$$ | . . X X O . O O . . X X X O O X X X . |
$$ | . . . O . . O . X X . . O X X , X . X |
$$ | . . O O O O O X . O X . O . . X . X . |
$$ | . . X X . O O O . O . X O . X . . . X |
$$ | . . . X . . O X . . . . O O O X X X . |
$$ | . . X O X . O X O O O . O . X O O O . |
$$ | . X O O . O O X . X O X X O O . . . . |
$$ | . X X O . . X . X X O . . O X X X X . |
$$ | . X O O . . X . X O O X X X O O O X . |
$$ | . . . . . . X X X O . X O O . O . O . |
$$ | . . . . . . . . X . . . . . . . . . . |
$$ ---------------------------------------[/go]

In each case, AlphaGo has accumulated nontrivial points in these in-between regions, and its position in these regions look qualitatively different than those of human pros. Human games tend to fill those areas in with strings of stones. There might be a few eyes made in those area, but generally there aren't a lot of hefty territories. In AlphaGo's case, its territorial boundaries extend much further in the centre (i.e., the in-between) than we would normally expect, but it also uses those area for safety. See the second position, for example, where that area forms the critical eyespace for the Black group. This would seem to line up with John's comment about extending boundaries into the centre.

Now in Shuei's case, he wasn't building territories quite like AlphaGo. But he did seem to play differently in these in-between regions than many pros. He certainly picked up points there, and the regions often turned out to be critical for group safety. I'm afraid my insights are like trying peer through a dense fog--it seems like there's something there, but it's tough to elucidate. Perhaps it has something to do with the fact that many of his games were handicapped, but I have to wonder if the ultra-strong players--the Meijins--have a deeper understanding of those areas than a lesser pro. Certainly I'm paying more attention to those areas now.

If some numerically inclined soul wanted to make an effort, it would be possible to load a series of SGF files and determine how often a given intersection on the board became territory (or should become territory, in the case of a resignation). There may be a difference in the way the AIs accumulate points, particularly the location of those points.

When I was an influence and center player, I surprised quite a few opponents with big center regions appearing seemingly out of nowhere. Such is possible by first creating options and second, if one has options for enough sides of an imagined region or one of several alternative imagined regions, building walls on two or more sides of the region. This uses sente, alternativ strategic options or the opponent's strategic acceptance when he thinks that his side territories would be sufficient. If the opponent does not expect my imagined region to appear, he may have thought to have won the game but then finds that he has to try winning the game again. The stronger the opponents the better they can reduce regions so eventually I have become a more territory-orientated player. Against weaker players, my moyo strategies often work though. I might resurrect my influence style against stronger players some time but predicting their best reductions and alternative exchanges before building center moyos will be mandatory for succeeding. I have not studied Shuei, KataGo etc. for their center region strategies yet but would expect to find such planning in time.

Phil: I wouldn't call what you spotted Go Seigen groups either, but of course they are related (and, in some unusual ways, the relationship is worth thinking about).

Since this thread is about Shuei, I think it is worth noting also his reputation for his famous L groups in the centre (i.e. based in the 5th to 8th line areas and facing the centre). These are often mentioned as an example of thickness (in Japanese). However, it is thickness of a different type from what some people still wrongly call influence (influence is one attribute of early thickness, but just one). Unfortunately, the Japanese use the same term (atsui) for both early and late thickness, except that for late thickness they never say atsumi (the -mi ending can be said to be what adds the attribute of influence to thickness). Late thickness is covered by what the Chinese call shou - and also by Shuei's L groups. From those L shapes you can see why the Japanese use the word 'thick' but for the western audience I think early and late thickness are best compartmentalised.

The fundamental meaning of shou 收, especially in its medieval senses which relate to go, start with 'bring in the harvest' and extend to gathering and then to binding up and completing (rather like yose as it happens!).

The (old) approach to go theory in Japan and China is affected in many unexpected ways because of group tax. For example, as regards strategic preparations for boundary plays in the central region we are talking about, the Japanese will often focus on kakoi (early surrounding moves) which are typically gote. But strategic preparations for shou in old Chinese games are often linked with tactics for keeping sente, and I assume this is because keeping sente is one way of (almost) ensuring long-range connections, which is such a central feature of group-tax games. I would hazard guess that the old no-komi players were so used to relying on safe plays as Black that they were often not as squeamish about accepting gote as Chinese players, and that it is possible that this attitude still lingers on on the probabilistic style characteristic still of many modern Japanese players. Maybe that's why the AI style of play seems to have been more of a shock for them than for other oriental players.

I did once try to come up with a name for the central area in question, mainly because of becoming familiar (obviously!) with Shuei. I felt that just saying 'centre' was inadequate, and I infer that Phil had the same reservations since he specifically mentions the 5th to 8th lines. That concurred with my own guess except that I was always tempted to include the 4th line, and that was confirmed in my mind when I saw all the 4th line shoulder hits AI made. I was also influenced by the many pros who would try very early moves in this area (people like Yoda) but would never play on the centre point. There were exceptions such as Yamashita, but I had a feeling from the comments the players made that they were looking for central moves that were not too far away from the main action. The centre point was just too far away and so a lot easier for the opponent to skirt round it.

I remember thinking of this area around the centre point as a belt (for some cosmological reason I've forgotten) and the term beltway drifted into my mind because of trips to Washington and the like. But I am not an American and I never really absorbed the nuances of that word in politics (which was what concerned me at the time). I eventually came up with a term that means a lot to me but will mean nothing to most people here - a barmkin (next to a pele tower). I posted about that here so explorers may wish to explore - but a barmkin was more specifically the area in the centre just beyond the corner point.

More recently, influenced by Chinese poems where old men (like me) cultivate medicinal herbs in their kitchen garden but still go foraging for rarer herbs in the foothills of the mountains - and there come across things like immortals hermits playing go! - I have tried to visualise the go board a corners = house and sides = kitchen garden = and the centre as the herb-filled forest going uphill to the bare peak of Mount Tengen. There are many fascinating and unexpected things to be found in that forest - sometimes dangerous things!

But I accept that most other players have a less fanciful way of looking at go. As we saw in a separate thread on tsume go, most players prefer to talk in terms of "I got 100 problems right in under 30 minutes" rather than "I learned 500 interesting new things by discussing 100 famous problems - and knowledge that will last me a lifetime." I prefer to hire other people to cultivate my garden while I go on trips into the fascinating forest.

So it seems that you mean connections in long-range connections. In KataGo's play, they are not a vital aspect because KataGo also considers sacrifices, incl. those of whole groups.

I'm not sure about this 收 being a Chinese Go concept but maybe it could just as well be a general concept of Go. Having a name to call things is useful.

Since 收官 is endgame I'd assume, probably like an ignoramus, that 收 is picking up or collecting points by closing boundaries and surrounding territory, and also reducing what could have become territory. However, you could be collecting or receiving almost anything, like the mail, if there isn't enough context.

This is a common character, number 351 by frequency, and you can expect roughly 1/2000 or so characters in modern corpus to be this one. It is used in other ways in Go context, it is so common that it has to be. 收气 is to take liberties, 收子 is to put away the stones after a game.

Shou is a Chinese concept in go, going back at least several centuries to the Ming dynasty. My reference is to a large corpus (of my own making) of the medieval language as used in go manuals, where the entire texts are go technical matter. In addition, there were essentially no binomes in the classical language and so reference to the modern language is a reference to a different language. Shou in medieval texts occurs only in commentaries and is not limited to the endgame.

The nearest Japanese term at that time was 結　which was then read kechi (as in Tale of Genji, Igo Shiki etc, all of which are in classical Japanese).

Your point about being willing to sacrifice is well made. I'm sure it's all in the KataGo neural net, but it's tough to tease out. The existence of long-range connection would help put a cap on how large the sacrifice would need to be, however. If we think about the long-range connection being between two outposts, sacrificing would mean given up one outpost. But you can play from the second outpost towards the sacrifice in order to limit its scope, hopefully building the outside position a little more in compensation.

This was at least part of what I took away from John's discussion of barmkins--the small defensive positions for use in troubled times. They provide a base from which to operate, and an attacker has to put in a bit more effort than just simple raiding for a few sheep.

Once I won a game in a tournament, in which I scored 4:0, by sacrificing 100 points to win by 2 points. Also KataGo does not have a limit to the size of a sacrifice and sometimes sacrifices huge groups in exchanges. Therefore, it might sacrifice both parts of what some might perceive as would-be long-range connection, or sacrifice one part in an exchange, then the other part in another exchange or more complicated. Or, as Iwamoto put it, connecting does not gain any points [in itself in the post group tax age].