translation of a 9d pro commentary

John Fairbairn · Post by **John Fairbairn** » Tue Jan 10, 2017 8:52 am

FWIW, I have had the feeling that there was some parallel between AlphaGo and Go Seigen, purely in terms of go. Some of the AlphaGo moves that Mi pointed out were moves that had struck me as Go Seigen moves. IIRC, the play against the large knight's enclosure, building up some -- ahem!-- thickness and then extending from it was something that I saw in a Go Seigen game when I first began to study go books. Also, Uberdude has pointed out the flexibility of AlphaGo's play, and I think that that is characteristic of Go Seigen, as well. Rui Naiwei told me that one thing about attending Go Seigen's study groups was that he often changed his mind, saying one thing one week and another thing the next. He was creative till the end.

I agree with all this and would like to add two things to expand on it.

1. The famous story by Go's doctor when he was in a sanatorium and studied go on his own - whenever the doctor entered, he noticed that there were only a few stones on the board and they were always in the centre (this was after the New Fuseki fad had faded).

2. I think it was Go's last joseki book and the content was to do with what aji lies beneath the joseki. He showed line after line of eye-popping variations in standard and even simple josekis that I have never seen elsewhere, in encyclopaedias, books or magazines, and what was more impressive, I would never have suspected that such possibilities even existed. One of the few books that made me physically gasp.

dhu163 · Post by **dhu163** » Tue Jan 10, 2017 6:58 pm

the thickness concept vs two eyes comes up again here see

Bill Spight · Post by **Bill Spight** » Tue Jan 10, 2017 7:46 pm

John Fairbairn wrote:I think it was Go's last joseki book and the content was to do with what aji lies beneath the joseki. He showed line after line of eye-popping variations in standard and even simple josekis that I have never seen elsewhere, in encyclopaedias, books or magazines, and what was more impressive, I would never have suspected that such possibilities even existed. One of the few books that made me physically gasp.

Is that this book? 定石の考え方―有段者のための集中講義 Thanks.

John Fairbairn · Post by **John Fairbairn** » Wed Jan 11, 2017 2:24 am

Is that this book? 定石の考え方―有段者のための集中講義 Thanks.

Yes.

Resuming the Go Seigen connection, move 91 in the Pak Cheng-hwan game reminded me that Go apparently made more contact plays than other players. I haven't followed the AlphaGo games closely enough to know, but does it likewise have a predilection for contact plays?

The commentaries are fascinating but raise awkward questions. E.g. are the pros in danger of anthropomorphising AlphaGo and so seeing things that just aren't there. In particular, they seem to believe that its superiority is largely strategic. They may be right, but that does seem to fly in the face of chess experience. Brutal tactics may not be as important in go as in chess, but is there not the possibility that the go equivalent is the Monte Carlo playout. In other words, are pros looking at the subtle early part of the game for their mistakes when they should be looking at the more mechanical later parts.

I also find it hard to imagine, programmatically, that AlphaGo is seeing opening lines 40 move ahead and evaluating them, or at least evaluating them in our terms. It seems more likely that it has stumbled upon a previously unsuspected basic go truth such as two eyes early on being even more important than we have ever suspected. Where the human neglects early eye shape, the long-term power of the Monte Carlo method, say, manages to exploit that over a long period.

The fact that other programs have been able to make a very sharp increase in strength by following AlphaGo's example seems to suggest there is something mechanical that can be reproduced. It beggars belief that the other programmers have discovered and copied AlphaGo's evaluation function.

Uberdude · Post by **Uberdude** » Wed Jan 11, 2017 3:06 am

John Fairbairn wrote:I also find it hard to imagine, programmatically, that AlphaGo is seeing opening lines 40 move ahead and evaluating them, or at least evaluating them in our terms.

It does read lines in the opening a long way ahead and evaluate them, there were some examples in the commentaries released several months ago. E.g vs Lee game 2, move 11 see variation to move 56.

John Fairbairn wrote: It seems more likely that it has stumbled upon a previously unsuspected basic go truth such as two eyes early on being even more important than we have ever suspected. Where the human neglects early eye shape, the long-term power of the Monte Carlo method, say, manages to exploit that over a long period. The fact that other programs have been able to make a very sharp increase in strength by following AlphaGo's example seems to suggest there is something mechanical that can be reproduced. It beggars belief that the other programmers have discovered and copied AlphaGo's evaluation function.

DeepZen (and probably the Xing Tian Chinese AI too) have implemented a Value (aka board evaluation function) neural network too. Not as good as AlphaGo's, but it has made it a lot stronger. (Also I expect the engineering cleverness and fine-tuning glue that sticks together the Value network, Policy network and MCTS parts of AlphaGo are better than in the other bots and will be where a lot of the secret-sauce not present in the Nature paper that makes AlphaGo so strong lies).

In the Park Junghwan game, I really liked Master's combination of simple patient one point jumps (o11, o5) and kosumi with the m6 splitting move. As Matt Cocke UK 5d says, Go is a topological game.

I too see some similarity with Go Seigen, the flexibility, appreciation of the centre, willingness to make unorthodox moves. Things like that shoulder hit into Mi Yuting's top right shimari make we wonder if most pros are reluctant to do so even if they consider it because they fear it is aji keshi and lack the confidence in their judgement to play it, whereas Go and AlphaGo have more confidence (yeah I'm anthropomorphising). Maybe it's no coincidence that AlphaGo appeared shortly after Go Seigen was reported to have died

.

John Fairbairn · Post by **John Fairbairn** » Wed Jan 11, 2017 4:58 am

As Matt Cocke UK 5d says, Go is a topological game.

Interesting and much underused concept* - possibly because it's too hard for humans but not for AlphaGo?. I first came across it through Paul Prescott at the London Go Centre (can't remember if he became British champion but he was at least a challenger). He studied topology (or a mathematical kind of it) at Cambridge, but in go he boiled it down in a simplistic but very powerful, which was something like preferring moves that did at least two things to moves that did one thing (or maybe better, moves that affected two or more groups instead of one).

How do you see topology working in go? And do you know how Matt sees it?

*There is a word, not common but not rare, in Japanese go 碁形 which is a nuisance to translate well: 'go shape' is obviously silly. It refers to the shapes in the game as a whole, but not in the sense of katachi. The sense is in fact the topography of the game, and the analogous word for topography in the geographical sense is 地形. I'm wondering, in the light of your comment and the changes in thinking being wrought by AlphaGo, whether we should be introducing the word 'topography' to go. What do you think?

dhu163 · Post by **dhu163** » Wed Jan 11, 2017 5:30 am

Contact plays: Partial yes: I seem remember thinking after Alphago vs Lee Sedol that the bot made some contact plays when a human wouldn't necessarily. I only remember a few now in joseki - game 1 and game 2 to quickly settle a weak stone. And there is the famous 4-4 attach to the 3-4 stone in the Chinese opening that it played in a self-play game.

Go seigen is also famous for shoulder hit small knight enclosure and tenuki like Master vs Mi Yuting

but is there not the possibility that the go equivalent is the Monte Carlo playout.

I don't really think this is significant. I have experimented with Leelabot 5d with and without the neural net, and the skill level is horribly bad without it, only depending on Monte Carlo. Also, I do think the human early game is a weak point compared to perfect play, while the analysable endgame we are much better at. (e.g. Conway's surreal numbers). I find that bots have a much more accurate sense of logic to guide where to play in the opening, and as I have said elsewhere, it is the opening where we can learn a lot from them. (Pros normally say the opening is irrelevant but I think it is more because no-one has a clue about what value judgement is best) I also think their use of middlegame probes is particularly educational, or being prepared to make a large exchange, or their value judgment when the opponent makes a probe/overplay.

that AlphaGo is seeing opening lines 40 move ahead and evaluating them

I don't know if Uberdude is making an implication about general reading. It looks at some simple variations very far ahead to help it judge the position but it can hardly say it is the best variation for both sides. However, given any one particular variation (e.g. crawling on first line, Lee Sedol game 5), I think the probability it was considered among the myriad of variations is drastically low, especially when the neural net is unlikely to recommend crawling so far on the first line. The key point is in the quality of the neural net.

It seems more likely that it has stumbled upon a previously unsuspected basic go truth such as two eyes early on being even more important than we have ever suspected. Where the human neglects early eye shape, the long-term power of the Monte Carlo method, say, manages to exploit that over a long period.

I agree with the basic meaning expressed here, though it is clearly very complicated. You are describing the "settle fast and simplify" feature of bots. While this may often be a good idea that humans can learn from, I suspect that much of it is due to the nature of the programming of the bot and the neural net:

It likes to play simply because it is bad at reading, and also may be bad at life and death (not sure how difficult alphago finds this, but other bots certainly have issues). And it fears uncertainty. Its skill will always be limited if it is has such fear.
Its reading is guided by its own neural net, so it fears what it thinks are good moves, but may miss good moves from the opponent that it wouldn't play. It is used to playing good shape moves, and can miss severe moves from the other side. Though this seems to argue in the opposite direction. But in any case, it can miss local things easily, and this inevitably affects its style. After playing with leelabot a lot for 3 months, though it is a lot stronger than me, I am starting to see the subtle weaknesses in its neural net beyond the basic semeai, reading issues.

Things like that shoulder hit into Mi Yuting's top right shimari make we wonder if most pros are reluctant to do so even if they consider it because they fear it is aji keshi and lack the confidence in their judgement to play it,

for sure

Many pros believe in the value of the centre, but it is so uncertain, few dare to use it, so it becomes impractical for humans to study it. At least my own attempts at the London open to imitate a bot style didn't go so well. And a mutual friend mentioned how he actually gets worse from playing Alphago

Cassandra · Post by **Cassandra** » Wed Jan 11, 2017 6:33 am

Some thoughts of my wife, an Ikebana teacher:

As a matter of course, the game of Go is affected by the "laws of harmony", thus "the flow of lines" matters, thus "topology" will be an issue.

What she learned from Japanese Ikebana masters ...

-- Become perfect in "technique".
-- Learn to "see" and "feel".
-- Be convinced that what you are going to do is "right".
-- Now you are free.

Uberdude · Post by **Uberdude** » Wed Jan 11, 2017 7:14 am

I believe Matt's PhD is in singularity theory (the Mathsy kind not the new AI "gonna kill us all" meaning) and he tends to like large scale fighting games where you can get such double threats, so that's an interesting double anecdote. He is/was playing a (dormant) Malkovich game with Joaz here in which he pushed into a knight's move and Joaz didn't connect so broke though, writing about the topology aspect: http://www.lifein19x19.com/forum/viewto ... al#p172969. I cringed when Joaz didn't block as I imagine most other dan players would (he did consider it but then did his clever/dubious tenuki with the comment "If feels really DDKish, but it is hard to see how white gets a win here."). A few other posts from Matt on it listed here: http://www.lifein19x19.com/forum/search ... thor=drmwc.

I see topology in Go as what I sometimes call "large-scale connectedness" of the big picture of white group here, black group there, another white one on that side, which sounds similar to your 碁形. 'Connected groups good, separate groups' bad is hardly a new idea, but different players have different preferences, for example Kim Jiseok seems quite willing to make lots of little groups in messy fighting whilst Takemiya likes to connect (I particularly like his connecting weakish stones to form a moyo, turning liability to asset). Of course sometimes connecting on dame is slow and inefficient, you can place too much trust in the topological aspect and hopeful future territorial gains (e.g. picking up endgame points thanks to thick connected positions, for example moves that are locally double sente for territory are no longer sente for life and death so you can more easily ignore them (or get them yourself) and make beneficial trades). Another nice move from AlphaGo stressing topology/large-scale connectedness was 81 vs Lee in game 2. Kim Myungwan 9p really liked this move, as if it was good, which it seems to be, shows how the normal human thinking of directly answering the invasion directly was misguided. I seem to recall though that Lee could have played better afterwards (one point jump too simple?) so maybe it wasn't as good as the game result made it look.

Click Here To Show Diagram Code: [go]$$Wcm80 AlphaGo (black) vs Lee Sedol $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . 1 . X . . . . . . . . . . | $$ | . . X , X . . . . , X . . X . X . X . | $$ | . . . . . . . . . . . . . . . . X O . | $$ | . . . 2 . . . . . . . . . . . O O . O | $$ | . . O . . . . . . . . . . . . . . O . | $$ | . . . . X . . . . . . . O . . . . . . | $$ | . . . O . X . . . . . . . . O O O . . | $$ | . . . O . . . . . , . . . X X X . . . | $$ | . . . O X . O . . . . . . . . . . . . | $$ | . . O X . X . X . . . . . . . . . . . | $$ | . . O O X X . X . . . . . . X . . . . | $$ | . O X X X O O O O . . . . . . . O . . | $$ | . O O X O X O X . . O . . . . X O . . | $$ | . X X O O . X O . X X . X . X , O . . | $$ | . X O X . O O . . . . . . X O O . . . | $$ | . . . X . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

Bill Spight · Post by **Bill Spight** » Wed Jan 11, 2017 1:42 pm

Ah! An unusual ko and the endgame. Just my cup of tea.

I have added a few comments to Meng's on the Park game.

Go to move 177.

Edit: I am going to add another variation. Please wait.

Edit: Done!

Edit: Got Black and White mixed up. Corrected.

Gomoto · Post by **Gomoto** » Wed Jan 11, 2017 3:03 pm

E.g. are the pros in danger of anthropomorphising AlphaGo and so seeing things that just aren't there.

No.

AlphaGo is winning by "anthropomorphing" into a new kind of AI. Anthropomorphism is no sin in this case, it is just adequate.

Alpha Go has a very good "intuition" and is very fast at "reading" as well.

sorin · Post by **sorin** » Wed Jan 11, 2017 8:29 pm

John Fairbairn wrote:
Is that this book? 定石の考え方―有段者のための集中講義 Thanks.
Yes.

Can somebody who has it give an example from this book? I became very curious after reading John's description.

John Fairbairn · Post by **John Fairbairn** » Thu Jan 12, 2017 2:07 am

The book, "How to think about Joseki," is a waste of time unless you read idiomatic Japanese. The format is writer Mihori Sho asking Go questions and Go answering at some length. Go never dumbs down and all the meat is in the text. "Meat" seems especially appropriate as it's a slim book (160 pages) but very high protein content. The subtitle in fact is "Concentrated course for dan players."

Formally it is divided into four sections: How to Use Josekis, Bad Josekis, Within the Joseki, and What You Need to Know after the Joseki (more specifically the things amateurs don't know). There are also three insightful mini-essays.

As those section headings indicate, Go is never afraid to call a spade a spade, and the very first example involves a simple joseki that is common in the AlphaGo and other pro games, and which we all know and play. Go says it's bad, expounds at some length. Mihori, a 6-dan amateur, says "I don't understand" and so Go has to launch into further explanation. Again Mihori declares himself astonished and unable to comprehend and so Go has to try again. That may indicate the level he's aiming at. Rather than teach you nifty traps or new lines, he wants to show you how a pro thinks about joseki.

At the beginning, with this example, you get a glimmer of how pros think. You may just have a feeling of enlightenment, like Mihori ("Hmm, finally I may just be getting an inkling of understanding") but, as the book wears on, you realise it was only a glimmer and as an amateur you are still in mirkest glen. I suspect you already need to think like a pro to really understand this book.

You may also understand now why I didn't bother trying of give a sample - I'd have to end up giving a whole chapter.

sorin · Post by **sorin** » Thu Jan 12, 2017 9:43 am

John Fairbairn wrote:You may also understand now why I didn't bother trying of give a sample - I'd have to end up giving a whole chapter.

Thank you very much for the description! I wish we had an English translation of this book, I would like very much to read it.

wolfking · Post by **wolfking** » Thu Jan 12, 2017 5:31 pm

sorin wrote: Can somebody who has it give an example from this book? I became very curious after reading John's description.

John Fairbairn wrote:The book, "How to think about Joseki," is a waste of time unless you read idiomatic Japanese. The format is writer Mihori Sho asking Go questions and Go answering at some length. Go never dumbs down and all the meat is in the text. "Meat" seems especially appropriate as it's a slim book (160 pages) but very high protein content. The subtitle in fact is "Concentrated course for dan players."

I totally agree with John that it is a waste of time if you do not understand the language and try to read the book using only the diagrams.

However if you do understand Japanese or have a translation of the book (I have the Chinese translation) I do think it is still a good read for dan level players maybe even strong SDKs. The Q and A format of the book was actually designed to be more friendly to amateurs.

The whole book consists of series of conversations between a made up character Mr. Q (assumed to be amateur 2 dan) and Go Seigen. Mr. Q would ask questions any curious amateur would normally ask about josekis in pro games, and Go would explain why certain moves (often slight deviation or less common variations of standard josekis) were played using words and diagrams. To give you a taste here is the very first series.

Mr. Q asked Go sensei why

in Diagram 1 was not played at the usual low position "a".

Click Here To Show Diagram Code: [go]$$c Diagram 1 $$ +---------------------------------------+ $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . a . 4 . . . | $$ | . . . , . . . . . , . . . 0 . , 8 . . | $$ | . . . 5 . . . . . . . . . . . 1 6 . . | $$ | . . . . . . . . . . . . . . . . 7 . . | $$ | . . . . . . . . . . . . . . . 9 . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . 2 . . . . . , . . . . 3 , . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ +---------------------------------------+[/go]

Go: The usual "Kitani Joseki" is at

(Diagram 2). After that moves

to

immediately comes to mind. In this position you can probably tell that

would be better played at "a".

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

Mr. Q: I still don't see it.
Go: That is, if

was at "a" instead, then after

white would immediately attach at b, and now do you see a mojo developing on the top side? The effects of "a" and

are quite different. In this game, black played 4-5 points on both top left and bottom right, that is why white decided to play the higher knight's move at "a".

In the book Go continued to explain a little about the actual game where black played moves to avoid the outcome shown in Diagram 2 (Black did not play on the right side to finish the joseki, instead he immediately played a move to protect top left corner).

Life In 19x19

translation of a 9d pro commentary

Re: translation of a 9d pro commentary

Re: translation of a 9d pro commentary

Re: translation of a 9d pro commentary

Re: translation of a 9d pro commentary

Re: translation of a 9d pro commentary

Re: translation of a 9d pro commentary

Re: translation of a 9d pro commentary

Re: translation of a 9d pro commentary

Re: translation of a 9d pro commentary

Re: translation of a 9d pro commentary

Re: translation of a 9d pro commentary

Re: translation of a 9d pro commentary

Re: translation of a 9d pro commentary

Re: translation of a 9d pro commentary

Re: translation of a 9d pro commentary