It is currently Thu Mar 28, 2024 11:56 am

All times are UTC - 8 hours [ DST ]




Post new topic Reply to topic  [ 6 posts ] 
Author Message
Offline
 Post subject: Seki Kotaro
Post #1 Posted: Mon Jan 24, 2022 6:02 am 
Oza

Posts: 3647
Liked others: 20
Was liked: 4626
Seki Kotaro has recently made news through becoming the fast ever Japanese pro to progress from qualifying as 1-dan to winning a BIg 7 title (4 years 8 months; the 47th Tengen in Dec 2021, shortly after his 20th birthday).

Comments on games in the Japanese press do not often allude to AI (and graphs are almost totally absent), but comments on Seki's games do have a little more AI seasoning than most. Indeed, he has been described as what me might, in English, call the poster boy for AI. The Japanese phrase for this is 申し子 or moushigo. 申す is an honorific verb for speaking, and might be used for addressing the gods. So a moushigo is a child heaven-sent to a childless couple in response to their prayers. I suspect there is a hidden pun here, too, because the foreign player who is talked about in AI terms in Japan is Sin Chin-seo. It has been suggested that Sin is the human whose play most matches AI play. By coincidence, Sin's surname is 申.

Seki's play is likewise noted as following the AI tramlines, not quite to the terminus, but certainly for a long ride. In Game 4 of the Tengen title match, he as driver with Ichiriki Ryo as conductor, managed a trip of 40 stops before the first glitch appeared. Even then it was a small one, but highly instructive.

Since I am presenting this in my own way, a little bit of background is needed. The position after move 40 was the one below (White has just played L17). Seki made the comment that there was a severe move here according to AI, and as Black (Seki) missed it he got a small disadvantage. In AI terms it was a mistake of 3-4 percentage points, which is inconsequential to us but of concern to a pro. I have shown, unranked, the moves that were considered by my AI. You may care to guess the AI move, and adduce the reasoning, which we look at below.



What I think I have discovered is that whereas traditional go theory is heavily binary (profit/thickness, invade/reduce; attack defence, miai, etc), the theory being developed in the light of AI is about nexuses. Properly, the plural of nexus is nexus (long u in Latin) but I will grit my teeth and talk about nexuses (and nexusology as the theory).

It is my assertion that if you look at a typical position with AI's policy moves marked, as above, you can/should categorise them into three main (Tier 1) nexuses: settling moves, colonising moves and overconcentrating moves. There is a Tier 2, which includes a probes/preparation nexus and a trades/sacrifice nexus, for example. And so on.

Each concept is a nexus because it does not depend on a definition but is rather a network of associations. So settling, for example, covers making bases, making prophylactic defensive moves (e.g. mamori and honte), and settling boundaries (yose). 'Overconcentrating' is a nexus that includes efficiency but also e.g. forcing moves.

Some moves can fall into more than one nexus. I won't say much about that here (I do say a little more in Go Wisdom though that is not yet definitive), but it does add a time element to nexusology in that it can make such moves more urgent. Normally nexusology is not designed to predict the best move. It is rather designed to provide a rational narrative to the commentary on a game. It tries to predict the right area for the best move and the explain the rough reasons for that move. Choice of the exact right move depends also simply on reading, along with the undercurrent of timing.

What I would claim is that people attuned to AI moves would instinctively look at certain types of moves such as those triangled above. Even weakish amateurs would now readily look for shoulder hits, contact plays and centre moves. I believe that pros do the same but on a larger scale. That larger scale is, I think, describable as nexusology. I stress that the terms and categories are my own, and could be quite off-beam, but I have based them in what I have read.

Now back to specifics. What was so interesting for me about the right play for move 41 above was that Seki did not really miss it. He actually played it, but on move 51. His mistake was more a matter of timing. The right move could be defined as coming in to the overconcentrating and settling categories, but so could Seki's move. It is my strong impression that the best young pros of today are not having much problem seeing the the candidate moves. Their problems are with timing them.

Let us now look at Seki's comments in that light. The following is the AI recommendation. Seki's actual move was M16.

Attachment:
Capture1.JPG
Capture1.JPG [ 44.46 KiB | Viewed 3206 times ]


Seki's 'headline' comment on Black 1 (which he did play on move 51, I repeat, though the reasoning differed by then) was that, "It's rather difficult for a human to see it, but given the situation on the tight side, it seems to be the most urgent point." I infer from this, as I have already indicated above, that the flaw in his nexusology is not really a matter of noticing the move but a matter of timing or urgency.

Seki then goes on: If White 2 submits at 5, Black can be satisfied with then playing the jump to 'a'. That being so, I would want definitely to play the nobi for White 2, but then Black 3 and the hane 5 are severe. After White cuts at 6, Black 7 and 9 are a tesuji. They make miai of b and c [S7]. White has been suckered into this sequence. Therefore, White cannot resist Black 5 with 6 here but will have to do something like block at White 1 below." (10 at 1)

Attachment:
Capture2.JPG
Capture2.JPG [ 43.71 KiB | Viewed 3206 times ]


"After Black connects at 2, White makes a prophylactic move (mamori) on the right side at 3, but Black's hane next at 4 is very punishing. White has to defend submissively [ukeru as opposed to mamoru] and ends up shapeless. When he guilelessly plays 5, Black cuts at 6 and White 7 to 11 produce shapeless mess. In contrast, Black has very nice shape on the right side. The division of spoils favours him, doesn't it? I could understand that Black's contact play 1 in the first diagram was the urgent point once the AI showed it to me, but with just a cursory look at the position I didn't really register it."

In connection with that last remark, he said elsewhere that he prefers 8-hour games to 5-hour games. THta bucks the trend a bit. I wonder whether it indicates that the new breed of pro relaises they need more game-time to get to grips with the complexities nexusology implies.

In that separate interview, Seki says some very interesting things about AI percentages. There's not enough life on L19 to justify me spending the extra time on that, but if you want to hunt it out it's in Go Monthly 2022/2.


This post by John Fairbairn was liked by 2 people: bernds, jeromie
Top
 Profile  
 
Offline
 Post subject: Re: Seki Kotaro
Post #2 Posted: Tue Jan 25, 2022 1:17 pm 
Lives in sente

Posts: 902
Location: Fort Collins, CO
Liked others: 319
Was liked: 287
Rank: AGA 3k
Universal go server handle: jeromie
Quote:
Normally nexusology is not designed to predict the best move. It is rather designed to provide a rational narrative to the commentary on a game. It tries to predict the right area for the best move and the explain the rough reasons for that move. Choice of the exact right move depends also simply on reading, along with the undercurrent of timing.


This reminds me of a book I read with a group of faculty and staff when I was teaching at a community college. The book, The Righteous Mind by Jonathan Haidt, presents the theory that ethical decisions are typically made on instinct with a logical rationale applied afterward. According to his moral foundations theory, the initial decision a person arrives at is influenced by a person's stance in five clusters of moral thought: care, fairness, loyalty, authority, and sanctity. (He later proposes a sixth cluster, liberty.)

His description of ethical decisions is pretty similar to the nexus idea you propose. Obviously, the types of decisions and their stakes are different on a go board, but the basic idea that complex decisions are usually made by instinct first with a logical rationale applied later matches the idea of intuition first / reading & logic second that seems central to nexusology. Another element of his theory that applies here is that if you are self-aware enough to be cognizant of the initial instinctual decision and the logical factors that go into it, it can hone your intuition, help you understand others' decisions better (particularly important when we're trying to make sense of AI moves that they can't explain), and allow you to more effectively apply logic in situations where overriding the initial impulse is appropriate (a skill many amateurs, myself included, could apply more effectively on the go board).

Top
 Profile  
 
Offline
 Post subject: Re: Seki Kotaro
Post #3 Posted: Wed Jan 26, 2022 11:09 am 
Gosei

Posts: 1625
Liked others: 542
Was liked: 450
Rank: senior player
GD Posts: 1000
I don't know whether I understand nexus as used in go. Is it true that almost every move in a game involves nexuus (using the word processor romaji Japanese long vowel)? We try to play multipurpose moves. Even the first moves involve nexuus: starpoint vs. komoku would be influence vs. territory, for example.

Top
 Profile  
 
Offline
 Post subject: Re: Seki Kotaro
Post #4 Posted: Wed Jan 26, 2022 11:41 am 
Oza

Posts: 3647
Liked others: 20
Was liked: 4626
Quote:
I don't know whether I understand nexus as used in go. Is it true that almost every move in a game involves nexuus (using the word processor romaji Japanese long vowel)? We try to play multipurpose moves. Even the first moves involve nexuus: starpoint vs. komoku would be influence vs. territory, for example.


Until I came along I don't think anyone has seriously used nexus in go, and I've certainly never seen it used in Japanese in any domain.

I am using it in the ordinary English sense of a series or group of connections - of two or more things but in practice the minimal nexus of two is usually disregarded, and that is certainly what I emphasise. I don't regard star point vs. komoku as a true nexus. The true nexus would have to be starpoints AND komokus.

My underlying point is that, in go, we (including pros) accentuate the binary and also binary opposites, but AI is teaching us that we should think in broader and less oppositional grouping. That in turn means getting away from traditional go words such as thickness and territory, simply because they are too tinged by binary-ism. Hence ny attempt to kick start the process with what I regard as Tier 1 concepts: settling, colonising and overconcentrating. As in the example I gave, most moves shown as candidates can be said to fall in at least one category, and when a move does fall into a category of that type, it acquires ALL the attributes AND associations (i.e. it's a kind of neural networking). Nexusology, in my view, allows us to describe moves in a richer way that shows not just core behaviour but all the associations that are present. I gave a concrete 9though still limited) example elsewhere of how we can talk about tsumes (checking extensions) where the richness of the nexus way of looking at it explains (a) why Shusai was fond of such m+oves and (b) why checking extension is a poor equivalent.

In broader terms it also explains why pro vocabulary is changing in the AI age. I haven't seen a discussion of this, so I have to draw my own conclusions, but as a relatively trivial example of this is an apparent trend to refer to shoulder hits as forcing moves and not shoulder hits. The functions of moves seem to be emphasised now rather than the appearance.


This post by John Fairbairn was liked by: gowan
Top
 Profile  
 
Offline
 Post subject: Re: Seki Kotaro
Post #5 Posted: Sun Jan 30, 2022 2:52 am 
Lives in gote

Posts: 532
Liked others: 27
Was liked: 142
KGS: pajaro
IGS: pajaro
I find this very interesting.

Before AI, I used to think "stronger players tell me that this is better, this is safe, this is... but what if there is some flaw in what is thought to be right? what if there is actually another way?" But amateurs learnt from pros, pros spend a lot of time studying and discussing, and their living is at stake. So they must know best. Also, I remember Go Seigen's words about his chances if he played God: "with 2 stones, the result would be uncertain; with 3 stones, I would be confident; with 4 stones, I could bet my life on the game". How realistic is that now...?

AI has brought some ugly truths: pros (guess what) are not perfect, make mistakes, and are far from playing a flawless game. Sai needs some more reincarnations. I watched Michael Redmond's youtube videos about "Master". Master was an unknown internet player that played 60 games about top pros (some pros played more than once) and won all of them. In his comments, Redmond said several times "we thought that this (some sequence, some move) was right/wrong, but Master played different and proved us wrong". Something like that. "we"=humans, "Master"=AlphaGo. You must have already guessed.

So now humans, seeing that AI beats them effortlessly, have decided to learn from it. Here comes where what John said hit me.

I thought that humans would see the game with the pre-AI knowledge and theory, and improve it with AI. For example: thickness vs. territory. Go (as humans see it) is about balance. If in a certain situation you get thickness, your opponent get territory. Then you want to use this thickness to get territory elsewhere, your opponent will get thickness, and hopefully, when the game is over, exchanges have gone slightly your way and you win. In this pre-AI way of thinking, you need to estimate if the thickness you win is worth more or less than the territory you give to the opponent. This estimation is hard, of course, and at the same time, you need to balance more pairs of concepts: sente vs. gote, attack vs. defense,...

Now add AI analysis. I thought that, with AI analysis, those estimations would change. "We thought that this wall was equal to that territory, but the wall has a defect, it's not so good". Or "we thought that this sequence was even, and called it joseki. but we have found another variation, and nobody calls it joseki anymore". And so on. Humans would keep the same theory, but improve it with better moves, sequences, timing... That was my understanding.

But I see now that pros have had to change their way of thinking, with new and updated concepts. I think that they still could explain some (most? all?) moves in a simple way to a wide audience. The old concepts are still useful. But deep inside, things are more complicated, and new ways to explain the game are needed.

The success of Seki Kotaro may be seen as what can happen with players that learn after this revolution. The gap with pre-AI style is obvious.

Top
 Profile  
 
Offline
 Post subject: Re: Seki Kotaro
Post #6 Posted: Sun Jan 30, 2022 7:19 am 
Oza

Posts: 3647
Liked others: 20
Was liked: 4626
Quote:
Also, I remember Go Seigen's words about his chances if he played God: "with 2 stones, the result would be uncertain; with 3 stones, I would be confident; with 4 stones, I could bet my life on the game". How realistic is that now...?


I think he might still win the bet. It is true that some pros have been losing 3 and 4-stone games against bots, but there are pros and pros. I recall Go Seigen's comment on another pro, to the effect: "He's weak; he's only a 4-dan." And there is the related story that Yasunaga Hajime was a 4-dan pro and was known as the terror of other pros in handicap games. But Go beat him down to 4 stones. Another case where he pre-dated AI.

A more pertinent pro comment - I forget who said it but it may also have been Go - is "all amateurs are weak." I can imagine most people here would shrug and say, of course. But over long years I have found that people in many, many fields accept this at one level but not at other, more important, levels. By that I mean they refuse to act on the fact they are weak by sticking to what they think they know (or have been told). They are not properly addressing the causes of their weakness.

It's common on L19, as it is everywhere else, but two non-go examples have stuck in my mind. The first was in my very first job, as a translator. We took on a work-experience graduate who had just got a degree in German. Obviously we had to check her work. It was horrific. Despite her degree she did not know that werden is used in German to make the passive as well as the future tense. She thought it only made the future. Most of our work was translating German patents, and as in all technical texts the passive dominates. That was bad enough, but what was really bad was that she would not listen to any corrections. Her teacher, she said, had said werden made the future and that was that. She was soon made to leave.

The second example I most remember is more go-related. It was when I was working on the world's first shogi computer, and part of the project involved talking to top chess grandmasters, trying to interest them in shogi. Ignorance was not the problem here. Walter Browne's pithy answer was typical of the attitude of many: "I've spent years learning about chess; why should I give all that up for another game?" In other words, short-sightedness was the problem.

Other grandmasters listened with interest, though were disinclined to follow it through. However, one did tell us that Bent Larsen had looked at shogi and had been influenced by it: he was noted then for his unusual edge-pawn pushes. Shogi happens to have a proverb: "Push the edge pawn!" I have been told since that the fondness for king walks that some chess players later developed was likewise influenced by shogi (where the king cannot castle except by taking a long walk).

It is, of course, perfectly understandable that go amateurs who treat the game as a time-filling or stress-escaping hobby do not wish to fall, Alice like, down every new rabbit-hole (unless, perhaps, it belongs to a "Wonderland" labelled AI coney). But that doesn't alter the underlying truth that "all amateurs are weak" and that what got them to that state of weakness was a mode of thinking that they have not updated.

Young players like Seki have the advantage that they have not much to update: they are almost a tabula rasa. In such a competitive profession, they are highly unlikely to tell us directly what secrets they are discovering. My own approach is to try to glean what I can from the linguistic crumbs that fall from their tabula whenever they are obliged to say something about their games. Lean pickings still, but I think I can see patterns emerging.


This post by John Fairbairn was liked by: gowan
Top
 Profile  
 
Display posts from previous:  Sort by  
Post new topic Reply to topic  [ 6 posts ] 

All times are UTC - 8 hours [ DST ]


Who is online

Users browsing this forum: No registered users and 1 guest


You cannot post new topics in this forum
You cannot reply to topics in this forum
You cannot edit your posts in this forum
You cannot delete your posts in this forum
You cannot post attachments in this forum

Search for:
Jump to:  
Powered by phpBB © 2000, 2002, 2005, 2007 phpBB Group