Life In 19x19

Hi everyone,

I've been working on a small Go-related project and decided to share it with the community.

Please check out my article:
https://not-perfect.ghost.io/sensaigo/

And the demo video:
https://www.youtube.com/watch?v=evyP_wBoC8c

I hope this is of interest to anyone, let me know your thoughts

Best wishes,
Sascha

"the best moves in Go are actually not recent developments"

As a general statement, this is wrong. In AI study, I have discovered very many moves that AI analysis determines as correct and that previous human play has not known.

"All these types of feedback happen after the actual game"

Why only after games and not also after sequences or individual moves?

"The feedback should be a very simple signal that can be subconsciously understood"

AI often distinguishes between correct and mistake in ways that cannot be perceived "intuitively" but can also be understood by explanations of non-statistical reasoning. (There are also cases accessible by subconsicious thinking, but by far not all.)

"engine live-evaluates the position from viewpoint of the player [...] immediately knowing when you've made a mistake [...] this direction of development and research is worth following up on, not only for the game of Go, but for other domains"

Good ideas... but what you suggest requires invention of an extremely powerful expert system, more powerful, in particular, than the combined literature and the literature of the next at least 20 (maybe hundreds of) years.

"This claim is of course impossible to verify"

The claim is easily refuted on the short time scale by counter-examples of players, such as me, having gained necessarily non-subconscious learning insights from AI study.

Hi Robert,

first thanks for engaging, although it seems you misunderstood me in some points, so let me correct this.

RobertJasiek wrote:"the best moves in Go are actually not recent developments"

You're misquoting here. In the sentence I'm talking about search algorithms - which have been Monte-Carlo based since many years before neural networks were introduced.

RobertJasiek wrote:"All these types of feedback happen after the actual game"
Why only after games and not also after sequences or individual moves?

Sure we might realize during the game that an earlier move/sequence was not correct, but I was talking about reviews, analysis and discussions which we don't usually do during a live game...

RobertJasiek wrote: Good ideas... but what you suggest requires invention of an extremely powerful expert system, more powerful, in particular, than the combined literature and the literature of the next at least 20 (maybe hundreds of) years.

An extremely powerful expert system, more powerful than all the combined literature... Well this has already been invented That's the point.

RobertJasiek wrote: "This claim is of course impossible to verify..."
The claim is easily refuted on the short time scale by counter-examples of players, such as me, having gained necessarily non-subconscious learning insights from AI study.

You should brush up on your understanding of logic, Robert, that is no refutation at all.

The thesis is "Longterm training with live AI feedback (in addition to "classical" training) has the potential to improve evaluation, judgement, intuition and playing strength in general". A counter-example would be if someone had actually studied this way for a long time but seen no improvement. And even one counter-example wouldn't be a refutation since I'm not making all-or-nothing claims here.
The only scientifically valid way to approach this would be to have a group of players (big enough to be statistically relevant) train longterm (months/years) with this method and compare their improvement with a control group that followed the exact same training regimen, only leaving out the live feedback during games.

Anyway my point is not to devalue any existing methods for improvement (play, study, tsumego etc.), on the contrary, they are absolutely necessary. I'm just proposing that we can add something more to get a little extra boost.

The human brain is an all-purpose learning system, and it might be able to incorporate AI feedback signals as well in order to improve it's internal model of the game of Go (or other domains).

That is my theory and only time will tell us if there's anything to it.

Oh, so with "move" you have meant "progress", I see.

In my AI use, I mostly analyse single moves or short sequences. Therefore, one need not always consider whole game sequences.

What? Current AI an "expert system"? Rather close to the contrary.

And no, current AI is not more powerful than all the combined literature. It only has - most of the time - stronger play. With "powerful", however, I also understand "power of providing reasoning". There, current AI is extremely weak while all the combined literature is comparatively strong.

"The thesis is 'Longterm training with live AI feedback (in addition to 'classical' training) has the potential to improve evaluation, judgement, intuition and playing strength in general'."

I see. This is not outrageous but reasonable. It does not even claim to be remotely as effective as long-term training with the literature etc. (Of course, claims can be adjusted...)

"Anyway my point is not to devalue any existing methods for improvement (play, study, tsumego etc.), on the contrary, they are absolutely necessary. I'm just proposing that we can add something more to get a little extra boost."

Fair enough.

First, I'm totally with the OP idea: how to bring "immediate feedback" into "many repetitions". This is the conundrum of Go improvement: the usual feedback mechanism is win/lose but then you don't get many repetitions. You can evaluate the moves but only after the game, so the feedback is no longer instantaneous.

Yet there remains a problem: the "many repetitions" of subsequent moves are rarely actual repetitions. They vary in nature. Even if you get immediate feedback on move N, that teaches you little about move N+1, so you can't really take the instant feedback into the next repetitions, unlike a practice session on your forehand topspin in tennis, for example.

My solution is somewhere halfway: suppose the thing you want to work on is "invading/reducing moyos". Then you can select a number of games (by Takemiya e.g.) having such moyos and then try solving the question on game 1 + analysis with AI. Using that feedback you go into game 2. Again, these are not exactly repetitions of the same challenge (which would train your memory, not your solving skills), but they are similar enough to provide some kind of deliberate practice.

Secondly, reacting to Robert pointing out AI's weakness at articulated knowledge. I agree that articulated knowledge brings any type of expertise to a higher level, collectively speaking. Novices will benefit from books. Amateurs up to my level still benefit from that too. But there's IMO a threshold. At some point, articulation slows down the learning. Pros can look at AI sequences and figure out what they mean, applying it to their game, without articulating it in any form. Expert knowledge goes beyond language (yes, I'm defying the great Feynman here). You can see this even when pros perform post-game analysis. There's not much talking going on, mostly laying out sequences. The stones do the talking, words would slow them down.

What this means for us is that we shouldn't indulge in the absence of articulated AI understanding. We can look into the sequences offered, and compare them to the sequences we thought of during the game, or during our non-AI-review of the game. Occasionally, we will discover a pattern that might deserve articulation, cfr. the "slonection" I once coined, even if that helped only me.

Knotwilg wrote:Novices will benefit from [articulated knowledge in] books. Amateurs up to my level still benefit from that too.

Not "up to your level". At my higher level, the fastest way of learning go theory still is articulated knowledge in books. The problem is not any limitation of my ability to learn from that but the problem is the only tiny amounts of articulated knowledge in books suitable for at least my level! I do not have the slightest doubt that articulated knowledge in books would also be my fastest way of learning go theory at any higher level. (Of course, such is insufficient for overall strength increase, which must also improve skills of calculation, tactics etc.)

But there's IMO a threshold. At some point, articulation slows down the learning.

My opinion is contrary.

Pros can look at AI sequences and figure out what they mean, applying it to their game, without articulating it in any form.

I (amateur player) look at AI sequences, figure out what they mean and apply it to my game play. I can do so without articulating it to others. Doing so with also articulating it in writing to myself or others eases my application to my game play because I can look up the contents again whenever useful.

That pro players do not (or only hardly) articulate it to others is no evidence that they (unless of the subconscious only type) would not reflect their insight explicitly in their own thinking.

You can see this even when pros perform post-game analysis. There's not much talking going on, mostly laying out sequences. The stones do the talking, words would slow them down.

Such weak teaching is not evidence that explicit explanations could not be very much better.

When I study (e.g., AI) sequences, looking at the raw sequences is only the first step of learning. Inferring articulated go theory is a later, and extraordinarily more effective, step of learning.

What this means for us is that we shouldn't indulge in the absence of articulated AI understanding.

Terrible advice.

It is sometimes hard to tell what people mean when they talk about learning Go. To me that means improving at playing the game from start to finish and to the best of my knowledge you need to play games, from start to finish, without any external aid if you wish to improve at this. Everything else that you do is support for the main activity of playing games. However, there are other views.

To me playing games is practice, exercise and repetition. The proposed method of constantly giving hints to the player is something very different from what I'd wish to call a game (but what to call it ). Add hints about the progress and now we are no longer practicing how to play a game without hints and we don't have an uninterrupted repetition of the mental process that goes into playing a game. Presumable this activity replaces actually playing Go with the opponent and that is the main harm.

As far as I know such signaling of how an exercises is going isn't considered conductive for learning in general. If so, then this must be especially true for Go. Much of the skill of playing Go is in perseverance when facing difficult situations. The best players are after all the ones that are the most dangerous when they should be losing, not the ones that declare the game over one way or the other. When you have hints during a game it has to imped your ability to see the game in the same way as you would if you had no aid. By doing this you remove your most important sense during the exercise, the sense of how the game is going. Yet, the goal of the whole exercise is to exercise your own senses of Go and learn to rely on yourself.

I think it is counterproductive to have any such feedback during a game. Between games is a different story.

Probably the best thing you do besides playing games is a brief review of your games while the memory of playing them is fresh. Possibly the second thing is another brief review when the memory of playing the game is starting to fade. I won't pretend to offer scientific proof, but remembering what you were doing must be crucial for improving on it, right? Possibly remembering what you were doing for longer is correlated to how much you improve. Of course, there may also be games that are better forgotten

Studying books and games (as opposed to brief reviews) is undoubtedly good but it is tertiary activity (if playing games is primary and brief reviews are secondary). The same goes for studying life and death, shape or anything else, it is tertiary activity that should accompany playing games and brief reviews of what you were doing in those games.

Computer programs are of course great for exploring shapes and testing your reasoning about some positions, their main drawbacks is that they still tend to be fairly opinionated and they often prefer going down the rabbit hole over the simple solutions that often work better in practice. Their evaluation often isn't objective for human vs. human game and it gets worse the farther you are from their level of play.

Books have similar problem, the advice often isn't objective when transferred from the pages of the book, especially by someone who is still learning what the book presents. At least books are (still) written by humans that can try to put themselves in the shoes of the reader.

Amateurs up to my level still benefit from that too. But there's IMO a threshold. At some point, articulation slows down the learning. Pros can look at AI sequences and figure out what they mean, applying it to their game, without articulating it in any form. Expert knowledge goes beyond language (yes, I'm defying the great Feynman here). You can see this even when pros perform post-game analysis. There's not much talking going on, mostly laying out sequences. The stones do the talking, words would slow them down.

I suspect most people here reading this would says, "Yes, yes", but I think that's because they will mostly have learned go by a similar, western route. In reality, I think there's a fundamental flaw.

Most players in Japan, Korea and China learn go by playing with family or friends at home or school. They will typically have absorbed a massive amount of go knowledge into their intuition before they even own their first go book. But when they do encounter their first books, they will find that they are written specifically for them, which means the writer will assume this large bank of intuitive knowledge exists in his readers, and he will further assume that when he uses certain words, his readers will understand them in the way he meant them.

In stark contrast, the typical western player skips the early intuition-building playing stage and learns from books instead (other book-fed western players). These books are more often than not translation of Oriental books, and so are full of strange concepts and strange words. Some communication happens, of course, but it is gappy and blighted in various ways. Go intuition is then typically replaced the dead-end path of logic.

There is nothing magic in the intuition the Oriental players rely on. We all can, and do, use intuition in our daily lives - even to a 9-dan level. It's just that in go we skipped the intuition stage. To give a personal example from yesterday, I met my son-in-law after not seeing him for a while, and the first words he says to me were, "2-nil up." I know instantly what he means, not just literally but all the nuances. I might reply, "Ah, 1953." He understands perfectly, even though he is a Chelsea supporter. Then we switch to talking about something else. What I have understood is that Newcastle United are doing well against a fancied team in the FA Cup (soccer) and I can now hope for their glory days of 1953 again. If we had that conversation surrounded by Newcastle football fans, most of them would understand instantly what we are talking about, even if they miss a few nuances. Everyone here can cite similar examples from their own experience.

It should therefore be no surprise when a go pro, talking to other go players - an audience he knows well - just says, "The hane looks good." He knows they will spot the hane on move 13 because they all share the same intuition of how moves will flow to reach that point. Communication can break down sometimes, of course, such as when Kitani hummed and hawed about a particular sequence by chuntering about a yose-ko. His audience was mystified. But the ko in question was 40 moves ahead! Still, once he demonstrated it, the audience didn't have to be tutored as to what a yose-ko was and why it mattered.

The underlying point there is that communication (and thus articulation) is still taking place, even if it is in shorthand. And intuition-based communication supplemented by knowledge from books that all mean the same thing to the audience is so efficient that such shorthand is possible.

If you look at Oriental books and analyse the language, you will find that a great deal of knowledge is assumed. That is, it is not articulated in the book because it is assumed to have been articulated, via the intuition at an earlier stage.

I just happen to have in front of me an example of some of the threads implied in what I have just said. It is a commentary by Go Seigen on one of his games against Iwamoto. It is 1948-11-16s if you want to look it up in the GoGo0D database. A crucial part of the game is shown below (White to play; no komi; both have captured 4).

Click Here To Show Diagram Code

[go]$$W
$$ -----------------------------------------
$$ | . . . . . . . . . . . . O . O . O X a |
$$ | . f . . . . . . . . . O X O O O . O X |
$$ | . . X O O X . O . X . O X X X O O X . |
$$ | . . X , . . . . . , . O X . X X X . . |
$$ | . . . O . O . O . X X O O X X X . X . |
$$ | . . O . . . . . b O O X X O O O O O . |
$$ | . . . . . . . . . X X . . . . . . . . |
$$ | . . . . . . . . . . O X X . . . . . . |
$$ | . . . . . . . . . . . X O O O . O . . |
$$ | . . . , . . . . . , . X X X X O . . . |
$$ | . . . X X . . . . . . . . O X O X . . |
$$ | . . . . O . O . . . . . . . O X X . . |
$$ | . . X . O . . c . . . . . . O O X . . |
$$ | . e . O X . . . . . O . . . . . . . . |
$$ | . . O O . . . . . . . . . . . . X . . |
$$ | . . X , . X . . . , X . . d . X . . . |
$$ | . . . X . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ----------------------------------------|[/go]

A ko has been going on in the upper-right corner. The size of it clearly affects the whole game significantly. However, there is no commentary relating to its size (this is normal - the size of any play is rarely discussed). What has been discussed is the fact that this is a yose-ko that is not a yose-ko, i.e. there are extra liberties, and the possibility of a two-stage ko was discussed. At no point was the definition of either a yose-ko (of either type) or a two-stage ko discussed. And so all this, including the ability to count these freaky kos, was knowledge the reader was assumed to have (or have access to). The only count that was mentioned (or, rather, implied) was the number of ko threats, but only in the sense that White had more of them. The significance of this was not directly discussed, and it came up only in the sense of Black deciding to look for gains elsewhere. More assumptions.

In this specific position, White's move 102 was discussed. The discussion centred on whether White could take the ko at 'a'. But that was not considered a valid option because Black would play the "two-stone ponnuki" (二目ポンウキ) at 'b' (which is the move White correctly played). This is an example where most western players would think it's not a ponnuki because it's not a diamond shape - an example of how their faulty understanding of terms holds them back. The Japanese reader, however, would get the correct nuance that this is not really anything to do with shape but signifies a wow capture (nuki) - he would metaphorically hear the plop (pon) of a champagne cork being popped, and his eyes would light up at the coming feast. In that frame of mind (and instead of of muttering "that's not a diamond shape"), he would then seamlessly grasp the next point of the commentary, which was that Black would next be able to gorge himself by playing at 'c' in return for surrendering the upper-right corner.

Furthermore, the reader would understand (intuitively) the significance of the description of move 'c' as a karami, often used in English but sometimes rendered as something like a 'twisting attack' in English. But Go Seigen didn't use the locution 'to attack' (semeru). Instead, he used the verb semaru (simply 'to put pressure on'). And again the Japanese reader would realise instantly that he needn't expect to kill any White group - he just needs to pressurise the two White groups simultaneously to bring in a bumper harvest.

Stronger western players would grasp these nuances, of course, but (I believe) they would not come to mind so intuitively and naturally - they'd be part of a conscious thought process instead. And weaker players would just flail in the darkness - "what's a karami?"

The continuation of the commentary follows the same path of discussing mainly tactics, with strategy being left largely unspoken but assumed to be obvious. For example, the upper-right ended in a live-live situation (again, size not mentioned) and a bit later White played at 'd'. The commentary then said Black took the opportunity to tenuki (another term that means subtly different things to Japanese and English readers - the real sense is not 'playing elsewhere' but 'skipping a move', which implies having to come back later). In fact he played two "skips" at 'e' and 'f', and the argument in favour of this way of play by Black (which was Go, the commentator) was that he was "thick" (another loaded term) at the top and so could fight in the corner. He ended winning by 12 points (no komi).

My point is that the typical western reader of such commentaries (this style really is the norm) is approaching them from the disadvantaged standpoint of not having much in the bank in the way of intuition, and even when he is reading the same book as a Japanese person, he is not quite getting quite the same information from it.

I'm sure all this goes a long way to explain why western players find it hard to get to the top. Realistically, we can't expect a situation where many western kids will play the game at home or school in the way Oriental players, do. As to books, I happen to believe that, ironically, many of the early English translations may have done a disservice to western players (the concepts of thickness = walls and yose = endgame are the glaring examples, but there are many others).

If the goal is to produce stronger western players, it may, therefore, be that case that our best opportunity is to use AI as the springboard for a new vocabulary - our own, based on a tabula rasa - and a new way of looking at go theory. Personally, I think the lack of opportunity to play just to build up one's intuitive play early in one's go career may be too much of a handicap for most people to contribute, and too few contributors would not create a groundswell. But there's no harm in trying, and just the effort of trying will lead to some improvement for the individual.

John Fairbairn wrote:[Western] books are more often than not translation of Oriental books, and so are full of strange concepts and strange words.

Not 'strange' but often 'non-Oriental'.

the dead-end path of logic.

Contrary to your metaphor, logic starts from and creates truths.

[Oriental] books that all mean the same thing

If only they did.

the size of any play is rarely discussed

In ko fights during the middle game, there are alternative evaluations to move values, such as judgements after ko dissolutions.

For other moves, move values are often not discussed because a) explanations skip the (late) endgame and b) calculation consumes more time and space than teachers spend when they have enough other things to explain so can avoid the tedious aspects.

Stronger western players would grasp these nuances, of course, but (I believe) they would not come to mind so intuitively and naturally - they'd be part of a conscious thought process instead.

Yes, likely.

all this goes a long way to explain why western players find it hard to get to the top.

There are several reasons. (One of them has become less severe: missing strong opponents are mitigated by AI opponents.)

If the goal is to produce stronger western players, it may, therefore, be that case that our best opportunity is to use AI as the springboard for a new vocabulary - our own, based on a tabula rasa - and a new way of looking at go theory.

I have been doing the tabula rasa approach to new vocabulary. I do not see how AI (pure moves and their statistics) provides any vocabulary - please suggest a related approach!

John Fairbairn wrote: In stark contrast, the typical western player skips the early intuition-building playing stage and learns from books instead (other book-fed western players).

This may have been true for us, who learned go before servers came into play. The difference between real-life go and server go aside, nowadays most western players play many more games than we did, and they often skip books. Instead they might turn to youtube for information, where at least Michael Redmond, then again the female Korean pro Yeon-Woo and Baduk Doctor have channels of varying educational quality. I dare say Nate (TelegraphGo) is doing a very good job in bringing pro level games within reach of amateur understanding. There's of course some lower quality content available too and overall the Internet poses the big threat of too much information, or rather infotainment.

What I have understood is that Newcastle United are doing well against a fancied team in the FA Cup (soccer) and I can now hope for their glory days of 1953 again. If we had that conversation surrounded by Newcastle football fans, most of them would understand instantly what we are talking about, even if they miss a few nuances. Everyone here can cite similar examples from their own experience.

You are making a good analogy here: despite the availability of online content on football/soccer, our collective intuition about the game is still levels above that of the average Chinese football afficionado, because we've been playing it since childhood.

If you look at Oriental books and analyse the language, you will find that a great deal of knowledge is assumed. That is, it is not articulated in the book because it is assumed to have been articulated, via the intuition at an earlier stage.

Much like a commentary on a football game might scream "OMG what's he doing???" when a defender heads the ball into the center rather than to the outside. We all know that's risky.

$$W
$$ -----------------------------------------
$$ | . . . . . . . . . . . . O . O . O X a |
$$ | . f . . . . . . . . . O X O O O . O X |
$$ | . . X O O X . O . X . O X X X O O X . |
$$ | . . X , . . . . . , . O X . X X X . . |
$$ | . . . O . O . O . X X O O X X X . X . |
$$ | . . O . . . . . b O O X X O O O O O . |
$$ | . . . . . . . . . X X . . . . . . . . |
$$ | . . . . . . . . . . O X X . . . . . . |
$$ | . . . . . . . . . . . X O O O . O . . |
$$ | . . . , . . . . . , . X X X X O . . . |
$$ | . . . X X . . . . . . . . O X O X . . |
$$ | . . . . O . O . . . . . . . O X X . . |
$$ | . . X . O . . c . . . . . . O O X . . |
$$ | . e . O X . . . . . O . . . . . . . . |
$$ | . . O O . . . . . . . . . . . . X . . |
$$ | . . X , . X . . . , X . . d . X . . . |
$$ | . . . X . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ----------------------------------------|

Click Here To Show Diagram Code

[go]$$W
$$ -----------------------------------------
$$ | . . . . . . . . . . . . O . O . O X a |
$$ | . f . . . . . . . . . O X O O O . O X |
$$ | . . X O O X . O . X . O X X X O O X . |
$$ | . . X , . . . . . , . O X . X X X . . |
$$ | . . . O . O . O . X X O O X X X . X . |
$$ | . . O . . . . . b O O X X O O O O O . |
$$ | . . . . . . . . . X X . . . . . . . . |
$$ | . . . . . . . . . . O X X . . . . . . |
$$ | . . . . . . . . . . . X O O O . O . . |
$$ | . . . , . . . . . , . X X X X O . . . |
$$ | . . . X X . . . . . . . . O X O X . . |
$$ | . . . . O . O . . . . . . . O X X . . |
$$ | . . X . O . . c . . . . . . O O X . . |
$$ | . e . O X . . . . . O . . . . . . . . |
$$ | . . O O . . . . . . . . . . . . X . . |
$$ | . . X , . X . . . , X . . d . X . . . |
$$ | . . . X . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ----------------------------------------|[/go]

Thanks for bringing a diagram into the discussion!

I was expecting you to give a reasoning by Go Seigen why the ko would be better than B. With one glance, I identified B as the point to play. Even evaluating the ko as being about 50 points (not going into the subtlety of approach kos reducing the size) I still wanted to play B. The only doubt I had was the livelihood of the white group on the right side. Anyway I would play B without much hesitation and only then start thinking about the ko fight.

I'm sure all this goes a long way to explain why western players find it hard to get to the top.

That, and the opportunity to play strong opponents from an early age, also in live circumstances.

If the goal is to produce stronger western players, it may, therefore, be that case that our best opportunity is to use AI as the springboard for a new vocabulary - our own, based on a tabula rasa - and a new way of looking at go theory. Personally, I think the lack of opportunity to play just to build up one's intuitive play early in one's go career may be too much of a handicap for most people to contribute, and too few contributors would not create a groundswell. But there's no harm in trying, and just the effort of trying will lead to some improvement for the individual.

Agree. I may seem to contradict my earlier post but my point is rather that, YES, we need good theory, concepts, "articulation" to provide for the fertile soil where western players can grow. But NO, after a certain point (5d?), the expertise no longer requires vocabulary driven knowledge but rather a word-less environment where the concepts are articulated through moves, sequences and positional evaluation, much like the interface AI is providing us with. I also don't think a deep study of e.g. miai values, articulated via game trees and fractional numbers, will make you leap from 5d to pro level. I can fathom though how a 9p with total knowledge of endgame values gets an edge over a 9p who doesn't.

I know there is distrust of youtubers for various reasons, but when I look at Baduk Doctor's game reviews, I see him mostly running through sequences and pointing out moves that he would never have thought of, or polish his intuition about why a group can be left alone despite its looking heavy. I'll make an exception for the concept that surfaces most in high level analyses by the likes of BadukDoctor or TelegraphGo: "a favorable exchange". This is of course almost a meaningless phrase in itself but underneath that superficial concept lies an understanding that we articulators may want to explore more.

kvasir wrote: As far as I know such signaling of how an exercise is going isn't considered conductive for learning in general.

Do you have some source for that? I'm genuinely interested.

kvasir wrote: [...] remembering what you were doing must be crucial for improving on it, right?

That's exactly why I think training with immediate feedback is worth pursuing.
From my own limited experience using the system I can say that the instant feedback for some huge mistakes and my natural emotional reaction of "Damn, I just f***ed up" has really "burned" those shapes into my mind.
Of course, the review afterwards was still necessary to find out how I should've played.

Concerning the discussion of articulated knowledge/reasoning vs. intuition:
I think articulated knowledge is immensely helpful for improvement, I credit a lot of my personal progress to some amazing books, but it definitely has its limits.
Firstly, it cannot cover all situations, as it's by nature a simplification/generalization of an immensely complex domain. And secondly - maybe more important - it is too slow.
We cannot possibly reason about all possible moves on the board in the limited time we have during a game, we dismiss probably 95% of all legal moves intuitively (depending on the stage of the game).
Our focus is drawn to some candidate moves that just "feel good", and from there we try to read continuations and evaluate the possible outcomes.
While reasoning is important for decision making in Go, intuition is essential for pruning the search space and also for evaluation.

On a side note:
The kind of "expert system" Robert seems to envision is actually how AI/Machine Learning was first approached in the 70s/80s, by trying to encode human knowledge in rule-based software,
I think the first chess programs (maybe also Go ones?) were implemented like this.
However that approach never went anywhere as it is too general and can't properly deal with unknown situations.

A last real-world example:
I've become very passionate about playing basketball over the past few years. I'm a "study" type and have invested a lot of time learning about the game.
I have so much knowledge accumulated in my head and it has definitely helped a lot with improvement, but actually applying this knowledge during a game is a whole other beast.
Decisions have to be taken in fractions of a second, ball and players are moving constantly, there is just no time to consciously process everything. Not to mention that your body has to be capable
of actually executing everything you want it to. But with repetition and experience these processes become more automated/intuitive and decision making becomes much faster.

The analogy might not be perfect, although - believe it or not - I actually discovered some related concepts in Basketball and Go.
I might write a post about that sometime

kvasir wrote:It is sometimes hard to tell what people mean when they talk about learning Go. To me that means improving at playing the game from start to finish and to the best of my knowledge you need to play games, from start to finish, without any external aid if you wish to improve at this. Everything else that you do is support for the main activity of playing games. However, there are other views.

I agree with your first statement: the goal is to improve at playing the game from start to finish without external aid. Let's call this "performance". The only way to know if you have improved your performance, is to perform more (play) and measure the outcome (review).
Your second statement is probably equivalent to my previous sentence but the categorization of everything else as tertiary activities suggests that all training which is NOT by means of performance is of a much lower importance. Here I'm not so sure.

It's always dangerous to draw from other fields, but it can also be instructive. Besides Go I play the guitar and table tennis. It is definitely NOT true that the main way to improve at the guitar is playing more songs. The main method of training is to isolate the tough parts in the song and focus training sessions on improving those. Then you bring it all together again to see if the performance is improving. Likewise in table tennis one doesn't improve efficiently by always playing matches. You work on your serve and other strokes in isolation. I can dwell on training programs in both areas but you'll agree with me just "performing" is not enough and you may agree that it doesn't constitute the bulk of training.

I do have a harder time to make similar constructions in Go. The common "isolated training" consists of joseki, life and death, and endgame. However, the impact on one's performance (winning games) of such training is not as strong as in the abovementioned skills. There are even proverbs saying (rote) learning joseki makes you weaker, suggesting the whole (board) always defies the sum of the parts.

I think it is counterproductive to have any such feedback during a game. Between games is a different story.

Not sure. Yes, if you are always doing that. Perhaps not if it's just a part of your training program. Why would "supervised learning" during a game be so bad? Why would you always have to figure out everything for yourself?

Probably the best thing you do besides playing games is a brief review of your games while the memory of playing them is fresh.

That's what we're having today. It's not bad. But I understand the OP is looking for an improvement on that.

Computer programs are of course great for exploring shapes and testing your reasoning about some positions, their main drawbacks is that they still tend to be fairly opinionated and they often prefer going down the rabbit hole over the simple solutions that often work better in practice. Their evaluation often isn't objective for human vs. human game and it gets worse the farther you are from their level of play.

http://www.neuralnetgoproblems.com is doing a good job at countering your argument.

At the risk of distracting from the main point, this Western kyu-level player has learned that the tortoise-shell shape formed by capturing at B is incredibly strong ("the tortoise shell is worth 60 points" is in John Power's proverbs book) and my eye is drawn to it even more than a ponnuki.

Knotwilg wrote:But NO, after a certain point (5d?), the expertise no longer requires vocabulary driven knowledge but rather a word-less environment where the concepts are articulated through moves, sequences and positional evaluation, much like the interface AI is providing us with.

Absolutely disagree.

I also don't think a deep study of e.g. miai values, articulated via game trees and fractional numbers, will make you leap from 5d to pro level.

Of course not. It is never only one aspect of the game that would be sufficient for much improvement. It is always the combination of all aspects of the game! Counting AND tactical reading AND positional judgement AND strategy AND psychology AND time management etc.

golem7 wrote:articulated knowledge [...] is too slow.

For complete analysis of early positions, it is too slow. For correct analysis of late positions or almost correct analysis of positions with certain characteristics enabling simplified analysis, it is fast; as fast as subconscious thinking but furthermore almost correct while the latter can be correct or wrong.

intuition is essential for pruning the search space and also for evaluation.

"Intuition" or analytical means enable this pruning. Therefore, "intuition" is non-essential. From the POV of analytics, means such as proximity or greater number (such as of stones) enable fast pruning.

The kind of "expert system" Robert seems to envision is actually how AI/Machine Learning was first approached in the 70s/80s, by trying to encode human knowledge in rule-based software,

No. The old computer go expert systems relied on very weak human knowledge, such as shape names. Contrarily I describe strong human knowledge, such as mathematically proved theory or much more generally applicable concepts. My knowledge is not explored far enough to allow strong computer expert systems, except for certain classes of already solved simplified / late game positions. Currently my knowledge is much more useful for human players than computer players because we can cope with knowledge gaps while computers still need to be programmed to combine expert input with deep learning whereever gaps remain.

However that approach never went anywhere as it is too general

"Too general" is a wrong description. Rather you mean something like: pretended to be more general than it was.

and can't properly deal with unknown situations.

Of course, it could deal with them if there was the fall-back to deep learning whenever generalising knowledge cannot solve something because of violated assumptions for its application.