Life In 19x19

Reviewing my game with Katago, I noticed how indifferent it is to where to play in very early game. Out of interest, for the example below I checked every move on 3rd line or above, up to tengen. (Well, not every move because I did it manually, but every move near the stones on board plus a sampling of the center.)

Anyway, in the position below it is WHITE turn, and the triangled moves are the only ones on 3rd line or higher that lose more than 1.5 points in Katago estimation. In fact there are few moves that lose even 1 point, such as the 5th line nobis from the white star points, as well as L3, N4 which appear to be bad shape. The point comparison is to F17, which is presumably the best move for white.

Don't know how interesting this is but since I did the exercise, thought I'd plop this here too.

Yes, this is of some interest,

However, I think that for anything goes, the correct measure is a loss to par of less than ½ point on average, not 1½ points.

That's by territory scoring, since a loss on average of ½ point by territory scoring can become, given perfect play, a loss of 2 points by area scoring.

I have noticed when analysing my own games that barring any obvious blunders, it is very difficult to lose more than 2-3 points per move in the opening. In a way, this makes sense since the middle game makes or breaks a game, however, your experiment has made me wonder something. I find it difficult to understand how some white moves could only lead to a 1 point loss. Perhaps I'm thinking too much like a human but many white moves would surrender the control of the game to black and my instinct is that this should lead to a loss of at least several points. Maybe the explanation is that AI would tenuki afterwards and leave the suboptimal stone as aji for later, thus keeping the control. Human players, on the other hand, often prefer responding locally and are also more likely to get frazzled when continuous pressure is applied.

How many playouts did you do? I did 10k playouts for that position in KaTrain (Katago Engine) and the corner move evaluation looks like this: With Passing as -11 as a kind of floor on how bad a move can be here. I find it useful with early game moves like this to think of them in terms of fractions of komi. R5 in particular here is like saying "I wanted to play with no komi." Most of the others like O3 are more "well how much worse is this than N3 if you want to play around here." The difference between P2 and N2 in evaluation here also makes a lot of sense. Both are bad but one of them really has the wrong idea. N2 is at least trying to go in the right direction. Most of this here makes some sense to me, though I doubt I grasp a lot of the finer details.

My weak human brain thinks N3 is good here because the combination of F3 and R15 make us not want black to play here and press us down.

Second though on using these things for training. For kyu players especially, figuring out why some mistakes here are considering much worse than others is perhaps quite valuable. Why is P2 so much worse than N2 or P3? Or better why is P2 so much worse than L3? I think there's some value in trying to figure out why the AI evaluates all these moves differently. Especially when it comes to moves we'd not consider.

I have a reversed/reworded kind of insight: only the middle game really matters (at my level). This year I have played and analyzed more than 80 games and invariably the 3 biggest issues occurred in the middle game.

The reason why there's more instructional material on the opening and the end game is that they are more easily wrapped into patterns and common shapes. The big middle game tactics often involve a large part of the board with a particular shape at the heart of it.

So even if the middle game holds more potential for affecting the end result, the more "patternized" stages may allow for easier learning.

I think for humans it not only matters how many points a move gains or loses if you can play almost perfectly.
Many positions will be hard to play for a human, even though the AI says it's even. And this matters a lot to humans.

I think it's hard to quantify how easy or difficult a position is to play for a human, but I expect that it will be much harder to play well from a random game position, than it is to play well from an Edo period pro game position.

I have a reversed/reworded kind of insight: only the middle game really matters (at my level). This year I have played and analyzed more than 80 games and invariably the 3 biggest issues occurred in the middle game.

The reason why there's more instructional material on the opening and the end game is that they are more easily wrapped into patterns and common shapes. The big middle game tactics often involve a large part of the board with a particular shape at the heart of it.

Given that you've done a lot of thinking about this, could you tell us what you now think about the following propositions:

1. We tend to learn openings first which we mostly think means learning patterns.

2. Imbued with a habit of looking for patterns, we carry this over to the second phase, the middle-game (and since there are actually some patterns in the m-g, we get sucked in further.

3. We soon get out of our depth with m-g patterns.

4. The patterns we think we see are actually chimeras.

5. We would do better to forget patterns and focus more on general advice and common sense. For example, follow the Ten Go Maxims, such as Reject the Small and Take the Large. Or: Where the Opponent is Strong, Defend in Advance. Or, in more modern terms: Thickness is only thickness if you can use it as thickness.

I ask because my hunch is that almost all amateurs almost all of the time, at least up to mid-dan level, forget basic such common-sense. Furthermore, such common-sense applies throughout the game, and even if you do sometimes use common-sense, by eliminating the time you don't, you will get a far bigger (and instant) improvement, relative to patterns, by applying it more often.

I would go further and suggest that those who do focus on shape tend to be lopsided in their approach: they stress creating shapes and looking for their weaknesses, whereas AI stresses destroying shapes and creating overconcentration (i.e. efficiency over prettiness).

My reading of the OP is that this is essentially what the poster discovered - he mentions applying common sense. What he was left with was a huge array of good possibilities, all nothing to do with shapes or patterns and everything to do with plain old run-of-the-mill no-nonsense down-to-earth CS.


In the same way that BS baffles brains, maybe CS baffles pattern recognition.

gennan wrote:I think for humans it not only matters how many points a move gains or loses if you can play almost perfectly.
Many positions will be hard to play for a human, even though the AI says it's even. And this matters a lot to humans.

I think it's hard to quantify how easy or difficult a position is to play for a human, but I expect that it will be much harder to play well from a random game position, than it is to play well from an Edo period pro game position.

20th century opening theory was derived from 19th century opening theory, passing through the New Fuseki. One thing AI bots are telling us is that 20th century theory overvalued sides, given 4-4 point openings. Humans had already begun to suspect that in the late 20th century, but today's top bots really prefer to have two stones in at least one adjacent corner before extending to a side. OC, humans in the 19th century for the most part had those stones in the corner before extending to a side. So, yes, 19th century pro positions are easier to play for humans than 20th century pro positions.

But humans are good at learning patterns, and before long we will see top pros who have grown up in the AI era. I think that pro play will have advanced by at least one stone during this decade.

gennan wrote:I think for humans it not only matters how many points a move gains or loses if you can play almost perfectly.
Many positions will be hard to play for a human, even though the AI says it's even. And this matters a lot to humans.

I think it's hard to quantify how easy or difficult a position is to play for a human, but I expect that it will be much harder to play well from a random game position, than it is to play well from an Edo period pro game position.

Well it's similar to a joseki dictionary before AI. For a pro lines A, B and C may all be playable in a tricky joseki position. If you can only make sense of A then B and C get shelved for later. If the AI said a move is 0.3 points better that the move you like here and you cannot see why it's better, then changing your move is like similar to playing the B line or C line above "because the pros play it."


From a *study* point of view it's different in both. Figuring out why B and C work is similar to studying the 0.3 point playouts to try and grasp what's going on. In both cases you may end up sticking with your first choice, the point of studying isn't to make you a B or C line player after all.

John Fairbairn wrote:

I have a reversed/reworded kind of insight: only the middle game really matters (at my level). This year I have played and analyzed more than 80 games and invariably the 3 biggest issues occurred in the middle game.

The reason why there's more instructional material on the opening and the end game is that they are more easily wrapped into patterns and common shapes. The big middle game tactics often involve a large part of the board with a particular shape at the heart of it.

Given that you've done a lot of thinking about this, could you tell us what you now think about the following propositions:

1. We tend to learn openings first which we mostly think means learning patterns.

2. Imbued with a habit of looking for patterns, we carry this over to the second phase, the middle-game (and since there are actually some patterns in the m-g, we get sucked in further.

3. We soon get out of our depth with m-g patterns.

4. The patterns we think we see are actually chimeras.

I give us more credit than that. The main advantage today's bots have over us is that they play the averages better than we do. As far as reading goes, we are better at depth first search, they are better at choosing candidate plays. But we can learn how to choose many candidate plays from them.

John Fairbairn wrote:5. We would do better to forget patterns and focus more on general advice and common sense.

The two are not mutually exclusive. In fact, they go well together for humans, IMHO. We are good at both.

John Fairbairn wrote:I ask because my hunch is that almost all amateurs almost all of the time, at least up to mid-dan level, forget basic such common-sense. Furthermore, such common-sense applies throughout the game, and even if you do sometimes use common-sense, by eliminating the time you don't, you will get a far bigger (and instant) improvement, relative to patterns, by applying it more often.

This sounds like what I dubbed the kyu disease, i.e., sticking to learned patterns instead of reasoning about the position.

John Fairbairn wrote:I would go further and suggest that those who do focus on shape tend to be lopsided in their approach: they stress creating shapes and looking for their weaknesses, whereas AI stresses destroying shapes and creating overconcentration (i.e. efficiency over prettiness).

This is something that I noticed, as well. When bots choose what I consider shape plays, it is typically preventing the opponent from making good shape.

John Fairbairn wrote: 1. We tend to learn openings first which we mostly think means learning patterns.

That's one side of it I think, the other side being the available material. It's somehow easier to teach openings, in terms of patterns (joseki, the Chinese opening patterns) than middle game heuristics.

But you're right: even a heuristic oriented book like Attack & Defence has turned many into adopting the boshi (which has been partially denounced by AI).

John Fairbairn wrote: 2. Imbued with a habit of looking for patterns, we carry this over to the second phase, the middle-game (and since there are actually some patterns in the m-g, we get sucked in further.

Yes, learning the proper invasion points of a mixed 3 space extension, gives more immediate satisfaction and grip than "attack for profit".

John Fairbairn wrote: 3. We soon get out of our depth with m-g patterns.

Yes. None of those 240 mistakes surfacing from my 2021 review so far had to do with a pattern. That's self fulfilling of course: since I know many of those patterns, that's not where my errors come from.

John Fairbairn wrote: 4. The patterns we think we see are actually chimeras.

Not sure about the metaphor . I know what a Chimera is but not what you mean by it in this context. We do have the metaphor of a "seven headed monster" in Dutch, which means, hard to defeat, hard to master ... That's not what you mean I guess.

John Fairbairn wrote:
We would do better to forget patterns and focus more on general advice and common sense. For example, follow the Ten Go Maxims, such as Reject the Small and Take the Large. Or: Where the Opponent is Strong, Defend in Advance. Or, in more modern terms: Thickness is only thickness if you can use it as thickness.

Well my research was a very personal one. I find little guidance in proverbs which leave a lot of room for interpretation. I never liked "Lessons in the fundamentals" for that reason (it IS witty). Although I must admit "the stones go walking" has been inspirational throughout, as a maxim that dictates many patterns - and then again, sometimes I keep walking for too long.

I haven't concluded on the research but a few "patterns" in my mistakes have surfaced: slow connections for example. Likewise, cutting hot headedly, resulting in a weak group and spontaneous territory for my opponent. I probably overestimate the value of connecting and cutting, which is a very good beginner heuristic. But does that hold for good advice regardless of your personal history? I don't have the intention to write new proverbs. I'll publish my personal corrective research eventually on SL - if it serves others, good.

John Fairbairn wrote: I ask because my hunch is that almost all amateurs almost all of the time, at least up to mid-dan level, forget basic such common-sense. Furthermore, such common-sense applies throughout the game, and even if you do sometimes use common-sense, by eliminating the time you don't, you will get a far bigger (and instant) improvement, relative to patterns, by applying it more often.

The governing issue in my (middle) game is that I get carried away with what I'd like to happen. I see a sequence, I think or wish that it gives me a certain result and then I stop checking whether it actually does. That's where my overplays come from. My underplays come from complacency. The situation is difficult and instead of thinking I play "honte" which is actually very slow.

John Fairbairn wrote: I would go further and suggest that those who do focus on shape tend to be lopsided in their approach: they stress creating shapes and looking for their weaknesses, whereas AI stresses destroying shapes and creating overconcentration (i.e. efficiency over prettiness).

That's a good way of looking at it. I like the "efficiency over prettiness" maxim - as a corrective tool. In my case, the table shape is something I need to unlearn. Many of my mistakes are slow or otherwise inefficient table shapes.

John Fairbairn wrote: My reading of the OP is that this is essentially what the poster discovered - he mentions applying common sense. What he was left with was a huge array of good possibilities, all nothing to do with shapes or patterns and everything to do with plain old run-of-the-mill no-nonsense down-to-earth CS.

I think any amateur can be liberated by playing opening moves which they never considered but which can't be wrong by common sense, only marginally inferior by AI/pro knowledge or fad. But the heuristics in the opening are more absolute than in the middle game: "avoid being surrounded" vs "be ready to sacrifice the small for the large at any point". Being surrounded is almost binary. Being small isn't.

Thanks for the good list!

Not sure about the metaphor . I know what a Chimera is but not what you mean by it in this context.

The English metaphor refers to a thing which is "hoped for but is illusory or impossible to achieve" (OED). I confess I've never seen how that derives from the Greek for a "she-goat," but in the modern world we do talk a lot about illusions and rarely about nanny goats.

Being surrounded is almost binary. Being small isn't.

And this is where I have the chimera of understanding. I know what all the words mean individually, but there are no receptors in my brain for this sort of language to latch on to, and so the whole sentence just floats on by in a big blob far into the ether.

The situation is difficult and instead of thinking I play "honte" which is actually very slow.

Earlier this week a friend of mine came up, out of nowhere, with the sentence "Well, ending up in gote all the time never did Shuei any harm." I though this was the best sentence about go I had ever head. It was true, It was insightful. It was unexpected. It was novel. It was valuable.

I therefore immediately read your sentence through that filter. I may have been primed to do that also by the fact that I've been pondering about such matters in the context of old Chinese go. They don't have a word for honte but they express similar ideas all the time. For example, a very common locution is "Move 10 is for safety." They also talk a lot about a player "sets himself at ease" or "gets peace of mind." These are general ideas which I class as "common sense" as opposed to "pattern recognition." They don't point to a specific move, but to a specific idea. And they are nearly always gote (or "slow")!

Nevertheless, there is a way to be more specific about such moves in old Chinese. They use the character 扼 (which you will never see in a Japanese go book). I have seen it rendered as "vital point" but that is confusing. It does not refer to a nakade move or an eye-stealing tesuji, but to a defensive move that "guards" a vital point. This is superficially similar to a honte. However, 扼 has another meaning. The other sense is one of restraining, which obviously means restraining the opponent, and so such moves are effectively "binary"! They occupy a vital strategic point and work for you and against the opponent. The imagery that flashed into my mind when I first got a grasp on this kind of move was NOT of a honte as a castle that is mainly designed to keeps the occupants safe (which is how I think most people view a honte), but rather of a fort in the Wild West. Yes it had a defensive function to keep the US Cavalry safe, but it was also (and mainly) the base for raids on the indians. This imagery may not mean much to young people now that PC rules in Hollywood, but it was a staple of films when I was young. As it happens I always supported the Indians, but I would also spend hours making forts out of matchsticks...

The comment about Shuei made a lot of sense to me in the light of this Chinese concept. I think it is worth everyone here having a re-think about honte. In particular getting away from the ideas that this is solid "shape" and/or is "slow", and considering it more along the lines of whether it has a "binary" effect.

Incidentally, I haven't got a good translation for 扼 in old go, but what I am working with at the moment is "important restraining move." Important naturally implies that it's for you and restraining implies it's acting on him.

I think any amateur can be liberated by playing opening moves which they never considered but which can't be wrong by common sense

This use of the word 'liberated' is magical and the whole sentence has a shattering Tristan-chord kind of effect relative to the usual environment of go thinking. I include myself and pros. Almost every commentary I read leads to me saying "Oh, I didn't realise you could that," but even pros are "unliberated", as AI has shown. I even think this sentence belongs on a congress T-shirt!