Life In 19x19

Originally featured on my study journal, I'm turning this into a series of exercises. This year I'm playing 1 game a day (10 games behind) and reviewing them with Katago, in particular the biggest mistakes, either swinging the probability of winning or losing a major number of points. It's those that I'm turning into ABC-exercises. One is my move, thus the mistake. Another is KataGo's move. And then there's another one, just to make it more interesting.

Click Here To Show Diagram Code

[go]$$ Mistake 1
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . b . . . . . |
$$ | . . . . O . O a O . X X . X O O O O . |
$$ | . . O O X O O X O . O X X O O X O . O |
$$ | . . O X X X X X . O O O X X O X O O X |
$$ | . . . . . . . O . O O X X . X X X X X |
$$ | . . O O X . . . O X O O O X . . . . . |
$$ | . . . . X . . X X X X X X . . . . . . |
$$ | . . . O X . . . . . . . . . . . . . . |
$$ | . . . O X . . . . . . O . X . X X X . |
$$ | . . . O X . . . . X X O . . . X O O . |
$$ | . . O X X . . . X O O . c X X O O . . |
$$ | . . O O X . . . X X O O . O O . . . . |
$$ | . O . . . . . . X O O . . . . . . . . |
$$ | . O X X . . . . . X X O . . . . O . . |
$$ | . O O X . . . . . X O . . . . X O . . |
$$ | X O X O . . . . . X . O . O O O . O . |
$$ | . X X X X . . . . X O O X . X X O . O |
$$ | . . . . . . X X X O . O X . X . X O . |
$$ | . . . . . . X O O O O . . . . X X O . |
$$ ---------------------------------------[/go]

Should White connect at ''a'', atari at ''b'' or play ''c''?

In light of this post viewtopic.php?p=264182#p264182 the correct move must be "a". Move "c" is just 2 points reverse sente I think (edit: actually M9 is not a point, unless White plays at N10 later, but anyway it is not very big), so much smaller than saving the two stones H15.

I think c is good. Threatening either to cut off two stones or to break into the white territory with two peeps from inside.

It's White to play...

Let me toss an idea out for discussion: the biggest mistake here is to make an ABC of mistakes.

I obviously don't know what will be featured in later examples, but I have very long experience of mistakes. I even made one in 1999.

My experience of mistakes, once you strip out those made because of carelessness, tiredness and such like, is that the most important are those where you an say: "Oh, I didn't realise you could do that!" In the same way, the moves in commentaries, joseki books. L&D and everything else that have made me stop in surprise (and still remember to this day) are moves where my reaction is: "Oh, I didn't realise you could do that!"

When I look at the response of pros to the unexpected plays by AI bots, over and over again what I see is that their response boils down to: "Oh, I didn't realise you could do that!"

I am inferring that Dieter's responses in his examples were of the same type.

What this seems to mean, to me, is that our most fundamental mistake is not including the surprising moves in our list of candidate moves. I suspect there is a subtle difference between the pro and amateur cases: pros have blind spots because they have not built these moves into their intuitions yet; amateurs seem caught out most by not realising how big certain moves can be. But, either way, the problem is a shortage of candidate moves in our heads.

If that thinking is along the right lines, it seems that restricting the candidates to ABC is the very opposite of the right way to go. The first step should surely be to extend our list of candidates moves to ABCDEFGH and beyond. The next, and more important, step of course is to rank those moves, but you can't rank them unless they are there in the first place.

I wonder, therefore, whether a more fruitful approach might be to list "all the moves I considered were" against "all the moves katago considered" and see if some pattern emerges, "all" being simplified to e.g. top ten rather than top three.

And if you didn't consider as many as a top ten, is that not part of the problem?.

John, this series intends to consistently look at the data of a 2d player (me) and understand what kinds of mistakes they are (I am) making. Rarely they fall into the category "Oh I didn't realize you could do that". They are rather "I know this but failed to apply my knowledge" or "I should have read this because the situation looks suspicious indeed".

The "Oh I didn't realize you could do this" is ubiquitous but the swing values are rather small.

If you think I am making a mistake by looking at my mistakes, fine. I don't. I can do it in private if you think I'm littering the forum. However, when looking into it privately, I thought "this is the kind of material you rarely see in instructional articles". So I gave it a shot. If there's low interest, I'll stop posting.

Please keep posting, I am interested. All my games are like that: I have several options A, B, C and I never know which one is best.

If you think I am making a mistake by looking at my mistakes, fine. I don't. I can do it in private if you think I'm littering the forum. However, when looking into it privately, I thought "this is the kind of material you rarely see in instructional articles". So I gave it a shot. If there's low interest, I'll stop posting.

That's an oddly prickly reaction. Most of my own posts can be construed as a desperate attempt to stimulate discussion here, so I am hardly likely to be suggesting you are littering the forum. Since my attempts to get discussions going are almost always rebuffed, maybe I should be the one to stop posting.

The "Oh I didn't realize you could do this" is ubiquitous but the swing values are rather small.

This I actually found very interesting (if it's genuinely reliable), and illustrates perfectly the value of discussing, adding to my own (and others'?) knowledge. It engenders the further point for discussion: if that type of mistake is so ubiquitous, why is it not being looked at more often?

And yet another point that flows from discussing: if you find the ABC approach valuable (and I never said that was a mistake - I just suggested widening the scope) but rare, I can suggest you look at some of the Ishi Press books. I think Breakthrough to Shodan was of that type, and (I memory serves) there were several series of tat type in Go World.

John Fairbairn wrote: maybe I should be the one to stop posting.

Please don't. In this case you started out by calling the theme/approach of the series a mistake. So yeah, that prickled me a little bit. But it fades out fast.

John Fairbairn wrote:
This I actually found very interesting (if it's genuinely reliable), and illustrates perfectly the value of discussing, adding to my own (and others'?) knowledge. It engenders the further point for discussion: if that type of mistake is so ubiquitous, why is it not being looked at more often?

I'm not sure if I expressed myself well.

In this ABC series, I'm zooming in on moves I played, which were evaluated way worse than its preferred move, i.e. big mistakes. I just checked the 12 examples I had prepared

In 7/12 cases I had the preferred move on my radar but thought something else was more urgent or affordable
In 3 cases I had no radar, I wasn't playing on autopilot.
Only in 2 cases KataGo's move was of the category "Oh I didn't realize you could do that"

This small sample suggests that my candidate selection is not as bad as I thought, or that it's not coming to the surface as much when sampling big mistakes.

Your question begs an exercise on my end: during reviews keep a tally of 1) I played the best move 2) I considered the best move but didn't play it 3) I didn't consider the best move + 2a/b-3a/b my move was significantly/not significantly worse

John Fairbairn wrote: What this seems to mean, to me, is that our most fundamental mistake is not including the surprising moves in our list of candidate moves. I suspect there is a subtle difference between the pro and amateur cases: pros have blind spots because they have not built these moves into their intuitions yet; amateurs seem caught out most by not realising how big certain moves can be. But, either way, the problem is a shortage of candidate moves in our heads.

I will echo this idea, although I think amateurs may also be caught by evaluating areas as being larger than they really are because they don't know of the existence of a critical weakness. I'll give an example where "I didn't realize that". I saw it about 20 years ago and I still remember it vividly.

The basic position is here, and the question is what kind of scope for action there is on the right side.

Click Here To Show Diagram Code

[go]$$
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . O . O . . . |
$$ | . . . . . . . . . . . . O . O . O . . |
$$ | . . . . . . . . . . . X . O X O O . . |
$$ | . . . . . . . . . . . . X O X X X . . |
$$ | . . . . . . . . . . . . . X . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . a . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . a . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . X . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . O . . . |
$$ | . . . . . . . . . . . . . O . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]

Clearly there's scope for play and I naturally focused on the points marked 'a' as well as few others. But in two pro games the initial move was a deep invasion, but elsewhere. I'll hide the continuation to let people think.

I think what we don't see is this ko: Now that I'm out of the initial publishing frenzy, I can take a step back to appreciate John's point.

For educational purposes I limited the choices to 3 because I wanted to highlight my big mistake vs the proper move and then all good things come in 3.

That last bit is where the space opened up. When it's about playing good go, for educational purposes it might be better to offer all the possibilities. But that's another exercise than what I wanted to show.

I thought that if I would show the mistake and the good move only, the exercise would become too easy.

To comment on this:

Knotwilg wrote:In this ABC series, I'm zooming in on moves I played, which were evaluated way worse than its preferred move, i.e. big mistakes.

While it's important to eliminate big mistakes, I think that to improve it is necessary to look at small mistakes as well. Typically when a player is several stones stronger than another, Katago's graph has more or less a constant slope, plus or minus a few accidents, showing that the stronger player consistently makes better moves. Imagine if all your moves were 1 point better on average, what would be the difference after 100 moves?

To confirm that impression I analyzed a game I played last Sunday. First, the game without comments (186 moves). This game was EGF-rated, with 1h main time + Canadian byo-yomi 15 stones/5 minutes:
And then the game with comments (only up to move 126) with a tentative explanation for each mistake by White.
Almost all the game was quite balanced: until move 176, none of the players made 10+ point mistakes but both made a roughly equal number of 1-5 point mistakes, which could be categorized as followitis, failure to consider tenuki, failure to look at the opponent's weaknesses, etc.

So I would say that correcting our way of thinking about the game could potentially eliminate a large number of small/medium mistakes, which overall would have more impact than eliminating a small number of big mistakes.

Now, I estimate that we are about 4 (or maybe 5?) stones weaker than Knotwilg, so our fundamentals are not as good, but still I suspect that the above conclusion still holds at low dan level.

Hi JLT

The major reason for me to zoom in on "big mistakes" is that KataGo (or any AI) has a certain margin of error, which is higher in case of few playouts. When doing a full game review, I'm not regularly pausing to get huge playouts, although I will pause around the spikes to verify where the problem really started.

My threshold is around 5 points AND 10%. Why AND? Because a >5 mistake with a low change in % means the game is unbalanced already. And a >10% for a 1 point mistake means the game is still playable.

Next I'm looking into all games, finding the three biggest mistakes. I'm at 80 games now so I already have 240 mistakes to learn from. Patterns emerge and if I can correct these, I'll become better. That's the big IF.

Learning from KataGo's half point improvements would require lifting my level of understanding to pro level. At my level, it is better to regard them as noise.

Thanks for giving so much attention to the thread! It encourages me to continue.

Knotwilg wrote:I think what we don't see is this ko:

$$
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . O . O . . . |
$$ | . . . . . . . . . . . . O . O . O . . |
$$ | . . . . . . . . . . . X . O X O O . . |
$$ | . . . . . . . . . . . . X O X X X 3 . |
$$ | . . . . . . . . . . . . . X . 6 5 4 8 |
$$ | . . . . . . . . . . . . . . . 9 . 7 . |
$$ | . . . . . . . . . . . . . . . . 1 . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . 2 . . |
$$ | . . . . . . . . . . . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . a . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . X . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . O . . . |
$$ | . . . . . . . . . . . . . O . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------

Click Here To Show Diagram Code

[go]$$
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . O . O . . . |
$$ | . . . . . . . . . . . . O . O . O . . |
$$ | . . . . . . . . . . . X . O X O O . . |
$$ | . . . . . . . . . . . . X O X X X 3 . |
$$ | . . . . . . . . . . . . . X . 6 5 4 8 |
$$ | . . . . . . . . . . . . . . . 9 . 7 . |
$$ | . . . . . . . . . . . . . . . . 1 . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . 2 . . |
$$ | . . . . . . . . . . . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . a . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . X . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . O . . . |
$$ | . . . . . . . . . . . . . O . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]

Hmmm. I think that the colors of the moves are reversed in that diagram. Here is the diagram with the first move colored White.

Knotwilg wrote: Learning from KataGo's half point improvements would require lifting my level of understanding to pro level. At my level, it is better to regard them as noise.

Certainly what works for different people is different. But for most people, my feeling is that this is a common misconception. I mean, consider this:

Click Here To Show Diagram Code

[go]$$c Is this a mistake?
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . X , . . . . . , . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . O . . |
$$ | . . . . . . . . . . . . . . . . O X . |
$$ | . . . . . . . . . . . . . . . . O X . |
$$ | . . . O . . . . . , . . . . . O X . . |
$$ | . . . . . . . . . . . . . . . O X . . |
$$ | . . . . . . . . . . . . . . O X a . . |
$$ | . . . . . . . . . . . . . . . 1 . . . |
$$ ---------------------------------------[/go]

This move is judged to lose only about a point compared to connecting at "a" (which is more-or-less the best move on this board, in fact). But only a point! So are we wrong to spend the time to teach beginners to connect solidly on the second line as the default move except when there is a specific need for another move? I don't think so. This is a basic educational shape that is common - it occurs multiple times per game, and is easy to learn, and in combination with some other common edge patterns, helps build some foundational intuition for beginners about basic tactics near the edge.

In general, I think focusing only on the magnitude of a mistake in points misses these two critical elements:

1. Commonness of the shape - if a move in a particular shape is only a 1 point mistake but the shape occurs once every other game, then it could easily be many times more important than a tactical blunder that loses 20 points but where that particular tactic won't show up in that particular way in the next 100 games.

2. Ease of recognition and learning - if a move or concept is recognizable once you see it and fits well with your existing intuitions and knowledge so that you can learn it easily, it could easily be far more gain per time or effort expended than other scorewise-larger mistakes.

The ability to help with these two things is of course one of the reasons why having a stronger human player or teacher review your game can be far more effective than self-review with AI. But I can think of still a number of ways to partially get at these things with AI, if not so hasty to immediately filter by mistake size in points. I haven't played as much recently, but some things I've felt were useful personally in the past include:

* Scroll through the game-being-reviewed quickly and just see what the AI wanted to play and what its instincts were, and look for any moves the AI suggests that differ from yours, even if only better by a point, but where upon seeing the move and pondering it a little, it's like "oh duh, that shape is a lot better, I like that move a lot" or otherwise where it's clear upon seeing the move why it's better. You DON'T need high playouts for this, you're looking for major differences of instinct that you can absorb. Of course you might be wrong and not *really* understand, but if your brain is saying "I can intuit this move, I can learn it", that move probably *is* on average easier to intuit and learn.

* Also in that quick scan, looking for moves with recognizable shapes that are less "oh duh" and more like "wow I didn't know you could do that, but I see why it works" or even "I don't fully understand this move, but I understand enough to know that if I were the opponent facing this move, I'd find it extremely annoying and hard to find a good response to". Those could be good to try in future games. Again, you don't need high playouts for this either.

* Before looking at "the answers", doing a quick self-review pass to note positions where you were genuinely uncertain about the right shape, or the right tactic, or the right direction. Those might be good points for learning. Of course, if the AI recommended variation is arcane and incomprehensible (and it's still arcane and incomprehensible even when you interactively take the opponent's side and play "what if" a little bit, which often clears it up), move on. But if it's clear, then that's a good place to start. For me, a lot of mistakes are fall into the category of "things you don't know you didn't know", but if you *do* know you don't know, why not take advantage of that? And if in fact you got it right, the fact you were uncertain maybe makes it still a good learning sample, even though it wasn't a mistake.

* Focusing mildly more on mistakes, even if smaller, in semi-common corner patterns or invasions or josekis or whatever. Patterns that are more common are more likely to be relevant in future games. Similarly for things in midgame fights where despite the fight itself being unique the shape or tesuji within that fight that was missed is recognizable and familiar in retrospect - being more common makes it more useful.