Life In 19x19

I don't know if it pertains to this topic but it relates to the latest tangent: I think many people just don't fathom how good programs like KataGo are at Go. I will give an example.

I played a 5 stone handicap game with the handicap KataGo. The entire set of my mistakes larger or equal to 1 point was 4.5, 2.9, 1.9, 1.8, 1.7, 1.6, 1.6, 1.6, 1.6, 1.5, 1.3, 1.3, 1.1, 1.0 or about 25.4 points in total. Nevertheless I went from being ahead by 57.5 points to losing by 6.5. The noticeable mistakes are only 25.4 points or 40% of the total of 64 lost points.

This is based on the 1000 playouts done during the game. When I reanalyze the game with ten thousand playouts there are slightly more noticeable errors: 4.4, 2.5, 2.1, 2.0, 1.7, 1.6, 1.6, 1.6, 1.5, 1.3, 1.3, 1.2, 1.2, 1.1, 1.1, 1.0 but still only 43% of the total lost points.

Basically, even after eliminating all noticeable errors it would still be hard to keep up with KataGo in this setup at 3 stones (let alone actually win). I find the idea of gaining 5 stones by correcting moves that immediately stand out as bad as you see the response to be funny. With my setup at least I can't say I'd really notice the moment I am getting creamed. If I notice an abrupt change in the course of the game it is always way too late.

I thought the reviving-comment was a bit dismissive. Even if these lectures are not up to AI's standards, and even ignoring the AI and comparing these lectures to other better lectures, and even if pros are not playing this way anymore, it's not as if these diagrams are useless to all learners. I had never reviewed these before and so I went back and while there might be something better out there, I learned something. At the very least the lectures reminded me of concepts I should be considering during the game.

I find AI analysis especially useless and many of the AI's top suggestions are actually bad moves for me because they lead to a sequence that I could not hope to complete. It's one thing for a Pro to follow AI suggestions. But Pros know better than to teach using AI suggestions.

Anyway, I thought by 2022 that we had grown past pointing to old ideas and laughing with the AI.

CDavis7M wrote:At the very least the lectures reminded me of concepts I should be considering during the game.

I agree. Throwing away all previous theories is a bit like throwing away Plato in Philosophy or Marx in Political Theory or Keynes in Economics.

When learning a subject, if you're going to gain a wide proficiency, you need to gain experience of using, considering and evaluating concepts - some of these may have been overtaken. But it's useful to be aware of them. And sometimes they get revived.

When learning a subject

At last! Sanity. Too much of debates like the current one are obsessed with the word 'strong'. Go has survived and flourished over two millennia because it is game we can play to win, yes, but also appreciate.

There is nothing new about this. 99.9% of people who follow baseball would never even get a whiff of a cup of coffee in the Major Leagues, but look how many of those fans are AVID fans. A very high proportion of those fans can't toss a ball more than 10 feet (witness celeb starting pitch after starting pitch), they might have hit a ball with a baseball bat half a dozen times in their school lives (and still bask in the glow of achievement). But they turn out in their thousands every day and follow every game, memorising stat after stat. Some are so keen they even try to understand what a baulk is. That's commitment for you.

We need such commitment in go. We need people who want to learn about the game, in all its facets, and not just obsess with getting strong enough to solve a 6-kyu tsumego problem once in a while.

(Takes tongue out of cheek - but may re-insert it soon.)

.

John Fairbairn wrote:
There is nothing new about this. 99.9% of people who follow baseball (...)

But you see ... baseball is not like Go. You can appreciate sports even as a couch potato. You can't really enjoy Go if you don't know what's going on. No need to be as strong as the players but the more you understand, the more you appreciate. Pro commentaries of pro games help me understanding what's going on and appreciate them better. The nature of appreciation and the act itself are much closer in a mind sport than in a physical sport.

I'm not obsessed with getting stronger. I am indeed kind of obsessed with improving my understanding. But that boils down to the same thing, in Go, except that - as you pointed out - there's some discipline required during the operation which isn't needed during the appreciation.

BTW, this series is called "how to become dan" so I criticize it on the claimed merits.

John Fairbairn wrote: We need such commitment in go. We need people who want to learn about the game, in all its facets, and not just obsess with getting strong enough to solve a 6-kyu tsumego problem once in a while.

So who in this call is NOT like that?

kvasir wrote: I find the idea of gaining 5 stones by correcting moves that immediately stand out as bad as you see the response to be funny.

Thanks for your analysis BTW

The above statement reminds of another debate: regardless of how many stones you can gain at once, do you think it's the right approach to improve your game by inspecting the big errors first? This is what I'm doing in 2021 and 2022. Or do you believe, like I believe Uberdude does, that the size of the error should not be a heuristic, compound small errors are just as important, and the filter should come from what you want to learn about or any pattern that comes from the review? Which is what I gather from your reply.

My view on that is that you should focus on the most frequent errors and that you should not focus on the same thing for too long until you change your focus. Using the computer to find the errors and ignoring very small errors is mostly a practical matter. What is a 0.3 error in the opening? I don't really know but it seems that 1.5 point errors are already something that can be considered.

The reason to focus on the most frequent errors is that these are the errors that happen in most of your games. The very large errors may not impact your average game at all unless they are frequent. Fixing frequent errors should compound because once you work through many such mistakes it is sure that there will a lot fewer mistakes. Maybe not so if you only fix large errors without considering if that particular error is a symptom of something more general. Basically the points lost are not an indicator of importance but at least thinking about which errors are common and which are not should give some indication of what is important.

The reason for not focusing on the same thing for too long is that if you can't get rid of a mistake from your game after some moderate effort it might just be difficult to improve this part of your game. In those case it is better to move on and focus on something else, you can revisit it later.

Another thing that is important is that the computer analysis doesn't find all types errors. It really seems to hide somethings and only reveal over a number of moves that something was lost. That handicap game I posted is an example. It is not clear where many of the points were lost, it appears as if black's lead is slowly adjusted down without moves being marked as losing any points. I have also had this experience in even games were my opponent seems to gain on me but I can't find specific moves that the computer things are bad. That is there are some intangibles, maybe somethings that the computer doesn't evaluate correctly until it starts to affect the position more immediately.

Another possible defect of the computer evaluation is that when comparing two positions the evaluation is often suspiciously similar to just counting the points and ignoring other things (thickness, safety, potential, ...). I don't know how to refute the computer on such things but it sometimes appears to me that giving 2 points for better outside positions is exactly 2 points worse to the computer. This is not always the case but sometimes it appears to me that the computer doesn't care about anything except a rough count (a few tens of moves deep into a variation I guess, it is really strong in that regard).

Some of the largest mistakes that one makes are related to computer idiosyncrasy. The computer just never defends dragons, if you make eyes it is sure to be a point loss. The computer wants to fight almost every ko and it thinks it can invade any moyo. The point loss values in these cases are often not useful but they are often the largest that I see.

kvasir wrote: Fixing frequent errors should compound because once you work through many such mistakes it is sure that there will a lot fewer mistakes. Maybe not so if you only fix large errors without considering if that particular error is a symptom of something more general. Basically the points lost are not an indicator of importance but at least thinking about which errors are common and which are not should give some indication of what is important.

(...)

The reason for not focusing on the same thing for too long is that if you can't get rid of a mistake from your game after some moderate effort it might just be difficult to improve this part of your game.

(..)

Another thing that is important is that the computer analysis doesn't find all types errors. It really seems to hide somethings and only reveal over a number of moves that something was lost.

(...)

Some of the largest mistakes that one makes are related to computer idiosyncrasy. The computer just never defends dragons, if you make eyes it is sure to be a point loss. The computer wants to fight almost every ko and it thinks it can invade any moyo. The point loss values in these cases are often not useful but they are often the largest that I see.

These four points are among the best correcting advice I have received on my "mistakes" series.

Another possible defect of the computer evaluation is that when comparing two positions the evaluation is often suspiciously similar to just counting the points and ignoring other things (thickness, safety, potential, ...).

Really? Now that would be rather disappointing. I'll check that on occasion. It has never occurred to me that it could be anything else than the average score of the end positions of each evaluation, although lightvector at some point has argued it is something else indeed.

Aren't you confusing the point evaluation by KataGo with point evaluation of traditional programs?

kvasir wrote: Some of the largest mistakes that one makes are related to computer idiosyncrasy. The computer just never defends dragons, if you make eyes it is sure to be a point loss. The computer wants to fight almost every ko and it thinks it can invade any moyo. The point loss values in these cases are often not useful but they are often the largest that I see.

Great post. The only minor quibble I'd have is that the AI dragon tenuki tesuji is probably something we can learn from rather than just an idiosyncrasy. It is easy to get into 'can't quite read out if there's scope for 2 eyes -no harm adding a move to be sure' over-passive mindset.

I remember being surprised when a Chinese pro said in a game lecture he hadn't fully read out whether a group was safe to tenuki from before he played - but he estimated the group was "already around 80% alive, 90% alive in a fast game - but that a passive move felt definitely wrong."

Isn't these puzzling tenukis and refusals to defend seemingly large groups really a matter of positional evaluation? If an AI plays tenuki instead of defending the AI considers its choice more valuable, no?

We can learn from how the computer treats eyeless dragons, for sure, but our decision can still be correct even if it has a large point loss. The point loss is not a good indicator here. We can still be making a correct decision if we didn't have the slightest idea how to live or connect a large group or if it is was just too risky. Go is a game played by humans and using clocks, and we have limitations.

When it is an error it can be almost binary. It could be that you will obviously lose if you defend or that it is just too hard to kill -- no need to worry. The point loss in these cases can be all over the place. It is not that we should ignore such errors, it is rather that we should figure out what is important and what is not. I think considering how frequently this kind of thing happens is more useful than thinking about how many points it lost.

I call it idiosyncrasy because it is something that is very hard to replicate and the computer seems to go overboard in terms of what human players could replicate. KataGo often doesn't resolve the life and death until 20-30 moves into variations if you try to play it out, it is rather confident, and it does seems to see very quickly if it needs to defend or not. It is perplexing to me how it can so quickly decide that it doesn't need to defend when the struggle to live is long and hard. Every such position is unique in a way and KataGo finds many different reasons and techniques to avoid living dragons. We can study that but as to the question of which mistakes in our games are most important it is not necessarily these mistakes even though they often have a high point loss value. That depends on how frequent such mistakes are and on your ability to solve such problems.

To make it clear I don't think you should ONLY study your own mistakes. These are in many ways reflection of who you play frequently and how good they are at the game.

When reviewing my games, I never care about mistakes worth 2 points or less, but after reading this

kvasir wrote:The noticeable mistakes are only 25.4 points or 40% of the total of 64 lost points.

I am wondering if a better method would be to study AI alternative moves, even when the game move isn't considered as a mistake by AI.

I am wondering if a better method would be to study AI alternative moves, even when the game move isn't considered as a mistake by AI.

I think it's a psychological problem. It feels natural to try to spot a mistake and correct, and the bigger the mistake the more correction we think we will achieve. But there's a world of difference between avoiding mistakes and making good moves.

In my experience, in most disciplines, you learn mostly by being shown the correct way to do something - by imitation in other words. You are told to read the best books - you are not told to spot the mistakes in Shakespeare. You are told to keep your fingers on the right keys on the piano. You are not encouraged to play bum notes and sort out the mess later. In calligraphy you learn from following the best models - until modern times at lest, you are not supposed to indulge in a "self-expressive" scrawl.

There are disciplines where your task is to spot the flaw and fix it, e.g. medicine. But even there, there is good practice and bad practice in applying the remedy, and the good practice is learned by imitating the experts.

This advice is also the perennial advice of go experts, i.e. pros. They tell us things like "Play over a 1,000 games". Although it's been decades since I studied go seriously, I can still remember vividly my Eureka moments when playing over a pro game and suddenly realising, "Oh, I didn't know you could that!" I then imitated that move and suddenly found I had added an extra arrow to my quiver.

Given both the frequency of such pro advice and my own experiences in following it, I have been staggered over the years at how little most people seem to play over whole games. Most seem to obsess over portions of a game: josekis, the fusekis, tsumegos or the endgame. That's like trying to become a good tennis player without trying a backhand, or without learning to move around on court.

I think if an AI bot could talk, it would also say to us, "I learned to play well by playing over millions of complete games and imitating the best in each case." Aside from corrections imposed from the outside by programmers (e.g. for reading ladders), I don't think bots ever try to learn by correcting mistakes. Mistakes are spotted, of course, put are just put in the dustbin.

But even a commitment to imitation has to be carefully considered, in various ways. Choosing the best model to follow, for example, may be a matter of temperament - you want to play in a certain style, say. But, also, do you want quick results, which tend to be easy-come-easy-go, or are your prepared to take your time and achieve good habits that stick.

In go, the traditional pro advice has been to study a single player thoroughly (and then move onto another). Again going by my own experience, amateurs tend to make the choice on the basis of a style they like, and also they very rarely go onto a second player. I think that's a mistake, and not what is behind the pro advice.

There is a problem with playing over game records. Commentaries are actually quite rare. So you play over a game record partly blind, and on trust. The trust is that the moves are all good ones, or at least acceptable to a pro. But, despite our much vaunted claim that humans can tell you what is happening and bots can't, in 99% of games you play over, there is no commentary and so there is no human actually telling you what is going on. Nor do you get any advice on alternative good moves, or even what other moves the pro looked at. The only way you can acquire an understanding in such cases may be to see repeated examples by the same player in game after game. In other words, you try to absorb his habits. Doing what neural-network bots do but on a very small scale.

This is where AI scores. In a separate post I put forward my ideas on nexusology. That was just my own way of looking at something we can all see and interpret in our own ways: a bot will reveal to us all the main moves it considers as candidates, typically about 10 for much of the game. We still have to trust that these are all decent moves, but we are no longer blind.

Trust can be dealt with fairly easily. You need only to play against a bot by choosing a non-blue move, and you'll get a feel for how much that loses. My guess is that you may lose the game but you'll play far better than you did as a pure human. So you trust would not be misplaced unless you wanted to be Sin Chin-seo or Ke Jie.

So you can extrapolate from that and say to yourself, it's good enough if I can create a palette of candidate moves to choose from that more or less matches the palette that the bot creates.

My sense of how to do evolved from trying to categorise all the candidate moves a bot showed. Because I don't study go to become stronger (I prefer to be a fan who "appreciates" the game), I have not tested my method properly, but provisionally I got good results from categorising candidates moves at a first level as settling, colonising and overconcentrating moves. But these were all nexuses with multiple components (definitely not binary as in profit/thickness. At a second level I had other nexuses. E.g. one included trading/sacrificing/miai moves/tenuki.

By looking at positions through that prism and producing a list of ten candidate moves, I found I could produce a list that married up pretty well (moye around 75% much of the time) with the bot candidates (Lizzie/Leela's in my case). I could not say with any confidence which was the best move, but as I say, it's probably sufficient at my level to play any one of those moves. On the other hand, because of the categories I had imposed, I felt I had some human understanding of the moves listed. And at no point did I try to fix mistakes. My goal was exclusively to get a matching list. I stopped once I felt I had got satisfactory results, but I feel confident I could get better results if I continued on the same path (mainly by refining the categories).

Knotwilg wrote:
The above statement reminds of another debate: regardless of how many stones you can gain at once, do you think it's the right approach to improve your game by inspecting the big errors first? This is what I'm doing in 2021 and 2022. Or do you believe, like I believe Uberdude does, that the size of the error should not be a heuristic, compound small errors are just as important, and the filter should come from what you want to learn about or any pattern that comes from the review? Which is what I gather from your reply.

On bleak days, I think the core determinant of playing strength is pure reading ability. Without improving one's own reading ability, tackling errors is likely to lead to small changes within a similar rating band of overall strength. Having said that, I am enjoying Knotwilg's 2021/2022 series.

Seeing AI save groups and carry off near impossible invasions seems to reinforce that for me. I may have a nice palette of moves to choose from by studying AI, but when the stones come together it's going to be reading strength that determines the result.

Incidentally, don't pros say "the way to get strong/become a dan" is to solve lots of problems rather than play through games? The one's I've run into seem to anyway, though maybe it's not the traditional advice over the last 3000 years.