Intro

There’s a class of problems called “Local Choice”. “Local” implies it’s not a whole board problem. “Choice” means it’s a multiple choice problem: you choose between enumerated options. In this problem, you choose between a, b, and c for the best move.

The goal is to create a class of problems that represent a common weakness in players: making local tactical decisions that aren’t life and death or tesuji situations – they are just about getting the best local result possible in terms of points, sente, influence, weaknesses, etc.

These problems are generated by KataGo (with the tool autoprob). They are taken from amateur games, so represent real world situations. The generator looks for a variety of characteristics:

• Not too many stones, not too few (about 10-30)
• One move is clearly the best by a good margin (approx 4 points at least)
• Other moves look at least reasonably tempting to a human
• The top move is the best place to play even if the rest of the board is empty

Part of the generation process creates responses for mistakes, as well as text that tries to explain why some moves are good and some are bad. This process relies on a lot of heuristics, which are not guaranteed to always work well or give good human hints.

I want the site to have a large amount of these problems. Enough that we can use them in things like tournaments and have them not repeat or be memorized, and so people can drill on them endlessly. We want a wide variety of difficulties on these problems too as much as possible.

One nice property is these problems are always correct, in the sense that thanks to superhuman KataGo nets, the moves are always the right moves. However, there are still many challenges in importing and creating problems in terms of making them good for users.

Currently the correct answer is a single node. No further sequence gets played out. This may be improved over time but it’s not always simple, trying to make a good tree from these positions. The downside is that it’s hard to see why the move works sometimes.

These problems are collected here:

https://www.goproblems.com/problems/collections/252

Ongoing improvements

It’s possible for problems in this category to be improved over time, while live. With Suggested Edits especially, users may contribute things like better explanatory text.

It’s also okay if we end up deleting some that are confusing or have other issues. Better to keep the quality high.

Problematic problems

There are a variety of reasons some of the generated problems may be a bad fit and not a good idea to add to the site. The generator tries to screen out these problems as much as possible, but it’s not perfect. Human oversight is needed. In fact, probably only ⅓ of the generated problems are worth adding. The good news is that it’s easy to generate infinite options so it’s best to err on the side of rejection.

Let’s go over some of the common challenges, and what to do about it.

Joseki position

Anything that is directly from a joseki shouldn’t be added (or even a single wrong move in a joseki). These situations are best handled in separate joseki testing which will be created independently. Unfortunately there’s no automated way to detect this (yet). So you have to eyeball it from experience. It’s not the end of the world if some get through, but try to do the best possible.

Too obvious answers

We try to screen this automatically, but some slip through. For example, the previous move is an atari, and the response is a completely obvious connect or capture. Some of these are okay but too many are a waste of time.

Just a life and death or tesuji problem

These problems shouldn’t involve the life and death of major groups. That’s a different kind of problem and not suited here.

Ko Fight

Generally ko fights don't make good local choice problems since they often require sequences of playing away and evaluating threats or tenukis in other parts of the board.

Reasoning relies on complicated tesujis

The situation may have an answer that only makes sense if you see a complicated tesuji or life and death sequence to be played later. These aren’t ideal because the user doesn’t really get why a move is correct.

If the followup is a simpleish move, the text can explain it.

Example 1

D is correct but it’s not obvious why S3 for black next doesn’t refute it. It’s because Q3 is then sente above so white can next turn at S2. This would need to be spelled out in the problem text, or even better, add this as a variation manually since it’s a forced sequence.

Example 2

The correct answer is A. It looks like C is sente to kill the above group, but after white responds at A in the original diagram, this is sente to save the group above.

Confusing text about life status

Katago will think some groups are alive or dead in unexpected ways. For example, some stones that could easily live it might say are dead because it knows the optimal strategy is to tenuki, or perhaps sacrifice them. This can be fixed in the text box typically.

Example:

This is after the correct move. The text says “Playing here is worth about 14 points over playing away in an empty corner. The stones marked with a triangle are dead. Black would have played at A if you played away. Black would likely play at B next.”

But of course, it’s not clear why the triangled stones are dead because black can obviously save them if black wants to. This could just be removed from the text to fix the problem potentially.

Katago continuations are too high level

Sometimes the refutation move for a wrong path can be really advanced and hard to understand for a human. This may be fixable in text or by changing the move.

A is correct. Hard to understand why. Complicated life and death as well as shape.

Just doesn’t feel good

The editor should use their gut. If a problem doesn’t feel interesting or convincing for whatever reason, skip it. For me this happens around ¼ problems.

Too many options

Somewhat of a judgment call but usually only 3-4 max.

Position is overall too complicated

Sometimes there can be a problem where there is way too much going on, almost like it’s a mini full board problem. That can force the user to look at a bunch of unnecessary parts instead of focusing on the one real problem. This isn’t to say problems should be easy, they just shouldn’t be messy if not needed.

Text missing good explanations

This is common, and can happen for a million reasons. Sometimes it’s possible to fix. Almost all problems can benefit from editing the text, but we have to balance how much time it can take.

Let’s do some examples.

Example 1

https://www.goproblems.com/problems/36096#mode=intro

C is the right answer. A and b are too slow.

After playing A and white responds:

The text for a says “This loses about 8 points. Black makes about 5 corner points.” We could add something like “The black move was too slow.” Or “The corner was the wrong direction, the black wall was already strong. Playing on the side is bigger.”

Correct move:

The correct move text says “Playing here is worth about 1 points over playing away in an empty corner. White would have played at the same place if you played away. White would likely play at B next.” It could have something added like “Pushing white down is sente and big.”

Example 2

Correct is B.

The correct text says: “Playing here is worth about 9 points over playing away in an empty corner. The stones marked with a triangle are stronger. White would have played at A if you played away. White would likely play at B next.” Should probably add something about how even though white gets to press down at B, it’s worth it to take the points in the corner.

Example 3

C is correct. “Playing here is worth about 3 points over playing away in an empty corner. Black would have played at A if you played away. Black would likely play at B next.” Should add to the text “Black can’t block without losing the stone at B11.”