Reading coach software?

[Bill Spight]

This topic is inspired by my recent reflections on the fact that players use tsumego problems to improve their reading skills. As I did in part, once upon a time. I used to think that problem software would play against the solver, offering stubborn resistance, backtracking to play other options that the solver should be able to refute. But that is not so good if the purpose is to improve reading skills, because a major blind spot in reading is missing the opponent's toughest replies. The reader should find those moves, too. So for training reading skills, you don't want an opponent, you want a coach.

Now, developing reading coach software is a project that I would find interesting, but one that would lie years in the future. I. e, maybe never. But perhaps others would be interested in doing so. It might even turn into an open source project. We certainly have people with the requisite skills in our community.

Supposing that such software were to be developed. What features would you like it to have?

[Bill Spight]

One question that arises is what is reading? I don't think that there is a single answer. Sakata said that a one lane road does not involve reading. OTOH, Segoe wrote a book of problems, most of which involved fairly long one lane roads, with the purpose of improving reading.

At the least, reading involves the construction of a search (or solution) tree (or graph), the selection of candidate moves, pruning the tree, deciding when to stop, and evaluating the leaves of the tree.

A while back I mused about go pedagogy in terms of the trivium: grammar, logic, and rhetoric. In those terms I think that reading falls under logic. It constructs an argument. This play is correct (or best, or good) because, even if the opponent makes one of these replies, I can refute them to achieve these results. (OC, reading may not produce such an argument. Too bad. ;( ) Reading may produce a proof, but that is neither necessary nor even desirable, for reasons of efficiency.

[Bill Spight]

Many who use tsumego to train reading dispense with providing the answers. What is the problem with providing answers?

One problem, the main one, I think, with providing answers is memory. If I remember the answer to a problem, it is useless to me for reading practice. (It might be nice to know the answer, but that is another question. )

But there is also a problem with not providing answers. Suppose that I think that I have solved a problem and review it later. Once more I do the same reading that I did before, only more quickly and easily than the first time. I believe that this is a good thing, because I need to read quickly in real games. However, I actually got the problem wrong. What I have done is reinforce my bad reading, making it more likely that I will make the same kind of errors in a real game, and be none the wiser.

A good coach can tell me when I have erred, without providing the answer.

A good coach can also tell me when I have succeeded. It is very important for learning to get feedback as soon as possible, especially positive feedback.

[oren]

One question that arises is what is reading? I don't think that there is a single answer. Sakata said that a one lane road does not involve reading. OTOH, Segoe wrote a book of problems, most of which involved fairly long one lane roads, with the purpose of improving reading.

Yamada Kimio also published one recently that is a bunch of one lane road problems that I find very useful. It seems to be 'reading' to me if only the difficulty of keeping all the stones in the right spots through a very long sequence.

[RobertJasiek]

Such a software must also have a provision for cases of too great complexity beyond reading logic.

[RBerenguel]

I've been thinking a lot lately about "what is reading" and what it involves, I'm still trying to figure out the most important parts and how to train them separately. I.e. the main set of "qualities" I'm pretty sure about is depth/breadth, clarity, speed, but there are side skills like focus, so I'm still thinking in circles around this.

I think to train reading you need to improve all these areas, and I feel there are ways to isolate each skill and concentrate on it, to try to even out the field. But... First, it is hard to figure out what the skills are, but also how to isolate and train them effectively.

[Polama]

I ran into a research paper that set me thinking along similar lines:
Automatic Discovery of Cognitive Skills to Improve the Prediction of Student Learning

It claims equivalent or often superior results to experts for dividing problems into skills, where how good you are at some problems in a skill predict how you'll perform on other problems based on that skill. It does seem like a very important role for a coach: providing appropriates problems that will strengthen your weaknesses without overwhelming you.

If someone really wanted to pursue this and gathered some problem & performance data I'm sure the authors would provide code. Or for that matter, the algorithm described doesn't look particularly difficult to implement.

[EdLee]

Sakata said that a one lane road...

Segoe wrote...involved fairly long one lane roads

Yamada Kimio also published one recently that is a bunch of one lane road problems

Another pro also expresses the same sentiment.
Seems many or most pros feel the same way about "forced sequences"...

[RBerenguel]

I ran into a research paper that set me thinking along similar lines:
Automatic Discovery of Cognitive Skills to Improve the Prediction of Student Learning

It claims equivalent or often superior results to experts for dividing problems into skills, where how good you are at some problems in a skill predict how you'll perform on other problems based on that skill. It does seem like a very important role for a coach: providing appropriates problems that will strengthen your weaknesses without overwhelming you.

If someone really wanted to pursue this and gathered some problem & performance data I'm sure the authors would provide code. Or for that matter, the algorithm described doesn't look particularly difficult to implement.

Crap, another paper to my to-read

[Bill Spight]

Tsumego problems typically tell who is to play and what their goal is. There are pluses and minuses to that. On the plus side, if a line does not reach the goal, you can tell that it is incorrect. OTOH, reaching the goal does not mean that it is correct, or complete. You may have missed the best or most stubborn replies of the opponent.

Also, part of reading is evaluating the leaves of the tree, and telling what can be accomplished does not train that skill. IMO, status problems are better for training reading than tsumego per se.

[xed_over]

I used to think that problem software would play against the solver, offering stubborn resistance, backtracking to play other options that the solver should be able to refute.
...
a major blind spot in reading is missing the opponent's toughest replies. The reader should find those moves, too.

When I first started looking at tsumego, I couldn't figure out how to solve them on my own, because I couldn't figure out how white should reply to black's attempts. I kept making moves as white that were still in black's favor. I could only solve them when using a program that auto-replied to my every attempt.

It wasn't until I read Davies' book Tesuji, where he explained how with each move, you have to mentally switch sides and find their best move. That completely changed the game for me (pardon the pun). Now it wasn't just one problem, but many, many problems in one. (Its funny that I don't recall seeing the same explanation in he's other book, Life and Death, where it seems it would be more appropriate)

Even now, I may want to show off some novel or interesting problem to someone, but if they don't see the answer right away, I can't fully refute their incorrect attempts, because I only know the "correct" solution, and not the full problem. So they may think they've solved it differently, when they haven't yet.

[RBerenguel]

Many who use tsumego to train reading dispense with providing the answers. What is the problem with providing answers?

One problem, the main one, I think, with providing answers is memory. If I remember the answer to a problem, it is useless to me for reading practice. (It might be nice to know the answer, but that is another question. )

But there is also a problem with not providing answers. Suppose that I think that I have solved a problem and review it later. Once more I do the same reading that I did before, only more quickly and easily than the first time. I believe that this is a good thing, because I need to read quickly in real games. However, I actually got the problem wrong. What I have done is reinforce my bad reading, making it more likely that I will make the same kind of errors in a real game, and be none the wiser.

A good coach can tell me when I have erred, without providing the answer.

A good coach can also tell me when I have succeeded. It is very important for learning to get feedback as soon as possible, especially positive feedback.

But... there are several sides to this Bill:

As you get better, you read quicker. Either:

• You read exactly the same, faster (relying mostly on memory and move on,
• You got actually better at reading speed, so you can visualise the same patterns faster. Given the same time as in first try, you can explore more variations (breadth) or get deeper (depth)

What I've done in problems without answers (the sets I keep doing and doing of Cho tsumego) is reading more variations as I get faster. I.e. on pass 1 I may find a solution I'm happy about. When I see it again, I can quickly visualise this in around 1/3 and 2/3 of the original reading time (give or take, just a ballpark estimate.) Then I explore some side variations. If I don't find any problem I mark it as "ok," if I see something fishy or find a mistake I mark it as wrong so I can read it again in full.

[Boidhre]

GoChild: https://gochild2009.appspot.com/

Does this to some extent. You solve tsumego by playing them out against the AI and there are usually multiple "wrong paths" where the AI plays out a refutation. It's particularly interesting in something like endgame problems where it's not necessarily obvious why a particular solution is inferior. It doesn't show you the answers but will indicate when a line is incorrect or when you've found the solution (which means there's a definite ceiling above while problems will be unsolvable for you until you try and brute force them).

[Bill Spight]

What I've done in problems without answers (the sets I keep doing and doing of Cho tsumego) is reading more variations as I get faster. I.e. on pass 1 I may find a solution I'm happy about. When I see it again, I can quickly visualise this in around 1/3 and 2/3 of the original reading time (give or take, just a ballpark estimate.) Then I explore some side variations. If I don't find any problem I mark it as "ok," if I see something fishy or find a mistake I mark it as wrong so I can read it again in full.

That's great!

But you still miss the immediate feedback and reinforcement. And it is still possible to be wrong.

Also you are missing information that may be of value for gauging the difficulty of problems and for overlearning.

[RBerenguel]

What I've done in problems without answers (the sets I keep doing and doing of Cho tsumego) is reading more variations as I get faster. I.e. on pass 1 I may find a solution I'm happy about. When I see it again, I can quickly visualise this in around 1/3 and 2/3 of the original reading time (give or take, just a ballpark estimate.) Then I explore some side variations. If I don't find any problem I mark it as "ok," if I see something fishy or find a mistake I mark it as wrong so I can read it again in full.

That's great!

But you still miss the immediate feedback and reinforcement. And it is still possible to be wrong.

Also you are missing information that may be of value for gauging the difficulty of problems and for overlearning.

Totally agree... I just had to pick the ebst I could do easily