Tsumego difficulty estimation and AI

Hey folks,

I wondered the last couple of days how ranks are attributed to life-and-death problems.

Obviously there is the empirical way - I’m not sure, but people have told me that is the way how it is done on 101weiqi.com for example.

Before Internet it were mostly educated guesses by strong players I suppose.

With todays advancement would it be possible to build an AI capable of reliably estimating life-and-death problem difficulties?

It would be a step towards the development of a problem generator, if I’m not mistaken.

Just throwin’ the idea out there, and curious to see what folks with some knowledge in AI think about it. :slight_smile:

2 Likes

some tiny tsumego are hard because counter-intuitive , so computer will not be good way to estimate difficulty to humans

OGS has users with rank - it could give rank to tsumego as well, then rank of tsumego would change same way as rank of opponent, but no one coded this

tsumego generators exist long ago, all problems on sahwal were generated

4 Likes

Yeah, it would have to be trained on the data of human solved problem sets. I’m convinced an AI could recognize these tricky shapes.

That sounds interesting, never heard of sahwal?! O.o

1 Like

You can use a rating algorithm to learn the difficulty for each tsumego. If the user solved it, you count it like a win for the user, if she failed to solve it, count it like a win for the tsumego.

4 Likes

I’m pretty sure this would work ok. Training data could be the initial board state plus variants and the difficulty for example as measured by goproblems.com. Great project for a machine learning student.

1 Like

A nice feature might be an AI assisted tsumego solver where users could enter their own tsumego problem or select a group to analyze at a specific move in one of their games. AI might go for a bigger move elsewhere, but you still might want to know if it was even possible to save that little invading group.

It looks like a program called Xuan Xuan Go has something like that, but you have to register for the full version (buy for 69 Euro) to get the tsumego solver; I have not tried it. I’m sure it can’t solve everything. DDK players would benefit from even a very limited tsumego solver. I’ve toyed with the idea of trying to write a simple brute-force tsumego solver, but as you get into larger problems, you would probably need a Neural Net to trim down the decision tree. I’m sure that would be quite a large undertaking (understatement). I’m no AI expert either; Just an electrical engineer who does a bit of coding on the side.

1 Like

I think katrain has features going in that direction.

1 Like