Has anyone ever considered using ai to create tsumego or tesuji problems. It seems like that could be a great thing to do. You might be able to get them from the games they play, or specifically program the ai to create the problems.

What do you all think?

Nice idea!

- For this you would need to define what tsumego and tesuji problems are. Perhaps both share the property that they involve a local move with no alternatives?
- Note that it might be hard to define what 'local' means. You could put a bounding box around each group and check if there's a single best move in it.
- One approach that will definitely come up with some examples is take it one step further and check if a board position has only one move for the current player. You could then construct a problem set from those moves by checking them manually.

Using AI to extract problems from human games (by this forum's lightvector):

https://neuralnetgoproblems.com

Building that site was a fun experiment.

I hope eventually someone has the time and inclination to do it far better than I did. That site was built back when all I had was pure policy nets, with the one key detail being to give the player rank as an input to the net to try to teach the net about how players of different ranks play and how their moves might differ on average, to try to determine some idea of the "rank" of a problem. I haven't revisited this idea since.

In addition to improving the "rank" heuristics, maybe bigger nets and more data could be much better. The nets weren't that large by today's standards. And it was before I started seriously experimenting with how to integrate a net with search to actually produce high-quality evaluations. So the site is relying heavily on just "this was the move the pro or high-dan amateur played" without the ability to back it up with independent AI evaluation - whether to verify the goodness of the move, or even just to show more context to the user and/or give feedback about answers besides the pro move.

I do think it still came out well given the amount of effort it involved and the limitations I had at the time.

And I hope at least a few people have found it useful. A lot of the problems turned out to be good "follow basic instinct" practice (https://senseis.xmp.net/?BasicInstinct) which hopefully can be educational if you're a beginner unfamiliar some of the basic instinct moves in practical open-space situations.

AFAIK, human pros are at least as good as current top bots in tsumego. However, the bots are not trained to solve or create tsumego. If AI programs were trained to create tsumego, they might create some fantastic ones.

Tesuji are another matter. In reviews of pro games, top bots often find (probably) better local plays than the pros did. The positions where that happens could be good for tesuji problems.

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

Visualize whirled peas.

Everything with love. Stay safe.

_________________
http://tchan001.wordpress.com
A blog on Asian go books, go sightings, and interesting tidbits
Go is such a beautiful game.

Every position is a problem but we often want something more specific. The more specific the more difficult constructing problems becomes. It does not require non-AI or AI programs to create problems. If one uses programs for creation, we must evaluate their problem positions whether they meet our specifications.

Hashimoto Utaro was said to have invented a problem every day. That impressed me before I started to create problems.

Nowadays, I create more than one problem per day on average so that I also solve them and describe the solution. Creating a problem takes between seconds and 4 days. The more specific the more time I need. 4 days rarely occur when I need a very particular (counter-)example for given presuppositions, e.g. when I have proved a theorem of existence or need to disprove a conjecture of non-existence.

Creating problems is the easy part. The difficult part is to solve them and verify the solutions by tactical reading (incl. verification of its decision-making) or equivalent means (such as assessing numbers of liberties in a capturing race). Computer-sampling of variations is NOT a verification of a solution! E.g., if a computer tests 99 of 99 failure variations for one particular next move, the program might be wrong if the 100th variation would be successful.

Amaaaaazing I really like the problems. Thanks, lightvector!