Automatic generation of studies

[fkarger]

In the age of AI, it is perhaps an obvious question whether the automatic generation of studies is possible.
So far, however, I have not found anything on this.

This could be interesting for at least 2 reasons:
- bringing out previously unknown beauties of chess
- enriching test suites for engines

Does anyone know of developments that go in this direction?

Chesstempo and Co have developed methods to automatically extract tactics,
but the aesthetic requirements for studies are probably more demanding.

[fkarger]

Just for fun, I asked Grok if he could compose a study for me that would be a challenge for engines.

Grok now explains in detail and confidently why the following position fulfills the requirement.

[d]r4rk1/pp1qbppp/2p1pn2/3pP3/3P4/2P2NP1/P3QPBP/1R1R2K1 w - - 0 1

My opinion: There is still some room for improvement

[jefk]

there exists a program to (creatively) generate ('beautiful'' ?) studies/problems;
mostly endgame studies, i thought. But it's exposure (download) was withdrawn
and it's now a private program; also not for sale, apparently
https://en.chessbase.com/post/chestheti ... stom-mates

Some of the studies it generated are still online; at various places. even Amazon
(5 books with Esthetica's studies)

[fkarger]

Thank you for the hint!

I think I once read that article but could not recall it.

[MyChessPosters]

I really like your mindset here. Taking time to understand the “why” behind each move, even without memorizing full lines, often leads to deeper and more practical understanding. Especially at the club level, it's more important to know plans and ideas than to recall long sequences. Your approach feels sustainable and smart. Curious—do you also revisit the same openings regularly or rotate through different ones to keep things fresh?

[towforce]

Just for fun, I asked Grok if he could compose a study for me that would be a challenge for engines...

If you look at the AI forum, you'll see that I am a chatbot enthusiast - but it's the wrong tool for this job.

What makes a good puzzle:

1. One move wins or draws

2. All the other moves draw or lose

3. Until a few correct moves have been made, the plan to save the situation is not obvious

This tells me that the tool for the job is a modified engine. When it has found a candidate puzzle, it would switch its own search depth from 20 to 3, and it should now fail to find the solution.