There was a piece here on HN about a week ago on Kolmagorov Arnold network (KANs) as an alternative to NNs.
<a href="https://news.ycombinator.com/item?id=40219205">https://news.ycombinator.com/item?id=40219205</a>
If I understood it correctly, a KAN would be a suitable network to work on this problem, as it's activation functions are in the links (and can be any function).
That's a really hard problem, I wonder how stuff like topology and combinatorial mathematics are related to this, do they apply to this kind of problems?