[yazr]

Many had the same thought after AlphaZero i.e. some sort of expert-iteration for any (formal/combinatorial) domain what can be tree-searched

But it turns out, that interesting problems have existing refined algorithms that are difficult to beat

I agree that (DL + something) is the next step

Do you lean more towards symbolic search + neural heuristic or neural representation of symbolic domain ?

[mark_l_watson]

I lean towards symbolic AI + neural models. There is an architecture diagram in the article showing one good example.