In pursuing some machine learning studies for harmony, it strikes me that the input/output relationships are often too far from one another.
Especially: it may not always be clear what the input/output relationships are. But, perhaps more importantly, it seems excessively difficult to make the inputs change their relationships to one another.
Parameters which seem to be working independently should be able to suddenly seem related: that an increase in harmonic tension might suddenly and palpably correspond to an increase in dynamic or tempo or some other identifiable parameter. The parameters might not have so much meaning on their own (like a single note – fairly arbitrary). But the relatedness of the parameters to each other is something which can bring a strong musical impression (like a sempre PP in Beethoven, which seems to suppress another compositional development, until it can be suppressed no longer).