I've been wondering if machine learning or metaheuristics could be used to get close to an input sound. I've been toying with this idea using MIDI CC with a Rytm.