It has been suggested that part of our appreciation for music comes from dopamine responses triggered by building-up then breaking sensory expectations, i.e. that music manufactures continuous surprise during listening, an effect that appears sustainable even after repeated exposure to a given musical pattern.
At the same time, techniques for music generation have recently made great progress, with WaveNet  in particular appearing to be a promising avenue.
We propose to combine such music generation techniques with a “critic” component that would analyze the degree of sustained surprise (i.e. enjoyment) that a musical track generates during listening. This critic may be a model attempting to predict the short-term evolution of the musical track; the error signal of such a model would provide visibility into listening surprise. This component would guide the music generation process so as to produce tracks that not only sound “like music” but also sound enjoyable.