Machine Improvisation in Music: Information-Theoretical Approach

This chapter introduces the methods and techniques of machine improvisation based on information-theoretical modeling of music, starting from the first 1998 universal classification modeling of music as an information source, style mixing using joint information source, variable-length motif dictionary improvisation based on universal prediction, and use of information dynamics for symbolic approximation in the factor oracle machine improvisation algorithm. Later developments include query-guided machine improvisation, free-energy modeling of music cognition, and reformulating of variational generative neural music models in terms of rate-distortion theory. This information-theoretical framework offers a novel view of man–machine creative music interaction as a communication problem between an artificial agent and a musician, seeking optimal trade-off between novelty and stylistic imitation under scarcity constraints.