Artificial intelligence methods for music generation: a review and future perspectives

Applying artificial intelligence (AI) methods for music generation has been attracting the interest of researchers since the first days of computing. Music has self-referentiality, i.e., does not necessarily relate to extramusical concepts, and is organized on multiple interrelated levels of abstraction, a fact that makes it a peculiar test-bed for AI methods. This chapter focuses on how different AI methods model abstractions from musical surfaces, how they model knowledge, and knowledge acquisition mechanisms. The main categories presented herein are (i) nonadaptive methods that rely on user modeling for generating interesting and meaningful music, (ii) probabilistic approaches that learn, either implicitly or explicitly, from data, and (iii) evolutionary methods that allow intuitive user interaction. Based on the review of such methods, this chapter attempts to highlight some recent algorithmic developments that might lead future algorithmic approaches to more intuitive AI methods that incorporate deep abstract representations of music and are able to generate music that interpolates between or even extrapolates from learned styles.