Reinforcement Learning and the Creative, Automated Music Improviser

Abstract

Automated creativity, giving a machine the ability to originate meaningful new concepts and ideas, is a significant challenge. Machine learning models make advances in this direction but are typically limited to reproducing already known material. Self-motivated reinforcement learning models present new possibilities in computational creativity, conceptually mimicking human learning to enable automated discovery of interesting or surprising patterns. This work describes a musical intrinsically motivated reinforcement learning model, built on adaptive resonance theory algorithms, towards the goal of producing humanly valuable creative music. The capabilities of the prototype system are examined through a series of short, promising compositions, revealing an extreme sensitivity to feature selection and parameter settings, and the need for further development of hierarchical models.