Integrating Machine Learning and Rule-Based Approaches in Symbolic Music Generation

When composing a piece of artistic or experimental music, a composer often tries to create a new style of music that differs from existing styles. This contradicts the idea of conventional machine-learning-based generation systems because they learn from existing music pieces and imitate their styles. Therefore, generating music beyond the assumed range of styles would be difficult for these systems. For example, to generate atonal music by learning from tonal music would be difficult. On the other hand, in the context of computer-assisted composition, especially in classical contemporary music, rule-based approaches based on constraint programming or genetic programming have been relatively successful. With these approaches, composers can define original musical styles by designing rules in the form of constraints or fitness functions. However, the difficulty of these approaches is that composers must program every detail of the music, which is extremely burdensome. As an early step to creating a computer-assisted composition system that overcomes both difficulties, this paper proposes a new neural network model that integrates the style definition capability of rule-based approaches and the automation power of machine learning. The key idea is to generalize the GAN discriminator to repurpose it as a constraint solver beyond the GAN’s usual function of imitating the original. In addition, as a proof of concept, this paper shows that the proposed model can actually output solutions satisfying some musical rules that are not necessarily relevant to the dataset.