On the impact of domain-specific knowledge in evolutionary music composition

In this paper we investigate the effect of embedding different levels of musical knowledge in the virtual machine (VM) architectures and phenotype representations of an algorithmic music composition system. We examine two separate instruction sets for a linear genetic programming framework that differ in their knowledge of musical structure: one a Turing-complete register machine, unaware of the nature of its output; the other a domain-specific language tailored to operations typically employed in the composition process. Our phenotype, the output of the VM, is rendered as a musical model comprising a sequence of notes represented by duration and pitch. We compare three different pitch schemes with differing embedded knowledge of tonal concepts, such as key and mode. To derive a fitness metric, we extract musical features from a corpus of Hungarian folk songs in the form of n-grams and entropy values. Fitness is assessed by extracting those same attributes from the phenotype and finding the maximal similarity with representative corpus features. With two different VM architectures and three pitch schemes, we present and compare results from a total of six configurations, and analyze whether the domain-specific knowledge impacts the results and the rate of convergence in a beneficial manner.