Automatic music composition by genetic programming

Abstract

Automatic music composition is an area of research widely studied nowadays and many approaches have been proposed for this problem. This work is based on an existing project developed by the GRFIA which uses genetic programming for generating music melodies without human supervision. The project utilises a general-purpose library which is in charge of the genetic programming logic. The task of supervising the melodies is accomplished by a set of machine learning algorithms that are trained using a corpus of songs in order to select the best melodies generated. This final degree project develops a new library which replaces the one used by the original project. This new library implements some of the logic of genetic programming but the part in charge of selecting the best individuals has been developed using the multi-objective optimization algorithm NSGA-III. On the other hand, this project extends the binary tree structure used by the software. The current data model is able to store melodic and rhythm information and the proposed model is able to store harmonic information too. This change improves the way new melodies are generated. Finally, a comparative has been made using performance data and the overall score of the melodies generated. The result of the analysis is positive, but it has slightly improved in comparison to the original project. Even though, the two main goals, developing a new library and extending the model, have been successfully completed.