Composição musical algorítmica utilizando redes geradoras adversárias e algoritmos genéticos

• Ano de defesa: 2023
• Autor(a) principal: Matheus Bitarães de Novaes
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Minas Gerais
• Palavras-chave em Português: Engenharia elétrica; Inteligência computacional; Redes neurais (Computação); Algoritmos genéticos; Música; Redes geradoras adversárias; GANs; Redes neurais; Música algorítmica
• Link de acesso: https://hdl.handle.net/1843/53936

The application of Computation Intelligence for musical pieces generation is present in literature since the early moments of this research field. Since then, algorithmic art has been following the technological advances in the field and, since it is a subject that can be approached by many sides, there are a diverse set of approaches in literature to emulation of the artistic process by computers. Among the research field explored for computational musical pieces generation, Genetic Algorithms and Neural Networks have significant presence and, as GANs have become more widely used, there has been an increase in the use of them for creating art. This work proposes an architecture composed of a genetic algorithm whose initial population is fed by generative adversarial networks (GANs) specialized in generating melodies for certain harmonic functions. The fitness function of the genetic algorithm is a weighted sum of heuristic methods for evaluating quality, where the weights of each function are assigned by the user, before requesting the melody. A data augmentation statregy for the GAN training data was proposed and experimentally validated. Another experiment performed was a comparison between the quality of the melodies generated by the proposed architecture, a GAN and an LSTM network. The effects of utilizing the Discriminator’s evaluation on the fitness function of the generic algorithm were also experimented in a third experiment. The statistical comparison give evidences that this approach enhances melody quality in comparisson with using the fitness function without Discriminator’s evaluation.