Melody generation and optimization with genetic algorithm

Music composition has been a significant area of human creativity for centuries, and recent advancements in computer technology have led to the widespread adoption of Automatic Music Composition. Within this field, Genetic Algorithms (GAs), inspired by natural selection, stand out as a powerful technique for finding optimal or near-optimal solutions in vast search spaces. However, the subjective nature of music and the difficulty of objectively evaluating best music complicate the design of the GA's fitness function. This thesis aimed to empirically investigate the potential of using Genetic Algorithms to generate and optimize musically meaningful and aesthetically pleasing melodies by addressing these challenges. This thesis designed and implemented a GA-based system including an appropriate chromosome structure to represent melodies, genetic operators (selection, crossover, mutation) to manipulate musical structures, and a rule-based fitness function based on music theory principles. The effects of different population sizes (100, 200, 300) and tournament selection sizes (3, 5) on algorithm performance (convergence speed, fitness values) and the musical qualities of generated melodies were examined. The findings indicated that GAs can generate musically meaningful melodies with appropriate design and parameters. Larger populations demonstrated the potential to increase diversity and achieve higher final fitness values, while larger tournament sizes could lead to faster convergence but risked reducing diversity. The use of multiple musical criteria in the fitness function was found to improve output quality and help reduce repetition. In conclusion, GAs can be considered as interactive assistants supporting the creative process in music composition.