Automatic Jazz Melody Composition Through a Learning-Based Genetic Algorithm

In this study, we automate the production of good-quality jazz melodies through genetic algorithm and pattern learning by preserving the musically important properties. Unlike previous automatic composition studies that use fixed-length chromosomes to express a bar in a score, we use a variable-length chromosome and geometric crossover to accommodate the variable length. Pattern learning uses the musical instrument digital interface data containing the jazz melody; a user can additionally learn about the melody pattern by scoring the generated melody. The pattern of the music is stored in a chord table that contains the harmonic elements of the melody. In addition, a sequence table preserves the flow and rhythmic elements. In the evaluation function, the two tables are used to calculate the fitness of a given excerpt. We use this estimated fitness and geometric crossover to improve the music until users are satisfied. Through this, we successfully create a jazz melody as per user preference and training data.