Research on deep reinforcement learning-driven music melody generation algorithms

Abstract

A novel hybrid melody-generation framework which integrates a Variational Autoencoder (VAE) with Reinforcement Learning (RL) through Proximal Policy Optimization (PPO) to enhance musical structure, coherence, and statistical expressiveness. For the purpose of providing a clean training corpus, the Lakh MIDI Dataset Clean was pre-processed by performing melody-line extraction, normalization and piano-roll encoding. The VAE learns latent musical representations of pitch, rhythm and harmonic relationships while the PPO-based RL module refines generated melodies based on reward signals rooting in tonal consonance, rhythmic stability, note diversity and melodic contour stability. The experimental evaluations show that there are significant improvements compared to the baselines that only used VAE. The VAE-PPO system produces structurally coherent melodies with quantifiable improvements in tonal consonance (14.6%), rhythmic stability (12.3%), and note diversity (18.9%). The pitch distribution corroborates this stable melodic behaviour by continuously being concentrated on musically acceptable ranges. Moreover, the PCA and t-SNE based latent-space analyses reveal well-formed clusters, showing low reconstruction error and thus successful learning of both local and global musical patterns. The proposed VAE-PPO framework performs data-driven learning and optimization of symbolic music representations, demonstrating the application of machine learning and analytics to complex, high-dimensional music data. The PCA and t-SNE based latent-space analyses reveal well-formed clusters, showing low reconstruction error and thus successful learning of both local and global musical patterns. The reward progression curves confirm robustness in the PPO optimization process, showing consistent increases during episodes. Overall, the findings show that melodies with considerably improved structural consistency, harmonic quality, and melodic variation are produced when latent representation learning and fine-grained RL optimization are combined. This makes the suggested architecture a viable option for next-generation AI-assisted music composing tools.