Multi-agent Reinforcement Learning for Structured Symbolic Music Generation

Abstract

Generating structured music using deep learning methods with symbolic representation is challenging due to the complex relationships between musical elements that define a musical composition. Symbolic representation of music, such as MIDI or sheet music, can help overcome some of these challenges by encoding the music in a format that allows manipulation and analysis. However, the symbolic representation of music still requires interpretation and understanding of musical concepts and theory. In this paper, we propose a method for symbolic music generation using a multi-agent structure built on top of growing hierarchical self-organizing maps and recurrent neural networks. Our model primarily focuses on music structure. It operates at a higher level of abstraction, enabling it to capture longer-term musical structure and dependency. Our approach involves using reinforcement learning as a self-learning method for agents and the human user as a musical expert to facilitate the agents’ learning of global dependency and musical characteristics. We show how agents can learn and adapt to the user’s preferences and musical style. Furthermore, we present and discuss the potential of our approach for agent communication, learning and adaptation, and distributed problem-solving in music generation.