Fine-Tuning Music Generation with Reinforcement Learning Based on Transformer

Deep supervised learning is the most common way of automatically music generation. However, this sort of model only learns probabilities from dataset, and such pattern does not leave much room for manully control, which could result in an out of expectation result. In this paper, we have proposed a novel approach of polyphonic music generation using Deep Reinforcement Learning base on Transformer skeleton. The principal novelty of our approach centres on having a well trained music Transformer network as basement, then using Reinforcement Learning to fine tune it to impose music theory into the model. The reward of RL consists of probability learned from training data and music theory proposed to follow. We analyzed quantitatively and qualitatively, and results show that the proposed model enhances the performance of deep supervised learning who only learns from data and the music generated comes out to be more creative. Eventually, we discussed the usefulness of our model.