Comparing a Q-Learning agent’s and human-generated piano fingerings

Abstract

This paper reports the findings of a Q-Learning reinforcement learning agent when given the task to generate the fingering for piano sheet music. Although some studies on automatically generating piano fingering have been done, these researches had not focused on reinforcement learning and Q-Learning, specifically. An environment, a set of states and actions, and a reward scheme had been created for the Q-Learning agent, and it had been given four piano sheets, which were used to generate fingerings. The input for the algorithm contained only right-handed, single-note melodies. The results showed that the algorithm had an overall better performance than previous research done on this topic, with some limitations. These results lend support to the idea that the agent learns not only to optimally place its fingers but also some hand-ergonomic rules, which it was not taught. The research demonstrates that reinforcement learning is a tool that can be used for newly-beginning pianists to help them with an understanding of piano fingerings.