Tunes using Deep RL
This project is was created to explore the role of AI in the creative arts (music). Using Deep RL (DQN) and the rules of music theory, I tried to emulate what might be "creative" or "original" music.
Music Generation specifically is a task that machine learning should be good at as it contains sequences of notes and fairly structured overall. Models should and do learn it fairly well, but a lot of RNN-based models attempt to extract this structure and use it to learn what sequences should be played. It does not actually choose any new notes on anything other than what has been previously found. Here's where RL comes in. Reinforcement Learning by definition is choosing actions to maximize rewards and learning which notes best follow others. But the most important aspects, in my opinion, is how RL can be used to add exploration to the mix. RL models/agents can choose to ignore the best possible option and instead choose a random action (or note). This feature is amazing because it intuitively allows the agent to keep trying new notes to see which is the most rewarding. Regardless, these notes still exist within the composition and forever change it. This is the creative aspect (which is almost human) that tries new things to find a better way.
If you liked this repo, please give it a ⭐!
Here are a few samples of what this can generate:
Use the colab notebook to play around with the settings and generate your own tunes! To learn more about the libraries used and how to change the settings, please visit the Magenta Project's Github Page.
If you find any problems within the project, please create an issue.