Interactive Exploration-Exploitation Balancing for Generative Melody Composition
summary
A web app (React + Python, MIDI pipeline) couples a pre-trained MusicVAE (512‑dim, reduced) latent space to human-in-the-loop Bayesian optimization using GP‑UCB beta slider that explicitly controls exploration versus exploitation. Candidates are sampled along a 1D subspace and the chosen melody populates the editor; sliding beta makes next candidates more similar or more diverse. A 12‑participant study shows 11/12 prefer this method, with higher enjoyment, exploration and expressiveness and improved search performance.
AI-written summary · about
chapters (8)
- 0:00Introduction
- 0:18Overview
- 0:48Previous Work
- 1:28System Overview
- 2:31Algorithm Implementation
- 3:04Problem formulation
- 3:30User study
- 4:28Conclusions
related talks
01h 04mKeynote 3. Interactive generative models for music
19mGenerative music, playful visualizations and where to find them
41mMusic and Machine Learning
43mA Progressive-Adaptive Music Generator (PAMG): An Approach to Interactive Procedural(…)
36mMusic and Art with ML
21mMusical Deep Neural Networks in the Browser
33mContext-Aware Masking for Controlled Music Generation in MIDI
34mModelling emotional nuance of musical performance