×

Exploring the musicVAE space with an Interactive Genetic Algorithm

This project uses the musicVAE neural network provided by Google through magenta.js to generate random sounds. An interactive genetic alogrithm is used to explore the latent sound space.

To do so, 12 random vectors are sampled in the 256-dimensional latent space and we can only move along these vectors. This dramatically reduces the search space but also limits the possible outcomes, but makes the use of an IGA much more feasible. The IGA is driven by manual user feedback.

The tool samples six random points in this 12-dimensional space and decodes their latent represenation into a 4-bar, chord-conditioned, midi sequence. You can then provide feedback on the sound quality by moving your cursor left/right on the screen. A click sends the "fitness" to the algorithm and triggers the next sound. After all six sounds have played, new sounds are prepared based on the fitness of observed sounds through crossover and mutation.

Developed by Adrian Goedeckemeyer in June 2019.

Thanks to the Magenta team and Dr. Eric Bonabeau for inspiration.