Planting and Shaping the Neural Tree: Growing a Musical Collaborator Through Objective-Free Neuroevolution in Open-Ended Exploration

Abstract

We present a live performance for experimental electronic music in which a continuously evolving neural network joins two human performers to form an improvising trio. Rather than training a generative model on pre-existing works, we use neuroevolution strategies to grow a feed-forward neural graph in real time, with no formal optimization objective. Following a tree metaphor, the performers engage the network in two complementary roles: one "plants" its seed by controlling parameter initialization and mutation, while the other "shapes" its canopy by mapping MIDI note and control events to different synthesizers and their parameters. Using the Godot game engine, the evolving topology is rendered visually as a live, wind-blown tree, giving the audience a direct view of the organism’s growth. Built on the open-source BbMuse framework, our approach foregrounds lightweight, real-time AI that evolves alongside human musicians without the use of pre-curated data, treating the machine as a creative collaborator.