Neural Boids

2 days ago

Noids are neural boids that use a small neural network to output steering forces based on local perception, with 1,922 learned parameters.
Real starlings flock by tracking 6-7 topological neighbors, adjusting velocity to stay close, match heading, and avoid collisions, creating murmurations without global signals.
Craig Reynolds' boids (1986) use three hand-tuned rules: separation, alignment, and cohesion, which have been the basis for flocking simulations since.
Noids eliminate hand-written rules, instead learning steering functions from data, using topological neighbors like real birds.
The neural network in a noid takes 24 inputs (own velocity, heading, and 5 nearest neighbors' relative positions and velocities) and outputs a 2D acceleration.
Training involves imitation learning, where noids mimic Reynolds' rules through gradient descent, optimizing 1,922 weights.
The transition from random weights to flocking behavior is a phase transition in weight space, emerging from interactions of all parameters.
Noids perform well on GPUs due to matrix multiplications aligning with GPU operations, making them efficient for large flocks.
The noid library is open-source, written in Rust, and supports training and weight interpolation, enabling flexible flocking simulations.

Hasty Briefsbeta