Applies the paper's Rebound Winner-Take-All (RWTA) neuron motif to
cart-pole balancing as a minimal proof-of-concept. A 2-neuron
Half-Centre Oscillator (HCO) controller is evolved with CMA-ES —
no RL, no backprop, no hand-tuning.
- cartpole_rwta/physics.py: cart-pole ODE in SI units
- cartpole_rwta/controller.py: HCO controller (rebound slow-spiking
neurons + mutual WTA inhibition), 15 evolved parameters
- cartpole_rwta/evolve.py: CMA-ES loop with CSV logging and
checkpointing; falls back to a built-in (mu,lambda)-ES if cma
is not installed
- cartpole_rwta/analyze.py: fitness curve and episode trace plots
- cartpole_rwta/NOTES.md: detailed architecture notes including
plain-language explanation of rebound excitability, WTA, and
how the HCO maps onto the cart-pole action space
Setup: uv pip install numpy numba matplotlib cma
Run: python -m cartpole_rwta.evolve
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
oscillatory.net