A policy learns by climbing its reward — nudging its parameters a little further up the hill each episode. How far it steps is the learning rate. Too small and it crawls; too big and it overshoots the peak, oscillates, and eventually blows up. Scrub it and watch the learning curve go from a smooth rise to a screaming divergence.
Gradient ascent on a reward R(θ)=−(θ−θ*)²; θ ← θ + α·∇R. Node-verified: small α converges slowly, α≈0.5 almost instantly, α>1 diverges. Built on the reusable probe.js primitive. Runs on your device. · Institute for Physical AI @ BMI