DiffusionControl / results /pendulum /baselines_stats.txt
matCercola18's picture
Add files using upload-large-folder tool
929e74d verified
Raw History Blame Contribute Delete
942 Bytes
Pendulum baselines + statistics (held-out N=40, seed=123)
DIFFUSION (amortized, certified) — bootstrap 95% CIs:
regret vs per-plant oracle: median +1.0% CI [+0.9, +1.2]
≥1 certified: 1.00 CI [1.00, 1.00]
≥1 stabilising: 1.00 CI [1.00, 1.00]
median cost: 1.178 CI [0.802, 1.859]
CLASSICAL per-plant LQR (on the linearisation):
stabilises the NON-LINEAR pendulum (in-region ICs): 0.95
median cost (stable): 1.082 CI [0.717, 1.779]
per-plant design time (Riccati): 0.79 ms (NOT amortised: one solve/plant)
CONTRAST:
- LQR is solved PER PLANT; the diffusion model amortises the whole family (2 ms/sample,
one forward pass, no per-plant solve).
- LQR uses FULL STATE u=-Kx (more info than our output sector φ) and is tuned to the
SAME quadratic cost — yet the amortised model matches per-plant-optimal to +1% and is
ABSOLUTE-STABILITY CERTIFIED in-region, which the linearisation-based LQR is not.