DiffusionControl / results /baselines /fixed_controller_pendulum.txt
matCercola18's picture
Add files using upload-large-folder tool
c67b25a verified
Raw History Blame Contribute Delete
1.87 kB
=== PENDULUM fixed dimensionless controller (N=40, seed=123 held-out) ===
Justification: theta'' = (g/l) sin(theta) - (b/(m l^2)) omega + u/(m l^2), controller u = -phi(y), y = c1 theta + c2 omega (c FIXED at (1.0, 0.35) for the whole family, as in the amortized model). The natural dimensionless rescaling of phi is by the holding torque m*g*l (the torque needed to hold the pendulum horizontal); a natural TIME scale is sqrt(l/g) (the small-oscillation period / sqrt(g/l) natural frequency), though this codebase does not rescale omega's contribution to y by it (c2 is fixed, not per-plant) -- so the fixed controller below rescales phi by m*g*l only, exactly mirroring how eval_pendulum.strong_oracle's own smart init (2.5*m*l*g*xp) and random-restart shapes ((U[0.5,2.5]*m*l*g)*xp*(|xp|/5+0.3)) are built.
Config space = SAME as the oracle's restarts: shape in {linear xp, nonlinear xp*(|xp|/5+0.3)}, K_hat in linspace(0.5,2.5,9). Tuned on a FRESH N=200 set (seed=555, disjoint from the held-out test set) by (certified_rate, success_rate); NOT re-tuned per plant (one fixed controller for the whole family). Cost = the paper's ABSOLUTE-weight J = integral (theta^2 + 0.1 omega^2) + u^2 dt (simulate_pendulum_batch default weights) -- not a scale-invariant metric.
Chosen fixed controller: shape=nonlinear K_hat=2.500 (phi(y) = 2.500 * m*g*l * shape(y))
success (empirical stability, N=40): 1.000
certified (regional LMI, N=40): 1.000
cost regret vs cached per-plant oracle: median +56.69% p90 +88.86% max +124.15% (n=40)
synthesis time: 81.2865 ms/plant (no per-plant solve -- table lookup + rescale)
top 5 tuning configs (shape, K_hat, certified_rate, success_rate) on the tuning set:
('nonlinear', 2.5, 1.0, 1.0)
('nonlinear', 2.25, 1.0, 1.0)
('linear', 2.5, 1.0, 1.0)
('linear', 2.25, 1.0, 1.0)
('linear', 2.0, 1.0, 1.0)