File size: 4,671 Bytes
68a167f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63

==============================================================================
PENDULUM retrieval/interpolation baselines
==============================================================================
[pendulum plants] seeds: 123
[pendulum diffusion] seeds: (0, 1, 2)
oracle_J backed out from cache: finite for 40/40 plants
/home/cercola/matte_server/DiffusionControl/nonlinear/data_generation/pendulum_lmi.py:149: UserWarning: Solution may be inaccurate. Try another solver, adjusting the solver settings, or solve with verbose=True for more information.
  prob.solve(solver=solver, verbose=False)
diffusion generation: 21.39 ms/controller (mean over 3 seeds)
=== PENDULUM retrieval/interpolation baselines (N=40, seed=123) ===
verifier: certify_controller (regional LMI) + simulate_pendulum_batch rollout (n_ic=3, dt=0.02, duration=6.0, energy_thresh=0.02, theta0=0.35)
oracle: CACHED results/pendulum/eval.npz (eval_pendulum.strong_oracle, 24 restarts x 100 steps), oracle_J backed out from best_costs/regrets
K_std=8, diffusion seeds=(0, 1, 2)

--- K=8 ---
arm             cert_any  stable_any   persamp  reg_med%  reg_p90%  reg_max%  ms/plant   McNemar(n10,n01,p)    Wilcoxon(n,p)
diffusion           1.00        1.00      1.00      1.09      2.46      3.36    171.13          (reference)      (reference)
  (sd across 3 seeds: cert_any 0.000, stable_any 0.000, per-samp 0.001)
retrieval           1.00        1.00      1.00      0.40      0.86      1.70     0.375              (0,0,1)     (40,2.8e-07)
interpolation       1.00        1.00      1.00      0.40      0.86      1.70     0.375              (0,0,1)     (40,2.8e-07)

--- K=1 ---
arm             cert_any  stable_any   persamp  reg_med%  reg_p90%  reg_max%  ms/plant   McNemar(n10,n01,p)    Wilcoxon(n,p)
diffusion           1.00        1.00      1.00      1.54      3.59      4.80    171.13          (reference)      (reference)
  (sd across 3 seeds: cert_any 0.000, stable_any 0.000, per-samp 0.000)
retrieval           1.00        1.00      1.00      1.84      3.70      5.85     0.127              (0,0,1)        (40,0.13)
interpolation       1.00        1.00      1.00      1.84      3.70      5.85     0.127              (0,0,1)        (40,0.13)



==============================================================================
PENDULUM fixed dimensionless controller (arm e)
==============================================================================
[fixed-controller tuning] seeds: 555
[fixed-controller tuning] N_tune=200 seed=555: best config shape=nonlinear K_hat=2.50 certified=1.00 success=1.00
  top 5 configs (shape, K_hat, certified_rate, success_rate):
    ('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)
=== 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)