File size: 2,029 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
/home/cercola/matte_server/DiffusionControl/nonlinear/experiments/timing_vs_classical_stages.py:594: RuntimeWarning: invalid value encountered in subtract
  reg = (J[:, 0] - oracle) / oracle * 100
/home/cercola/matte_server/DiffusionControl/nonlinear/experiments/timing_vs_classical_stages.py:594: RuntimeWarning: invalid value encountered in divide
  reg = (J[:, 0] - oracle) / oracle * 100
{
 "hurwitz_first_gain/b1": {
  "n": 240,
  "median_ms": 0.32864371314644814,
  "p99_ms": 1.3357699383050188,
  "mean_ms": 0.3679626078034441,
  "min_ms": 0.2974970266222954,
  "max_ms": 3.758460283279419,
  "note": "stabilize_repair_fast.hurwitz_gain (first Hurwitz gain on 41-gain grid, F=1) incl. realize_loop, per plant; open-loop-stable plants return g=0 (trivial); the 0.4 ms eigenvalue repair"
 },
 "gain_margin_design/b1": {
  "n": 240,
  "median_ms": 2.6366771198809147,
  "p99_ms": 3.9378471672534943,
  "mean_ms": 2.694359033678969,
  "min_ms": 2.495504915714264,
  "max_ms": 4.250047728419304,
  "note": "realize_loop + 41-gain grid: Hurwitz check + Lyapunov-solve LQ cost, argmin (F=1, phi(y)=g*y); certificate = eigenvalues; per plant"
 },
 "gain_margin_design/batch60": {
  "n": 40,
  "median_ms": 2.406177382605771,
  "p99_ms": 4.753681387131411,
  "mean_ms": 2.5085222022607923,
  "min_ms": 2.3125936121990285,
  "max_ms": 6.146816428129871,
  "note": "looped over batch=60; ms per plant"
 },
 "gain_margin_design/quality_heldout60": {
  "coverage": 0.8833333333333333,
  "n_with_oracle": 46,
  "regret_median_pct": 1274.6151773239899,
  "regret_p90_pct": 6298.871143265807,
  "n_designed": 60,
  "note": "regret vs bank_density_siso.npz draw0_oracle (60-step refinement of the model's winner), rollout verifier of retrieval_baselines; F=1 fixed, phi(y)=g*y. NOTE the SISO cost is charged on the LOOP output y of L=G*F, so it is not invariant to the filter scale: a design that shrinks F reaches J~0 (checked: F gain 1e-2 gives J=0.009), so classical regret against this oracle is only meaningful with F held at 1"
 }
}