matCercola18 commited on
Commit
e664825
·
verified ·
1 Parent(s): ed12bd0

X19 SISO re-score results (npz, logs, reports)

Browse files
results/icml_r3/X19/FROZEN.json ADDED
@@ -0,0 +1,10 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "starts": 16,
3
+ "steps": 400,
4
+ "ckpt_every": 25,
5
+ "lr": 0.1,
6
+ "bcap": 10000.0,
7
+ "bisect_iters": 10,
8
+ "dt": 0.03,
9
+ "f_margin": 0.01
10
+ }
results/icml_r3/X19/PREREG.md ADDED
@@ -0,0 +1,129 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # X19 pre-registration: SISO re-scored with the corrected Popov certificate and an arm-independent cost reference
2
+
3
+ Written 2026-10-14, before any test plant was scored. Part A (design) was written before the validation runs;
4
+ Part B (frozen configuration) is appended after validation and before test scoring. Code:
5
+ `nonlinear/experiments/siso_x19.py` (+ `certify()` factored out of `nonlinear/experiments/siso_certify.py`).
6
+
7
+ ## Part A: design
8
+
9
+ ### Plants
10
+ - Validation: N = 40, seed 9191. Used to tune the reference's step budget and learning rate, and nothing else.
11
+ - Test: N = 150, seed 4343, drawn by `retrieval_baselines.gen_siso_plants` (= `stats_siso_regret.phase_gpu`,
12
+ `np.random.seed(seed)` then `testing.test.generate_random_sys(4, stable=True)`). Checked: identical to the E4
13
+ refresh draw (`results/icml_r2/E_stats/refresh/siso_phase1.pkl`). No test plant is in the training bank
14
+ (smallest max-abs raw coefficient difference to any of the 5,635 bank plants is 0.059).
15
+
16
+ ### Loop, realization, certificate
17
+ - Controller: order-4 filter F and a piecewise-linear spline phi (11 knots on [-5, 5]); u = -phi(clip(y_L, -5, 5)),
18
+ y_L = F y_G. Plant G strictly proper (relative degree 1), so L = G F is strictly proper for any proper F.
19
+ - Realization of every loop, every arm: `k_sweep.realize_pad` (python-control series of the scipy
20
+ controllable-canonical realizations of G and F; state [x_G (4); x_F (4)], padded to 8 states). This is the
21
+ realization the certificate and every prior SISO number use.
22
+ - Verifier call = one `siso_certify.certify(A, B, C, phi)`: L Hurwitz, beta_need = max phi(x)/x over the nonzero
23
+ knots, and `PopovControllerDataset.solve_popov_lmi(L, bmax=max(beta_need, 1e-4), iters=0)` (CLARABEL, SCS
24
+ fallback, numerically re-verified) feasible at beta_need. No rollout ever stands in for the certificate.
25
+
26
+ ### Cost
27
+ - J = mean over 4 initial states of int_0^60 [q(y_G) + u^2] dt, q = `closed_loop.q_state` = y^2 (1 + 1/(1 + y^2)),
28
+ RK4 dt = 0.03 s, trapezoid quadrature, float64.
29
+ - Initial states: plant block x_G(0) ~ N(0, I_4), drawn with `np.random.default_rng((1919, plant_index))`, the
30
+ same 4 states for every arm on that plant; controller (filter / observer) state at rest.
31
+ - Horizon 60 s, not 12 s: open-loop plant poles have real part in [-1.1, -0.1] and generated filters have poles
32
+ as slow as -0.008, so 12 s cuts off a large part of the cost of slow but stable loops and favours fast ones.
33
+ - Three choices differ from the letter of the task; each is forced by the arm-independent reference:
34
+ 1. Cost on the plant output y_G, not the filter output y_L. With y_L, scaling F by c and the spline argument by
35
+ 1/c leaves u(t) and x_G(t) unchanged but scales int q(y_L) by about c^2, so an optimizer that moves F can
36
+ lower J without changing the closed loop. Older protocols fixed F, so this never showed.
37
+ 2. Initial state on the plant block only. In realize_pad the filter state is in F's own canonical coordinates;
38
+ a random filter state injects an F-dependent disturbance (the generated F have C_F entries up to ~100), which
39
+ is not a property of the plant. With the controller at rest, every arm regulates the same plant states.
40
+ 3. The reference differentiates through `siso_x19.realize_torch`, a torch twin of realize_pad for strictly
41
+ proper F (checked equal to 1e-10 in `siso_x19 selfcheck`), not through `diff_realize.realize_loop_diff`:
42
+ realize_loop_diff's single canonical form of G F mixes plant and filter coordinates, so a shared x0 would
43
+ be a different plant initial state for every F.
44
+
45
+ ### Budget and arm result
46
+ - Candidate budget b in {1, 2, 4, 8, 16}. At b, an arm is certified if any of its first b candidates is; its cost
47
+ is that of the cheapest certified candidate among the first b.
48
+ - Fallback (primary regret): if none of the first b is certified, the arm deploys phi = 0, which is always
49
+ certified because G is open-loop stable; its cost is then the open-loop cost. Regret conditional on
50
+ certification is reported as a secondary column.
51
+
52
+ ### Arms
53
+ 1. Generator: shipped `Best_nonlinear_cond_popov_L2.pth`, DDPM-500 (the checkpoint's 500-step squaredcos
54
+ scheduler, `retrieval_baselines.sample_siso`), 16 draws per plant, sampler seeds 0, 1, 2; per-seed results and
55
+ the per-plant mean over seeds.
56
+ 2. Retrieval: the stored controller (first stored row) of each of the b nearest distinct bank plants under the
57
+ per-dimension z-scored metric of `audit_siso_retrieval.py` (the `nn_z` arm). Bank: 5,635 distinct plants,
58
+ 180,160 rows, `data/popov_dataset_normalized_L.pt`.
59
+ 3. Regressor `models/regressor_siso.pt` (deterministic, b = 1). CVAE `models/Best_cvae_siso.pth`
60
+ (`eval_ablation_generative.load_cvae`) at every b, seeds 0, 1, 2.
61
+ 4. Classical, no learning:
62
+ a. phi = 0 (open loop, F = 1).
63
+ b. Static output gain: F = 1, phi(y) = k y. One Popov-LMI bisection on L = G (bmax = 100, 16 iterations:
64
+ 18 LMI solves) gives the certified range [0, beta_G]; k minimizes J over a 41-point grid on that range
65
+ plus a 21-point refinement around the grid minimum (62 rollouts, no verifier calls); one confirming
66
+ certificate. Cost: 19 verifier calls per plant; reported at every b as a b-independent number.
67
+ c. LQG-style observer-based linear controller on G: LQR gain with Q = 2 C'C (q(y) ~ 2y^2 near 0), R = 1;
68
+ observer gain from the dual Riccati equation with W = B B' + 1e-3 I, V = 1; observer at rest. Checked by
69
+ closed-loop eigenvalues. A linear-class reference, not the same controller class (dynamic linear output
70
+ feedback, no sector nonlinearity).
71
+ 5. Independent reference ("oracle"): per plant, 16 cold starts drawn from the dataset prior
72
+ (`PopovControllerDataset._generate_random_sys(4, stable=True, strictly_proper=True)` for F; its certified
73
+ sector beta from `solve_popov_lmi(bmax=1e4, iters=10)`; spline phi(x_k) = x_k r_k, r_k ~ U(0, beta), the
74
+ dataset's `generate_spline`), seeded by (2929, plant index); nothing from any arm. Each start optimized
75
+ jointly over F (n2, n1, n0 and the 4 monic-denominator coefficients) and the 11 knots with Adam through
76
+ realize_torch and the batched simulator on the exact test objective. Parametrization: num = start (1 + t),
77
+ den = start exp(t), knot = start exp(t), so every knot stays in its quadrant, phi(0) = 0 and the denominator
78
+ coefficients stay positive. A step that puts a pole of F at real part >= -0.01 is rejected for that start.
79
+ After the last step the final iterate is certified; if it fails, snapshots (every 25 steps) are certified
80
+ from the last backwards, and the start itself is certified by construction. The cheapest certified start is
81
+ the plant's reference cost. Learning rate and step budget are chosen on validation (Part B).
82
+
83
+ ### Metrics and tests
84
+ - Per arm and b: certified rate (95% percentile bootstrap CI over plants, 10,000 resamples); median and mean
85
+ regret = (J_arm - J_best) / J_best x 100, with J_best = min over the reference and every arm, seed and b on that
86
+ plant (bootstrap CIs); median cost ratio J_arm / J_(phi=0). LQG is left out of J_best because it is a different
87
+ controller class (dynamic linear feedback); how often it is cheaper than J_best is reported. The static gain
88
+ (F = 1) is in J_best. Secondary: regret against the reference alone, (J_arm - J_ref) / J_ref.
89
+ - Paired differences, generator minus each other arm, at b = 1, 4, 16, on per-plant regret (generator value =
90
+ mean over its 3 sampler seeds; plants are the resampling unit, never sampler seeds): bootstrap CI of the mean
91
+ and median difference, two-sided Wilcoxon signed-rank p, Holm correction over the family {retrieval, CVAE,
92
+ regressor (b = 1 only), phi = 0, static gain, LQG} x {1, 4, 16} (16 tests). Certified-rate differences:
93
+ bootstrap CI and exact McNemar (secondary, not Holm-corrected).
94
+ - Old criterion re-run on the test plants for the generator (3 seeds) and retrieval: the pre-X19 verifier
95
+ (`retrieval_baselines.siso_batch_verify`: 12 s, final energy < 0.3, x0 ~ N(0, I_8) on the loop state with torch
96
+ seed 777, state clamp 20), "any of the first b passes", next to the certified rate.
97
+ - No directional hypothesis is pre-set; the outcome is reported as measured.
98
+
99
+ ## Part B: validation protocol for the reference (fixed before the validation sweep finished)
100
+ - Sweep on the 40 validation plants: Adam learning rate in {0.02, 0.05, 0.1} (scale-free units: relative change
101
+ per step), 600 steps each, dt = 0.03 s (the exact test objective), 16 starts, snapshots every 25 steps.
102
+ - Rule: lr = the value with the lowest mean over plants of log(best-of-16 training loss at step 600). Step budget
103
+ = the smallest multiple of 100 at which, for that lr, the plant-mean best-of-16 loss is within 0.5% of its
104
+ step-600 value; if the wall-time estimate for 150 test plants on 12 workers exceeds 90 min, take the largest
105
+ multiple of 100 that fits and log the deviation.
106
+ - Static gain, LQG, phi = 0, retrieval, regressor, CVAE and the generator have nothing tuned.
107
+
108
+ ## Part B result: frozen reference configuration (validation only; written before the test reference ran)
109
+ - `results/icml_r3/X19/val_selection.txt`, `FROZEN.json`: lr = 0.1, 400 steps, 16 starts, snapshots every 25 steps,
110
+ prior bisection bmax = 1e4 with 10 iterations, F pole margin 0.01, dt = 0.03 s. Plant-mean best-of-16 training
111
+ loss at steps 300 / 400 / 600: 25.879 / 25.846 / 25.723 (lr 0.1); lr 0.02 and 0.05 end at 25.863 and 25.783.
112
+ - On validation the reference's final iterate is certified for 99.2% of starts (lr 0.1); best-of-8 starts is a
113
+ median 0.07% above best-of-16 (lr 0.05).
114
+ - Test wall estimate: about 50-60 min on 12 workers (13 plants per worker, ~7 s per Adam step measured under the
115
+ current host load). The estimate printed by `siso_x19 select` (43 min) assumed 10 plants per worker for every lr
116
+ and is low for lr 0.1, which ran on 8 workers with 5 plants each. The user lifted the 6 h cap for CPU work, so
117
+ the preregistered 16 starts stay.
118
+
119
+ ## Deviations and events (logged before test scoring of the reference)
120
+ 1. The first lr = 0.1 validation run crashed: a diverging start gave a non-finite gradient, Adam wrote NaN into a
121
+ filter coefficient, and `np.roots` raised. Fix: non-finite gradients are zeroed before the step, and a step that
122
+ leaves any F coefficient or knot non-finite is rejected for that start, like an unstable F. The fix is a no-op
123
+ for runs that never produce a non-finite value, so the finished lr = 0.02 and 0.05 runs are unchanged; lr = 0.1
124
+ was rerun with the fix on 8 workers (log of the crash: `logs/val_oracle_lr0.1_crashed_nan.log`).
125
+ 2. The test arms (generator, retrieval, regressor, CVAE, phi = 0, static gain, LQG, old criterion) were scored
126
+ before the reference configuration was frozen. None of them has a tunable setting (Part A fixes all of them),
127
+ and the selection rule reads only `val/oracle_lr*.npz`, so no test number entered the selection.
128
+ 3. Post hoc, at the coordinator's request and NOT part of this preregistration: 200 random training-bank rows were
129
+ scored under the X19 cost (`posthoc_labels.txt`). Reported separately from the preregistered table.
results/icml_r3/X19/logs/posthoc_labels.log ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ /home/cercola/matte_server/DiffusionControl/myenv/lib/python3.13/site-packages/scipy/signal/_lti_conversion.py:77: BadCoefficients: Badly conditioned filter coefficients (numerator): the results may be meaningless
2
+ num, den = normalize(num, den) # Strips zeros, checks arrays
3
+ /home/cercola/matte_server/DiffusionControl/myenv/lib/python3.13/site-packages/scipy/signal/_lti_conversion.py:77: BadCoefficients: Badly conditioned filter coefficients (numerator): the results may be meaningless
4
+ num, den = normalize(num, den) # Strips zeros, checks arrays
5
+ /home/cercola/matte_server/DiffusionControl/myenv/lib/python3.13/site-packages/scipy/signal/_lti_conversion.py:77: BadCoefficients: Badly conditioned filter coefficients (numerator): the results may be meaningless
6
+ num, den = normalize(num, den) # Strips zeros, checks arrays
7
+ /home/cercola/matte_server/DiffusionControl/myenv/lib/python3.13/site-packages/scipy/signal/_lti_conversion.py:77: BadCoefficients: Badly conditioned filter coefficients (numerator): the results may be meaningless
8
+ num, den = normalize(num, den) # Strips zeros, checks arrays
9
+ /home/cercola/matte_server/DiffusionControl/nonlinear/data_generation/build_dataset.py:131: UserWarning: Solution may be inaccurate. Try another solver, adjusting the solver settings, or solve with verbose=True for more information.
10
+ cp.Problem(cp.Minimize(0), cons).solve(solver=solver, verbose=False)
11
+ X19 POST HOC (not preregistered): training labels under the X19 cost
12
+ 200 random rows (seed 1234) of data/popov_dataset_normalized_L.pt (180,160 rows, 5,635 plants); rows with a clipped G skipped
13
+ each row's own controller (F, spline) on its own plant; same x0 convention, 60 s cost on y_G, real Popov certificate
14
+ row controller certified (corrected LMI): 99.0% [97.5, 100.0]
15
+ median J_row / J(phi=0): 0.999 [0.999, 1.000]; rows cheaper than phi=0: 73.0%
16
+ median J_row / J(best certified static gain): 1.189 [1.163, 1.269]; rows cheaper than the static gain: 0.0%
17
+ same, certified rows only (n=198): J/J0 0.999, J/Jsg 1.187
18
+ median J(static gain) / J(phi=0) on these plants: 0.837; static gain certified 100.0%
19
+ dataset objective (code): none. build_dataset.PopovControllerDataset.create_entry draws a random stable strictly proper F,
20
+ computes the Popov sector beta of L = G F, and draws the spline as phi(x_k) = x_k r_k with r_k ~ U(0, beta)
21
+ (generate_spline); no cost is evaluated or minimised. CIs: 95% percentile bootstrap over rows (10,000).
results/icml_r3/X19/logs/test_arms.log ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ generator seed 0: 99.8 ms/controller (DDPM-500, CPU 4 threads)
2
+ generator seed 1: 100.4 ms/controller (DDPM-500, CPU 4 threads)
3
+ generator seed 2: 97.0 ms/controller (DDPM-500, CPU 4 threads)
4
+ /home/cercola/matte_server/DiffusionControl/nonlinear/data_generation/build_dataset.py:131: UserWarning: Solution may be inaccurate. Try another solver, adjusting the solver settings, or solve with verbose=True for more information.
5
+ cp.Problem(cp.Minimize(0), cons).solve(solver=solver, verbose=False)
6
+ /home/cercola/matte_server/DiffusionControl/nonlinear/data_generation/build_dataset.py:131: UserWarning: Solution may be inaccurate. Try another solver, adjusting the solver settings, or solve with verbose=True for more information.
7
+ cp.Problem(cp.Minimize(0), cons).solve(solver=solver, verbose=False)
8
+ /home/cercola/matte_server/DiffusionControl/nonlinear/data_generation/build_dataset.py:131: UserWarning: Solution may be inaccurate. Try another solver, adjusting the solver settings, or solve with verbose=True for more information.
9
+ cp.Problem(cp.Minimize(0), cons).solve(solver=solver, verbose=False)
10
+ /home/cercola/matte_server/DiffusionControl/nonlinear/data_generation/build_dataset.py:131: UserWarning: Solution may be inaccurate. Try another solver, adjusting the solver settings, or solve with verbose=True for more information.
11
+ cp.Problem(cp.Minimize(0), cons).solve(solver=solver, verbose=False)
12
+ /home/cercola/matte_server/DiffusionControl/myenv/lib/python3.13/site-packages/scipy/signal/_lti_conversion.py:77: BadCoefficients: Badly conditioned filter coefficients (numerator): the results may be meaningless
13
+ num, den = normalize(num, den) # Strips zeros, checks arrays
14
+ /home/cercola/matte_server/DiffusionControl/myenv/lib/python3.13/site-packages/scipy/signal/_lti_conversion.py:77: BadCoefficients: Badly conditioned filter coefficients (numerator): the results may be meaningless
15
+ num, den = normalize(num, den) # Strips zeros, checks arrays
16
+ /home/cercola/matte_server/DiffusionControl/myenv/lib/python3.13/site-packages/scipy/signal/_lti_conversion.py:77: BadCoefficients: Badly conditioned filter coefficients (numerator): the results may be meaningless
17
+ num, den = normalize(num, den) # Strips zeros, checks arrays
18
+ /home/cercola/matte_server/DiffusionControl/myenv/lib/python3.13/site-packages/scipy/signal/_lti_conversion.py:77: BadCoefficients: Badly conditioned filter coefficients (numerator): the results may be meaningless
19
+ num, den = normalize(num, den) # Strips zeros, checks arrays
20
+ scored 16950 candidates in 91 s
21
+ /home/cercola/matte_server/DiffusionControl/nonlinear/data_generation/build_dataset.py:131: UserWarning: Solution may be inaccurate. Try another solver, adjusting the solver settings, or solve with verbose=True for more information.
22
+ cp.Problem(cp.Minimize(0), cons).solve(solver=solver, verbose=False)
23
+ /home/cercola/matte_server/DiffusionControl/myenv/lib/python3.13/site-packages/scipy/signal/_lti_conversion.py:77: BadCoefficients: Badly conditioned filter coefficients (numerator): the results may be meaningless
24
+ num, den = normalize(num, den) # Strips zeros, checks arrays
25
+ arms done in 891 s -> results/icml_r3/X19/test/arms.npz
results/icml_r3/X19/logs/test_oracle.log ADDED
@@ -0,0 +1,121 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ /home/cercola/matte_server/DiffusionControl/nonlinear/data_generation/build_dataset.py:131: UserWarning: Solution may be inaccurate. Try another solver, adjusting the solver settings, or solve with verbose=True for more information.
2
+ cp.Problem(cp.Minimize(0), cons).solve(solver=solver, verbose=False)
3
+ /home/cercola/matte_server/DiffusionControl/nonlinear/data_generation/build_dataset.py:131: UserWarning: Solution may be inaccurate. Try another solver, adjusting the solver settings, or solve with verbose=True for more information.
4
+ cp.Problem(cp.Minimize(0), cons).solve(solver=solver, verbose=False)
5
+ /home/cercola/matte_server/DiffusionControl/nonlinear/data_generation/build_dataset.py:131: UserWarning: Solution may be inaccurate. Try another solver, adjusting the solver settings, or solve with verbose=True for more information.
6
+ cp.Problem(cp.Minimize(0), cons).solve(solver=solver, verbose=False)
7
+ /home/cercola/matte_server/DiffusionControl/nonlinear/data_generation/build_dataset.py:131: UserWarning: Solution may be inaccurate. Try another solver, adjusting the solver settings, or solve with verbose=True for more information.
8
+ cp.Problem(cp.Minimize(0), cons).solve(solver=solver, verbose=False)
9
+ /home/cercola/matte_server/DiffusionControl/nonlinear/data_generation/build_dataset.py:131: UserWarning: Solution may be inaccurate. Try another solver, adjusting the solver settings, or solve with verbose=True for more information.
10
+ cp.Problem(cp.Minimize(0), cons).solve(solver=solver, verbose=False)
11
+ /home/cercola/matte_server/DiffusionControl/nonlinear/data_generation/build_dataset.py:131: UserWarning: Solution may be inaccurate. Try another solver, adjusting the solver settings, or solve with verbose=True for more information.
12
+ cp.Problem(cp.Minimize(0), cons).solve(solver=solver, verbose=False)
13
+ /home/cercola/matte_server/DiffusionControl/nonlinear/data_generation/build_dataset.py:131: UserWarning: Solution may be inaccurate. Try another solver, adjusting the solver settings, or solve with verbose=True for more information.
14
+ cp.Problem(cp.Minimize(0), cons).solve(solver=solver, verbose=False)
15
+ /home/cercola/matte_server/DiffusionControl/nonlinear/data_generation/build_dataset.py:131: UserWarning: Solution may be inaccurate. Try another solver, adjusting the solver settings, or solve with verbose=True for more information.
16
+ cp.Problem(cp.Minimize(0), cons).solve(solver=solver, verbose=False)
17
+ /home/cercola/matte_server/DiffusionControl/nonlinear/data_generation/build_dataset.py:131: UserWarning: Solution may be inaccurate. Try another solver, adjusting the solver settings, or solve with verbose=True for more information.
18
+ cp.Problem(cp.Minimize(0), cons).solve(solver=solver, verbose=False)
19
+ /home/cercola/matte_server/DiffusionControl/nonlinear/data_generation/build_dataset.py:131: UserWarning: Solution may be inaccurate. Try another solver, adjusting the solver settings, or solve with verbose=True for more information.
20
+ cp.Problem(cp.Minimize(0), cons).solve(solver=solver, verbose=False)
21
+ /home/cercola/matte_server/DiffusionControl/nonlinear/data_generation/build_dataset.py:131: UserWarning: Solution may be inaccurate. Try another solver, adjusting the solver settings, or solve with verbose=True for more information.
22
+ cp.Problem(cp.Minimize(0), cons).solve(solver=solver, verbose=False)
23
+ /home/cercola/matte_server/DiffusionControl/nonlinear/data_generation/build_dataset.py:131: UserWarning: Solution may be inaccurate. Try another solver, adjusting the solver settings, or solve with verbose=True for more information.
24
+ cp.Problem(cp.Minimize(0), cons).solve(solver=solver, verbose=False)
25
+ [plants 143-149] step 50/400 246 s mean best-of-16 10.7532
26
+ [plants 117-129] step 50/400 338 s mean best-of-16 8.8347
27
+ [plants 104-116] step 50/400 342 s mean best-of-16 30.2747
28
+ [plants 78-90] step 50/400 342 s mean best-of-16 14.1117
29
+ [plants 130-142] step 50/400 342 s mean best-of-16 89.8495
30
+ [plants 39-51] step 50/400 343 s mean best-of-16 21.5019
31
+ [plants 143-149] step 100/400 517 s mean best-of-16 10.1936
32
+ [plants 65-77] step 50/400 343 s mean best-of-16 25.5908
33
+ [plants 0-12] step 50/400 343 s mean best-of-16 26.9942
34
+ [plants 13-25] step 50/400 342 s mean best-of-16 40.1623
35
+ [plants 52-64] step 50/400 343 s mean best-of-16 47.0111
36
+ [plants 91-103] step 50/400 341 s mean best-of-16 34.2187
37
+ [plants 26-38] step 50/400 344 s mean best-of-16 32.4092
38
+ [plants 143-149] step 150/400 783 s mean best-of-16 9.9911
39
+ [plants 117-129] step 100/400 679 s mean best-of-16 8.7743
40
+ [plants 104-116] step 100/400 681 s mean best-of-16 29.8859
41
+ [plants 78-90] step 100/400 682 s mean best-of-16 13.9679
42
+ [plants 130-142] step 100/400 684 s mean best-of-16 85.3213
43
+ [plants 39-51] step 100/400 685 s mean best-of-16 21.3309
44
+ [plants 65-77] step 100/400 686 s mean best-of-16 25.3684
45
+ [plants 0-12] step 100/400 683 s mean best-of-16 26.3377
46
+ [plants 52-64] step 100/400 685 s mean best-of-16 46.3254
47
+ [plants 13-25] step 100/400 684 s mean best-of-16 39.1500
48
+ [plants 91-103] step 100/400 682 s mean best-of-16 33.5267
49
+ [plants 26-38] step 100/400 684 s mean best-of-16 31.6725
50
+ [plants 143-149] step 200/400 1053 s mean best-of-16 9.9867
51
+ [plants 117-129] step 150/400 1023 s mean best-of-16 8.7450
52
+ [plants 104-116] step 150/400 1019 s mean best-of-16 28.9234
53
+ [plants 78-90] step 150/400 1021 s mean best-of-16 13.9046
54
+ [plants 130-142] step 150/400 1024 s mean best-of-16 84.0869
55
+ [plants 39-51] step 150/400 1025 s mean best-of-16 21.2518
56
+ [plants 65-77] step 150/400 1026 s mean best-of-16 25.2922
57
+ [plants 0-12] step 150/400 1024 s mean best-of-16 26.1220
58
+ [plants 13-25] step 150/400 1024 s mean best-of-16 37.3522
59
+ [plants 52-64] step 150/400 1025 s mean best-of-16 46.1132
60
+ [plants 91-103] step 150/400 1023 s mean best-of-16 33.3868
61
+ [plants 26-38] step 150/400 1020 s mean best-of-16 31.4557
62
+ [plants 143-149] step 250/400 1319 s mean best-of-16 9.9827
63
+ [plants 117-129] step 200/400 1363 s mean best-of-16 8.7244
64
+ [plants 104-116] step 200/400 1360 s mean best-of-16 28.5828
65
+ [plants 78-90] step 200/400 1359 s mean best-of-16 13.8803
66
+ [plants 130-142] step 200/400 1365 s mean best-of-16 82.8415
67
+ [plants 39-51] step 200/400 1366 s mean best-of-16 21.2101
68
+ [plants 65-77] step 200/400 1367 s mean best-of-16 25.2350
69
+ [plants 0-12] step 200/400 1365 s mean best-of-16 26.0456
70
+ [plants 13-25] step 200/400 1366 s mean best-of-16 36.9915
71
+ [plants 52-64] step 200/400 1367 s mean best-of-16 46.0113
72
+ [plants 26-38] step 200/400 1359 s mean best-of-16 31.3876
73
+ [plants 91-103] step 200/400 1364 s mean best-of-16 33.1737
74
+ [plants 143-149] step 300/400 1586 s mean best-of-16 9.7882
75
+ [plants 117-129] step 250/400 1703 s mean best-of-16 8.7128
76
+ [plants 104-116] step 250/400 1700 s mean best-of-16 28.3972
77
+ [plants 143-149] step 350/400 1852 s mean best-of-16 9.5329
78
+ [plants 78-90] step 250/400 1700 s mean best-of-16 13.8681
79
+ [plants 130-142] step 250/400 1706 s mean best-of-16 81.1111
80
+ [plants 39-51] step 250/400 1707 s mean best-of-16 21.1881
81
+ [plants 65-77] step 250/400 1709 s mean best-of-16 25.0977
82
+ [plants 0-12] step 250/400 1706 s mean best-of-16 26.0054
83
+ [plants 13-25] step 250/400 1707 s mean best-of-16 36.9439
84
+ [plants 52-64] step 250/400 1708 s mean best-of-16 45.9413
85
+ [plants 26-38] step 250/400 1702 s mean best-of-16 31.3410
86
+ [plants 91-103] step 250/400 1707 s mean best-of-16 32.3650
87
+ [plants 143-149] step 400/400 2122 s mean best-of-16 9.5313
88
+ [plants 117-129] step 300/400 2046 s mean best-of-16 8.7069
89
+ [plants 104-116] step 300/400 2043 s mean best-of-16 28.1922
90
+ [plants 78-90] step 300/400 2042 s mean best-of-16 13.8545
91
+ [plants 130-142] step 300/400 2047 s mean best-of-16 81.0208
92
+ [plants 39-51] step 300/400 2048 s mean best-of-16 21.1652
93
+ [plants 65-77] step 300/400 2049 s mean best-of-16 24.8493
94
+ [plants 0-12] step 300/400 2045 s mean best-of-16 25.9745
95
+ [plants 52-64] step 300/400 2048 s mean best-of-16 45.8897
96
+ [plants 13-25] step 300/400 2047 s mean best-of-16 36.7279
97
+ [plants 26-38] step 300/400 2041 s mean best-of-16 31.2946
98
+ [plants 91-103] step 300/400 2046 s mean best-of-16 32.3195
99
+ [plants 117-129] step 350/400 2383 s mean best-of-16 8.7028
100
+ [plants 104-116] step 350/400 2380 s mean best-of-16 28.1257
101
+ [plants 78-90] step 350/400 2379 s mean best-of-16 13.8588
102
+ [plants 130-142] step 350/400 2384 s mean best-of-16 80.8192
103
+ [plants 39-51] step 350/400 2385 s mean best-of-16 21.1375
104
+ [plants 65-77] step 350/400 2386 s mean best-of-16 24.8225
105
+ [plants 0-12] step 350/400 2383 s mean best-of-16 25.9445
106
+ [plants 52-64] step 350/400 2386 s mean best-of-16 45.8506
107
+ [plants 13-25] step 350/400 2385 s mean best-of-16 36.6736
108
+ [plants 26-38] step 350/400 2378 s mean best-of-16 31.2569
109
+ [plants 91-103] step 350/400 2383 s mean best-of-16 32.2955
110
+ [plants 117-129] step 400/400 2719 s mean best-of-16 8.6978
111
+ [plants 104-116] step 400/400 2715 s mean best-of-16 28.1225
112
+ [plants 78-90] step 400/400 2714 s mean best-of-16 13.8567
113
+ [plants 130-142] step 400/400 2720 s mean best-of-16 80.6878
114
+ [plants 39-51] step 400/400 2719 s mean best-of-16 21.1291
115
+ [plants 0-12] step 400/400 2716 s mean best-of-16 25.8845
116
+ [plants 65-77] step 400/400 2720 s mean best-of-16 24.8109
117
+ [plants 13-25] step 400/400 2718 s mean best-of-16 36.6355
118
+ [plants 52-64] step 400/400 2719 s mean best-of-16 45.8161
119
+ [plants 26-38] step 400/400 2710 s mean best-of-16 31.2233
120
+ [plants 91-103] step 400/400 2715 s mean best-of-16 32.2876
121
+ oracle: 3081 s
results/icml_r3/X19/logs/val_arms.log ADDED
@@ -0,0 +1,33 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ generator seed 0: 28.0 ms/controller (DDPM-500, CPU 12 threads)
2
+ generator seed 1: 29.8 ms/controller (DDPM-500, CPU 12 threads)
3
+ generator seed 2: 30.0 ms/controller (DDPM-500, CPU 12 threads)
4
+ /home/cercola/matte_server/DiffusionControl/myenv/lib/python3.13/site-packages/scipy/signal/_lti_conversion.py:77: BadCoefficients: Badly conditioned filter coefficients (numerator): the results may be meaningless
5
+ num, den = normalize(num, den) # Strips zeros, checks arrays
6
+ /home/cercola/matte_server/DiffusionControl/myenv/lib/python3.13/site-packages/scipy/signal/_lti_conversion.py:77: BadCoefficients: Badly conditioned filter coefficients (numerator): the results may be meaningless
7
+ num, den = normalize(num, den) # Strips zeros, checks arrays
8
+ /home/cercola/matte_server/DiffusionControl/myenv/lib/python3.13/site-packages/scipy/signal/_lti_conversion.py:77: BadCoefficients: Badly conditioned filter coefficients (numerator): the results may be meaningless
9
+ num, den = normalize(num, den) # Strips zeros, checks arrays
10
+ /home/cercola/matte_server/DiffusionControl/myenv/lib/python3.13/site-packages/scipy/signal/_lti_conversion.py:77: BadCoefficients: Badly conditioned filter coefficients (numerator): the results may be meaningless
11
+ num, den = normalize(num, den) # Strips zeros, checks arrays
12
+ /home/cercola/matte_server/DiffusionControl/myenv/lib/python3.13/site-packages/scipy/signal/_lti_conversion.py:77: BadCoefficients: Badly conditioned filter coefficients (numerator): the results may be meaningless
13
+ num, den = normalize(num, den) # Strips zeros, checks arrays
14
+ /home/cercola/matte_server/DiffusionControl/myenv/lib/python3.13/site-packages/scipy/signal/_lti_conversion.py:77: BadCoefficients: Badly conditioned filter coefficients (numerator): the results may be meaningless
15
+ num, den = normalize(num, den) # Strips zeros, checks arrays
16
+ /home/cercola/matte_server/DiffusionControl/myenv/lib/python3.13/site-packages/scipy/signal/_lti_conversion.py:77: BadCoefficients: Badly conditioned filter coefficients (numerator): the results may be meaningless
17
+ num, den = normalize(num, den) # Strips zeros, checks arrays
18
+ /home/cercola/matte_server/DiffusionControl/myenv/lib/python3.13/site-packages/scipy/signal/_lti_conversion.py:77: BadCoefficients: Badly conditioned filter coefficients (numerator): the results may be meaningless
19
+ num, den = normalize(num, den) # Strips zeros, checks arrays
20
+ /home/cercola/matte_server/DiffusionControl/myenv/lib/python3.13/site-packages/scipy/signal/_lti_conversion.py:77: BadCoefficients: Badly conditioned filter coefficients (numerator): the results may be meaningless
21
+ num, den = normalize(num, den) # Strips zeros, checks arrays
22
+ /home/cercola/matte_server/DiffusionControl/myenv/lib/python3.13/site-packages/scipy/signal/_lti_conversion.py:77: BadCoefficients: Badly conditioned filter coefficients (numerator): the results may be meaningless
23
+ num, den = normalize(num, den) # Strips zeros, checks arrays
24
+ /home/cercola/matte_server/DiffusionControl/myenv/lib/python3.13/site-packages/scipy/signal/_lti_conversion.py:77: BadCoefficients: Badly conditioned filter coefficients (numerator): the results may be meaningless
25
+ num, den = normalize(num, den) # Strips zeros, checks arrays
26
+ /home/cercola/matte_server/DiffusionControl/myenv/lib/python3.13/site-packages/scipy/signal/_lti_conversion.py:77: BadCoefficients: Badly conditioned filter coefficients (numerator): the results may be meaningless
27
+ num, den = normalize(num, den) # Strips zeros, checks arrays
28
+ scored 4520 candidates in 7 s
29
+ /home/cercola/matte_server/DiffusionControl/nonlinear/data_generation/build_dataset.py:131: UserWarning: Solution may be inaccurate. Try another solver, adjusting the solver settings, or solve with verbose=True for more information.
30
+ cp.Problem(cp.Minimize(0), cons).solve(solver=solver, verbose=False)
31
+ /home/cercola/matte_server/DiffusionControl/myenv/lib/python3.13/site-packages/scipy/signal/_lti_conversion.py:77: BadCoefficients: Badly conditioned filter coefficients (numerator): the results may be meaningless
32
+ num, den = normalize(num, den) # Strips zeros, checks arrays
33
+ arms done in 86 s -> results/icml_r3/X19/val/arms.npz
results/icml_r3/X19/logs/val_oracle_lr0.02.log ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ /home/cercola/matte_server/DiffusionControl/nonlinear/data_generation/build_dataset.py:131: UserWarning: Solution may be inaccurate. Try another solver, adjusting the solver settings, or solve with verbose=True for more information.
2
+ cp.Problem(cp.Minimize(0), cons).solve(solver=solver, verbose=False)
3
+ /home/cercola/matte_server/DiffusionControl/nonlinear/data_generation/build_dataset.py:131: UserWarning: Solution may be inaccurate. Try another solver, adjusting the solver settings, or solve with verbose=True for more information.
4
+ cp.Problem(cp.Minimize(0), cons).solve(solver=solver, verbose=False)
5
+ /home/cercola/matte_server/DiffusionControl/nonlinear/data_generation/build_dataset.py:131: UserWarning: Solution may be inaccurate. Try another solver, adjusting the solver settings, or solve with verbose=True for more information.
6
+ cp.Problem(cp.Minimize(0), cons).solve(solver=solver, verbose=False)
7
+ /home/cercola/matte_server/DiffusionControl/nonlinear/data_generation/build_dataset.py:131: UserWarning: Solution may be inaccurate. Try another solver, adjusting the solver settings, or solve with verbose=True for more information.
8
+ cp.Problem(cp.Minimize(0), cons).solve(solver=solver, verbose=False)
9
+ oracle lr=0.02: 3882 s
results/icml_r3/X19/logs/val_oracle_lr0.05.log ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ /home/cercola/matte_server/DiffusionControl/nonlinear/data_generation/build_dataset.py:131: UserWarning: Solution may be inaccurate. Try another solver, adjusting the solver settings, or solve with verbose=True for more information.
2
+ cp.Problem(cp.Minimize(0), cons).solve(solver=solver, verbose=False)
3
+ /home/cercola/matte_server/DiffusionControl/nonlinear/data_generation/build_dataset.py:131: UserWarning: Solution may be inaccurate. Try another solver, adjusting the solver settings, or solve with verbose=True for more information.
4
+ cp.Problem(cp.Minimize(0), cons).solve(solver=solver, verbose=False)
5
+ /home/cercola/matte_server/DiffusionControl/nonlinear/data_generation/build_dataset.py:131: UserWarning: Solution may be inaccurate. Try another solver, adjusting the solver settings, or solve with verbose=True for more information.
6
+ cp.Problem(cp.Minimize(0), cons).solve(solver=solver, verbose=False)
7
+ /home/cercola/matte_server/DiffusionControl/nonlinear/data_generation/build_dataset.py:131: UserWarning: Solution may be inaccurate. Try another solver, adjusting the solver settings, or solve with verbose=True for more information.
8
+ cp.Problem(cp.Minimize(0), cons).solve(solver=solver, verbose=False)
9
+ oracle lr=0.05: 3938 s
results/icml_r3/X19/logs/val_oracle_lr0.1.log ADDED
@@ -0,0 +1,113 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ /home/cercola/matte_server/DiffusionControl/nonlinear/data_generation/build_dataset.py:131: UserWarning: Solution may be inaccurate. Try another solver, adjusting the solver settings, or solve with verbose=True for more information.
2
+ cp.Problem(cp.Minimize(0), cons).solve(solver=solver, verbose=False)
3
+ /home/cercola/matte_server/DiffusionControl/nonlinear/data_generation/build_dataset.py:131: UserWarning: Solution may be inaccurate. Try another solver, adjusting the solver settings, or solve with verbose=True for more information.
4
+ cp.Problem(cp.Minimize(0), cons).solve(solver=solver, verbose=False)
5
+ /home/cercola/matte_server/DiffusionControl/nonlinear/data_generation/build_dataset.py:131: UserWarning: Solution may be inaccurate. Try another solver, adjusting the solver settings, or solve with verbose=True for more information.
6
+ cp.Problem(cp.Minimize(0), cons).solve(solver=solver, verbose=False)
7
+ /home/cercola/matte_server/DiffusionControl/nonlinear/data_generation/build_dataset.py:131: UserWarning: Solution may be inaccurate. Try another solver, adjusting the solver settings, or solve with verbose=True for more information.
8
+ cp.Problem(cp.Minimize(0), cons).solve(solver=solver, verbose=False)
9
+ /home/cercola/matte_server/DiffusionControl/nonlinear/data_generation/build_dataset.py:131: UserWarning: Solution may be inaccurate. Try another solver, adjusting the solver settings, or solve with verbose=True for more information.
10
+ cp.Problem(cp.Minimize(0), cons).solve(solver=solver, verbose=False)
11
+ /home/cercola/matte_server/DiffusionControl/nonlinear/data_generation/build_dataset.py:131: UserWarning: Solution may be inaccurate. Try another solver, adjusting the solver settings, or solve with verbose=True for more information.
12
+ cp.Problem(cp.Minimize(0), cons).solve(solver=solver, verbose=False)
13
+ /home/cercola/matte_server/DiffusionControl/nonlinear/data_generation/build_dataset.py:131: UserWarning: Solution may be inaccurate. Try another solver, adjusting the solver settings, or solve with verbose=True for more information.
14
+ cp.Problem(cp.Minimize(0), cons).solve(solver=solver, verbose=False)
15
+ /home/cercola/matte_server/DiffusionControl/nonlinear/data_generation/build_dataset.py:131: UserWarning: Solution may be inaccurate. Try another solver, adjusting the solver settings, or solve with verbose=True for more information.
16
+ cp.Problem(cp.Minimize(0), cons).solve(solver=solver, verbose=False)
17
+ [plants 5-9] step 50/600 240 s mean best-of-16 5.9222
18
+ [plants 35-39] step 50/600 242 s mean best-of-16 30.3595
19
+ [plants 25-29] step 50/600 242 s mean best-of-16 29.5510
20
+ [plants 0-4] step 50/600 243 s mean best-of-16 18.6369
21
+ [plants 10-14] step 50/600 242 s mean best-of-16 31.6388
22
+ [plants 30-34] step 50/600 243 s mean best-of-16 15.4700
23
+ [plants 20-24] step 50/600 244 s mean best-of-16 46.7941
24
+ [plants 15-19] step 50/600 244 s mean best-of-16 31.8007
25
+ [plants 5-9] step 100/600 488 s mean best-of-16 5.8819
26
+ [plants 35-39] step 100/600 488 s mean best-of-16 30.0417
27
+ [plants 25-29] step 100/600 487 s mean best-of-16 29.4066
28
+ [plants 0-4] step 100/600 489 s mean best-of-16 18.5308
29
+ [plants 10-14] step 100/600 486 s mean best-of-16 31.4163
30
+ [plants 30-34] step 100/600 490 s mean best-of-16 15.3861
31
+ [plants 20-24] step 100/600 488 s mean best-of-16 46.2581
32
+ [plants 15-19] step 100/600 489 s mean best-of-16 31.6393
33
+ [plants 5-9] step 150/600 734 s mean best-of-16 5.8682
34
+ [plants 25-29] step 150/600 729 s mean best-of-16 29.3565
35
+ [plants 35-39] step 150/600 731 s mean best-of-16 29.9257
36
+ [plants 0-4] step 150/600 732 s mean best-of-16 18.4839
37
+ [plants 10-14] step 150/600 728 s mean best-of-16 31.3648
38
+ [plants 30-34] step 150/600 731 s mean best-of-16 15.3668
39
+ [plants 20-24] step 150/600 728 s mean best-of-16 45.9615
40
+ [plants 15-19] step 150/600 729 s mean best-of-16 31.4649
41
+ [plants 5-9] step 200/600 971 s mean best-of-16 5.8591
42
+ [plants 35-39] step 200/600 965 s mean best-of-16 29.8832
43
+ [plants 25-29] step 200/600 963 s mean best-of-16 29.3150
44
+ [plants 0-4] step 200/600 965 s mean best-of-16 18.4461
45
+ [plants 10-14] step 200/600 961 s mean best-of-16 31.3286
46
+ [plants 30-34] step 200/600 964 s mean best-of-16 15.3519
47
+ [plants 20-24] step 200/600 959 s mean best-of-16 45.8052
48
+ [plants 15-19] step 200/600 964 s mean best-of-16 31.4185
49
+ [plants 5-9] step 250/600 1205 s mean best-of-16 5.8458
50
+ [plants 25-29] step 250/600 1196 s mean best-of-16 29.2901
51
+ [plants 35-39] step 250/600 1198 s mean best-of-16 29.8635
52
+ [plants 0-4] step 250/600 1199 s mean best-of-16 18.4170
53
+ [plants 10-14] step 250/600 1196 s mean best-of-16 31.3043
54
+ [plants 30-34] step 250/600 1196 s mean best-of-16 15.3416
55
+ [plants 20-24] step 250/600 1187 s mean best-of-16 45.7218
56
+ [plants 15-19] step 250/600 1195 s mean best-of-16 31.3958
57
+ [plants 5-9] step 300/600 1438 s mean best-of-16 5.8425
58
+ [plants 25-29] step 300/600 1428 s mean best-of-16 29.2733
59
+ [plants 35-39] step 300/600 1431 s mean best-of-16 29.8505
60
+ [plants 0-4] step 300/600 1431 s mean best-of-16 18.3946
61
+ [plants 10-14] step 300/600 1428 s mean best-of-16 31.2900
62
+ [plants 30-34] step 300/600 1427 s mean best-of-16 15.3325
63
+ [plants 20-24] step 300/600 1420 s mean best-of-16 45.6683
64
+ [plants 15-19] step 300/600 1426 s mean best-of-16 31.3821
65
+ [plants 5-9] step 350/600 1672 s mean best-of-16 5.8397
66
+ [plants 35-39] step 350/600 1660 s mean best-of-16 29.8397
67
+ [plants 25-29] step 350/600 1662 s mean best-of-16 29.2615
68
+ [plants 0-4] step 350/600 1665 s mean best-of-16 18.3769
69
+ [plants 10-14] step 350/600 1661 s mean best-of-16 31.2794
70
+ [plants 30-34] step 350/600 1664 s mean best-of-16 15.3170
71
+ [plants 20-24] step 350/600 1653 s mean best-of-16 45.6276
72
+ [plants 15-19] step 350/600 1658 s mean best-of-16 31.3737
73
+ [plants 5-9] step 400/600 1907 s mean best-of-16 5.8373
74
+ [plants 35-39] step 400/600 1897 s mean best-of-16 29.8309
75
+ [plants 25-29] step 400/600 1900 s mean best-of-16 29.2528
76
+ [plants 0-4] step 400/600 1902 s mean best-of-16 18.3629
77
+ [plants 10-14] step 400/600 1896 s mean best-of-16 31.2699
78
+ [plants 20-24] step 400/600 1885 s mean best-of-16 45.5980
79
+ [plants 30-34] step 400/600 1901 s mean best-of-16 15.2943
80
+ [plants 15-19] step 400/600 1895 s mean best-of-16 31.3234
81
+ [plants 5-9] step 450/600 2145 s mean best-of-16 5.8347
82
+ [plants 35-39] step 450/600 2130 s mean best-of-16 29.8244
83
+ [plants 25-29] step 450/600 2136 s mean best-of-16 29.2473
84
+ [plants 10-14] step 450/600 2130 s mean best-of-16 31.2607
85
+ [plants 0-4] step 450/600 2139 s mean best-of-16 18.3531
86
+ [plants 20-24] step 450/600 2118 s mean best-of-16 45.5763
87
+ [plants 30-34] step 450/600 2137 s mean best-of-16 15.2476
88
+ [plants 15-19] step 450/600 2132 s mean best-of-16 31.0600
89
+ [plants 5-9] step 500/600 2370 s mean best-of-16 5.8323
90
+ [plants 35-39] step 500/600 2351 s mean best-of-16 29.8195
91
+ [plants 25-29] step 500/600 2358 s mean best-of-16 29.2422
92
+ [plants 0-4] step 500/600 2360 s mean best-of-16 18.3494
93
+ [plants 10-14] step 500/600 2352 s mean best-of-16 31.2514
94
+ [plants 20-24] step 500/600 2338 s mean best-of-16 45.5603
95
+ [plants 30-34] step 500/600 2358 s mean best-of-16 15.2004
96
+ [plants 15-19] step 500/600 2350 s mean best-of-16 30.8120
97
+ [plants 5-9] step 550/600 2601 s mean best-of-16 5.8304
98
+ [plants 35-39] step 550/600 2579 s mean best-of-16 29.8161
99
+ [plants 25-29] step 550/600 2590 s mean best-of-16 29.2395
100
+ [plants 0-4] step 550/600 2592 s mean best-of-16 18.3462
101
+ [plants 10-14] step 550/600 2584 s mean best-of-16 31.2428
102
+ [plants 20-24] step 550/600 2568 s mean best-of-16 45.5806
103
+ [plants 30-34] step 550/600 2589 s mean best-of-16 15.1689
104
+ [plants 15-19] step 550/600 2580 s mean best-of-16 30.6866
105
+ [plants 5-9] step 600/600 2836 s mean best-of-16 5.8289
106
+ [plants 35-39] step 600/600 2812 s mean best-of-16 29.8131
107
+ [plants 25-29] step 600/600 2824 s mean best-of-16 29.2373
108
+ [plants 10-14] step 600/600 2817 s mean best-of-16 31.2341
109
+ [plants 20-24] step 600/600 2801 s mean best-of-16 45.5650
110
+ [plants 0-4] step 600/600 2826 s mean best-of-16 18.3434
111
+ [plants 30-34] step 600/600 2825 s mean best-of-16 15.1420
112
+ [plants 15-19] step 600/600 2811 s mean best-of-16 30.6211
113
+ oracle lr=0.1: 2973 s
results/icml_r3/X19/logs/val_oracle_lr0.1_crashed_nan.log ADDED
@@ -0,0 +1,58 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ /home/cercola/matte_server/DiffusionControl/nonlinear/data_generation/build_dataset.py:131: UserWarning: Solution may be inaccurate. Try another solver, adjusting the solver settings, or solve with verbose=True for more information.
2
+ cp.Problem(cp.Minimize(0), cons).solve(solver=solver, verbose=False)
3
+ /home/cercola/matte_server/DiffusionControl/nonlinear/data_generation/build_dataset.py:131: UserWarning: Solution may be inaccurate. Try another solver, adjusting the solver settings, or solve with verbose=True for more information.
4
+ cp.Problem(cp.Minimize(0), cons).solve(solver=solver, verbose=False)
5
+ /home/cercola/matte_server/DiffusionControl/nonlinear/data_generation/build_dataset.py:131: UserWarning: Solution may be inaccurate. Try another solver, adjusting the solver settings, or solve with verbose=True for more information.
6
+ cp.Problem(cp.Minimize(0), cons).solve(solver=solver, verbose=False)
7
+ /home/cercola/matte_server/DiffusionControl/nonlinear/data_generation/build_dataset.py:131: UserWarning: Solution may be inaccurate. Try another solver, adjusting the solver settings, or solve with verbose=True for more information.
8
+ cp.Problem(cp.Minimize(0), cons).solve(solver=solver, verbose=False)
9
+ concurrent.futures.process._RemoteTraceback:
10
+ """
11
+ Traceback (most recent call last):
12
+ File "/home/cercola/miniconda3/lib/python3.13/concurrent/futures/process.py", line 254, in _process_worker
13
+ r = call_item.fn(*call_item.args, **call_item.kwargs)
14
+ File "/home/cercola/miniconda3/lib/python3.13/concurrent/futures/process.py", line 203, in _process_chunk
15
+ return [fn(*args) for args in chunk]
16
+ ~~^^^^^^^
17
+ File "/home/cercola/matte_server/DiffusionControl/nonlinear/experiments/siso_x19.py", line 335, in oracle_chunk
18
+ bad = np.array([np.roots(np.concatenate([[1], d])).real.max() >= -cfg["f_margin"] for d in dn])
19
+ ~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^
20
+ File "/home/cercola/miniconda3/lib/python3.13/site-packages/numpy/lib/_polynomial_impl.py", line 247, in roots
21
+ roots = eigvals(A)
22
+ File "/home/cercola/miniconda3/lib/python3.13/site-packages/numpy/linalg/_linalg.py", line 1206, in eigvals
23
+ _assert_finite(a)
24
+ ~~~~~~~~~~~~~~^^^
25
+ File "/home/cercola/miniconda3/lib/python3.13/site-packages/numpy/linalg/_linalg.py", line 207, in _assert_finite
26
+ raise LinAlgError("Array must not contain infs or NaNs")
27
+ numpy.linalg.LinAlgError: Array must not contain infs or NaNs
28
+ """
29
+
30
+ The above exception was the direct cause of the following exception:
31
+
32
+ Traceback (most recent call last):
33
+ File "<frozen runpy>", line 198, in _run_module_as_main
34
+ File "<frozen runpy>", line 88, in _run_code
35
+ File "/home/cercola/matte_server/DiffusionControl/nonlinear/experiments/siso_x19.py", line 476, in <module>
36
+ print(f" oracle: {time.time() - t:.0f} s", flush=True)
37
+ ~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
38
+ File "/home/cercola/matte_server/DiffusionControl/nonlinear/experiments/siso_x19.py", line 436, in stage_val
39
+ if os.path.exists(p):
40
+
41
+ File "/home/cercola/matte_server/DiffusionControl/nonlinear/experiments/siso_x19.py", line 370, in oracle
42
+ per = per or max(1, int(np.ceil(len(gs) / nw)))
43
+ ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
44
+ File "/home/cercola/miniconda3/lib/python3.13/concurrent/futures/process.py", line 617, in _chain_from_iterable_of_lists
45
+ for element in iterable:
46
+ ^^^^^^^^
47
+ File "/home/cercola/miniconda3/lib/python3.13/concurrent/futures/_base.py", line 619, in result_iterator
48
+ yield _result_or_cancel(fs.pop())
49
+ ~~~~~~~~~~~~~~~~~^^^^^^^^^^
50
+ File "/home/cercola/miniconda3/lib/python3.13/concurrent/futures/_base.py", line 317, in _result_or_cancel
51
+ return fut.result(timeout)
52
+ ~~~~~~~~~~^^^^^^^^^
53
+ File "/home/cercola/miniconda3/lib/python3.13/concurrent/futures/_base.py", line 449, in result
54
+ return self.__get_result()
55
+ ~~~~~~~~~~~~~~~~~^^
56
+ File "/home/cercola/miniconda3/lib/python3.13/concurrent/futures/_base.py", line 401, in __get_result
57
+ raise self._exception
58
+ numpy.linalg.LinAlgError: Array must not contain infs or NaNs
results/icml_r3/X19/posthoc_labels.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4e80741c4504ac7f6b0b0618ad0634ff8a43cf8fc848e9e1038afa753fdb9713
3
+ size 6535
results/icml_r3/X19/posthoc_labels.txt ADDED
@@ -0,0 +1,11 @@
 
 
 
 
 
 
 
 
 
 
 
 
1
+ X19 POST HOC (not preregistered): training labels under the X19 cost
2
+ 200 random rows (seed 1234) of data/popov_dataset_normalized_L.pt (180,160 rows, 5,635 plants); rows with a clipped G skipped
3
+ each row's own controller (F, spline) on its own plant; same x0 convention, 60 s cost on y_G, real Popov certificate
4
+ row controller certified (corrected LMI): 99.0% [97.5, 100.0]
5
+ median J_row / J(phi=0): 0.999 [0.999, 1.000]; rows cheaper than phi=0: 73.0%
6
+ median J_row / J(best certified static gain): 1.189 [1.163, 1.269]; rows cheaper than the static gain: 0.0%
7
+ same, certified rows only (n=198): J/J0 0.999, J/Jsg 1.187
8
+ median J(static gain) / J(phi=0) on these plants: 0.837; static gain certified 100.0%
9
+ dataset objective (code): none. build_dataset.PopovControllerDataset.create_entry draws a random stable strictly proper F,
10
+ computes the Popov sector beta of L = G F, and draws the spline as phi(x_k) = x_k r_k with r_k ~ U(0, beta)
11
+ (generate_spline); no cost is evaluated or minimised. CIs: 95% percentile bootstrap over rows (10,000).
results/icml_r3/X19/test/arms.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:eae79910a385f2d92ce1323b13cbd9ad70329d0efcdbd2d3b58401f940f85896
3
+ size 2276298
results/icml_r3/X19/test/oracle.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f57bf4a9961794ab0d6df5f19d423e2d6f5c35361f6378a8d6ff586d2cd6686d
3
+ size 7210495
results/icml_r3/X19/test_report.txt ADDED
@@ -0,0 +1,79 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ === X19 SISO re-score (test, N = 150 plants), corrected Popov certificate, 60 s cost on y_G ===
2
+ bank (retrieval): 5635 distinct plants, 180160 rows; generator 99.1 ms/controller (DDPM-500, CPU)
3
+ regret = (J - J_best)/J_best x 100, J_best = min over the reference and every same-class arm/seed/b; uncertified at b -> phi = 0 (open-loop cost). CIs: 95% percentile bootstrap over plants (10,000).
4
+ stochastic arms (generator, CVAE): per-plant mean over sampler seeds 0,1,2; [per-seed values]
5
+
6
+ arm b certified median regret % mean regret % cond. med % vs ref med % J/J(phi=0)
7
+ generator (DDPM-500) 1 97.8 [ 96.4, 99.1] +21.19 [ +16.65, +23.75] +47.75 [ +35.03, +63.43] +19.59 +20.65 0.995 [99.3%/+20.0, 95.3%/+20.3, 98.7%/+19.7]
8
+ generator (DDPM-500) 2 99.6 [ 98.9,100.0] +16.08 [ +13.67, +22.03] +41.98 [ +30.24, +56.35] +16.31 +15.62 0.986 [100.0%/+16.1, 99.3%/+16.4, 99.3%/+16.4]
9
+ generator (DDPM-500) 4 100.0 [100.0,100.0] +14.85 [ +12.30, +20.33] +38.11 [ +27.58, +51.08] +15.16 +14.59 0.976 [100.0%/+15.4, 100.0%/+15.2, 100.0%/+14.7]
10
+ generator (DDPM-500) 8 100.0 [100.0,100.0] +14.29 [ +11.33, +20.02] +35.64 [ +25.52, +48.16] +13.96 +13.82 0.967 [100.0%/+14.5, 100.0%/+14.4, 100.0%/+13.7]
11
+ generator (DDPM-500) 16 100.0 [100.0,100.0] +12.87 [ +10.38, +17.46] +31.91 [ +23.20, +42.58] +12.58 +12.36 0.953 [100.0%/+12.4, 100.0%/+11.7, 100.0%/+13.6]
12
+
13
+ retrieval (nn_z, distinct) 1 99.3 [ 98.0,100.0] +20.09 [ +14.45, +24.49] +46.40 [ +34.04, +61.77] +20.03 +20.09 0.999
14
+ retrieval (nn_z, distinct) 2 100.0 [100.0,100.0] +18.71 [ +14.40, +22.90] +44.72 [ +32.68, +59.62] +18.71 +18.50 0.998
15
+ retrieval (nn_z, distinct) 4 100.0 [100.0,100.0] +18.43 [ +14.27, +22.61] +42.66 [ +31.06, +57.04] +18.43 +18.27 0.995
16
+ retrieval (nn_z, distinct) 8 100.0 [100.0,100.0] +18.25 [ +13.18, +22.04] +41.90 [ +30.26, +56.30] +18.25 +18.04 0.992
17
+ retrieval (nn_z, distinct) 16 100.0 [100.0,100.0] +16.59 [ +12.73, +21.42] +38.44 [ +27.25, +52.59] +16.59 +16.58 0.983
18
+
19
+ CVAE 1 98.4 [ 96.2,100.0] +16.63 [ +14.05, +24.18] +43.91 [ +32.89, +57.08] +16.69 +16.43 0.987 [98.0%/+18.0, 98.7%/+16.7, 98.7%/+16.6]
20
+ CVAE 2 98.7 [ 96.7,100.0] +15.72 [ +12.82, +22.95] +41.19 [ +30.67, +53.87] +15.88 +15.71 0.982 [98.7%/+15.7, 98.7%/+16.1, 98.7%/+15.7]
21
+ CVAE 4 98.7 [ 96.7,100.0] +15.47 [ +11.87, +21.68] +39.38 [ +29.18, +51.55] +15.59 +15.47 0.980 [98.7%/+15.7, 98.7%/+15.6, 98.7%/+15.5]
22
+ CVAE 8 98.7 [ 96.7,100.0] +15.28 [ +11.76, +20.48] +38.26 [ +28.19, +50.11] +15.30 +15.28 0.978 [98.7%/+15.4, 98.7%/+15.1, 98.7%/+15.3]
23
+ CVAE 16 98.7 [ 96.7,100.0] +14.98 [ +11.60, +20.39] +37.21 [ +27.36, +48.80] +15.09 +14.89 0.975 [98.7%/+14.7, 98.7%/+15.0, 98.7%/+15.3]
24
+
25
+ regressor 1 98.7 [ 96.7,100.0] +16.06 [ +13.22, +21.43] +43.01 [ +31.94, +56.24] +16.06 +16.02 0.983
26
+
27
+ phi = 0 1 100.0 [100.0,100.0] +18.92 [ +14.87, +23.73] +48.16 [ +34.89, +64.26] +18.92 +18.71 1.000
28
+
29
+ static gain (F = 1) 1 100.0 [100.0,100.0] +0.20 [ +0.10, +0.34] +1.02 [ +0.76, +1.30] +0.20 +0.20 0.849
30
+
31
+ LQG (linear class) 1 100.0 [100.0,100.0] +11.70 [ +10.00, +13.40] +14.35 [ +12.87, +15.92] +11.70 +11.16 0.943
32
+
33
+ reference (16 starts) 1 100.0 [100.0,100.0] +0.00 [ +0.00, +0.00] +0.23 [ +0.15, +0.32] +0.00 +0.00 0.842
34
+
35
+ --- paired: generator (seed-mean) minus arm, per-plant regret (pct points); Wilcoxon signed-rank, Holm over 16 ---
36
+ arm b mean diff median diff p Holm p gen better/worse/tie
37
+ retrieval (nn_z, distinct) 1 +1.34 [ -1.54, +4.39] -0.06 [ -0.29, +0.37] 0.54 0.54 78/72/0
38
+ retrieval (nn_z, distinct) 4 -4.55 [ -7.50, -1.85] -1.60 [ -2.30, -1.23] 9e-13 7.2e-12 116/34/0
39
+ retrieval (nn_z, distinct) 16 -6.53 [ -10.83, -2.83] -2.11 [ -2.64, -1.38] 1.8e-10 1.3e-09 121/29/0
40
+ CVAE 1 +3.84 [ +0.57, +7.69] +1.04 [ +0.61, +1.81] 4.6e-05 0.00018 47/103/0
41
+ CVAE 4 -1.27 [ -3.38, +0.90] -0.41 [ -0.92, -0.11] 0.012 0.037 90/60/0
42
+ CVAE 16 -5.30 [ -7.60, -3.23] -1.80 [ -2.24, -1.31] 7.2e-13 6.5e-12 115/35/0
43
+ regressor 1 +4.73 [ +1.15, +8.65] +1.47 [ +0.97, +2.34] 1.3e-08 6.4e-08 41/109/0
44
+ phi = 0 1 -0.41 [ -2.44, +1.50] -0.62 [ -1.01, -0.23] 0.094 0.19 97/53/0
45
+ phi = 0 4 -10.05 [ -13.73, -6.89] -2.91 [ -3.52, -2.44] 7.8e-25 8.6e-24 145/5/0
46
+ phi = 0 16 -16.25 [ -21.95, -11.40] -5.38 [ -6.38, -3.93] 2.3e-26 3.7e-25 149/1/0
47
+ static gain (F = 1) 1 +46.73 [ +34.09, +62.23] +20.81 [ +16.28, +22.80] 2.3e-26 3.7e-25 0/150/0
48
+ static gain (F = 1) 4 +37.09 [ +26.67, +50.00] +14.65 [ +11.82, +19.81] 2.3e-26 3.7e-25 0/150/0
49
+ static gain (F = 1) 16 +30.90 [ +22.28, +41.50] +12.40 [ +9.74, +16.40] 2.3e-26 3.7e-25 0/150/0
50
+ LQG (linear class) 1 +33.40 [ +21.66, +48.22] +8.52 [ +6.02, +11.00] 6.3e-26 7.5e-25 3/147/0
51
+ LQG (linear class) 4 +23.76 [ +14.20, +35.84] +3.00 [ +2.01, +6.30] 1.2e-20 1.2e-19 20/130/0
52
+ LQG (linear class) 16 +17.57 [ +9.80, +27.55] +1.39 [ +0.26, +3.20] 4.4e-09 2.6e-08 58/92/0
53
+
54
+ --- paired certified-rate difference, generator (seed-mean) minus arm (pct points); exact McNemar per sampler seed ---
55
+ retrieval (nn_z, distinct) 1 -1.6 [ -3.3, +0.4] McNemar (gen-only, arm-only, p): (1,1,1); (1,7,0.07); (1,2,1)
56
+ retrieval (nn_z, distinct) 4 +0.0 [ +0.0, +0.0] McNemar (gen-only, arm-only, p): (0,0,1); (0,0,1); (0,0,1)
57
+ retrieval (nn_z, distinct) 16 +0.0 [ +0.0, +0.0] McNemar (gen-only, arm-only, p): (0,0,1); (0,0,1); (0,0,1)
58
+ CVAE 1 -0.7 [ -2.2, +1.1] McNemar (gen-only, arm-only, p): (3,1,0.62); (0,5,0.062); (2,2,1)
59
+ CVAE 4 +1.3 [ +0.0, +3.3] McNemar (gen-only, arm-only, p): (2,0,0.5); (2,0,0.5); (2,0,0.5)
60
+ CVAE 16 +1.3 [ +0.0, +3.3] McNemar (gen-only, arm-only, p): (2,0,0.5); (2,0,0.5); (2,0,0.5)
61
+ regressor 1 -0.9 [ -2.7, +0.9] McNemar (gen-only, arm-only, p): (2,1,1); (0,5,0.062); (2,2,1)
62
+
63
+ --- reference (arm-independent, dataset-prior cold starts) ---
64
+ config: {"starts": 16, "steps": 400, "ckpt_every": 25, "lr": 0.1, "bcap": 10000.0, "bisect_iters": 10, "dt": 0.03, "f_margin": 0.01}
65
+ final iterates certified: 98.9% of 2400 starts; selected start used a fallback snapshot on 2.0% of plants; plants with a certified reference: 100% (start certified by construction)
66
+ reference is the best known on 64.0% of plants; reference regret median +0.00%, max +3.70%
67
+ best-of-16 vs best-of-8 / -4 / -1 starts (median % above best-of-16): 0.00 / 0.29 / 1.94
68
+ F-stability rejections: 69365; LMI calls in the fallback walk: 2602; wall 3080 s (20.5 s/plant on <= 12 CPU workers)
69
+ X9a re-check (certified): plants with >= 2 distinct certified reference end points within +5% of the best: 100.0% [100.0, 100.0] (bootstrap), mean count 10.78 (16 starts)
70
+ LQG (linear class): closed-loop Hurwitz on 100.0% of plants; cheaper than J_best on 0.0% of plants; median J_lqg/J_best 1.117
71
+ static gain: median k* 0.50, median certified range beta_G 100.00 (cap 100 reached on 86.7%), 19 verifier calls/plant; certified 100.0%
72
+ horizon check |x(60 s)| (mean over ICs) of the chosen b=16 controller: generator median 1.9e-07, >1e-2 on 0.9%; reference median 9.5e-04, >1e-2 on 4.7%
73
+ horizon tail: 101 chosen controllers with |x(60 s)| > 1e-3 (71 reference, 30 generator); J(600 s) exceeds J(60 s) by max 0.546% (median 0.0003%); max |x(600 s)| 6.4e-04
74
+
75
+ --- what changed: old criterion (12 s rollout, final energy < 0.3, loop-state x0) vs certified, same candidates ---
76
+ arm b=1 b=2 b=4 b=8 b=16 (old / certified, %)
77
+ generator (DDPM-500) 57.6 / 97.8 67.6 / 99.6 71.1 / 100.0 72.4 / 100.0 74.0 / 100.0
78
+ retrieval (nn_z, distinct) 46.0 / 99.3 60.0 / 100.0 68.7 / 100.0 69.3 / 100.0 72.0 / 100.0
79
+ generator per candidate: certified 96.7%, old-criterion pass 57.5%, certified but failing the old test 39.4%, passing it but not certified 0.1%
results/icml_r3/X19/test_stats.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4a67464de33f2aa2f820cd0d849d6e0991468f5036c83c8a75e6d8307cb50f01
3
+ size 150458
results/icml_r3/X19/val/arms.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:02c03037af375d08b7b7224f71f0a3c41d878524ac1d5b07771d14ffcc392f41
3
+ size 624706
results/icml_r3/X19/val/oracle_lr0.02.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e4517fc431ceaba898051b8595d884812a52c068823d139964b897f19911f957
3
+ size 2983648
results/icml_r3/X19/val/oracle_lr0.05.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5e1d93dd41499bd2241755a40ad20f7b7f57d0e092abd97a95f8d79e9b7841ba
3
+ size 2888521
results/icml_r3/X19/val/oracle_lr0.1.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:41712a71cc5a5605ce7642325ba270308c1ca9fd0ac13ec99327e5758a055498
3
+ size 2722173
results/icml_r3/X19/val_report.txt ADDED
@@ -0,0 +1,79 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ === X19 SISO re-score (val, N = 40 plants), corrected Popov certificate, 60 s cost on y_G ===
2
+ bank (retrieval): 5635 distinct plants, 180160 rows; generator 29.3 ms/controller (DDPM-500, CPU)
3
+ regret = (J - J_best)/J_best x 100, J_best = min over the reference and every same-class arm/seed/b; uncertified at b -> phi = 0 (open-loop cost). CIs: 95% percentile bootstrap over plants (10,000).
4
+ stochastic arms (generator, CVAE): per-plant mean over sampler seeds 0,1,2; [per-seed values]
5
+
6
+ arm b certified median regret % mean regret % cond. med % vs ref med % J/J(phi=0)
7
+ generator (DDPM-500) 1 96.7 [ 93.3, 99.2] +15.97 [ +11.77, +20.94] +32.43 [ +19.83, +47.66] +13.89 +15.96 0.993 [100.0%/+13.9, 95.0%/+13.6, 95.0%/+16.5]
8
+ generator (DDPM-500) 2 99.2 [ 97.5,100.0] +14.15 [ +8.88, +19.28] +28.73 [ +17.68, +41.86] +12.45 +14.15 0.986 [100.0%/+11.7, 97.5%/+12.9, 100.0%/+12.1]
9
+ generator (DDPM-500) 4 100.0 [100.0,100.0] +12.31 [ +7.19, +15.95] +24.89 [ +15.17, +36.39] +11.47 +12.27 0.978 [100.0%/+11.7, 100.0%/+11.9, 100.0%/+11.3]
10
+ generator (DDPM-500) 8 100.0 [100.0,100.0] +10.10 [ +6.29, +15.76] +21.75 [ +13.53, +31.42] +9.46 +10.06 0.967 [100.0%/+10.5, 100.0%/+9.1, 100.0%/+9.6]
11
+ generator (DDPM-500) 16 100.0 [100.0,100.0] +9.01 [ +5.87, +13.78] +19.56 [ +12.17, +28.36] +8.53 +8.97 0.958 [100.0%/+8.7, 100.0%/+8.7, 100.0%/+8.3]
12
+
13
+ retrieval (nn_z, distinct) 1 100.0 [100.0,100.0] +13.50 [ +9.01, +24.20] +30.42 [ +18.73, +44.48] +13.50 +13.46 0.999
14
+ retrieval (nn_z, distinct) 2 100.0 [100.0,100.0] +12.08 [ +8.69, +18.04] +28.72 [ +17.20, +42.61] +12.08 +12.08 0.998
15
+ retrieval (nn_z, distinct) 4 100.0 [100.0,100.0] +11.09 [ +8.59, +17.96] +23.98 [ +15.84, +33.36] +11.09 +11.09 0.992
16
+ retrieval (nn_z, distinct) 8 100.0 [100.0,100.0] +10.59 [ +8.43, +17.90] +23.44 [ +15.32, +32.84] +10.59 +10.59 0.987
17
+ retrieval (nn_z, distinct) 16 100.0 [100.0,100.0] +10.59 [ +8.30, +17.25] +22.80 [ +14.76, +32.15] +10.59 +10.59 0.983
18
+
19
+ CVAE 1 100.0 [100.0,100.0] +13.89 [ +10.15, +20.96] +28.47 [ +18.14, +40.65] +14.87 +13.85 0.990 [100.0%/+15.7, 100.0%/+12.9, 100.0%/+14.1]
20
+ CVAE 2 100.0 [100.0,100.0] +13.16 [ +9.53, +19.96] +25.97 [ +16.82, +36.78] +13.06 +13.11 0.987 [100.0%/+14.8, 100.0%/+12.9, 100.0%/+12.7]
21
+ CVAE 4 100.0 [100.0,100.0] +12.90 [ +8.50, +19.81] +24.94 [ +15.99, +35.47] +12.85 +12.74 0.986 [100.0%/+13.0, 100.0%/+12.8, 100.0%/+12.6]
22
+ CVAE 8 100.0 [100.0,100.0] +12.54 [ +8.42, +19.27] +23.85 [ +15.33, +33.93] +12.51 +12.50 0.985 [100.0%/+12.7, 100.0%/+12.4, 100.0%/+12.5]
23
+ CVAE 16 100.0 [100.0,100.0] +12.54 [ +8.37, +18.59] +23.19 [ +14.88, +32.97] +12.51 +12.50 0.984 [100.0%/+12.7, 100.0%/+12.4, 100.0%/+12.4]
24
+
25
+ regressor 1 100.0 [100.0,100.0] +14.04 [ +9.44, +20.92] +27.76 [ +17.55, +39.90] +14.04 +14.04 0.987
26
+
27
+ phi = 0 1 100.0 [100.0,100.0] +13.98 [ +9.19, +21.65] +31.93 [ +19.15, +47.29] +13.98 +13.94 1.000
28
+
29
+ static gain (F = 1) 1 100.0 [100.0,100.0] +0.51 [ +0.11, +0.79] +1.03 [ +0.56, +1.56] +0.51 +0.51 0.879
30
+
31
+ LQG (linear class) 1 100.0 [100.0,100.0] +9.44 [ +7.15, +12.50] +12.02 [ +9.73, +14.61] +9.44 +9.40 0.971
32
+
33
+ reference (16 starts) 1 100.0 [100.0,100.0] +0.00 [ +0.00, +0.00] +0.11 [ +0.04, +0.21] +0.00 +0.00 0.878
34
+
35
+ --- paired: generator (seed-mean) minus arm, per-plant regret (pct points); Wilcoxon signed-rank, Holm over 16 ---
36
+ arm b mean diff median diff p Holm p gen better/worse/tie
37
+ retrieval (nn_z, distinct) 1 +2.01 [ -1.18, +6.44] -0.64 [ -1.34, +0.72] 0.96 1 25/15/0
38
+ retrieval (nn_z, distinct) 4 +0.92 [ -2.37, +5.25] -1.14 [ -2.12, -0.57] 0.05 0.3 28/12/0
39
+ retrieval (nn_z, distinct) 16 -3.24 [ -6.26, -0.81] -2.11 [ -3.58, -1.20] 0.00045 0.0036 32/8/0
40
+ CVAE 1 +3.96 [ +0.58, +8.27] +0.76 [ -0.76, +1.99] 0.15 0.62 17/23/0
41
+ CVAE 4 -0.05 [ -1.97, +2.55] -0.91 [ -1.54, -0.19] 0.062 0.31 27/13/0
42
+ CVAE 16 -3.62 [ -6.64, -1.18] -2.34 [ -2.92, -1.25] 0.00025 0.0022 30/10/0
43
+ regressor 1 +4.67 [ +1.36, +8.77] +0.76 [ +0.19, +2.48] 0.016 0.11 12/28/0
44
+ phi = 0 1 +0.50 [ -1.28, +2.31] -0.77 [ -1.82, +1.06] 0.85 1 25/15/0
45
+ phi = 0 4 -7.04 [ -11.84, -3.56] -2.42 [ -4.44, -2.00] 6e-11 7.2e-10 39/1/0
46
+ phi = 0 16 -12.37 [ -19.23, -6.80] -4.80 [ -6.94, -3.22] 1.8e-12 2.9e-11 40/0/0
47
+ static gain (F = 1) 1 +31.40 [ +19.04, +46.34] +15.30 [ +11.25, +20.94] 1.8e-12 2.9e-11 0/40/0
48
+ static gain (F = 1) 4 +23.87 [ +14.39, +35.15] +11.96 [ +6.88, +15.83] 1.8e-12 2.9e-11 0/40/0
49
+ static gain (F = 1) 16 +18.54 [ +11.35, +26.99] +8.58 [ +5.50, +11.72] 1.8e-12 2.9e-11 0/40/0
50
+ LQG (linear class) 1 +20.41 [ +9.75, +33.46] +6.07 [ +3.97, +10.09] 5.6e-10 6.1e-09 4/36/0
51
+ LQG (linear class) 4 +12.88 [ +5.31, +22.01] +2.22 [ +0.69, +4.08] 1.3e-05 0.00013 9/31/0
52
+ LQG (linear class) 16 +7.54 [ +2.24, +14.04] -0.12 [ -1.07, +1.40] 0.25 0.74 21/19/0
53
+
54
+ --- paired certified-rate difference, generator (seed-mean) minus arm (pct points); exact McNemar per sampler seed ---
55
+ retrieval (nn_z, distinct) 1 -3.3 [ -6.7, -0.8] McNemar (gen-only, arm-only, p): (0,0,1); (0,2,0.5); (0,2,0.5)
56
+ retrieval (nn_z, distinct) 4 +0.0 [ +0.0, +0.0] McNemar (gen-only, arm-only, p): (0,0,1); (0,0,1); (0,0,1)
57
+ retrieval (nn_z, distinct) 16 +0.0 [ +0.0, +0.0] McNemar (gen-only, arm-only, p): (0,0,1); (0,0,1); (0,0,1)
58
+ CVAE 1 -3.3 [ -6.7, -0.8] McNemar (gen-only, arm-only, p): (0,0,1); (0,2,0.5); (0,2,0.5)
59
+ CVAE 4 +0.0 [ +0.0, +0.0] McNemar (gen-only, arm-only, p): (0,0,1); (0,0,1); (0,0,1)
60
+ CVAE 16 +0.0 [ +0.0, +0.0] McNemar (gen-only, arm-only, p): (0,0,1); (0,0,1); (0,0,1)
61
+ regressor 1 -3.3 [ -6.7, -0.8] McNemar (gen-only, arm-only, p): (0,0,1); (0,2,0.5); (0,2,0.5)
62
+
63
+ --- reference (arm-independent, dataset-prior cold starts) ---
64
+ config: {"starts": 16, "steps": 600, "ckpt_every": 25, "lr": 0.1, "bcap": 10000.0, "bisect_iters": 10, "dt": 0.03, "f_margin": 0.01}
65
+ final iterates certified: 99.2% of 640 starts; selected start used a fallback snapshot on 0.0% of plants; plants with a certified reference: 100% (start certified by construction)
66
+ reference is the best known on 75.0% of plants; reference regret median +0.00%, max +1.35%
67
+ best-of-16 vs best-of-8 / -4 / -1 starts (median % above best-of-16): 0.04 / 0.26 / 1.56
68
+ F-stability rejections: 40257; LMI calls in the fallback walk: 736; wall 2973 s (74.3 s/plant on <= 12 CPU workers)
69
+ X9a re-check (certified): plants with >= 2 distinct certified reference end points within +5% of the best: 100.0% [100.0, 100.0] (bootstrap), mean count 11.28 (16 starts)
70
+ LQG (linear class): closed-loop Hurwitz on 100.0% of plants; cheaper than J_best on 0.0% of plants; median J_lqg/J_best 1.094
71
+ static gain: median k* 0.45, median certified range beta_G 100.00 (cap 100 reached on 85.0%), 19 verifier calls/plant; certified 100.0%
72
+ horizon check |x(60 s)| (mean over ICs) of the chosen b=16 controller: generator median 2.9e-08, >1e-2 on 0.0%; reference median 9.2e-04, >1e-2 on 5.0%
73
+ horizon tail: 28 chosen controllers with |x(60 s)| > 1e-3 (19 reference, 9 generator); J(600 s) exceeds J(60 s) by max 0.034% (median 0.0000%); max |x(600 s)| 5.2e-05
74
+
75
+ --- what changed: old criterion (12 s rollout, final energy < 0.3, loop-state x0) vs certified, same candidates ---
76
+ arm b=1 b=2 b=4 b=8 b=16 (old / certified, %)
77
+ generator (DDPM-500) 66.7 / 96.7 75.8 / 99.2 80.0 / 100.0 80.0 / 100.0 80.8 / 100.0
78
+ retrieval (nn_z, distinct) 52.5 / 100.0 67.5 / 100.0 77.5 / 100.0 80.0 / 100.0 80.0 / 100.0
79
+ generator per candidate: certified 96.1%, old-criterion pass 67.7%, certified but failing the old test 28.7%, passing it but not certified 0.2%
results/icml_r3/X19/val_selection.txt ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ X19 validation: reference lr / step selection (40 plants, 16 starts)
2
+ lr 0.02: mean log best-of-16 at step 600 2.6493; plant-mean best-of-16 at steps 1/100/200/300/400/500/600: 35.147 / 26.470 / 26.197 / 26.060 / 25.961 / 25.898 / 25.863; final certified 99.4%, evaluated best-of-16 J mean 25.863; wall 3882 s
3
+ lr 0.05: mean log best-of-16 at step 600 2.6453; plant-mean best-of-16 at steps 1/100/200/300/400/500/600: 35.147 / 26.165 / 25.968 / 25.869 / 25.828 / 25.804 / 25.783; final certified 99.1%, evaluated best-of-16 J mean 25.783; wall 3938 s
4
+ lr 0.1: mean log best-of-16 at step 600 2.6433; plant-mean best-of-16 at steps 1/100/200/300/400/500/600: 35.147 / 26.070 / 25.926 / 25.879 / 25.846 / 25.758 / 25.723; final certified 99.2%, evaluated best-of-16 J mean 25.723; wall 2973 s
5
+ selected lr = 0.1, steps = 400 (first multiple of 100 within 0.5% of step 600); test wall estimate 43 min
results/icml_r3/X19/val_stats.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:88f77b89053fcc69be85697bfc88281b354bf3258f3f17adc4c5cb50c394a3b5
3
+ size 57714