matCercola18 commited on
Commit
43352e9
·
verified ·
1 Parent(s): 9e5e7dc

X28 SISO S1: gate, control, diagnostics, symmetric-knot gate; MIMO self-checks and validation runs; PREREG with addenda

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. results/icml_r3/X28/PREREG.md +172 -0
  2. results/icml_r3/X28/control/cfg.json +10 -0
  3. results/icml_r3/X28/control/plant_00.npz +3 -0
  4. results/icml_r3/X28/control/plant_01.npz +3 -0
  5. results/icml_r3/X28/control/plant_02.npz +3 -0
  6. results/icml_r3/X28/control/plant_03.npz +3 -0
  7. results/icml_r3/X28/control/plant_04.npz +3 -0
  8. results/icml_r3/X28/control/plant_05.npz +3 -0
  9. results/icml_r3/X28/control/plant_06.npz +3 -0
  10. results/icml_r3/X28/control/plant_07.npz +3 -0
  11. results/icml_r3/X28/control/plant_08.npz +3 -0
  12. results/icml_r3/X28/control/plant_09.npz +3 -0
  13. results/icml_r3/X28/control/plant_10.npz +3 -0
  14. results/icml_r3/X28/control/plant_11.npz +3 -0
  15. results/icml_r3/X28/control/plant_12.npz +3 -0
  16. results/icml_r3/X28/control/plant_13.npz +3 -0
  17. results/icml_r3/X28/control/plant_14.npz +3 -0
  18. results/icml_r3/X28/control/plant_15.npz +3 -0
  19. results/icml_r3/X28/control/plant_16.npz +3 -0
  20. results/icml_r3/X28/control/plant_17.npz +3 -0
  21. results/icml_r3/X28/control/plant_18.npz +3 -0
  22. results/icml_r3/X28/control/plant_19.npz +3 -0
  23. results/icml_r3/X28/control/plant_20.npz +3 -0
  24. results/icml_r3/X28/control/plant_21.npz +3 -0
  25. results/icml_r3/X28/control/plant_22.npz +3 -0
  26. results/icml_r3/X28/control/plant_23.npz +3 -0
  27. results/icml_r3/X28/control/plant_24.npz +3 -0
  28. results/icml_r3/X28/control/plant_25.npz +3 -0
  29. results/icml_r3/X28/control/plant_26.npz +3 -0
  30. results/icml_r3/X28/control/plant_27.npz +3 -0
  31. results/icml_r3/X28/control/plant_28.npz +3 -0
  32. results/icml_r3/X28/control/plant_29.npz +3 -0
  33. results/icml_r3/X28/control/plant_30.npz +3 -0
  34. results/icml_r3/X28/control/plant_31.npz +3 -0
  35. results/icml_r3/X28/control/plant_32.npz +3 -0
  36. results/icml_r3/X28/control/plant_33.npz +3 -0
  37. results/icml_r3/X28/control/plant_34.npz +3 -0
  38. results/icml_r3/X28/control/plant_35.npz +3 -0
  39. results/icml_r3/X28/control/plant_36.npz +3 -0
  40. results/icml_r3/X28/control/plant_37.npz +3 -0
  41. results/icml_r3/X28/control/plant_38.npz +3 -0
  42. results/icml_r3/X28/control/plant_39.npz +3 -0
  43. results/icml_r3/X28/control_report.txt +60 -0
  44. results/icml_r3/X28/diag.npz +3 -0
  45. results/icml_r3/X28/diag_report.txt +6 -0
  46. results/icml_r3/X28/gate/cfg.json +10 -0
  47. results/icml_r3/X28/gate/plant_00.npz +3 -0
  48. results/icml_r3/X28/gate/plant_01.npz +3 -0
  49. results/icml_r3/X28/gate/plant_02.npz +3 -0
  50. results/icml_r3/X28/gate/plant_03.npz +3 -0
results/icml_r3/X28/PREREG.md ADDED
@@ -0,0 +1,172 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # X28 pre-registration: generate the spline nonlinearity on SISO and MIMO
2
+
3
+ Written 2026-11-05, before any X28 code ran or any X28 number existed.
4
+ Decision and setup: Matteo, 2026-11-05 (`docs/HANDOFF_2026-11-05_movehpc.md` section 1). He confirmed the task and the July-faithful MIMO definition directly in the move-HPC session.
5
+
6
+ ## 0. Question
7
+
8
+ Does a diffusion model that generates the nonlinearity phi (11-knot spline) together with the linear part beat the October setting, where phi is fixed (SISO, X21) or an unknown actuator (MIMO, X27)? Either answer is a result. If the spline loses, the measured loss documents why the setting changed.
9
+
10
+ ## 1. SISO (X28-S)
11
+
12
+ Everything is as X21 (`docs/icml_r3_x21.md` section 2) except the generated object: plants (train seed 8181, 1971 kept; validation = 40 X20 pilot plants, seed 7171; test = 199 plants, seed 8282), loop u = -phi(clip(F y, -5, 5)), |phi| <= 1, regional circle certificate `rstar` certified at s1 = 0.5 sigma_p, F stable with margin 0.01, cost J = mean over 4 directions of the integral over 60 s of q(y) + u^2.
13
+
14
+ ### 1.1 Spline class fix (new function, X20/X21 code untouched)
15
+
16
+ `siso_x20_unstable.ref_optimize` uses phi = clamp(phi0 exp(ts), -1, 1) from phi0 = sat_1, which has two defects:
17
+ - A knot pushed above 1 once has zero gradient from then on, so it is dead even if the gradient later points down.
18
+ - Every knot keeps its sign, and every nonzero knot starts on the bound.
19
+
20
+ The handoff suggested phi_k = tanh(raw_k). It cannot represent sat_1 exactly (tanh never reaches 1), and near +-1 its gradient vanishes, so it does not meet the requirement that the joint optimum contains the F-only optimum. X28 uses projected Adam on the knots directly: phi_k free in [-1, 1], phi(0) = 0, projected after each step. The forward pass has no clamp, so gradients stay alive at the bound and knots move both ways.
21
+
22
+ Gauge starts: F -> c F with phi = sat(y / c) is exactly the same controller. For c = 2 and 5 this puts knots in the interior, so the optimizer can shape the knee at finer resolution than c = 1 allows. The X20b class never had this.
23
+
24
+ ### 1.2 S1, label gate (CPU node, 40 validation plants, s1)
25
+
26
+ Per plant, joint Adam on (F, knots), 400 steps, snapshots every 25, `results/icml_r3/X19/FROZEN.json` otherwise:
27
+ - Warm arm W: from the plant's F-only optimum (X20b `fonly_num`, `fonly_den`) at c in {1, 2, 5}; F lr 0.02, knot lr 0.02.
28
+ - Cold arm H: from the 16 HIFOO starts of `ref_starts` (seed (2929, i, 0)); F lr 0.1 (frozen X19 value), knot lr 0.02.
29
+
30
+ Selection per start: the cheapest snapshot certified at s1. J_joint = min(all W and H starts, J_Fonly), so J_joint <= J_Fonly by construction.
31
+
32
+ Check before reading results: recomputed J at the c = 1 start must equal the stored `fonly_J` within 1e-6 relative.
33
+
34
+ **Rule.** d = (J_Fonly - J_joint) / J_joint per plant. "The spline pays at the optimum" iff median d >= 2% and the 95% bootstrap CI lower bound (10,000 resamples over plants) > 0. Same threshold as X20b.
35
+
36
+ Report: median d with CI, W and H separately, per stratum, number of plants with d >= 2%, knot distance from the equivalent sat at the best start. The gate sets expectations and explains later results. It does not stop S2 to S4 (Matteo's instruction).
37
+
38
+ ### 1.3 S2, labels
39
+
40
+ Joint labels on the 1971 training plants and the test plants, warm-started from X21 F-only labels. The exact starts and steps are fixed by addendum after S1, using S1's W versus H result and validation plants only.
41
+
42
+ ### 1.4 S3, generators
43
+
44
+ X21 recipe (`train_x21.py`, DDPM-500, MLP, EMA 0.999, cost-slot conditioning, CFG). clS: cl encoding of F (9 numbers) + 10 free knots, the twin of X21 clL. coefS: slog(theta) + 10 knots. 3 seeds for clS, 1 for coefS at least. Sampler clip range chosen on validation only, as X21.
45
+
46
+ ### 1.5 S4, evaluation (199 test plants, used once)
47
+
48
+ `x21_eval.py` extended to decode phi. Reference J* = min(all F-only and joint labels) per plant. Budgets K = 16, 64; b = 1, 4, 16, 64.
49
+
50
+ **Primary comparison:** clS versus X21 clL, b = 16, each at its validation-chosen clip, 3 seeds, paired by plant, regret against the same J*.
51
+ - "Spline better": median paired regret difference (clL - clS) >= 1 point, Holm-corrected Wilcoxon p < 0.05, and certified rate not lower by more than 2 points (McNemar reported).
52
+ - "Spline worse": the same with clS and clL swapped.
53
+ - Otherwise "no measurable difference".
54
+
55
+ Also reported: regret against J_Fonly (continuity with X21), coefS, and the distance of generated knots from sat at the nearest gauge c, so "the model learned to output sat" is visible.
56
+
57
+ ## 2. MIMO (X28-M)
58
+
59
+ ### 2.1 Definition (Matteo's choice 2026-11-05: July-faithful)
60
+
61
+ Plants: `family_sofm.py`, m = 2 and 3 first, m = 5 only if cheap. Banks and held-out sets as X27.
62
+
63
+ Controller: u = -Phi(K y), Phi = diag(psi_i). Each psi_i is piecewise linear on the 11 knots {-5, ..., 5} with psi_i(0) = 0.
64
+
65
+ **Ray extension (soundness, differs from July).** July clipped y to [-5, 5]. Then psi(y)/y -> 0 as |y| grows, so no global sector with alpha_i > 0 exists, and the plants here are unstable. X28 extends each psi_i past +-5 along the ray through the origin: psi(y) = y psi(+-5)/5. It is continuous, and the sector is exact from the knots: alpha_i = min_k psi_i(x_k)/x_k, beta_i = max_k psi_i(x_k)/x_k. Unbounded actuator, global certificate, as in X27's family.
66
+
67
+ Spline class: alpha_i >= 0.1 and beta_i <= 10 at every knot (projection after each step). Identity is inside. K and the knot scale share a gauge; any normalization for the generator is fixed by addendum before training.
68
+
69
+ Certificate: circle LMI with per-channel sectors. A_K = A - B diag(alpha) K C, off-diagonal block C'K' diag(beta - alpha) T - P B, block (2,2) -2T, T diagonal, CLARABEL, eps = 1e-4 and the numerical re-check of `family_sofm.verify`. Self-check before any label: at alpha_i = 0.4, beta_i = 2.2 it must agree with `verify` on the X27 bank gains.
70
+
71
+ Cost: mean over the 2m unit coordinate initial directions of the integral over 20 s of sum_i q(y_i) + sum_i u_i^2, with q(y) = y^2 (1 + y^2 / (1 + y^2)) (`nonlinear/costs/mimo.py:68`), RK4 dt = 0.01. Rollouts with final |x| > 1e-2 are flagged in the report.
72
+
73
+ ### 2.2 Arms
74
+
75
+ Same plants, same cost:
76
+ - (a) generated K with phi = identity. Certificate: the LMI at alpha = beta = 1.
77
+ - (b) generated (K, splines). Certificate: section 2.1.
78
+ - Per-plant optimizers of both, as references.
79
+ - X27's robust-K setting as context only. It solves a different problem, so no head-to-head.
80
+
81
+ ### 2.3 Gate and labels
82
+
83
+ Per-plant optimizer: GPU-batched Adam through the rollout, warm-started from the bank's best K with phi = identity, with gauge starts as in 1.1 where the sector class allows, keeping the cheapest certified snapshot. J_spline includes the phi = identity optimum, so J_spline <= J_id by construction.
84
+
85
+ **Gate rule.** d = (J_id - J_spline) / J_spline on the held-out plants. "The spline pays at the optimum" iff median d >= 2% and CI lower bound > 0. Does not stop the generator (Matteo's instruction).
86
+
87
+ ### 2.4 Generators and evaluation
88
+
89
+ As X27's trainer and DDPM-500, 3 seeds each for (a) and (b), trained on labels from the same bank plants.
90
+
91
+ **Primary comparison:** (b) versus (a), best of 16 certified, paired by plant, regret against J* = min(all labels), with the same "better / worse / no measurable difference" rule as 1.5.
92
+
93
+ Details fixed only after this file (steps, label counts, batch sizes, gauge) go in dated addenda before the affected result exists.
94
+
95
+ ## 3. Compute and reporting
96
+
97
+ CPU jobs on `move-hpc-01-cpu`; at most 2 GPUs at once. Code on `icml-experiments`. Results to HF `matCercola18/DiffusionControl` by targeted upload of `results/icml_r3/X28/`: `utils/hf_backup.py` from a worktree holding the re-timed X7 oracles would overwrite the virgo timings at the X7 paths. Write-up in `docs/icml_r3_x28.md`. No paper edits until Matteo agrees the claim.
98
+
99
+ ## Addendum 2026-11-05: deviations approved
100
+
101
+ While the S1 gate was running and before any gate result was read, Matteo approved the two departures from the handoff draft: projected knots with gauge starts instead of tanh(raw) (section 1.1), and the ray extension past +-5 with knot ratios in [0.1, 10] instead of July's clip (section 2.1). No rule or threshold changed.
102
+
103
+ ## Addendum 2026-11-05: control arm for the S1 gate
104
+
105
+ Written after reading the gate result and before running anything below. The gate rule in 1.2 passed as written (median d = 2.84%, CI [1.06, 5.61], `gate_report.txt`). That decision stays on record and the rule is not changed.
106
+
107
+ The rule has a flaw I did not see when writing it: its baseline J_Fonly came from a different optimizer run than J_joint. The warm arm W gives F another 400 steps at a smaller learning rate, and selection takes the cheapest certified snapshot where the X20b label took the last certified one. So d mixes the value of the spline with the value of more or better optimization of F. Two facts from the gate data make this concrete: on plants 31 and 38 (d = 74.68% and 28.72%) the winning F-only label is an unmoved start (snapshot 0), and the winning spline solutions move only the knots at +-1, by a median of 0.11.
108
+
109
+ Control, same code path and budget as the gate (`spline_optimize` with knot learning rate 0, so knots stay at sat_1):
110
+ - W0: from the F-only label at c = 1, F lr 0.02. Gauges c = 2, 5 are skipped because with frozen knots and the multiplicative F parametrization they are the same trajectory.
111
+ - H0: the same 16 starts as H, seed (2929, i, 0), F lr 0.1.
112
+ - Selection: cheapest snapshot certified at s1, as in the gate.
113
+
114
+ Both arms run on all 40 validation plants whatever the first one shows.
115
+
116
+ Reported number: e = (J_F - J_J) / J_J, signed, with J_F = min(W0, H0, J_Fonly) and J_J = min(W, H, J_Fonly). Same-optimizer comparisons reported too: W against W0 and H against H0.
117
+
118
+ Attribution rule: the gate pass is attributed to the spline iff median e >= 2% and the 95% bootstrap CI lower bound (10,000 resamples over plants) is > 0. Otherwise the result is stated as "passes 1.2 as written, gain not attributable to the spline". This does not stop S2 to S4, but S2 and S3 are decided with it in hand (next addendum).
119
+
120
+ Also reported, at each plant's best joint controller: J and the certificate with its knots replaced by sat_1 (F kept), to separate what F moved from what the knots moved. And J at dt 0.01 for the same controllers, against dt 0.03, to rule out a discretization artifact.
121
+
122
+ ## Addendum 2026-11-05: MIMO optimizer details (section 2.3)
123
+
124
+ Written before any MIMO optimizer result exists. Checks so far: self-check of section 2.1 passes on m = 2 and 3 (200 gains each, half perturbed, no disagreement with `family_sofm.verify`), and J is gauge invariant to machine precision. Code: `nonlinear/mimo_unstable/x28_mimo.py`.
125
+
126
+ Lesson from the SISO gate, applied here from the start: both arms use the same K start, the same optimizer and the same budget, so that d does not mix the spline with extra optimization.
127
+
128
+ - Start: the plant's best bank K. `id` arm at c = 1, knots frozen at identity. `sp` arm at gauges c = 1, 2, 5 (K / c with all knot ratios = c, the same controller), knots free. Identity in `id` is one start because with frozen knots the gauges coincide.
129
+ - Parametrization: K = K_start + s theta, s = max(1, max|K_start|) per plant; knot ratio r_ik = psi_i(x_k) / x_k = exp(z_ik), z projected to [log 0.1, log 10] after each step, no clamp in the forward pass.
130
+ - Adam, 400 steps, snapshot every 25 plus snapshot 0, lr 0.02 on z. Cheapest certified snapshot per start. J_id is the `id` arm's; J_spline = min(`sp` starts, J_id).
131
+ - K learning rate: chosen from {0.002, 0.005, 0.02} on validation plants only (the first 64 bank plants, which the held-out plants never touch), as the value with the lowest median certified J_spline. The gate uses the same value for both arms.
132
+ - Plants: m = 3 bank `D2_scale/m3_bank.npz`; m = 2 bank `C_mimo/sof/bank.npz` (the 12-parameter layout is the same as `family_sofm` at m = 2); held-out sets `D2_scale/m{m}_heldout_{4242,9001}.npz`, 120 plants per m. Gate on those 120. The gate is the per-plant optimum, no generator is selected on it.
133
+
134
+ ## Addendum 2026-11-05: control result, S2 and S3 decisions, two diagnostics
135
+
136
+ Written after the control arm and before any S2 label, S3 model or diagnostic below exists. Only validation plants were used.
137
+
138
+ **Control result** (`control_report.txt`). Same optimizer and budget, knots frozen at sat_1: e median 2.82% [1.01, 4.73], e >= 2% on 23 of 40. F-only refinement alone: median 0.00% [0.00, 0.03]. Attribution rule met. Plants 31 and 38 were optimizer gaps (d 74.68% and 28.72% shrink to e 3.71% and 3.79%).
139
+
140
+ Re-check: 3 of 40 best spline controllers (plants 33, 37, 38) do not re-certify in a fresh process. The certificate is order dependent: plant 38 gave r*/s1 = 1.21 and 0.61 in two fresh runs, depending on which plants were solved before it. A pass is backed by a numerically checked witness (`_radius` re-check), so a pass stands; a fail is not proof of anything. Counting the 3 plants as no gain: e median 2.04% [0.43, 4.46]. The same procedure scores every arm in S4, so the order dependence does not favour one arm.
141
+
142
+ **Diagnostics (no rule, reported next to the gate):**
143
+ - Symmetrized knots: phi_sym(x) = (phi(x) - phi(-x)) / 2 at each plant's best spline controller, F kept. Its local sector lies inside phi's at every level, so phi's certificate covers it. Reported: J on the 4 cost directions against the spline and against J_F.
144
+ - Fresh directions: 64 unit directions per plant (seed (7474, i)) at s1, best spline controller against best frozen-knot controller. No certificate covers these directions, so non-convergent rollouts (final |x| > 1e-2) are counted per arm. Reported: per plant the median over directions of the paired relative difference, then the median over plants with CI.
145
+
146
+ **S2 (labels).** For each training plant (1,971) and each test plant with labels: from each X21 F-only label, the warm arm at c = 1 (F lr 0.02, knot lr 0.02, 400 steps, cheapest snapshot certified at s). If no snapshot certifies, the X21 label itself (sat_1 knots) is the row. Test plants also get c = 2 and 5 from their best X21 label; these only enter J*. Starts of several plants run in one optimizer call. The math per start is unchanged (Adam is elementwise and the loss is a sum over starts). Before the run, this is checked against single-plant `spline_optimize` on validation plants.
147
+ Why (validation only): W carries the gain (2.82% alone against 2.84% with H). c = 1 alone gives 2.04% against W0, and keeps every training row in one gauge.
148
+
149
+ **S3 (generators).** clS trains on the c = 1 joint rows: one per X21 label, the same row count as clL, J_best = the plant's cheapest joint row, target = cl encoding of F (9) + the 10 nonzero knots. coefS: slog(theta) + the 10 knots. Recipe and step budget as X21 clL. Checkpoints in `models/x28_clS*` and `models/x28_coefS*` (X21 already has an unrelated `x21_fgen_clS`). Sampler clip chosen on the 40 validation plants, as X21 did, against a validation J* = min(J_Fonly, all gate and control labels).
150
+
151
+ ## Addendum 2026-11-05: diagnostic result, symmetric-knot gate
152
+
153
+ Written after `diag_report.txt` and before the symmetric run below. S2 labels were started and stopped after one minute, before any plant was saved. They resume unchanged unless Matteo changes the knot class.
154
+
155
+ **Diagnostic result.** With the knots made odd, phi_sym(x) = (phi(x) - phi(-x)) / 2, and F kept, the spline's gain is gone: (J_sym - J_spline) / J_spline median 3.85% [1.17, 5.78], and the frozen-knot optimum beats phi_sym by a median 0.19% [0.02, 0.47]. On 64 fresh directions at s1, the spline is not cheaper than the frozen-knot controller: median over plants -0.36% [-4.87, 0.00], cheaper on 12 of 39 plants. Non-convergent rollouts: 365 of 2,560 for the spline, 300 for frozen knots. Reading: under the X21 cost (4 fixed directions per plant), the gain comes from asymmetric knots that fit those directions, and it does not carry over to new initial conditions.
156
+
157
+ **Symmetric-knot gate (S1-sym).** The question: does a spline that cannot fit sign patterns pay anything? Odd knots: 5 free values at x = 1..5, mirrored, in [-1, 1]. Same warm arm as the gate (F-only label, gauges c = 1, 2, 5, F lr 0.02, knot lr 0.02, 400 steps, cheapest snapshot certified at s1). Compared with the frozen-knot optimum J_F of the control.
158
+ Number: e_sym = (J_F - J_S) / J_S, signed, J_S = min(W_sym, J_F). Rule, as before: "a symmetric spline pays" iff median e_sym >= 2% and the 95% bootstrap CI lower bound > 0.
159
+ Diagnostic: the 64-fresh-direction comparison, best symmetric controller against best frozen-knot controller.
160
+
161
+ ## Addendum 2026-11-06: MIMO gate with the SISO lessons
162
+
163
+ Written after the MIMO validation runs (64 bank plants) and before any held-out MIMO number.
164
+
165
+ - K learning rate, by the rule of the addendum "MIMO optimizer details": median certified J_spline on validation 12.342 at 0.002, 12.386 at 0.005, 12.282 at 0.02, so 0.02. m = 2 uses the same value (no m = 2 validation run).
166
+ - Validation reading, not a gate result: free knots beat identity by a median 5.6% to 5.8% at every rate.
167
+ - The SISO risk applies here. The MIMO cost starts only from the 2m positive unit directions, so asymmetric knots could fit them. Two additions, filed before the held-out run:
168
+ 1. Odd-knot arm `so`: psi(-x) = -psi(x), equal ratios at -x_k and x_k. Same starts (c = 1, 2, 5) and budget, in the same Adam run as `id` and `sp`. Same rule: "an odd spline pays" iff median d_odd >= 2% and the CI lower bound > 0.
169
+ 2. Fresh directions: each arm's best certified controller is scored on 64 random unit directions in R^(2m) per plant, both signs, seed (7575, plant id). The MIMO certificate is global, so it covers these directions. Reading rule: an arm's gain counts as a control improvement only if the per-plant median of (J_id - J_arm) / J_arm over the fresh directions has a median over plants whose 95% CI lower bound is > 0.
170
+ - All starts of all arms run in one Adam call, checked equal to the per-arm optimizer.
171
+
172
+ SISO results that this addendum follows (in `diag_report.txt` and `gate_sym_report.txt`): odd knots give e_sym median 0.00% [0.00, 0.00], >= 2% on 0 of 40, and no difference on 64 fresh directions. S2 to S4 stay as registered (free knots), because Matteo's instruction is that the gate does not stop the generator. S4 adds one diagnostic, filed here: the S4 controllers are also scored on the 64 fresh directions of each test plant (seed (7474, i)), each arm against clL.
results/icml_r3/X28/control/cfg.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/X28/control/plant_00.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:404bc3761a4377c90233dc24624850096f17c2540cdfc28101ec8b9a7f5d0ae8
3
+ size 9644
results/icml_r3/X28/control/plant_01.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e48e72bdd8f1b0239a4d03abb07f475532274c4fe2d8c6b6076c53468e300b80
3
+ size 9501
results/icml_r3/X28/control/plant_02.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:989e8df35363ca36396ca9cf4d4ab094556e6ec58e8248d44c03cfc5748926a0
3
+ size 9579
results/icml_r3/X28/control/plant_03.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:46e06e56a0c80d9e81ae967e14693b9f2463ae1e4b275c8e5b2278f01f516e19
3
+ size 9638
results/icml_r3/X28/control/plant_04.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3b2c5d71ab5a47cddfb010dfa552918c80630fe1bad68965596dfd1ca270de53
3
+ size 9549
results/icml_r3/X28/control/plant_05.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4eecbf206c09270b80bd905082e03b95279434bed220c3dba61bb05531c29005
3
+ size 9554
results/icml_r3/X28/control/plant_06.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:63bf0494abc9e6e481473e601ace34dedee400d8aab2444ab16010bb08796d0b
3
+ size 9486
results/icml_r3/X28/control/plant_07.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6f9a77491ae479861775193ff04455f36521572caeeaac35e6d2e5c92a511069
3
+ size 9628
results/icml_r3/X28/control/plant_08.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fedd92d31d5dca6d12e361390815bbe3f3b8b4766426c841ce135a6bc27bd8ac
3
+ size 9503
results/icml_r3/X28/control/plant_09.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9770be89fd365b70c49a73e765a8afcbdfee2e0a7ab0dae5e4e4af7693b25e3c
3
+ size 9585
results/icml_r3/X28/control/plant_10.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:206fdcc3651dae922006a5d63dde23026096a52ebef440a729320ef0ce6a57a6
3
+ size 9634
results/icml_r3/X28/control/plant_11.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2b17296b2f927827cec87a87c6394e32753d27a4086e2781fef2eab19e260767
3
+ size 9598
results/icml_r3/X28/control/plant_12.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ceef5ac8173587f23b936309f8dd66817f4659fd9278d7aecf6f04715d2ff403
3
+ size 9613
results/icml_r3/X28/control/plant_13.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bdedc476ecd7c6ad814320cab96deb74596fb0b87035519ca796bfda685268c6
3
+ size 9597
results/icml_r3/X28/control/plant_14.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2aa80555babb4269bf1f3183e71e7f5d7dd863bf8fcbcbd9d558b3680aced081
3
+ size 9585
results/icml_r3/X28/control/plant_15.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:613ddfa5e7827bd9bca3c036d2e07fa04c5299da4a9bd82af2b1060d6cded175
3
+ size 9524
results/icml_r3/X28/control/plant_16.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:91f6ad849c0270b164abaf501a43aa1eae3e507ba18cddf808f4f5a8a17c5a51
3
+ size 9574
results/icml_r3/X28/control/plant_17.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3e6b753291be66c2b7cdb53440451150971df0d7f8c217059aa3bd6043fe1d99
3
+ size 9579
results/icml_r3/X28/control/plant_18.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:17972fa2ad7d22924f43553f582a175c2a53f862a0b201cd6c9fe4c15fe2d054
3
+ size 9596
results/icml_r3/X28/control/plant_19.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0e8e83ffd08e35231890c6b4f8b11a3c141b43bb1a0a7778d4163b7b19dd8c81
3
+ size 9438
results/icml_r3/X28/control/plant_20.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:be59495721481cca584e83d09db95bbd8fc0bff04d9630c7a3e9424cc14a2020
3
+ size 9523
results/icml_r3/X28/control/plant_21.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3351fe6f101f75b3f0183034204f8ccb4869f5c2759961edb49257140888edc3
3
+ size 9495
results/icml_r3/X28/control/plant_22.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4327ddb7e5645c024cd1b262efcdca93a30440c3004a6b26170b821de675ae5d
3
+ size 9539
results/icml_r3/X28/control/plant_23.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5f8733f07e67bfc0da6b5d9d925b30da3932b2c7e2e85703e1ddf11a55041d52
3
+ size 9528
results/icml_r3/X28/control/plant_24.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:befd841e318424c5874cdfffaccc61d7f2196cab9e3a8bff862fc581ee0d4404
3
+ size 9599
results/icml_r3/X28/control/plant_25.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:42c9cbb051863f543435ee3155b86870efac181a5224a79730af1bb8a33cb69a
3
+ size 9524
results/icml_r3/X28/control/plant_26.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:18bf40508cb37a22a46088c05659b0dc2728fbc4a465d91c91b3fb0c6baa4060
3
+ size 9632
results/icml_r3/X28/control/plant_27.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4ede095c413766faf68da61dd0e2dc01423a516aa4dc57d822f0b8fe4378194e
3
+ size 9556
results/icml_r3/X28/control/plant_28.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6f070dbdd7b42bfebecd90756107a57de5d0467db14587d7abd95509175d0846
3
+ size 9560
results/icml_r3/X28/control/plant_29.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8b7d867a889a34313877833aa8aeed5bb7beacde408d33dba2a9d6a78b5ec3a5
3
+ size 9572
results/icml_r3/X28/control/plant_30.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f56a1598dc846895dbfa44bb68d7ff1a14c6e2435fa01e5114c5f3b13b185725
3
+ size 9530
results/icml_r3/X28/control/plant_31.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7290c857d3844786e438e88d07648c258312df1b75410b080b60f0bbac2e6d5b
3
+ size 8436
results/icml_r3/X28/control/plant_32.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e8121bcae114bcda62c5b6f54ea6576ae74571215e81b4ca95dfff7767d673bb
3
+ size 9506
results/icml_r3/X28/control/plant_33.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ee4afa352a3709ad39959481ff89901ca0e58dfaded534c0c0b3ebb05e5cc07f
3
+ size 9668
results/icml_r3/X28/control/plant_34.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7bd8462a162e1b4a1e6601bd0fa0d226d8c03155afa7f94d22637496da32be02
3
+ size 9626
results/icml_r3/X28/control/plant_35.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:29967affb751c0209f01c506ef7e8c307e8abc47f3246756e993cf75159dcb11
3
+ size 9643
results/icml_r3/X28/control/plant_36.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:99245b58c9af2bc091c00fd23e1855bca53af95ccfebce51874d5b4040f3baa3
3
+ size 9623
results/icml_r3/X28/control/plant_37.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3eaa6c45f94cb8902292080d56fc74a39cf3facf7c2b00036275a001127b42dd
3
+ size 9585
results/icml_r3/X28/control/plant_38.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:be2c21ab4325af0407d7b5192a40bf1ab2ae7733fdc814a1c8ef75cd8986b6b2
3
+ size 9614
results/icml_r3/X28/control/plant_39.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ce30ff60d9d289e67a92115369ecaebfc483e5631008f8520c06bb2302659e8d
3
+ size 9487
results/icml_r3/X28/control_report.txt ADDED
@@ -0,0 +1,60 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ X28 S1 control arm (PREREG addendum), n = 40 of 40
2
+ H0 starts equal H starts on 36 of 40 plants
3
+
4
+ gate d = (J_Fonly - J_J) / J_J, median 2.84% [1.06, 5.61]
5
+ e = (J_F - J_J) / J_J, signed, median 2.82% [1.01, 4.73], mean 4.14% [2.56, 6.29]; e >= 2% on 23 of 40, e < 0 on 1
6
+ stratum A: median 3.87% [1.21, 6.41]
7
+ stratum B: median 1.59% [0.43, 4.98]
8
+ F-only refinement alone: (J_Fonly - J_F) / J_F, median 0.00% [0.00, 0.03]
9
+ same optimizer W at c = 1 vs W0 (one start, same lr): (J_frozen - J_free) / J_free, median 2.04% [0.94, 4.16], n = 39
10
+ same optimizer W best of 3 gauges vs W0: (J_frozen - J_free) / J_free, median 3.56% [1.09, 4.75], n = 39
11
+ same optimizer H vs H0: (J_frozen - J_free) / J_free, median 1.46% [0.41, 4.01], n = 40
12
+
13
+ best joint controller recomputed: J vs gate max rel 0.0e+00, re-certified at s1 on 37 of 40
14
+ knot removal (F kept, knots -> sat_1): (J_sat - J_spline) / J_spline, median 10.96% [8.42, 18.12]; loop with sat_1 certified at s1 on 33 of 40
15
+ dt check: |J(dt 0.01) - J(dt 0.03)| / J, spline median 1.6e-04, max 1.1e-03
16
+ same gain at dt 0.01: (J_sat - J_spline) / J_spline, median 10.96% [8.41, 18.13]
17
+
18
+ ATTRIBUTION: the gate pass is attributed to the spline (median e >= 2% and CI lower bound > 0).
19
+
20
+ i s J_Fonly J_W J_H J_W0 J_H0 e% knotrm% satcert
21
+ 0 A 0.070657 0.069999 0.069999 0.070657 0.070657 0.94 2.20 1
22
+ 1 A 0.38974 0.38974 0.4109 0.38974 0.38974 0.00 0.00 1
23
+ 2 A 0.13499 0.1296 0.13002 0.13499 0.13499 4.16 9.09 1
24
+ 3 A 0.2795 0.2739 0.27459 0.2795 0.2795 2.04 24724.69 0
25
+ 4 A 0.68583 0.68583 0.70086 0.68583 0.68583 0.00 0.00 1
26
+ 5 A 0.1105 0.094083 0.10079 0.10101 0.1105 7.36 13.41 1
27
+ 6 A 1.3606 1.3606 1.4547 1.3606 1.3606 0.00 0.00 1
28
+ 7 A 0.12 0.11587 0.11584 0.12 0.12 3.59 13.69 1
29
+ 8 A 0.074069 0.05373 0.053743 0.073992 0.074069 37.71 49.19 1
30
+ 9 A 0.10265 0.10114 0.10213 0.10264 0.10265 1.49 725.83 0
31
+ 10 A 0.081738 0.078435 0.078435 0.081738 0.081738 4.21 8.39 1
32
+ 11 A 0.16231 0.159 0.16222 0.16225 0.16225 2.05 3.46 1
33
+ 12 A 0.1853 0.17285 0.17285 0.1853 0.1853 7.20 10.18 1
34
+ 13 A 0.49868 0.49813 0.4983 0.49853 0.49868 0.08 0.29 1
35
+ 14 A 0.071794 0.066323 0.066322 0.071793 0.071794 8.25 13.09 1
36
+ 15 A 0.29173 0.26758 0.26834 0.2917 0.29173 9.01 817.43 0
37
+ 16 A 0.14761 0.13975 0.1398 0.14761 0.14761 5.62 59.93 1
38
+ 17 A 0.13627 0.13013 0.13232 0.13627 0.13627 4.72 78.34 1
39
+ 18 A 0.23848 0.17945 0.20849 0.19377 0.23848 7.98 45.30 1
40
+ 19 A 0.77037 0.76988 0.77021 0.77036 0.77037 0.06 116.78 1
41
+ 20 B 2.1194 2.1194 2.1622 2.1194 2.1194 0.00 0.00 1
42
+ 21 B 0.42494 0.42459 0.42555 0.42475 0.42494 0.04 3.14 1
43
+ 22 B 2.4229 2.3077 2.3813 2.4174 2.4229 4.75 6255.82 0
44
+ 23 B 0.061683 0.061606 0.061606 0.061683 0.061683 0.12 86824.25 0
45
+ 24 B 0.040465 0.037751 0.037784 0.040383 0.040465 6.97 8.45 1
46
+ 25 B 2.0998 2.0998 2.122 2.0998 2.0998 0.00 2594.93 0
47
+ 26 B 0.024577 0.023242 0.023252 0.024575 0.024577 5.74 11.74 1
48
+ 27 B 0.13865 0.12787 0.12852 0.13454 0.13865 5.21 14.92 0
49
+ 28 B 0.24806 0.22929 0.22946 0.24796 0.24806 8.14 8.74 1
50
+ 29 B 0.079573 0.077978 0.077978 0.079573 0.079573 2.05 3.36 1
51
+ 30 B 0.019654 0.019414 0.019428 0.019625 0.019654 1.09 10.01 1
52
+ 31 B 0.18762 0.10741 0.18762 0.11139 0.18762 3.71 21.32 1
53
+ 32 B 0.25918 0.23587 0.23587 0.25918 0.25918 9.88 14.23 1
54
+ 33 B 0.091918 0.087037 0.087037 0.091918 0.091918 5.61 8.95 1
55
+ 34 B 0.20699 0.20579 0.20579 0.20699 0.20699 0.58 1.43 1
56
+ 35 B 0.1507 0.15023 0.15023 0.1507 0.1507 0.32 0.59 1
57
+ 36 B 1.7212 1.7212 1.712 1.7212 1.7212 0.54 2.49 1
58
+ 37 B 0.28926 0.28587 0.28595 0.28911 0.28926 1.13 1.71 1
59
+ 38 B 0.15583 0.12107 0.15583 0.12566 0.15583 3.79 40.28 1
60
+ 39 B 0.27761 inf 0.28356 0.2756 0.27761 -0.72 36.29 1
results/icml_r3/X28/diag.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fb1e83f98841331f3cb0b55f572c202b8dff133d2c776210ca4c370aa4a02664
3
+ size 77126
results/icml_r3/X28/diag_report.txt ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ X28 S1 diagnostics (PREREG addendum 'control result, S2 and S3'), n = 40 of 40 plants with a certified spline
2
+ recomputed spline J vs gate: max rel 0.0e+00
3
+ symmetrized knots, F kept: (J_sym - J_spline) / J_spline median 3.85% [1.17, 5.78]; (J_F - J_sym) / J_sym median -0.19% [-0.47, -0.02]
4
+ for reference (J_F - J_spline) / J_spline median 2.82% [1.01, 4.73]
5
+ 64 fresh directions at s1: non-convergent rollouts, spline 365 of 2560 (plants with any 11), frozen knots 300 (plants with any 9)
6
+ per plant median over directions where both converge of (J_F - J_spline) / J_spline; median over plants -0.36% [-4.87, 0.00], > 0 on 12 of 39
results/icml_r3/X28/gate/cfg.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/X28/gate/plant_00.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:34e62a52677e144992bfa6bb50071aee4934d88501bbe8e99a435377a7d6d11e
3
+ size 7828
results/icml_r3/X28/gate/plant_01.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:40777e05be0c5a475f0e2ecdb56505faba8eadbd88c2314ac4a4585675fb5a24
3
+ size 7567
results/icml_r3/X28/gate/plant_02.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e75bce70119dfc97d7e7b098b2dcf32d77dfb39a603701bb369c01aef82ce6f9
3
+ size 7768
results/icml_r3/X28/gate/plant_03.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f785ecf4bd0fbf2c8b427892b82230a9f69c6e61889cc65afc8b54e7197b7804
3
+ size 7882