matCercola18 commited on
Commit
c6dfe66
·
verified ·
1 Parent(s): 268d7a3

X28: SISO S4 test (eval_test, paired, fresh), validation with t-repair, MIMO spline-class study, PREREG 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 +43 -0
  2. results/icml_r3/X28/eval_test/x21_fgen_clL_best_tr_K16_clip1.5.npz +3 -0
  3. results/icml_r3/X28/eval_test/x21_fgen_clL_best_tr_K64_clip1.5.npz +3 -0
  4. results/icml_r3/X28/eval_test/x21_fgen_clL_s1_best_tr_K16_clip1.5.npz +3 -0
  5. results/icml_r3/X28/eval_test/x21_fgen_clL_s2_best_tr_K16_clip1.5.npz +3 -0
  6. results/icml_r3/X28/eval_test/x28_clS_s0_best_K16_clip1.npz +3 -0
  7. results/icml_r3/X28/eval_test/x28_clS_s0_best_tr_K16_clip1.5.npz +3 -0
  8. results/icml_r3/X28/eval_test/x28_clS_s0_best_tr_K64_clip1.5.npz +3 -0
  9. results/icml_r3/X28/eval_test/x28_clS_s1_best_K16_clip1.5.npz +3 -0
  10. results/icml_r3/X28/eval_test/x28_clS_s1_best_tr_K16_clip1.5.npz +3 -0
  11. results/icml_r3/X28/eval_test/x28_clS_s2_best_K16_clip1.npz +3 -0
  12. results/icml_r3/X28/eval_test/x28_clS_s2_best_tr_K16_clip1.5.npz +3 -0
  13. results/icml_r3/X28/eval_test/x28_coefS_s0_best_tr_K16_clip3.npz +3 -0
  14. results/icml_r3/X28/eval_test_fresh.npz +3 -0
  15. results/icml_r3/X28/eval_test_fresh.txt +8 -0
  16. results/icml_r3/X28/eval_test_paired.txt +33 -0
  17. results/icml_r3/X28/eval_test_report.txt +62 -0
  18. results/icml_r3/X28/eval_val/x28_clS_s0_best_tr_K16_clip1.5.npz +3 -0
  19. results/icml_r3/X28/eval_val/x28_clS_s0_best_tr_K16_clip1.npz +3 -0
  20. results/icml_r3/X28/eval_val/x28_clS_s0_best_tr_K16_clip2.npz +3 -0
  21. results/icml_r3/X28/eval_val/x28_clS_s0_best_tr_K16_clip3.npz +3 -0
  22. results/icml_r3/X28/eval_val/x28_clS_s1_best_tr_K16_clip1.5.npz +3 -0
  23. results/icml_r3/X28/eval_val/x28_clS_s1_best_tr_K16_clip1.npz +3 -0
  24. results/icml_r3/X28/eval_val/x28_clS_s1_best_tr_K16_clip2.npz +3 -0
  25. results/icml_r3/X28/eval_val/x28_clS_s1_best_tr_K16_clip3.npz +3 -0
  26. results/icml_r3/X28/eval_val/x28_clS_s2_best_tr_K16_clip1.5.npz +3 -0
  27. results/icml_r3/X28/eval_val/x28_clS_s2_best_tr_K16_clip1.npz +3 -0
  28. results/icml_r3/X28/eval_val/x28_clS_s2_best_tr_K16_clip2.npz +3 -0
  29. results/icml_r3/X28/eval_val/x28_clS_s2_best_tr_K16_clip3.npz +3 -0
  30. results/icml_r3/X28/eval_val/x28_coefS_s0_best_tr_K16_clip1.5.npz +3 -0
  31. results/icml_r3/X28/eval_val/x28_coefS_s0_best_tr_K16_clip1.npz +3 -0
  32. results/icml_r3/X28/eval_val/x28_coefS_s0_best_tr_K16_clip2.npz +3 -0
  33. results/icml_r3/X28/eval_val/x28_coefS_s0_best_tr_K16_clip3.npz +3 -0
  34. results/icml_r3/X28/eval_val_report.txt +80 -0
  35. results/icml_r3/X28/mimo/fresh_study_m2.npz +3 -0
  36. results/icml_r3/X28/mimo/fresh_study_m2.txt +6 -0
  37. results/icml_r3/X28/mimo/fresh_study_m3.npz +3 -0
  38. results/icml_r3/X28/mimo/fresh_study_m3.txt +6 -0
  39. results/icml_r3/X28/mimo/gate_study_m2.txt +27 -0
  40. results/icml_r3/X28/mimo/gate_study_m3.txt +27 -0
  41. results/icml_r3/X28/mimo/study_choice.txt +8 -0
  42. results/icml_r3/X28/mimo/study_m2/cert_cf_c1.npz +3 -0
  43. results/icml_r3/X28/mimo/study_m2/cert_cf_c2.npz +3 -0
  44. results/icml_r3/X28/mimo/study_m2/cert_cf_c5.npz +3 -0
  45. results/icml_r3/X28/mimo/study_m2/cert_co_c1.npz +3 -0
  46. results/icml_r3/X28/mimo/study_m2/cert_co_c2.npz +3 -0
  47. results/icml_r3/X28/mimo/study_m2/cert_co_c5.npz +3 -0
  48. results/icml_r3/X28/mimo/study_m2/cert_id_c1.npz +3 -0
  49. results/icml_r3/X28/mimo/study_m2/cert_so3_c1.npz +3 -0
  50. results/icml_r3/X28/mimo/study_m2/cert_so3_c2.npz +3 -0
results/icml_r3/X28/PREREG.md CHANGED
@@ -181,3 +181,46 @@ Written after the S3 validation runs (`eval_val_report.txt`) and before any test
181
  - **Paired test (section 1.5).** Seed j of clS against seed j of clL, plant-paired, at b = 16, regret against J* with uncertified = 1e3 (`x21_paired.py` conventions): median per-plant difference clL - clS with bootstrap CI and Wilcoxon signed-rank p, Holm over the 3 seed pairs; certified rate difference with exact McNemar p. A verdict ("spline better" or "spline worse", section 1.5 thresholds) holds iff it holds on at least 2 of the 3 seed pairs. All b = 1, 4, 16 (64 where present) are reported.
182
  - **Fresh-direction diagnostic** (addendum 2026-11-06): each arm's best certified candidate at b = 16 per test plant, scored on 64 unit directions at s (seed (7474, i)), against clL seed 0's; non-convergent rollouts counted per arm.
183
  - Compute: the CPU node is down, so this runs on idle `gpu-light` node 02 (Matteo's 2026-11-06 fallback).
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
181
  - **Paired test (section 1.5).** Seed j of clS against seed j of clL, plant-paired, at b = 16, regret against J* with uncertified = 1e3 (`x21_paired.py` conventions): median per-plant difference clL - clS with bootstrap CI and Wilcoxon signed-rank p, Holm over the 3 seed pairs; certified rate difference with exact McNemar p. A verdict ("spline better" or "spline worse", section 1.5 thresholds) holds iff it holds on at least 2 of the 3 seed pairs. All b = 1, 4, 16 (64 where present) are reported.
182
  - **Fresh-direction diagnostic** (addendum 2026-11-06): each arm's best certified candidate at b = 16 per test plant, scored on 64 unit directions at s (seed (7474, i)), against clL seed 0's; non-convergent rollouts counted per arm.
183
  - Compute: the CPU node is down, so this runs on idle `gpu-light` node 02 (Matteo's 2026-11-06 fallback).
184
+
185
+ ## Addendum 2026-11-13: MIMO spline-class study (Matteo: "choose the best one ... let's do a study on which one to use")
186
+
187
+ Written after the MIMO held-out gate (`gate_heldout_m{2,3}.txt`, `fresh_heldout_m{2,3}.txt`) and before any number below exists. Held-out reading so far: free piecewise-linear knots pay at the optimum (median d 5.19% at m = 3, 6.36% at m = 2) but lose on fresh directions (-4.55%, -1.58%); odd piecewise-linear knots pay (4.36%, 5.52%) and keep a fresh-direction gain (+0.54% [0.20, 1.27], +2.71% [1.88, 3.39]). Matteo asked for the spline class to be chosen by a study. The study runs on validation plants only; the held-out plants then confirm the chosen class.
188
+
189
+ **Classes** (all with knot ratios r = psi(x)/x in [0.1, 10] at x in {-5, ..., -1, 1, ..., 5}, ray extension past +-5, sector exactly [min r, max r] per channel, circle LMI of section 2.1):
190
+ - `sp` piecewise linear, free (10 per channel), the July class;
191
+ - `so` piecewise linear, odd (5 per channel);
192
+ - `so3` piecewise linear, odd, knots at |x| = 1, 3, 5 only (3 per channel; psi linear in between);
193
+ - `cf` cubic, free: psi(v) = v rho(v), rho the Fritsch-Carlson PCHIP interpolant of the 10 ratios (end slopes = end secants). PCHIP stays between its two end values on every interval, so the sector is exact (checked on a grid of 4001 points);
194
+ - `co` cubic, odd (5 per channel).
195
+ Baseline: `id` (identity knots).
196
+
197
+ **Run.** The first 64 bank plants for m = 3 (`D2_scale/m3_bank.npz`) and m = 2 (`C_mimo/sof/bank.npz`), the K learning rate of the gate (0.02), knot lr 0.02, 400 steps, gauge starts c = 1, 2, 5 for every class, all starts of all classes in one Adam run, cheapest certified snapshot (`x28_mimo.py`, tag `study`). Fresh directions: 64 random unit directions in R^(2m) per plant, both signs, seed (7676, plant id) (distinct from the held-out seed 7575, because validation and held-out ids overlap).
198
+
199
+ **Selection rule.** For class c and size m, f_c(m) = median over plants of the per-plant median over the 64 fresh directions of (J_id - J_c) / J_c (`stage_fresh`, both rollouts convergent). Eligible: the 95% bootstrap CI lower bound of f_c(m) is > 0 at both m, and the non-convergent fresh rollouts of c exceed those of `id` by at most 1% of the rollouts (41 of 4,096) at both m. Winner: the eligible class with the highest mean of f_c over m = 2, 3. Parsimony: among eligible classes within 0.25 points of the winner's mean, the one with the fewest parameters per channel (ties: piecewise linear). The gate number d at the optimum is reported for every class but does not select. If no class is eligible, the MIMO generator arm (b) waits for Matteo.
200
+
201
+ **Confirmation.** The winner on the 120 held-out plants per m with the gate rule of 2.3 and the fresh-direction reading rule of addendum 2026-11-06 (seed 7575). If the winner is `so`, that held-out result already exists and was seen before this study (above). The winner becomes the spline class of MIMO arm (b) (sections 2.2 and 2.4); labels and generators follow in a later addendum.
202
+
203
+ SISO S4 continues as registered (free piecewise-linear knots). The SISO odd-knot gate gave 0.00%, so the class study is run on MIMO, where a gain exists.
204
+
205
+ Compute: inside Matteo's GPU hold job on node 05 (he allowed its use on 2026-11-12), with the SISO S4 job on node 02: 2 GPU nodes.
206
+
207
+ ## Addendum 2026-11-13: MIMO study result, labels, generators and evaluation (sections 2.2 to 2.4)
208
+
209
+ Written after the study (`study_choice.txt`) and before the held-out so3 confirmation, any MIMO label, model or generator result.
210
+
211
+ **Study result** (`study_choice.txt`, 64 validation plants per m). Fresh-direction gain f over identity: `so3` +3.99% [2.86, 5.13] at m = 2 and +1.82% [0.74, 2.57] at m = 3; `so` +3.11%, +1.32%; `co` +2.75%, +0.91%; free classes negative at both m (`sp` -2.11%, -5.94%; `cf` -3.09%, -6.06%) with 180 to 188 (`sp`) and 13 to 43 (`cf`) more non-convergent rollouts than identity. Choice by the rule: **so3**, odd piecewise linear with knots at |x| = 1, 3, 5. The held-out confirmation runs `so3` (starts c = 1, 2, 5) in its own Adam call next to the stored `id`, `sp`, `so` runs; starts never interact, so this equals one joint call.
212
+
213
+ **Labels (S2-M).** Bank plants without the 64 validation plants: m = 3 `D2_scale/m3_bank.npz` [64, 8000) (7,936 plants), m = 2 `C_mimo/sof/bank.npz` [64, 4000) (3,936 plants). Optimizer of the gate: `id` at c = 1 and `so3` at c = 1, 2, 5, K lr 0.02, knot lr 0.02, 400 steps, snapshot every 25, cheapest certified snapshot (`x28_mimo.py` optimize and certify, tag `lab`), with gradient checkpointing of the rollout (`X28M_SEG=100`, identical values and gradients, checked). Label (a): the `id` arm's controller K. Label (b): the cheapest certified of the three `so3` starts and `id` (identity is in the so3 class, all ratios 1). Snapshot 0 is the bank gain, certified for identity, so every plant has both labels.
214
+
215
+ **Generators (S3-M).** One row per plant per arm. Conditioning: plant parameters scaled to [-1, 1] by the family box (as X27). Target (a): K (m^2 numbers). Target (b): K and the knots in a fixed gauge: per channel i, K row i times r_i(1) and ratios divided by r_i(1), so r_i(1) = 1 and the controller is unchanged; target = K (m^2) and log r_i(3), log r_i(5) (2m). Targets min-max scaled to [-0.96, 0.96] (`make_dataset_sofm` convention). Trainer: X27's `train.py` (DDPM-500, MLP 256 x 5, 96,000 steps, batch 64, lr 2e-4, EMA 0.999), seeds 0, 1, 2 per arm and m. No checkpoint selection: the final EMA 0.999 weights.
216
+
217
+ **Evaluation (S4-M).** The 120 held-out plants per m (`m{m}_heldout_{4242,9001}`). 16 samples per plant, DDPM with clip range 1 (as X27 `sample.py`), sampler seed = model seed. Decode, then for (b) project the ratios to [0.1, 10]. A sample counts iff certified: (a) the Lyapunov LMI of A - B K C at alpha = beta = 1 (as the `id` arm); (b) the circle LMI of section 2.1 at its sector. Cost J: 2m positive unit directions, 20 s, RK4 dt 0.01 (the label cost). J* = min over all held-out labels (`id`, `sp`, `so`, `so3`). Regret at b = 1, 4, 16 with uncertified = 1e3.
218
+ Primary, per m: (b) seed j against (a) seed j at b = 16, plant-paired; "spline better" iff median paired regret difference (a) - (b) >= 1 point, Holm (over the 3 seed pairs) Wilcoxon p < 0.05, and certified rate of (b) not lower by more than 2 points; "spline worse" with the arms swapped; verdict iff on >= 2 of 3 pairs, otherwise "no measurable difference" (the SISO rule of 1.5).
219
+ Diagnostic: each arm's best certified b = 16 sample per plant on the 64 fresh directions of seed (7575, plant id), against (a) seed 0, as in the gate.
220
+
221
+ Compute: inside GPL_move's job 17025 on node 03 (Matteo 2026-11-13: share that session's node), within its 12 GB VRAM and 3 CPU budget.
222
+
223
+ ## Addendum 2026-11-13: correction to the arm (b) target
224
+
225
+ Written before any MIMO label run, dataset or model exists, after a round-trip check of the code on the study's validation labels. The gauge of the previous addendum (K row i times r_i(1), ratios divided by r_i(1)) is not a symmetry of the so3 class: the knots sit at fixed positions, and a row scaling c is only matched by psi_i(w / c), whose knots move to c x_k. On the 64 validation labels the round trip changed J by up to 72% (m = 2) and 225% (m = 3); without it the round trip is exact (3e-16). The gauge starts c = 1, 2, 5 of the optimizer stay valid, because they start from a linear psi, for which the scaling is exact.
226
+ Target of arm (b), replacing the previous one: K (m^2) and log r_i(1), log r_i(3), log r_i(5) per channel (3m), each ratio projected to [0.1, 10] after decoding. Nothing else changes.
results/icml_r3/X28/eval_test/x21_fgen_clL_best_tr_K16_clip1.5.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8c08bb2cb874310cf82830ac8d387a8e15f509f82f81f3f969fbadc1017540c9
3
+ size 283098
results/icml_r3/X28/eval_test/x21_fgen_clL_best_tr_K64_clip1.5.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:801b6a9a8313ed566bd74b1a3fab6710789b72a9008f2fabdeb04e7a8b700df2
3
+ size 1000130
results/icml_r3/X28/eval_test/x21_fgen_clL_s1_best_tr_K16_clip1.5.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:144a92e411113600c40b95c7dbf6aa9b1f13c3d8874a41dab783dc4cc6d50d10
3
+ size 282911
results/icml_r3/X28/eval_test/x21_fgen_clL_s2_best_tr_K16_clip1.5.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b1ec5ec6cc977ff1960ba830e4f70a9867faf493ef8300babca9d4ea023d3eee
3
+ size 283059
results/icml_r3/X28/eval_test/x28_clS_s0_best_K16_clip1.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c7ae2df412a230999495c6874507edcbd7f90d31337d411a19d26135ee89936d
3
+ size 406038
results/icml_r3/X28/eval_test/x28_clS_s0_best_tr_K16_clip1.5.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6a8ee07d59e0ae3d5ad03767c388c221312b88e53af75407c64e92dc3e833f6e
3
+ size 408708
results/icml_r3/X28/eval_test/x28_clS_s0_best_tr_K64_clip1.5.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d7010fc5e83dcd362d8b7528a4573d66aa4186c2b7fe0199153968228a6f1b99
3
+ size 1501865
results/icml_r3/X28/eval_test/x28_clS_s1_best_K16_clip1.5.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9ecb26bbfc89abae03cc98bd1a99199e26f6f8f2cd47ab572e4a9032ad6cb9aa
3
+ size 406750
results/icml_r3/X28/eval_test/x28_clS_s1_best_tr_K16_clip1.5.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:dcb17bdf400b35ede7c9c6224305a566ed9663d53dcaad680e1575954cd20071
3
+ size 410425
results/icml_r3/X28/eval_test/x28_clS_s2_best_K16_clip1.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:92be1ce6a89259cc00acea1e30544976146a05de70c9578bb9fefb9992966679
3
+ size 406592
results/icml_r3/X28/eval_test/x28_clS_s2_best_tr_K16_clip1.5.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0c51c3ccdafc56f271d46e306c92435995c1f05e27575c47e89f14568dc03212
3
+ size 409386
results/icml_r3/X28/eval_test/x28_coefS_s0_best_tr_K16_clip3.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:66bd5cfa12c56f42cb22b277db1e7f2919340ea6f3fea65b6a25ec049b64afe6
3
+ size 408282
results/icml_r3/X28/eval_test_fresh.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f57f48c9a3f3be96ead682ee9a999054c1083a76cb0df103db7cdbd598013e5d
3
+ size 1319384
results/icml_r3/X28/eval_test_fresh.txt ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ == S4 fresh directions: best certified b=16 candidate per test plant on 64 unit directions at s (seed (7474, i)), vs x21_fgen_clL_best_tr_K16_clip1.5
2
+ x21_fgen_clL_best_tr_K16_clip1.5 plants 196 | median over plants of per-plant median (J_clL0 - J_arm) / J_arm +0.00% [+0.00, +0.00], > 0 on 0 | non-convergent rollouts 626 of 12544
3
+ x28_clS_s0_best_tr_K16_clip1.5 plants 188 | median over plants of per-plant median (J_clL0 - J_arm) / J_arm -1.41% [-2.40, -0.30], > 0 on 74 | non-convergent rollouts 1160 of 12096
4
+ x28_clS_s1_best_tr_K16_clip1.5 plants 190 | median over plants of per-plant median (J_clL0 - J_arm) / J_arm -1.50% [-2.68, -0.30], > 0 on 77 | non-convergent rollouts 1775 of 12288
5
+ x28_clS_s2_best_tr_K16_clip1.5 plants 190 | median over plants of per-plant median (J_clL0 - J_arm) / J_arm -1.31% [-2.14, -0.30], > 0 on 80 | non-convergent rollouts 1189 of 12160
6
+ x28_coefS_s0_best_tr_K16_clip3 plants 186 | median over plants of per-plant median (J_clL0 - J_arm) / J_arm -4.44% [-5.77, -2.89], > 0 on 47 | non-convergent rollouts 2089 of 12096
7
+ x21_fgen_clL_s1_best_tr_K16_clip1.5 plants 193 | median over plants of per-plant median (J_clL0 - J_arm) / J_arm +0.22% [-0.09, +0.53], > 0 on 107 | non-convergent rollouts 568 of 12480
8
+ x21_fgen_clL_s2_best_tr_K16_clip1.5 plants 193 | median over plants of per-plant median (J_clL0 - J_arm) / J_arm +0.23% [-0.08, +0.57], > 0 on 107 | non-convergent rollouts 1154 of 12416
results/icml_r3/X28/eval_test_paired.txt ADDED
@@ -0,0 +1,33 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ == PRIMARY: clS (t-repair) vs clL (t-repair), seed j vs seed j; Holm over the seed pairs of each b
2
+ x28_clS_s0_best_tr_K16_clip1.5 vs x21_fgen_clL_best_tr_K16_clip1.5 b= 1 | regret 100000.00% vs 54.90% | clL - spline median +0.00 pts [-10.29, +0.00] p_holm 0.001 | certified 49.2% vs 65.3%, diff -16.1 [-24.6, -7.5] McNemar p 0.00045 | spline cheaper on 29.1%, dearer on 48.2%
3
+ x28_clS_s1_best_tr_K16_clip1.5 vs x21_fgen_clL_s1_best_tr_K16_clip1.5 b= 1 | regret 135.35% vs 49.63% | clL - spline median +0.00 pts [-1.69, +0.00] p_holm 0.0063 | certified 52.8% vs 65.3%, diff -12.6 [-21.1, -4.5] McNemar p 0.0059 | spline cheaper on 32.7%, dearer on 45.7%
4
+ x28_clS_s2_best_tr_K16_clip1.5 vs x21_fgen_clL_s2_best_tr_K16_clip1.5 b= 1 | regret 100000.00% vs 42.49% | clL - spline median +0.00 pts [-3.61, +0.00] p_holm 0.0029 | certified 49.7% vs 61.3%, diff -11.6 [-19.6, -3.0] McNemar p 0.0095 | spline cheaper on 28.1%, dearer on 45.7%
5
+ x28_clS_s0_best_tr_K16_clip1.5 vs x21_fgen_clL_best_tr_K16_clip1.5 b= 4 | regret 16.47% vs 12.85% | clL - spline median -1.25 pts [-3.69, +0.00] p_holm 0.0055 | certified 85.9% vs 92.0%, diff -6.0 [-10.6, -1.0] McNemar p 0.023 | spline cheaper on 41.7%, dearer on 53.3%
6
+ x28_clS_s1_best_tr_K16_clip1.5 vs x21_fgen_clL_s1_best_tr_K16_clip1.5 b= 4 | regret 15.09% vs 13.25% | clL - spline median -1.12 pts [-2.08, +0.00] p_holm 0.0055 | certified 85.4% vs 94.0%, diff -8.5 [-14.1, -3.5] McNemar p 0.0023 | spline cheaper on 41.2%, dearer on 55.8%
7
+ x28_clS_s2_best_tr_K16_clip1.5 vs x21_fgen_clL_s2_best_tr_K16_clip1.5 b= 4 | regret 19.78% vs 11.93% | clL - spline median -1.27 pts [-4.65, +0.00] p_holm 0.0027 | certified 83.4% vs 89.9%, diff -6.5 [-12.1, -1.0] McNemar p 0.035 | spline cheaper on 39.7%, dearer on 55.3%
8
+ x28_clS_s0_best_tr_K16_clip1.5 vs x21_fgen_clL_best_tr_K16_clip1.5 b=16 | regret 7.14% vs 6.79% | clL - spline median -0.04 pts [-0.61, +0.48] p_holm 1 | certified 95.0% vs 98.5%, diff -3.5 [-6.5, -1.0] McNemar p 0.039 | spline cheaper on 48.7%, dearer on 50.3%
9
+ x28_clS_s1_best_tr_K16_clip1.5 vs x21_fgen_clL_s1_best_tr_K16_clip1.5 b=16 | regret 6.33% vs 6.62% | clL - spline median +0.19 pts [-0.42, +0.91] p_holm 1 | certified 96.5% vs 98.0%, diff -1.5 [-4.5, +1.5] McNemar p 0.51 | spline cheaper on 50.8%, dearer on 48.7%
10
+ x28_clS_s2_best_tr_K16_clip1.5 vs x21_fgen_clL_s2_best_tr_K16_clip1.5 b=16 | regret 6.30% vs 6.81% | clL - spline median -0.32 pts [-1.19, +0.19] p_holm 1 | certified 95.5% vs 97.5%, diff -2.0 [-4.5, +0.0] McNemar p 0.22 | spline cheaper on 45.2%, dearer on 52.8%
11
+ VERDICT b=16 (3 seed pairs): better on 0, worse on 0 -> no measurable difference
12
+ == secondary: clS without t-repair vs clL
13
+ x28_clS_s0_best_K16_clip1 vs x21_fgen_clL_best_tr_K16_clip1.5 b= 1 | regret 100000.00% vs 54.90% | clL - spline median -4.71 pts [-60.06, +0.00] p_holm 5.4e-06 | certified 43.7% vs 65.3%, diff -21.6 [-30.7, -12.1] McNemar p 1.5e-05 | spline cheaper on 25.6%, dearer on 53.3%
14
+ x28_clS_s1_best_K16_clip1.5 vs x21_fgen_clL_s1_best_tr_K16_clip1.5 b= 1 | regret 100000.00% vs 49.63% | clL - spline median -1.81 pts [-35.18, +0.00] p_holm 6.1e-08 | certified 35.7% vs 65.3%, diff -29.6 [-38.2, -21.1] McNemar p 4.4e-10 | spline cheaper on 21.6%, dearer on 52.3%
15
+ x28_clS_s2_best_K16_clip1 vs x21_fgen_clL_s2_best_tr_K16_clip1.5 b= 1 | regret 100000.00% vs 42.49% | clL - spline median +0.00 pts [-17.29, +0.00] p_holm 0.00015 | certified 44.2% vs 61.3%, diff -17.1 [-26.1, -8.0] McNemar p 0.00037 | spline cheaper on 27.6%, dearer on 47.2%
16
+ x28_clS_s0_best_K16_clip1 vs x21_fgen_clL_best_tr_K16_clip1.5 b= 4 | regret 26.65% vs 12.85% | clL - spline median -7.51 pts [-12.91, -3.98] p_holm 1.5e-10 | certified 82.4% vs 92.0%, diff -9.5 [-15.1, -4.0] McNemar p 0.0013 | spline cheaper on 26.1%, dearer on 69.3%
17
+ x28_clS_s1_best_K16_clip1.5 vs x21_fgen_clL_s1_best_tr_K16_clip1.5 b= 4 | regret 24.18% vs 13.25% | clL - spline median -5.23 pts [-10.55, -1.82] p_holm 4.6e-09 | certified 74.4% vs 94.0%, diff -19.6 [-26.1, -13.1] McNemar p 1.8e-08 | spline cheaper on 30.2%, dearer on 66.8%
18
+ x28_clS_s2_best_K16_clip1 vs x21_fgen_clL_s2_best_tr_K16_clip1.5 b= 4 | regret 33.60% vs 11.93% | clL - spline median -5.95 pts [-13.14, -2.90] p_holm 6.2e-10 | certified 79.9% vs 89.9%, diff -10.1 [-16.1, -4.0] McNemar p 0.0017 | spline cheaper on 26.1%, dearer on 68.3%
19
+ x28_clS_s0_best_K16_clip1 vs x21_fgen_clL_best_tr_K16_clip1.5 b=16 | regret 10.57% vs 6.79% | clL - spline median -2.92 pts [-4.67, -1.72] p_holm 3.8e-09 | certified 93.5% vs 98.5%, diff -5.0 [-8.5, -2.0] McNemar p 0.0063 | spline cheaper on 27.6%, dearer on 71.4%
20
+ x28_clS_s1_best_K16_clip1.5 vs x21_fgen_clL_s1_best_tr_K16_clip1.5 b=16 | regret 7.37% vs 6.62% | clL - spline median -0.75 pts [-1.67, -0.00] p_holm 0.0014 | certified 91.0% vs 98.0%, diff -7.0 [-11.1, -3.5] McNemar p 0.00052 | spline cheaper on 41.2%, dearer on 57.3%
21
+ x28_clS_s2_best_K16_clip1 vs x21_fgen_clL_s2_best_tr_K16_clip1.5 b=16 | regret 10.28% vs 6.81% | clL - spline median -2.85 pts [-4.08, -1.62] p_holm 1.5e-08 | certified 94.0% vs 97.5%, diff -3.5 [-6.5, -1.0] McNemar p 0.039 | spline cheaper on 29.6%, dearer on 68.3%
22
+ == secondary: coefS seed 0 (t-repair) vs clL seed 0
23
+ x28_coefS_s0_best_tr_K16_clip3 vs x21_fgen_clL_best_tr_K16_clip1.5 b= 1 | regret 100000.00% vs 54.90% | clL - spline median -6.54 pts [-18.65, +0.00] p_holm 0.0012 | certified 48.7% vs 65.3%, diff -16.6 [-26.6, -6.5] McNemar p 0.002 | spline cheaper on 29.1%, dearer on 55.3%
24
+ x28_coefS_s0_best_tr_K16_clip3 vs x21_fgen_clL_best_tr_K16_clip1.5 b= 4 | regret 21.51% vs 12.85% | clL - spline median -8.35 pts [-12.66, -5.04] p_holm 2.4e-08 | certified 83.4% vs 92.0%, diff -8.5 [-14.6, -2.0] McNemar p 0.012 | spline cheaper on 30.7%, dearer on 67.3%
25
+ x28_coefS_s0_best_tr_K16_clip3 vs x21_fgen_clL_best_tr_K16_clip1.5 b=16 | regret 10.58% vs 6.79% | clL - spline median -3.23 pts [-5.46, -2.24] p_holm 4.2e-10 | certified 95.0% vs 98.5%, diff -3.5 [-7.0, +0.0] McNemar p 0.092 | spline cheaper on 28.1%, dearer on 71.9%
26
+ == secondary: K = 64, seed 0
27
+ x28_clS_s0_best_tr_K64_clip1.5 vs x21_fgen_clL_best_tr_K64_clip1.5 b= 1 | regret 228.92% vs 53.11% | clL - spline median -0.24 pts [-8.79, +0.00] p_holm 0.0068 | certified 51.8% vs 67.8%, diff -16.1 [-25.6, -7.0] McNemar p 0.0013 | spline cheaper on 33.2%, dearer on 50.3%
28
+ x28_clS_s0_best_tr_K64_clip1.5 vs x21_fgen_clL_best_tr_K64_clip1.5 b= 4 | regret 14.50% vs 12.85% | clL - spline median -0.81 pts [-2.51, +0.00] p_holm 0.066 | certified 88.4% vs 91.0%, diff -2.5 [-6.5, +1.5] McNemar p 0.33 | spline cheaper on 41.7%, dearer on 52.3%
29
+ x28_clS_s0_best_tr_K64_clip1.5 vs x21_fgen_clL_best_tr_K64_clip1.5 b=16 | regret 6.61% vs 6.60% | clL - spline median -0.03 pts [-0.80, +1.00] p_holm 0.95 | certified 96.0% vs 97.5%, diff -1.5 [-3.5, +0.5] McNemar p 0.38 | spline cheaper on 47.7%, dearer on 50.3%
30
+ x28_clS_s0_best_tr_K64_clip1.5 vs x21_fgen_clL_best_tr_K64_clip1.5 b=64 | regret 2.96% vs 4.07% | clL - spline median +0.44 pts [+0.16, +0.91] p_holm 7.5e-05 | certified 99.0% vs 98.5%, diff +0.5 [+0.0, +1.5] McNemar p 1 | spline cheaper on 62.3%, dearer on 36.7%
31
+ knots of the best b=16 candidate, x28_clS_s0_best_tr_K16_clip1.5: max |phi - sat_1| median 0.086, > 0.05 on 66.8% of plants, max |phi(x) + phi(-x)| median 0.066
32
+ knots of the best b=16 candidate, x28_clS_s1_best_tr_K16_clip1.5: max |phi - sat_1| median 0.083, > 0.05 on 62.8% of plants, max |phi(x) + phi(-x)| median 0.064
33
+ knots of the best b=16 candidate, x28_clS_s2_best_tr_K16_clip1.5: max |phi - sat_1| median 0.077, > 0.05 on 60.3% of plants, max |phi(x) + phi(-x)| median 0.055
results/icml_r3/X28/eval_test_report.txt ADDED
@@ -0,0 +1,62 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ == x28_clS_s0_best_tr_K16_clip1.5 (step 24000), 199 plants; sampling 5 s, scoring 944 s
2
+ certified per candidate 52.2%; candidates with knots off sat_1 100.0%
3
+ b= 1: certified 49.2% | regret vs J* median 100000.00% [91.98, 100000.00] | vs J_Fonly 100000.00% [79.81, 100000.00] | beats tuned linear 23.1%
4
+ b= 4: certified 85.9% | regret vs J* median 16.47% [12.15, 20.50] | vs J_Fonly 9.81% [6.84, 13.43] | beats tuned linear 60.3%
5
+ b=16: certified 95.0% | regret vs J* median 7.14% [5.60, 8.64] | vs J_Fonly 2.31% [1.60, 3.33] | beats tuned linear 86.9%
6
+ == x28_clS_s1_best_tr_K16_clip1.5 (step 26000), 199 plants; sampling 5 s, scoring 939 s
7
+ certified per candidate 52.1%; candidates with knots off sat_1 100.0%
8
+ b= 1: certified 52.8% | regret vs J* median 135.35% [62.33, 100000.00] | vs J_Fonly 123.46% [49.75, 100000.00] | beats tuned linear 31.2%
9
+ b= 4: certified 85.4% | regret vs J* median 15.09% [11.08, 20.07] | vs J_Fonly 9.06% [6.10, 12.85] | beats tuned linear 60.8%
10
+ b=16: certified 96.5% | regret vs J* median 6.33% [5.00, 7.42] | vs J_Fonly 2.07% [1.38, 2.84] | beats tuned linear 83.4%
11
+ == x28_clS_s2_best_tr_K16_clip1.5 (step 24000), 199 plants; sampling 5 s, scoring 898 s
12
+ certified per candidate 52.7%; candidates with knots off sat_1 100.0%
13
+ b= 1: certified 49.7% | regret vs J* median 100000.00% [92.40, 100000.00] | vs J_Fonly 100000.00% [82.90, 100000.00] | beats tuned linear 26.6%
14
+ b= 4: certified 83.4% | regret vs J* median 19.78% [14.31, 24.44] | vs J_Fonly 11.97% [8.56, 15.73] | beats tuned linear 57.3%
15
+ b=16: certified 95.5% | regret vs J* median 6.30% [5.67, 7.26] | vs J_Fonly 2.34% [1.70, 3.23] | beats tuned linear 84.4%
16
+ == x21_fgen_clL_best_tr_K16_clip1.5 (step 26000), 199 plants; sampling 6 s, scoring 407 s
17
+ certified per candidate 65.9%; candidates with knots off sat_1 0.0%
18
+ b= 1: certified 65.3% | regret vs J* median 54.90% [35.80, 97.49] | vs J_Fonly 40.21% [24.37, 69.08] | beats tuned linear 39.7%
19
+ b= 4: certified 92.0% | regret vs J* median 12.85% [10.18, 15.87] | vs J_Fonly 6.32% [4.91, 7.28] | beats tuned linear 73.4%
20
+ b=16: certified 98.5% | regret vs J* median 6.79% [5.57, 8.07] | vs J_Fonly 2.16% [1.83, 2.65] | beats tuned linear 87.4%
21
+ == x21_fgen_clL_s1_best_tr_K16_clip1.5 (step 26000), 199 plants; sampling 5 s, scoring 466 s
22
+ certified per candidate 65.1%; candidates with knots off sat_1 0.0%
23
+ b= 1: certified 65.3% | regret vs J* median 49.63% [33.95, 92.94] | vs J_Fonly 36.62% [19.29, 53.54] | beats tuned linear 41.7%
24
+ b= 4: certified 94.0% | regret vs J* median 13.25% [11.42, 16.88] | vs J_Fonly 6.56% [4.79, 8.06] | beats tuned linear 73.4%
25
+ b=16: certified 98.0% | regret vs J* median 6.62% [5.45, 8.06] | vs J_Fonly 2.07% [1.65, 2.48] | beats tuned linear 88.4%
26
+ == x21_fgen_clL_s2_best_tr_K16_clip1.5 (step 24000), 199 plants; sampling 5 s, scoring 319 s
27
+ certified per candidate 65.5%; candidates with knots off sat_1 0.0%
28
+ b= 1: certified 61.3% | regret vs J* median 42.49% [31.99, 115.60] | vs J_Fonly 32.65% [22.81, 78.49] | beats tuned linear 37.2%
29
+ b= 4: certified 89.9% | regret vs J* median 11.93% [10.03, 15.00] | vs J_Fonly 6.62% [5.47, 7.84] | beats tuned linear 69.3%
30
+ b=16: certified 97.5% | regret vs J* median 6.81% [5.59, 7.99] | vs J_Fonly 2.02% [1.67, 2.51] | beats tuned linear 87.4%
31
+ == x28_coefS_s0_best_tr_K16_clip3 (step 28000), 199 plants; sampling 5 s, scoring 1111 s
32
+ certified per candidate 49.7%; candidates with knots off sat_1 100.0%
33
+ b= 1: certified 48.7% | regret vs J* median 100000.00% [103.27, 100000.00] | vs J_Fonly 100000.00% [73.09, 100000.00] | beats tuned linear 21.6%
34
+ b= 4: certified 83.4% | regret vs J* median 21.51% [16.58, 30.32] | vs J_Fonly 19.05% [13.32, 23.13] | beats tuned linear 54.3%
35
+ b=16: certified 95.0% | regret vs J* median 10.58% [8.29, 12.93] | vs J_Fonly 6.04% [4.70, 8.22] | beats tuned linear 78.4%
36
+ == x28_clS_s0_best_K16_clip1 (step 24000), 199 plants; sampling 5 s, scoring 759 s
37
+ certified per candidate 47.2%; candidates with knots off sat_1 99.9%
38
+ b= 1: certified 43.7% | regret vs J* median 100000.00% [509.71, 100000.00] | vs J_Fonly 100000.00% [351.64, 100000.00] | beats tuned linear 25.1%
39
+ b= 4: certified 82.4% | regret vs J* median 26.65% [19.34, 33.41] | vs J_Fonly 17.62% [13.49, 23.73] | beats tuned linear 55.3%
40
+ b=16: certified 93.5% | regret vs J* median 10.57% [8.47, 12.13] | vs J_Fonly 5.56% [4.26, 7.62] | beats tuned linear 79.4%
41
+ == x28_clS_s1_best_K16_clip1.5 (step 26000), 199 plants; sampling 5 s, scoring 678 s
42
+ certified per candidate 36.8%; candidates with knots off sat_1 100.0%
43
+ b= 1: certified 35.7% | regret vs J* median 100000.00% [100000.00, 100000.00] | vs J_Fonly 100000.00% [100000.00, 100000.00] | beats tuned linear 22.1%
44
+ b= 4: certified 74.4% | regret vs J* median 24.18% [18.72, 30.83] | vs J_Fonly 15.82% [9.80, 21.77] | beats tuned linear 51.8%
45
+ b=16: certified 91.0% | regret vs J* median 7.37% [6.21, 9.00] | vs J_Fonly 3.23% [2.42, 4.50] | beats tuned linear 76.9%
46
+ == x28_clS_s2_best_K16_clip1 (step 24000), 199 plants; sampling 5 s, scoring 718 s
47
+ certified per candidate 46.9%; candidates with knots off sat_1 100.0%
48
+ b= 1: certified 44.2% | regret vs J* median 100000.00% [295.81, 100000.00] | vs J_Fonly 100000.00% [295.46, 100000.00] | beats tuned linear 26.6%
49
+ b= 4: certified 79.9% | regret vs J* median 33.60% [25.08, 42.93] | vs J_Fonly 22.60% [14.87, 32.28] | beats tuned linear 53.8%
50
+ b=16: certified 94.0% | regret vs J* median 10.28% [8.61, 12.43] | vs J_Fonly 5.42% [4.21, 7.19] | beats tuned linear 77.4%
51
+ == x28_clS_s0_best_tr_K64_clip1.5 (step 24000), 199 plants; sampling 18 s, scoring 3915 s
52
+ certified per candidate 51.2%; candidates with knots off sat_1 100.0%
53
+ b= 1: certified 51.8% | regret vs J* median 228.92% [93.75, 100000.00] | vs J_Fonly 227.68% [68.18, 100000.00] | beats tuned linear 29.6%
54
+ b= 4: certified 88.4% | regret vs J* median 14.50% [12.28, 17.42] | vs J_Fonly 9.04% [7.54, 11.41] | beats tuned linear 67.3%
55
+ b=16: certified 96.0% | regret vs J* median 6.61% [5.82, 7.35] | vs J_Fonly 2.28% [1.60, 3.18] | beats tuned linear 84.4%
56
+ b=64: certified 99.0% | regret vs J* median 2.96% [2.61, 3.51] | vs J_Fonly 0.45% [0.00, 0.75] | beats tuned linear 94.0%
57
+ == x21_fgen_clL_best_tr_K64_clip1.5 (step 26000), 199 plants; sampling 14 s, scoring 1299 s
58
+ certified per candidate 66.6%; candidates with knots off sat_1 0.0%
59
+ b= 1: certified 67.8% | regret vs J* median 53.11% [37.09, 83.62] | vs J_Fonly 35.18% [27.02, 74.13] | beats tuned linear 37.7%
60
+ b= 4: certified 91.0% | regret vs J* median 12.85% [11.10, 16.72] | vs J_Fonly 6.55% [4.76, 8.68] | beats tuned linear 70.4%
61
+ b=16: certified 97.5% | regret vs J* median 6.60% [5.88, 8.24] | vs J_Fonly 2.44% [1.86, 2.86] | beats tuned linear 86.9%
62
+ b=64: certified 98.5% | regret vs J* median 4.07% [3.54, 4.77] | vs J_Fonly 0.81% [0.57, 1.03] | beats tuned linear 95.0%
results/icml_r3/X28/eval_val/x28_clS_s0_best_tr_K16_clip1.5.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c3c8481b1318f690bf8ec523bc6bf123b8e2934d5f35cec0efc988d6e0c35f84
3
+ size 84662
results/icml_r3/X28/eval_val/x28_clS_s0_best_tr_K16_clip1.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b97bd547f57e43e87f7ee929db6f050552e9a4104362ebdc945fb78fcf76cc3c
3
+ size 84773
results/icml_r3/X28/eval_val/x28_clS_s0_best_tr_K16_clip2.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bad6516313bb689e0859b570b385c498eee08954cfb12b4b4378c0438d1d63a3
3
+ size 84597
results/icml_r3/X28/eval_val/x28_clS_s0_best_tr_K16_clip3.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f4dbff1fd7632c4f45506493142a7c20eba19d4aa13271cce6cd8b70d9290328
3
+ size 84667
results/icml_r3/X28/eval_val/x28_clS_s1_best_tr_K16_clip1.5.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2790e2334528a7e1fd347e2fe49f123dfcb908aa18e44743ac846fa4a7077cd6
3
+ size 85300
results/icml_r3/X28/eval_val/x28_clS_s1_best_tr_K16_clip1.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:28ec6d8eba839bb68503be32b9a9dbd8e16e594d568852f05564533617993774
3
+ size 85525
results/icml_r3/X28/eval_val/x28_clS_s1_best_tr_K16_clip2.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7275f33679eb9e86820e78cdb69a4f2668348544365bfc678c9378c86f8848ba
3
+ size 85138
results/icml_r3/X28/eval_val/x28_clS_s1_best_tr_K16_clip3.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:dbc31268e9a0cebf1d175202689a9f85b55260ff1ff6ea1d39324ce74253d8b3
3
+ size 85259
results/icml_r3/X28/eval_val/x28_clS_s2_best_tr_K16_clip1.5.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fc4a4d78bd89c1c1745680d58c41db5988cf6b16cfa6fbd15918d817a0e35bb7
3
+ size 84783
results/icml_r3/X28/eval_val/x28_clS_s2_best_tr_K16_clip1.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:269e46e92e307bcff82fdb5b5e3f48884ba8bc19749efaad7049ecbf43f4022f
3
+ size 84976
results/icml_r3/X28/eval_val/x28_clS_s2_best_tr_K16_clip2.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6b06a42dddc57a77247692e35fa94ca969557729d261de36f7f4c7a6d30a5750
3
+ size 84587
results/icml_r3/X28/eval_val/x28_clS_s2_best_tr_K16_clip3.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:44b1f4182338f9ca63c1c02a6c93f27aad6af8ae0a755dbfb742f6f4740ba990
3
+ size 84777
results/icml_r3/X28/eval_val/x28_coefS_s0_best_tr_K16_clip1.5.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5029c66fb3d58987bfa2e801ec5545b7ea63d106374f21a05b37758314872974
3
+ size 82275
results/icml_r3/X28/eval_val/x28_coefS_s0_best_tr_K16_clip1.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9f6bbe9969d0b036e8511b98b6525a3f6a6c84e7619792f5933ac326fcf729ac
3
+ size 78666
results/icml_r3/X28/eval_val/x28_coefS_s0_best_tr_K16_clip2.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:328715e2ab7cd958ee883f01c1ead8834dc8d9f97874f811b291d02b0d21d8fe
3
+ size 83609
results/icml_r3/X28/eval_val/x28_coefS_s0_best_tr_K16_clip3.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:587f8e35fce888e03f4436dbff11b9a3d8c6d3efb1afd0db83437ee363077bb4
3
+ size 84638
results/icml_r3/X28/eval_val_report.txt CHANGED
@@ -138,3 +138,83 @@
138
  b= 1: certified 62.5% | regret vs J* median 51.01% [18.44, 100000.00] | vs J_Fonly 35.53% [13.49, 100000.00] | beats tuned linear 47.5%
139
  b= 4: certified 92.5% | regret vs J* median 13.84% [7.81, 22.33] | vs J_Fonly 7.84% [4.72, 15.62] | beats tuned linear 67.5%
140
  b=16: certified 100.0% | regret vs J* median 7.12% [4.39, 10.44] | vs J_Fonly 2.19% [1.44, 4.58] | beats tuned linear 95.0%
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
138
  b= 1: certified 62.5% | regret vs J* median 51.01% [18.44, 100000.00] | vs J_Fonly 35.53% [13.49, 100000.00] | beats tuned linear 47.5%
139
  b= 4: certified 92.5% | regret vs J* median 13.84% [7.81, 22.33] | vs J_Fonly 7.84% [4.72, 15.62] | beats tuned linear 67.5%
140
  b=16: certified 100.0% | regret vs J* median 7.12% [4.39, 10.44] | vs J_Fonly 2.19% [1.44, 4.58] | beats tuned linear 95.0%
141
+ == x28_clS_s0_best_tr_K16_clip1 (step 24000), 40 plants; sampling 2 s, scoring 293 s
142
+ certified per candidate 60.5%; candidates with knots off sat_1 100.0%
143
+ b= 1: certified 55.0% | regret vs J* median 382.55% [92.19, 100000.00] | vs J_Fonly 362.53% [90.27, 100000.00] | beats tuned linear 27.5%
144
+ b= 4: certified 92.5% | regret vs J* median 25.29% [11.90, 33.24] | vs J_Fonly 17.79% [5.93, 28.58] | beats tuned linear 70.0%
145
+ b=16: certified 100.0% | regret vs J* median 10.83% [6.51, 16.31] | vs J_Fonly 4.84% [2.77, 9.43] | beats tuned linear 87.5%
146
+ == x28_clS_s0_best_tr_K16_clip1.5 (step 24000), 40 plants; sampling 2 s, scoring 353 s
147
+ certified per candidate 49.8%; candidates with knots off sat_1 100.0%
148
+ b= 1: certified 42.5% | regret vs J* median 100000.00% [119.31, 100000.00] | vs J_Fonly 100000.00% [113.17, 100000.00] | beats tuned linear 25.0%
149
+ b= 4: certified 87.5% | regret vs J* median 19.14% [13.44, 28.47] | vs J_Fonly 17.02% [8.91, 23.23] | beats tuned linear 62.5%
150
+ b=16: certified 100.0% | regret vs J* median 9.79% [5.46, 13.59] | vs J_Fonly 4.88% [3.38, 8.01] | beats tuned linear 80.0%
151
+ == x28_clS_s0_best_tr_K16_clip2 (step 24000), 40 plants; sampling 2 s, scoring 279 s
152
+ certified per candidate 46.1%; candidates with knots off sat_1 100.0%
153
+ b= 1: certified 40.0% | regret vs J* median 100000.00% [213.65, 100000.00] | vs J_Fonly 100000.00% [207.37, 100000.00] | beats tuned linear 17.5%
154
+ b= 4: certified 82.5% | regret vs J* median 28.09% [14.87, 55.97] | vs J_Fonly 25.39% [9.51, 47.75] | beats tuned linear 50.0%
155
+ b=16: certified 97.5% | regret vs J* median 9.70% [5.87, 13.47] | vs J_Fonly 4.64% [3.47, 7.66] | beats tuned linear 77.5%
156
+ == x28_clS_s0_best_tr_K16_clip3 (step 24000), 40 plants; sampling 2 s, scoring 294 s
157
+ certified per candidate 42.5%; candidates with knots off sat_1 100.0%
158
+ b= 1: certified 37.5% | regret vs J* median 100000.00% [399.86, 100000.00] | vs J_Fonly 100000.00% [389.72, 100000.00] | beats tuned linear 15.0%
159
+ b= 4: certified 77.5% | regret vs J* median 24.52% [14.29, 35.88] | vs J_Fonly 20.89% [12.63, 29.21] | beats tuned linear 50.0%
160
+ b=16: certified 97.5% | regret vs J* median 9.84% [6.39, 15.65] | vs J_Fonly 5.05% [3.76, 8.45] | beats tuned linear 77.5%
161
+ == x28_clS_s1_best_tr_K16_clip1 (step 26000), 40 plants; sampling 3 s, scoring 314 s
162
+ certified per candidate 60.3%; candidates with knots off sat_1 100.0%
163
+ b= 1: certified 65.0% | regret vs J* median 129.01% [47.40, 50243.95] | vs J_Fonly 126.61% [41.23, 50229.98] | beats tuned linear 32.5%
164
+ b= 4: certified 92.5% | regret vs J* median 24.19% [12.61, 36.32] | vs J_Fonly 15.04% [6.37, 29.74] | beats tuned linear 70.0%
165
+ b=16: certified 100.0% | regret vs J* median 12.00% [7.00, 15.72] | vs J_Fonly 5.20% [2.50, 10.85] | beats tuned linear 87.5%
166
+ == x28_clS_s1_best_tr_K16_clip1.5 (step 26000), 40 plants; sampling 2 s, scoring 337 s
167
+ certified per candidate 51.2%; candidates with knots off sat_1 100.0%
168
+ b= 1: certified 45.0% | regret vs J* median 100000.00% [69.67, 100000.00] | vs J_Fonly 100000.00% [65.51, 100000.00] | beats tuned linear 22.5%
169
+ b= 4: certified 95.0% | regret vs J* median 17.88% [10.15, 41.27] | vs J_Fonly 13.22% [7.53, 39.24] | beats tuned linear 60.0%
170
+ b=16: certified 100.0% | regret vs J* median 9.57% [5.74, 12.88] | vs J_Fonly 4.33% [2.53, 8.85] | beats tuned linear 90.0%
171
+ == x28_clS_s1_best_tr_K16_clip2 (step 26000), 40 plants; sampling 2 s, scoring 293 s
172
+ certified per candidate 46.4%; candidates with knots off sat_1 100.0%
173
+ b= 1: certified 37.5% | regret vs J* median 100000.00% [195.37, 100000.00] | vs J_Fonly 100000.00% [191.09, 100000.00] | beats tuned linear 15.0%
174
+ b= 4: certified 85.0% | regret vs J* median 17.56% [11.54, 42.77] | vs J_Fonly 15.80% [7.97, 40.06] | beats tuned linear 57.5%
175
+ b=16: certified 95.0% | regret vs J* median 9.08% [5.87, 14.64] | vs J_Fonly 5.15% [3.80, 8.98] | beats tuned linear 80.0%
176
+ == x28_clS_s1_best_tr_K16_clip3 (step 26000), 40 plants; sampling 2 s, scoring 304 s
177
+ certified per candidate 42.5%; candidates with knots off sat_1 100.0%
178
+ b= 1: certified 35.0% | regret vs J* median 100000.00% [50092.27, 100000.00] | vs J_Fonly 100000.00% [50085.86, 100000.00] | beats tuned linear 17.5%
179
+ b= 4: certified 90.0% | regret vs J* median 19.00% [10.67, 39.55] | vs J_Fonly 13.53% [8.30, 33.41] | beats tuned linear 62.5%
180
+ b=16: certified 97.5% | regret vs J* median 9.38% [5.51, 13.31] | vs J_Fonly 6.38% [3.35, 8.98] | beats tuned linear 80.0%
181
+ == x28_clS_s2_best_tr_K16_clip1 (step 24000), 40 plants; sampling 4 s, scoring 295 s
182
+ certified per candidate 61.3%; candidates with knots off sat_1 100.0%
183
+ b= 1: certified 62.5% | regret vs J* median 160.16% [73.22, 100000.00] | vs J_Fonly 157.45% [71.40, 100000.00] | beats tuned linear 32.5%
184
+ b= 4: certified 92.5% | regret vs J* median 26.81% [13.57, 38.50] | vs J_Fonly 16.55% [6.75, 30.69] | beats tuned linear 67.5%
185
+ b=16: certified 100.0% | regret vs J* median 12.62% [5.37, 19.79] | vs J_Fonly 3.88% [2.58, 13.60] | beats tuned linear 87.5%
186
+ == x28_clS_s2_best_tr_K16_clip1.5 (step 24000), 40 plants; sampling 2 s, scoring 322 s
187
+ certified per candidate 48.9%; candidates with knots off sat_1 100.0%
188
+ b= 1: certified 52.5% | regret vs J* median 368.79% [62.28, 100000.00] | vs J_Fonly 353.62% [59.85, 100000.00] | beats tuned linear 27.5%
189
+ b= 4: certified 90.0% | regret vs J* median 17.08% [9.03, 36.35] | vs J_Fonly 13.24% [6.10, 28.48] | beats tuned linear 62.5%
190
+ b=16: certified 100.0% | regret vs J* median 8.56% [3.82, 12.98] | vs J_Fonly 4.33% [2.00, 8.47] | beats tuned linear 77.5%
191
+ == x28_clS_s2_best_tr_K16_clip2 (step 24000), 40 plants; sampling 2 s, scoring 325 s
192
+ certified per candidate 43.6%; candidates with knots off sat_1 100.0%
193
+ b= 1: certified 37.5% | regret vs J* median 100000.00% [415.67, 100000.00] | vs J_Fonly 100000.00% [405.35, 100000.00] | beats tuned linear 17.5%
194
+ b= 4: certified 77.5% | regret vs J* median 26.27% [18.38, 51.95] | vs J_Fonly 23.15% [10.00, 46.89] | beats tuned linear 52.5%
195
+ b=16: certified 100.0% | regret vs J* median 12.54% [3.76, 16.87] | vs J_Fonly 6.44% [2.50, 11.09] | beats tuned linear 77.5%
196
+ == x28_clS_s2_best_tr_K16_clip3 (step 24000), 40 plants; sampling 2 s, scoring 280 s
197
+ certified per candidate 40.5%; candidates with knots off sat_1 100.0%
198
+ b= 1: certified 37.5% | regret vs J* median 100000.00% [343.82, 100000.00] | vs J_Fonly 100000.00% [334.93, 100000.00] | beats tuned linear 15.0%
199
+ b= 4: certified 80.0% | regret vs J* median 29.34% [16.08, 83.13] | vs J_Fonly 26.24% [12.19, 71.00] | beats tuned linear 47.5%
200
+ b=16: certified 95.0% | regret vs J* median 12.05% [4.23, 16.16] | vs J_Fonly 7.26% [2.51, 11.54] | beats tuned linear 72.5%
201
+ == x28_coefS_s0_best_tr_K16_clip1 (step 28000), 40 plants; sampling 2 s, scoring 187 s
202
+ certified per candidate 38.1%; candidates with knots off sat_1 100.0%
203
+ b= 1: certified 37.5% | regret vs J* median 100000.00% [180.00, 100000.00] | vs J_Fonly 100000.00% [161.88, 100000.00] | beats tuned linear 22.5%
204
+ b= 4: certified 50.0% | regret vs J* median 50126.73% [51.91, 100000.00] | vs J_Fonly 50113.26% [47.20, 100000.00] | beats tuned linear 32.5%
205
+ b=16: certified 67.5% | regret vs J* median 137.13% [22.05, 662.38] | vs J_Fonly 97.82% [14.90, 647.99] | beats tuned linear 37.5%
206
+ == x28_coefS_s0_best_tr_K16_clip1.5 (step 28000), 40 plants; sampling 2 s, scoring 186 s
207
+ certified per candidate 44.4%; candidates with knots off sat_1 100.0%
208
+ b= 1: certified 52.5% | regret vs J* median 148.36% [40.57, 100000.00] | vs J_Fonly 140.26% [39.68, 100000.00] | beats tuned linear 25.0%
209
+ b= 4: certified 65.0% | regret vs J* median 46.44% [21.01, 50229.23] | vs J_Fonly 39.68% [14.63, 50219.54] | beats tuned linear 45.0%
210
+ b=16: certified 80.0% | regret vs J* median 22.97% [9.87, 57.48] | vs J_Fonly 14.10% [4.40, 54.46] | beats tuned linear 60.0%
211
+ == x28_coefS_s0_best_tr_K16_clip2 (step 28000), 40 plants; sampling 2 s, scoring 280 s
212
+ certified per candidate 46.2%; candidates with knots off sat_1 100.0%
213
+ b= 1: certified 50.0% | regret vs J* median 50081.59% [49.10, 100000.00] | vs J_Fonly 50076.27% [44.56, 100000.00] | beats tuned linear 32.5%
214
+ b= 4: certified 75.0% | regret vs J* median 41.63% [17.19, 71.88] | vs J_Fonly 33.16% [9.69, 59.76] | beats tuned linear 60.0%
215
+ b=16: certified 87.5% | regret vs J* median 10.47% [6.23, 16.43] | vs J_Fonly 4.99% [3.15, 6.38] | beats tuned linear 80.0%
216
+ == x28_coefS_s0_best_tr_K16_clip3 (step 28000), 40 plants; sampling 2 s, scoring 258 s
217
+ certified per candidate 44.5%; candidates with knots off sat_1 100.0%
218
+ b= 1: certified 42.5% | regret vs J* median 100000.00% [66.37, 100000.00] | vs J_Fonly 100000.00% [63.51, 100000.00] | beats tuned linear 25.0%
219
+ b= 4: certified 75.0% | regret vs J* median 34.01% [13.29, 55.61] | vs J_Fonly 21.54% [8.30, 47.26] | beats tuned linear 62.5%
220
+ b=16: certified 90.0% | regret vs J* median 9.07% [6.62, 14.19] | vs J_Fonly 4.63% [2.86, 8.85] | beats tuned linear 85.0%
results/icml_r3/X28/mimo/fresh_study_m2.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4fd7aad4decf538d206e38e0fe4fc8a23979801da5da7467651363b5da635bf0
3
+ size 504450
results/icml_r3/X28/mimo/fresh_study_m2.txt ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ X28-M fresh directions, set study, m = 2: 64 random unit directions per plant, seed (7676, plant id)
2
+ sp: per plant median over directions of (J_id - J_sp) / J_sp; median over plants -2.11% [-4.80, -1.32], > 0 on 19 of 64; non-convergent rollouts id 0, sp 180
3
+ so: per plant median over directions of (J_id - J_so) / J_so; median over plants 3.11% [2.48, 4.11], > 0 on 55 of 64; non-convergent rollouts id 0, so 0
4
+ so3: per plant median over directions of (J_id - J_so3) / J_so3; median over plants 3.99% [2.86, 5.13], > 0 on 56 of 64; non-convergent rollouts id 0, so3 0
5
+ cf: per plant median over directions of (J_id - J_cf) / J_cf; median over plants -3.09% [-4.50, -2.32], > 0 on 15 of 64; non-convergent rollouts id 0, cf 43
6
+ co: per plant median over directions of (J_id - J_co) / J_co; median over plants 2.75% [2.13, 3.45], > 0 on 54 of 64; non-convergent rollouts id 0, co 1
results/icml_r3/X28/mimo/fresh_study_m3.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:97614f56288c2dddc9881663b7cd23e66f709f32c27b88163baf9f694b2c0cbc
3
+ size 565124
results/icml_r3/X28/mimo/fresh_study_m3.txt ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ X28-M fresh directions, set study, m = 3: 64 random unit directions per plant, seed (7676, plant id)
2
+ sp: per plant median over directions of (J_id - J_sp) / J_sp; median over plants -5.94% [-8.80, -3.77], > 0 on 8 of 64; non-convergent rollouts id 0, sp 188
3
+ so: per plant median over directions of (J_id - J_so) / J_so; median over plants 1.32% [0.64, 1.94], > 0 on 47 of 64; non-convergent rollouts id 0, so 0
4
+ so3: per plant median over directions of (J_id - J_so3) / J_so3; median over plants 1.82% [0.74, 2.57], > 0 on 50 of 64; non-convergent rollouts id 0, so3 0
5
+ cf: per plant median over directions of (J_id - J_cf) / J_cf; median over plants -6.06% [-8.40, -4.46], > 0 on 8 of 64; non-convergent rollouts id 0, cf 13
6
+ co: per plant median over directions of (J_id - J_co) / J_co; median over plants 0.91% [0.45, 1.51], > 0 on 47 of 64; non-convergent rollouts id 0, co 0
results/icml_r3/X28/mimo/gate_study_m2.txt ADDED
@@ -0,0 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ X28-M gate, set study, m = 2, n = 64, id certified on 64, median J_id 11.8
2
+ id: snapshot 0 kept on 0, last snapshot on 15
3
+ piecewise linear, free (sp, 10 per channel): d = (J_id - J_spline) / J_spline median 6.61% [5.46, 7.52], >= 2% on 60 of 64, mean 7.57% -> PAYS
4
+ start sp_c1: certified 64, snapshot 0 kept on 0, median d 5.48%
5
+ start sp_c2: certified 64, snapshot 0 kept on 1, median d 6.13%
6
+ start sp_c5: certified 64, snapshot 0 kept on 0, median d 4.85%
7
+ winning splines 64: max beta / alpha of a channel median 5.411; max |log r(-x) / r(x)| median 1.346; rollouts with final |x| > 1e-2: 0
8
+ piecewise linear, odd (so, 5 per channel): d = (J_id - J_spline) / J_spline median 5.40% [4.48, 6.41], >= 2% on 53 of 64, mean 6.02% -> PAYS
9
+ start so_c1: certified 64, snapshot 0 kept on 0, median d 4.02%
10
+ start so_c2: certified 64, snapshot 0 kept on 0, median d 5.08%
11
+ start so_c5: certified 64, snapshot 0 kept on 0, median d 3.71%
12
+ winning splines 64: max beta / alpha of a channel median 3.728; max |log r(-x) / r(x)| median 0.000; rollouts with final |x| > 1e-2: 0
13
+ piecewise linear, odd, knots at |x| = 1, 3, 5 (so3, 3 per channel): d = (J_id - J_spline) / J_spline median 5.22% [4.37, 6.29], >= 2% on 53 of 64, mean 5.57% -> PAYS
14
+ start so3_c1: certified 64, snapshot 0 kept on 0, median d 4.36%
15
+ start so3_c2: certified 64, snapshot 0 kept on 0, median d 4.86%
16
+ start so3_c5: certified 64, snapshot 0 kept on 0, median d 3.99%
17
+ winning splines 64: max beta / alpha of a channel median 3.437; max |log r(-x) / r(x)| median 0.000; rollouts with final |x| > 1e-2: 0
18
+ cubic (PCHIP ratio), free (cf, 10 per channel): d = (J_id - J_spline) / J_spline median 6.39% [5.07, 7.52], >= 2% on 58 of 64, mean 7.47% -> PAYS
19
+ start cf_c1: certified 64, snapshot 0 kept on 1, median d 5.21%
20
+ start cf_c2: certified 64, snapshot 0 kept on 1, median d 5.87%
21
+ start cf_c5: certified 64, snapshot 0 kept on 0, median d 4.90%
22
+ winning splines 64: max beta / alpha of a channel median 5.454; max |log r(-x) / r(x)| median 1.198; rollouts with final |x| > 1e-2: 0
23
+ cubic (PCHIP ratio), odd (co, 5 per channel): d = (J_id - J_spline) / J_spline median 5.37% [4.41, 6.24], >= 2% on 53 of 64, mean 5.67% -> PAYS
24
+ start co_c1: certified 64, snapshot 0 kept on 0, median d 4.09%
25
+ start co_c2: certified 64, snapshot 0 kept on 0, median d 4.70%
26
+ start co_c5: certified 64, snapshot 0 kept on 0, median d 3.67%
27
+ winning splines 64: max beta / alpha of a channel median 3.671; max |log r(-x) / r(x)| median 0.000; rollouts with final |x| > 1e-2: 0
results/icml_r3/X28/mimo/gate_study_m3.txt ADDED
@@ -0,0 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ X28-M gate, set study, m = 3, n = 64, id certified on 64, median J_id 13.19
2
+ id: snapshot 0 kept on 0, last snapshot on 15
3
+ piecewise linear, free (sp, 10 per channel): d = (J_id - J_spline) / J_spline median 6.31% [4.87, 7.24], >= 2% on 55 of 64, mean 6.41% -> PAYS
4
+ start sp_c1: certified 64, snapshot 0 kept on 0, median d 4.62%
5
+ start sp_c2: certified 64, snapshot 0 kept on 0, median d 5.46%
6
+ start sp_c5: certified 64, snapshot 0 kept on 0, median d 4.83%
7
+ winning splines 62: max beta / alpha of a channel median 6.247; max |log r(-x) / r(x)| median 1.389; rollouts with final |x| > 1e-2: 0
8
+ piecewise linear, odd (so, 5 per channel): d = (J_id - J_spline) / J_spline median 4.80% [4.05, 5.97], >= 2% on 51 of 64, mean 5.47% -> PAYS
9
+ start so_c1: certified 64, snapshot 0 kept on 0, median d 4.01%
10
+ start so_c2: certified 64, snapshot 0 kept on 0, median d 4.44%
11
+ start so_c5: certified 64, snapshot 0 kept on 0, median d 3.35%
12
+ winning splines 64: max beta / alpha of a channel median 4.856; max |log r(-x) / r(x)| median 0.000; rollouts with final |x| > 1e-2: 0
13
+ piecewise linear, odd, knots at |x| = 1, 3, 5 (so3, 3 per channel): d = (J_id - J_spline) / J_spline median 4.54% [4.12, 5.59], >= 2% on 51 of 64, mean 5.24% -> PAYS
14
+ start so3_c1: certified 64, snapshot 0 kept on 0, median d 4.06%
15
+ start so3_c2: certified 64, snapshot 0 kept on 0, median d 4.20%
16
+ start so3_c5: certified 64, snapshot 0 kept on 0, median d 3.41%
17
+ winning splines 64: max beta / alpha of a channel median 4.361; max |log r(-x) / r(x)| median 0.000; rollouts with final |x| > 1e-2: 0
18
+ cubic (PCHIP ratio), free (cf, 10 per channel): d = (J_id - J_spline) / J_spline median 5.39% [4.77, 6.42], >= 2% on 53 of 64, mean 5.90% -> PAYS
19
+ start cf_c1: certified 64, snapshot 0 kept on 0, median d 4.50%
20
+ start cf_c2: certified 64, snapshot 0 kept on 1, median d 5.12%
21
+ start cf_c5: certified 64, snapshot 0 kept on 0, median d 4.10%
22
+ winning splines 61: max beta / alpha of a channel median 5.582; max |log r(-x) / r(x)| median 1.225; rollouts with final |x| > 1e-2: 0
23
+ cubic (PCHIP ratio), odd (co, 5 per channel): d = (J_id - J_spline) / J_spline median 5.02% [4.30, 5.81], >= 2% on 50 of 64, mean 5.59% -> PAYS
24
+ start co_c1: certified 64, snapshot 0 kept on 0, median d 3.79%
25
+ start co_c2: certified 64, snapshot 0 kept on 0, median d 4.53%
26
+ start co_c5: certified 64, snapshot 0 kept on 0, median d 3.14%
27
+ winning splines 62: max beta / alpha of a channel median 5.394; max |log r(-x) / r(x)| median 0.000; rollouts with final |x| > 1e-2: 0
results/icml_r3/X28/mimo/study_choice.txt ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ X28-M spline-class study, selection rule of PREREG addendum 2026-11-13 (64 validation plants per m, fresh seed (7676, id))
2
+ sp piecewise linear, free params/ch 10 | m=2: f -2.11% [-4.80, -1.32], non-conv excess 180 | m=3: f -5.94% [-8.80, -3.77], non-conv excess 188 | mean f -4.02% | not eligible
3
+ so piecewise linear, odd params/ch 5 | m=2: f +3.11% [+2.48, +4.11], non-conv excess 0 | m=3: f +1.32% [+0.64, +1.94], non-conv excess 0 | mean f +2.22% | eligible
4
+ so3 piecewise linear, odd, knots at |x| = 1, 3, 5 params/ch 3 | m=2: f +3.99% [+2.86, +5.13], non-conv excess 0 | m=3: f +1.82% [+0.74, +2.57], non-conv excess 0 | mean f +2.90% | eligible
5
+ cf cubic (PCHIP ratio), free params/ch 10 | m=2: f -3.09% [-4.50, -2.32], non-conv excess 43 | m=3: f -6.06% [-8.40, -4.46], non-conv excess 13 | mean f -4.58% | not eligible
6
+ co cubic (PCHIP ratio), odd params/ch 5 | m=2: f +2.75% [+2.13, +3.45], non-conv excess 1 | m=3: f +0.91% [+0.45, +1.51], non-conv excess 0 | mean f +1.83% | eligible
7
+ highest mean f: so3 (+2.90%); within 0.25 points: so3
8
+ CHOICE: so3 (piecewise linear, odd, knots at |x| = 1, 3, 5, 3 parameters per channel)
results/icml_r3/X28/mimo/study_m2/cert_cf_c1.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:dae6840d2c16bab58f6068905e2845ccc19e5874f8d7a003891f231c882464f0
3
+ size 13157
results/icml_r3/X28/mimo/study_m2/cert_cf_c2.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:193b78e354d2b3b8df683c0309b4b691a73de2b0d093eb54e52eb5b02e9ab343
3
+ size 12506
results/icml_r3/X28/mimo/study_m2/cert_cf_c5.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7c81b65b2bda71cab5d8aa859d8fb246b657066333be57304a8adeafbeb80040
3
+ size 10662
results/icml_r3/X28/mimo/study_m2/cert_co_c1.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9f5aed427832a7e28df22bbe0e437889d73f7a7224fd7f7fe5b743ee563a3e91
3
+ size 10428
results/icml_r3/X28/mimo/study_m2/cert_co_c2.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f660b2e230916a4b26409df0ec51f17fff36b2b494a90a52e5dd8c2a0f833531
3
+ size 10313
results/icml_r3/X28/mimo/study_m2/cert_co_c5.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a475533fb570be544e66ebd696b6a6d8219fae7c368609527e7ff56b4ccb37a5
3
+ size 9431
results/icml_r3/X28/mimo/study_m2/cert_id_c1.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a65eaa8b6036193a293eade0fc030012c984d9323253634c7561897c59c308d0
3
+ size 4640
results/icml_r3/X28/mimo/study_m2/cert_so3_c1.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1486b8be5691efc688b16d9a1c4738f33baef7da5d907bf2c2f5fd0a03574255
3
+ size 10442
results/icml_r3/X28/mimo/study_m2/cert_so3_c2.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c266bd9aceb424566dd1815186ac9949a2f5a66f0389a8cf7fd16a899c83c1a0
3
+ size 10363