Update DSM to reduced comp complexity
#21
by odunbar - opened
- .cache/known_methods_snapshot.json +15 -2
- data/opt_dsm_score_adam/{leaderboard_sbadam_dsm_parallel_l96_flux-force_2026-09-14.nc → leaderboard_sbadam_dsm_parallel_l96_flux-force_2026-11-04.nc} +1 -1
- data/opt_dsm_score_adam/{leaderboard_sbadam_dsm_parallel_l96_vec-force_2026-09-14.nc → leaderboard_sbadam_dsm_parallel_l96_vec-force_2026-11-02.nc} +1 -1
- data/opt_dsm_score_lm/{leaderboard_sblm_dsm_parallel_l96_flux-force_2026-08-12.nc → leaderboard_sblm_dsm_parallel_l96_flux-force_2026-11-12.nc} +2 -2
- data/opt_dsm_score_lm/{leaderboard_sblm_dsm_parallel_l96_vec-force_2026-08-12.nc → leaderboard_sblm_dsm_parallel_l96_vec-force_2026-11-12.nc} +2 -2
- src/data_store.py +4 -4
- src/pages/MethodDetails.py +34 -16
.cache/known_methods_snapshot.json
CHANGED
|
@@ -279,7 +279,7 @@
|
|
| 279 |
},
|
| 280 |
"eks": {
|
| 281 |
"abbreviation": "EKS",
|
| 282 |
-
"Method": "Ensemble Kalman Sampler",
|
| 283 |
"parallelism": "parallel-interacting",
|
| 284 |
"update_type": "gradient (Kalman)",
|
| 285 |
"method_goal": "uq",
|
|
@@ -291,7 +291,7 @@
|
|
| 291 |
},
|
| 292 |
"ces-eks": {
|
| 293 |
"abbreviation": "CES-EKS",
|
| 294 |
-
"Method": "Calibrate Emulate Sample (
|
| 295 |
"parallelism": "parallel-interacting",
|
| 296 |
"update_type": "gradient (Kalman)",
|
| 297 |
"method_goal": "uq",
|
|
@@ -311,6 +311,19 @@
|
|
| 311 |
"ies-eks",
|
| 312 |
"boed-eks"
|
| 313 |
]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 314 |
}
|
| 315 |
},
|
| 316 |
"observed_methods": [
|
|
|
|
| 279 |
},
|
| 280 |
"eks": {
|
| 281 |
"abbreviation": "EKS",
|
| 282 |
+
"Method": "Ensemble Kalman Sampler (EKS a.k.a ALDI)",
|
| 283 |
"parallelism": "parallel-interacting",
|
| 284 |
"update_type": "gradient (Kalman)",
|
| 285 |
"method_goal": "uq",
|
|
|
|
| 291 |
},
|
| 292 |
"ces-eks": {
|
| 293 |
"abbreviation": "CES-EKS",
|
| 294 |
+
"Method": "Calibrate Emulate Sample (EKS)",
|
| 295 |
"parallelism": "parallel-interacting",
|
| 296 |
"update_type": "gradient (Kalman)",
|
| 297 |
"method_goal": "uq",
|
|
|
|
| 311 |
"ies-eks",
|
| 312 |
"boed-eks"
|
| 313 |
]
|
| 314 |
+
},
|
| 315 |
+
"gboed-eks": {
|
| 316 |
+
"abbreviation": "GBOED-EKS",
|
| 317 |
+
"Method": "Goal-oriented Bayesian Optimal Experimental Design (EKS)",
|
| 318 |
+
"parallelism": "parallel-interacting",
|
| 319 |
+
"update_type": "general",
|
| 320 |
+
"method_goal": "uq",
|
| 321 |
+
"emulator_use": "within-optimize",
|
| 322 |
+
"aliases": [
|
| 323 |
+
"gboed-eks",
|
| 324 |
+
"gboed-eks-eig",
|
| 325 |
+
"boed-eks-eig"
|
| 326 |
+
]
|
| 327 |
}
|
| 328 |
},
|
| 329 |
"observed_methods": [
|
data/opt_dsm_score_adam/{leaderboard_sbadam_dsm_parallel_l96_flux-force_2026-09-14.nc → leaderboard_sbadam_dsm_parallel_l96_flux-force_2026-11-04.nc}
RENAMED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
size 15752
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:7c94130718fc17fb117d7f8be86f284321df1855aead5e4460bfcf49368cdb7c
|
| 3 |
size 15752
|
data/opt_dsm_score_adam/{leaderboard_sbadam_dsm_parallel_l96_vec-force_2026-09-14.nc → leaderboard_sbadam_dsm_parallel_l96_vec-force_2026-11-02.nc}
RENAMED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
size 15752
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:35427ab4b7669ee97df5c8e344b08f8371151e2a2e1401a155d39a2c7faa6ed9
|
| 3 |
size 15752
|
data/opt_dsm_score_lm/{leaderboard_sblm_dsm_parallel_l96_flux-force_2026-08-12.nc → leaderboard_sblm_dsm_parallel_l96_flux-force_2026-11-12.nc}
RENAMED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:1f4bfc91dff5e2da8d51095fae7f25080e05465425a7a08662ac04d47be37674
|
| 3 |
+
size 15752
|
data/opt_dsm_score_lm/{leaderboard_sblm_dsm_parallel_l96_vec-force_2026-08-12.nc → leaderboard_sblm_dsm_parallel_l96_vec-force_2026-11-12.nc}
RENAMED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:5c4ab848a4546fe0ade4c79749b868af845796fb0531805f4f07a73ed9880d7e
|
| 3 |
+
size 15752
|
src/data_store.py
CHANGED
|
@@ -68,9 +68,9 @@ DATASET_FILES = {
|
|
| 68 |
"adam_results/leaderboard_adam_l96_flux-force_2026-10-21.nc",
|
| 69 |
"levenberg_marquardt_results/leaderboard_lm_l96_flux-force_2026-10-20.nc",
|
| 70 |
"opt_kgmm_score_lm/leaderboard_sblm_kgmm_parallel_l96_flux-force_2026-09-02.nc",
|
| 71 |
-
"opt_dsm_score_lm/leaderboard_sblm_dsm_parallel_l96_flux-force_2026-
|
| 72 |
"opt_kgmm_score_adam/leaderboard_sbadam_kgmm_parallel_l96_flux-force_2026-09-11.nc",
|
| 73 |
-
"opt_dsm_score_adam/leaderboard_sbadam_dsm_parallel_l96_flux-force_2026-
|
| 74 |
"rl-ddpg/leaderboard_rl-ddpg-short_l96_flux-force_2026-10-04.nc",
|
| 75 |
"rl-ddpg/leaderboard_rl-ddpg-long_l96_flux-force_2026-10-04.nc",
|
| 76 |
"rl-td3/leaderboard_rl-td3-short_l96_flux-force_2026-10-04.nc",
|
|
@@ -84,9 +84,9 @@ DATASET_FILES = {
|
|
| 84 |
"adam_results/leaderboard_adam_l96_vec-force_2026-06-26.nc",
|
| 85 |
"levenberg_marquardt_results/leaderboard_lm_l96_vec-force_2026-06-29.nc",
|
| 86 |
"opt_kgmm_score_lm/leaderboard_sblm_kgmm_parallel_l96_vec-force_2026-09-02.nc",
|
| 87 |
-
"opt_dsm_score_lm/leaderboard_sblm_dsm_parallel_l96_vec-force_2026-
|
| 88 |
"opt_kgmm_score_adam/leaderboard_sbadam_kgmm_parallel_l96_vec-force_2026-09-11.nc",
|
| 89 |
-
"opt_dsm_score_adam/leaderboard_sbadam_dsm_parallel_l96_vec-force_2026-
|
| 90 |
"rl-ddpg/leaderboard_rl-ddpg-short_l96_vec-force_2026-10-04.nc",
|
| 91 |
"rl-ddpg/leaderboard_rl-ddpg-long_l96_vec-force_2026-10-04.nc",
|
| 92 |
"rl-td3/leaderboard_rl-td3-short_l96_vec-force_2026-10-04.nc",
|
|
|
|
| 68 |
"adam_results/leaderboard_adam_l96_flux-force_2026-10-21.nc",
|
| 69 |
"levenberg_marquardt_results/leaderboard_lm_l96_flux-force_2026-10-20.nc",
|
| 70 |
"opt_kgmm_score_lm/leaderboard_sblm_kgmm_parallel_l96_flux-force_2026-09-02.nc",
|
| 71 |
+
"opt_dsm_score_lm/leaderboard_sblm_dsm_parallel_l96_flux-force_2026-11-12.nc",
|
| 72 |
"opt_kgmm_score_adam/leaderboard_sbadam_kgmm_parallel_l96_flux-force_2026-09-11.nc",
|
| 73 |
+
"opt_dsm_score_adam/leaderboard_sbadam_dsm_parallel_l96_flux-force_2026-11-04.nc",
|
| 74 |
"rl-ddpg/leaderboard_rl-ddpg-short_l96_flux-force_2026-10-04.nc",
|
| 75 |
"rl-ddpg/leaderboard_rl-ddpg-long_l96_flux-force_2026-10-04.nc",
|
| 76 |
"rl-td3/leaderboard_rl-td3-short_l96_flux-force_2026-10-04.nc",
|
|
|
|
| 84 |
"adam_results/leaderboard_adam_l96_vec-force_2026-06-26.nc",
|
| 85 |
"levenberg_marquardt_results/leaderboard_lm_l96_vec-force_2026-06-29.nc",
|
| 86 |
"opt_kgmm_score_lm/leaderboard_sblm_kgmm_parallel_l96_vec-force_2026-09-02.nc",
|
| 87 |
+
"opt_dsm_score_lm/leaderboard_sblm_dsm_parallel_l96_vec-force_2026-11-12.nc",
|
| 88 |
"opt_kgmm_score_adam/leaderboard_sbadam_kgmm_parallel_l96_vec-force_2026-09-11.nc",
|
| 89 |
+
"opt_dsm_score_adam/leaderboard_sbadam_dsm_parallel_l96_vec-force_2026-11-02.nc",
|
| 90 |
"rl-ddpg/leaderboard_rl-ddpg-short_l96_vec-force_2026-10-04.nc",
|
| 91 |
"rl-ddpg/leaderboard_rl-ddpg-long_l96_vec-force_2026-10-04.nc",
|
| 92 |
"rl-td3/leaderboard_rl-td3-short_l96_vec-force_2026-10-04.nc",
|
src/pages/MethodDetails.py
CHANGED
|
@@ -422,9 +422,16 @@ method_meta = {
|
|
| 422 |
"benchmark rather than annealed, chosen empirically as the smallest noise level at which the "
|
| 423 |
"network still receives a usable learning signal. Same Stein recalibration and collapsed-window "
|
| 424 |
"`ForwardDiff.jl` fallback as LM-KGMM; same Tukey-tapered response integral and evaluation-"
|
| 425 |
-
"budget cap."
|
|
|
|
|
|
|
|
|
|
|
|
|
| 426 |
),
|
| 427 |
"params": [
|
|
|
|
|
|
|
|
|
|
| 428 |
{"Parameter": "σ (DSM noise scale)", "Value": "0.2 (L63) / 0.025 (L96 variants)"},
|
| 429 |
{"Parameter": "tukey_alpha (response taper)", "Value": "0.75"},
|
| 430 |
{"Parameter": "budget_total (outer_iter × N_ens)", "Value": "500"},
|
|
@@ -449,8 +456,8 @@ method_meta = {
|
|
| 449 |
"holds on the calibration ensemble, `kgmm_ridge` shrinks per-cluster covariances toward the "
|
| 450 |
"global covariance, and `is_collapsed_window` falls back to a direct `ForwardDiff.jl` gradient "
|
| 451 |
"on windows that have collapsed onto a fixed point; total cost is capped at "
|
| 452 |
-
"`outer_iter × N_ens ≤ budget_total` per cell,
|
| 453 |
-
"iteration count under a fixed evaluation budget; beyond that, Adam's own adaptive per-parameter "
|
| 454 |
"step scaling is the only guardrail on the step itself. Unlike `β₁`/`β₂`/`ε` (shared across all "
|
| 455 |
"cases), the step size `α` is set per-experiment from the case's prior — sized to ~0.25× the "
|
| 456 |
"mean prior standard deviation, so a run of consistent-sign steps crosses ~1 prior std in ~4 "
|
|
@@ -462,7 +469,8 @@ method_meta = {
|
|
| 462 |
{"Parameter": "β₁, β₂", "Value": "0.9, 0.999"},
|
| 463 |
{"Parameter": "kgmm_ridge", "Value": "0.1"},
|
| 464 |
{"Parameter": "tukey_alpha (response taper)", "Value": "0.75"},
|
| 465 |
-
{"Parameter": "budget_total (outer_iter × N_ens)", "Value": "
|
|
|
|
| 466 |
],
|
| 467 |
"notes": "**Inherits Adam's local-minimum susceptibility** (no trust region, restart, or ensemble diversity to escape a bad basin) **and LM-KGMM's cheaper-than-DSM score estimation** — the k-means/Gaussian-mixture score fit is far less costly than training a neural score network, so ADAM-KGMM stayed tractable on problems where a DSM-based Adam variant would not.",
|
| 468 |
},
|
|
@@ -477,17 +485,27 @@ method_meta = {
|
|
| 477 |
"density's score (Vincent, 2011); `σ` is fixed per benchmark rather than annealed, chosen "
|
| 478 |
"empirically as the smallest noise level at which the network still receives a usable learning "
|
| 479 |
"signal. Same Stein recalibration and collapsed-window `ForwardDiff.jl` fallback, and the same "
|
| 480 |
-
"Tukey-tapered response integral and evaluation-budget cap as
|
|
|
|
| 481 |
"step size `α` (shared `β₁`/`β₂`/`ε`) as ADAM-KGMM — sized to ~0.25× the case's prior standard "
|
| 482 |
"deviation, since a single global value tuned for one benchmark's residual scaling left Adam "
|
| 483 |
-
"unable to traverse even the L63/L96-const prior within its iteration budget."
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 484 |
),
|
| 485 |
"params": [
|
| 486 |
{"Parameter": "α (step size)", "Value": "Per-case, ~0.25× prior std (0.025–0.75 across benchmarks)"},
|
| 487 |
{"Parameter": "β₁, β₂", "Value": "0.9, 0.999"},
|
|
|
|
|
|
|
|
|
|
| 488 |
{"Parameter": "σ (DSM noise scale)", "Value": "0.2 (L63) / 0.025 (L96 variants)"},
|
| 489 |
{"Parameter": "tukey_alpha (response taper)", "Value": "0.75"},
|
| 490 |
-
{"Parameter": "budget_total (outer_iter × N_ens)", "Value": "
|
|
|
|
| 491 |
],
|
| 492 |
"notes": "**Inherits both LM-DSM's slow DSM-training cost and Adam's local-minimum susceptibility** — training the DSM score network is the binding wall-clock cost (as with LM-DSM), and Adam's lack of a trust region or ensemble diversity leaves it as prone to getting stuck in a bad basin as ADAM-KGMM.",
|
| 493 |
},
|
|
@@ -735,9 +753,9 @@ method_meta = {
|
|
| 735 |
),
|
| 736 |
},
|
| 737 |
"IES-EKS": {
|
| 738 |
-
"citation": "Elsaesser et al., J. Adv. Model. Earth Syst., 2025; Garbuno-Inigo,
|
| 739 |
"url": "https://doi.org/10.1029/2024MS004713",
|
| 740 |
-
"source": "`uq_experiments/BayesianOptimalExperimentalDesign` — `boed_core.jl` (`run_eks`, `acquire_batch`), variant `:eks` (leaderboard key `boed-eks`); `
|
| 741 |
"math": (
|
| 742 |
"The same iterated emulate-sample loop as IES-IEKF and IES-TMCMC (see those entries), with the "
|
| 743 |
"posterior sampler swapped for the Ensemble Kalman Sampler (EKP `Sampler`). Iteration 1 "
|
|
@@ -749,8 +767,8 @@ method_meta = {
|
|
| 749 |
"subsample of that ensemble."
|
| 750 |
),
|
| 751 |
"guardrails": (
|
| 752 |
-
"The sampler is `EKP.Sampler`
|
| 753 |
-
"
|
| 754 |
"`EKSStableScheduler(alpha = 0.25)` for a fixed 500 steps. `alpha = 0.25` was chosen because it "
|
| 755 |
"sat closer to the ST-MCMC reference than the default 1, and `alpha >= 3` is unstable. Like the IEKF "
|
| 756 |
"variant it ignores the GP predictive variance. GP fits use the same data-relative hyperparameter "
|
|
@@ -758,7 +776,7 @@ method_meta = {
|
|
| 758 |
"shortfall is padded with NaN rows that the leaderboard stage drops."
|
| 759 |
),
|
| 760 |
"params": [
|
| 761 |
-
{"Parameter": "sampler_type", "Value": "\"
|
| 762 |
{"Parameter": "EKS particles (eks_n_particles)", "Value": "200"},
|
| 763 |
{"Parameter": "scheduler", "Value": "EKSStableScheduler(alpha = 0.25)"},
|
| 764 |
{"Parameter": "steps (eks_max_iters / eks_T)", "Value": "500 / Inf (fixed step count)"},
|
|
@@ -773,9 +791,9 @@ method_meta = {
|
|
| 773 |
),
|
| 774 |
},
|
| 775 |
"GBOED-EKS": {
|
| 776 |
-
"citation": "Holthuijzen et al. 2026 (EIG acquisition); Garbuno-Inigo,
|
| 777 |
"url": "https://arxiv.org/abs/2508.13071",
|
| 778 |
-
"source": "`uq_experiments/BayesianOptimalExperimentalDesign` — `boed_core.jl` (`run_eks`, `optimize_batch`, `acquire_batch`), variant `:eks_eig` (leaderboard key `boed-eks-eig`); `
|
| 779 |
"math": (
|
| 780 |
"GBOED with the posterior sampler swapped for the Ensemble Kalman Sampler (see the GBOED-IEKF and "
|
| 781 |
"IES-EKS entries). Iteration 1 forward-evaluates an `N_ens`-point Latin-hypercube design in the "
|
|
@@ -788,7 +806,7 @@ method_meta = {
|
|
| 788 |
),
|
| 789 |
"math_eq": r"\mathrm{EIG}(X) = \frac{1}{k}\sum_j \frac{1}{2}\log\frac{\det K_{pp}^{(j)}}{\det \Sigma'^{(j)}}, \quad \Sigma'^{(j)} = K_{pp}^{(j)} - K_{pc}^{(j)}\left(K_{cc}^{(j)}\right)^{-1}K_{cp}^{(j)}",
|
| 790 |
"guardrails": (
|
| 791 |
-
"The sampler is `EKP.Sampler` with `sampler_type = \"
|
| 792 |
"fixed 500 steps of `EKSStableScheduler(alpha = 0.25)` with 200 particles, ignoring the GP "
|
| 793 |
"predictive variance. The batch is optimized with `Optim.jl`'s `LBFGS` inside the χ² ball of the "
|
| 794 |
"prior-whitened space (radial `tanh` squash, so every candidate has radius below √χ²_k(0.999)), with "
|
|
@@ -799,7 +817,7 @@ method_meta = {
|
|
| 799 |
"caps as IES-IEKF."
|
| 800 |
),
|
| 801 |
"params": [
|
| 802 |
-
{"Parameter": "sampler_type", "Value": "\"
|
| 803 |
{"Parameter": "EKS particles (eks_n_particles)", "Value": "200"},
|
| 804 |
{"Parameter": "scheduler", "Value": "EKSStableScheduler(alpha = 0.25)"},
|
| 805 |
{"Parameter": "steps (eks_max_iters / eks_T)", "Value": "500 / Inf (fixed step count)"},
|
|
|
|
| 422 |
"benchmark rather than annealed, chosen empirically as the smallest noise level at which the "
|
| 423 |
"network still receives a usable learning signal. Same Stein recalibration and collapsed-window "
|
| 424 |
"`ForwardDiff.jl` fallback as LM-KGMM; same Tukey-tapered response integral and evaluation-"
|
| 425 |
+
"budget cap. **Score model & training:** a small network (MLP for L63, periodic 1-D U-Net for "
|
| 426 |
+
"L96) is trained with Adam on the current statistics window each outer iteration — a full "
|
| 427 |
+
"initial fit, then short warm-started refits that carry the weights across iterations; on the "
|
| 428 |
+
"larger L96 vec/flux cases only every 20th window sample is used for training (the GFDT score "
|
| 429 |
+
"is still evaluated on the full window)."
|
| 430 |
),
|
| 431 |
"params": [
|
| 432 |
+
{"Parameter": "Score network", "Value": "3×128 swish MLP (L63) / circular 1-D U-Net, ~66k params (L96 variants)"},
|
| 433 |
+
{"Parameter": "DSM training epochs (initial → warm refit)", "Value": "500 → 50 (L63) / 150 → 20 (L96 const) / 150 → 10 (L96 vec, flux)"},
|
| 434 |
+
{"Parameter": "DSM optimizer", "Value": "Adam, lr 1e-3 / 8e-4 initial → 3e-4 warm; batch 256 (L63) / 128 (L96); 10% validation hold-out"},
|
| 435 |
{"Parameter": "σ (DSM noise scale)", "Value": "0.2 (L63) / 0.025 (L96 variants)"},
|
| 436 |
{"Parameter": "tukey_alpha (response taper)", "Value": "0.75"},
|
| 437 |
{"Parameter": "budget_total (outer_iter × N_ens)", "Value": "500"},
|
|
|
|
| 456 |
"holds on the calibration ensemble, `kgmm_ridge` shrinks per-cluster covariances toward the "
|
| 457 |
"global covariance, and `is_collapsed_window` falls back to a direct `ForwardDiff.jl` gradient "
|
| 458 |
"on windows that have collapsed onto a fixed point; total cost is capped at "
|
| 459 |
+
"`outer_iter × N_ens ≤ budget_total` per cell (5,000 here, vs 500 for the LM variants), "
|
| 460 |
+
"trading ensemble size against iteration count under a fixed evaluation budget; beyond that, Adam's own adaptive per-parameter "
|
| 461 |
"step scaling is the only guardrail on the step itself. Unlike `β₁`/`β₂`/`ε` (shared across all "
|
| 462 |
"cases), the step size `α` is set per-experiment from the case's prior — sized to ~0.25× the "
|
| 463 |
"mean prior standard deviation, so a run of consistent-sign steps crosses ~1 prior std in ~4 "
|
|
|
|
| 469 |
{"Parameter": "β₁, β₂", "Value": "0.9, 0.999"},
|
| 470 |
{"Parameter": "kgmm_ridge", "Value": "0.1"},
|
| 471 |
{"Parameter": "tukey_alpha (response taper)", "Value": "0.75"},
|
| 472 |
+
{"Parameter": "budget_total (outer_iter × N_ens)", "Value": "5000"},
|
| 473 |
+
{"Parameter": "collapse tolerance (min variance)", "Value": "1e-2"},
|
| 474 |
],
|
| 475 |
"notes": "**Inherits Adam's local-minimum susceptibility** (no trust region, restart, or ensemble diversity to escape a bad basin) **and LM-KGMM's cheaper-than-DSM score estimation** — the k-means/Gaussian-mixture score fit is far less costly than training a neural score network, so ADAM-KGMM stayed tractable on problems where a DSM-based Adam variant would not.",
|
| 476 |
},
|
|
|
|
| 485 |
"density's score (Vincent, 2011); `σ` is fixed per benchmark rather than annealed, chosen "
|
| 486 |
"empirically as the smallest noise level at which the network still receives a usable learning "
|
| 487 |
"signal. Same Stein recalibration and collapsed-window `ForwardDiff.jl` fallback, and the same "
|
| 488 |
+
"Tukey-tapered response integral and evaluation-budget cap (5,000 evaluations) as ADAM-KGMM. "
|
| 489 |
+
"Same per-case "
|
| 490 |
"step size `α` (shared `β₁`/`β₂`/`ε`) as ADAM-KGMM — sized to ~0.25× the case's prior standard "
|
| 491 |
"deviation, since a single global value tuned for one benchmark's residual scaling left Adam "
|
| 492 |
+
"unable to traverse even the L63/L96-const prior within its iteration budget. **Score model & "
|
| 493 |
+
"training:** a small network (MLP for L63, periodic 1-D U-Net for L96) is trained with Adam on "
|
| 494 |
+
"the current statistics window each outer iteration — a full initial fit, then short "
|
| 495 |
+
"warm-started refits that carry the weights across iterations; on the larger L96 vec/flux "
|
| 496 |
+
"cases only every 20th window sample is used for training (the GFDT score is still evaluated "
|
| 497 |
+
"on the full window)."
|
| 498 |
),
|
| 499 |
"params": [
|
| 500 |
{"Parameter": "α (step size)", "Value": "Per-case, ~0.25× prior std (0.025–0.75 across benchmarks)"},
|
| 501 |
{"Parameter": "β₁, β₂", "Value": "0.9, 0.999"},
|
| 502 |
+
{"Parameter": "Score network", "Value": "3×128 swish MLP (L63) / circular 1-D U-Net, ~66k params (L96 variants)"},
|
| 503 |
+
{"Parameter": "DSM training epochs (initial → warm refit)", "Value": "500 → 50 (L63) / 150 → 20 (L96 const) / 150 → 10 (L96 vec, flux)"},
|
| 504 |
+
{"Parameter": "DSM optimizer", "Value": "Adam, lr 1e-3 / 8e-4 initial → 3e-4 warm; batch 256 (L63) / 128 (L96); 10% validation hold-out"},
|
| 505 |
{"Parameter": "σ (DSM noise scale)", "Value": "0.2 (L63) / 0.025 (L96 variants)"},
|
| 506 |
{"Parameter": "tukey_alpha (response taper)", "Value": "0.75"},
|
| 507 |
+
{"Parameter": "budget_total (outer_iter × N_ens)", "Value": "5000"},
|
| 508 |
+
{"Parameter": "collapse tolerance (min variance)", "Value": "1e-2"},
|
| 509 |
],
|
| 510 |
"notes": "**Inherits both LM-DSM's slow DSM-training cost and Adam's local-minimum susceptibility** — training the DSM score network is the binding wall-clock cost (as with LM-DSM), and Adam's lack of a trust region or ensemble diversity leaves it as prone to getting stuck in a bad basin as ADAM-KGMM.",
|
| 511 |
},
|
|
|
|
| 753 |
),
|
| 754 |
},
|
| 755 |
"IES-EKS": {
|
| 756 |
+
"citation": "Elsaesser et al., J. Adv. Model. Earth Syst., 2025; Garbuno-Inigo, Nüsken and Reich, SIAM J. Appl. Dyn. Syst., 2020 (ALDI)",
|
| 757 |
"url": "https://doi.org/10.1029/2024MS004713",
|
| 758 |
+
"source": "`uq_experiments/BayesianOptimalExperimentalDesign` — `boed_core.jl` (`run_eks`, `acquire_batch`), variant `:eks` (leaderboard key `boed-eks`); `EKP.Sampler(prior)` with EnsembleKalmanProcesses' default `sampler_type = \"aldi\"`.",
|
| 759 |
"math": (
|
| 760 |
"The same iterated emulate-sample loop as IES-IEKF and IES-TMCMC (see those entries), with the "
|
| 761 |
"posterior sampler swapped for the Ensemble Kalman Sampler (EKP `Sampler`). Iteration 1 "
|
|
|
|
| 767 |
"subsample of that ensemble."
|
| 768 |
),
|
| 769 |
"guardrails": (
|
| 770 |
+
"The sampler is the default `EKP.Sampler(prior)`, i.e. `sampler_type = \"aldi\"` (the same ALDI "
|
| 771 |
+
"variant used by the EKS and CES-EKS entries), driven by "
|
| 772 |
"`EKSStableScheduler(alpha = 0.25)` for a fixed 500 steps. `alpha = 0.25` was chosen because it "
|
| 773 |
"sat closer to the ST-MCMC reference than the default 1, and `alpha >= 3` is unstable. Like the IEKF "
|
| 774 |
"variant it ignores the GP predictive variance. GP fits use the same data-relative hyperparameter "
|
|
|
|
| 776 |
"shortfall is padded with NaN rows that the leaderboard stage drops."
|
| 777 |
),
|
| 778 |
"params": [
|
| 779 |
+
{"Parameter": "sampler_type", "Value": "\"aldi\" (EKP default, Sampler(prior))"},
|
| 780 |
{"Parameter": "EKS particles (eks_n_particles)", "Value": "200"},
|
| 781 |
{"Parameter": "scheduler", "Value": "EKSStableScheduler(alpha = 0.25)"},
|
| 782 |
{"Parameter": "steps (eks_max_iters / eks_T)", "Value": "500 / Inf (fixed step count)"},
|
|
|
|
| 791 |
),
|
| 792 |
},
|
| 793 |
"GBOED-EKS": {
|
| 794 |
+
"citation": "Holthuijzen et al. 2026 (EIG acquisition); Garbuno-Inigo, Nüsken and Reich, SIAM J. Appl. Dyn. Syst., 2020 (ALDI posterior sampler)",
|
| 795 |
"url": "https://arxiv.org/abs/2508.13071",
|
| 796 |
+
"source": "`uq_experiments/BayesianOptimalExperimentalDesign` — `boed_core.jl` (`run_eks`, `optimize_batch`, `acquire_batch`), variant `:eks_eig` (leaderboard key `boed-eks-eig`); `EKP.Sampler(prior)` with EnsembleKalmanProcesses' default `sampler_type = \"aldi\"`.",
|
| 797 |
"math": (
|
| 798 |
"GBOED with the posterior sampler swapped for the Ensemble Kalman Sampler (see the GBOED-IEKF and "
|
| 799 |
"IES-EKS entries). Iteration 1 forward-evaluates an `N_ens`-point Latin-hypercube design in the "
|
|
|
|
| 806 |
),
|
| 807 |
"math_eq": r"\mathrm{EIG}(X) = \frac{1}{k}\sum_j \frac{1}{2}\log\frac{\det K_{pp}^{(j)}}{\det \Sigma'^{(j)}}, \quad \Sigma'^{(j)} = K_{pp}^{(j)} - K_{pc}^{(j)}\left(K_{cc}^{(j)}\right)^{-1}K_{cp}^{(j)}",
|
| 808 |
"guardrails": (
|
| 809 |
+
"The sampler is the default `EKP.Sampler(prior)` with `sampler_type = \"aldi\"` (as in the EKS entry), run for a "
|
| 810 |
"fixed 500 steps of `EKSStableScheduler(alpha = 0.25)` with 200 particles, ignoring the GP "
|
| 811 |
"predictive variance. The batch is optimized with `Optim.jl`'s `LBFGS` inside the χ² ball of the "
|
| 812 |
"prior-whitened space (radial `tanh` squash, so every candidate has radius below √χ²_k(0.999)), with "
|
|
|
|
| 817 |
"caps as IES-IEKF."
|
| 818 |
),
|
| 819 |
"params": [
|
| 820 |
+
{"Parameter": "sampler_type", "Value": "\"aldi\" (EKP default, Sampler(prior))"},
|
| 821 |
{"Parameter": "EKS particles (eks_n_particles)", "Value": "200"},
|
| 822 |
{"Parameter": "scheduler", "Value": "EKSStableScheduler(alpha = 0.25)"},
|
| 823 |
{"Parameter": "steps (eks_max_iters / eks_T)", "Value": "500 / Inf (fixed step count)"},
|