Adaptive sampling strategies from uncertainty-aware reduced-order models can improve the efficiency of amortized optimization by focusing label acquisition on high-uncertainty parameter regions.

This discovery has a Claude-generated validation package with a full experimental design.

Precise Hypothesis

Spine & Adversarial Read

• highUncertainty sampling is a decades-old active learning technique (query-by-committee, GP-based Bayesian optimization); the discovery as framed offers no clear mechanistic novelty beyond applying known active learning to a new domain (amortized/surrogate optimization), and evidence strength (0.63) and verification confidence (0.48) suggest this has not been rigorously distinguished from prior art.
  The EVP does not yet cite or benchmark against specific prior active-learning-for-BO methods (e.g., BoTorch's qNIPV, entropy search); without a live prior-art search this gap is acknowledged and unresolved — CLOSEST_EXISTING_WORK is empty due to lack of search results, which is a genuine limitation of this package, not a resolved question.
• highReported gains could be an artifact of increased sample diversity rather than uncertainty per se (a well-known confound in active learning), especially in batch acquisition settings where top-k uncertainty selection clusters points.
  Methodology explicitly includes a diversity-only (Sobol/LHS) baseline and ablations across batch sizes to help disentangle diversity from uncertainty effects, but does not include a formal diversity+uncertainty hybrid (e.g., coreset-uncertainty combination) baseline, which would be needed for a fully conclusive causal claim.
• mediumWhy these specific benchmarks (Branin/Ackley, a generic PDE, one BBOB/aero dataset) and these three surrogate types (GP, ensemble, evidential) rather than others — the methodology choice is not justified against alternative surrogate families (e.g., random forests, Bayesian neural nets with MC-dropout) or more realistic industrial-scale simulators, risking a narrow/cherry-picked validation.
  Benchmark choice is partially justified by covering low/medium dimensionality and both synthetic and physics-based cases, but no explicit justification is given for excluding MC-dropout BNNs or random forest quantile regression, nor for the specific dimensionality cutoffs (2-5D, 10-20D); this should be addressed by adding a methodology-justification appendix or expanding surrogate diversity in a follow-up iteration.

Experimental Protocol

ROI Projection

Implementation Sketch

Initialize D_labeled = seed_sample(param_space, n=seed_size)
Y_labeled = simulate(D_labeled)  # expensive oracle calls
surrogate = init_surrogate(type=[GP|Ensemble|Evidential])

for round in range(n_rounds):
    surrogate.fit(D_labeled, Y_labeled)
    candidates = sample_candidate_pool(param_space, n=large_pool)
    uncertainty = surrogate.predict_uncertainty(candidates)
    
    if method == "uncertainty":
        new_points = top_k(candidates, key=uncertainty, k=batch_size)
    elif method == "random":
        new_points = random_sample(candidates, k=batch_size)
    elif method == "sobol":
        new_points = sobol_sample(param_space, k=batch_size)
    
    new_labels = simulate(new_points)  # oracle call, cost-tracked
    D_labeled.append(new_points); Y_labeled.append(new_labels)
    
    log(labels_used, surrogate.eval(held_out_test_set), calibration_metrics)

report(labels_to_target_curve, AUC_efficiency, calibration_diagnostics)