ADAPTIVE INFILL POINT GENERATOR

Reads an existing design and/or a trained ml_learn_auto model and proposes the next batch of points to run, balancing exploitation (committee-disagreement uncertainty) against exploration (distance to existing points). Reports current out-of-fold accuracy versus a target and estimates how many more points are needed via a learning-curve fit or confidence-interval-width formula. Supports accuracy and optimization (Expected-Improvement) modes; composes with the Sampling Point Generator and Sampling Domain Reducer.

When to use

Tagged: doe, adaptive sampling, active learning, infill, sequential, surrogate, optimization loop, accuracy.

Inputs

Label ID Type Default Required Description
Design Dataset source dataset —   Existing design: input columns plus response columns. Provide this, or a trained model via mfile/mathmodel.
Model File mfile text —   Trained model handle: .pkl path or per-target CSV manifest. Feature names, targets and bounds are recovered via ml_predict_info.
Math Model mathmodel text —   Saved math-model id; resolved to its .pkl model file.
Variables variables dataset —   Variable definitions with name/min/max/type. If absent, recovered from the model schema or inferred from the dataset column ranges.
Inputs inputs text —   Comma-separated input (feature) column names. Defaults to all non-target columns of the dataset, or the model’s registered inputs.
Targets targets text —   Comma-separated response (target) column names. Defaults to the model’s registered targets.
Mode mode select accuracy   accuracy = raise global surrogate accuracy; optimization = Expected-Improvement infill toward a better optimum; diagnose = model-free design-adequacy check (no surrogate run).
Adequacy Model Order adequacy_formula select linear   Model order assumed when mode = diagnose; sets the coefficient count used for the adequacy floor.
Objective objective select minimize   Optimization direction (used when mode = optimization).
Number Of Points num_points scalar 5   How many new points to propose this round.
Target Accuracy target_accuracy scalar 0.9   Desired out-of-fold R2 (CoP). Drives target_met and recommended_more.
Sizing Method sizing_method select learning_curve   How recommended_more is estimated: learning-curve extrapolation or confidence-interval width.
Target CI Width target_ci_width scalar —   Desired confidence-interval half-width (used when sizing_method = ci_width).
Confidence Level confidence_level scalar 0.95   Confidence level for the CI-width sizing method.
Candidate Method candidate_method select latin-hypercube   How the dense candidate pool is generated before scoring.
Candidate Pool Size candidate_pool_size scalar 2000   Number of candidates generated and scored before selection.
Exploration Weight exploration_weight scalar 0.5   Weight w in score = w*local + (1-w)*global. Lower favors exploitation (model-weak regions).
Regression Types regression_types text linear_regression,rfr_regression,gp_regression   Committee members trained by ml_learn_auto; their prediction spread is the model-agnostic uncertainty signal.
CV Folds cv_folds scalar 5   Folds for the out-of-fold accuracy estimate.
Constraints constraints dataset —   Optional constraints (variable, condition, target) filtering proposed points; same format as the Sampling Point Generator.

Outputs

Label ID Type Description
Experiments experiments dataset Proposed new design points to run, one row each with the input variable columns and a score.
Current Accuracy current_accuracy scalar Best out-of-fold CoP/R2 of the current design.
Accuracy Metrics accuracy_metrics dataset Per-target CoP / RMSE / MAE for the current design.
Target Met target_met scalar yes/no — whether current accuracy meets target_accuracy.
Recommended More Points recommended_more scalar Estimated additional points to reach the target (learning-curve or CI-width). An estimate, not a guarantee.
Learning Curve learning_curve dataset Sub-sample sizes vs accuracy plus fitted parameters and asymptote; empty when sizing_method = ci_width.
Infill Scores infill_scores dataset Per-candidate local/global/total score breakdown for transparency.
Convergence convergence json Summary: current, target, met, delta, recommended_more, sizing_method, asymptote, plateaued, global_term_basis.
Design Adequacy adequacy json mode = diagnose: model-free adequacy {n_samples, n_inputs, formula, coefficients, rank_floor, recommended, residual_dof, sample_to_coeff_ratio, deficit, status}.
Adequacy Status status scalar mode = diagnose: under-determined / minimal / ok.
Adequacy Message message scalar mode = diagnose: human-readable adequacy verdict and recommended sample count.

Disciplines

  • ai_ml.prognosis
  • design_exploration.doe
  • design_exploration.optimization
  • design_exploration.sensitivity

Auto-generated from platform schema. Worker id: doe_adaptive_infill. Schema hash: 771e6223956b. Hand-curated docs in workerexamples/ override this page when present.