adaptive_sampling

(Experimental) Adaptively refine a Gaussian process surrogate

Topics

uncertainty_quantification

Specification

• Alias: nond_adaptive_sampling

• Arguments: None

Child Keywords:

Required/Optional	Description of Group	Dakota Keyword	Dakota Keyword Description
Optional		initial_samples	Initial number of samples for sampling-based methods
Optional		seed	Seed of the random number generator
Optional		samples_on_emulator	Number of samples at which to evaluate an emulator (surrogate)
Optional		fitness_metric	(Experimental) Specify the fitness_metric used to select the next point
Optional		batch_selection	(Experimental) How to select new points
Optional		refinement_samples	Number of samples used to refine a probabilty estimate or sampling design.
Optional		import_build_points_file	File containing points you wish to use to build a surrogate
Optional		export_approx_points_file	Output file for surrogate model value evaluations
Optional		misc_options	(Experimental) Hook for algorithm-specific adaptive sampling options
Optional		max_iterations	Number of iterations allowed for optimizers and adaptive UQ methods
Optional		response_levels	Values at which to estimate desired statistics for each response
Optional		probability_levels	Specify probability levels at which to estimate the corresponding response value
Optional		gen_reliability_levels	Specify generalized relability levels at which to estimate the corresponding response value
Optional		distribution	Selection of cumulative or complementary cumulative functions
Optional		rng	Selection of a random number generator
Optional		model_pointer	Identifier for model block to be used by a method

Description

This is an experimental capability that is not ready for production use at this point. It was part of an investigation into computational topology-based approaches to feature identification and surrogate refinement.

The goal in performing adaptive sampling is to construct a surrogate model that can be used as an accurate predictor to some expensive simulation, thus it is to one’s advantage to build a surrogate that minimizes the error over the entire domain of interest using as little data as possible from the expensive simulation. The adaptive part alludes to the fact that the surrogate will be refined by focusing samples of the expensive simulation on particular areas of interest rather than rely on random selection or standard space-filling techniques.

At a high-level, the adaptive sampling pipeline is a four-step process:

• Evaluate the expensive simulation (referred to as the true model) at initial sample point

  1. Fit a surrogate model

  2. Create a candidate set and score based on information from surrogate

  3. Select a candidate point to evaluate the true model

  4. Loop until done

In terms of the Dakota implementation, the adaptive sampling method currently uses Latin Hypercube sampling (LHS) to generate the initial points in Step 1 above. For Step 2, we use a Gaussian process model.

The default behavior is to add one point at a time. At each iteration (e.g. each loop of Steps 2-4 above), a Latin Hypercube sample is generated (a new one, different from the initial sample) and the surrogate model is evaluated at this points. These are the candidate points that are then evaluated according to the fitness metric. The number of candidates used in practice should be high enough to fill most of the input domain: we recommend at least hundreds of points for a low-dimensional problem. All of the candidates (samples on the emulator) are given a score and then the highest-scoring candidate is selected to be evaluated on the true model.

The adaptive sampling method also can generate batches of points to add at a time using the batch_selection and batch_size keywords.