Dakota Reference Manual  Version 6.4
Large-Scale Engineering Optimization and Uncertainty Analysis
 All Pages
batch_selection


(Experimental) How to select new points

Specification

Alias: none

Argument(s): none

Default: naive

Required/Optional Description of Group Dakota Keyword Dakota Keyword Description
Required
(Choose One)
Group 1 naive Take the highest scoring candidates
distance_penalty Add a penalty to spread out the points in the batch
topology In this selection strategy, we use information about the topology of the space from the Morse-Smale complex to identify next points to select.
constant_liar Use information from the existing surrogate model to predict what the surrogate upgrade will be with new points.

Description

adaptive_sampling is an experimental capability that is not ready for production use at this time.

With batch or multi-point selection, the true model can be evaluated in parallel and thus increase throughput before refitting our surrogate model. This proposes a new challenge as the problem of choosing a single point and choosing multiple points off a surrogate are fundamentally different. Selecting the n best scoring candidates is more than likely to generate a set of points clustered in one area which will not be conducive to adapting the surrogate.

We have implemented several strategies for batch selection of points. These are described in the User's manual and are the subject of active research.

The batch_selection strategies include:

  1. naive:
  2. distance_penalty
  3. constant_liar
  4. topology