search_model_graphs

For weighted multilevel Monte Carlo, this option activates a search over possible hierarchical model graphs

Specification

• Alias: None

• Arguments: None

• Default: NO_GRAPH_RECURSION

Child Keywords:

Required/Optional	Description of Group	Dakota Keyword	Dakota Keyword Description
Optional		model_selection	Perform a recursion of admissible model subsets for a given model ensemble
Required (Choose One)	DAG Ensemble Generation Option	no_recursion	Do not recur over admissible DAGs for a given model ensemble
		full_recursion	Perform a full recursion of all admissible DAGs for a given model ensemble

Description

Referring to acv_recursive_diff, weighted MLMC is a special case of ACV-RD, resulting in a forward to the generalized ACV solver for the case of a fixed hierarchical DAG.

While the use of a hierarchical DAG is required in MLMC, the approximation selections and orderings within this DAG can be varied, so generalized ACV capabilities for model graph search (see full_recursion) and model selection (see model_selection) are available.

Examples

Note that the default for weighted MLMC is no search,

method,
    multilevel_sampling
      pilot_samples = 20 seed = 1237
      weighted
      max_function_evaluations = 500

with options to activate search over model ordering,

method,
    multilevel_sampling
      pilot_samples = 20 seed = 1237
      weighted
        search_model_graphs full_recursion
      max_function_evaluations = 500

over model subset selection,

method,
    multilevel_sampling
      pilot_samples = 20 seed = 1237
      weighted
        search_model_graphs no_recursion model_selection
      max_function_evaluations = 500

or over both.

method,
    multilevel_sampling
      pilot_samples = 20 seed = 1237
      weighted
        search_model_graphs full_recursion model_selection
      max_function_evaluations = 500