inform_search

Surrogate informs evaluation order in mesh adaptive search

Specification

• Alias: None

• Arguments: None

Description

When inform_search is specified with use_surrogate, mesh_adaptive_search uses the surrogate to sort list of trial points and subsequently the true function is evaluated on the most promising points first. Both true function and surrogate are used interchangeably within the method.

Default Behavior

inform_search is not the default surrogate usage mode.

Expected Output

The user can expect to see both the number of true model evaluations and the number of approximation (i.e., surrogate) evaluations reported in the Dakota screen output. The former captures the sum of truth evaluations done for the surrogate construction and for the optimization.

Usage Tips

When inform_search is specified, the maximum_function_evaluations keyword applies to only the optimization method and does not account for evaluations needed to construct the surrogate. If the user has a strict evaluation budget, they should set maximum_function_evaluations such that evaluation budget = number of evaluations to construct surrogate + maximum_function_evaluations.

Examples

The following example shows the syntax used to set use_surrogate to optimize.

method,
 mesh_adaptive_search
 model_pointer = 'SURROGATE'
 use_surrogate inform_search

model,
 id_model = 'SURROGATE'
 surrogate global
 polynomial quadratic
 dace_method_pointer = 'SAMPLING'

variables,
 continuous_design = 3
   initial_point   -1.0    1.5   2.0
   upper_bounds    10.0   10.0  10.0
   lower_bounds   -10.0  -10.0 -10.0
   descriptors      'x1'  'x2'  'x3'
 discrete_design_range = 2
   initial_point      2     2
   lower_bounds       1     1
   upper_bounds       4     9
   descriptors      'y1'   'y2'
       discrete_design_set
         real = 2
           elements_per_variable = 4 5
           elements = 1.2 2.3 3.4 4.5 1.2 3.3 4.4 5.5 7.7
           descriptors      'y3'   'y4'
         integer = 2
           elements_per_variable = 2 2
           elements = 4 7 8 9
           descriptors      'z1'   'z2'

method,
 id_method = 'SAMPLING'
 model_pointer = 'TRUTH'
 sampling
   samples = 55

model,
 id_model = 'TRUTH'
 single
   interface_pointer = 'TRUE_FN'

interface,
 id_interface = 'TRUE_FN'
       direct
   analysis_driver = 'text_book'

responses,
 objective_functions = 1
 no_gradients
 no_hessians

The following will appear toward the end of the screen output when Dakota is run on this example. The number of true function evaluations includes the 55 evaluations that were done to construct the surrogate (as specified in the SAMPLING method block) plus the number of truth evaluations done by mesh_adaptive_search.

<<<<< Function evaluation summary (APPROX_INTERFACE): 1660 total (1660 new, 0 duplicate)
<<<<< Function evaluation summary (TRUE_FN): 795 total (795 new, 0 duplicate)