Initial number of samples for sampling-based methods


  • Alias: None

  • Arguments: INTEGER

  • Default: model-dependent


The initial_samples keyword is used to define the number of initial samples (i.e., randomly chosen sets of variable values) at which to execute a model. The initial samples may later be augmented in an iterative process.

Default Behavior

By default, Dakota will use the minimum number of samples required by the chosen method.

Usage Tips

To obtain linear sensitivities or to construct a linear response surface, at least dim+1 samples should be used, where “dim” is the number of variables. For sensitivities to quadratic terms or quadratic response surfaces, at least (dim+1)(dim+2)/2 samples are needed. For uncertainty quantification, we recommend at least 10*dim samples. For variance_based_decomp, we recommend hundreds to thousands of samples. Note that for variance_based_decomp, the number of simulations performed will be N*(dim+2).


    sample_type random
    initial_samples = 20
    refinement_samples = 5