Dakota Reference Manual  Version 6.12
Explore and Predict with Confidence
 All Pages

Required or expected accuracy in optimization variables.


Alias: none

Argument(s): REAL

Default: 1.0e-4 (COBYLA), 1.0e-5 (PS), 1.0e-6 (SW)


Unlike it's use as a stopping criteria in other methods, variable_tolerance is used to communicate the accuracy of the optimization variables to COBYLA. It represents the mimimum distance from the simplex within which linear approximation subproblems can be trusted to be sufficiently representative of the true problem. It is analagous to the minimum trust-region size used in trust-region methods. Note that, per COBYLA documentation, the level of accuracy is not guaranteed.

Default Behavior

The default value is 0.0001.

Additional Discussion

While variable_tolerance is not a stopping criteria for COBYLA, we note that it may have a side effect regarding convergence. In particular, if lower accuracy is required of the optimization variables, it may stop sooner than if higher accuracy is required.