convergence_tolerance
Stopping criterion based on relative error
Specification
Alias: None
Arguments: REAL
Default: 1.e-4
Description
Multilevel sampling seeks an error balance between the estimator variance and the remaining bias error at the highest level, the two contributors to mean squared error (MSE). Since the remaining bias error is generally unknown, the convergence_tolerance is used to provide a error target relative to the Multifidelity Monte Carlo estimator variance resulting from the pilot sample. If the pilot samples are not shaped for the low-fidelity model, i.e. the number of low-fidelity evaluations is equal to the number of high-fidelity evaluations for each level, the Multifidelity estimator falls back to a Multilevel Monte Carlo estimator which is used to assess the estimator pilot samples variance. The samples allocated at each level are proportional to \(\frac{1}{\epsilon^2}\) , so each order of magnitude reduction in convergence_tolerance will tend to increase the sample allocation by two orders of magnitude. Therefore, this control should be used with care to avoid allocation of huge sample sets that could overrun available memory.
Default Behavior
The default value for convergence_tolerance is currently .0001, which may be too resolved for expensive simulations or high variance QoI.
