snowpac
Stochastic version of NOWPAC that incorporates error estimates and noise mitigation.
Specification
Alias: None
Arguments: None
Child Keywords:
Required/Optional |
Description of Group |
Dakota Keyword |
Dakota Keyword Description |
---|---|---|---|
Optional |
Seed of the random number generator |
||
Optional |
Use trust region as the globalization strategy. |
||
Optional |
Number of iterations allowed for optimizers and adaptive UQ methods |
||
Optional |
Number of function evaluations allowed for optimizers |
||
Optional |
Turn on scaling for variables, responses, and constraints |
||
Optional |
Identifier for model block to be used by a method |
Description
NOWPAC is a provably-convergent gradient-free optimization method from MIT that solves a series of trust region surrogate-based subproblems to generate improving steps. The stochastic version is SNOWPAC, which incorporates noise estimates in its objective and inequality constraints. SNOWPAC modifies its trust region controls and adds smoothing from a Gaussian process surrogate in order to mitigate noise. SNOWPAC also supports a feasibility restoration mode, so it is not necessary to start from a feasible design.
Note: (S)NOWPAC is not configured with Dakota by default and requires a separate installation of the NOWPAC distribution from MIT, combined with its TPLs of Eigen and NLOPT.
Examples
Relative to the NOWPAC specification, SNOWPAC supports a seed control for repeatability of runs and also requires the return of error estimates from the underlying evaluator (e.g., UQ method such as Monte Carlo sampling).
method,
snowpac
seed = 2504
max_function_evaluations = 1000
convergence_tolerance = 1e-4
trust_region
initial_size = 0.10
minimum_size = 1.0e-6
contract_threshold = 0.25
expand_threshold = 0.75
contraction_factor = 0.50
expansion_factor = 1.50