method
Begins Dakota method selection and behavioral settings.
Topics
block
Specification
Alias: None
Arguments: None
Child Keywords:
Required/Optional |
Description of Group |
Dakota Keyword |
Dakota Keyword Description |
|---|---|---|---|
Optional |
Name the method block; helpful when there are multiple |
||
Optional |
Control how much method information is written to the screen and output file |
||
Optional |
Number of designs returned as the best solutions |
||
Required (Choose One) |
Method (Iterative Algorithm) |
Strategy in which a set of methods synergistically seek an optimal design |
|
Multi-Start Optimization Method |
|||
Pareto set optimization |
|||
(Experimental Capability) Solves a mixed integer nonlinear optimization problem |
|||
Local Surrogate Based Optimization |
|||
Adaptive Global Surrogate-Based Optimization |
|||
DOT conjugate gradient optimization method |
|||
DOT modified method of feasible directions |
|||
DOT BFGS optimization method |
|||
DOT Sequential Linear Program |
|||
DOT Sequential Quadratic Program |
|||
CONMIN conjugate gradient optimization method |
|||
CONMIN method of feasible directions |
|||
(Experimental) Dynamically-loaded solver |
|||
NPSOL Sequential Quadratic Program |
|||
Sequential Quadratic Program for nonlinear least squares |
|||
NLPQL Sequential Quadratic Program |
|||
A conjugate gradient optimization method |
|||
Quasi-Newton optimization method |
|||
Finite Difference Newton optimization method |
|||
Newton method based least-squares calbration |
|||
Newton method based optimization |
|||
Simplex-based derivative free optimization method |
|||
Third-party optimization library integration demonstration. |
|||
Rapid Optimization Library (ROL) is a large-scale optimization package within Trilinos. |
|||
Pattern search, derivative free optimization method |
|||
Finds optimal variable values using adaptive mesh-based search |
|||
Gradient-free inequality-constrained optimization using Nonlinear Optimization With Path Augmented Constraints (NOWPAC). |
|||
Stochastic version of NOWPAC that incorporates error estimates and noise mitigation. |
|||
Multi-objective Genetic Algorithm (a.k.a Evolutionary Algorithm) |
|||
Single-objective Genetic Algorithm (a.k.a Evolutionary Algorithm) |
|||
Pattern search, derivative free optimization method |
|||
Simple greedy local search method |
|||
Constrained Optimization BY Linear Approximations (COBYLA) |
|||
DIviding RECTangles method |
|||
Evolutionary Algorithm |
|||
(Experimental) Coliny beta solver |
|||
Trust-region method for nonlinear least squares |
|||
(Experimental) nonlinear conjugate gradient optimization |
|||
DIviding RECTangles method |
|||
Voronoi-based high-dimensional global Lipschitzian optimization |
|||
Classical high-dimensional global Lipschitzian optimization Classical high-dimensional global Lipschitzian optimization |
|||
Global Surrogate Based Optimization, a.k.a. EGO |
|||
Generic UQ method for constructing and interrogating a surrogate model. |
|||
UQ method leveraging a functional tensor train surrogate model. |
|||
Multifidelity uncertainty quantification using function train expansions |
|||
Multilevel uncertainty quantification using function train expansions |
|||
Uncertainty quantification using polynomial chaos expansions |
|||
Multifidelity uncertainty quantification using polynomial chaos expansions |
|||
Multilevel uncertainty quantification using polynomial chaos expansions |
|||
Uncertainty quantification with stochastic collocation |
|||
Multifidelity uncertainty quantification using stochastic collocation |
|||
Randomly samples variables according to their distributions |
|||
Multilevel sampling methods for UQ |
|||
Multifidelity sampling methods for UQ |
|||
Multilevel-Multifidelity sampling methods for UQ |
|||
Approximate control variate (ACV) sampling methods for UQ |
|||
Importance sampling |
|||
Gaussian Process Adaptive Importance Sampling |
|||
(Experimental) Adaptively refine a Gaussian process surrogate |
|||
Probability-of-Failure (POF) darts is a novel method for estimating the probability of failure based on random sphere-packing. |
|||
Recursive k-d (RKD) Darts: Recursive Hyperplane Sampling for Numerical Integration of High-Dimensional Functions. |
|||
Evidence theory with evidence measures computed with global optimization methods |
|||
Interval analysis using global optimization methods |
|||
Bayesian calibration |
|||
Design and Analysis of Computer Experiments |
|||
Design of Computer Experiments - Centroidal Voronoi Tessellation |
|||
Morris One-at-a-Time |
|||
Evidence theory with evidence measures computed with local optimization methods |
|||
Interval analysis using local optimization |
|||
Local reliability method |
|||
Global reliability methods |
|||
Design of Computer Experiments - Quasi-Monte Carlo sampling |
|||
Samples variables along a user-defined vector |
|||
Samples variables as a specified values |
|||
Samples variables along points moving out from a center point |
|||
Samples variables on full factorial grid of study points |
|||
Estimate order of convergence of a response as model fidelity increases |
|||
Description
The method keyword signifies the start of a block in the Dakota
input file. A method block contains the various keywords necessary to
select a method and to control its behavior.
Method Block Requirements
At least one method block must appear in the Dakota input
file. Multiple method blocks may be needed to fully define
advanced analysis approaches.
Each method block must specify one method and, optionally, any
associated keywords that govern the behavior of the method.
The Methods
Each method block must select one method.
Starting with Dakota v6.0, the methods are grouped into two types: standard methods and multi-component methods.
The standard methods are
stand-alone and self-contained in the sense that they only require
a model to perform a study. They do not call other methods.
While methods such as
polynomial_chaos and efficient_global internally utilize multiple
iterator and surrogate model components, these components are
generally hidden from user control due to restrictions on modularity;
thus, these methods are stand-alone.
The multi-component group
of methods provides a higher level “meta-algorithm” that points to
other methods and models that support sub-iteration.
For example, in a sequential hybrid method, the hybrid
method specification must identify a list of subordinate methods, and
the “meta-algorithm” executes these methods in sequence and transfers
information between them. Surrogate-based minimizers provide another
example in that they point both to other methods (e.g. what
optimization method is used to solve the approximate subproblem) as
well as to models (e.g. what type of surrogate model is employed).
Multi-component methods generally provide some level of “plug and
play” modularity, through their flexible support of a variety of
method and model selections.
Component-Based Iterator Commands
Component-based iterator specifications include hybrid, multi-start,
pareto set, surrogate-based local, surrogate-based global, and branch
and bound methods. Whereas a standard iterator specification only
needs an optional model pointer string (specified with
model_pointer), component-based iterator specifications can include
method pointer, method name, and model pointer specifications in order
to define the components employed in the “meta-iteration.” In
particular, these specifications identify one or more methods (by
pointer or by name) to specify the subordinate iterators that will be
used in the top-level algorithm. Identifying a sub-iterator by name
instead of by pointer is a lightweight option that relaxes the need
for a separate method specification for the sub-iterator; however, a
model pointer may be required in this case to provide the
specification connectivity normally supported by the method pointer.
Refer to these individual method descriptions for specific
requirements for these advanced methods.
Method Independent Controls
In addition to the method, there are 10 optional keywords, which are referred to as method independent controls. These controls are valid for enough methods that it was reasonable to pull them out of the method dependent blocks and consolidate the specifications, however, they are NOT universally respected by all methods.
Examples
Several examples follow. The first example shows a minimal specification for an optimization method.
method
dot_sqp
This example uses all of the defaults for this method.
A more sophisticated example would be
method,
id_method = 'NLP1'
dot_sqp
max_iterations = 50
convergence_tolerance = 1e-4
output verbose
model_pointer = 'M1'
This example demonstrates the use of identifiers and pointers
as well as some method independent and method
dependent controls for the sequential quadratic programming (SQP)
algorithm from the DOT library. The max_iterations,
convergence_tolerance, and output settings are method independent
controls, in that they are defined for a variety of methods.
The next example shows a specification for a least squares method.
method
optpp_g_newton
max_iterations = 10
convergence_tolerance = 1.e-8
search_method trust_region
gradient_tolerance = 1.e-6
Some of the same method independent controls are present along with
several method dependent controls ( search_method and
gradient_tolerance) which are only meaningful for OPT++ methods (see
topic-package_optpp).
The next example shows a specification for a nondeterministic method with several method dependent controls (refer to method-sampling).
method
sampling
samples = 100
seed = 12345
sample_type lhs
response_levels = 1000. 500.
The last example shows a specification for a parameter study method where, again, each of the controls are method dependent (refer to method-vector_parameter_study).
method
vector_parameter_study
step_vector = 1. 1. 1.
num_steps = 10
