rol

Rapid Optimization Library (ROL) is a large-scale optimization package within Trilinos.

Topics

local_optimization_methods

Specification

  • Alias: None

  • Arguments: None

Child Keywords:

Required/Optional

Description of Group

Dakota Keyword

Dakota Keyword Description

Optional

max_iterations

Number of iterations allowed for optimizers and adaptive UQ methods

Optional

variable_tolerance

Step length-based stopping criteria for derivative-free optimizers

Optional

gradient_tolerance

Stopping critiera based on L2 norm of gradient

Optional

constraint_tolerance

Maximum allowable constraint violation still considered feasible

Optional

options_file

File containing advanced ROL options

Optional

scaling

Turn on scaling for variables, responses, and constraints

Optional

model_pointer

Identifier for model block to be used by a method

Description

ROL is used for the solution of optimal design, optimal control and inverse problems in large-scale engineering applications.

Usage Tips

ROL is a general gradient-based library designed to scale well to very large problem sizes. For large problem sizes (i.e. number of variables), ROL’s trust region method and conjugate gradient methods exhibit good scalability for unconstrained problems. ROL handles equality constraints natively but converts inequality constraints into equality constraints with bounded slack variables. This has might degrade convergence for problems involving large number of inequality constraints. ROL has traditionally been applied to problems with analytic gradients (and Hessians)but can can be used with Dakota-provided finite-differencing approximations to the gradient of both objective and constraints. However, a user employing these is advised to use alternative optimizers such as DOT until performance of ROL improves in future releases.

Expected HDF5 Output

If Dakota was built with HDF5 support and run with the hdf5 keyword, this method writes the following results to HDF5: