Dakota Reference Manual
Version 6.4
LargeScale Engineering Optimization and Uncertainty Analysis

The OPT++ library[60] contains primarily gradientbased nonlinear programming optimizers for unconstrained, boundconstrained, and nonlinearly constrained minimization: PolakRibiere conjugate gradient (Dakota's optpp_cg
method), quasiNewton (Dakota's optpp_q_newton
method), finite difference Newton (Dakota's optpp_fd_newton
method), and full Newton (Dakota's optpp_newton
method).
The conjugate gradient method is strictly unconstrained, and each of the Newtonbased methods are automatically bound to the appropriate OPT++ algorithm based on the user constraint specification (unconstrained, boundconstrained, or generallyconstrained). In the generallyconstrained case, the Newton methods use a nonlinear interiorpoint approach to manage the constraints. The library also contains a direct search algorithm, PDS (parallel direct search, Dakota's optpp_pds
method), which supports bound constraints.
Controls
max_iterations
max_function_evaluations
convergence_tolerance
output
speculative
Concurrency
OPT++'s gradientbased methods are not parallel algorithms and cannot directly take advantage of concurrent function evaluations. However, if numerical_gradients
with method_source
dakota
is specified, a parallel Dakota configuration can utilize concurrent evaluations for the finite difference gradient computations.
Constraints
Linear constraint specifications are supported by each of the Newton methods (optpp_newton
, optpp_q_newton
, optpp_fd_newton
, and optpp_g_newton
)
optpp_cg
must be unconstrained
optpp_pds
can be, at most, boundconstrained.