Sandia National Labs
A strong need for derivative-free algorithms exists in the context of real-world optimization problems
where function evaluations can be computationally expensive and noisy. The objective and constraint functions
commonly exist as simple script interfaces to CPU intensive model analysis software. A single evaluation may
involve invoking cumbersome simulation codes whose run time is measured in hours. In this context, we present
an asynchronous parallel implementation of a derivative-free augmented Lagrangian algorithm for handling general
nonlinear constraints. The method requires approximate minimizers to a series of linearly constrained subproblems
involving the augmented Lagrangian of the nonlinear constraints. These subproblem are solved using a generating
set search algorithm capable of handling degenerate linear constraints. The objective and nonlinear constraint
functions are computed asynchronously in parallel.
A description and theoretical analysis of the algorithm will be given followed by numerical results.