The tool can be used for testing nonlinear optimizers on a few well-known functions. It has been used for nonlinear operators generated from AnalyticFunctions using the AnalyticFunctionOperator wrapper class. Have a look at the Analytic::Common namespace candidates. There must be an implementation of a helper traits class in kernel/solver/test_aux/function_traits.hpp specifying the real minima and a starting point.

Some part of the solver configuration can be done on the command line (solvertype, direction update, preconditioner), while other options must be specified at compile time (the linesearch, memory architecture and floating point precision). The solver by default logs all iterates and produces a vtk output of the domain of interest (named dbg_nlopt_[FUNCTIONNAME]_domain.vtu) and a polyline of all iterates (named dbg_nlopt_[FUNCTIONNAME]_iterates.vtu) so one can study the convergence behavior of the solver.

This can easily be extended to use nonlinear operators from other sources, but the whole plotting procedures require that we optimize a scalar function in two or three variables.

If FEAT3 was compiled with alglib in the buildid, the ALGLIBMinCG solver is available as well.

Usage:

dbg-nlopt [options...]

where options are

--help Displays usage information
--solver Supported solver are NLCG(default) and ALGLIBMinCG
--precon Supported preconditioners for NLCG are
- ApproximateHessian
- Hessian
- none(default).
  Preconditioning for ALGLIBMinCG is not supported due to limitations of ALGLIB itself (it appears one cannot recompute the preconditioner in every iteration).
--direction_update Supported direction updates for NLCG are
- DaiYuan
- DYHSHybrid
- FletcherReeves
- HestenesStiefel
- PolakRibiere(default).
  Supported direction updates for ALGLIBMinCG are
- DaiYuan
- DYHSHybrid
- automatic(default)

See also: FEAT::Solver::NLCG, FEAT::Solver::Linesearch