Return to search

Direct Search Methods for Nonsmooth Problems using Global Optimization Techniques

This thesis considers the practical problem of constrained and unconstrained local optimization. This subject has been well studied when the objective function f is assumed to smooth. However, nonsmooth problems occur naturally and frequently in practice. Here f is assumed to be nonsmooth or discontinuous without forcing smoothness assumptions near, or at, a potential solution. Various methods have been
presented by others to solve nonsmooth optimization problems, however only partial convergence results are possible for these methods.

In this thesis, an optimization method which use a series of local and localized global optimization phases is proposed. The local phase searches for a local minimum
and gives the methods numerical performance on parts of f which are smooth. The localized global phase exhaustively searches for points of descent in a neighborhood of cluster points. It is the localized global phase which provides strong theoretical convergence results on nonsmooth problems.

Algorithms are presented for solving bound constrained, unconstrained and constrained nonlinear nonsmooth optimization problems. These algorithms use direct search methods in the local phase as they can be applied directly to nonsmooth problems because gradients are not explicitly required. The localized global optimization phase uses a new partitioning random search algorithm to direct random sampling into promising subsets of ℝⁿ. The partition is formed using classification and regression trees (CART) from statistical pattern recognition. The CART partition defines desirable subsets where f is relatively low, based on previous sampling, from which further samples are drawn directly. For each algorithm, convergence to an essential local minimizer of f is demonstrated under mild conditions. That is, a point x* for which the set of all feasible points with lower f values has Lebesgue measure zero for all sufficiently small neighborhoods of x*. Stopping rules are derived for each algorithm giving practical convergence to estimates of essential local minimizers. Numerical results are presented on a range of nonsmooth test problems for 2 to 10 dimensions showing the methods are effective in practice.

Identiferoai:union.ndltd.org:canterbury.ac.nz/oai:ir.canterbury.ac.nz:10092/5060
Date January 2010
CreatorsRobertson, Blair Lennon
PublisherUniversity of Canterbury. Mathematics and Statistics
Source SetsUniversity of Canterbury
LanguageEnglish
Detected LanguageEnglish
TypeElectronic thesis or dissertation, Text
RightsCopyright Blair Lennon Robertson, http://library.canterbury.ac.nz/thesis/etheses_copyright.shtml
RelationNZCU

Page generated in 0.0019 seconds