Low-Dimensional Curvature Methods in Derivative-Free Optimization on Shared Computing Networks

Macklem, Mason S. (2009) Low-Dimensional Curvature Methods in Derivative-Free Optimization on Shared Computing Networks. PhD thesis, Dalhousie University.

Download (3248Kb) | Preview


    This thesis presents work on building an in-house optimization package for use on shared computing networks, such as WestGrid in Western Canada, SHARCnet in Ontario, and ACEnet in Atlantic Canada, where users from various institutions and a variety of research projects compete with each other for access to computing resources. We introduce a parallel implementation of a direct-search method that is designed for use on a small number of processors, for situations where limiting the number of computing resources requested will yield a higher priority placement within the job queue. We ¯rst describe an approach to load balancing by using a particular method of partitioning the set of search directions, using objects from graph decomposition and graph factorization. We then present a new algorithm which uses this partitioning of the problem into independent subproblems, with local curvature information from within each subproblem used to re-align the search directions. We then consider direct-search and evolutionary strategies, comparing these two classes of methods on several standard test problems and varying numbers of processors, and discussing where each of these two communities can learn lessons from the other.

    Item Type: Thesis (PhD)
    Additional Information: pubdom FALSE
    Subjects: 65-xx Numerical analysis > 65Kxx Mathematical programming, optimization and variational techniques
    Faculty: UNSPECIFIED
    Depositing User: lingyun ye
    Date Deposited: 05 Aug 2009
    Last Modified: 13 Jan 2015 12:37

    Actions (login required)

    View Item