Performance evaluation on optimisation of 200 dimensional numerical tests - results and issues

    Research output: Chapter in Book/Report/Published conference proceedingChapterpeer-review

    Abstract

    Abstract: Many tasks in science and technology require optimisation. Resolving such tasks could bring great benefits to community. Multidimensional problems where optimisation parameters are hundreds and more face unusual computational limitations. Algorithms, which perform well on low number of dimensions, when are applied to high dimensional space suffers insuperable difficulties. This article presents an investigation on 200 dimensional scalable, heterogeneous, real-value, numerical tests. For some of these tests optimal values are dependent on dimensions’ number and virtually unknown for variety of dimensions. Dependence on initialisation for successful identification of optimal values is analysed by comparison between experiments with start from random initial locations and start from one location. The aim is to: (1) assess dependence on initialisation in optimisation of 200 dimensional tests; (2) evaluate tests complexity and required for their resolving periods of time; (3) analyse adaptation to tasks with unknown solutions; (4) identify specific peculiarities which could support the performance on high dimensions (5) identify computational limitations which numerical methods could face on high dimensions. Presented and analysed experimental results can be used for further comparison and evaluation of real value methods.
    Original languageMultiple
    Title of host publicationМеждународна научна конференция "70 години МТФ"
    Subtitle of host publicationМТФ '2015 28-ма Международна научна конференция на Машинно-технологичния факултет на ТУ - София, Созопол, България 11-13 Септември 2015. Сборник доклади
    Place of PublicationSofia, Bulgari
    PublisherИздателство на Техническия Университет София
    Edition1
    ISBN (Print)978-619-167-178-6
    Publication statusPublished - 30 Sept 2015

    Cite this