Library Open Repository

Escaping the bounds of generality : unbounded bi-objective optimisation

Downloads

Downloads per month over past year

Berry, Adam Michael (2008) Escaping the bounds of generality : unbounded bi-objective optimisation. PhD thesis, University of Tasmania.

[img] PDF (Whole thesis)
whole_BerryAdam...pdf | Request a copy
Full text restricted
Available under University of Tasmania Standard License.

Abstract

The vast majority of contemporary evolutionary multiobjective optimisation research
is grounded in the ideals of generality — that is, the capacity to perform well
irrespective of the number of objectives that exist in a given task. Despite this
resolute focus, the fact remains that a large portion of the reported real-world
applications and existing algorithmic studies are exclusively two-dimensional. By
remaining fixated on generality, research has failed to explore the unique properties
of these hi-objective tasks and, in so doing, has sacrificed potential performance
gains in lieu of an often-unnecessary level of flexibility. As a response, this work
focuses on the bi-objective domain, endeavouring to elucidate and then harness the
special characteristics of non-dominated bi-objective sets to produce powerful and
efficient specialist techniques.
Central to this work is the hi-objective specialisation of the powerful elite archiving
mechanism. Where conventional modern systems limit the size of their archives to
curb the inefficiencies of naive list constructs (despite the potential for degradation in
both the quality of archival members and crowding estimates caused by such
artificial thresholding), this thesis describes a new construct (the MakTree) that
releases artificial size bounds while maintaining high levels of efficiency. Indeed,
both theoretical and empirical results illustrate that the Mak_Tree performs better
than other generalist unbounded methodologies and is often more efficient than, or at
least competitive with, tightly bound truncated techniques.
Moreover, this thesis indicates that the use of unbounded archives imparts a real
practical performance benefit to the optimisation of hi-objective problems. The
extension of modern evolutionary optimisers to incorporate the efficient unbounded
Mak_Tree construct — be it in a passive, active or hybridised manner — results in
significant performance improvements across a host of disparate test functions. The
creation of novel algorithms designed to capitalise on the properties of the Mak Tree
further emphasises the power of specialisation, with significant improvements again
noted when results are compared against those produced by the contemporary
bounded approaches.
Outside of the improvement of optimiser efficacy, efficient access to unbounded elite
sets also offers potential for the development of enhanced meta-processes.
Specifically, this thesis proposes and then explores amongst the first fully-realised,
intuitive, autonomous and reliable termination systems available to the field — a
system that is simply infeasible with access only to bounded archival sets.
Additionally, the work examines a new end-of-run presentation system that distills
complete unbounded archives into well-distributed collections that are suitable for
decision-maker analysis. Empirical results suggest that the proposed system
produces significantly better distributions than pre-existing techniques and offers
guarantees of solution quality that simply cannot be made when using bounded
stores.

Item Type: Thesis (PhD)
Copyright Holders: The Author
Copyright Information:

Copyright 2008 the Author - The University is continuing to endeavour to trace the copyright
owner(s) and in the meantime this item has been reproduced here in good faith. We
would be pleased to hear from the copyright owner(s).

Additional Information:

Available for library use only and copying in accordance with the Copyright Act 1968, as amended. Thesis (PhD)--University of Tasmania, 2008. Includes bibliographical references

Date Deposited: 25 Nov 2014 00:56
Last Modified: 11 Mar 2016 05:53
Item Statistics: View statistics for this item

Actions (login required)

Item Control Page Item Control Page