Characterising fitness landscapes with fitness-probability cloud and its applications to algorithm configuration

Lu, Guanzhou (2014). Characterising fitness landscapes with fitness-probability cloud and its applications to algorithm configuration. University of Birmingham. Ph.D.

Preview

Lu14PhD.pdf
PDF - Accepted Version
Download (1MB)

Abstract

Metaheuristics are approximation optimisation techniques widely applied to solve complex optimisation problems. Despite a large number of developed metaheuristic algorithms, a limited amount of work has been done to understand on which kinds of problems the proposed algorithm will perform well or poorly and why. A useful solution to this dilemma is to use fitness landscape analysis to gain an in-depth understanding of which algorithms, or algorithm variants are best suited for solving which kinds of problem instances, even to dynamically determine the best algorithm configuration during different stages of a search algorithm.

This thesis for the first time bridges the gap between fitness landscape analysis and algorithm configuration, i.e., finding the best suited configuration of a given algorithm for solving a particular problem instance. Studies in this thesis contribute to the following:

a. Developing a novel and effective approach to characterise fitness landscapes and measure problem difficulty with respect to algorithms.

b. Incorporating fitness landscape analysis in building a generic (problem-independent) approach, which can perform automatic algorithm configuration on a per-instance base, and in designing novel and effective algorithm configurations.

c. Incorporating fitness landscape analysis in establishing a generic framework for designing adaptive heuristic algorithms.

Type of Work:

Thesis (Doctorates > Ph.D.)

Award Type:

Doctorates > Ph.D.

Supervisor(s):