AbstractsComputer Science

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.