Back to list of 2000/01 seminars

Evolutionary algorithms and other neighbourhood search algorithms, such as tabu search, have proved to be effective heuristic optimisation techniques for a wide range of pro blem domains such as engineering design, operational research, telecommuni-cations, and bioin formatics. However, the approach to their design is currently ad hoc. This introduces a num ber of barriers to the future adoption and scalability of these methods. For example, if the lin k between the optimiser design and the problem domain is not explicit then maintainability can be compromised, and it becomes difficult to rationalise optimiser performance in te rms that the end user can understand.

My research is aimed at resolving this issue of principled design, by proposing a Knowledge Based Systems approach. To this end, I will present a knowledge-level analysis t hat describes these optimisers in terms of the problem domain knowledge that they embody. I will also outline an accompanying formalism that allows domain knowledge to be formally expressed and then used to derive appropriate neighbourhood operators, such as t he mutation and recombination operators in evolutionary algorithms.

Finally, I will also show that a consideration of the pragmatics of the above an alysis provides an effective experimental protocol that exploits the commonalites of these metho ds, allows hillclimbing experiments to evaluate design decisions, and allows these decision s to be transferred to the design of other optimisers, including evolutionary algorithms . As well as providing an overview of the above work, this presentation will highl ight future research directions, and their relevance to practice.