- author = "Martin Hughes",
- title = "Metaheuristics for black-box robust optimisation problems",
- school = "Management Science, Lancaster University",
- year = "2020",
- address = "Bailrigg, Lancaster, United Kingdom",
- month = may,
- keywords = "genetic algorithms, genetic programming, Grammar-Guided Genetic Programming, PSO, Largest Empty Hypersphere Metaheuristic, LEH",
-
URL = "
https://www.research.lancs.ac.uk/portal/en/publications/metaheuristics-for-blackbox-robust-optimisation-problems.html",
-
URL = "
https://eprints.lancs.ac.uk/id/eprint/145732/1/2020hughesphd.pdf",
-
DOI = "
doi:10.17635/lancaster/thesis/1036",
- size = "176 pages",
-
abstract = "Our interest is in the development of algorithms
capable of tackling robust black-box optimisation
problems, where the number of model runs is limited.
When a desired solution cannot be implemented exactly
(implementation uncertainty) the aim is to find a
robust one. Here that is to find a point in the
decision variable space such that the worst solution
from within an uncertainty region around that point
still performs well.
This thesis comprises three research papers. One has been published, one accepted for publication, and one submitted for publication. We initially develop a single-solution based approach, largest empty hypersphere (LEH), which identifies poor performing points in the decision variable space and repeatedly moves to the centre of the region devoid of all such points. Building on this we develop population based approaches using a particle swarm optimisation (PSO) framework. This combines elements of the LEH approach, a local descent directions (d.d.) approach for robust problems, and a series of novel features. Finally we employ an automatic generation of algorithms technique, genetic programming (GP), to evolve a population of PSO based heuristics for robust problems. We generate algorithmic sub-components, the design rules by which they are combined to form complete heuristics, and an evolutionary GP framework. The best performing heuristics are identified.
With the development of each heuristic we perform experimental testing against comparator approaches on a suite of robust test problems of dimension between 2D and 100D. Performance is shown to improve with each new heuristic. Furthermore the generation of large numbers of heuristics in the GP process enables an assessment of the best performing sub-components. This can be used to indicate the desirable features of an effective heuristic for tackling the problem under consideration. Good performance is observed for the following characteristics: inner maximisation by random sampling, a small number of inner points, particle level stopping conditions, a small swarm size, a Global topology, and particle movement using a baseline inertia formulation augmented by LEH and d.d. capabilities.",
- notes = "Supervisors: Marc Goerigk and Vikam Dokka and Mike Wright",
