Created by W.Langdon from gp-bibliography.bib Revision:1.7954
Existing discrete truss optimisation methods focus primarily on optimising global topology using a ground structure approach, with all possible node and beam locations being specified a priori and the algorithm selecting the most appropriate configuration from the given options. The standard method is to explore this search space, while seeking minimum cross-sectional areas for all elements in order to reduce the self-weight of the structure. In doing so, critical knowledge of section geometry and orientation is omitted. This leads to inaccurate stress calculations and structures failing to meet codes of practice. These issues can be addressed by constraining the optimisation method to only use standard construction elements. It is shown in this thesis that solutions close to the theoretical optimum can be achieved using commercially available elements.
The classical ground structure discrete optimisation method has furthermore been shown to be inherently restrictive, as it severely limits the representation space to what is explicitly defined; a larger representation space can more effectively navigate through the search space. However, a larger representation space can potentially lead to difficulties in evolving any fit solution. Unfit individuals must be handled carefully in order to successfully evolve any fit solution in early generations. It is therefore imperative to design the fitness function in such a way as to minimise the risk of the algorithm becoming stuck in a local optimum, before a single fit solution has been evolved.
The application of Grammatical Evolution (GE), a grammar-based form of Genetic Programming (GP), has shown that it is not only capable of generating innovative engineering designs, but that the recursive properties of formal grammars allows GE to define its own node locations for any number of nodes within a pre-specified design envelope, thereby vastly increasing its representation capabilities. Nodes are then connected via a Delaunay triangulation algorithm, leading to fully triangulated, kinematically stable structures. The net result is that discrete beam-truss structures can be optimised in a continuum manner, in a black-box fashion, without the need to know any information about the problem other than the design envelope. Existing discrete optimisation techniques are compared and contrasted, and notable savings in structure self-weight are demonstrated over traditional methods.",
Supervisor: Ciaran McNally",
Genetic Programming entries for Michael Fenton