| Summary: | This dissertation investigates the support of conceptual engineering design through the decomposition of multi-dimensional search spaces into regions of high performance. Such decomposition helps the designer identify optimal design directions by the elimination of infeasible or undesirable regions within the search space. Moreover, high levels of interaction between the designer and the model increases overall domain knowledge and significantly reduces uncertainty relating to the design task at hand. The aim of the research is to develop the archetypal Cluster Oriented Genetic Algorithm (COGA) which achieves search space decomposition by using variable mutation (vmCOGA) to promote diverse search and an Adaptive Filter (AF) to extract solutions of high performance [Parmee 1996a, 1996b]. Since COGAs are primarily used to decompose design domains of unknown nature within a real-time environment, the elimination of apriori knowledge, speed and robustness are paramount. Furthermore COGA should promote the in-depth exploration of the entire search space, sampling all optima and the surrounding areas. Finally any proposed system should allow for trouble free integration within a Graphical User Interface environment. The replacement of the variable mutation strategy with a number of algorithms which increase search space sampling are investigated. Utility is then increased by incorporating a control mechanism that maintains optimal performance by adapting each algorithm throughout search by means of a feedback measure based upon population convergence. Robustness is greatly improved by modifying the Adaptive Filter through the introduction of a process that ensures more accurate modelling of the evolving population. The performance of each prospective algorithm is assessed upon a suite of two-dimensional test functions using a set of novel performance metrics. A six dimensional test function is also developed where the areas of high performance are explicitly known, thus allowing for evaluation under conditions of increased dimensionality. Further complexity is introduced by two real world models described by both continuous and discrete parameters. These relate to the design of conceptual airframes and cooling hole geometries within a gas turbine. Results are promising and indicate significant improvement over the vmCOGA in terms of all desired criteria. This further supports the utilisation of COGA as a decision support tool during the conceptual phase of design.
|
|---|
