In recent years, the theories of natural selection and biological evolution have proved popular metaphors for understanding and solving optimization problems in engineering design. This thesis identifies some fundamental problems associated with this use of such metaphors. Key objections are the failure of evolutionary optimization techniques to represent explicitly the goal of the optimization process, and poor use of knowledge developed during the process. It is also suggested that convergent behaviour of an optimization algorithm is an undesirable quality if the algorithm is to be applied to multimodal problems. An alternative approach to optimization is suggested, based on the explicit use of knowledge and/or assumptions about the nature of the optimization problem to construct Bayesian probabilistic models of the surface being optimized and the goal of the optimization. Distinct exploratory and exploitative strategies are identified for carrying out optimization based on such models—exploration based on attempting to reduce maximally an entropy-based measure of the total uncertainty concerning the satisfaction of the optimization goal over the space, exploitation based on evalutation of the point judged most likely to achieve the goal—together with a composite strategy which combines exploration and exploitation in a principled manner. The behaviour of these strategies is empirically investigated on a number of test problems. Results suggest that the approach taken may well provide effective optimization in a way which addresses the criticisms made of the evolutionary metaphor, subject to issues of the computational cost of the approach being satisfactorily addressed.

Document Type


Publication Date