Zeroth-order shape optimization utilizing a learning classifier system

A methodology to perform generalized zeroth-order two- and three-dimensional shape optimization utilizing a learning classifier system is developed and applied. To this end, the applicability of machine learning to mechanical engineering is investigated. Specifically, the methodology has the objective of determining the optimal boundary to minimize mass while satisfying constraints on stress and geometry.; Even with the enormous advances in shape optimization no method has proven to be satisfactory across the broad spectrum of optimization problems facing the modern engineer. The methodology developed in this dissertation is based upon a classifier system (CS) and exploits the CS's adaptability and generality. It thereby overcomes many of the limitations of today's conventional shape optimization techniques. A CS learns rules, postulated as if-then statements, in order to improve its performance in an arbitrary environment, (which for this investigation consists of stress and mass information from components). From this input, and from a population of initially randomly generated rules, the classifier system is expected to learn to make the appropriate component shape modifications to reach a minimum mass design while satisfying all stress constraints. The CS learns by utilizing the design improvement success or failure feedback.; After a review of mechanical engineering shape optimization methods, an explanatory presentation of CSs and their underlying genetic algorithm (GA) describes how classifier systems learn from feedback and the GA. With this foundation set, the coupling of the shape optimization domain with the classifier system proceeds to form, the Shape oPtimization via Hypothesizing Inductive classifier system compleX (SPHINcsX). The complex learns shape optimization by its application to a suite of sizing optimization problems.; The most tangible artifact of this research is the successful development of the zeroth-order shape optimization complex. The complex proved adept at solving both two- and three-dimensional shape optimization problems. The research also provides a demonstrative example of the power and flexibility of machine learning in general and CSs in particular--how they may be leveraged as tools for mechanical engineering design, and insights into their proper application.