Coevolutionary synthesis of fuzzy decision support systems

Many essential applications in finance, medicine, engineering, and science require increasingly complex decision-making capabilities. There is accordingly a growing demand for decision support systems (DSSs) to assist humans in their tasks. To provide accurate and reliable decision support, a DSS needs not only to be robust in the face of the uncertainty but also to model the decision-making logic in a form that is understandable. Compared with other machine learning methods, fuzzy rule-based systems possess the merits of providing strong approximate reasoning in the presence of imprecise data while representing domain knowledge as a set of interpretable semantic rules. Using them to realize DSSs is thus a most suitable approach yielding powerful fuzzy decision support systems (FDSSs). However, the synthesis of an optimal FDSS with well-balanced accuracy and interpretability is an arduous task. Experience shows that it is very difficult for human experts to manually design its two most important components, the fuzzy membership functions and fuzzy rule base, which directly affect system performance. Ad-hoc architectures, which must be redesigned anew for every application, and improperly chosen parameters typically introduce unwanted biases and unavoidably result in suboptimal systems. Ideally, the decision-making logic should therefore be induced automatically from example and further optimized for the problem at hand. To achieve this goal, a generic approach is needed that can automatically synthesize an accurate and interpretable FDSS, while requiring minimal or no human effort.DOCTOR OF PHILOSOPHY (SCE