Optimal data-driven rule extraction using adaptive fuzzy-neural models

Neural network and fuzzy rule-based approaches to data-driven modeling have recently gained a lot of attention. The property of universal approximation makes it possible to imitate a large class of complex nonlinear systems with a certain degree of accuracy, while the incorporation of rules and fuzzy sets allows for concise representation of expert knowledge. Most of these hybrid, data-driven modeling methods emphasize accurate approximation, while little attention is paid to the issue of transparency, which represents the ease of understanding the model behavior. Very recently, attempts have been made to compromise these two aspects. This dissertation approaches the problem of optimal data-driven modeling from an original point of view. The most important optimality factors in fuzzy modeling are first introduced and their influence is discussed. Based on this discussion, several key ideas are proposed and elaborated to provide a unified framework for optimal neuro-fuzzy model construction. The use of relational rules in fuzzy modeling is described. A relational rule incorporates relationships between the model inputs into the reasoning process, which improves the balance between the accuracy and the transparency of the model. An identification method is then proposed for data-driven determination of the model structure. In an easy and straightforward manner, the method automatically detects strongly correlated inputs. Finally, an optimality criterion and a search method are proposed as means for optimal model selection.