The first stage in creating data-driven soft computing models is to derive information from data, towards developing the structure of computational models. An effective way to extract information from data is through granular computing (GrC), which is inspired by the way human naturally groups similar objects together. Using GrC and Fuzzy Sets in modelling engineering systems, one can describe systems in a very transparent and interpretable way. GrC has been applied in the literature, however, there are still issues with data uncertainty and its impact on the interpretability of Fuzzy Logic systems. The consideration for data uncertainty is very important for real applications, where data usually comes from measurements/sensing, with inherent noise and uncertainty. Hence, this thesis aims to investigate methods to address uncertainty in GrC, methods to address variable importance in GrC, methods to address granular overlapping in GrC, and methods to evaluate the resulting impact on the interpretability.
A new data-driven modelling framework based on GrC, Fuzzy Logic, and uncertainty measure is proposed, in which the uncertainty (in this thesis captured via conflict) between information granules is modelled using Shannon entropy. The issue of interpretability due to overlapping is addressed with a new iterative data granulation mechanism that controls the amount of granule overlapping using R-value, a metric that represents the ratio of overlapping areas among categories in a data cluster. In order to characterise the importance of data features, Weighted GrC (W-GrC), a new iterative data granulation technique with evolving feature weighting is proposed. The feature weights are determined based on the current information granules (within-granule variances) and adaptively change in each iteration. Since W-GrC is studied in both Type-1 and Type-2 systems, a new interpretability index for Type-2 Fuzzy Logic systems is proposed based on Nauck’s index, taking into account both upper and lower membership functions.
A thorough set of simulations based on UCI datasets are conducted to demonstrate the effectiveness of each of the frameworks proposed in this thesis. The simulation results demonstrate the potential of all proposed frameworks in improving the predictive accuracy while maintaining good level of interpretability.
