Antecedent selection in fuzzy rule interpolation using feature selection techniques

Fuzzy rule interpolation offers a useful means for enhancing the robustness of fuzzy models by making inference possible in systems of only a sparse rule base. However in practical applications, the rule bases provided may contain irrelevant, redundant, or even misleading antecedents, which makes the already challenging tasks such as inference and interpolation even more difficult. The majority of the techniques developed in the literature assumes equal significance of rules and their antecedents, which may lead to biased or incorrect reasoning outcomes. This paper investigates similar problems being tackled in the area of feature selection, in an attempt to identify techniques that can be applied to measure the significance of rule antecedents. In particular, two feature evaluation methods based on correlation analysis and fuzzy-rough set theory have been examined, in order to reveal their effectiveness in determining the importance of individual antecedents, and their capabilities for discovering subsets of antecedents that provide similar reasoning accuracies as a larger set of antecedents used in the original rules. In addition, the significance values measured by the proposed method are treated as the weights associated with the relevant rule antecedents, in an effort to facilitate more appropriate selection of, and interactions with the rules in performing both forward and backward fuzzy rule interpolation via scale and move transformation-based methods. Experimental studies based on a practical scenario concerning terrorist activities and also synthetic random data are conducted, demonstrating the potential and efficacy of the proposed work.