Interpretability Improvement of Data-Driven Evolving Fuzzy Systems

In this paper complexity reduction strategies and algorithms are investigated in connection with data-driven incremental learning approaches for specific Takagi-Sugeno fuzzy systems, also called {\em Fuzzy Basis Function Networks}. %The main focus will be attended to {\em FLEXFIS}, a specific variant for an incremental learning of this %kind of fuzzy systems, but in principle %also called {\em FLEXFIS} \cite{LughoferKlement05}. Hereby, the task lies in the improvement of linguistic interpretability of the evolving fuzzy systems not only within an offline post processing phase, i.e. after the complete data-driven modelling process, but also in online mode, i.e. after each incremental learning step. The latter ensures a transparency of the so far trained fuzzy models, which can be an essential requirement for online monitoring tasks, where transparent rules may act as useful explanation for actual system states. The paper is concluded with an evaluation of the proposed methods on high dimensional data coming from an industrial measuring processes, where purely evolving fuzzy systems are compared against linguistically improved evolving fuzzy systems with respect to approximation quality and model complexity.