On-Line Interpretability by Fuzzy Rule Base Simplification and Reduction

Authors

Abstract

With the advent of data-driven fuzzy modelling algorithms interpretability of the models became a major concern since it is difficult to avoid some degree of redundancy and unnecessary complexity when fuzzy models are acquired from data. Mechanisms for merging membership functions and rule base simplification are studied in order to improve the interpretability of the fuzzy models. The main innovation introduced in our work is the on-line implementation of these mechanisms, i.e., they are incorporated in the learning process and not only applied when the learning process is concluded, as it is common practice. This allows the minimization of redundancy and complexity of the models during its development, achieving transparency automatically. The on-line learning technique used is the evolving Takagi-Sugeno (eTS) fuzzy models, based on a novel on-line learning algorithm that recursively develops the model structure and parameters. Results for a benchmark data set, the Box-Jenkins time-series prediction, are presented.

Keywords

On-line learning, eTS fuzzy models, recursive fuzzy clustering, interpretability, transparency, similarity measures, rule base simplification and reduction.

Subject

On-line Learning Algorithms

Conference

European Symposium on Intelligent Technologies, Hybrid Systems and their implementation on Smart Adaptive Systems, EUNITE2004, June 2004