AFCGD: an adaptive fuzzy classifier based on gradient descent

Abstract

In traditional fuzzy classification systems, learning is done from a stationary data distribution. In online rule learning, however, data are non-stationary and change dynamically over time. It confronts the learning process with some new challenges including concept drift. Evolving fuzzy schemes are common solutions in this field which try to handle these issues by in-time modification of their structures. In this regard, a basic challenge is how to apply a fast and simple scheme to modify the rule-base regarding each new sample. This paper introduces an efficient adaptive mechanism named adaptive fuzzy classifier based on gradient descent (AFCGD) for online learning of an evolving fuzzy model. We derive online rule update formulas for modification of the classifier's structure regarding the concept of data to minimize the misclassification error through gradient descent. The updating formulas, which are computationally cheap, allow AFCGD to adjust the rule-base after emergence of new incoming sample. Therefore, it always remains up-to-date and can handle any alteration in the concept of data. AFCGD has simple structure to build; thus, it is so effective in memory usage and computational time. The efficacy of our proposed algorithm has been assessed by some synthetic data and several real-world benchmark problems while comparing with some recent evolving and state-of-the-art classifiers. The proposed method achieves comparable and even better results against other fuzzy and non-fuzzy classifiers in terms of accuracy and run-time.