GT2FC: An Online Growing Interval Type-2 Self-Learning Fuzzy Classifier

Abstract

In this paper, we propose a Growing Type-2 Fuzzy Classifier (GT2FC) for online rule learning from real-time data streams. While in batch rule learning, the training data are assumed to be drawn from a stationary distribution, in online rule learning, data can dynamically change over time becoming potentially nonstationary. To accommodate dynamic change, GT2FC relies on a new semi-supervised online learning algorithm called Growing Gaussian Mixture Model (2G2M). In particular, 2G2M is used to generate the type-2 fuzzy membership functions to build the type-2 fuzzy rules. GT2FC is designed to accommodate data online and to reconcile labeled and unlabeled data using self-learning. Moreover, GT2FC maintains low complexity of the rule base using online optimization and feature selection mechanisms. GT2FC is tested on data obtained from an ambient intelligence application, where the goal is to exploit sensed data to monitor the living space on behalf of the inhabitants. Because sensors are prone to faults and noise, type-2 fuzzy modeling is very suitable for dealing with such an application. Thus, GT2FC offers the advantage of dealing with uncertainty in addition to self-adaptation in an online manner. For illustration purposes, GT2FC is also validated on synthetic and classic UCI data-sets. The detailed empirical study shows that GT2FC performs very well under various experimental settings.