A multilayered neuro-fuzzy classifier with self-organizing properties

Abstract

A novel self-organizing neuro-fuzzy multilayered classifier (SONeFMUC) is suggested in this paper which is composed of small-scale interconnected fuzzy neuron classifiers (FNCs) arranged in layers. The model provides a different perspective for generating a new class of hierarchical classifiers with multilevel classifiers combination. At each layer, parent FNCs are combined to construct a descendant FNC at the next layer with enhanced classification qualities. The generic FNCs exhibit an original structure including four modules: (1) The fuser aggregates the decision support outputs of the parent FNCs using a fusion operator. (2) The data splitting module divides data into correctly classified patterns with high degree of confidence which are handled by the fuser and ambiguous ones. (3) The last two modules are realized by fuzzy rule-based systems and implement a neuro-fuzzy classifier within each FNC, used to improve the classification accuracy of ambiguous patterns. The former performs feature transformations to an intermediate output space while the latter provides the decision supports of non-confident patterns. Unlike traditional classifiers, the classification mapping is accomplished in SONeFMUC by performing successive decision fusions and feature transformations. SONeFMUC structure is determined sequentially via a self-constructing learning algorithm. Structure learning inherently implements feature selection, considering the most informative attributes as model inputs. To improve classification accuracy, the models obtained after structure learning are optimized using a parameter learning scheme based on genetic algorithms. The effectiveness of our approach is tested on a set of benchmark classification problems. Experimental results show that the proposed neuro-fuzzy classifier is favorably compared to other well-known classification techniques of the literature.