A neuro-fuzzy classification of objects and their states

Abstract

In this paper, we present the development and some investigation results of a neural fuzzy network (NFN) to solve classification problems of biological objects and their states. A general approach to the problem of automatic classification of objects is studied. In this work, we use a Fisher's Iris data set for the object classification problem, and the assessment of the emotional state of objects is given on the data set of the balance scale. The NFN is a combina-tion of fuzzy inference systems and a neural network. The structure of the NFN takes the form of a multilayer unidirectional network consisting of an input layer, a fuzzy activation function (AF) layer, a fuzzification layer, a defuzzification layer, a normalization layer and an output layer. In the fuzzy AF layer, we use fuzzy AFs that were developed and investigated in our previous studies. Herein, four fuzzy AFs corresponding to four types of NFN are investigated. The training process of the NFN is conducted by the use of the K-mean cluster method to determine antecedent network parameters. During the training process, a scaled conjugate gradient (SCG) algorithm is used to reduce the computational effort of every learning iteration, and therefore, enhance the learning speed. The testing process is carried out with Fisher's Iris data sets and the balance scale. These data sets are classical and often used to illustrate the performance of various statistical classification algorithms. The results obtained by various types of fuzzy AF have confirmed the validity of the current NFN to solve classification problems. A comparative analysis of the current four NFN types and the NFN using the Gaussian AF from a previous pa-per in terms of efficiency and accuracy is performed in this work.