Abstract: A Radial Basis Function Neural Network (RBFNN) is a general approximator. In this paper a granular activation function is proposed to improve its learning under the noisy conditions. The granular activation function is also named the interval activation function and it is typicaly the Gaussian function which benefits from having a fixed mean and an uncertain standard deviation. The hidden layer of the proposed network has a total of three parameters to train that it consists the means, the lower bounds of the standard deviations and the higher bounds of the standard deviations of the Gaussian functions. The output layer parameters for training are the means of the interval weights and the intervals of the weights. “K-Means clustering algorithm” method is used to train these parameters. The purpose of the above learning method is regarded as one of the granular method presenting the bottom-up granulation which causes the input vectors clustered in the larger granules in the hidden layer. Gradient descend method is also used to train these parameters to compare with this novel method. The structure is tested with or without noisy data to identify a nonlinear dynamic system with multiple time delays and to predict a chaotic model, Mackey-Glass. It has been shown that the sensibility related to input alterations reduces because of using the granular activation function in RBF Neural Network structure and the response of Granular RBF Neural Network with noisy data is better than RBF Neural Network.
Rights and permissions | |
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License. |