Sequential Learning Algorithm of Neural Networks Systems for Time Series

摘要

This article describes a new structure to create a RBF neural network that uses regression weights to replace the constant weights normally used. These regression weights are assumed to be functions of input variables. In this way the number of hidden units within a RBF neural network is reduced. A new type of nonlinear function is proposed: the pseudo-gaussian function. With this, the neural system gains flexibility, as the neurons possess an activation field that does not necessarily have to be symmnetric with respect to the centre or to the location of the neuron in the input space. In addition to this new structure, we propose a sequential learning algorithm, which is able to adapt the structure of the network; with this, it is possible to create new hidden units and also to detect and remove inactive units. We have presented conditions to increase or decrease the number of neurons, based on the novelty of the data and on the overall behaviour of the neural system, (for example pruning the hidden units that have lowest relevance to the neural system using Orthogonal Least Squares (OLS) and other operators), respectively. The feasibility of the evolution and learning capability of the resulting algorithm for the neural network is demonstrated by predicting time series.