Chaotic time-series prediction using resource-allocating RBF networks

One of the main problems associated with artificial neural networks on-line learning methods is the estimation of model order. In this paper, we report about a new approach to constructing a resource-allocating network exploiting weights adaptation using QRD-based recursive least-squares technique. Further, we studied the performance of Dynamic Cell Structures algorithm for on-line adaptation of centers positions. The proposed method was tested on the task of Mackey-Glass time-series prediction. Order of resulting networks and their prediction abilities were superior to those previously reported by Platt [6]. 1 Introduction The resource-allocating network (RAN) was introduced by Platt [6] and further extended by McLachlan and Lowe [5]. Similar approach was also outlined in [1]. RANs were usually based on radial basis function (RBF) networks. Associated with these, two essential problems -- weights adaptation and center selection -- need to be solved. Our approach to this task can be c..