Learning capabilities of ELM-trained Time-Varying Neural Networks

Abstract

System identification in nonstationary environments surely represents a challenging problem. The authors have recently proposed an innovative neural architecture, namely Time-Varying Neural Network (TV-NN), which has shown remarkable identification capabilities in this kind of scenarios. It is characterized by time-varying weights, each being a linear combination of a certain set of basis functions. This inevitably increases the network complexity with respect to the stationary NN counterpart and in order to keep the training time low, an Extreme Learning Machine (ELM) approach has been proposed by the same authors for TV-NN learning, instead of Back-Propagation based techniques. However the learning capabilities of TV-NN trained by means of ELM have not been investigated in the literature and in this contribution such a lack is faced: the theoretical foundations of ELM usage for TV-NN are analytically discussed, by extending the corresponding results obtained in the stationary case study. © Springer International Publishing Switzerland 2014.