A data-reusing gradient descent algorithm for complex-valued recurrent neural networks

Abstract

A class of data-reusing (DR) learning algorithms for real-valued recurrent neural networks (RNNs) employed as nonlinear adaptive filters is extended to the complex domain to give a class of data-reusing learning algorithms for complex valued recurrent neural networks (CRNNs). For rigour, the derivation of the data-reusing complex real time recurrent learning (DRCRTRL) algorithm is undertaken for a general complex activation function. The analysis provides both error bounds and convergence conditions for the case of contractive and expansive complex activation functions. The improved performance of the data-reusing algorithm over the standard one is verified by simulations on prediction of complex valued signals.