Real-time independent component analysis based on gradient learning with simultaneous perturbation stochastic approximation

Abstract

We present a novel algorithm for independent component analysis (ICA) based on gradient learning with simultaneous perturbation stochastic approximation (SPSA). This algorithm can work well both in batch mode and in on-line mode of ICA processing. It converges very fast even for non-stationary, and/or non-identically independent distributed (non-I.I.D.) signals, so that the algorithm is very suitable for most real-time applications. In this paper, theories and implementations of the algorithm are described. Results of computer simulation are also presented to demonstrate the effectiveness. © Springer-Verlag 2004.