Blind Multiuser Detection for DS-CDMA Using Independent Component Analysis Neural Network

Abstract

In this paper, the use of Independent Component Analysis (ICA) neural networks for multiuser detection in multipath DS-CDMA communication systems operating in convolutive channels is examined, in both synchronous and asynchronous cases. To take advantage of known parameters, the ICA detector system was initialized by a detector that uses a subspace-based method of estimating channel noise followed by a multistage refinement based on the Hopfield neural network. In order to operate in a totally blind environment, the detector then makes use of an independent component analysis neural network. A number of different ICA learning algorithms were applied to the CDMA detection problem. We explored the use of nonlinear-Hebbian and natural gradient learning, which we believe to be a unique application to the multiuser detection problem. Nonlinear Hebbian learning was found to yield superior results in the most benign cases while natural gradient learning yielded superior results in the harshest environments. This detector gives superior performance over both the conventional single user detector and the LMMSE detector.