Blind multiuser detection for DS-CDMA using independent component analysis and prior knowledge

In this work, new neural network structures employing Independent Component Analysis (ICA) are developed for blind multiuser detection in multipath DS-CDMA communication systems. A new scheme is developed to allow for the incorporation of prior knowledge into the ICA neural learning algorithms, coupling knowledge specific to the problem of multiuser detection in DS-CDMA communications systems with the adaptive capabilities of advanced, nonlinear, self-learning neural network structures. Making use of a priori knowledge specific to the problem environment, and coupling it with neural network structures in a hybrid manner, results in significant improvement in the bit-error-rate performance of multiuser detectors.; Four approaches to the multiuser detection problem are presented: transversal filters, statistical methods, signal subspace methods and neural network methods. This work makes contributions in each of these areas.; A new semi-blind multistage detector based on the Hopfield neural network is presented. In order to operate in a totally blind environment, ICA neural networks are then employed. The ICA structures displayed superior bit-error-rate performance when compared to the single-user detector, the linear minimum mean squared error detector and the Hopfield neural network-based detector. In addition, a mechanism for incorporating prior probabilities and likelihoods is derived, allowing for knowledge specific to the CDMA multiuser detection problem to be incorporated into the ICA learning algorithm.