Abstract:
Over the last decade, access to personal wireless communication networks has evolved to a point of necessity. Attached to the phenomenal growth of the telecommunications industry in recent times is an escalating demand for higher data rates and efficient spectrum utilization. This demand is fuelling the advancement of third generation (3G), as well as future, wireless networks. Current 3G technologies are adding a dimension of mobility to services that have become an integral part of modem everyday life. Wideband code division multiple access (WCDMA) is the standardized multiple access scheme for 3G Universal Mobile Telecommunication System (UMTS). As an air interface solution, CDMA has received considerable interest over the past two decades and a great deal of current research is concerned with improving the application of CDMA in 3G systems. A factoring component of CDMA is multiuser detection (MUD), which is aimed at enhancing system capacity and performance, by optimally demodulating multiple interfering signals that overlap in time and frequency. This is a major research problem in multipoint-to-point communications. Due to the complexity associated with optimal maximum likelihood detection, many different sub-optimal solutions have been proposed. This focus of this dissertation is the application of neural networks for MUD, in a direct sequence CDMA (DS-CDMA) system. Specifically, it explores how the Hopfield recurrent neural network (RNN) can be employed to give yet another suboptimal solution to the optimization problem of MUD. There is great scope for neural networks in fields encompassing communications. This is primarily attributed to their non-linearity, adaptivity and key function as data classifiers. In the context of optimum multiuser detection, neural networks have been successfully employed to solve similar combinatorial optimization problems. The concepts of CDMA and MUD are discussed. The use of a vector-valued transmission model for DS-CDMA is illustrated, and common linear sub-optimal MUD schemes, as well as the maximum likelihood criterion, are reviewed. The performance of these sub-optimal MUD schemes is demonstrated. The Hopfield neural network (HNN) for combinatorial optimization is discussed. Basic concepts and techniques related to the field of statistical mechanics are introduced and it is shown how they may be employed to analyze neural classification. Stochastic techniques are considered in the context of improving the performance of the HNN. A neural-based receiver, which employs a stochastic HNN and a simulated annealing technique, is proposed. Its performance is analyzed in a communication channel that is affected by additive white Gaussian noise (AWGN) by way of simulation. The performance of the proposed scheme is compared to that of the single-user matched filter, linear decorrelating and minimum mean-square error detectors, as well as the classical HNN and the stochastic Hopfield network (SHN) detectors. Concluding, the feasibility of neural networks (in this case the HNN) for MUD in a DS-CDMA system is explored by quantifying the relative performance of the proposed model using simulation results and in view of implementation issues.
