A Comparative Study of Three Different Topologies of Neural Network–based Multiuser Detectors of WCDMA signals

One of the main obstructions in Multi-Access systems is the Multi-Access Interference (MAI) between signals that sharing the same channel. Specially, in the CDMA systems where all users share the same channel all the time. The objective of this work is to compare three different Artificial Neural Network (ANN)-based multiuser detectors for Wideband Code Division Multiple Access WCDMA system built to combat the effect of the MAI. The system employs ANN Multiuser Detectors (MUD) for a synchronous Wideband Code Division Multiple Access (WCDMA) mobile system uplink channel with variable number of paths and complex-valued Rayleigh fading and uses a bank of Rake receivers as pre-detectors. The three types of the ANNs employed are the Backpropagation FeedForward (FF), Radial Basis Function (RBF), and The Self Organizing Feature Maps (SOM) which is an unsupervised ANN involves clustering of patterns into similar groups. Comparing results for the different types of NN MUDs, the FFNN MUDs give better performance when trained properly, but their training is the main difficult issue and takes very long time. While RBF MUDs’ training is easy and fast, attention is needed to set the training parameters. On the other hand SOM MUDs don’t need any training since they are unsupervised NNs, but their implementation is not easy. RBF and SOM MUDs give lower performance compared to the FFNN MUDs.