On the training of DS-CDMA neural-network receivers

In this paper we prove formally that the optimum (nonlinear) DS-CDMA single-user decision boundary exhibits the following properties: (i) it is symmetric with respect to the origin and (ii) as it is traversed away from the origin, it converges to a hyperplane parallel to the MF decision boundary. Then, we translate properties (i) and (ii) to a set of constraints that can be used by any optimization algorithm for the selection (training) of the parameters of a general multi-layer-perceptron neural-network receiver. Using these constraints, the number of parameters to be optimized is reduced by nearly 50% for large-size networks, which effectively doubles the speed of any training procedure. Furthermore, we utilize properties (i) and (ii) to develop a new initialization scheme that provides additional improvements on the convergence rate and can be used by any recursive optimization algorithm. As a representative case study we consider the back-propagation (BP) algorithm and develop a constrained version of it that incorporates both the proposed constraints and the proposed initialization. The convergence rate enhancement achieved by constrained-BP is illustrated by simulations.