Improved neural network equalization by the use of maximum covariance weight initialization

In this paper we focus on adaptive equalization of binary telecommunication signals in a baseband digital communication system. We have studied the use of multilayer perceptron (MLP) networks for equalizing binary data bursts in a channel that introduces both intersymbol interference and noise to the transmitted signal. Conventionally the weights of the MLP network are initialized randomly. Here we have studied the use of maximum covariance (MC) initialization scheme in the weight initialization. By applying MC initialization we have been able to speed up the convergence and decrease the total computational load of the system. This is very important in telecommunications, where it is often not possible to use systems that require a lot of computation.