Minimal resource allocation networks for adaptive noise cancellation

Abstract

This thesis focuses on developing a dynamic minimal radial basis function (RBF) network referred to as Minimal Resource Allocation Network (MRAX) for adaptive noise cancellation. Unlike most of the classical RBF networks in which the number of hidden neurons are fixed a priori, the network structure here is dynamic based on the observation data. The problem of using MRAN for adaptive noise cancellation is developed. MRAX has the same structure as a common RBF but uses a sequential learning algorithm in which hidden neurons are added or pruned depending on certain criteria. If no hidden neuron is added to the network, the exiting network parameters are updated by an Extended Kalman Filter (EKF). Both the growth criterion and the pruning strategy as well as the adjustment the network parameters are performed sequentially with the arrival each input data so as to produce a compact RBF network. A comparison made with the recurrent radial basis function (RRBF) network of Bilings and Fung shows that MRAX produces better noise reduction than the recurrent RBF network with a more compact RBF network architecture.