NEURAL NETWORK CONTROLLED ADAPTIVE FILTERS

Abstract

Recently SZTIPÁNOVITS has proposed [2,3] an adaptive processing system - an adaptive FIR filter structure - that consists of a resonator based digital filter (RBDF) and a neural network. The RBDF is a highly parallel structure with several structural and implementational advantages, which in the proposed case performs a recursive transformation as well. This paper focuses on the advantages and disadvantages of the proposed composite structure. Some improvements are suggested, for example extending the parallelism of the structure to the neural network as well, which results in a better convergence rate of the training procedure. On the other hand, the convergence rate can be improved by using the combination of the output error and the transform domain component error. Finally the structure is extended to adaptive IIR filtering problems which requires the modification of the resonator based IIR structure.