Improvements of Analog Neural Networks Based on Kalman Filter

Abstract

In the paper, original improvements of recurrent analog neural networks, which are based on Kalman filter, are presented. These improvements eliminate some disadvantages of the classical Kalman neural network and enable a real time processing of quickly changing signals, which appear in adaptive antennas and similar applications. This goal is reached using such circuit elements, which increase the convergence rate of the network and decrease the dependence of convergence rate on the ratio of eigenvalues of the correlation matrix of input signals.