Fast learning algorithms for training of feedforward multilayer perceptrons based on extended Kalman filter

The new algorithm based on network decomposition into layers and estimation of the local weights by using extended Kalman filter (EKF) derived from the local optimality criteria is proposed in this paper. Local optimality criteria are formulated on the basis of specific output error backpropagation. Simulation examples show a high efficiency of the proposed algorithm from the point of view of both convergence rate and generalization capabilities.