Practical considerations for Kalman filter training of recurrent neural networks

General recurrent neural networks for application studies have not been widely used, possibly due to the relative ineffectiveness of existing gradient-based training algorithms. An overview of a decoupled extended Kalman filter (DEKF) algorithm for training of recurrent neural network architectures is presented, with special emphasis on application to control problems. Qualitative differences between the DEKF algorithm, which only performs updates to a recurrent network's weight parameters, and a recent EKF formulation of R.J. Williams (1992) that performs parallel estimation of both the network's weights and recurrent node outputs are discussed.<
>