A comparison between Kalman filters and recurrent neural networks

The performance of a recurrent neural network signal estimator is compared to that of the basic discrete time Kalman filter for a number of simulated systems. The selected systems diverge from the assumptions upon which the Kalman filter is based. The architecture of the recurrent neural network is described. The training algorithm is based on the conjugate gradient optimization method. The neural network was found to provide improved performance over the Kalman filter in several cases. In all cases tried, the neural net was found to never perform significantly worse than the Kalman filter.<
>