Nonlinear state-space modeling using recurrent multilayer perceptrons with unscented Kalman filter

The most important characteristics required in dynamic systems modeling using neural networks are fast convergence and generalization capability. To achieve these, this paper presents an approach to nonlinear state-space modeling using recurrent multilayer perceptrons (RMLP) trained with the unscented Kalman filter (UKF). The recently proposed UKF, which is proper to state-space representation, offers not only fast convergence but also derivative-free computations and an easy implementation, compared with the extended Kalman filter (EKF) widely used for neural networks. Through modeling experiments of nonlinear systems, the effectiveness of the RMLP with the UKF is demonstrated.