Identification of nonlinear state space models using an MLP network trained by the EM algorithm

Abstract

Identification of nonlinear state space models when no information is available from the state transition or output model has played an important role in the recent research. In this paper, we propose a new approach for modeling a discrete time nonlinear state space system with a multi layer perceptron (MLP) neural network. The expectation maximization (EM) algorithm is used for joint parameter and state estimation of the proposed structure where the particle smoothing algorithm will be applied for estimating hidden states. Because of the non-affine structure of MLP networks compared with some other models such as radial basis functions, the gradient method is used at the M phase of the EM algorithm for parameter and noise estimation. Simulation studies show the superiority and fast convergence of our proposed structure in identification of nonlinear state space models.