A Neuro-based EM-Particle Smoothing Algorithm For Identification of Nonlinear State Space Models

Abstract

The expectation maximization(EM) algorithm and particle filtering have been greatly used in many estimation problems. In this paper, we propose a combination of the EM algorithm and particle smoothing for identification of nonlinear state space models using artificial neural networks. After representing a radial basis function(RBF) neural network as a parametric structure for describing the state transition and output equations of a state space model, the EM algorithm is applied for updating parameters and estimating states of the nonlinear system. Moreover, the particle smoothing algorithm is used at the E phase for state estimation. Simulation studies show the fast convergence rate and satisfactory accuracy of the proposed method in identification of nonlinear plants whose state transition function, output structure or both are unknown.