Nonlinear state space learning with EM and neural networks
Jordan Freitas, Mahesan Niranjan, AH Gee · 2002
In this paper, we derive the EM algorithm for nonlinear state space models. We show how this algorithm, in conjunction with the well known techniques of Kalman smoothing, can be used for nonlinear system identification. A multilayer perceptron, whose derivatives are computed by backpropagation, is used to generate the measurements mapping. We found that the methodic is intrinsically very powerful, simple, elegant and stable. However, it exhibits very slow convergence.