Identification of nonlinear systems using misspecified predictors

Identification of nonlinear systems is an important albeit difficult task. This work considers parameter estimation, using the prediction error method, of the class of models that fit into a nonlinear state space formulation. Finding the optimal predictor for such nonlinear models, if at all possible, often requires significant effort. As an alternative, techniques from indirect inference are used to circumvent this problem. A misspecified predictor, parameterized by a new set of parameters, is used in lieu of the optimal predictor. These new parameters are found numerically by using simulations of the model to be identified. The proposed method is applied to simulation examples and real process data with encouraging results.