A novel randomized approach to nonlinear system identification

Classical incremental approaches for the identification of polynomial NARX/NARMAX models often yield unsatisfactory results in terms of structure selection, which is crucial for model reliability over long-range prediction horizons. This paper embeds the nonlinear identification problem into a probabilistic framework and presents a novel randomized algorithm for structure selection. The approach is validated over different models by means of Monte Carlo simulations, and is shown to outperform competitor probabilistic methods in terms of both reliability and computational efficiency.