EFFICIENT IMPLEMENTATION OF SEPARABLE LEAST SQUARES FOR THE IDENTIFICATION OF COMPOSITE LOCAL LINEAR STATE-SPACE MODELS

The efficient implementation of separable least squares identification of nonlinear systems using composite local linear state-space models is discussed in this paper. A full parametrization of system matrices combined with projected gradient search is used to identify the model. This combined approach reduces the number of iterations and improves the numerical condition of the optimization algorithm. Further enhancements result from using numerical tools, e.g. the QR-decomposition, and efficient approaches to compute the separable least squares matrices and gradients of the cost functions. Monte-Carlo simulations are used to show the effectiveness of the proposed approach.