Separate Initialization of Dynamics and Nonlinearities in Nonlinear State-Space Models

This work focuses on the identification of nonlinear dynamic systems. In particular the problem of obtaining good starting values for the identification of nonlinear state-space models is addressed. A fast and efficient initialization algorithm is proposed, combining the use of methods from the statistical learning community to model the nonlinearities and classic system identification tools to capture system dynamics. The performance of the method is evaluated on simulation examples.