On Model Selection for State Estimation for Nonlinear Systems

State filters can be used to produce online estimates of the state of a process. If an exact model for the true process is not known, but multiple candidate models are available to describe the current behavior of the true system, it is necessary to select that model that leads to the optimal state estimates. This paper describes a novel approach for model selection for state estimation by comparing the expected weighted prediction error using estimated states of different candidate models. The expected prediction error can not be computed exactly, but can be estimated using a newly derived generalized version of the FPE selection criterion. A simulation example of a time varying system is used to illustrate the performance of the selection method.