On transient dynamics, off-equilibrium behaviour and identification in blended multiple model structures

The use of multiple-model techniques has been reported in a variety of control and signal processing applications. However, several theoretical analyses have recently appeared which outline fundamental limitations of these techniques in certain domains of application. In particular, the identifiability and interpretability of local linear model parameters in transient operating regimes is shown to be limited. Some modifications to the basic paradigm are suggested which overcome a number of the problems. As an alternative to parametric identification of blended multiple model structures, nonparametric Gaussian process priors are suggested as a means of providing local models, and the results compared to a multiple-model approach in a Monte Carlo simulation on some simulated vehicle dynamics data.