Nonlinear Hybrid Systems Identification using Kernel-Based Techniques

Abstract

Hybrid systems can describe in an unified setting many processes which combine continuous/discrete dynamics and logic rules. Their identification from input/output data is however difficult since it requires to jointly solve a classification and estimation problem. Restricting the attention to piecewise linear models, recent research has shown how these difficulties can be successfully faced combining Gaussian regression and stochastic simulation techniques. In this paper we extend this approach to systems composed by nonlinear submodels. Numerical examples regarding estimation of discontinuous functions and identification of piecewise nonlinear dynamic systems are then included to illustrate the potential of the new approach.