Sparse Generalized Kernel Modeling for Nonlinear Systems

A generalized kernel modeling approach is proposed for identification of discrete-time nonlinear systems. Each kernel regressor in the generalized kernel model has an individually fitted diagonal covariance matrix which is determined by maximizing the correlation between the regressor and training data. A state-of-the-art construction algorithm based on orthogonal least squares regression with leave-one-out test statistic and local regularization is applied to select a parsimonious generalized kernel model from the full regression matrix. The effectiveness of the proposed nonlinear modeling approach is demonstrated by the experimental results involving one simulated system and two real data sets.