SVMs for System Identification: The Linear Case

In this work we deal with the application of Support Vector Machines for Regression (SVRs) to the problem of identifying linear dynamic systems on the basis of a set of Input/Output samples. Three different examples of simple linear systems will be considered, taking into account both non-recursive and recursive models. When defining the SVR estimating function, several examples of kernels will be employed, with emphasis on the ones that may be more appropriate for describing linear models. As a further step, the exact parameters of the model will be directly estimated from the Support Vector values resulting from the SVR training phase.