Smoothness priors support vector method for robust system identification

Abstract

Support vector regression (SVR) is applied to identify linear dynamical systems. The system model is described in terms of basis functions, such as Laguerre or Kautz filters, and the coefficients of the expansion are determined using support vector machine regression. In SVR, the variance of the parameter estimates is bounded by the inclusion of a quadratic regularisation term. Here, model complexity is efficiently reduced by taking the regularisation term as a frequency-domain smoothness prior, defined as the square of the ℒ2-norm of the mth order derivative of the frequency response function.