Empirical data modelling algorithms: additive spline models and support vector machines

Empirical data modelling techniques are widely used in the control field, from simple white-box, linear parameter identification schemes to black-box nonlinear models. Non-linear, semi-parametric model building algorithms have been extensively studied over the past ten years, and despite their success in many applications where prior information is lacking or incorrect, verification and validation is notoriously difficult. One of the key aspects of verification and validation is transparency, where the network's generalisation abilities are explicitly represented. The paper describes two approaches for building an ANOVA representation of non-linear, multivariate data: one based on forwards selection and backwards elimination spline models and the other using a support vector machine with an ANOVA-kernel decomposition.