Abstract:
In the context of nonlinear system identification and forecasting of time-series, important challenges are related to the incorporation of prior knowledge and to the estimation of such models from large scale datasets. In this talk, estimation techniques based on Least-Squares Support Vector Machines (LS-SVM) are used in the context of nonlinear system identification. The primal-dual optimization framework of LS-SVMs can be exploited and extended to incorporate structured elements available from prior knowledge about the identification problem at hand.
It will be shown that the prior knowledge becomes part of the kernel function of the model, such that it can be used directly to evaluate the models at new datapoints. This property makes a contribution in terms of modularity of the model formulation, in the sense that different types of prior knowledge can be tested in practice simply by changing the kernel function being used. Furthermore, large scale versions of the different LS-SVM extensions can be formulated in primal space by using the Nystrom method. By considering each of the developed extensions as building blocks, a modular framework for the case of nonlinear system identification is further proposed. Practical examples show the benefits of this approach.
Anout the Speaker
Marcelo Espinoza has obtained the degrees of Civil Engineer and M.Sc. in Applied Economics from Universidad de Chile, and the degrees of Master in Artificial Intelligence and Ph.D in Electrical Engineering from the Katholieke Universiteit Leuven, Belgium. Currently he is a postdoctoral researcher at the SISTA Research Division of the Electrical Engineering Dept. at the K.U.Leuven, with interests in nonlinear system identification, time series prediction, kernel methods and dynamical systems