Least Squares Support Vector Machines Time: 010912 10:15-12:00 Lecture
010912 13:15-15:00 Discussion
010913 10:15-12:00 Lecture
010913 13:15-15:00 Discussion
010914 10:15-12:00 Lecture
010914 13:15-15:00 Discussion

Speaker: Johan Suykens, Katholieke Universiteit Leuven Suitable reading Nonlinear Modeling and Support Vector Machines
Johan Suykens
IEEE Instrumentation and Measurement Technology Conference
Budapest, Hungary, May 21st-23rd, 2001
and the references therein
Language: English Abstract: In the last decade, neural networks have proven to be a powerful methodology in a wide range of fields and applications. Although neural nets have often been presented as a kind of miracle approach, reliable training methods exist nowadays mainly thanks to interdisciplinary studies and insights from several fields including statistics, circuit-, systems and control theory, signal processing, information theory, physics and others. Despite many of these advances, there still remain a number of weak points such as the existence of many local minima solutions and how to choose the number of hidden units. Major breakthroughs are obtained at this point with a new class of neural networks called support vector machines (SVMs), developed within the area of statistical learning theory and structural risk minimization. SVM solutions are characterized by convex optimization problems (typically quadratic programming). Moreover, the model complexity (e.g. number of hidden units) also follows from solving this convex optimization problem. The method is kernel based and allows for linear, polynomial, spline, RBF, MLP kernels and others.

In the first part of this course we explain the theory of linear and non-linear SVM for solving classification and nonlinear function estimation problems. In the second part, we focus on least squares support vector machines (LS-SVMs) which involve solving linear systems instead of QP problems. The method is capable of solving highly nonlinear and noisy black-box modelling problems, even in high dimensional input spaces. Issues of robust nonlinear estimation and sparse approximation will be discussed, together with hyperparameter selection methods. Several frameworks will be explained, including Bayesian learning and VC theory. In the third part we present first extensions of LS-SVM methods to recurrent networks and use in optimal control problems. The huge potential of (LS)-SVM methodologies will be continuously illustrated on a large variety of examples and case studies.