Title: Optimal kernel filtering for system identi
Authors: Jose Principe xv
Abstract: This talk will summarize recent advances in nonlinear adaptive ltering. Designing adaptive lters in Reproducing Kernel Hilbert Spaces (RKHS) bridges the established procedures of adaptive lter theory with kernel methods. The end result is a family of lters that are universal approximators in the input space, that have convex performance surfaces (no local minima for large number of samples), and that are on-line, i.e. they adapt with every new sample of the input. Moreover, we will show that contrary to common believe some of its members do not need explicit regularization, e.g. the Kernel Least Mean Squares (KLMS) is well posed in the sense of Hadamard. They are however growing structures therefore special techniques need to be included to curtail their growth. Although the talk will focus on system identication, similar techniques can be applied to the kernel algorithms of machine learning.
Publication date: 2012-07-16
Online entry date: 2013-05-12
Conference: CONTROLO’2012
Publication pdf: Click for full publication