Tracking of multivariate time-variant systems based on on-line variable selection

Sung-Phil Kim, Y.N. Rao, D. Erdogmus, J.C. Principe
Proceedings of the 2004 14th IEEE Signal Processing Society Workshop Machine Learning for Signal Processing, 2004.  
Tracking time-variant systems has been of great interest in many engineering fields. Specifically, when system statistics change both in space (multivariate) and time with a short stationary regime, conventional adaptive algorithms suffer from the tradeoff between convergence rate and accuracy. I n this paper, we propose a tracking system consisting of a linear adaptive system accompanied by an on-line variable selection algorithm that is based on the least angle regression algorithm. This
more » ... ithm explicitly employs local (in time) correlation between the input and the output of an unknown system to select a subset of input variables at every time step. Therefore, it enables the multivariate adaptive Wter to track the temporal changes of correlated variables. Simulations involving tracking of multi-channel timevariant systems demonstrate superior performance of the proposed approach when compared with the Conventional methods.
doi:10.1109/mlsp.2004.1422966 fatcat:xxbhi7wdeba2lj52qbehprr76a