Systems Seminar

Slow-Fast Adaptive Algorithms and Averaging Theory

Prof. Jim Bucklew
UW ECE Department

Abstract

This talk will give an overview and introduction to stochastic averaging theory and its application to the analysis of the long term behavior of certain types of stochastic adaptive systems. The vast majority of adaptive systems of engineering interest are nonlinear, stochastic, feedback systems: a fact which complicates considerably the task of analytically predicting their performance. The usual tool to investigate these systems is the so-called averaging theory. We will give a simple heuristic explanation of what averaging theory is in the context of certain types of recursive algorithms (with much arm waving) and then discuss a couple of simple examples (one of which will be everyone's favorite ---the Least Mean Square Algorithm (LMS)). We then will give several examples of algorithms that do not fall within the standard averaging framework e.g. standard adaptive IIR algorithms and recurrent neural networks training algorithms. We will then finish the audience off with a discussion of some partial results that extend the current theory.

Time and Place: Wed., Mar. 17, 3:30-4:30 pm in 4610 Engr. Hall.

SYSTEMS SEMINAR WEB PAGE: http://www.cae.wisc.edu/~gubner/seminar/