Ready, Get Set, Switch!
by Kim Levy
Abstract: We introduce the problem of real-time tracking of a moving target, such as a signal, through noisy measurements. We propose an iterative method, stochastic approximation, to successively obtain better estimates of the target. A weight, or step size, is given to the feedback provided by the measurements. For tracking in a non-stationary environment, we use a constant step size. A large step size leads to faster convergence towards the target but higher variance around the target, while on the contrary a small step size will provide a more precise approximation but requires a greater number of iterations. The aim of this talk is to demonstrate how change detection can significantly improve the selection of the stepsize and the performance of the algorithm during the transitions between regimes. We present results for a new regression based hypothesis test and for the well-known Page-Hinkley test which allow us to get set for a switch.
For More Information: Owen Jone O.D.Jones@ms.unimelb.edu.au