Tracking period M discrete time series

Abstract

In the field of signal processing and communications, many time series are period M, discrete and noisy in nature. Not only do we often want to accurately estimate the M parameters of such a digital signal, we also require to learn the firing sequence of the parameters so that we can predict the next event in time. This is achieved by combining two neural nets. The first net clusters on the time difference between successive events to accurately estimate the parameters and the second net learns to predict which parameter is to be next in the sequence. Hence we are effectively able to track a period M discrete time series. This neural method, in general, requires only a few complete frames or cycles of the time series in order to converge, even for complicated sequences.