Characterizing signal behaviour using genetic programming

Abstract

Our overall goal is to detect automatically that a signal begins to deviate from its previous behaviours, using no other information than a sequence of samples of the signal. In order to detect such changes we use genetic programming to evolve an expression describing how the signal varies over time. One major difficulty when observing such signals is that they typically contain noise and other disturbances. Such disturbances makes it more difficult to find a useful expression characterizing the signal. We have derived a new method that simultaneously evolves a numeral denoting the number of neighbours to use in a moving average of the signal, and an expression characterizing the smoothed signal.