Evolutionary adaptive filtering based on competing filter structures

This paper presents a novel filtering scheme that realizes a general, fully adaptive structure where both coefficients and required memory size are identified automatically. In particular, no distinction between linear or nonlinear models is made, since the filter structure can evolve into either a linear or a second-order Volterra filter. This is achieved by monitoring the mixing variables of various combinations where differently-sized competing filters are used. Using a set of intuitive rules along with desired step sizes for memory size changes, a dynamically growing/shrinking model structure is realized. The effectiveness of the approach for a fast-converging identification of arbitrary unknown systems is shown by means of an acoustic echo cancellation task where realistic linear and nonlinear systems as well as stationary and nonstationary input signals are considered.