Approximating large-input space functions with fine localised detail using hybrid neuro-fuzzy systems and multi-layer perceptrons

Abstract

Automated currency condition screening requires an approximate model of the way human experts classify bank notes in terms of their fitness to be loaded into Automatic Teller Machines (ATMs). Part of this model must encapsulate the relationship between features describing aberrations on a bank note and a damage level dictated by the human expert. One general type of relationship is of particular interest. This function has eight inputs, is non-linear but has a generally smooth topography. However, it does contain specific points of localised detail, which Multi-Layer Perceptron (MLP) models tend to smooth out. A new architecture, called Apprentice, is described which uses both MLP and hybrid neuro-fuzzy architectures to overcome this problem and provide the required approximation. The sub-division of work within the new architecture is such that each model carries out the tasks to which it is most suited. Simple fusion algorithms combine the outputs of the two models so that the desired function is approximated with an appropriate accuracy. Results show how well the Apprentice architecture approximates the function in comparison with stand-alone MLPs.