Learning hybrid neuro-fuzzy classifier models from data: to combine or not to combine?

To combine or not to combine? Though not a question of the same gravity as the Shakespeare’s to be or not to be, it is examined in this paper in the context of a hybrid neuro-fuzzy pattern classifier design process. A general fuzzy min-max neural network with its basic learning procedure is used within six different algorithm independent learning schemes. Various versions of cross-validation, resampling techniques and data editing approaches, leading to a generation of a single classifier or a multiple classifier system, are scrutinised and compared. The classification performance on unseen data, commonly used as a criterion for comparing different competing designs, is augmented by further four criteria attempting to capture various additional characteristics of classifier generation schemes. These include: the ability to estimate the true classification error rate, the classifier transparency, the computational complexity of the learning scheme and the potential for adaptation to changing environments and new classes of data. One of the main questions examined is whether and when to use a single classifier or a combination of a number of component classifiers within a multiple classifier system.