Neurofuzzy Approximator based on Mamdani's Model

Neurofuzzy approximators can take on numerous alternatives, as a consequence of the large body of options available for defining their basic operations. In particular, the extraction of the rules from numerical data can be conveniently based on clustering algorithms. The large number of clustering algorithms introduces a further flexibility. Neurofuzzy approximators can treat both numerical and linguistic sources. The analysis of approximator sensitivity to the previous factors is important in order to decide the best solution in actual applications. This task is carried out in the present paper by recurring to illustrative examples and exhaustive simulations. The results of the analysis are used for comparing different learning algorithms. The underlying approach to the determination of the optimal approximator architecture is constructive. This approach is not only very efficient, as suggested by learning theory, but it is also particularly suited to combat the effect of noise that can deteriorate the numerical data.