Construction of a multilayer perceptron for a piecewise linearly separable classification problem

The principle behind fuzzy control is rule-based function approximation. Fuzzy rules can also be used to design classification systems. However, the common max-min inference leads to quite restricted classification systems that decide locally on the basis of only two variables. In order to build more flexible systems, it is recommendable to use other operators, for instance the Lukasiewicz t-norm instead of the minimum. It can be shown that such a fuzzy classifier basically constructs a set of (hyper-)planes to separate the classes. Since multilayer perceptrons rely in principle on the same strategy, our idea is to construct a fuzzy classifier on the basis of expert knowledge and then to transform it into a multilayer perceptrons in order to apply learning techniques to further reduce the classification error. The paper concentrates on the construction of a multilayer perceptron, based on a classification that uses hyperplanes for class separation.