An efficient method for constructing fuzzy rules

Recent advances have merged artificial neural networks with fuzzy logic to generate automatically and to tune membership functions, rules and inference systems. However, these tools are not simple and can generate very complicated error surfaces with multiple local optimums that are traps for the learning algorithm. With the clustering methods automatic rule generation and optimal shape of membership functions can be generated. In this paper a different approach is considered. Instead of generating cluster centers, some vectors are chosen by using certain described criteria. The structure of the learning machine is defined during training. The Vapnik Chervonenkis (VC) dimension is introduced as a measure of the capacity of the learning machine. A prediction of the expected error on the yet unseen examples can be estimated with the help of the VC dimension. The structural risk minimization principle is introduced to construct a machine with the lowest expected error.