A new approach for fuzzy neural network weight initialization
F. Abed Rouai, M. Ben Ahmed · 2002
We develop a method for extracting a fuzzy model directly from input-output data. Our approach is based on three fundamental factors: (1) The use of entropy theory for feature selection, (2) the identification of the fuzzy model structure in one single step by the incremental applying of the fuzzy-c-means algorithm directly to the Cartesian input-output data space, (3) the introduction of a new method "semi-Lambda-cut-density" based on the /spl lambda/-cut concept, for setting the initial weights in neurofuzzy networks (NFN). The NFN is trained by a backpropagation algorithm. A comparative study on benchmark examples is conducted and shows that our method solves the trade-off between the use of a small number of rules and the achievement of a fuzzy model best performance index.