Use of fractional powers to moderate neuronal contributions

A learning algorithm is described which permits the incorporation of nodes in the network which may contribute to fractional powers, rather than at full strength. This approach has implications for the implementation of fuzzy neural networks in which membership functions can be used to determine the appropriate fractional exponents. In turn, this structure leads to the possibility of a variety of network architectures, where each layer can be viewed as a specific fractional layer. The method is illustrated in a medical application, in which a decision model is developed for the analysis of time series data obtained through chromatographic analysis of urine taken from patients with melanoma. The resulting model shows good results in its ability to predict the presence of metastasis in these patients