Approaches to the handling of fuzzy input data in neural networks

Neural networks in general lend themselves well to dealing with uncertainty, in that weights are adjusted according to input data. A number of issues arise in neural network research in the handling of uncertain or fuzzy information. These can be divided into several areas: input data; propagation of results through the network: and interpretation of final results. In terms of the fuzzy implementation of neural networks each area is discussed in turn, with possible approaches summarized for each. The introduction of fuzzy input causes substantial problems in most neural network learning algorithms. The learning algorithm must be able to handle interval data. A number of approaches to this problem are outlined. These fall into two main categories: (1) introduction of a preprocessor of some sort in order to handle the fuzzy input; and (2) direct modification of the learning algorithm to handle interval data.<<ETX>>