The use of fuzzy neural networks for feature/sensor selection

In diagnostic and fuzzy pattern recognition applications it is very difficult to find out which features to use to achieve the optimum performance. This paper describes a PC-based feature selection system that solves this problem. The system uses a real-time fuzzy neural network. By using the numerical data about the membership functions and by testing thousands of feature subset combinations, the system searches for a subset that increases the separation between classes. If such a subset exists, its use makes it easier to identify the classes. The use of fewer features also results in smaller array sizes and a faster operation. The results of applying this technique to two different systems are discussed