Neural-Network-Based Multiclass Approximate Classification

Abstract

This paper proposes a neural-network-based approximate classification method to cope with classification problems with no clear cut-off boundaries between classes. The proposed method assumes that a boundary area (i.e., fuzzy boundary) divides the pattern space into decision areas of the given classes. First, the concept of approximate classification based on the possibility theory is introduced to multiclass classification problems. Next, a learning algorithm for the possibility analysis is proposed by modifying the back-propagation algorithm of multilayer feedforward neural networks. The decision area of each class and the boundary area between classes are derived from the output values of the neural network trained by the proposed learning algorithm. Last, the learning ability to training data and the generalization ability to test data are examined for the proposed approximate classification method by computer simulations on the iris data of Fisher.