2002 年 15 巻 2 号 p. 85-91
Since a fuzzy classifier with ellipsoidal regions is based on the Mahalanobis distance, the generalization ability is degraded when the associated covariance matrices become singular. In this paper, we discuss two methods for improving the generalization ability : 1) during the symmetric Cholesky factorization of the covariance matrix we replace the input of the root with a prescribed positive value when it is smaller than the prescribed value, and 2) we tune the slopes of the membership functions so that the margins are maximized. We demonstrate the validity of our methods by computer simulations.