【摘要】 线性判别函数理论是线性分类器的分析基础,并不适合非线性分类器。本文把非线性激励函数视为隶属度函数,将非线性神经元及多层感知器分类行为的分析建筑在模糊集理论基础上,提出模糊线性判别函数与模糊判别边界、模糊分类等概念。并引出将隐层初始权向量均匀分布在权空间超球面上的初始化方法,明显提高了多层感知器的收敛性能,并提出了一种在多层感知器的类空间中构造最优超平面的简易新方法。更多还原

【Abstract】 The classical theory of linear discriminant functions provides an analysis for linear classifiers, but it is not applicable for nonlinear classifiers. The analysis of internal behavior of MLP--classifiers is based on the fuzzy sets in this paper and the nonlinear activation functions of neurons are considered as the membership functions of the output class--set. Based on it, consequently, the fuzzy linear discriminant functions, fuzzy decision (or indicating) surface, and fuzzy pattern classification are given as the extensions of those in crisp sets. A new approach of initializing-- weights is introduced, by which the hidden weight vectors are initialized and distributed uniformly on a hypersphere in weight space with a moderate radius. Many experiments have justified its usefulness for a significant improvement in the performance of MLP--classifiers. And a facile method of constructing optimal hyperplane in the class space of MLP--classifiers is proposed.更多还原