【摘要】 本文提出一种概率映射网络（PMN）的CEM（GeneralExpectationMaximization）训练算法，它是EM（ExpectionMaximization）算法的一种改进算法。PMN网为一个四层前馈网。它构成一个贝叶斯分类器，实现多类分类的贝叶斯判别，把输入的样本模式经网络变换为输出的分类判决，其网络节点对应于贝叶斯后验概率公式的各个变量。此PMN网络用高斯校函数作为密度函数，网络参数的训练由GEM算法实现，其学习方式为类间的监督学习和类内的非监督学习。最后的实验表明了此网络及其学习算法在分类应用中的有效性。更多还原

【Abstract】 A General Expectation Maximization(GEM) training algorithm for estimating the parameters of a Probability Mapping Network(PMN) is proposed in this paper. This is an improvement of EM (Expectation Maximization)algorithm. A PMN is a four-layer, Feed-forward Neural Network (FNN) with its nodes adopting Gaussian probability density function. As a multi-category Bayes classifier, A PMN outputs classification result after inputting sample patterns. In this way, EM algorithm is generalized to deal with supervised learning from inter-categories or unsupervised learning from intracategory. The effectiveness of the proposed network and it’s GEM algorithm are verified by two experiments.更多还原