基于最小分类错误训练的神经网络分类器设计

提出了一种基于最小分类错误（MCE）训练的采用多层感知器（MLP）结构的模式分类器设计方法。这是一种以分类错误率最小化为目标的模式分类器设计方法，将它用于MLP分类器设计能够进一步提高分类器的性能。采用MLP实现MCE训练中的分类损失计算，从而将MCE训练过程与MLP分类器设计统一在一个神经网络结构中，通过BP算法予以实现。这不仅能达到提高MLP分类器性能的目的，而且简化了它的设计过程。

Design of Neural Networks Classifier Based on Minimum Classification Error Training

An improved design method on pattern classifier based on multi-layer perceptrons (MLP) by means of minimum classification error (MCE) training was proposed. MCE training is a design method of pattern classifier, which takes minimum classification error rate as an objective. It will improve the classification performance of MLP classifier adopting MCE training method. This paper proposed that the classification losses are calculated by multi-layer perceptrons in MCE training. So, MCE training can be incorporated with MLP classifier design in an integrated neural network structure, and realized by means of an effective error back-propagation (BP) algorithm. This method can not only improve the classification performance of MLP classifier, but also simplify its design process.