摘要： 用一组训练样本对神经网络进行训练后，网络对训练阶段未曾见过的样本也能正确分类。但传统的神经网络模式分类方法泛化能力不十分理想，而且不稳定。对同一个分类任务，训练样本改变，分类器泛化能力的大小也会改变。该文提出一种基于最优分类面的神经网络模式分类方法。通过寻找并训练最优分类面，提高网络的泛化能力，增强泛化能力的稳定性。用异或问题和双螺旋线问题验证该新方法的有效性和泛化能力，取得了令人满意的结果。

关键词: BP神经网络, 模式分类, 最优分类面, 泛化能力

Abstract: Trained using a set of training examples, the neural network can also correctly classify the examples that never occur when training. But generalization capability of the traditional methods are not quite satisfying. The stability of generalization is also not quite good, trained with different training examples, the stability also changes. This paper presents a new approach of improving the generalization and the stability of neural networks. Proved by the XOR and double helices problems, the new approach is effective and satisfying.

Key words: BP neural network, Pattern classification, Optimal classification face, Generalization