BP神经网络学习算法的联合优化

The Joint Optimization of BP Learning Algorithm

针对ＢＰ网络学习速度的缓慢性，本文提出了一种联合优化后的快速学习算法。其改进具体表现在以下方面：１）采用Ｃａｕｃｈｙ误差估计器代替传统的ＬＭＳ误差估计器；２）对常规的Ｓｉｇｍｏｉｄ函数引入形态因子：３）采用非单调线性搜索法实现学习步长的自适应变化。最后，本文以模式分类、函数逼近和数据压缩的典型应用为例分别与标准ＢＰ和常规改进算法进行比较，验证了该算法的优越性。

To accelerate the training speed of BP network, a joint-optimized fast BP learning algorithm is proposed. improvements can be achieved as following: 1) Conventional LMS error estimator is replaced by Cauchy error estimator: 2) Shape factor is adopted in Sigmoid function, 3) The adaptive changes of learning rate is realized by using non-sole linear search. Effectiveness of this algorithm is verified by comparison of performance of this algorithm with conventional modified and standard BP algorithms in typical BP network applications as pattern-classification, approximation of mathematical function and data compression.