中文摘要：

隐层数和隐层结点数直接关乎BP网络的学习能力,但目前对隐层结点数的选择尚无适用的理论,一般凭经验或试凑确定。本文提出一种分段式自适应确定隐层结点数的算法,它通过评估网络输出相对误差相应调整隐层结点数,通过迭代运算在使网络输出相对误差逐步减小的同时,逼近可能的最优隐层结点数。通常这个最优结点数即网络输出相对误差出现震荡的起点对应的结点数,以这个结点数决定的网络结构能够在网络输出精度与运算开销之间取得较佳平衡。

英文摘要：

The learning ability of a BP network has a close relationship with the number of hidden layers and hidden nodes.However,there is no applicable theory for the decision of the number of hidden nodes,and it is usually determined by the experience.This paper presents a segmented self-adaptive algorithm to determine the number of hidden nodes.A corresponding adjustment to the number of hidden layer nodes will run relying on the evaluation of network output relative errors during an iterative computation.As the network output relative errors are gradually reduced,the system would approximate the best possible number of hidden nodes.Usually this node corresponds to the starting point of vibration of the network output relative errors.Hence the network will get a good balance between the output precision and operation cost.