前馈神经网络中隐层神经元数目的一种直接估计方法

目前还没有一个行之有效的方法直接估计前馈网络隐层神经元的数目。该文首先提出一种利用单调指数直接估算三层前馈网络隐层经元数目的方法，以保证网络近似逼近任意给定的训练数据。理论分析和计算实验表明，此方法能够在训练之前预先确定最优（最少）或接近最优的隐层神经元数目，使得网络在训练之后不仅可以较好地反映训练数据的变化趋势，而且有较为满意的逼近精度。

A METHOD TO DIRECTLY ESTIMATE THE NUMBER OF THE HIDDEN NEURONS IN THE FEEDFORWARD NEURAL NETWORKS

At the present, there are no effective methods to directly estimate the number of hidden neurons in feedforward neural networks. In this paper, a monotone index based method is presented to directly estimate the number of hidden neurons in a three layer feedforward network so that the network can approximate any given training data. Theoretical analyses and computer simulations show that by this method the optimal (least) or near optimal number of hidden neurons can be predetermined which guarantee that the network can approximate the training data to a satisfactory degree after training.