A New Method to Estimate the Generalization Error of Artificial Neural Network
一种估计人工神经网络泛化误差的新方法

The constructional learning is used to determine the architecture of neural network such that the network holds a satisfactory generalization. This paper considers the constructional learning in the case where the training set is randomly chosen from an input output space. A new objective function of constructional learning is presented. It is illustrated the reason why this objective function is superior to other functions. The learning algorithm for this objective function is also analyzed. Finally, a simulation example is given to show the efficiency of the method presented in this paper.