A Constrained Approximation Algorithm by Encoding Second-Order Derivative Information into Feedforward Neural Networks

摘要

In this paper, a constrained learning algorithm is proposed for function approximation. The algorithm incorporates constraints into single hidden layered feedforward neural networks from the a priori information of the approximated function. The activation functions of the hidden neurons are specific polynomial functions based on Taylor series expansions, and the connection weight constraints are obtained from the second-order derivative information of the approximated function. The new algorithm has been shown by experimental results to have better generalization performance than other traditional learning ones.