一种解决BP网络过拟合的学习算法

针对BP网络的过拟合问题,提出了一种基于验证误差的学习算法,该算法主要从构建最佳网络结构及获得最适训练次数两个角度进行了考虑.在学习过程中,通过不断检验验证误差及误差变化量,自动调整隐层节点数目,并适时停止网络训练.针对给定问题,该算法不仅能自动构建最佳网络结构和相应权系数,而且同时对网络的泛化能力进行了检验,从而有效避免了网络的过拟合.最后通过实验证明了算法的有效性.

A Learning Algorithm for Solving the Over-Fitting Problem in BPANN

To counter the over-fitting problem in BPANN, this paper puts forward a validation error-based learning algorithm, which takes into consideration building of the optimum net structure and obtaining the most appropriate training epochs. In the process of learning, the algorithm can automatically adjust the number of hidden layers and break net training in due course through constant test of validation error and error variation. Given the samples, this learning algorithm can not only build the optimum net structure and corresponding coefficient, it can also test the generalization ability of the net, thus effectively avoiding the over-fitting problem. In the end, an experiment is made to prove the validity of this algorithm.