BP网络隐层神经元数自调节算法的研究

The Research of BP Network Hidden Layer Neuron Self-Regulating Algorithm

针对传统BP算法中隐层神经元数不易选取的问题,本文提出了BP算法的改进模型———DBP(DoubleBP)网络模型.该模型可以动态自调节隐层神经元数,即通过一个给定先验知识的BP网络动态调节另一个BP网络中隐层神经元数,并且通过选取合适的权值和阈值使训练误差曲线迅速下降.解决了BP网络拓扑结构中隐层神经元个数以及新增加权值和阈值的选取问题,还对BP网络在陷入假饱和区如何逃逸提出了一种新的方法.最后通过仿真模拟取得了比较好的效果.

In traditional BP algorithm, the number of hidden layer neurons is uneasy to select. Aiming at sorting out this problem, this article points out the updated BP arithmetic model, which is DBP （Double BP） network model. The new model can adjust hidden layer neurons dynamically and automatically, which means, adjust the hidden layer neurons in another BP network dynamically via a experience and knowledge BP network, make the training error curve decrease rapidly by choosing the proper weight and thresholds. It sorts out the problem in the quantity of hidden layer neurons in BP network topology, as well as the issue of choosing the new added weight and threshold. In addition, it suggests a new approach against how the BP network escapes from the fake saturated area. Finally, it gets a good result through simulation.