BP神经网络的优化研究与应用

Optimization Research and Application of BP Neural Network

针对BP神经网络自身的一些局限性,诸如易陷于局部极小、网络收敛速度慢、训练时间长等,提出一种改进BP神经网络的研究方案,通过改变传统的固定学习率,引入动态变化,根据均方误差的变化而改变学习率。在误差曲面平坦区域增大学习率,在误差变化剧烈的区域减小学习率,从而加快算法的收敛速度,避免陷入局部极小值。文中在传统BP神经网络中使用动态学习速率,并融合参数可调激活函数来改进BP神经网络。采用公认完备、性能优异的KDD Cup99数据集,分别对改进算法和传统BP算法进行了对比实验。实验结果表明,与传统BP神经网络算法相比,改进算法极大地提高了训练速度,具有训练误差更小、预测精度更高的优点。

Aiming at some limitations of BP neural network,such as easy to be trapped in local minimum,slow convergence speed and long training time,we propose a scheme of improving BP neural network.By changing the traditional fixed learning rate and introducing dynamic change,the learning rate can be changed according to the change of mean square error.The learning rate is increased in the flat area of the error surface,and decreased in the area where the error changes sharply,so as to accelerate the convergence speed of the algorithm and avoid falling into the local minimum.In the traditional BP neural network,we use the dynamic learning rate with the parameter adjustable activation function to improve the BP neural network.The improved algorithm and the traditional BP algorithm are compared by using the KDD Cup99 data set,which is recognized to be complete and excellent in performance.Experiment shows that compared with traditional algorithms based on BP neural network,the proposed method has greatly improved training speed and has the advantages of smaller training error and higher prediction accuracy.