一种前馈神经网络的变误差主动式学习算法

Active Back propagation Algorithm Based on Adjusting Error for Multilayer Feed forward Neural Network

研究误差反向传播多层前馈神经网络的主动式学习方法．文章分析了目前用于训练前馈神经网络改进ＢＰ算法的特点和存在的不足，在此基础上提出逐次主动调整网络学习误差的网络训练思想，根据网络输出误差趋势，主动变化输出层的调整误差δｐｌ，使Ｗｋｊｉ和θｋｊ在调整过程中受到每次学习效果信息的控制，从而得到一种主动式变误差的学习算法．实验表明，在训练多层前馈神经网络时，变误差主动式算法的学习效率比改进ＢＰ算法的学习效率有明显提高．

The back propagation (BP) algorithm was used as a learning algorithm in training multilayer feed forward neural networks (MLFNN) in past years, and some improved BP algorithms have recently been developed to speed up MLFNN learning. However, the effeciency of these improved BP algorithms are limited due to ignoring the activity of adjusting error during training MLFNN. In this paper, an active back propagation (ABP) algorithm based on improved BP algorithm is developed for MLFNN trained. The ABP algorithm alters the adjusting errors of MLFNN during the network trained, according to the error tendency of the network, and aimed to enhance rapidity of the network trained. The paper describes experiments that compare the performance of ABP algorithm with improved BP algorithms. The experiment results have shown that the ABP algorithm gives more efficient than improved BP algorithm for MLFNN trained.