用于过程建模的BP网络的M训练法及其改进

Marquardt Algorithm and Development for Training BP Neural Network in Chemistry Processing Engineering

以Sigmoid为传递函数的BP网络在过程系统工程领域已经得到了广泛的应用 ,但是一般的GDR训练算法在极小点附近易发生振荡 ,收敛速度慢。本文提出了人工神经网络M法训练的新途径 ,并且通过不同算例和工业实际数据建模应用证实了M算法的收敛速度大约是GDR算法的 5 - 10倍左右 。

Backpropagation artificial neural network has been widely used in the field of process systems engineering. For a long time the BP network is trained by the GDR algorithm. But the network is limited in industrial application because of the much low convergent speed of the algorithm. In this paper, The Marquardt algorithm is presented and is incorporated into the backpropagation algorithm for training feedforward BP neural network. The new algorithm is tested on several function approximation problems and an industrial data fitting problem, and is compared with the GDR algorithm. It is found that the Marquardt algorithm takes a faster convergent speed and is much more efficient than the GDR algorithm.