摘要
采用了一种基于贝叶斯方法的前向神经网络训练算法以提高网络的泛化能力,并在网络的目标函数中引入了表示网络结构复杂性的惩罚项,避免了网络的过拟合。采用Levenberg-Marquardt算法训练网络,并使用Gauss-Newton的数值方法来近似求解Hessian矩阵,以减少计算量,从而提高了网络的收敛速度。将上述网络应用于冷轧过程的轧制力预报中,预报结果的精度远远高于解析模型,与基于传统BP神经网络的冷轧轧制力预报模型相比,在收敛的速度和预报的精度上均优于后者。
Bayesian regularization was applied to the training of feedforward neural networks in order to improve their generalization capabilities. A penalty item which represents the network complexity was introduced into the performance function to avoid ＂overfitting＂. A Gauss-Newton numerical method was used to solve the Hessian matrix approximately, which was implemented within the framework of the Levenberg-Marquardt algorithm to reduce the complexity of the calculation, and a fast convergence rate of the network was achieved. The network was applied to rolling force prediction in cold rolling process, exhibiting higher precision than the physical models. Compared with the prediction model based on traditional back-propagation neural network, the Bayesian network has a faster convergence rate and better precision.
出处
《钢铁研究学报》
CAS
CSCD
北大核心
2008年第10期59-62,共4页
Journal of Iron and Steel Research
基金
国家自然科学基金资助项目(50504007)
关键词
贝叶斯方法
神经网络
冷轧
Bayesian method
neural network
cold rolling
