多层前向网络的动态结构设计方法及其在回弹预测中的应用

Dynamic Structure Design Method of Multilayer Feedforward Network and Its Application in Springback Prediction

从构造的角度,开展神经网络的动态结构设计研究,提出一种基于泛化的多层前向网络动态结构设计方法,编制了相应的计算程序。在该方法中,基于Ockhams RAZOR原则,从一个较小的基本网络开始,通过动态增加隐结点或隐层,综合运用网络泛化能力的多种改进方法,改进的BP算法以及快速搜索机制和全局搜索机制相结合确定学习速率、动量系数、跳跃因子和正则化系数的方法,采用网络权值的局部和全局调节方案,对多层前向网络进行动态结构设计。上述方法在凸凹弧翻边回弹预测中的应用实例表明,运用该方法设计的网络具有较好的计算精度。

From the view point of construction, research on dynamic structure design of neural network is carried out and a dynamic structure design method of multilayer feedforward network （DYNSDMFN） based on generalization performance is proposed and the corresponding calculation program is worked out. Based on Ockhams RAZOR principle, DYNSDMFN starts from a basic architecture and designed structure of MFN dynamically, corresponding to given training sample set and test sample set, by means of adding new neuron of hidden layer or new hidden layer dynamically, comphrehensive application of several improving methods of generalization performance, an improved BP algorithm, the determination method of learning ratio, momentum coefficient, jumping factor and regularization coefficient adopting the combined searching mechanism in quick and global manner, the combination of local and global weight adjustment project, dynamic structure design of multilayer feedforward network is carried out The application results in the flanging springback prediction indicates that the network designed by using that method has a relative high calculation accuracy.