基于互补遗传算子的前馈神经网络三阶段学习方法

论文提出了一种新的基于互补遗传算子的前馈神经网络三阶段学习方法。该方法把神经网络的学习过程分为三个阶段。第一阶段为结构辨识阶段,采用遗传算法进行神经网络隐层节点数目的选择和初始参数的设定,并基于发现的遗传算子的互补效应设计高效互补遗传算子。第二阶段为参数辨识阶段,采用效率较高的神经网络算法如L-M算法进行神经网络参数的进一步学习。第三阶段为剪枝阶段,通过获得最小结构的神经网络以提高其泛化能力。在整个学习过程中,学习过程的可控性以及神经网络的逼近精度、复杂度和泛化能力之间得到了满意平衡。仿真试验结果证明了该方法的有效性。

A Mutual-Genetic-Operator-Based Three-Stage Method for Feedforward Neural Networks Training

A new mutual genetic operator based three stages feedforward neural network training method is proposed in this paper,which divides neural networks training procedure into three stages.In the first stage,which is called structure identification stage,the selection of network structure and initial parameters is carried out by instead of human,and mutual effect of combinatorial genetic operator is discovered to design more efficient genetic operators.In the second stage,which is called parameter identification stage,the traditional optimization method such as L-M algorithm is adopted to make refinements of parameters.In the third stage,which is called prune stage,minimal structure of neural network is achieved by prune so as to improve its generalization capability.Through the entire process,compromise is satisfactorily made among the control difficulty of training procedure,network complexity,approximation precision and generalization capability.Simulation results confirm these findings.