一种改进的量子神经网络训练算法  

An Improved Training Algorithm for Quantum Neural Networks

作　　者：张翼鹏[1] 陈亮[1] 郝欢[1] 

机构地区：[1]解放军理工大学通信工程学院,南京210007

出　　处：《电子与信息学报》2013年第7期1630-1635,共6页Journal of Electronics & Information Technology

基　　金：国家自然科学基金(61072042)资助课题

摘　　要：针对训练多层激励函数量子神经网络(MAF-QNN)时权值与量子间隔的目标函数存在冲突,导致收敛速度和网络性能下降的问题,该文提出一种改进的量子神经网络的训练算法。通过设计输出均方误差和这一目标函数对权值和量子间隔进行统一训练,同时引入Levenberg-Marquardt(LM)算法降低目标函数陷入局部极小值的概率,实现了对量子神经网络的高效训练。实验结果表明,该文提出的训练算法有效减少了迭代次数,显著提高了网络收敛精度,可应用于数据分类、函数逼近等场合,扩展了多层激励函数量子神经网络的应用领域。An improved training algorithm is proposed to solve the conflict of objective functions between weights and quantum interval during the training process of Multilevel Activation Functions-Quantum Neural Network（MAF-QNN）,which will result in the degradation of training speed and network performance.By the criterion of least mean square error,the objective functions of the weights and quantum interval are unified and trained simultaneously.And then,by introducing the Levenberg-Marquardt（LM） algorithm,the probability of which the training results fall into local minimum is reduced.Therefore,the MAF-QNN can be trained effectively and efficiently.Simulation results show that,the proposed algorithm can decrease the iteration times efficiently and improve convergence precision significantly.In this way,it can be applied to data classification,function approximation and so on,expanding the application fields of MAF-QNN.

关 键 词：量子神经网络 多层激励函数 Levenberg-Marquardt(LM)算法 最速下降 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]