基于Tanh多层函数的量子神经网络算法及其应用的研究  

Research on Quantum Neural Network and its Applications Based on Tanh Activation Function

作　　者：罗鹏[1] 李会方[1] 

机构地区：[1]西北工业大学电子信息学院,西安710129

出　　处：《计算机与数字工程》2012年第1期4-6,39,共4页Computer & Digital Engineering

基　　金：航空科学基金项目(2007ZC53032)资助

摘　　要：该文提出一种新的改进激励函数的量子神经网络模型。首先为了提高学习速率,在网络权值训练过程中引入了动量项。然后为了有效实现相邻类之间具有覆盖和不确定边界的分类问题,新网络采用区分度更大的双曲正切函数的叠加作为其隐层激励函数。最后将该算法用于字符识别,将双曲正切激励函数的量子神经网络应用于数字、字母和汉字样本的多次实验,并且与原多层激励函数量子神经网络和BP网络的实验效果进行比较,发现改进后量子神经网络不仅具有较高的识别率,而且在样本训练次数上相对原多层激励函数量子神经网络有明显减少。仿真结果证明该方法的优越性。In this paper an improved quantum neural network（QNN） is presented based on multilevel activation function to solve the problems of precision inadequacy and low convergence rate of the BP neural network used in character recognition.Firstly,momentum term is used to updating the weights to accelerate the convergence rate of learning algorithm in QNN.Secondly,a linear superposition of hyperbolic tangent function is used as activation function of hidden unit in the new networks to classify pattern recognition problems that have uncertainty and overlapping data between two patterns.Finally,The experimental results for number,letter,and Chinese character recognition is provided and compared with both BP neural networks and the original Qnn.Results indicate that the improved Qnn not only takes great recognition rate,but also decreases the number of the training time comparing with the original one.The superiority is demonstrated by simulation.

关 键 词：量子神经网络 多层激励函数 双曲正切函数 字符识别 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]