量子门Elman神经网络及其梯度扩展的量子反向传播学习算法  

Quantum Gate Elman Neural Network and Its Quantized Extended Gradient Back-propagation Training Algorithm

作　　者：李鹏华[1] 柴毅[1] 熊庆宇[2] 

机构地区：[1]重庆大学自动化学院输配电装备及系统安全与新技术国家重点实验室,重庆400044 [2]重庆大学软件学院,重庆400044

出　　处：《自动化学报》2013年第9期1511-1522,共12页Acta Automatica Sinica

基　　金：国家自然科学基金(60974090);高等学校博士学科点专项科研基金(20100191110037)资助~~

摘　　要：针对Elman神经网络的学习速度和泛化性能,提出一种具有量子门结构的新型Elman神经网络模型及其梯度扩展反向传播(Back-propagation)学习算法,新模型由量子比特神经元和经典神经元构成.新网络结构采用量子映射层以确保来自上下文单元的局部反馈与隐藏层输入之间的模式一致;通过量子比特神经元输出与相关量子门参数的修正互补关系以提高网络更新动力.新学习算法采用搜索然后收敛的策略自适应地调整学习率参数以提高网络学习速度;通过将上下文单元的权值扩展到隐藏层的权值矩阵,使其在与隐藏层权值同步更新过程中获取时间序列的额外信息,从而提高网络上下文单元输出与隐藏层输入之间的匹配程度.以峰值检波为例的数值实验结果显示,在量子反向传播学习过程中,量子门Elman神经网络具有较快的学习速度和良好的泛化性能.A novel Elman neural model with hybrid quantum gate structure and a quantized extended-gradient back- propagation （BP） training algorithm are proposed for improving the performance of the conventional Elman network. The novel model is comprised of qubit neurons and classical neurons. The quantum map layer is employed to address the pattern mismatch between the context layer and the input layer. The complementary relationships between the outputs of qubit neurons and the quantum gate parameters are applied to improve the updated ability of the conventional Elman network. The learning rate is adaptively adjusted by the searching and convergent learning strategy, which makes the new neural model achieving convergence with high speed. The context-layer weights are extended into the hidden-layer weights matrix for obtaining the extra gradient information, such that the context-layer patterns match the input-layer patterns with high level. The numerical experiments are carried out to verify the theoretical results and clearly show that the hybrid quantized Elman network using quantized training offers a good performance in terms of both weight convergence and generalization ability.

关 键 词：量子比特神经元 ELMAN神经网络 梯度扩展 反向传播 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]