摘要
根据Pad近似理论,构造出一类前向有理式神经网络.该网络采用四层结构,其中第一层(输入层)和第三层采用线性激励函数,第二层采用幂激励函数,第四层(输出层)采用分数函数(或称除法函数)作为激励函数.依据梯度下降法思想,推导了其权值修正的迭代公式.针对迭代方法收敛速度慢、易陷入局部极小点等缺点,进一步推导出了基于伪逆的权值直接确定方法,该方法避免了冗长的迭代过程.仿真和预测结果均表明Pad有理式神经网络及其权值直接确定法具有较好的计算速度和更高的逼近与预测精度.
Based on Padé approximation theory, a kind of Padé rational neural network is constructed. The neural network model adopts a four-layer structure, its first and third layers employing linear activation functions, while the second and fourth layers are activated by a group of power functions and rational function respectively. Based on BP and gradientdescent method, we firstly derive the weights-updating formula of the Padé neural network. More importantly, for resolving the inherent weakness of BP algorithms such as slow convergence and local minima, a pseudoinverse-based method is further proposed which could determine the neural weights directly without lengthy iterative BP training. Simulation results show that the weights-direct-determination method could be much more efficient and accurate than conventional BP iterative-training algorithms.
出处
《微电子学与计算机》
CSCD
北大核心
2009年第1期12-15,20,共5页
Microelectronics & Computer
基金
国家自然科学基金项目(60643004,60775050)
中山大学科研启动费、后备重点课题
关键词
Padé近似
有理式神经网络
权值修正
权值直接确定法
Padé approximation
rational neural network
weights updating
weights direct determination
