机构地区: 中山大学信息科学与技术学院电子与通信工程系
出 处: 《计算机仿真》 2009年第1期157-161,共5页
摘 要: 经典的BP神经网络学习算法是基于误差回传的思想。而对于特定的网络模型,采用伪逆思想可以直接确定权值进而避免以往的反复迭代修正的过程。根据多项式插值和逼近理论构造一个切比雪夫正交基神经网络,其模型采用三层结构并以一组切比雪夫正交多项式函数作为隐层神经元的激励函数。依据误差回传(BP)思想可以推导出该网络模型的权值修正迭代公式,利用该公式迭代训练可得到网络的最优权值。区别于这种经典的做法,针对切比雪夫正交基神经网络模型,提出了一种基于伪逆的权值直接确定法,从而避免了传统方法通过反复迭代才能得到网络权值的冗长训练过程。仿真结果表明该方法具有更快的计算速度和至少相同的工作精度,从而验证了其优越性。 Standard BP neural network is based on the error back - propagation method. For a special neural network model, the weights could be determined directly without lengthy iterative updating by using a pseudo - inverse method. Based on polynomial interpolation and approximation theory, a Chebyshev orthogonal basis neural network is constructed in this paper. The model adopts a three - layer structure, where the hidden - layer neurons are activated by a group of Chebysbev orthogonal polynomial functions. The weight updating formula is derived by following the standard BP training method. More importantly, the pseudo - inverse based method is then proposed, which could immediately determine the neural - network weights. Computer simulation results show that the one - step weight - determination method could be more efficient than the conventional BP iterative -training method, in addition to the equally -high working- precision at least, which reveals its advantages.
关 键 词: 切比雪夫正交多项式 人工神经网络 激励函数 权值修正公式 权值一步确定 伪逆
领 域: [自动化与计算机技术] [自动化与计算机技术]