摘要
构建一类新型基函数神经网络。依据梯度下降法思想,给出该神经网络的权值迭代公式,证明迭代序列能全局收敛到网络的最优权值,并由此推导出基于伪逆的最优权值一步计算公式——简称为权值直接确定法。理论分析表明,该新型神经网络具有最佳均方逼近能力和全局收敛性质,其权值直接确定法避免了冗长迭代计算、易陷于局部极小点、学习率难选取等传统BP神经网络难以解决的难题,仿真验证显示其相对BP神经网络的各种改进算法具有运算速度快、计算精度高等优势,且对于噪声有良好的滤除特性。
Constructs a new type of basis-function feedforward neural network. The weights-updating formula for the constructed neural network is derived based on the gradient-descent method, with global-convergence and least-squares approximation explored then. Moreover, presents a weights--direct--determination method by using pseudoinverse. Theoretical analysis demonstrates that the constructed neural network could remedy the weakness of conventional BP neural networks, such as, the local-minima phenomenon and difficulty of choosing learning rate, as the weights-direct-determination method could obtain the optimal weights just in one step without any lengthy BP iterative-training. Computer-simulation results substantiate the advantages of our neural network and its weights-direct-determination method, in the sense of speedy computation and high precision.
出处
《现代计算机》
2009年第2期4-8,共5页
Modern Computer
基金
国家自然科学基金(No60643004
60775050)
中山大学科研启动费
后备重点课题资助
关键词
基函数神经网络
权值直接确定
全局收敛性
逼近性能
Basis-Function Neural Network
Weights Direct Determination
Global Convergence
Approximation Performance
