点击次数：

论文类型：期刊论文

发表时间：2013-05-01

发表刊物：NEURAL COMPUTING & APPLICATIONS

收录刊物：SCIE、EI、Scopus

卷号：22

期号：SUPPL.1

页面范围：S379-S394

ISSN号：0941-0643

关键字：Neural networks; LCFNNs; Convergence; Constant learning rate; Gaussian basis function

摘要：Local coupled feedforward neural networks (LCFNNs) help address the problems of slow convergence and large computation consumption caused by multi-layer perceptrons structurally. This paper presents a modified gradient-based learning algorithm in an attempt to further enhance the capabilities of LCFNNs. Using this approach, an LCFNN can achieve quality generalisation with higher learning efficiency. Theoretical analysis of the convergence property of this algorithm is provided, indicating that the gradient of the error function monotonically decreases and tends to zeros and the weight parameter sequence converges to a minimum of the given error function with respect to the number of learning iterations. Conditions for the use of a constant learning rate in order to guarantee the convergence are also specified. The work is verified with numerical experimental results.