A Comparative Study of Dynamic Learning Rate BPN and Wavelet Neural Networks
Yifeng Zhao, Jianxin ZHANG, Anton J.R. Aendenroomer · 2007
This paper presents an improved back propagation network (iBPN) with dynamic learning rates to accelerate the network learning and convergence speed. As compared with conventional BPN, the improved BPN is able to approximate the complex non-linear functions with higher efficiency and accuracy. Wavelet neural network (WNN) is a comparatively novel universal tool for functional approximation, and is effective in solving the inherent problems of poor convergence or even divergence encountered in other kinds of neural networks. This paper, through a comparative study, shows that iBPN has the same generalization performance as wavelet neural networks. While WNN shows the highest efficiency, it lacks consistency. In contrast, results obtained from iBPN are highly consistent and are quite comparable with those obtained from WNN.