The Local Linear Adaptive Wavelet Neural Network with Hybrid EP/Gradient Algorithm and Its Application to Nonlinear Dynamic System Identification
Ting Wang, Yasuo Sugai · IEEJ Transactions on Electronics Information and Systems · 2002
Wavelet neural networks are networks employing nonlinear wavelet basis functions as the activation functions of the neurons. This paper presents a new type of wavelet-based neural network: the local linear adaptive wavelet neural network. A hybrid evolutionary programming and gradient descent algorithm is introduced to the learning of the proposed network. The local linear models which are used in some neurofuzzy systems are introduced as powerful weights instead of straightforward weights employed in the previous wavelet neural networks. Training is performed by using the evolutionary programming algorithm at first to search a good region in the parameter space and then employing the gradient descent algorithm to find a near optimal solution in that region. The experiments on a number of nonlinear dynamic system identification problems indicates that the proposed network with the hybrid EP/Gradient algorithm can successfully identify and describe the input/output relationship for an unknown complex system with a small number of wavelet basis functions and compared favorably to the traditional neural networks with the sigmoid activation functions and the previous wavelet neural networks with straightforward weights.