RESEARCH FOR MULTI-INPUT WAVELET NEURAL NETWORK
This work is supported by the State 863 Program (2003AA148040), the National Natural Science Foundation of China (grant number 10471151, 60216263, 6990312), Chongqing Science Technology Cooperative Program (2004-8770).
In this paper, a new multi-input wavelet network with one hidden layer and single output is proposed. The structure of this network is similar to that of the radial basis function (RBF) network, except that the radial basis functions are replaced by 1-D orthonormal wavelets. In the learning algorithm, training data and forecasting data are important factor in wavelets selecting. The weights are determined by linear equations. The learning error of network is zero when the forecasting data is a compact subset. This kind of network is able to approximate an arbitrary nonlinear function in multi-dimensional space and has robustness is demonstrated through theoretical analysis.
