A modified algorithm based on wavelet networks

Combining wavelet networks with a kind method of model structure design and parameter estimation of NARMAX model, a modified wavelet network training scheme is proposed. The scheme based on a great extent the numerical stability of MGS algorithm determines the node number of the networks under a [ERR] rule. The [err] selecting rule is used to judge the importance of each term compared to all the other terms, deleting spare terms and obtaining the true optimal structure for the wavelet neural networks. In additional, this scheme overcomes the storage problem resulted from the number explosion of the terms of the assumed wavelet networks whole model while applying the traditional MGS algorithm and avoids the backward substitution calculating process. The example shows that the wavelet networks are of small scale and small storage.