A new approach is proposed to improve the general identification algorithm of multidimensional systems using wavelet networks. The general algorithm involves mapping vector input into its norm to avoid problem of dimensionality in construction multidimensional wavelet basis functions. Thus, the basis functions are spherically symmetric without direction selectivity. In order to restore the direction selectivity, the improved approach weights the input variables before mapping it into a scalar form. The weights can be obtained using universal optimization algorithms. Generally, only local optimal weights are obtained. Even so, performance of identification can be improved.
展开▼
