HIERARCHICAL RBF NETWORKS AND LOCAL PARAMETERS ESTIMATE

The method presented here is aimed to a direct fast setting of the par ameters of a RBF network for function approximation. It is based on a hierarchical gridding of the input space; additional layers of Gaussia ns at lower scales are added where the residual error is higher. The n umber of the Gaussians of each layer and their variance are computed f rom considerations grounded in the linear filtering theory. The weight of each Gaussian is estimated through a maximum a posteriori estimate carried out locally on a sub-set of the data points. The method shows a high accuracy in the reconstruction, it can deal with non-evenly sp aced data points and can be fully parallelizable. Results on the recon struction of both synthetic and real data are presented and discussed. (C) 1998 Elsevier Science B.V. All rights reserved.