Basis functions with local support has two problems: 1) you get the "bed-of-nails" when the distance between them is much larger than their radius, and 2) they have to decay as you reach (and extend beyond) the boundary of the training data domain. The latter may cause the automatic radius optimization to choose too large radius. A good alternative may be to implement normalized RBF functions that sum to 1 everywhere in space (see e.g. Hastie). This could be turned on or off via an optional flag to RBFnet.train().
Basis functions with local support has two problems: 1) you get the "bed-of-nails" when the distance between them is much larger than their radius, and 2) they have to decay as you reach (and extend beyond) the boundary of the training data domain. The latter may cause the automatic radius optimization to choose too large radius. A good alternative may be to implement normalized RBF functions that sum to 1 everywhere in space (see e.g. Hastie). This could be turned on or off via an optional flag to
RBFnet.train().