Multiresolution radial basis function networks for the adaptive control of robotic systems

In this paper we review recent developments and their applications to robotic systems, and analyze several different variations of the basis function selection mechanism. In particular, we analyze algorithms in which the network is explicitly limited to employ no more than a specified fraction of the potential basis function set, and algorithms in which the actual dimensionality of each basis function employed is gradually expanded similar to the MARS algorithm. The results of these new algorithms is compared with previous results contrasting both the ultimate tracking accuracy and evolved network size for a canonical robotic application.