Integrated approximation and non-convex optimization using radial basis function networks

The authors consider the problem of learning inverse maps x ´=f¹(y) within the framework of radial basis function networks. If the forward function y=f( x) is approximated using a radial basis function network, it is found that the linear weights can be a good indicator of the network output. Centers then correspond to classes in the input space of the function, and the superposed weights correspond to properties associated with the respective classes. This provides suitable grounds for implementing efficient search strategies, for nonconvex and constrained or unconstrained optimization. The authors highlight the advantages of this scheme over other proposed methods for nonconvex optimization and present experimental results