Function Minimization for Dynamic Programming Using Connectionist Networks

Learning controllers based on dynamic programming require some means of storing arbitrary functions and finding global minima within cross sections of those functions. There are many general methods for learning and representing functions, including polynomials, multi-layer perceptrons with backpropagation, and radial basis functions, but these systems do not allow the minima to be found easily. A method is presented here for learning and finding the minima of all cross sections of an arbitrary, smooth function. This method is applicable to any general function approximation system that learns smooth functions from examples. Mathematical properties of this approach are described, applications to learning control are discussed, and simulation results are presented.