Direction-Dependent Learning Approach for Radial Basis Function Networks

Abstract

Direction-dependent scaling, shaping, and rotation of Gaussian basis functions are introduced for maximal trend sensing with minimal parameter representations for input output approximation. It is shown that shaping and rotation of the radial basis functions helps in reducing the total number of function units required to approximate any given input-output data, while improving accuracy. Several alternate formulations that enforce minimal parameterization of the most general radial basis functions are presented. A novel "directed graph" based algorithm is introduced to facilitate intelligent direction based learning and adaptation of the parameters appearing in the radial basis function network. Further, a parameter estimation algorithm is incorporated to establish starting estimates for the model parameters using multiple windows of the input-output data. The efficacy of direction-dependent shaping and rotation in function approximation is evaluated by modifying the minimal resource allocating network and considering different test examples. The examples are drawn from recent literature to benchmark the new algorithm versus existing methods.