An Adaptive Perturbation Scheme in Finite Difference Gradient Approximation

Abstract

The idea presented in this paper is to make perturbation constant adaptive in finite difference gradient approximation. In the proposed method we keep perturbation constant relatively large in the beginning and reduce it as the number of iterations proceed. It is also shown that large perturbation may cause false gradient information included in the approximated gradient. Problem of false gradient information is solved by placing a check not to include it in the resultant gradient. Numerical experiments show practical usefulness and improved convergence of the new approach.