Hill Climbing Methods for the Optimization of Multiparameter Noise Disturbed Systems

Abstract

Modifications of the statistical technique of Stochastic Approximation are used as a means of experimentally optimizing multiparameter systems on which the only available information is obtained from noise perturbed samples of the performance. Varieties of both Gradient and Relaxation Processes are considered and methods are given for experimentally adjusting certain undetermined process constants for most rapid (mean square) convergence to the optimum parameter settings.