Abstract

A novel adaptive random search algorithm for the optimization of functions of continuous variables is presented. The scheme does not require any assumptions about the function to be optimized, apart from the availability of evaluations f(x) at selected test points. We assume that the main computational cost lies in the function evaluations and the main design criteria of the RASH scheme consists of the adaptation of a search region by an affine transformation which takes into account the local knowledge derived from trial points generated with a uniform probability. The aim is to scout for local minima in the attraction basin where the initial point falls, by adapting the step size and direction to maintain heuristically the largest possible movement per function evaluation. The design is complemented by the analysis of some strategic choices (like the double-shot strategy and the initialization) and by experimental results showing that, in spite of its simplicity, RASH is a promising building block to consider for the development of more complex optimization algorithms. The developed software is built to facilitate the scientific experimentation and the integration of RASH as a component in more complex schemes.

Citations