A stochastic coordinate descent for bound constrained global optimization

Abstract

This paper presents a stochastic coordinate descent algorithm for solving bound constrained global optimization problems. The algorithm borrows ideas from some stochastic optimization methods available for the minimization of expected and empirical risks that arise in large-scale machine learning. Initially, the algorithm generates a population of points although only a small subpopulation of points is randomly selected and moved at each iteration towards the global optimal solution. Each point of the subpopulation is moved along one component only of the negative gradient direction. Preliminary experiments show that the algorithm is effective in reaching the required solution.