A Population-Based Stochastic Coordinate Descent Method

Abstract

This paper addresses the problem of solving a bound constrained global optimization problem by a population-based stochastic coordinate descent method. To improve efficiency, a small subpopulation of points is randomly selected from the original population, at each iteration. The coordinate descent directions are based on the gradient computed at a special point of the subpopulation. This point could be the best point, the center point or the point with highest score. Preliminary numerical experiments are carried out to compare the performance of the tested variants. Based on the results obtained with the selected problems, we may conclude that the variants based on the point with highest score are more robust and the variants based on the best point less robust, although they win on efficiency but only for the simpler and easy to solve problems.