Parallel global optimization : structuring populations in differential evolution
Matthieu Weber · Jyväskylä University Digital Archive (University of Jyväskylä) · 2010
Differential Evolution is a versatile and powerful optimization algorithm that can be applied to a wide range of continuous problems. Under some conditions however, the algorithm may suffer from stagnation, where it stops improving upon its current solutions, especially when applied to medium and large scale problems. The replacement of the usual panmictic population by a structured population composed of interacting subpopulations has been found to improve the performance of the Differential Evolution. This work presents a few algorithmic components build upon such algorithms, aimed at preventing stagnation by acting on the subpopulations themselves or by modifying the search logic of the algorithm. One important finding in this work is that the introduction of small amounts of randomness in structured population algorithm, controlled by using very simple rules leads to significant improvements in performance.