Subpopulation initialization driven by linkage learning for dealing with the Long-Way-To-Stuck effect

Abstract The maintenance of many subpopulations is an important technique employed in evolutionary methods. However, the use of a multi-population approach has its drawbacks. Among all, it requires spending high amounts of available resources. Therefore, the methods that dynamically manage the number of subpopulations gain an increasing interest due to their capability of adjusting the subpopulation number to their current state. They are shown to be capable of reaching excellent results. In this paper, we identify the Long-Way-To-Stuck effect and show it on the base of the practical, NP-complete problem. Such a phenomenon occurs when a randomly initialized population of an evolutionary method must be processed through many iterations before it is incapable of improving the best-found solution. If so, then a large amount of computational resources must be spent on subpopulation initialization, which may turn multi-population methods ineffective. Therefore, in this paper, we propose the Linkage Learning-Driven Subpopulation Initialization (LLDSI) that limits the costs of subpopulation initialization and significantly improves the effectiveness methods dynamically managing the subpopulation number.