Locality in decoder-based EAs for the multidimensional knapsack problem

Abstract

The performance of decoder-based evolutionary algorithms (EAs) strongly depends on the locality of the used decoder and opera- tors. While many approaches to characterize locality are based on the fitness landscape, we emphasize the explicit relation between genoty- pes and phenotypes. Statistical measures are demonstrated to reliably predict locality properties of selected decoder-based EAs for the multi- dimensional knapsack problem. Empirical results indicate that (i) strong locality is a necessary condition for high performance, (ii) the concept of heuristic bias also strongly affects solution quality, and (iii) it is im- portant to maintain population diversity, e.g. by phenotypic duplicate elimination.