Characterizing Locality in Decoder-Based EAs for the Multidimensional Knapsack Problem

The performance of decoder-based evolutionary algorithms (EAs) strongly depends on the locality of the used decoder and operators. While many approaches to characterize locality are based on the fitness landscape, we emphasize the explicit relation between genotypes and phenotypes. Statistical measures are demonstrated to reliably predict locality properties of selected decoder-based EAs for the multidimensional 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 important to maintain population diversity, e.g. by phenotypic duplicate elimination.

• Type: Conference
• Year: 1999
• Where: AE
• Authors: Jens Gottlieb, Günther R. Raidl
• Keywords: AE 1999, Artificial Intelligence, Decoder-based Evolutionary Algorithms, Multidimensional Knapsack Problem, Strong Locality