Unwanted feature interactions between the problem and search operators in evolutionary multi-objective optimization

Abstract

Providing self-reconfiguration at run-time amidst adverse environmental conditions is a key challenge in the design of dynamically adaptive systems (DASs). Prescriptive approaches to manually preload these systems with a limited set of strategies/solutions before deployment often result in brittle, rigid designs that are unable to scale and cope with environmental uncertainty. Alternatively, a more scalable and adaptable approach is to embed a search process within the DAS capable of exploring and generating optimal reconfigurations at run time. The presence of multiple competing objectives, such as cost and performance, means there is no single optimal solution but rather a set of valid solutions with a range of trade-offs that must be considered. In order to help manage competing objectives, we used an evolutionary multi-objective optimization technique, NSGA-II, for generating new network configurations for an industrial remote data mirroring application. During this process, we observed the presence of a hidden search factor that restricted NSGA-II’s search from expanding into regions where valid optimal solutions were known to exist. In follow-on empirical studies, we discovered that a variable-length genome design causes unintended interactions with crowding distance mechanisms when using discrete objective functions.