Many-objective visual analytics: In search of search-as-a-service

Newly available computational power and sophisticated visualization tools make it possible to use optimization in more sophisticated ways. In particular, optimizing a variety of problem formulations permits the exploration of competing hypotheses about how the design problem is framed, engaging in a process of constructive learning that develops problem insight and leads to the discovery of new design alternatives. This enables a many objective visual analytics (MOVA) process but requires a multi-objective evolutionary algorithm (MOEA) that works without significant user intervention. This creates a nead for search-as-a-service for which many MOEAs are not well suited. To determine MOEAs' suitability for use within the MOVA context, a comparative study is presented, in which five leading MOEAs are evaluated on their effectiveness, efficiency, reliability, and controllability on four different formulations of the same benchmark conceptual design problem. For researchers proposing new MOEAs, this study provides a baseline for MOEA effectiveness, efficiency, reliability and controllability across multiple problem formulations. For designers, the massive computational scale of this study yields results that point the way to reducing their own computational requirements, or equivalently making the most of a fixed computational budget. Based on our findings, the new Borg MOEA is established as a leading algorithm for search-as-a-service.