New study reveals how objective interactions affect massive optimization algorithms

A new study from researchers at Michigan State University (Shahbandegan, Hernandez, Dolson) tackles two underexplored questions in many-objective optimization: what happens when the number of objectives grows beyond the typical 15 into the massive regime, and how do problem characteristics like objective interactions influence algorithm performance? Using a diagnostic benchmark suite that controls problem features and scales to extremely high objective counts, the team evaluated four state-of-the-art evolutionary algorithms: NSGA-II, NSGA-III, MOEA/D, and lexicase selection.

The experiments revealed that problem characteristics — especially the nature of interactions between objectives — dramatically affect algorithm performance. Importantly, lexicase selection, originally designed for genetic programming, compared favorably with the other algorithms while avoiding dependence on predefined reference directions. These results highlight that understanding objective interactions is critical before selecting an algorithm for a specific massive-objective problem, challenging the assumption that more objectives merely demand scaling existing methods.