Exploiting Overlap When Searching for Robust Optima

This paper introduces a new view on fitness evaluation when searching for robust optima. It proposes to compare solutions in (successive) populations with respect to how they rank in robustness instead of aiming for accurate robustness estimates. This can be done by focusing on the non-overlapping parts of the regions of uncertainty of each pair of candidate solutions. An initial step toward a scheme implementing this view is made with the analysis and experiments on a simple scenario comparing two solutions on uniform input noise.