Are NSGA-III Implementations Really Stable? A Cross-Platform Analysis of Anomalous Behaviors of NSGA-III

NSGA-III is widely used as a baseline algorithm in multi- and many-objective optimization. Since no official source code was released, researchers rely on third-party implementations, with PlatEMO and pymoo being the most frequently used. In this paper, we examine whether these implementations exhibit different search behaviors on continuous multi-objective optimization problems. By visualizing the solution distributions of several examples, we confirm clearly divergent search behaviors that were previously not reported on continuous multi-objective optimization problems. We also note the difference between the implementations of PlatEMO's NSGA-III and pymoo's NSGA-III, and attribute their difference in search behavior to the different implementations.