Dynamic Landscape Analysis for Constrained Multiobjective Optimization Problems

Landscape analysis is a data-driven approach that involves sampling the search space of an optimization problem to generate a range of statistical features. These features serve to characterize the 'problem difficulty'. However, the computational costs associated with offline independent sampling can be excessive, and this approach often overlooks valuable information accumulated by the optimization algorithm. This paper aims to expand our understanding of landscape analysis in the domain of black-box constrained multiobjective optimization problems. We demonstrate the potential of leveraging optimization algorithm trajectories to measure landscape features. Our findings underscore the significance of utilizing landscape features as a means to approximate algorithm performance, particularly in cases involving new instances lacking a known reference set. Ultimately, our goal is to employ landscape analysis to dynamically adapt algorithm constraint handling techniques.