VS-RLS algorithm boosts multi-objective combinatorial optimization by adapting search stepsize

A team led by Xuepeng Ren has introduced a new local search method called VS-RLS (Variable Stepsize Randomized Local Search) for multi-objective combinatorial optimization problems (MOCOPs). While most evolutionary multi-objective optimization research has focused on continuous domains, combinatorial problems like scheduling, routing, and resource allocation have different structures and landscapes. Recent studies show that simple randomized local search can actually outperform complex multi-objective evolutionary algorithms (MOEAs) on MOCOPs. However, existing local search methods rely on a fixed neighborhood size, which limits exploration and often traps the search in local optima.

VS-RLS tackles this by dynamically adjusting the stepsize during the search process. In early phases, it uses a larger stepsize to explore broadly across the search space. As the search progresses, it gradually reduces the stepsize for fine-grained exploitation. This adaptive approach balances exploration and exploitation without complex metaheuristics. The researchers demonstrated VS-RLS's effectiveness and generalizability through extensive evaluations against both local search methods and MOEAs on diverse MOCOPs. The paper is available on arXiv (2602.05675) and represents a significant practical advance for real-world combinatorial optimization tasks where multiple objectives must be balanced.