Elite quality and diversity in GA initialization: A comparative evaluation of constraint-adaptive initialization strategies for genetic algorithms in CVRP

Abstract

The effectiveness of a Genetic Algorithm (GA) in solving the Capacitated Vehicle Routing Problem (CVRP) is strongly influenced by population initialization. This study investigates the quality–diversity trade-off at the initialization stage by comparatively evaluating nine strategies, including random generation, Nearest Neighbor (NN), k -Nearest Neighbor ( k -NN), and hybrid seeding approaches that combine random and NN methods. All strategies are assessed within a controlled GA framework on selected Set X benchmark instances and analyzed using non-parametric statistical procedures. Three initial population metrics — Initial Objective (Best), Initial Objective (Mean), and Initial Genetic Diversity — are examined in relation to final performance indicators, including Global Best Objective, Final Objective (Mean), Final Genetic Diversity, convergence behavior, and computational time. The results indicate that, within the evaluated framework and benchmark set, Initial Objective (Best) is the strongest observed predictor of final solution quality, whereas Initial Genetic Diversity does not show statistically significant evidence of acting as a primary predictor at the method-rank level, suggesting instead a secondary and supportive role in this experimental setting. Consequently, purely random initialization consistently underperforms due to the lack of strong exploitable solution seeds. Furthermore, although k -NN does not produce the strongest final solutions on its own, it consistently generates structured and high-quality diversity, making it a reliable diversity-enhancing component for initialization within the evaluated GA framework.

The authors' abstract, as published at the source. Swarm and Evolutionary Computation, 2026 · DOI ↗