Learning-assisted improvements in Adaptive Variable Neighborhood Search

Abstract

This study presents the design and integration of novel adaptive components within the Double-Adaptive General Variable Neighborhood Search (DA-GVNS) algorithm, aimed at improving its overall efficiency. These adaptations utilize iteration-based data to refine the search process, with enhancements such as an adaptive reordering mechanism in the refinement phase and a knowledge-guided approach to adjust the search strategy. Additionally, an adaptive mechanism for dynamically controlling the shaking intensity was introduced. The proposed knowledge-guided adaptations demonstrated superior performance over the original DA-GVNS framework, with the most effective scheme selected for further evaluation. Initially, the symmetric Traveling Salesman Problem (TSP) was used as a benchmark to quantify the impact of these mechanisms, showing significant improvements through rigorous statistical analysis. A comparative study was then conducted against six advanced heuristics from the literature. Finally, the most promising knowledge-guided GVNS (KG-GVNS) was tested against the original DA-GVNS on selected instances of the Quadratic Assignment Problem (QAP), where detailed statistical analysis highlighted its competitive advantage and robustness in addressing complex combinatorial optimization problems.