Reproduce “Large Language Model-Driven Full-Component Evolution of Adaptive Large Neighborhood Search” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shaohua Yu, Tianyu Chen, Linyan Liu, Jakob Puchinger
Researchers introduce a large language model (LLM)–driven evolutionary framework that automatically redesigns all seven core components of Adaptive Large Neighborhood Search (ALNS). This full-component evolution consistently surpasses carefully tuned human-designed ALNS baselines on Traveling Salesman Problem (TSP) and Capacitated Vehicle Routing Problem (CVRP) benchmarks, achieving lower optimality gaps and improved computational efficiency while also revealing novel algorithmic design principles.
Key insights
- LLM-driven full-component evolution can discover novel, counterintuitive algorithmic principles like scale-normalized acceptance, non-monotone perturbation control, and diversity-preserving adaptive learning mechanisms.
- Cross-component compatibility is a crucial factor, as multi-expert paradigms often produce individually strong components, but single-expert evolution sometimes yields better integrated, full ALNS systems.
- Control policies evolved on more constrained problems (e.g., CVRP) demonstrate better transferability and efficiency to less constrained problems (e.g., TSP) than the reverse.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.