Large Language Model-Driven Full-Component Evolution of Adaptive Large Neighborhood Search
A new framework uses LLMs to automatically rebuild a key logistics algorithm, cutting the optimality gap by 77%.
A team of researchers has introduced a novel framework that uses large language models (LLMs) to fully automate the design and evolution of a critical optimization algorithm. The work focuses on Adaptive Large Neighborhood Search (ALNS), a metaheuristic widely used for complex production and logistics problems like vehicle routing. Traditionally, crafting an effective ALNS algorithm requires significant manual effort from domain experts to design components like 'destroy' and 'repair' operators. This new framework, however, decouples ALNS into seven core modules and uses an LLM within a closed-loop evolutionary process to automatically generate and rebuild all of them from scratch.
By incorporating the MAP-Elites mechanism, the system maintains a diverse archive of high-performing algorithm designs, evolving both solution quality and strategic diversity. When tested on standard TSPLIB benchmarks, the AI-evolved algorithms consistently outperformed optimized, hand-crafted ALNS baselines. The improvement was most dramatic on large-scale instances, where the average optimality gap plummeted from 3.18% to just 0.74%—a 77% reduction. The evolution also uncovered novel, counterintuitive design patterns that provide new theoretical insights. Furthermore, the study compared different LLMs, revealing clear differences in their ability to support this complex engineering task, offering practical guidance for model selection in real-world applications.
- The framework automates the design of all 7 components of the ALNS algorithm, a task traditionally done manually by experts.
- On large-scale optimization benchmarks, it reduced the average optimality gap by 77%, from 3.18% to 0.74%.
- The evolutionary process discovered novel algorithm designs and provided a benchmark for comparing LLMs on complex engineering tasks.
Why It Matters
This demonstrates AI's potential to automate the design of complex software systems, drastically accelerating development for logistics, supply chain, and operational research.