LLMs evolve destruction operators for flow shop scheduling
New IG-DOE algorithm uses LLM-evolved operators to beat state-of-the-art QIG.
The permutation flow shop scheduling problem (PFSP) is a classic NP-hard combinatorial optimization challenge in manufacturing. Classical iterated greedy (IG) algorithms rely on a single fixed destruction operator, leading to search stagnation on large instances. To overcome this, researchers introduce IG-DOE (Iterated Greedy with Destruction Operator Ensemble), which switches among heterogeneous destruction operators only when stagnation is detected. The core mechanism—stagnation-triggered sequential switching—enriches perturbation behavior without manual tuning.
To automate the design of the destruction operator ensemble, the team proposes SCOE (Stagewise Cooperative Operator Evolution), an LLM-assisted framework. SCOE evolves a high-quality set of destruction operators through stagewise evolution, state-awareness, and cooperative evaluation. On the challenging VRF-hard-large benchmark, the evolved DOE generalizes from small to large instances, achieving significantly better average performance than QIG (a state-of-the-art IG variant) under identical CPU-time limits. Experiments on real-world industrial data further confirm that the evolved operators adapt to different data distributions without extra tuning, demonstrating practical viability for intelligent manufacturing.
- IG-DOE uses stagnation-triggered sequential switching among multiple destruction operators to avoid search stagnation.
- The SCOE framework leverages LLMs to automatically evolve destruction operator ensembles without expert crafting.
- On VRF-hard-large benchmark, IG-DOE outperforms state-of-the-art QIG under same CPU time and generalizes to larger unseen instances.
Why It Matters
Automated LLM-guided operator evolution can solve complex scheduling problems faster, reducing manual tuning in manufacturing.