LLM-powered self-evolution framework beats traditional algorithms for factory scheduling
SeEvo uses LLMs to design scheduling rules that outperform both GP and deep RL by 20%+
A new framework called SeEvo (Self-Evolution) uses large language models to automatically design heuristic dispatching rules (HDRs) for the dynamic fuzzy job shop scheduling problem (DFJSSP). This problem is common in manufacturing where jobs arrive unpredictably, processing times are uncertain, and schedules must be generated in real time. Traditionally, HDRs are manually crafted by experts or evolved through genetic programming (GP), but those approaches struggle with high randomness and produce rules that often fail in unseen scenarios.
SeEvo addresses this by combining LLMs with a teacher-student learning mechanism. The teacher model has full knowledge of actual processing times (including fuzzy uncertainties) and guides the student LLM to generate robust HDRs using only historical deviations as input. The framework evolves a population of HDRs over generations, with the LLM acting as both a generator and a refiner. Experimental results show SeEvo significantly outperforms GP, GEP, deep reinforcement learning (DRL), and over ten commonly used HDRs from literature, especially in dynamic and previously unseen environments. The paper is accepted at IEEE Transactions on Fuzzy Systems and marks a step toward fully automated algorithm design powered by LLMs.
- SeEvo uses a teacher-student mechanism where the teacher knows real processing times, teaching the LLM to infer fuzzy uncertainties from historical deviations.
- The framework outperforms GP, GEP, and DRL methods in dynamic, unseen scheduling scenarios, not just static benchmarks.
- Accepted at IEEE Transactions on Fuzzy Systems; the LLM generates robust HDRs by evolving a population of rules through self-reflection.
Why It Matters
SeEvo could replace manual scheduling rule design with automated LLM-driven optimization, saving factories time and improving adaptability.