The dynamic flexible job shop scheduling problem (DFJSSP) presents a challenge for modern flexible manufacturing systems, where heuristic dispatching rules (HDRs) are widely adopted for their rapid response capabilities. While automated design methods such as genetic programming have advanced the field, their reliance on random search operators motivates more structured, knowledge-driven approaches. This paper proposes Evo-TextGrad, a framework combining population-based evolutionary search with large language model (LLM)-driven feedback aggregation. Rather than relying on single-point feedback, Evo-TextGrad collects critique signals across an entire population of candidate HDRs and aggregates them to identify consistent improvement patterns, enhancing rule quality even when local LLMs are used. To enable cost-effective local deployment, a knowledge distillation pipeline automatically constructs a dataset of evolutionary success patterns by capturing prompt-HDR pairs that yield performance improvements. This curated dataset transfers HDR design capability to a local 7B model via LoRA-based fine-tuning, producing ModouGPT. Experiments on 1000 dynamic FJSSP instances against strong load-balancing baselines and evolutionary hyper-heuristic methods show that Evo-TextGrad achieves favourable performance over classical rules and state-of-the-art LLM-based frameworks, while ModouGPT matches larger models at lower resource cost. A real-world flexible manufacturing case study validates industrial effectiveness.
Huang et al. (Wed,) studied this question.