Dynamic hybrid flow shop scheduling problems (DHFSP) are critical in modern manufacturing systems, where uncertainties such as order fluctuations and equipment failures pose significant challenges. Traditional exact methods and metaheuristics struggle to meet real-time decision-making requirements under such dynamic conditions. Heuristic dispatching rules (HDRs) have been widely adopted for their rapid response capabilities. In recent years, large language models (LLMs) have demonstrated remarkable capabilities in code generation and logical reasoning, showing promising potential for automated HDR design. However, existing LLM-based methods predominantly adopt single-population evolution strategies, which suffer from insufficient population diversity, limited semantic-level reasoning, and premature convergence, thereby frequently becoming trapped in local optima. To address these challenges, this paper proposes a personalized multi-island reflective evolution framework that assigns distinct exploration personalities to multiple parallel sub-populations and incorporates an LLM-driven semantic reflection mechanism to achieve efficient search space coverage and enhanced exploration depth. The framework employs a two-stage strategy: offline training constructs a robust rule library through diversified dynamic scenarios, while online application enables rapid real-time decision-making. Experimental results on 300 test instances demonstrate that the proposed method outperforms traditional HDRs, conventional evolutionary rule generation methods, and state-of-the-art LLM-based approaches, exhibiting superior stability and generalization capability.
Lei et al. (2026) studied this question.