A Large Language Model-Assisted Reinforcement Learning Framework with Evolutionary Algorithm for Hybrid Flow-Shop Scheduling
针对混合流水车间调度问题，提出一种结合大语言模型、进化算法和强化学习的混合框架，在峰值功率约束下同时最小化完工时间和总能耗，实验表明优于现有算法。
With the growing emphasis on sustainable manufacturing, green scheduling has gained prominence from practitioners and researchers in the industry manufacturing enterprises. This article addresses the hybrid flow shop scheduling problem with peak power consumption constraints (HFSPP) to minimize makespan and total energy consumption (TEC), while ensuring real-time power usage does not exceed a predefined threshold. A hybrid framework, LLM-MODPPO-Evo, is proposed to solve HFSPP by simultaneously minimizing makespan and TEC. The framework consists of three parts: 1) large language models (LLMs) are introduced to automatically design algorithms and generate initial solutions; 2) evolutionary algorithm based on adaptive evolutionary strategy of covariance matrix (CMA-ES) is adopted to diversify solutions; 3) multi-objective distributed proximal policy optimization (MODPPO) is utilized to refine solutions. Firstly, initial scheduling algorithms generated by LLMs via tailored prompt frameworks produce diverse heuristic operators and decoding methods. Secondly, after screening feasible algorithms and producing the initial schemes, these solutions are extended through an EA to create a population of the improved schemes. Finally, the populations are adaptively optimized by MODPPO to yield the best Pareto-optimal solution set to balance makespan and TEC under peak power consumption constraints. The comparative results demonstrate that the framework outperforms that of the existing algorithms in solving the HFSPP.