发表机构
McGill University; COSMO – Stochastic Mine Planning Laboratory, McGill University; Ghent University; University College London; Vlerick Business School; Universidade Aberta; INESC TEC（麦吉尔大学; 麦吉尔大学 COSMO 随机矿山规划实验室; 根特大学; 伦敦大学学院; 弗莱里克商学院; 开放大学; INESC TEC）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
本文提出大语言模型引导的基于种群的框架，用于自动设计资源受限项目调度问题的优先级规则，实验表明该框架设计的规则性能优于传统规则及遗传编程设计的规则。
AI 中文摘要
资源受限项目调度问题（RCPSP）的目标是在满足 precedence（优先顺序）和可再生资源约束的前提下最小化总工期。优先级规则启发式算法能以较低的计算成本构建可行调度方案，并具备明确的决策逻辑，因此在实际应用中被广泛采用，是计算密集型方法的理想替代方案。然而，没有任何传统规则能在所有项目中表现一致，因此研究人员对自动优先级规则设计展开了研究。遗传编程（GP）超启发式算法是该任务的主流方法，但要进化出高性能规则，可能需要在训练项目上评估大量候选规则。近年来大语言模型（LLM）的进展使得利用性能反馈生成并迭代修改优先级规则成为可能。本文提出一种大语言模型引导的基于种群的自动优先级规则设计框架。在离线搜索过程中，大语言模型生成并修改候选规则，训练项目上的调度质量决定候选规则的适应度并指导后续修改。搜索结束后，将返回最优规则并直接应用于未见过的项目，无需进一步搜索或调用大语言模型。实验表明，大语言模型设计的规则在测试集上优于传统单一规则，且在可比搜索工作量下优于遗传编程设计的规则，同时与大规模遗传编程搜索得到的规则具有竞争力。在大型项目中，选定的大语言模型设计规则优于所有考虑的传统规则，以及高度并行项目上的两种遗传算法配置。一项 ablation study（消融研究）考察了主要搜索组件的变化如何影响性能，而规则分析则描述了大语言模型设计规则的结构和决策行为。
英文摘要
The objective of the resource-constrained project scheduling problem (RCPSP) is to minimize makespan while satisfying precedence and renewable-resource constraints. Priority-rule heuristics construct feasible schedules with low computational cost and explicit decision logic, making them widely used in practice and an attractive alternative to more computationally intensive methods. However, no traditional rule performs consistently well across projects, and researchers have therefore investigated automated priority-rule design. Genetic programming (GP) hyper-heuristics have been the predominant approach to this task, but evolving a high-performing rule may require evaluating many candidate rules on the training projects. Recent advances in large language models (LLMs) make it possible to generate and iteratively revise priority rules using performance feedback. This paper presents an LLM-guided population-based framework for automated priority-rule design. During the offline search, an LLM generates and revises candidate rules, while schedule quality on training projects determines candidate fitness and guides subsequent revisions. At the end of the search, the best rule is returned and applied directly to unseen projects without further search or LLM calls. Experiments show that the LLM-designed rules outperform traditional single rules across the test sets and outperform GP-designed rules obtained under comparable search effort, while remaining competitive with rules obtained from a substantially larger GP search. On large projects, selected LLM-designed rules outperform all considered traditional rules and two genetic algorithm configurations on highly parallel projects. An ablation study examines how changes to the main search components affect performance, while rule analyses describe the structure and decision behavior of the LLM-designed rules.
CommentsSupplementary material available as an ancillary file
