大模型引导的仿真轨迹启发式设计：动态生产与AGV调度案例研究

文章背景与核心概要

传统的基于仿真优化（SBO）的方法通常将仿真器视为黑盒，完全依赖聚合性能分数来评估候选策略。虽然这种方法能够对候选策略进行排序，但它无法深入揭示策略失败的原因，也无法提供如何修改其底层逻辑的洞察。

为了克服这一局限性，本文引入了一种大模型引导的启发式设计框架，将重复的仿真选择与事件级轨迹诊断相结合。通过结合 Gemini-3.1-Pro 在动态生产和自动导引车（AGV）调度离散事件仿真中的测试，该框架在多个运行周期和随机干扰场景下，显著优于传统的滚动混合整数线性规划（MILP）、基于规则以及元启发式基线方法，为工业复杂调度问题开辟了新途径。

LLM-Guided Heuristic Design from Simulation Traces: A Case Study in Dynamic Production and AGV Scheduling

Authors: Jinbo Li, Chuanhao Li
Primary Subject: Artificial Intelligence (cs.AI)
arXiv Identifier: arXiv:2608.09343 [cs.AI]
Submitted: 10 August 2026
Links: View PDF | HTML Version | DOI

摘要 (Summary)

仿真基准优化（SBO）传统上将仿真器视为黑盒，完全依赖综合性能得分来评估候选策略。尽管这种方法能够对候选策略进行排名，但它无法提供关于策略为什么失败或应该如何修改其底层逻辑的深入见解。

为了克服这一局限性，本文引入了一个大模型引导的启发式设计框架，该框架将反复的仿真选择与事件级的轨迹诊断结合起来。 * 评估与诊断： 每个策略候选都在多个仿真复现（replications）中进行测试。得分最低的复现会被重新运行，以生成可查询的事件级轨迹。 * 改进优化： 管理员智能体（manager agent）分析轨迹以形成瓶颈假设，而并行编辑智能体则实现代码级别的修改。只有通过执行检查和反复评估验证的改进才会被保留。

在动态生产与自动引导车（AGV）调度的离散事件仿真中，利用 Gemini-3.1-Pro 对该框架进行了测试。结果表明，在多个运营时间范围和随机干扰场景下，该框架显著超越了传统的滚动MILP、基于规则以及元启发式基线方法。

Simulation-based optimization (SBO) traditionally treats simulators as black boxes, relying entirely on aggregate performance scores to evaluate candidate policies. While this approach ranks candidates, it offers no insight into why a policy fails or how its underlying logic should be modified.

To overcome this limitation, this paper introduces an LLM-guided heuristic design framework that pairs repeated simulation selection with event-level trace diagnosis. * Evaluation & Diagnosis: Each policy candidate is tested across multiple simulation replications. The lowest-scoring replication is replayed to generate a queryable event-level trace. * Refinement: A manager agent analyzes the trace to form bottleneck hypotheses, while parallel editing agents implement code-level revisions. Only improvements validated by execution checks and repeated evaluations are retained.

Tested in a discrete-event simulation of dynamic production and automated guided vehicle (AGV) scheduling using Gemini-3.1-Pro, the framework significantly outperformed conventional rolling-MILP, rule-based, and metaheuristic baselines across multiple operational horizons and randomized disruption scenarios.

核心发现与结果 (Key Findings & Results)

性能提升： 在五次独立的优化运行中，最终的平均得分达到 77.51（满分100分）。在得分最高的运行中，基于轨迹的诊断——例如主动充电、距离感知的AGV分配以及重新平衡的调度优先级——将历史最佳平均得分从 62.49 提升至 78.61。
鲁棒性： 针对 100 个匹配的随机种子进行评估时，表现最优的策略在每一个随机种子上的得分都持续超过所有基线方法。此外，它在随机故障发生时仍能保持性能优势，而无需重新进行优化。
消融实验洞察： 移除并行候选生成或取消对轨迹数据库访问的实验导致了较低的最终平均得分，这证明了这两个组件在引导有针对性的代码级策略改进方面具有至关重要的作用。

Performance Gains: Across five independent optimization runs, final mean scores averaged 77.51 (on a 0–100 scale). In the highest-scoring run, trace-based diagnoses—such as proactive charging, distance-aware AGV assignment, and rebalanced dispatch priorities—raised the best-so-far mean score from 62.49 to 78.61.

Robustness: Evaluated against 100 matched seeds, the top-performing policy consistently outscored all baseline methods on every single seed. Furthermore, it retained its performance advantage under random faults without requiring re-optimization.

Ablation Insights: Experiments removing either parallel candidate generation or access to the trace database resulted in lower final mean scores, demonstrating the critical importance of both components in guiding targeted code-level policy improvements.