Skip to content

Intelligent Human-Machine Partnership for Manufacturing: Enhancing Warehouse Planning through Simulation-Driven Knowledge Graphs and LLM Collaboration

🕒 Published (v1): 2025-12-20 08:09 UTC · Source: Arxiv · Venue: AAAI · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

This paper presents a two-stage framework that converts Discrete Event Simulation (DES) output from a warehouse model into a Neo4j Knowledge Graph, then exposes it to a dual-path LLM agent (operational QA chain + iterative reasoning chain) so planners can diagnose bottlenecks via natural language. The system achieves Pass@1 of 0.92 and perfect Pass@2 of 1.00 on 25 operational questions, outperforming single-pass and self-reflection baselines.

Problem

Discrete Event Simulation generates voluminous, granular operational logs that manufacturing planners cannot efficiently query without specialized scripting skills. Existing approaches use manual inspection or bespoke scripts, creating silos between operational experts and analytical tools, and failing to surface cascading root causes of bottlenecks.

Method

Stage 1 — KG Construction: DES event logs (timestamps, resource states, package flow) are parsed and materialized into a Neo4j graph using a warehouse-specific ontology (nodes: Supplier, Worker, AGV, Forklift, Package, Storage; edges: typed with timestamp annotations).

Stage 2 — Dual-Path LLM Agent (via LangGraph + GPT-4/5): - Query classifier routes incoming natural-language questions to one of two chains. - QA Chain (operational): Step Generation Module decomposes the query → Cypher Generation Module produces graph queries → Query Execution + Self-Reflection loop auto-corrects syntax/semantic errors → Answer Synthesis Module aggregates results. - Iterative Reasoning Chain (investigative): A Reasoning Module generates sequential sub-questions conditioned on accumulated KG evidence; each sub-question runs the full QA Chain; a Sufficiency Assessment decides when to stop and trigger a Findings Summarizer.

Temperature is set 0.0–0.3; self-reflection is embedded per step (not post-hoc).

Key Contributions

  • First application of KG + LLM agent pipeline specifically to DES output data for warehouse planning.
  • Dual-path architecture distinguishing direct operational retrieval from iterative diagnostic reasoning.
  • Step-wise query decomposition with per-step self-reflection, shown to outperform monolithic query generation and post-answer self-reflection baselines.
  • Validated on three injected-fault investigative scenarios (stage transfer delay, degraded forklift, supplier-specific processing delay) with human expert evaluation across seven quality dimensions.

Results

  • Operational QA (Pass@k, 25 questions):
  • Proposed framework: P@1 = 0.92, P@2 = 1.00
  • Baseline + self-reflection: P@1 = 0.70, P@2 = 0.82
  • Baseline (single-pass): P@1 = 0.59, P@2 = 0.71
  • Worst baseline category: "Worker" P@1 = 0.33; framework achieves 0.75 P@1 / 1.00 P@2 on same category.
  • Investigative QA (human expert scoring, 0–10, 12 analysis logs across 3 scenarios):
  • Necessity: 9.0; Completeness: 8.58; Factuality: 8.46; Logical Coherence: 8.21; Clarity: 8.04; Sufficiency: 8.0; Query Quality: 7.96.
  • Case study (Scenario 1): Framework correctly identified "Wait to Worker" as root bottleneck for CamelCargo (6,870 s vs. 4,955 s global avg; one package waited 3,654 s vs. 1,375 s avg) and traced downstream AGV over-utilization (54.8% vs. 41.7%) and forklift under-utilization (23.7% vs. 40.1%).

Limitations

  • KG schema design requires substantial domain expertise and upfront engineering investment.
  • Cypher generation reliability is not guaranteed for novel operational scenarios; continuous monitoring and confidence scoring are needed.
  • Validated only on a single warehouse unloading simulation; generalizability to picking, packing, inventory management, and other DES frameworks is undemonstrated.
  • Evaluation of investigative scenarios relies on human expert judgment (3 scenarios, qualitative), not a large-scale automated benchmark.
  • The paper references "GPT-5" (OpenAI 2025), which was not publicly released as of the paper's arXiv submission date (Dec 2025), making reproducibility unclear.

Relevance to Agentic AI / LLM Agents

This work demonstrates a concrete instantiation of tool-augmented, iterative reasoning agents in a high-stakes industrial domain: the LLM agent autonomously decomposes problems, generates and self-corrects structured graph queries (Cypher), evaluates sufficiency of gathered evidence, and loops until diagnosis is complete — all without human intervention in the loop during inference. The dual-path architecture (direct retrieval vs. multi-hop investigation) is a practically grounded example of adaptive planning under uncertainty. The per-step self-reflection mechanism directly addresses a known failure mode of monolithic LLM-to-query translation and adds to the growing body of evidence that decomposition + reflection significantly outperforms single-pass generation. For researchers tracking agentic AI, this paper offers a worked example of grounding LLM reasoning in a structured external knowledge base (KG-as-tool) to reduce hallucination while preserving natural-language access.