基於多代理人之供應鏈協同調度架構： 存貨異常診斷與大型語言模型輔助訂貨決策支援

Abstract

供應鏈營運環境日益動態且不確定，庫存補貨決策已不再只是補貨時點與數量之計算問題。當需求波動、資料不一致、供應延遲、跨倉支應限制與作業規則衝突同時發生時，管理者需整合異常脈絡、檢核限制條件、比較候選方案，並保留可供後續審核與追蹤之決策依據。傳統庫存控制與作業研究方法雖能提供數值計算與限制檢核基礎，但在語意理解、脈絡整合、角色協作與決策追蹤方面，仍需更完整的決策支援機制。

基於此，本研究提出一套以大型語言模型驅動之多代理系統為核心的庫存補貨異常決策支援架構。正式業務知識圖譜用以保存補貨規則、績效條件與作業限制，作為代理判斷之正式知識依據；經驗知識圖譜則保存系統建議、人工審核、最終決策及其差異。多代理模組透過共享狀態與狀態驅動工作流分工協作，並結合確定性工具、執行期控制、限制條件檢核、人機協同審核與查核追蹤，使候選結果須經驗證與人工確認後，方可進入後續流程。

為評估所提出架構之可行性，本研究以汽車輪胎補貨實務案例進行驗證，並以人工實際訂購結果作為比較基準，透過訂購判斷一致率、訂貨召回率、數量吻合率、漏訂件數與過訂件數，檢視系統建議與人工決策之接近程度。結果顯示，系統在是否訂貨之方向判斷上具有一定穩定性，後續重算階段之訂貨件號捕捉能力與數量吻合程度亦有所改善；惟部分案例仍存在漏訂、過訂及建議數量差異，顯示系統在貼近實際補貨判斷方面仍有改善空間。

整體而言，本研究之貢獻不在於以單一模型取代既有補貨方法或人工決策，而在於建立一套整合語意理解、正式知識查詢、確定性計算、多代理協作、人工審核與經驗回饋的受控決策支援流程，使補貨建議具備可執行、可檢核、可審核與可追溯之特性。此架構具有降低重複性人工檢查、提升審核一致性及保存專家決策經驗之應用潛力，其長期營運成效與知識回饋效果仍有待後續研究進一步驗證。

Inventory replenishment under dynamic and uncertain supply chain conditions involves more than determining order timing and quantity. Managers must interpret anomaly contexts, examine operational constraints, compare feasible alternatives, and retain decision evidence for subsequent review. Although conventional inventory-control and Operations Research methods provide a rigorous basis for quantitative analysis, they remain limited in integrating contextual information, coordinating multiple decision roles, and preserving traceable decision processes.

This study proposes an inventory replenishment anomaly decision-support framework centered on an LLM-driven multi-agent system. The Business Knowledge Graph provides formal replenishment rules, performance requirements, and operational constraints, while the Empirical Knowledge Graph records system recommendations, human reviews, final decisions, and decision differences. Shared state, deterministic tools, runtime control, constraint validation, Human-in-the-Loop review, and audit tracing are integrated to ensure that candidate outputs are verified before being committed to subsequent stages.

The framework is evaluated through an industry-based tire replenishment case study, using actual human ordering decisions as the comparison baseline. Ordering-decision agreement, order recall, quantity match rate, missed orders, and over-orders are adopted to assess the consistency between system recommendations and human decisions. The results show that the framework provides relatively stable order-or-no-order judgments and improves item identification and quantity matching after recalculation. Nevertheless, missed orders, over-orders, and quantity discrepancies remain in several cases.

The contribution of this study lies not in replacing existing replenishment methods or human judgment, but in establishing an executable, verifiable, reviewable, and traceable decision-support process. The framework offers practical potential for reducing repetitive manual checking, improving review consistency, and preserving expert decision knowledge.