Uncertainty-Aware Multi-Agent AI for Autonomous Inventory and Order Orchestration in Multi-Echelon Supply Chains
编号:218
访问权限:仅限参会人
更新:2026-11-16 10:12:48 浏览:5次
Online
摘要
Global multi-echelon supply networks face severe vulnerabilities from cascading demand shocks, lead-time volatility, and supplier capacity curtailments. Traditional centralized planning and decoupled heuristics trigger pervasive bullwhip instability or require capital-intensive inventory over-buffering. This paper introduces Uncertainty-Aware Multi-Agent Autonomous Orchestration (U-MAAO), a decentralized multi-agent framework integrating: (i) Adaptive Conformal Inference (ACI) for distribution-free quantile demand forecasting; (ii) a two-level coordination architecture combining a Constrained Allocation Consensus Kernel (CACK) that reconciles upstream replenishment proposals under shared vendor capacity via proximal ADMM, and an Inter-Warehouse Transshipment Agent (IWTA) executing a three-stage lexicographic multi-SKU linear program across 244 directed lateral corridors under route-specific freight costs ($0.35–$0.75/unit), transit SLAs (τ_{ij} ≤ 1 d), destination storage headroom, and shared corridor capacities; and (iii) local Differential Privacy (DP) preserving proprietary transaction volumes under bounded Laplace perturbation (ε = 0.50, δ = 0). Across continuous 100-day episodes on a calibrated 35-node, 35-SKU benchmark subjected to step surges (+22%/+25%), uniform +1 day transit delays across all network inbound arcs, and 50% supplier capacity collapses, U-MAAO maintains a 99.60% ± 0.01% on-hand unit fill rate and restricts order-line stockouts to 0.77% ± 0.02% (saving 0.61 percentage points in stockouts versus demand-chasing heuristics, p < 0.001), while saving $9.4k–$10.4k monthly in inventory holding costs and dampening bullwhip to 1.03 ± 0.00 (p < 0.001 vs. proportional smoothing heuristics). Reference median decision latency is 23.86 ms (P90: 25.10 ms, P99: 27.84 ms, tail bounded below 50 ms, with the three-stage lexicographic transfer LP executing in 19.03 ms median), demonstrating a resilient, computationally efficient architecture for autonomous multi-echelon orchestration.
关键词
Autonomous inventory orchestration, multi-agent AI, adaptive conformal inference, proximal ADMM consensus, lateral transshipments, supply chain resilience.
稿件作者
Sudheer Kumar Rallabandi
WiseTech Global
发表评论