Supply Chain Resilience Optimization Integrating Transformers and Multi-Agent Reinforcement Learning

Abstract

INTRODUCTION: Manufacturing supply chains face increasing vulnerabilities from delayed risk identification, conflicting node-level decisions, and escalating recovery costs, particularly under complex disruption scenarios. Existing approaches often lack proactive, coordinated responses across autonomous supply chain nodes, limiting their ability to anticipate disruptions and synchronize mitigation actions effectively. To address these gaps, this paper proposes an integrated resilience optimization framework that combines predictive analytics with multi-agent coordination.

OBJECTIVES: The primary objective is to develop a supply chain resilience framework, termed T-MARL, that enables autonomous suppliers, manufacturers, warehousing nodes, and distributors to: (1) predict future disruption probabilities using temporal risk precursor data, and (2) execute coordinated, proactive replenishment, production, and distribution decisions under inventory and capacity constraints. The framework aims to improve resilience index, reduce total operating costs, and shorten recovery time compared to baseline strategies, while also examining the sensitivity of performance gains to prediction noise and communication latency.

METHODS: The T-MARL framework models each supply chain node as an independent agent. A lightweight Transformer captures temporal features from 6-dimensional risk precursor sequences over 12 consecutive periods to forecast disruption probabilities. These predictions feed into a value-decomposition multi-agent reinforcement learning mechanism with centralized training–decentralized execution, where a value-decomposition network orchestrates cross-node actions. Risk-constrained action gating ensures that proactive responses adhere to inventory and capacity limits. A digital twin simulation environment was built with fixed random seeds, covering three scenarios—single-source supply disruption, demand surge, and compound disruption. Five strategies were tested, with 30 runs per unit using common random number pairing.

RESULTS: On the trajectory-held-out test set, the Transformer achieved an AUC of 0.722 and an F1-score of 0.586, while Random Forest achieved a slightly higher AUC of 0.724. Under the compound disruption scenario, T-MARL attained a resilience index of 0.849—a 1.63 percentage-point gain over VDN-MARL -reduced total cost by 14.7%, and shortened average recovery time from 1.67 periods to zero. T-MARL also achieved lower operating costs while maintaining service levels comparable to rule-based strategies. However, the paper also found that prediction noise and communication latency diminish these performance gains.

CONCLUSION: This research contributes a reproducible digital twin analysis framework for disruption prediction, cross-node collaborative response, and rapid recovery in manufacturing supply chains. While T-MARL demonstrates significant improvements in resilience and cost-efficiency under ideal conditions, practical deployment must account for real-world imperfections such as noisy forecasts and delayed communications. The framework provides a scalable foundation for future work on adaptive, data-driven supply chain resilience.