Cross-regional, cross-tier and nonlinear disruptions are common in current supply chains. In this paper, an EAIOA algorithm is developed. The algorithm to be described is built upon a closed-loop perception-action-learning: the perception layer fuses multi-source IoT, ERP and other system data so as to derive multimodal state representations; the adaptation layer dynamically adjusts multi-objective weights by quantifying real-time cost, service level, carbon emission fluctuations employing entropy method. The optimization layer represents the supply chain as a graph and uses GNN to capture higher-order relationships among nodes and induce continuous decision in DDPG model. Simulation experiments in a multi-level network with sudden demands, production break downs and transportation delays show that from the standpoint of the total weighted cost, EAIOA performs 12.8%, 9.7%, and 51.2% better than static weights, MADDPG (Multi-Agent Deep Deterministic Policy Gradient) and rule-based benchmarks respectively with the latter exceeding 96%. Weight curves have also retired service priority level in the disturbance period on their own, proving an excellent adaptive feature at a millisecond scale of the algorithm and multi-objective collaboration optimization under complex supply chain environment.
Zheng et al. (2026) studied this question.