Graph-based multi-agent reinforcement learning for industrial supply chain efficiency optimization

Industrial supply-chain decisions are affected by variable demand, transportation delays, and inventory uncertainty. This study evaluates an integrated Graph-MAPPO framework that combines graph-based state representation with cooperative multi-agent policy learning in a data-informed supply-chain simulation environment. A deterministic mapping of 1,873 development-period DataCo records produced 140 proxy nodes and 2,626 transaction-induced directed links; the subsequent operational trajectories were generated by explicit simulation rules rather than replayed as observed industrial trajectories. Normal, Surge, Disruption, and Mixed demand/lead-time scenarios were evaluated with fully specified distributions and reserved random seeds. For the controlled comparison, all learning-based baselines used 1,000 training episodes, the same cooperative reward, equivalent tuning effort, and common evaluation seeds; a 10-step rolling-horizon MPC baseline was also implemented. Graph-MAPPO obtained a mean episode reward of − 24.2559, a mean backlog of 41.05 units, and a service level of 99.70%. Several baselines achieved higher reward or lower backlog, and rolling-horizon MPC achieved the strongest scalar reward-backlog score but required substantially greater online computation. The ablation results support a contribution from graph-aware representation within the evaluated architecture, but the findings demonstrate a specific coordination-computation trade-off rather than general superiority or validated industrial deployment.