OR-Transformer Scales Real-Time Supply Chain Decisions to 1,000 Items.

Modern supply chain management often requires complex coordination of replenishment across thousands of diverse items, facing stochastic demand, varied lead times, and shared ordering costs. Traditional methods like rolling-horizon stochastic mixed-integer linear programs (MILPs) become too slow at this scale, while standard reinforcement learning (RL) struggles with high-dimensional action spaces. Researchers have introduced OR-Transformer, a novel deep reinforcement learning framework specifically tailored for joint replenishment under stochastic demand. It features an item-permutation-equivariant Transformer architecture and uses pathwise-gradient training to navigate inventory dynamics effectively. Evaluations demonstrate that OR-Transformer increasingly surpasses both learning-based and MILP baselines as the number of inventory items grows, handling up to 1,024 items. Crucially, it reduces online decision-making time by over 4 million times compared to MILP solvers, enabling real-time, large-scale deep RL applications in supply chain operations.