Amazon Tackles Uncertainty in Its Middle-Mile Logistics Network

Amazon's middle-mile logistics network is the segment connecting fulfillment centers and sortation centers, underpinning its same-day and next-day delivery promises. This network involves tens of millions of products, hundreds of facilities, limited operating hours, and countless routes that must each be decided open or closed, making it a mixed-integer optimization problem that's hard to solve even under deterministic conditions alone. With just 300 yes/no decisions, the number of possible combinations exceeds the number of atoms in the observable universe, and even a single planning problem can be difficult to parallelize, taking hours to solve.

Uncertainty shows up as day-to-day fluctuations in demand and transit times, as well as structural shocks like heavy snowstorms, power outages, and facility disruptions. Hardening every scenario one by one doesn't scale, so instead of pursuing perfect forecasts, Amazon aims for optionality. Even accounting for demand variability alone, the potential savings are around 0.5%, but Amazon views even small differences as translating into real customer experience.

To this end, the network design tool narrows down candidate routes and identifies consolidation points where freight from multiple origins can be combined onto a single truck.

• Operating hours are validated not as 24-hour blocks but using pre-computed 15-minute timing bounds for each candidate route, aligning them with actual feasible departure and arrival times.
• Risk-aware design combines hundreds of synthetic origin-destination scenarios generated via Monte Carlo with a graph attention network.
• The model uses a site graph, which captures facilities and route/geographic proximity, together with an origin-destination graph, which treats each origin-destination pair as a node, to learn spatial correlations.

This approach makes it possible to distinguish between a design that is efficient on average but fragile under shocks, and one that is slightly more expensive but more resilient. By validating the network across hundreds of possible futures, Amazon is building a middle-mile logistics network that can keep its delivery promises during peak season, actual disaster situations, and even unprecedented circumstances such as opening a new facility or a regional storm.