Plenary: Planning Transportation Capacity at Amazon
By Violet Chen
On Tuesday morning, a room full of participants joined Dr. Russell Allgor, Chief Scientist of WW Ops and AMZL at Amazon, to learn about how Amazon designs and operates its logistics network. Widely recognized for fast delivery and extensive product offerings, Amazon’s success is indispensable from an efficient, dynamic, and robust logistics network. In his talk, Dr. Allgor discussed technical and organizational challenges that Amazon encounters in network design, introduced the current solutions addressing these challenges, and highlighted some opportunities and challenges that remain.
Dr. Allgor started with an overview of the fulfillment and delivery operations at Amazon. A couple of services were featured; for instance, Prime customers enjoy a broad range of free delivery options: free two-day and one-day delivery from coast to coast, free same day delivery in select locations, and free two-hour delivery with Prime Now on daily essentials. Regarding fulfillment and delivery, Amazon has a three-fold mission: low cost, fast speed, and wide selection. While attaining two of these goals is not so hard, he noted that achieving all three is much more challenging. To set the stage for discussing potential solutions, Dr. Allgor introduced the infrastructure of the Amazon logistics network. The network consists of facilities including fulfillment centers (FC), sortation hubs, Prime Now hubs, delivery depots, and transportation vehicles (trailers, vans, planes), which together serve for storage, shipping, and delivery.
The transportation network design should primarily support decisions in three aspects: what infrastructure to install, how to leverage the installed capacity, and how to schedule freight to increase delivery speed. Dr. Allgor expanded on two specific design topics. The first topic is what connections to enable and how much capacity is required. He explained two candidate network models: a point-to-point method where each pair of FC and delivery depot has a designated truck for transportation, and a hub-to-spoke method where FCs and delivery depots are connected through central sortation hubs. Formulating a correct optimization model is typically not the main challenge in deciding optimal connections. Dr. Allgor pointed out several more difficult challenges, such as, how to design efficient metaheuristic to solve large models, and how to implement a deterministic model for execution systems operating in an online fashion.
The second topic discussed in detail was how to model and optimize time in planning. Dr. Allgor listed some examples of decisions to be made within this aspect: deciding sortation cycle at sortation hub, scheduling trucks at different delivery stages, and choosing operation methods for loading and unloading. The research team at Amazon developed a resource task network (RTN)-based approach to represent these planning decisions. Connecting their problem with RTN allowed Amazon to leverage the rich knowledge on RTN, which the chemical process engineering field has been developing for over two decades. With the RTN-based representation, they developed a mixed integer program formulation as primary framework for optimizing time-related decisions.
Dr. Allgor summarized his talk by emphasizing the importance of scalability and flexibility in practical modeling applications. An effective model should be maintainable and generalizable in the long run, and be robust to human errors.