We share an ideal inventory placement strategy for designing an effective inbound network of an online retailer. In an online retailer's fulfillment network, demand is ideally fulfilled from the nearest warehouse with inventory and ship capacity before other, further-away warehouses (defined as spillover). The spillover of demand leads to variable (inventory) turn (measured as the ratio of the average daily shipment units to the average on-hand units) for each item at different warehouses. We build a two-stage stochastic model to capture the variable turns of an item at different warehouses in inventory placement and a discrete-event simulation model to assess its impact on fulfillment. We generate two inventory allocation plans under two inventory placement strategies: assuming all warehouses have the same network turn (Constant Turn) for an item or incorporating the variable turns (Variable Turn) at warehouses for an item. Simulation is used to evaluate the fulfillment of the two inventory placement strategies, showing that Variable Turn reduces total fulfillment distance.

To align resourcing decisions in a large-scale supply chain network, Sales and Operations Planning generates optimal aggregate flow and capacity plans given customer demand and vendor supply projections. These plans must strike a balance between incentives of disparate business units. We evaluate common baseline-generation approaches for assessing plan and execution quality against instances of a network model. We show that our choice of methodology may lead to conflicting interpretations of plan quality.

In this work, we consider a broad class of dynamic resource allocation problems, and study the performance of practical algorithms. In particular, we focus on the interplay between the distribution of request types and achievable performance, given the broad set of configurations that can be encountered in practical settings. While prior literature studied either a small number of request types or a continuum of types with no gaps, our work allows for a large class of type distributions. Using initially the prototypical multi-secretary problem to explore fundamental performance limits as a function of type distribution properties, we develop a new algorithmic property “conservativeness with respect to gaps,” that guarantees near-optimal performance. In turn, we introduce a practically-motivated, simulation-based algorithm called RAMS, and establish its near-optimal performance, not only for multi-secretary problems, but also for general dynamic resource allocation problems.

Various Amazon systems optimize shipment-path assignment decisions with capacitated resources. In this process, every shipment is unique, the shipments are assigned to paths, and paths have many capacity constraints. As such, these are very large scale assignment problems, not easily solvable in their canonical form. They consume a lot of memory and they take a lot of time to solve due to their single-node memory and CPU footprint. As Amazon is constantly growing, the challenge is to keep the solution time and memory under control, without compromising the solution quality. In this talk, we present an algorithm to solve these large scale linear assignment problems with Lagrangian decomposition techniques. We discuss the strategies for fast convergence, including problem reformulation and mixed use of cutting planes and subgradient updates.