Two stage resource allocation decisions in modern distribution

Abstract

In CHAPTER 5, the bilevel framework is used to examine the concept of providing choice and its value when the platform does not have perfect knowledge of drivers’ selections. A full factorial design of experiments combines the developed bilevel optimization model with simulation to study the impact of choice for a specific peer-to-peer resource sharing system – a ride-sharing system. Our integrated recommendation model outperforms existing centralized, stable matching, and decentralized approaches in solving distributed resource allocation problems with ad hoc supplier discretion. The optimal number of choices to provide depends on many factors but we find that the most contributing factors are the degree of uncertainty the platform has about drivers’ selections, the willingness of drivers to participate, and the participants’ origin and destination distributions.

This dissertation develops new analytical and optimization models, as well as accompanying solution techniques for two-stage resource allocation problems found in modern distribution. First, analytical approaches are developed for proactive versus reactive order-fulfillment resource allocation decisions, in which responsiveness is important, but may require additional resources. Next, bilevel optimization approaches are created for resource allocation decisions when the resources are owned by decentralized independent suppliers. A new integrated, two-stage approach accommodates the needs of both demand requests and resource owners.

This is the first work to extend the two-stage multi-item newsvendor model to consider negative marginal shortage costs for a set of SKUs. Closed-form expressions are derived for finding the candidate SKUs, as well as their optimal quantity to handle in a proactive strategy. The impact of activity profile based fulfillment cost functions that vary for different SKUs is analyzed. It is optimal to apply a reactive strategy to only high-demanded SKUs in particular cases, which depends on the degree of required responsiveness and the application’s environment. Two operationally challenging applications are modeled. CHAPTER 2 focuses on sea-based logistics, where fulfilling personalized stochastic requests with little warning requires selective offloading of SKUs in a high dense storage area. A reactive strategy is a direct transfer of cargo to the receiving ship. A proactive strategy pre-stages cargo on the deck, which involves additional labor efforts and double handling of cargo, but can improve responsiveness. Results are illustrated with historical data from a sea-based logistics military application. CHAPTER 3 focuses on a ship-from-store environment, in which brick-and-mortar stores are employed as distribution centers to fulfill on-line demands in addition to serving in-store customers. Resource allocation decisions require trading off whether to retrieve a SKU from the shopping floor in advance or after demand materializes. Operational factors of a ship-from-store environment are identified and mapped to the input parameters of the model to determine how these factors influence the candidate SKUs for a proactive strategy. Our analytical models can aid decision makers with optimal item assignment quantities to proactively store in the backroom.

The second part of this dissertation (i.e., CHAPTER 4 and CHAPTER 5) studies an evolving distribution application. Peer-to-peer resource sharing systems use a centralized platform to match underutilized resource capacities owned by decentralized independent suppliers with demand requests. The platform’s resource allocation decisions are more challenging than traditional centralized allocation decisions because suppliers’ discretion and preferences need to be balanced with systematic resource considerations. We formally define a novel integrated approach, which recasts the platform's role as one providing personalized recommendations of multiple requests to decentralized suppliers. A new hierarchical bilevel optimization framework is created and efficient solution techniques (including exact methods and heuristics) are developed. In the first stage, the platform makes recommendations to decentralized suppliers. In the second stage, these decentralized suppliers have discretion to decide whether to provide access to their resources or not. In CHAPTER 4 our exact reformulation technique is shown to be computationally superior compared to classical solution techniques in bilevel optimization. Also, for the case where all drivers do not necessarily receive recommendations with the same number of ride requests, our heuristic method provides a close to optimal solution in a fraction of time compared to exact solution methods.