Thesis Title: Marketplace Design for Crowdsourced Delivery
Thesis Committee:
Dr. He Wang (advisor), Industrial and Systems Engineering, Georgia Institute of Technology
Dr. Martin Savelsbergh (advisor), Industrial and Systems Engineering, Georgia Institute of Technology
Dr. Alan Erera, Industrial and Systems Engineering, Georgia Institute of Technology
Dr. Pascal Van Hentenryck, Industrial and Systems Engineering, Georgia Institute of Technology
Dr. Greyson Daugherty, Director of Machine Learning and Operations Research, Roadie, Inc.
Date and Time: April 20th at 10:00am
In-Person Location: Groseclose 226A
Abstract:
The continued rise of e-commerce has shaped the landscape of logistics over the last decade. Customers consistently expect fast and reliable delivery to the tune of same-day and multi-hour service guarantees. This emerging and unique commercial landscape has led to the appearance of third-party logistics (3PL) companies that rely almost exclusively on crowdsourced couriers - crowdsourced delivery platforms. The main challenge of a crowdsourced delivery platform is to meet a service level for their customers (e.g., 95% on-time delivery) by serving dynamically arriving delivery tasks with time windows. The two critical courier management decisions for a platform are how to schedule couriers and how to assign delivery tasks to couriers. These two decisions can be centralized (i.e., decided by the platform) or decentralized (i.e., decided by the couriers). Centralizing these decisions produces a more reliable workforce while decentralizing them may come with cost savings to the platform and allows for more freedom to couriers in deciding when and where to work. Crowdsourced delivery platforms have begun to utilize multiple courier types (i.e., a hybrid system) with the hope of reaping the advantages of each. In this work, we address the challenge of managing crowdsourced delivery platforms that use committed (centralized) and ad-hoc (decentralized) couriers at the strategic, tactical, and operational levels.
In Chapter 2, we consider strategic decisions which are made with a long-term horizon in mind, and, in the context of a crowdsourced delivery platform, focus on the design of the labor force. We address the question of designing the labor supply by investigating the benefit of employing a hybrid fleet of ad-hoc and committed couriers. We model the planning problem of a crowdsourced delivery platform as a fluid model that jointly decides the optimal fleet size for committed couriers and static pricing policy for ad-hoc couriers. Analyzing our models, we answer the questions: under what conditions is a hybrid delivery system expected to outperform either pure system?; and in those cases, what are the main mechanisms responsible for cost savings? We find that committed couriers can be used to remove pressure from the ad-hoc pricing channel when the system is used near its capacity, all while balancing the spatial mismatch between ad-hoc supply and demand.
In Chapter 3, we move on to the tactical problem (e.g., made with short- to mid-term horizons in mind) of scheduling committed couriers under the influence of same-day demand and uncertain ad-hoc courier behavior and arrivals. We model this problem and present a sample average approximation and simulation optimization (SAA-SO) heuristic to solve it. Unlike other scheduling problems, this problem may need to be solved multiple times per day for multiple markets. As such, we propose a prescriptive machine learning method to approximate our SAA-SO heuristic. The main idea of our method is to utilize the SAA-SO method offline to create a set of solutions to a diverse set of problem instances (i.e., demand and ad-hoc courier forecasts), and use a trained machine learning model to generate solutions to new forecasts online. Thus, we bypass the computationally intensive optimization step and prescribe an approximate solution. Our ML method generates solutions that have a cost within .02-1.9% of the SAA-SO method for varying instances and is orders of magnitude faster than the offline method.
Finally, in Chapter 4 we tackle the operational decisions (e.g., those made repeatedly over the course of an operating period). In the context of the hybrid delivery system of the previous two chapters, we investigate the question: How should delivery tasks be allocated between the ad-hoc and committed courier subsystems dynamically? The two natural ways of viewing allocation policies (inspired by queuing theory) are delivery task splitting and pooling policies. We model the real-time operations of a hybrid system as a two-sided queue and find that either a splitting or a pooled policy can outperform the other, depending on the setting. Specifically, split policies perform better in situations with unexpectedly low traffic intensity, while pooled policies perform better in situations with unexpectedly high traffic intensity. In our simulation experiments we find that intelligently designed splitting policies that utilize future information tend to beat pooled policies in most realistic scenarios.