Optimization problems in last-mile and same-day delivery systems

Claudia Archetti

Robust, interpretable, and fair offline risk score and policy learning for responsible allocation of scarce resources from observational data

Phebe Vayanos

Last-mile delivery is defined as the last leg of the distribution plan, whereby products are delivered from the last point in the supply chain to the final destination, which could be either the customer’s place or a pickup point where the customer agreed to receive what he/she ordered. Optimizing last-mile delivery operations have become crucial, especially due to the explosion of e-commerce. In fact, e-commerce has completely re-shaped distribution systems, especially in last-mile, generating a huge volume of fragmented and low-volume deliveries and drastically reducing consolidation opportunities. A further element which exacerbated this effect is the pressure for fast deliveries: indeed, e-customers lack the instant gratification from shopping on-site, thus they are extremely demanding with respect to delivery time. Recent statistics show that e-customers are willing to pay more for a faster delivery service. Same-day delivery has nowadays become standard practice in many cities in the US and around Europe. Also, services like meal deliveries require extremely fast deliveries (within one hour) and their popularity has exploded especially after the COVID-19 pandemic. While this represents a huge business opportunity for companies in almost all fields, it also raises dramatic issues in terms of management of the operations related to satisfying customers' orders. From a scientific point of view, distribution problems arising in last-mile delivery systems are variants of routing problems. In this course, some of these challenging routing problems will be presented. We will focus on novel aspects of the context in which these problems arise, like the high degree of dynamism, with an emphasis on modeling aspects and solution methodologies. We will start with an introduction to more classical routing problems and see how we can adapt the knowledge related to those well-established problems (studied for decades already) to the new business environment.

Many of our social and health service systems that allocate resources to satisfy basic needs are severely under-resourced. These include hospitals that provide treatments, social services that offer preventive care, and homeless services that provide shelter and housing. We consider two data-driven approaches for deciding who should get access to these scarce resources and also study the ethical implications of each. The first one relies on learning a vulnerability risk score that predicts the chance of an adverse outcome under no resource and prioritizes those most at risk for the more supportive resources. The second takes an outcome perspective and aims to match people to resources as they become available in a way that achieves desirable long-term societal outcomes. In this series of talks, we study both of these problems and consider several challenges related to 1) the self-reported and observational nature of the data available to help learn these models in the real world, 2) the presence of unobserved confounders, and 3) the occurrence of distribution shifts caused by changes to the administration and wording of the tools used to collect the data. We propose solutions that meld robust and integer optimization, machine learning, and causal inference to address these challenges. We evaluate our approaches on the problem of allocating scarce housing to people experiencing homelessness using data from the homeless management information system. This work is the result of long term partnerships with homeless service providers, community members, and policy-makers in Los Angeles and Missouri.