Dynamic Workforce Acquisition for Crowdsourced Last-Mile Delivery Platforms

Yanzhe (Murray) Lei,Stefanus Jasin,Jingyi Wang,Houtao Deng,Jagannath Putrevu

DOI: https://doi.org/10.2139/ssrn.3532844

2020-01-01

SSRN Electronic Journal

Abstract:We consider a Dynamic Workforce Acquisition (DWA) problem for crowdsourced last-mile delivery platforms that need to match supply (the number of available workers) with demand (the number of requested deliveries). This need arises due to the fact that the initial number of scheduled workers does not always match the number of requested deliveries. If the supply level is much lower than demand, the platform cannot guarantee on-time deliveries for many customers, which is undesirable; if, on the other hand, supply level is much higher than demand, it leads to a low worker utilization, which is also undesirable since it may discourage the freelance workers to continue working for the platform in the future. The question is how to dynamically acquire new workers (e.g., by sending out requests to work extra hours) such that supply level matches demand as closely as possible by the end of the planning horizon (i.e., the targeted service time). This is not a trivial problem. A key challenge here is due to the uncertainties in workers' response (i.e., when they will respond and whether they will respond positively to the request) and the uncertainties in new demand for deliveries at the targeted service time within the planning horizon. We first study a base model where workers respond immediately to the request, there is no new demand arriving within the planning horizon, and the costs of under-staffing and over-staffing are equal. We propose a simple heuristic policy and derive its performance bound. For this setting, we also provide a (close to) matching theoretical lower bound for any feasible non-anticipating policy. These results show that our policy is near-optimal. We also discuss how to modify our first policy for the setting where the costs of under-staffing and over-staffing are not necessarily the same. We show that our modified policy is robust with respect to the magnitude of these costs. Finally, we use the insights from the policies in the base model to construct a heuristic policy for a more general model with uncertainties in both workers' responses and demand arrivals. We derive performance bounds for two intuitive dynamic notification schedules and test their empirical performances using numerical experiments.