Search

In ride-sharing systems, platform providers aim to distribute the drivers in the city to meet current and potential future demand and to avoid service cancellations. Ensuring such distribution is particularly challenging in the case of a crowdsourced fleet, as drivers are not centrally controlled but are free to decide where to reposition when idle. Thus, providers look for alternative ways to ensure a vehicle distribution that benefits users, drivers, and the provider. We propose an intuitive mean to improve idle ride-sharing vehicles' repositioning: repositioning heatmaps. These heatmaps highlight driver-specific earning opportunities approximated based on the expected future demand, current and expected future fleet distribution, and the location of the specific driver. Based on the heatmaps, drivers make decentralized yet better-informed repositioning decisions. As our heatmap policy changes the driver distribution in the future, we propose an adaptive learning algorithm for designing our heatmaps in large-scale ride-sharing systems. We simulate the system and generate heatmaps based on the previously learned policy in every iteration. We then update the policy based on the simulation's outcome and use it in the next iteration. We test our heatmap design in a comprehensive case study on New York ride-sharing data. We show that carefully designed heatmaps reduce service cancellations and therefore, revenue loss for the platform and drivers significantly while leading to a better service level for the users and to a fairer treatment of drivers. History: This paper has been accepted for the Transportation Science Special Issue on Machine Learning Methods and Applications in Large-Scale Route Planning Problems. Funding: This research is funded by the German Research Foundation (Deutsche Forschungsgemeinschaft) [Grant 494812908]. M. W. Ulmer's work is funded by the German Research Foundation Emmy Noether Programme [Grant 444657906]. Supplemental Material: The online appendix is available at https://doi.org/10.1287/trsc.2023.1202. [ABSTRACT FROM AUTHOR]

In recent years, innovative ride‐sharing systems have gained significant attention. In such systems, dynamic fleet management covers demand and fulfillment control to determine which stochastically incoming requests are to be satisfied and how vehicle resources are utilized for their fulfillment, respectively. Demand and fulfillment control can be implemented ranging from straightforward myopic to more sophisticated anticipatory. In this paper, our aim is twofold: (1) we want to classify how policies implement demand and fulfillment control in the related literature on dynamic fleet management; (2) we want to explore the effectiveness of demand and fulfillment control under varying conditions in order to identify benefits and risks for ride‐sharing systems. To this end, we define policies that differ in the optimization of demand and/or fulfillment control through the exploitation of either confirmed or complete information. Our experimental results demonstrate that demand and fulfillment control affect the performance and service quality of ride‐sharing systems quite differently. [ABSTRACT FROM AUTHOR]