Dynamic methods for matching requests to suppliers in transportation platforms

Abstract

Peer-to-peer transportation platforms dynamically match requests (e.g., a ride, a delivery) to independent suppliers. A central challenge is how to make these matching decisions that can accommodate the different needs of all three stakeholders (platform, suppliers, requests) and in environments where both demand requests and suppliers spontaneously arrive to the platform over time. Further, because suppliers are not employed nor controlled by the platform, the platform cannot be certain that a supplier will accept an offered request. To mitigate this selection uncertainty, this dissertation models and creates new optimization approaches for two strategies that provide more autonomy to suppliers while also providing low request match times for requests and high match rates by the platform.The first strategy is to have the platform offer each supplier a menu of requests to choose from. Such menus need to be created carefully because there is a trade-off between selection probability and duplicate selections. In addition to a complex decision space, supplier selection decisions are vast and have systematic implications impacting the platform’s revenue, other suppliers’ experiences (in the form of duplicate selections), and the request waiting times. Thus, we present a multiple scenario approach, repeatedly sampling potential supplier selections, solving the corresponding two-stage decision problems, and combining the multiple different solutions through a consensus algorithm. Extensive computational results using the Chicago Region as a case study illustrate that our method outperforms a set of benchmark policies. We quantify the value of anticipating supplier selection, offering menus to suppliers, offering requests to multiple suppliers at once, and holistically generating menus with the entire system in mind. Our method leads to more balanced assignments by sacrificing some "easy wins" towards better system performance over time and for all stakeholders involved, including increased revenue for the platform, and decreased match waiting times for suppliers and requests. Moreover, we extend the multiple scenario approach to consider forecasted events such as future supplier and request arrivals and departures and future supplier selection behavior. Computational results using the same case study show that our method with forecast can perform as well as our non-forecast approach, but never better. The second strategy is to have the platform learn suppliers' acceptance thresholds for requests from previous supplier interactions with the platform. The platform recommends single request offers to suppliers who can then either accept or reject these offered requests, and if a supplier is successfully matched and then completes a delivery request, they may re-enter the platform to attempt to be matched with another request. If the platform over-estimates a supplier's threshold, then that supplier may be offered a request that they will reject. Yet, if the platform is too conservative in its learning by under-estimating a supplier's threshold, then the supplier may not be offered requests that they would have otherwise accepted. To balance these trade-offs, we present a method with an objective-based optimization and a rejection-and-acceptance learning policy. Results from a computational design of experiments based on a food delivery platform show that our method performs better than a variety of benchmark methods. We quantify the value of learning and its impact on all three stakeholders (platform, suppliers, requests), and the value of learning from only supplier rejections, only supplier acceptances, and both. Our method yields lower wait times to get matched for both suppliers and requests, an increase in assignments, and it allows suppliers to be slightly pickier in their preferences, while still earning as much as, if not more than, the most agreeable suppliers.