Martin Savelsbergh and Marlin Ulmer were awarded one of the TSL/VeRoLog Cross-Region Collaboration Grants in 2018. The grant funded part of Marlin’s stay at Georgia Tech from October 2018 to March 2019. During that time, Martin and Marlin worked on Workforce Scheduling in the Era of Crowdsourced Delivery. The work was presented at last year’s TSL Workshop in Vienna and is now published in Transportation Science. During that stay, they also prepared Marlin’s (now successful) proposal for a DFG Emmy Noether Grant which will enable long-term collaborations and repeated visits of Marlin and his PhD-students to Martin and colleagues at ISyE at Georgia Tech.
The research project’s goal was to determine driver schedules when both demand and driver resources are uncertain. Such settings occur more and more often due to the gig economy and crowdsourcing, e.g., crowdsourced same-day delivery, restaurant meal delivery, and ridesharing. In such settings, customers request service spontaneously and expect service shortly after their request. Also, the times and the number of crowdsourced drivers entering the system, performing services, and leaving the system is uncertain. This leads to situations in which service promises cannot be met because there are too few drivers in the system. The result is high customer dissatisfaction. As a countermeasure, companies start to use scheduled drivers to complement the crowdsourced workforce. Determining the shifts for scheduled drivers is challenging because of the uncertain demand and crowdsourced delivery capacity as well as temporal and spatial consolidation effects, e.g., the number of drivers needed for service does not increase linearly with the number of requested services and customer demand is not served instantly at the time of request, but sometime later.
To determined cost-effective shifts for scheduled drivers, we combined continuous approximation (CA) with value function approximation (VFA). The continuous approximation captures the spatial consolidation effects by mapping the number of requests to the number of drivers required for high-quality service. The value function approximation captures the temporal consolidation effects by evaluating the driver scheduled in one hour with respect to the drivers needed in the following hours. The VFA, especially, is interesting for two reasons. First, it operates on a constrained Markov decision process because while a solution should be low-cost, it should also satisfy a service-level constraint. Second, we were able to prove that our combination of CA and VFA finds optimal solutions while none of our benchmark heuristics did. We used our policy to analyze schedules for different business models, amongst others different delivery promises, from very fast service as required in passenger transportation or restaurant meal delivery to less demanding service promises more common in “conventional” same-day delivery. We were able to show that a scheduled driver with the right shift can replace up to 6 crowdsourced drivers. We could also show that with less demanding service promises, shifts for scheduled drivers were usually in the second half of the day, when a lot of requests have already accumulated.
We would like to thank the TSL Society for supporting this research and providing Marlin with the opportunity to visit Atlanta. In addition to providing a great research experience, Marlin also had a chance to experience life in the US, from handing out candy during Halloween to the craziness of Valentine’s day and to the Super Bowl taking place in Atlanta (unfortunately without the Packers). Because Marlin received the Collaboration Grant for the second time and was told by Daniele Vigo that he will not get it a third time, he now encourages everyone else to apply – and hope that travel can happen again soon. Future grant winners will be announced at the INFORMS Annual Meeting.