In the last years, new technologies emerged in the field of mobility. Vehicles are connected, are partially autonomous, and collect vast amounts of data. These developments generate new possibilities particularly for urban logistic service providers. Service providers control a fleet of vehicles to serve customers or to transport goods to both businesses and customers within the city. Business models are manifold and comprise conventional parcel transport or less-than-truckload shipping, but also new models such as same-day delivery and shared mobility. Today, fleet operations are the major cost factor in companies’ supply chains. Furthermore, fleets often exhaust vast amounts of the limited urban street infrastructure causing congestion and emissions. To reduce costs, to avoid unnecessary traffic volumes, and to save resources, an anticipatory and dynamic control of these fleets is required. Due to the new communication and tracking possibilities, a suitable control should allow the real-time adaption of the fleet’s routes to newly revealed information such as new customer order requests or changing traffic conditions. Furthermore, a suitable control should anticipate potential future developments by exploiting collected data.
The challenges lead to the scientific field of stochastic dynamic vehicle routing problems (SDVRPs). In SDVRPs, vehicles are routed to serve customers. Planning is conducted under incomplete information due to uncertainty in, amongst others, travel times, service times, customer demands, or customer requests. Dispatchers are able to dynamically adapt their routing plans based on the revelation of new information. Furthermore, stochastic information on future uncertain developments such as potential future customer requests is available, provided by predictive analytics.
Our research focuses on modeling SDVRPs and on developing anticipatory methods for SDVRPs. Our methods draw on learning mechanisms and simulation from the field of approximate dynamic programming to estimate potential future outcomes in particular problem states. In our research, we use and extend existing methodology to address realistic problem models of high complexity. We also develop new and generic learning mechanisms to exploit problem structures and to allow efficient and effective learning.