Abstract:

In the home delivery industry, where profit margins are tight and customer service is a top priority, it is important for companies to solve the Vehicle Routing Problem (VRP) efficiently, yet consider real-world uncertainties, to minimize costs and ensure customer satisfaction through on-time deliveries. In this context, increasing the reliability of travel time estimations used while solving the VRP is crucial to obtain more robust planned routes, with less time windows violations, less idle times and consequently less realized costs for the companies. In this talk, we leverage 13 millions GPS observations coming from delivery vans traveling around the north-east US to train a deep learning model to predict travel time durations for each road segment and turn of the north-east US network. The trained model is then used to update road segment and turn durations of the network and its accuracy is validated by comparing the realized duration of 300.000 new travels with the predicted durations computed by solving a shortest path problem on the model-generated network.