A two-stage data-driven metaheuristic to predict last-mile delivery route sequences

Abstract

On delivery operations, vehicle routing problems regularly focus on optimizing route-solution quality. However, the real-life execution of routes often differs from the computed routing plan. Therefore, the routing algorithms should use information from historic routes and drivers’ decisions to make the planned route as similar as possible to the actual executed route. We study the problem of predicting route sequences that drivers follow in last-mile parcel delivery in the context of the 2021 Amazon Last-Mile Routing Research Challenge. We propose extracting information from historic routes using a probability estimation method that learns transition probabilities between clusters or zones, coupled with a simple yet effective two-stage Greedy Randomized Adaptive Search Procedure (GRASP) to incorporate the information from historical route data into route planning. Our GRASP includes components in the objective function that consider decisions observed on the historical route sequences supplied by Amazon. We evaluate the predicted route sequences using the score similarity metric provided in the Amazon challenge. The results show that our method obtains a similarity score of 0.0328, which is comparable to the scores achieved by the three winning teams of the Amazon challenge. The proposed method successfully uses historical routes to predict route sequences on a last-mile delivery setting.