EURO Meets NeurIPS 2022
Vehicle Routing Competition

This competition considers the vehicle routing problem with time windows (VRPTW) as well as a dynamic variant in which new orders arrive during a day. All submitted solvers compete on both problem variants, which is facilitated using baseline strategies provided.

The VRPTW problem considered in this competition is the same as in the DIMACS VRPTW challenge except that we use real-world instances provided by ORTEC, with an explicit duration matrix providing (non-euclidean) real world road driving times between customers. The number of customers is between 200 and 1000. Each customer has a coordinate, a demand, a time window and a service duration, where service must start within the time window. A vehicle can wait when it arrives too early at a customer, but cannot be late. The total demand of customers served by a vehicle should not exceed the given vehicle capacity. Following the DIMACS convention, the objective is to minimize total distance, as measured by pure driving duration (thus excluding waiting time). There is no practical constraint or objective regarding the number of vehicles.

For the dynamic variant of the problem, requests arrive at different 1-hour epochs during the day. The solver, or agent, can choose which requests to dispatch in the current epoch, for which it must create feasible routes, and which ones to postpone to consolidate with new requests that may arrive later. Some requests are must-go requests that must be dispatched during the current epoch as delaying them will render their time window infeasible. In the last epoch, all request are must-go. An environment will be provided that samples the locations, demands, service times and time windows of requests uniformly from the data of a static VRPTW instance. For the dynamic variant, the static instance thus provides a 'customer base' from which requests may arrive, and this information can be used by the solver. The solver should be able to interact with the environment through a controller script, which will be used for final evaluation of the solver.

Solvers should provide solutions (per epoch) within strict time limits and are allowed to use a single CPU core, as well as a single GPU. Details will be provided at the start of the competition on July 1st, when a quickstart repository will be released. This repository contains a set of instances, an implementation of the environment and a baseline, which can be used by participants to develop their solver.

The total prize money is more than €5000:

The detailed rules of the competition, including a detailed description of the problem, provided environment and baselines can be found here (version August 2nd).

The top 10 on this leaderboard at October 31st will be evaluated on the final dataset. Each submission gets two ranks: one based on static performance and one based on dynamic performance. The submissions are ranked by the average of these two ranks, with the average cost as tie-breaker.

This competition is a joint effort by the organizers of previous vehicle routing competitions: the VeRoLog 2017 and VeRoLog 2019 solver challenges, the 2021 AI for TSP Competition and the 2021/2022 DIMACS Vehicle Routing Implementation Challenge.

• Wouter Kool (ORTEC)
• Laurens Bliek (Eindhoven University of Technology)
• Danilo Numeroso (Università di Pisa)
• Robbert Reijnen (Eindhoven University of Technology)
• Reza Refaei Afshar (Eindhoven University of Technology)
• Yingqian Zhang (Eindhoven University of Technology)
• Tom Catshoek (Delft University of Technology)
• Kevin Tierney (Bielefeld University)
• Eduardo Uchoa (Universidade Federal Fluminense)
• Thibaut Vidal (Polytechnique Montréal)
• Joaquim Gromicho (ORTEC)