Learning to Optimize (L2O)

While route planning for small to medium numbers of vehicles has long been reliably solved using methods of classical operations research (OR), problems with more than 100 vehicles as well as complex constraints (relevant in the area of logistics companies and fleets of large mobile care services) pose challenges that normally require very long computing time and can often only be solved approximately. In addition, due to changing conditions (e.g. integration of new customers, longer travel times due to increased traffic, unknown construction sites or accidents), re-planning can often only be carried out in individual cases as required. If a dynamic route planning problem is solved by re-planning, this often results in tour plans that are far from optimal and therefore have a great potential for improvement.

The project aims at the effective and innovative combination of classical optimization methods of OR and techniques of Artificial Intelligence (AI). For the solution of dynamic route planning with many attributes and complex constraints, learning from existing data and experience will be used to improve the quality of the solutions. The core of the project is based on a parallel approach, in which machine learning is used to accelerate classical OR methods and to perform corrections and re-planning in almost real time. On the other hand, OR methods should support AI-based approaches to ensure the quality of the solutions achieved. For this purpose, planning data from various industrial partners will be analyzed, suitable solution methods will be developed on this basis, which then will be continuously adapted and improved.

DIMACS Implementation Challenge:
First place awarded to the BWOR-L2O solver, developped by Christian Ackermann, in the Dynamic Ride Hailing Problem of the Vehicle Routing Challenge: Result page
Announcement on Twitter (02.02.2022): Twitter post

Press releases:
Artikel im NDR vom 08.06.2020:
NDR: Hildesheimer Forscher optimieren Routenplanung
Artikel in der Hildesheimer Allgemeinen Zeitung vom 11.06.2020.:
HAZ: Uni hilft Fahrern auf die Sprünge

Project Sponsor:
BMBF

Partners:
Arbeitsgruppe Betriebswirtschaft und Operations Research - Universität Hildesheim
Spedition Hahne GmbH
SpediFix Logistiksoftware GmbH & Co. KG

Kontakt

Lars Schmidt-Thieme
Jonas Falkner

Publications

• Christian Ackermann, Julia Rieck (2021):
  New Optimization Guidance for Dynamic Dial-a-Ride Problems, in Operations Research Proceedings, Springer.
• Cornelius Rüther, Shabanaz Chamurally, Julia Rieck (2021):
  An a-priori Parameter Selection Approach to Enhance the Performance of Genetic Algorithms Solving Pickup and Delivery Problems, in Operations Research Proceedings, Springer.
• Maik Trott, Niels-Fabian Baur, Marvin Auf der Landwehr, Julia Rieck (2021):
  Evaluating the role of commercial parking bays for urban stakeholders on last-mile deliveries – A consideration of various sustainability aspects, in Journal of Cleaner Production, .
• Cornelius Rüther, Julia Rieck (2020):
  A grouping genetic algorithm for multi depot pickup and delivery problems with time windows and heterogeneous vehicle fleets, in Lecture Notes in Computer Science (LNCS), EvoCOP Proceedings 2020: European Conference on Evolutionary Computation in Combinatorial Optimization, Springer, pp. 148-163.