My journey at KEDGE Business School

This post has been brought to you by the “How's this for a work review?” department.

Can mathematical optimisation bridge the gap between last-mile logistics and sustainability? In two months, my contract with KEDGE Business School will be ending, finalising a postdoctoral research of one year in the ConStanCE project. The acronym means, in French, “Contenants Standards pour une Cyclologistique Efficace”.

The subject of the project is the encapsulation of parcels in medium-sized containers for a more efficient modal shift between traditional last-mile logistics and cyclologistics.

Prior to this postdoc, I had never worked in the domain of the supply chain.

Chapter 1: The Discovery

The discovery was quite the challenge, considering I had no real experience with supply chain problems, my main experiences were on more academic problems, and I still had my Ph.D. defense to prepare.

Nonetheless, I think my transition was quite smooth. I quickly learned what the research was about and integrated well in this multiactor project.

The team, CESIT, was nice, and the school building felt really modern compared to the old buildings of the University of Bordeaux.

Chapter 2: The Work

The work was the most difficult part. We were interested in computing, for each studied city, the optimal container capacity, measured in number of parcels, and the corresponding optimal territory design.

Mathematical models

Difficulties concerning this problem included the huge number of clients in the partner LSP's historical data, and thus how to model routing decisions, which are inherently operational decision, into a strategic decision model. We could either use routing algorithms from the literature and wait two millenia for the solution, or approximate the traveling costs via continuous approximations, methods based on demand density instead of accurate customer locations.

We chose the second solution, which gave us a quite small model containing only the container capacity selection constraints, the territory design assignment and contiguity constraints, and some other constraints related to the computation of the objective function.

However, this doesn't mean the problem was easy to solve. In fact, the continuous approximations and some other components of the objective function were non-linear and non-convex / non-concave. At first, I designed some ad-hoc primal and dual heuristic algorithms for the problem to obtain an initial feasible solution and an optimistic value for the objective function.

My objective was to use those heuristic algorithms to warm start exact solvers, and that's were the struggle started. To be honest, SCIP is bad. It crashed on even the simplest instances, and struggle to progress even with a time limit of one hour.

At the ROADEF conference in February 2026, Artelys told me to try Artelys Knitro, their in-house solver for non-linear programming. While it avoided crashes, results were only slightly better.

The breakthrough came when I managed to implement the model using Gurobi, something I had previously failed because, for some reason, it doesn't allow square roots in the objective but it allows a variable v and a constraint v = equation with square roots. Gurobi was able to greatly improve the initial solution in a minute. However, from the first minute to the one hour time limit, it didn't do anything more. Aside from the simplest instances, it never improved the dual bound at all.

Note: A Gurobi webinar highlighted that fixing an initial dual bound via a obj >= z constraint disrupts the variable selection algorithms. Although we didn't really confirm this was the reason, we confirmed that the added constraints did not help the solver.

Cartography tools

I also made a lot of cartography tools using MapBox, but I'm not sure I can display images about those, so I will leave this section for private discussions.

Synthesis and Reports

Finally, one of my favorite parts was writing the feedback reports to the ADEME, the French agency for ecological transition, because this part felt like the conclusion of the different blocks of tasks of the project. So I wrote, a lot, about different subjects:

All of those documents will be repackaged into a final report publicly available at the end of 2026. I will update this post once it's online!

Chapter 3: The Future

At the end of this project, I believed less in the potential of cyclologistics. While they offer clea environmental and social benefits in dense urban areas, their scalability is constrained by cargo bike capacities, human operational cycles, and geographic sparsity.

In contrast, I now believe deliveries by autonomous vehicles during nighttime is the better solution, especially with the rising numbers of parcel lockers, which solve the problem of customers being present for the delivery.

Although we already have very promising leads on autonomous vehicles of different sizes for this, the main bottlenecks are regulatory frameworks for sidewalks and multi-agent automatic synchronisation tools.

Key takeaway: I’ll be launching my independent research & optimization consultancy in October 2026. I’m particularly focused on autonomous last-mile logistics, network design, and large-scale mathematical programming, but remain open to projects in sustainable supply chains, operations research, or academic collaborations. Reach out at luismarques@luismarques.me.

Bonus: During this project I presented at two different conferences, ROADEF 2026 and IPIC 2026. I don't have much to say about these, but if you're interested, I can provide you with the submitted abstracts and slides on demand.

Thank you for reading. Reach me for further discussion and updates!