Master thesis: Fleet Logic Simulations

The goal is to extend the existing simulation approach and use it as a testbed for evaluating alternative fleet rebalancing strategies — comparing not just parameter tweaks to today's approach, but structurally different ways of deciding where vehicles should go.

An open methodological challenge that runs through this space, and that the thesis should engage with explicitly, is how to trust results produced inside a simulator at all:

• What it would take to validate simulated outcomes against real-world behaviour.

• How confident one can be that a strategy which wins inside the simulation would actually win in reality.

Possible directions to explore, not a fixed plan — to be shaped together with the mentors Weronika Sikorska and Christian von Koch:

• Extending the simulator's fidelity in whichever dimension turns out to matter most for comparing strategies fairly (demand modelling, vehicle depletion, field operations capacity).

• Building an evaluation harness that lets two or more rebalancing strategies be compared on equal footing inside the simulated environment.

• Implementing and comparing at least two fundamentally different strategies.

• As a stretch goal: training a reinforcement-learning policy directly against the simulator and evaluating it as one of the compared strategies.

By the end of the thesis, we expect the student to have:

• Extended the simulation engine in a way that's documented and usable beyond the thesis itself.

• Implemented and compared at least two structurally different rebalancing strategies inside it.

• Produced a reasoned position on how (or whether) the simulator's comparisons can be trusted to reflect reality.

• Optionally, a trained RL agent evaluated alongside the other strategies.

• Python, and ideally some exposure to discrete-event simulation or agent-based modelling.

• Operations research, control, or reinforcement learning fundamentals — a plus, not a requirement.

• Working with an open-ended problem and real, noisy operational context rather than a clean benchmark.

• A live, already-in-production problem — not a toy exercise invented for the thesis.

• A working codebase to build on rather than starting from a blank page.

• Close weekly mentorship from the team that owns this problem in production.

• A natural fit for students interested in simulation, operations research or reinforcement learning applied to a genuinely physical, real-world system.