Daniel Schermer

• This thesis addresses several challenges associated with introducing autonomous auxiliary vehicles, e.g., drones or robots, into a logistics system from the perspective of Operations Research. To this end, optimization models are formalized that enable the assessment of potential benefits of integrating such vehicles into last-mile delivery. As the resulting models are computationally challenging, the thesis continuously refines the formulations and develops appropriate algorithms that are capable of producing high-quality solutions with reasonable computational effort. This facilitates effective problem-solving and aids in shaping the design of future delivery fleets, laying the groundwork for future decision support systems. In Operations Research, Mixed-Integer Programming solvers play a pivotal role. As is evident by this thesis, this concerns solving Mixed-Integer Linear Programming formulations directly or integrating parts of them in matheuristic frameworks. This raises the fundamental question if the performance of such solvers can be enhanced in any meaningful way, by adjusting their default algorithmic behavior on a per-instance basis. We trace the roots of this problem to the Algorithm Selection Problem and develop a novel methodology that leverages Machine Learning to formalize a prescriptive optimization problem. By using a tailored Branch & Bound approach, this methodology enables us to effectively compute the (predictably) optimal configuration of a Mixed-Integer Programming solver on a per-instance basis with minimal computational overhead. The potential impact extends beyond specific problem domains, fostering a more comprehensive and synergistic approach to decision-making and optimization.