Prof. Dr. Michael Römer

Our research sits at the intersection of Operations Research and Machine Learning. We develop decision support methods for planning problems in which uncertainty matters. A common thread is the use of structured models, such as networks and decision diagrams, that are expressive enough for real-world problems and at the same time amenable to learning.
Many planning decisions have to be made before demand, arrivals or durations are known. We work on two-stage, robust and data-driven approaches, on quantifying the value of information, and on decision-focused learning, where predictive models are trained with respect to the decisions they inform.
Hospitals have to plan beds, staff and treatments while patient arrivals and lengths of stay are uncertain. With hospital partners and colleagues from medicine, we develop simulation-, learning- and optimization-based decision support: forecasting and managing intensive care occupancy, scheduling ICU admissions and clinical pathways, and rostering staff.
State-expanded networks encode complex rules, for example for shifts and rosters, in the structure of a graph, so that the remaining optimization model stays simple. Decision diagrams provide relaxations and bounds for discrete optimization. We develop such models and use machine learning to make them smaller and their bounds stronger.
Formulating an optimization model is often the bottleneck for using optimization in practice. In the project OptimizationChat we investigate how large language models can help to formalize and solve industrial planning problems through natural-language interaction.
Environmental decisions, such as where to monitor and what to protect, are made under substantial uncertainty and with limited budgets. With environmental economists and marine biodiversity researchers, we study the value of information in environmental monitoring and decision-focused learning for planning marine protected areas.
Many projects start from operational problems in logistics and retail: service network design, inventory routing, e-grocery fulfilment, truck loading and large-scale supply networks. Logistics also extends into space: our time-space network approach won the "Delivery Scheduling" category of the ESA/GECCO Space Optimisation Competition 2022.