MODUS Seminar Program

Weekly talks in the MODUS Seminar: Wednesday, 12:30-14:00, Room S102, FAN-B
Individual talks may be held via zoom or at different times, see the information, below.

Further talks will be announced as the semester progresses. To receive email notifications about upcoming talks, please subscribe to the MODUS elearning course.

Winter Semester 2026/2027

Friday, 9 October 2026, 10:00h, Room S 82 (NW II)
Sébastien Gros
Professor, Department of Engineering Cybernetics, NTNU Trondheim, Norway
Combining Reinforcement Learning and MPC - How to take optimal decisions with wrong models?

Abstract: Reinforcement Learning (RL) can be used to tune MPC schemes towards closed-loop optimality in the real world. That tuning is fundamentally different than using classical fitting approaches to adjust the model underlying MPC. These ideas are supported by a rich theory on how MPC is a universal approximation for Markov Decision Processes, provided that it is viewed in the right way. That theory proposes a bridge between the open-loop predictions of MPC and its closed-loop implementation. It also explains why using classical model fitting is sometimes good enough for achieving near-optimal closed-loop performance with MPC, and why it often falls short of achieving that. It finally opens a theoretical pathway on using (imperfect) simulation models to build optimal policies for the real-world, and on closing the “sim2real gap”. In this talk, we will provide an overview of these recent advances.

28 October 2026, 12:30h
Jonathan Schmalfuß
PhD Student, Chair of Scientific Computing, Universität Bayreuth
Mesh Partitioning with Explicit Communication Constraints

Abstract: Motivated by applications demand that existing tools could not support, we propose a method for enforcing explicit constraints on a partition’s quotient graph. At the coarsest level of multilevel k-way partitioning, we use several integer linear programming models to bound the graph’s maximum degree, constrain connectivity among groups of subdomains, or match a given reference quotient graph. We present models with linear complexity and investigate additional heuristics. Finally, we show that these constraints improve the performance of a shallow water solver.

11 November 2026, 12:30h
Jörg Rambau
Professor und Leiter des Lehrstuhls für Wirtschaftsmathematik, Universität Bayreuth
Grey-Box Attacks to ReLU Deep Neural Networks with Sample-Robust Optimization

Abstract: Deep neural networks (DNNs) can be attacked by adversarial examples, e.g., images that are misclassified because of just a few tweaked pixel values. White-box attacks (WBA) are based on complete knowledge of the DNN's internal parameters (i.e., weights and biases). In this talk, the concept of grey-box attacks (GBA) and a method of their a-priori evaluation is proposed. GBAs are only based on knowledge about the DNN's training method and data. Small standard benchmark images show how robustness concepts based on two metric spaces of DNNs lead to different GBAs with substantially differing success levels. This is joint work with my doctoral student Ronan Richter.

25 November 2026, 12:30h
Alberto De Marchi
Research Assistant, Engineering Mathematics, Universität der Bundeswehr München
Quadratic Optimization with Geometric Constraints: From Theory to Robot Control and Asset Allocation

Abstract: Many problems in optimization, control, and robotics involve convex quadratic costs subject to geometric constraints: sets with structured nonconvexity arising from logical conditions, cardinality restrictions, or physical complementarity relations. Model-based control of rigid and soft robots with frictional contacts can be formulated as quadratic programs with conic complementarity constraints, induced by the Signorini condition, Coulomb friction, and maximum dissipation. Likewise, the classical portfolio optimization problem, extended with minimum buy-in constraints, cardinality costs, or logical coupling conditions, leads to quadratic programs with the same geometric structure.
This talk presents two approaches to this challenging problem class, both built on the augmented Lagrangian framework combined with projections onto such structured nonconvex sets. By exploiting the underlying linear-quadratic structure, the two strategies aim to accelerate the augmented Lagrangian subproblem solves while preserving global convergence guarantees.
References: De Marchi, "A condensing approach for linear-quadratic optimization with geometric constraints" (2026), presented at ECC26
Ménager, Fabre, Bambade, Jallet, De Marchi, Carpentier, "Frictional contact-implicit inverse dynamics" (2025), submitted to IEEE T-RO.

9 December 2026, 12:30h
Boris Wembe Moafo
Research Associate, Numerical Mathematics and Control, Universität Paderborn
Exponential trim turnpikes for optimal control systems with symmetry: From continuous to discrete time

Abstract: Optimal control problems on long time horizons often exhibit the turnpike property: optimal trajectories spend most of the time close to a steady state, independently of the initial and final conditions. For systems with symmetry, such as vehicles, satellites or other mechanical systems, steady states are replaced by trims, i.e. motions that proceed at constant velocity along the symmetry directions. In this talk, we present recent results on the exponential trim turnpike property in continuous and discrete time. Splitting the state into shape variables and group variables (the coordinates along the symmetry), we show that the shape variables converge exponentially to an equilibrium, while the group variables evolve with constant velocity. We then present a geometric framework for nonlinear discrete-time problems with symmetry, based on discrete connections and discrete abelian Routh reduction, in which the same splitting is available and an analogous exponential trim turnpike holds. Finally, we show that the trim turnpike property is preserved under time discretization of the continuous problem, so that the discrete optimal solutions inherit the exponential structure of the continuous ones.
The talk is based on joint work with Sofya Maslovskaya, and Sina Ober-Blöbaum

13 January 2027, 12:30h
Tomáš Roubal
Department of Decision-Making Theory, Institute of Information Theory and Automation, The Czech Academy of Sciences, Prague, Czech Republic
Title will be announced later

3 February 2027, 12:30h
Niklas Kruse
Scientific Computing, University of Rostock
Steering Particles into Coherent Sets of Unknown Vector Fields

Abstract: A data-driven control scheme for steering tracer particles into coherent sets of unknown vector fields is presented. The control algorithm is based on coordinates obtained from a spacetime diffusion map that are adapted to the coherent structures of the unknown vector field. These coordinates are constructed from sampled trajectories of tracer particles initialized at different spatial locations. The coherence-adapted coordinates are then used to define a feedback law that steers particles into a desired coherent set of the uncontrolled vector field. The control method is demonstrated on a numerical example of a double gyre.

Further talks may be announced in the next weeks.