Vorträge im SoSe 2026

A Lyapunov theoretic approach to stability and robustness of optimal control is well-established for deterministic systems. Considering the rise of learning in control, the close links between optimal control and reinforcement learning (RL), as well as the fact that RL is usually formulated for stochastic systems, the question arises whether similar stability and robustness guarantees can be achieved in a stochastic setting. After exploring the relations between stability, optimal control, RL and dynamic programming in a general setting and how these concepts work in the stochastic case, the main result, which guarantees (robust) semiglobal practical recurrence under stabilizability and detectability assumptions, will be presented.

In this talk, we study exponential output-to-state stability (eOSS) for linear infinite-dimensional systems with bounded output operators. It is shown that exponential detectability guarantees the existence of a coercive eOSS Lyapunov function and therefore eOSS. A counterexample demonstrates that exponential zero-detectability does not imply eOSS, and thus does not imply exponential detectability. Under a state-space decomposition, it is established that zero-detectability implies detectability and leads to equivalent characterisations of eOSS. In addition, a structural sufficient condition for eOSS is provided. The results are illustrated with an example.

Distributed optimization problems comprise a number of subsystems, each with their own objective and local constraints, together with coupling conditions between the subsystems. If these coupling conditions are linear and the overall objective consists of the sum of the subsystem objectives, the overall problem becomes partially separable. Such systems are called tree-sparse or tree-coupled if, in addition, the coupling between subsystems is only enforced along an underlying tree graph. Prominent examples of this class of problems stem from scenario tree approaches for robust optimization. Standard solution methods typically lead to a sequence of highly structured linear subproblems of symmetric saddle-point type. We will show how to design efficient preconditioners for their solution using Krylov subspace methods.