|Zeit:||10. Januar 2023|
|Download als iCal:||
Prof. Andrea Iannelli
Institute for Systems Theory and Automatic Control
University of Stuttgart
Tuesday 2023-01-10 4 p.m.
IST Seminar Room 2.255 - Pfaffenwaldring 9 - Campus Stuttgart-Vaihingen
The increase in systems complexity caused by societal challenges and the push to address increasing challenging tasks makes the synthesis of actions to achieve certain closed-loop system’s performance a sequential decision making problem under uncertainty. This motivates us to rethink the standard paradigm in control design of synthesizing the control algorithm offline (e.g. a matrix of transfer functions as in loop-shaping or a static map between measured state and input as in model predictive control).
In this work we will present our on-going work towards framing control of adaptive systems in changing environments as an online learning problem, whereby the decision-maker takes sequential decisions by solving a series of time-varying optimization problems having a-priori only partial knowledge of the cost functions. On the one hand, the online learning viewpoint inherently takes into account the time-varying and uncertain nature of the problem. On the other hand, it considers a different performance metric than the others used in system theory and control, i.e. regret, which measures the accumulated suboptimality with respect to a clairvoyant decision maker.
Motivated by the goal to understand what online learning, traditionally used in game-theoretic or decision-making problems which have no dynamics, can offer in the context of systems theory and control, we analyse two classic control problems, i.e. Iterative Learning Control and control, from this viewpoint. Our findings show that regret characterizes fundamental limitations of non-adaptive algorithms (for the former) and captures the robustness associated with planning for the worst-case (for the latter).
Andrea Iannelli is a tenure-track junior professor in the Institute for Systems Theory and Automatic Control (IST) at the University of Stuttgart (Germany). Andrea's main research interests are at the intersection of control theory, optimization, and learning, with a particular focus on optimization-based control methods, system identification, uncertainty quantification, and sequential decision making problems. He obtained the Bachelor and Master degrees in Aerospace Engineering at the University of Pisa (Italy). In April 2019 he completed his PhD at the University of Bristol (UK), funded by the H2020 project FLEXOP, where he focused on the reconciliation between robust control theory and dynamical systems approaches, with application to uncertain aerospace systems. From May 2019 to September 2022 he was a PostDoctoral researcher in the Automatic Control Laboratory (IfA) at ETH Zürich (Switzerland). During his PostDoc he has been developing and demonstrating theoretical advances in data-driven control theory, optimization-based control, and system identification, with particular emphasis on the use of data to make reliable predictions and decisions.