Friday 18 January 2019 13.00 – 16.00, The Technical University of Denmark, Building 101, room S10.
Supervisor: Associate Professor John Bagterp Jørgensen, DTU Compute
Co-supervisor: Professor Henrik Madsen, DTU Compute
Co-supervisor: Associate Professor Niels Kjølstad Poulsen, DTU Compute
Co-supervisor: Associate Professor Bjarne Poulsen, DTU Compute
Examiners: Professor Michael Pedersen, DTU Compute
Professor Carsten W. Scherer, University of Stuttgart, Germany
Associate Professor Mark Cannon, University of Oxford, United Kingdom
Moderator: Associate Professor Martin Skovgaard Andersen, DTU Compute
Summary:
In daily life we are
surrounded by an increasing number of devices that each involve some element of automatic
control.
Drones and self-driving cars are just the most recent examples.
In case a model is available for the dynamics of the system to be controlled, an optimization-based predictive
control technique may well be an attractive option. However, the faster the systems dynamics, the more stringent
requirements it puts on the execution of the optimization. This thesis offers an efficient implementation of an
optimization problem occurring in the context of linear stochastic systems subject to constraints.
Modelling of the linear system is done in continuous time and comprises both a deterministic and stochastic
model part. To obtain a parsimonious parametrization, transfer functions are used. Also the cost function of the
control problem is formulated in continuous time. This thesis analyzes the behaviour of the corresponding
discretized control problems for increasingly fast sampling rate. Convergence theorems are proved for the case
of uniform sampling with sampling time tending to zero.
Often a single deterministic model is not sufficient to capture the behavior of a system. The description may then
be augmented with a stochastic part quantifying the degree of confidence in the deterministic model part. The
presentation bridges the gap between deterministic and stochastic descriptions for sufficiently regular linear
time-variant systems on state space form. It is shown how a description in terms of stochastic differential
equations (SDE) results if one requires the state equations to hold in the sense of distributions. In particular it
becomes clear why the stochastic integral appearing in the SDE must be defined the way it is.
READ MORE about this thesis in DTU Orbit.