Prof. Dr. Johannes Reuter and Hannes Homburger M. Eng.
Hochschule Konstanz (University of Applied Sciences)
Tuesday, January 11, 2022, 11:00 - 12:30
In the context of controlling complex nonlinear systems, model predictive control is about to become the de facto industrial standard. Recent increase in computational power allows for the use of non-standard methods for solving the underlying optimization problem in real-time. The use of Path Integrals (known e.g. from the solution of stochastic differential equations such as the Focker-Plank equation) therefore has become feasible. A major advantage of this approach is that the cost functional can be defined with fewer restrictions e.g. subj. to smoothness. This allows for a very simple, intuitive and effective way to design complex system behaviour into the cost functional. Moreover, the MPPI algorithm can excellently be parallelized without any inaccuracies or restrictions. Since the approach is model based, MPPI benefits from the combination of physical models with data driven models such as artificial neural networks. Further, because the data-driven part can be learned online, there is a possibility to adapt the controller to time-varying system dynamics. The talk will provide some background on MPPI control and will demonstrate the benefits by practical examples, such as swinging up a pendulum and autonomous docking of a vessel.
Meeting-ID: 627 9173 7415