Predictive Control Methods

Block diagram of a model predictive control scheme in a feedback loop with a plant.
Block diagram of a model predictive control scheme in a feedback loop with a plant.

Predictive control methods use finite-horizon model-based predictions to compute an optimal control input. In particular, the minimizing input is computed based on an optimization problem, such that some user-specified cost function is minimized while satisfaction of safety-critical constraints is ensured. Model predictive control (MPC) is an optimization-based control method that generates feedback by repeatedly solving such predictive control problems and applying the initial part of the resulting optimal input sequence.

We are specifically interested in MPC due to its unique ability to guarantee the satisfaction of safety-critical constraints for general nonlinear systems. Our research focuses on extending this safe-by-design property to consider uncertain systems, to account for limited computational resources, and also transfer this safety property to reinforcement learning algorithms.

 

Model uncertainty in predictive control

In practice, the model used for predictive control is typically affected by uncertainty and the identified model is potentially biased. Therefore, the model uncertainty needs to be taken into account when designing a predictive control method in order to retain the safe-by-design property of model predictive control (MPC).

Predictive Safety Filters
To account for model uncertainty, all possible trajectories are contained in a sequence of tubes/sets (black ellipses) around a nominal trajectory (dashed line).

Numerical methods for predictive control

Model predictive control is an advanced optimization-based control strategy that greatly facilitates complex control design tasks involving, e.g., nonlinear and constrained systems, large-scale systems and systems affected by uncertainty. However, its effectiveness and generality comes at the cost of solving an instance of a parametric mathematical program at every sampling time. Especially for systems with fast dynamics or when limited computational resources are available, this poses a challenging task. For this reason, we develop efficient algorithms and software implementations that allow one to speed up the computations and enable applicability of model predictive control to an increasingly large class of systems and control problems. 

 

CPU time for robust MPC with ellipsoidal reachable sets
The figure shows computation times associated with the solution of nominal (blue) optimal control problems and robustified ones with (zoRO - orange) and without (robust - green) the use of a tailored algorithm. Computation times can be reduced by up to 3 orders of magnitude by leveraging tailored numerical methods.

Predictive safety filter

Predictive safety filters enable the modularization of safety and performance in control design. Applications include the safe application of learning-based controllers, for example, imitation learning for autonomous driving applications, which then allows to safely reproduce a human driver behavior based on available data.

Model predictive safety certification
Predictive Safety Filter Example: Diagram (a) shows a possible vehicle trajectory from a safe desired learning control input, which is certified as safe for all future time steps. Diagram (b) shows the resulting vehicle trajectory from an unsafe desired learning controller input, where the vehicle ends up leaving the track. An alternate safe input is then applied by the safety filter, which keeps the vehicle inside the safety constraints.