Deep learning-based approximate MPC

We combine ideas from optimal control and deep learning to obtain easy and cheap to evaluate but nearly optimal performing approximate controllers.
The expressive capabilities of deep neural networks allow to closely replicate the behavior of complex and computationally challenging control and decision-making strategies. Since neural networks are composed of simple arithemetic functions, the usage of these complex algorithms on cheap hardware such as micro controllers is enabled and systems with very high sampling rates can be controlled.

In his recent work he proposed validation methods to obtain stochastic and analytic guarantees for the safe deployment of the neural network controllers.

Additionally, by exploiting ideas from reinforcement learning, the behavior of the controllers can be optimized such that bounds on stochastic safety guarantes can be improved or such that an arbitrary performance index is maximized, which might result in a controller that outperforms its optimization-based counterpart.

Coming soon.