Stability and feasibility of neural network-based controllers via output range analysis
in
- Research
Benjamin Karg and Sergio Lucia published their work at the 59th Conference on Decisions and Control on Jeju Island, South Korea.
The main contributions of this work are twofold. First, mixed-integer linear programs are used to guarantee that the neural network controller always provides feasible inputs and stabilizes the linear system. Second, a convex optimization problem is presented, which results in an modification of the neural network controller such that it behaves optimally in the neighborhood of the equilibrium point. This allows to deploy neural network controllers for systems that require an extremely high sampling rate with safety guarantees that perform nearly optimal.
The full paper can be found here.
