Stability Guarantee with Reinforcement Learning
[2019 - ongoing] - Not sure
One of the challenges facing the implementation of reinforcement learning algorithms, especially for safety-critical systems, is the lack of formal guarantees on performance. Indeed, from the controls engineer’s perspective, it is necessary to establish a certificate on control-theoretic stability and safety for real-world deployment. Inherently, this requires that the algorithm be reliably convergent for the search of an optimal Reinforcement Learning (RL) controller for a feasible system. Consequently, the goal of this research is to design a RL algorithm that reliably converges, and finds a controller solution, to a feasible continuous state and action space aircraft problem, that prescribes control-theoretic stability with analytical proofs.