A Generalized Model Reference Adaptive Controller-Reinforcement Learning Framework for Addressing Modeling Parametric Uncertainty

Recent advancements in reinforcement learning have made it possible to use model-free controllers for various applications. These controllers are trained based on simulated experiences for any given system. The controller's training involves mapping positive and adverse behaviors to the actions chosen by the agent. Through the exploration process, the agent enables the controller to learn the system's performance under various chosen actions and understand the system's behaviors in response to each action.However, a significant limitation is that the controller may not capture parametric variations within the system itself, such as variations in mass in a quadrotor system. To overcome this limitation, an integrated controller setup with a model reference adaptive controller and the reinforcement learning agent is hypothesized. This integrated setup is verified using a water-tank level control system and a parrot mambo quadrotor application.