Eventos Anais de eventos
COBEM 2023
27th International Congress of Mechanical Engineering
Comparing classic to novel flight control approaches to fixed-wing aircraft: feedback control versus Reinforcement Learning
Submission Author:
Joao Erick Fernandes , RJ
Co-Authors:
Joao Erick Fernandes, Flavio Luiz Cardoso-Ribeiro
Presenter: Joao Erick Fernandes
doi://10.26678/ABCM.COBEM2023.COB2023-0896
Abstract
This work aims to compare a classic control method - namely state feedback - to a Deep Q-Network Reinforcement Learning approach applied to the simple problem of altitude holding under disturbance on a longitudinal regional airplane model. A non-linear model was used for simulating the longitudinal behavior of the airplane after applying disturbances to the Angle of Attack and altitude from equilibrium. The disturbances applied were gradually increased in magnitude to compare the reaction of each controller to the point where it can no longer be considered a small disturbance. The same method is repeated with the introduction of a random model uncertainty. Elevator angle was the sole input used for both controllers. The controllers were compared by time domain criteria, such as damp, overshoot, and error as well as how easily those parameters can be directly controlled. A brief review of the Reinforcement Learning method from the control perspective is also presented. Data generated highlights the advantages of using a Reinforcement Learning control - particularly for such problems where the model is unknown, cannot be precisely derived, or changes over time.
Keywords
Reinforcement Learning, Flight dynamics, Flight control, Control Systems
