Autonomous control has become more reachable with recent advancements and new techniques in machine learning and is getting more and more popular in every field including flight control. The autonomous flight that is attained by using machine learning may become a crucial feature for an aircraft because it provides to get rid of external involvement in control and it has the potential to excel in human skills. In this work, it is desired to attain autonomous control for a case that demands complex dynamics like heading maneuvers of a fixed-wing aircraft. Especially deep reinforcement learning (DRL) which is a machine learning technique that brings together the power of reinforcement learning (RL) and deep learning (DL) stands out for autonomous flight control at the human level. It can be shown that the state-of-art DRL algorithms can provide a solution for the autonomous heading control of a fixed-wing aircraft. Therefore, in this work, a DRL algorithm is devised by combining DQN and Actor-Critic algorithms that are prominent methods in DRL. This new algorithm meets the requirement of dealing with continuous state and action spaces which is important for robot control. The algorithm has been tested for a simplified fixed-wing aircraft model to see its usefulness and whether it can be used for more complex tasks and dynamics. The promising results have shown that the algorithm can be enhanced to deal with also other flight tasks and it can offer the solution to complex real-world problems. (This study was produced from the master's thesis of the first ranked author.)
Anahtar Kelimeler: Deep Reinforcement Learning, Autonomous Control, Heading Control, Fixed-wing Aircraft
|
