Akademik Veri Yönetim Sistemi
AIAA Scitech Forum, 2020, Florida, Amerika Birleşik Devletleri, 6 - 10 Ocak 2020, cilt.1 PartF, (Tam Metin Bildiri)
Achieving autonomy for aerospace vehicles control has been an active area of research in the past; especially when the design of vehicles is becoming more challenging for achieving optimum performance. Usually, devised control strategies for these vehicles employ proportional integral (PI) or proportional integral derivative (PID) controllers using feedback loops. Though, these controllers have performed quite well with stable environments, however more intelligent and active flight control system is required, when dealing with unknown or harsh domains. Under the bigger umbrella of Machine Learning (ML), Reinforcement Learning (RL) has started to address these limitations of the conventional controllers and emerges as the most active, conceptually prudent and best suited machine learning category for autonomous control; which in recent years, has increasingly found use in aerospace control applications for platforms like aircraft, missile trajectory control, fixed wing UAVs, Quadcopters and so on. Additionally, RL coupled with neural nets has emerged as a robust methodology in solving complex domain control problems with continuous action and state spaces, thus, unleashing the hidden power of RL and outperforming both linear and orthodox nonlinear control strategies, which have their own inherent limitations. In the current research, we demonstrate the ability of the proposed RL agent or controller for an experimental glide vehicle to learn gliding strategies and to control the vehicle’s descent, thus eventually optimizing its glide range. At first, we model the experimental design vehicle using a 6-DoF model registering the translational and rotational response of aerial platform. Later, based on extensive literature review, two unique RL algorithms are developed namely ‘Modified Model-Free Dynamic Programming (MMDP) and Deep Deterministic Policy Gradient (DDPG)’; in relation with the current problem, same are applied and their results are compared and analyzed highlighting the success of RL over the performance of control policies obtained by classical approaches. Results are also verified by 6-DoF simulation of the experimental design vehicle.