Farias De Oliveira Nobrega, Luciana2024-10-242024-10-242024-10-24http://hdl.handle.net/10393/49787https://doi.org/10.20381/ruor-30639The advent of next-generation wireless networks has necessitated efficient and safe offloading of computational tasks to meet the demands of users for low latency and high reliability. This study focuses on task offloading with Machine Learning (ML) solutions in Unmanned Aerial Vehicle (UAV)-aided Multi-Access Edge Computing (MEC) systems, specifically applied to smart agriculture. The system involves closed-loop communication where sensors collect data, offload it to processors, and return commands to actuators. Despite efforts to minimize computational task waiting and processing times, few studies address these issues alongside both information delay and freshness, particularly in UAV systems. This thesis aims to jointly optimize data freshness and computational Turnaround Time (TAT) in UAV-aided MEC systems by proposing a UAV trajectory and UAV-MEC offloading strategy. We explore the trade-off between the Age of Loop (AoL) and computational TAT in dynamic network conditions, varying computational resources, data arrival rates, and packet workloads. We employ a hierarchical-based Deep Reinforcement Learning (DRL) approach to simultaneously evaluate and optimize the freshness of information and TAT. The solution decouples the original problem into two sub-problems, each managed by a distinct Markov Decision Process (MDP). This approach allows for the independent optimization of UAV trajectories and UAV-MEC offloading strategies, providing real-time, specific information about the environment to enhance decision-making. Our findings demonstrate that decoupling the problem and using a hierarchical-based DRL solution significantly improves performance. The hierarchical approach allows the agent to focus on learning optimal policies for trajectory and offloading decisions individually. Compared to the traditional, non-hierarchical DRL solution, it demonstrated excellent performance in unseen scenarios (i.e., those not included in the training process), highlighting its superior adaptability, stability and efficiency across various systems. In conclusion, the hierarchical DRL approach provides a robust and stable solution for optimizing UAV trajectories and MEC offloading in next-generation wireless networks. This strategy enables efficient use of computational resources while minimizing AoL and TAT, particularly in dynamic and resource-constrained environments. The results underscore the importance of hierarchical learning for real-time, specific environmental information, enhancing decision-making and performance in UAV-aided MEC systems.enAttribution-NonCommercial-NoDerivatives 4.0 Internationalhttp://creativecommons.org/licenses/by-nc-nd/4.0/Smart agricultureInternet of Things (IoT)Hierarchical Reinforcement Learning (HRL)information freshnessUAV-assisted networksThesis