Explainable AI for Intelligent Trajectory Design in UAV-Assisted Wireless Networks

Unmanned aerial vehicles (UAVs) serving as flying base stations promise to enhance the efficiency of future wireless systems. Their exceptional flexibility and adaptable deployment remarkably improve system capacity under elevated traffic demands. However, designing the UAV flight trajectory is crucial for maximizing the benefits of UAV-assisted wireless communications. Traditional methods for UAV trajectory design, such as exact and heuristic, are suitable for offline planning. They require precise system information and struggle with online planning and autonomous decision-making. To address these challenges, harnessing explainable artificial intelligence (AI) techniques in UAVs can significantly enhance self-awareness, interpretability, and generalizability, transforming wireless communications. Additionally, equipping UAVs with AI capabilities (AI-enabled UAVs) can enable them to excel in learning high-quality and explainable solutions for sequential decision-making problems such as trajectory design and sum-rate optimization in unpredictable wireless environments. This thesis presents novel explainable AI-driven methods based on active inference (AIn), and explainable reinforcement learning (XRL), to enhance wireless connectivity and optimize data transmission between the UAV and terrestrial users through strategic trajectory design. Using active inference, UAVs develop a comprehensive world model by learning from offline-optimized training examples. This world model, structured as a global dictionary, encapsulates the decision-making grammar of the traveling salesman problem with profits (TSPWP), representing hotspots, local paths, and complete trajectories. Equipped with this world model, UAVs can autonomously evaluate real-time situations, predict the outcomes of actions, and score policies based on expected surprise, enabling adaptive online planning. To ensure the interpretability and robustness of these policies, this thesis further presents an eXplainable reinforcement learning (XRL) methodology. By combining imitation learning (IL), inverse reinforcement learning (IRL), and reinforcement learning (RL), UAVs can analyze expert demonstrations, deduce data-driven reward functions, and employ probabilistic Q-tables to encode state-action correlations. This approach fosters the creation of lucid and adaptable policies that align with performance objectives, enabling UAVs to generalize effectively and operate seamlessly in unobserved environments. The resulting policies exhibit enhanced transparency, adaptability, and reliability, ensuring optimal performance in complex wireless communication networks. Simulation results demonstrate the superiority of the proposed frameworks, offering fast, stable, and reliable solutions with better generalization and adaptability than traditional Q-learning algorithms. Furthermore, preliminary findings highlight the promising potential of these approaches in achieving explainability in AI-based agents.