Explainable Reward Shaping via SHAP for Double-DQN in UAV-Assisted WSNs

Unmanned Aerial Vehicles (UAVs) are increasingly deployed in large-scale Wireless Sensor Networks (WSNs) for applications ranging from extending coverage to environmental monitoring and disaster response. Many machine learning techniques, particularly Reinforcement Learning (RL), have been employed to support efficient multi-UAV operations. However, existing approaches still struggle with effective coordination among multiple UAVs, and often rely on static, manually tuned reward functions that lack adaptability for dynamic environmental changes. Our paper presents a novel decentralized approach for multi-UAV sensor coverage that combines explainable reward shaping with Double Deep Q-Network (Double-DQN) reinforcement learning. We first adopt Double-DQN to allow each UAV to independently learn and act based on local observations while coordinating to maximize area coverage, conserve energy, minimize latency, and avoid collisions. We leverage SHapley Additive exPlanations (SHAP) to assess the impact of individual input features on UAV decisions, providing transparent insights into the rationale behind agent actions. We then propose a SHAP-based adaptive reward weighting mechanism that adjusts reward priorities dynamically based on feature importance. Extensive simulations, conducted under varied system parameters reflecting real-world conditions, demonstrate that our approach significantly improves coverage efficiency, energy utilization and collision avoidance, highlighting the promise of integrating explainability and adaptivity in real-time multi-agent UAV systems.