Direction Prediction Error-Prioritized Experience Replay (DPE-PER) for Adversarial UAV Interception Under Partial Observability

Autonomous interception of unmanned aerial vehicles (UAVs) is a critical capability in modern defense, requiring agents to navigate adversarial maneuvers under environmental uncertainty. While conventional path planning often assumes full observability and static behaviors, this study addresses the complex challenge of intercepting an evasive target within a partially observable Markov decision process (POMDP) framework. We propose a deep Q-network (DQN) framework enhanced by a novel sampling strategy: direction prediction error–prioritized experience replay (DPE-PER). Unlike standard replay mechanisms, DPE-PER utilizes an auxiliary prediction task to prioritize transitions based on direction prediction (DP) error and directional uncertainty. By emphasizing experiences where the agent’s internal model fails to anticipate target maneuvers, DPE-PER focuses learning on the most informative adversarial interactions. Extensive simulations demonstrate that our approach achieves a 48.2% improvement in target interception rates and a 25.4% increase in average return compared to standard experience replay (ER) and prioritized experience replay (PER), highlighting the effectiveness of integrating predictive signals into experience selection for robust UAV autonomy in adversarial environments.