Explainable Autonomous Maneuver-Penetration Strategy for UAVs via Continuous-Action Learning Automata

Modern air defense systems are becoming faster, more complex, and agile, rendering static evasive maneuvers ineffective for high-speed or hypersonic UAVs. This study proposes a lightweight, explainable Continuous-Action Learning Automata (CARLA) framework as an alternative to black-box DRL policies, reformulating UAV penetration as a probabilistic decision problem in human-interpretable maneuver primitives. Six primitives - straight acceleration, U-turn, S-turn, square wave, barrel roll, and inclined plane - are discretized with trigger range and commanded overload into a 3D action grid. Offline training simulates thousands of random 3-DOF engagement scenarios, evaluates sampled actions via Monte Carlo, and updates joint probability density using decaying Gaussian kernels for exploration-to-exploitation shift, converging to a peaked lookup table that prunes suboptimal tactics. Evaluation on 1,000 independent cases yields 88.8% penetration success - 50.3 points higher than random sampling - with barrel-roll dominating selections. Onboard execution via table lookup and scaling achieves sub-millisecond latency and full transparency for certification. The framework combines RL adaptability with rule-based interpretability for practical, real-time UAV penetration.