Tactical UAV path optimization under radar threat using deep reinforcement learning

The majority of the research efforts that aim to solve UAV path optimization problems in a Reinforcement Learning (RL) setting focus on closed spaces or urban areas as the operating environment. The problem of Tactical UAV (TUAV) path planning under hostile radar tracking threat has some peculiarities that distinguish it from other typical UAV path optimization problems. Particularly, 1-spatial regions delineated by threat probabilities may be legitimately penetrable under certain conditions that do not impair the survivability of the UAV and 2-A TUAV is detectable by a radar via its Radar Cross Section (RCS) which is a function of multiple parameters such as the radar operating frequency, the shape of the UAV and more importantly the engagement geometry between the radar and the UAV. The latter suggests that any maneuver performed by the UAV may change multiple angles that specify the engagement geometry. The work presented in this paper proposes a RL based solution to this complex problem in a novel way by 1-Implementing a Markov Decision Process (MDP) compliant RL environment with comprehensive probabilistic radar behavior models incorporated into it and 2-Integrating a core RL algorithm (namely DQN with Prioritized Experience Replay (DQN-PER) with a specific variant of transfer learning (namely learning from demonstrations (LfD)) in a single framework, demonstrating the utility of combining a core RL algorithm and a machine learning scheme toward boosting the performance of a learning agent, and more importantly to alleviate the sparse reward problem.