A Scientific Machine Learning Approach for Autonomous Maneuver Decision in Air Combat

This research presents a novel Scientific Machine Learning (SciML) approach to enable autonomous decision-making for maneuvering in air combat scenarios involving unmanned aerial vehicles (UAVs). While Reinforcement Learning (RL) has been widely applied in previous work, RL encounters a number of challenges ranging from the complexity of reward engineering to state-space expansion and convergence inefficiency when dealing with extremely dynamic environments. In contrast, the suggested SciML framework incorporates the dynamics of aircraft control within a Universal Ordinary Differential Equation (UODE) framework that allows for end-to-end optimization of tactical decision-making with better generalization and sample efficiency. The maneuver policy is represented by a neural network, where training is accomplished through gradient-based optimization with a physics-informed loss function based on tracking accuracy, collision avoidance, energy-efficient control, and boundary constraint satisfaction. Systematic simulation results for both stationary and evasive target scenarios validate the correctness and robustness of the methodology being considered. The SciML model exhibits competitive real-time performance compared to conventional reinforcement learning-based approaches while substantially lowering the training expenses, thus qualifying as a viable alternative for future autonomous aerial combat systems.