Active Defence Guidance for Hypersonic Vehicle with Incomplete Information Based on Reinforcement Learning

This paper investigates the active defense guidance problem for the hypersonic vehicle in target-interceptor-defender scenario. The active defense guidance problem of the hypersonic vehicle always subjects to the limitations of incomplete observation information. To tackle this issue, this paper proposes a cooperative active defense guidance based on Convolutional Deep Q-Network (CDQN) algorithm. By regarding the active defense scenario as the partially observable Markov decision process, the guidance problem is solved in the framework of reinforcement learning. In view of the spatiotemporal continuity properties of hypersonic vehicle, a stacking mechanism is proposed to process the incomplete information. Based on which, the convolutional neural networks are further employed to derive the cooperative active defense guidance law. Moreover, to tackle the sparse reward problem in CDQN's training, a continuous reward function is shaped based on environmental potential functions. Finally, numerical experiments are performed to demonstrate the performance and robustness of the proposed active defense guidance.