Three-Dimensional Guidance Law Design against Maneuvering Target via Deep Reinforcement Learning

To improve the engagement performance against a maneuvering target, a deep reinforcement learning (DRL)-based guidance law is proposed. An interceptor dynamic is first formulated which is trained by a deep deterministic policy gradient (DDPG) algorithm. To generate a robust RL-based guidance law, the interceptor is regarded as an agent. By interacting with a time-varying environment containing a maneuvering target, the hyperparameters of the DDPG algorithm are optimized by sampling a batch size of experience data offline. The reward shape function is properly designed to ensure a rapid and stable training process for the DDPG-based agent. The optimal guidance policy is extracted and used online to generate the overload commands to the interceptor. The simulation results show that an effective interception can be realized in spite of the change of target maneuvers and initial engagement scenarios, which proves the effectiveness of the proposed RL-based guidance law.