Realizing Midcourse Penetration With Deep Reinforcement Learning

A midcourse maneuver controller is obtained using deep reinforcement learning to maintain the survivability of a ballistic missile. First, abstracted as Markov decision process (MDP) with an unknown system state equation. Then, formed by Dueling Double Deep Q (D3Q) neural network used approximate state-action value function MDP. In order make controller’s intelligence improved learning, space, action and instant reward MDP are customized. The uses real-time situation input outputs ignition states pulse motors. Offline training shows that can achieve optimal strategy’s convergence after approximately 65 hours. Online tests demonstrate ability avoid interceptor intelligently account for entry error. scenarios multiple random factors, achieved penetration probability 100% mean re-entry error less than 5000 m.