Deep Reinforcement Learning Based 3D Integrated Guidance And Control For Hypersonic Missiles

The terminal guidance of hypersonic missiles in 3D space presents strong nonlinearity and coupling. The traditional dual loop guidance methods cannot meet the missile dynamic constraints with time delays. In response to the above issues, this paper proposes an improved deep reinforcement learning algorithm using a non-perfect classifier, trying to prevent invalid experience from entering the experience pool as much as possible, so that the agent can get rid of them when updating strategies. Firstly, the missile guidance and control problem is modeled as a Markov decision process. Then, a simulation environment is constructed based on the dynamic model of hypersonic missiles, an appropriate state space and a dense reward function based on non-perfect classifier are both designed. The soft actor-critic algorithm is utilized to train the agent. An integrated guidance and control strategy is finally obtained, which can generate real-time rudder angle instructions to hit the target based on the current state. The effectiveness, generality, and robustness of the method have been verified through several simulation experiments.