融合进化算法和深度强化学习的飞行器制导控制一体化

Aiming at the challenging problem of guidance and control for hypersonic flight vehicles under external disturbances and model uncertainties, this paper proposes an evolutionary reinforcement learning framework that integrates the twin delayed deep deterministic policy gradient and cross-entropy method (CEM). First, the motion model and integrated guidance and control model of the hypersonic flight vehicle are constructed. Second, the multi-constraint control problem in complex disturbed environments is transformed into a reinforcement learning decision optimization process. Leveraging the model-free, data-driven nature of deep reinforcement learning, an end-to-end mapping mechanism from state observations to rudder deflection commands is established. Meanwhile, a CEM-based action space sampling mechanism is introduced, which screens elite candidate action sets through the Q-value maximization criterion and uses the value function to guide the direction of evolutionary search. This effectively overcomes the defects of inefficient and highly blind exploration in traditional reinforcement learning and improves sample utilization efficiency. Finally, simulation results show that the proposed algorithm can adapt to variable mission flight conditions such as initial altitude deviations of ±300 m, velocity deviations of ±200 m/s, and aerodynamic parameter uncertainties of ±40%. It also significantly outperforms traditional control methods in core indicators such as terminal control accuracy and robustness.