Adaptive guidance for missile based on meta-learning and predictive control

To address the complex impacts of aerodynamic perturbations, target maneuvers, actuator failures, and missile parameter variations, this paper proposes an adaptive guidance framework based on deep reinforcement learning (DRL). First, the framework employs entropy-regularized DRL, utilizing a deep neural network (DNN) as the predictive model. By integrating a forgetting mechanism into meta-learning, it dynamically adjusts experience replay weights to prioritize the current environment. Subsequently, a dual-model predictive control architecture is designed, combining cross-entropy method model predictive control (CEM-MPC) and adaptive entropy model predictive path integral (MPPI) control. This architecture monitors the rate of change in cost standard deviation and automatically switches control strategies when fluctuations exceed a threshold, thereby leveraging the convergence speed and exploration advantages of both controllers. Simulations demonstrate that compared to traditional learning-based non-adaptive methods, the proposed approach stabilizes line-of-sight angular rates and improves interception success rates by 47.6 and 6.0 respectively, validating its effectiveness.