Enhancing Terminal Missile Capturability via Progressive Training and Deep Deterministic Policy Gradient-Based Reinforcement Learning

초록

This study investigates the enhancement of capturability for terminal homing missiles through reinforcement learning. A rigid-body missile model incorporating true proportional navigation guidance is utilized, coupled with a three-loop autopilot for aerodynamic fin control and a quaternion-based feedback system for thrust vector control. The deep deterministic policy gradient algorithm is employed due to its effectiveness in managing continuous action spaces. To further improve training efficiency and adaptability, progressive training is implemented by incrementally varying the missile's initial pitch and yaw angles, thereby progressively increasing scenario complexity. The reinforcement learning observation space includes control inputs and geometric parameters, while the action space involves the distribution ratios for aerodynamic and thrust vector controls. The designed reward function considers zero-effort-miss, control effort, and angular velocity. Numerical simulations demonstrate that the proposed reinforcement learning-based approach with progressive training achieves an enlarged capture region compared to methods employing static control allocation.

키워드