Research on Behavior Modeling of Computer Generated Forces Based on Improved DDQN

Current development of Computer Generated Forces (CGF) faces dual challenges in efficiency and intelligence. This study proposes a novel behavior modeling method driven by reinforcement learning (RL). We first manually construct a behavior tree (BT) for a typical anti-submarine helicopter search mission and conduct preliminary validation. Subsequently, RL dynamically optimizes the BT by encapsulating attack actions into Double Deep Q-Network (DDQN) nodes, significantly enhancing decision robustness in complex adversarial environments. Finally, through optimized environmental parameters, agent architecture improvements, and reward function refinement, the model achieves substantial performance leaps: target detection rate increases to 76% (from 18%) and attack success rate reaches 58% (from 9%). This work advances CGF development toward automation and intelligence.