Reinforcement Learning-Based Dynamic Event-Triggered Nonlinear Games Under Incomplete and Asymmetric Information

Abstract This article proposes a dynamic event-triggered reinforcement-learning-based composite game interception method for unknown maneuvering targets under incomplete and asymmetric information and maneuverability constraints. Unlike many existing game-based interception methods that rely on a worst-case target response assumption, the proposed method evaluates the real-time effect of target maneuvers and adaptively switches between a nongame optimal strategy and a game-based strategy. This reduces unnecessary policy conservatism while preserving robustness against strongly evasive maneuvers. To address unknown target intention preference, a disturbance observer is designed to reconstruct the target maneuvering input, and the intention weighting matrix is inferred through the target-side optimality condition. Critic, actor, and virtual actor neural networks are developed to approximate the value function, the interceptor policy, and the virtual target policy, respectively. Dynamic event-triggered mechanisms are further introduced to reduce redundant policy updates and situation evaluations. The closed-loop system state, neural-network weight estimation errors, and disturbance observation error are proved to be uniformly ultimately bounded, and Zeno behavior is excluded. Simulation results demonstrate the effectiveness of the proposed method.