Model-Based Actor–Critic Learning for Secure Tracking Control of Vehicle Systems With a Dynamic Event-Triggered Mechanism
提出一种结合模型和演员-评论家学习的框架，在动态事件触发和欺骗攻击下实现车辆安全跟踪，相比无模型方法提高了跟踪精度和收敛速度，并节省带宽。
This article proposes a synergistic model-based actor–critic (AC) learning framework for secure vehicle tracking under a dynamic event-triggered scheme (DETS) and subject to deception attacks. To explicitly resolve the conflict between communication conservation and learning convergence, the AC algorithm is structurally embedded within a discrete linear quadratic tracker (LQT) and backstepping architecture. This co-design utilizes the nominal linear structure as a rigorous prior to analytically derive update laws, ensuring stable Hamilton–Jacobi–Bellman approximation even when data are sparse and stochastically corrupted. By bounding the policy search space, it ensures a minimal number of learnable parameters and a clear physical interpretation. Lyapunov analysis rigorously proves the asymptotic convergence of the tracking errors to zero. Numerical simulations demonstrate superior tracking accuracy, faster convergence, and significant bandwidth savings compared to standard model-free reinforcement learning baselines.