Bias-compensated Q-learning for optimal tracking control under denial-of-service attacks

Networked control systems are widely adopted in modern applications. Nevertheless, learning optimal control policies from data under communication attacks remains challenging. In this paper, we propose a bias-compensated Q-learning algorithm for knowledge-informed optimal tracking control. The method considers discrete-time linear time-invariant systems subject to control-channel denial-of-service attacks. Through a bias compensation mechanism that addresses control-input losses, the proposed method eliminates the need for explicit expectation computation in Q-function updates and enables direct data-driven learning under DoS attacks. Moreover, the proposed method requires no knowledge of system dynamics beyond the known system structure and supports data collection under arbitrary behavior policies. This design preserves high data efficiency, computational simplicity, and robustness within a model-free framework. Numerical simulations on a planar quadrotor position-tracking task show that a Riccati-equivalent optimal policy can still be learned from data under 40% DoS attacks.