ABSTRACT In this paper we propose a model‐free reinforcement learning framework for adaptive control of continuous‐time linear systems without requiring prior system knowledge. By estimating system states from measurable outputs and iteratively minimizing the optimal cost function, both optimal and suboptimal output feedback control policies are derived. An integral reinforcement learning approach with an actor–critic structure enables real‐time estimation of Q‐function parameters, thereby facilitating adaptive policy optimization. In particular, the proposed approach can outperform traditional LQR‐based output‐feedback control by delivering superior results in complex systems, while maintaining good performance in simple scenarios. Simulation results on a Multi‐agent Network with Eliminated Constraints and an active vibration control system validate the framework's effectiveness and robustness under uncertainty.
Jitian et al. (Tue,) studied this question.