Reinforcement learning adaptive network for nonaffine discrete-time control systems: Managing implicit zero-gain

This paper investigates the phenomenon of implicit zero control gain in nonaffine discrete-time systems caused by dead-zone regions, utilizing an experimental setup focused on motor torque control. It challenges the conventional assumptions in adaptive controller designs, which typically require a nonzero control gain. An adaptive controller is subsequently developed using an actor–critic architecture integrated with fuzzy-rule networks and a customized surfacing reward function dealing both tracking error and control effort. The proposed controller achieves near-optimal tracking performance, improves control energy efficiency, and exhibits robustness against unknown disturbances. Learning laws are derived to optimize the controller’s performance without imposing a minimum control gain requirement. The controller’s effectiveness in addressing the dead-zone problem is validated through experimental and comparative results, especially under disturbance amplitudes of approximately 30% of the desired output trajectory, including high-frequency components. •DC motor experiments show zero-gain challenges in adaptive control designs.•A surfacing reward function replaces positive vectors with adjustable parameters.•Time-varying learning rates improve disturbance rejection and resolve zero-gain issues.