IJCAS

Two-time-scale (TTS) systems were proposed to describe accurately complex systems that include multiple variables running on two-time scales. Different response speeds of variables and incomplete model information affect the tracking performance of TTS systems. For tracking control of an unknown model, the practicability of reinforcement learning (RL) has been subject to criticism, as the method requires a stable initial policy. Based on singular perturbation theory (SPT), a composite sub-optimal tracking policy is investigated combining model information with measured data. Besides, a selection criterion for the initial stabilizing policy is presented by considering the policy as an input constraint. The proposed method integrating RL technique with convex optimization improves the tracking performance and practicability effectively. Finally, an emulation experiment in F-8 aircraft is given to demonstrate the validity of the developed method.

Keywords: Convex optimization, input constrained, reinforcement learning, sub-optimal tracking control, twotime-scale system.