Reinforcement learning-based near-optimal control for discrete time-delay singularly perturbed systems with unmeasurable states

This brief proposes a novel reinforcement learning-based composite near-optimal output feedback (OPFB) control method for discrete time-delay singularly perturbed systems (SPSs) with unknown dynamics and unmeasured system states. Firstly, an optimal control problem of the full-order time-delay SPS is formulated. The singular perturbation techniques are used to obtain the reduced-order subsystems and equivalent subproblems. Then, a state reconstruction method is presented to represent the state in terms of available input and output data. In addition, a reinforcement learning (RL) algorithm based on input and output data is developed. Theoretical analyses are discussed on the convergence of the proposed algorithm and the asymptotical stability of the system under the composite controller. Finally, numerical simulation results illustrate the effectiveness of the proposed approach.