Finite-horizon LQ control for unknown discrete-time linear systems via extremum seeking

We present a non-model based approach for asymptotic, locally exponentially stable attainment of the optimal open-loop control sequence for unknown, discrete-time linear systems with a scalar input, where not even the dimension of the system is known. This control sequence minimizes the finite-time horizon cost function, which is quadratic in the measured output and in the input. We make no assumptions on the stability of the unknown system, but we do assume that the system is reachable. The proposed algorithm employs the multi-variable discrete-time extremum seeking approach to minimize the cost function, extending results established for the scalar discrete-time extremum seeking method. Simulation results show that the Hessian's condition number, used as a measure of the optimization problem's level of difficulty, increases with both the system's level of instability and the length of the finite horizon for a scalar system. Thus, we suggest solving well-conditioned, shorter time horizon optimal control problems to obtain good initial control estimates for longer time horizon problems if possible. We also show that the algorithm accommodates input constraints by employing the projection operator and introduce a Newton-based discrete-time extremum seeking controller, which removes the convergence rate's dependence on the unknown Hessian.