We propose a novel flexible-step model predictive control algorithm for unknown linear time-invariant discrete-time systems. The goal is to asymptotically stabilize the system without relying on a pre-collected dataset that describes its behavior in advance. In particular, we aim to avoid a potentially harmful initial open-loop exploration phase for identification, since full identification is often not necessary for stabilization. Instead, the proposed control scheme explores and learns the unknown system online through measurements of inputs and states. The measurement results are used to update the prediction model in the finite-horizon optimal control problem. If the current prediction model yields an infeasible optimal control problem, then persistently exciting inputs are applied until feasibility is reestablished. The proposed flexible-step approach allows for a flexible number of implemented optimal input values in each iteration, which is beneficial for simultaneous exploration and exploitation. A generalized control Lyapunov function is included into the constraints of the optimal control problem to enforce stability. This way, the problem of optimization is decoupled from the problem of stabilization. For an asymptotically stabilizable unknown control system, we prove that the proposed flexible-step algorithm can lead to global convergence of the system state to the origin.
Pietschner et al. (Wed,) studied this question.