Linear Trajectory Tracking with Integral Action based on Reinforcement Learning

Abstract

This paper proposes a model-free linear quadratic tracking controller, with integral action and based on reinforcement learning, for linear plants in the presence of constant input disturbances. The novelty lies on the methodology to implement the reference generator, providing a wide class of reference trajectories to be tracked, alongside the integral action responsible for rejecting disturbances and driving the mean value of the plant output in steady-state to a desired value, while the state-feedback from the generator deals with the zero-mean desired trajectory component. Numerical simulations illustrate the closed-loop performance.