Steady-State Error Compensation in Reference Tracking and Disturbance Rejection Problems for Reinforcement Learning-Based Control

Abstract

Reinforcement learning (RL) is a promising, upcoming topic in automatic control applications. Where classical control approaches require a priori system knowledge, data-driven control approaches like RL allow a modelfree controller design procedure, rendering them emergent techniques for systems with changing plant structures and varying parameters. While it was already shown in various applications that the transient control behavior for complex systems can be sufficiently handled by RL, the challenge of non-vanishing steady-state control errors remains, which arises from the usage of control policy approximations and finite training times. To overcome this issue, an integral action state augmentation (IASA) for actor-critic-based RL controllers is introduced that mimics an integrating feedback, which is inspired by the delta-input formulation within model predictive control. This augmentation does not require any expert knowledge, leaving the approach model free. As a result, the RL controller learns how to suppress steady-state control deviations much more effectively. Two exemplary applications from the domain of electrical energy engineering validate the benefit of the developed method both for reference tracking and disturbance rejection. In comparison to a standard deep deterministic policy gradient (DDPG) setup, the suggested IASA extension allows to reduce the steady-state error by up to 52 % within the considered validation scenarios.

Deep reinforcement learning for PMSG wind turbine control via twin delayed deep deterministic policy gradient (TD3)

“…16 The authors of this study also implemented steady-state error compensation as a means of enhancing the efficacy of the DDPG agent. 17 Actor-critic methods such as DDPG are in general reinforcement learning techniques in which two neural networks, named "actor" and "critic" are present. 18,19 The actor learns a policy that is modeled by a parameterized distribution, while the critic learns either a value function (i.e., a function V(s) that provides the expected return when the initial state is s and a given policy is applied) or an action-value function (i.e., a function Q(s, a) that provides the expected return obtained when the initial state is s, an arbitrary action a which may not have been obtained from the given policy is taken at s, while the given policy is used for all subsequent discrete time instants) and uses it to evaluate the performance of the policy optimized by the actor.…”