Reinforcement Learning Motivated Feedforward Control Approach for Disturbance Rejection and Tracking

This paper presents a new discrete-time method for feedforward control for disturbance rejection and tracking to enhance the dynamic system response of linear time-invariant (LTI) systems. The proposed approach uses optimization techniques prevalent in Reinforcement Learning (RL), without requiring prior knowledge of the plant dynamics or a disturbance model. Unlike in usual RL or Adaptive Dynamic Programming frameworks, where the learning process is driven by temporal differences used to approximate an optimal cost or value function, this paper proposes a learning scheme based on immediate rewards. Thereby, the connection between reward based RL, system identification and adaptive control is addressed. It is further shown, how classical adaptive feedforward control problems can be transformed into a RL setting, and a sample efficient algorithm is presented including some practical implementation advice.