Data-Driven Reinforcement Learning-Based Forgetting Factor Iterative Learning Control
Published in IEEE Transactions on Automation Science and Engineering, 2025
Keywords: Iterative learning control (ILC), reinforcement learning (RL), cyber-physical systems (CPS), resilient data-driven control, K-armed bandit problem.
Abstract
We propose a data-driven control scheme that integrates reinforcement learning (RL) with iterative learning control (ILC). To overcome the shortcomings of conventional ILC under non-repetitive noise, disturbances, and model uncertainties, we introduce a dynamic Forgetting Factor ILC (FFILC). By leveraging past iteration data and optimizing the forgetting factor via RL, FFILC achieves faster convergence and greater adaptability. We provide a rigorous convergence proof and demonstrate, through simulations on a quadrotor and a drum-type boiler turbine plant, that RL-FFILC outperforms advanced baseline controllers in terms of tracking accuracy and robustness.
Note to Practitioners
Industrial automation often relies on repetitive tasks where ILC can enhance performance—but classical ILC falters in the presence of non-repetitive disturbances or poorly known dynamics. RL-FFILC offers a practical, data-driven solution that requires no precise model identification. By adaptively learning the optimal forgetting factor online, this method improves accuracy and resilience in Cyber-Physical Systems, such as UAV flight control and power-plant turbine regulation, making it well suited for real-world industrial deployment.