Model-Free Data-Driven Predictive Control Using Reinforcement Learning

This paper proposes a novel approach for Predictive Control utilizing Reinforcement Learning (RL) and Data-Driven techniques to derive optimal control policies for real systems. Using pure input-output multi-step predictors based on Subspace Identification and RL techniques, the resulting predictive control scheme can approximate the optimal control policy of a system with high accuracy, even if the predictor cannot accurately capture the true system dynamics. One of the key contributions of the proposed approach is the extension of the framework connecting Model Predictive Control (MPC) and RL to one that does not require explicit state-space models, nor to define a notion of state at all. The paper demonstrates the efficacy of the proposed approach through an illustrative example, highlighting the ability of our approach to provide an optimal control policy for a real system without requiring any prior knowledge about its internal dynamics.