Data-Driven vs Machine Learning MPC: A Comparative Study in Robust Control Systems

This study systematically compares Data-Driven Model Predictive Control (DD-MPC) and Machine Learning-based MPC (ML-MPC) for controlling systems with unknown dynamics under stochastic disturbances. DD-MPC leverages historical input–output data to construct predictive control laws, offering improved robustness in the absence of explicit system models, albeit with higher computational demands. In contrast, ML-MPC integrates neural network approximations to capture complex system behavior, enabling enhanced adaptability, computational efficiency, and predictive accuracy. Simulation results on a multi-input linear time-invariant system subject to Gaussian noise demonstrate that ML-MPC outperforms DD-MPC, achieving up to 15% reduction in tracking error and 20% fewer constraint violations. DD-MPC exhibits stable and reliable intermediate performance, particularly in noise-resilient control scenarios. Overall, the findings highlight ML-MPC as a highly promising approach for robust control in uncertain and data-rich environments, particularly in applications such as autonomous systems. The novelty of this work is to provide a quantitative validation of DD-MPC and ML-MPC under the same uncertain environments, which allows new insight into their respective merits for model-free data-driven control. Future research should focus on formal stability guarantees and real-world experimental validation.