Abstract

Learning-based surrogate models offer a powerful alternative to analytical models for model-based control of nonlinear dynamical systems with uncertain and context-dependent dynamics. A receding-horizon control framework is developed that exploits flow-matching models to generate state trajectories conditioned on the system’s initial state, uncertain parameters, and candidate input sequences. These models provide expressive, data-driven surrogate dynamics without requiring explicit analytical representations. To compute control inputs, Bayesian optimization minimizes a cost function evaluated on surrogate-generated trajectories, enabling efficient optimization despite the non-differentiable and computationally expensive nature of the generative model. The resulting inputs are applied in a receding-horizon fashion and re-optimized using updated state and parameter information, yielding an adaptive, learning-based control strategy. The effectiveness of the approach is demonstrated on a bioreactor system.

Link: https://ieeexplore.ieee.org/document/11625215