Stochastic MPC with Online-optimized Policies and Closed-loop Guarantees

Marcell Bartos, Alexandre Didier, Jerome Sieber, Johannes Köhler, Melanie N. Zeilinger

TL;DR

This paper introduces a novel Stochastic Model Predictive Control (SMPC) approach for linear systems with Gaussian disturbances. It achieves guaranteed closed-loop constraint satisfaction and recursive feasibility by optimizing affine feedback policies online, leading to improved performance and reduced conservatism compared to existing methods.

What Problem Does This Paper Address?

Model Predictive Control (MPC) is a powerful control strategy, but applying it to systems with uncertainties (like random disturbances) while guaranteeing safety (constraint satisfaction) is challenging. Traditional Stochastic MPC (SMPC) methods often struggle to balance performance with robust constraint satisfaction, especially when dealing with unbounded disturbances like Gaussian noise. Some methods achieve good performance by optimizing control policies online but sacrifice guaranteed constraint satisfaction. Others guarantee safety but are more conservative and don't leverage online optimization for better performance. This paper aims to bridge this gap by developing an SMPC scheme that simultaneously offers online-optimized policies for improved performance and strong guarantees for closed-loop constraint satisfaction.

Methods

The core of the proposed method lies in two key innovations:

  1. Online-optimized Affine Feedback Policies: Instead of using fixed feedback strategies, the method optimizes affine feedback policies online. This is achieved by leveraging the System Level Parameterization (SLP) framework, which allows for convex optimization of these policies. This directly addresses the performance limitation of fixed-feedback approaches.
  2. Reconditioning and Finitely Determined Terminal Sets: To ensure recursive feasibility and closed-loop constraint satisfaction, the paper introduces a "reconditioning" mechanism. This means that predicted probabilistic constraints are updated at each time step based on the most current knowledge of the system and past disturbances. Additionally, a finitely determined maximal admissible set is designed for the terminal part of the prediction horizon. This set acts as a safety net, ensuring that even with an optimized policy, the system remains within safe bounds in the long run.

The overall approach involves solving a finite-horizon optimization problem at each time step in a receding horizon fashion. The problem formulation is a Second-Order Cone Program (SOCP), which is computationally tractable.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • The Trade-off is Not Inevitable: The paper demonstrates that it's possible to achieve both online-optimized performance and guaranteed closed-loop constraint satisfaction, challenging the notion that one must be sacrificed for the other in SMPC.
  • Reconditioning is Key for Online Adaptation: By reconditioning probabilistic constraints on the latest information, the controller can adapt to disturbances more effectively while maintaining safety guarantees. This is crucial for online-optimized policies where feedback matrices become random variables.
  • Terminal Sets for Guaranteed Safety: Designing a specific type of terminal set (finitely determined maximal admissible set) is essential for ensuring that the system remains safe even beyond the finite prediction horizon, especially when combined with a fixed stabilizing controller.
  • SLP for Convexity: The use of System Level Parameterization is instrumental in reformulating the problem to allow for convex optimization of affine feedback policies, making the online optimization tractable.

Key Results

The paper's main contributions are theoretical guarantees and a numerical demonstration:

  • Theoretical Guarantees: The proposed SMPC scheme guarantees recursive feasibility and closed-loop satisfaction of probabilistic constraints for linear systems with additive Gaussian disturbances. It also provides a bound on the asymptotic average cost, comparable to static state-feedback controllers.
  • Numerical Example: A building temperature control example shows that the proposed method (RC and RC-mod) operates closer to the constraints and achieves a lower closed-loop cost compared to existing methods like CLP [17] and Indirect Feedback (IF) SMPC [13]. The RC-mod variation, which uses a slightly more conservative terminal set, shows improved performance over the standard RC approach by allowing more deviation from the initial optimized policy. The method also demonstrates tighter constraint satisfaction than CLP.

Limitations and Future Work

  • Computational Cost: While the problem is formulated as an SOCP, it is more computationally intensive than simpler methods like QP-based controllers (e.g., CLP). The offline design of the terminal set can also be time-consuming.
  • Linear Systems and Gaussian Disturbances: The current framework is limited to linear time-invariant systems and additive Gaussian disturbances.
  • Future Work: The authors suggest extending the results to more general settings, including:
  • Uncertain systems.
  • Nonlinear systems.
  • Output feedback.
  • Subgaussian or bounded-variance disturbance distributions.

Summary updated

Abstract

This paper proposes a stochastic model predictive control method for linear systems affected by additive Gaussian disturbances. Closed-loop satisfaction of probabilistic constraints and recursive feasibility of the underlying convex optimization problem is guaranteed. Optimization over feedback policies online increases performance and reduces conservatism compared to fixed-feedback approaches. The central mechanism is a finitely determined maximal admissible set for probabilistic constraints, together with the reconditioning of the predicted probabilistic constraints on the current knowledge at every time step. The proposed method's reduced conservatism and improved performance in terms of the achieved closed-loop cost is demonstrated in a numerical example.
BibTeX
@misc{bartos2025stochastic,
  author = {Marcell Bartos and Alexandre Didier and Jerome Sieber and Johannes Köhler and Melanie N. Zeilinger},
  title = {{Stochastic MPC with Online-optimized Policies and Closed-loop Guarantees}},
  year = {2025},
  eprint = {2502.06469},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2502.06469},
}
Figure