On convex problems in chance-constrained stochastic model predictive control

Eugenio Cinquemani, Mayank Agarwal, Debasish Chatterjee, John LygerosarXiv 2009 · DOI · arXiv

Abstract

We investigate constrained optimal control problems for linear stochastic dynamical systems evolving in discrete time. We consider minimization of an expected value cost over a finite horizon. Hard constraints are introduced first, and then reformulated in terms of probabilistic constraints. It is shown that, for a suitable parametrization of the control policy, a wide class of the resulting optimization problems are convex, or admit reasonable convex approximations.

Introduction

In real-world control applications—ranging from autonomous drone navigation to the management of complex chemical processes—systems are rarely isolated. They are constantly subjected to random disturbances, or stochasticity. To ensure safety and operational efficiency, controllers must keep system states (like velocity or temperature) within specific bounds.

Traditionally, Model Predictive Control (MPC) has relied on "hard" constraints. However, in a stochastic environment, a sufficiently large random disturbance will eventually cause any hard constraint to be violated, making the optimization problem infeasible. To address this, researchers utilize chance constraints, which require that limits are satisfied with a high probability (e.g., 95% or 99%). While theoretically sound, chance constraints are notoriously difficult to implement because they are usually non-convex, leading to high computational costs that are incompatible with the real-time requirements of MPC.

Methodology: Bridging Theory and Tractability

This research demonstrates that many stochastic model predictive control (SMPC) problems with probabilistic constraints can be reformulated as efficient convex optimization problems. The authors achieve this through two primary strategies: policy parameterization and convex approximation.

1. Noise-Feedback Policy Parameterization

A significant hurdle in stochastic control is that standard state-feedback laws ($u = Kx$) often lead to non-convexity when combined with chance constraints. The authors leverage a powerful "noise-feedback" trick: expressing the control input $u$ as an affine function of past noise realizations ($w$). This parameterization ensures that the system state and control inputs remain affine functions of the optimization variables, a fundamental requirement for maintaining convexity.

2. Convex Approximations of Chance Constraints

Since exact chance constraints remain non-convex in many settings, the paper introduces several conservative but tractable convex approximations:

  • Constraint Separation: This method uses Boole’s inequality to break a joint probability requirement (the probability that all constraints are met) into individual constraints for each time step.
  • Confidence Ellipsoids: By utilizing the geometric properties of Gaussian noise, the controller ensures the system remains within a "safe" ellipsoid. This maps the problem into Second-Order Cone Programs (SOCP) or Linear Matrix Inequalities (LMI).
  • Integrated Chance Constraints (ICC): Unlike standard chance constraints that focus on the probability of violation, ICCs penalize the expected magnitude of the violation. The authors prove that this approach is inherently convex under very general conditions and provides smoother numerical stability.

Experimental Validation

The proposed methods were evaluated using a synthetic benchmark: a discrete-time model of a mechanical system consisting of four masses connected by springs and dampers. This system provides a complex enough dynamic to test the limits of the optimization frameworks. The researchers conducted 1,000 simulated sample paths to compare how different constraint formulations handled random disturbances.

Key Results

  • Computational Efficiency: The authors successfully mapped various chance-constrained problems into standard convex optimization frameworks (SOCP and LMI). These can be solved in milliseconds using off-the-shelf solvers like CVX, making them suitable for real-time applications.
  • Performance and Safety: In the 4-mass system simulations, the conservative approximations were highly effective. Even when the allowed violation probability ($\alpha$) was set to a relatively high 10%, the system recorded zero violations across all 1,000 runs, highlighting the robustness of the conservative approach.
  • Feedback Sensitivity: The study found that LMI-based ellipsoidal constraints resulted in a control policy more sensitive to noise (active feedback), whereas polytopic constraint methods tended toward "open-loop" behavior with smaller feedback gains.

Discussion and Insights

A critical insight from the study involves the "concentration of measure" phenomenon. While ellipsoidal methods appear more sophisticated, they can become overly conservative as the number of constraints or the time horizon increases. In high-dimensional spaces, the volume of the "safe" region shrinks rapidly, potentially leading to suboptimal performance.

Furthermore, Integrated Chance Constraints (ICC) emerged as a superior alternative for many practical scenarios. Because ICCs account for the severity of a violation rather than just its occurrence, they allow for more flexible and stable control laws that do not "overreact" to minor, low-risk boundary crossings.

Limitations and Future Work

While this work advances the tractability of SMPC, several challenges remain:

  • Full State Observation: The current framework assumes the controller can perfectly observe the system state to reconstruct noise. Future research must extend these results to "output feedback" scenarios where states are estimated from noisy sensors.
  • Linearity Constraints: The results are currently limited to linear dynamical systems. Extending convexity results to nonlinear dynamics remains a significant open problem in the field.
  • Tuning Conservatism: Because the methods are conservative, they may sacrifice some performance for safety. Future work could focus on dynamically tuning parameters (such as the risk allocation in constraint separation) to close the performance gap.

Conclusion

By combining noise-feedback parameterization with robust convex approximations, this paper provides a roadmap for implementing chance-constrained control in real-time stochastic systems. The transformation of these difficult problems into SOCPs and LMIs brings high-assurance stochastic control one step closer to widespread industrial application.

Summary updated

Abstract

We investigate constrained optimal control problems for linear stochastic dynamical systems evolving in discrete time. We consider minimization of an expected value cost over a finite horizon. Hard constraints are introduced first, and then reformulated in terms of probabilistic constraints. It is shown that, for a suitable parametrization of the control policy, a wide class of the resulting optimization problems are convex, or admit reasonable convex approximations.
BibTeX
@misc{cinquemani2009convex,
  author = {Eugenio Cinquemani and Mayank Agarwal and Debasish Chatterjee and John Lygeros},
  title = {{On convex problems in chance-constrained stochastic model predictive control}},
  year = {2009},
  eprint = {0905.3447},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/0905.3447},
}
Figure