Closed-loop Analysis of ADMM-based Suboptimal Linear Model Predictive Control

Anusha Srikanthan, Aren Karapetyan, Vijay Kumar, Nikolai Matni

Abstract

Many practical applications of optimal control are subject to strict real-time computational constraints. When applying Model Predictive Control (MPC) in these settings, respecting timing constraints is often achieved by limiting the number of iterations of the optimization algorithm, resulting in "suboptimal" MPC. This paper proposes a suboptimal MPC scheme based on the Alternating Direction Method of Multipliers (ADMM). Focusing on the linear quadratic regulator (LQR) problem with state and input constraints, we demonstrate how ADMM can split the MPC problem into iterative updates of an unconstrained optimal control problem and a dynamics-free feasibility step. We prove that a warm-start approach combined with a sufficient number of iterations per time-step yields a scheme that asymptotically stabilizes the system and maintains recursive feasibility.

Introduction

Model Predictive Control (MPC) is a premier strategy for controlling complex systems due to its ability to explicitly handle state and input constraints. However, the requirement to solve a constrained optimization problem at every sampling interval poses a significant hurdle for high-speed applications, such as autonomous vehicles or aerial robotics. In these contexts, the available computation time is often insufficient to reach full numerical convergence.

Engineers typically address this via "early termination," stopping the solver after a fixed duration or iteration count. While common in practice, this creates a theoretical challenge: the suboptimal control signal acts as a disturbance, and the solver's internal state becomes coupled with the system's physical dynamics. This paper provides a formal stability analysis for such systems, specifically addressing the difficult case where constraints are active and the system-optimizer dynamics are nonlinear.

Methodology

ADMM for Operator Splitting

The proposed scheme employs the Alternating Direction Method of Multipliers (ADMM) to decompose the MPC optimization into two simpler sub-problems:

  1. Dynamics Update: An unconstrained optimal control problem that possesses an analytical solution.
  2. Feasibility Step: A dynamics-free projection that ensures state and input constraints are satisfied.

System-Optimizer Dynamics and Warm-Starting

Rather than treating the solver as an external black box, the authors model the combined evolution of the physical plant and the solver's internal variables as a single interconnected discrete-time dynamical system. A "warm-start" strategy is used, where the solver is initialized at each time step using the solution from the previous step. This initialization is critical for maintaining performance under low iteration counts.

Stability via the Small-Gain Theorem

To prove closed-loop stability, the authors utilize the Small-Gain Theorem—a fundamental tool in control theory. By analyzing how errors propagate through the feedback loop between the plant and the suboptimal optimizer, they derive an explicit lower bound ($\ell^*$) for the number of iterations required per time step to guarantee that the system remains asymptotically stable.

Key Insights and Results

Theoretical Guarantees

The primary contribution of this work is the rigorous proof of local asymptotic stability and recursive feasibility for ADMM-based suboptimal MPC. Recursive feasibility ensures that if the controller finds a valid solution at the first time step, it is guaranteed to find one at all future steps, even with early termination.

Numerical Validation

The approach was evaluated using a Double Integrator System, a standard benchmark representing basic Newtonian motion, under tight constraints. Key findings include:

  • Quantitative Bounds: The theoretical analysis yielded a requirement of $\ell^* = 85$ iterations for guaranteed stability.
  • Practical Performance: Simulations revealed the theoretical bound to be conservative; the system remained stable and satisfied all constraints with as few as 23 to 30 iterations. This significantly outperforms existing baseline methods, which often become unstable when constraints are active.
  • Region of Attraction: The authors demonstrated that as the number of iterations increases, the "Region of Attraction" (the set of initial states from which the system can safely recover) expands, eventually matching the performance of a fully optimal MPC.

Conclusion

This research provides the first formal proof that MPC can remain stable even when the underlying solver is stopped early, provided an ADMM-based warm-start strategy is used. By treating the solver as part of the dynamical system, the authors bridge the gap between theoretical control and real-time implementation. Future work will focus on reducing the conservativeness of the iteration bounds and extending the analysis to nonlinear system dynamics and trajectory tracking.

Summary updated

Abstract

Many practical applications of optimal control are subject to real-time computational constraints. When applying model predictive control (MPC) in these settings, respecting timing constraints is achieved by limiting the number of iterations of the optimization algorithm used to compute control actions at each time step, resulting in so-called suboptimal MPC. This paper proposes a suboptimal MPC scheme based on the alternating direction method of multipliers (ADMM). With a focus on the linear quadratic regulator problem with state and input constraints, we show how ADMM can be used to split the MPC problem into iterative updates of an unconstrained optimal control problem (with an analytical solution), and a dynamics-free feasibility step. We show that using a warm-start approach combined with enough iterations per time-step, yields an ADMM-based suboptimal MPC scheme which asymptotically stabilizes the system and maintains recursive feasibility.
BibTeX
@inproceedings{srikanthan2024closed,
  author = {Anusha Srikanthan and Aren Karapetyan and Vijay Kumar and Nikolai Matni},
  title = {{Closed-loop Analysis of ADMM-based Suboptimal Linear Model Predictive Control}},
  booktitle = {IEEE Control. Syst. Lett},
  year = {2024},
  url = {https://openreview.net/forum?id=0W1W6JWid7},
}
Figure