Abstract
To address feasibility issues in model predictive control (MPC), most implementations relax hard state constraints using additional slack variables with a suitable penalty. We propose an alternative strategy for open-loop asymptotically/Lyapunov stable nonlinear systems by relaxing the initial state constraint with a suitable penalty. The proposed MPC framework is globally feasible, ensures (semi-)global asymptotic stability, and (approximately) recovers the closed-loop properties of the nominal MPC on the feasible set. The proposed framework can be naturally combined with a robust formulation to ensure robustness subject to bounded disturbances while retaining input-to-state stability in case of arbitrarily large disturbances.
Introduction
Model Predictive Control (MPC) is a powerful tool because it handles constraints directly in the optimization loop. However, MPC suffers from a "recursive feasibility" problem: if a sudden, large disturbance pushes the system into a state where it is mathematically impossible to satisfy all constraints over the prediction horizon, the optimizer fails.
Engineers typically address this with "soft constraints"—adding slack variables to state constraints so they can be violated for a high price. However, soft constraints are difficult to tune; low penalties can lead to instability, while high penalties can cause numerical issues and lack strong stability guarantees for large disturbances.
Methodology: Initial State Relaxation
The core innovation of this paper is Initial State Relaxation. In standard MPC, the optimization starts exactly at the current measured state $x(k)$. In the proposed framework, the optimizer is allowed to choose a "nominal" initial state $\bar{x}$ that is near the actual state. A penalty is then added to the cost function based on the distance between the real state $x$ and this virtual state $\bar{x}$.
The "secret sauce" is the choice of the penalty function $V_\delta$. By using an incremental Lyapunov function, the authors prove Input-to-State Stability (ISS). This ensures that as long as disturbances eventually settle, the system is guaranteed to return to the target, regardless of how large the initial disturbance was.
Key Results and Evaluation
The authors evaluated their method using benchmark simulations, including a linear mass-spring-damper system and a nonlinear four-tank system.
Global Feasibility
In the linear mass-spring-damper test, standard soft-constraint methods (Soft-P and Soft-T) became infeasible when the initial condition was scaled by a factor of $c=4.00$. In contrast, the proposed method remained 100% feasible and stable across all tested scales.
Computational Efficiency
A major benefit of this formulation is its simplicity. For linear systems, it remains a simple Quadratic Program (QP). As shown in the performance data below, the average computation time for the proposed method (3.98ms) is nearly identical to nominal MPC (3.73ms) and significantly faster than other robust variants like Soft-T (10.79ms).
Conclusion
This framework provides a robust solution to the feasibility problem in MPC without the computational burden of traditional robust MPC or the stability risks of standard soft constraints. While the current theory assumes open-loop stability, its ability to handle unbounded disturbances while maintaining local optimality makes it a significant advancement for practical control applications. Future research will look to extend these results to unstable systems and account for parametric model mismatches.
Summary updated