Abstract
Research on sub-optimal Model Predictive Control (MPC) has led to a variety of optimization methods based on explicit or online approaches, or combinations thereof. Its foremost aim is to guarantee essential controller properties, i.e. recursive feasibility, stability, and robustness, at reduced and predictable computational cost, i.e. computation time and storage space. This paper shows how the input sequence of any (not necessarily stabilizing) sub-optimal controller and the shifted input sequence from the previous time step can be used in an optimal convex combination, which is easy to determine online, in order to guarantee input-to-state stability for the closed-loop system. The presented method is thus able to stabilize a wide range of existing sub-optimal MPC schemes that lack a formal stability guarantee, if they can be considered as a continuous map from the state space to the space of feasible input sequences.
The Challenge of Fast Control
Model Predictive Control (MPC) is a cornerstone of modern control engineering, particularly for complex systems with strict constraints. By solving an optimization problem at every time step, MPC identifies the optimal future trajectory. However, for high-speed applications like robotics or aerospace, the computational burden of solving these optimizations in real-time can be prohibitive.
Engineers often resort to "sub-optimal" MPC—using approximations or early solver termination to meet timing constraints. The critical drawback is that these approximations frequently sacrifice formal guarantees of stability and robustness. A system under sub-optimal control may fail to recover from unexpected disturbances. This research addresses this gap by providing a rigorous framework to stabilize fast, approximate controllers.
A Two-Step Solution: Tightening and Blending
The proposed method introduces a two-step "patch" to standard MPC frameworks:
- Constraint Tightening: The authors slightly reduce the allowed state and control boundaries by a safety margin ($\delta$). This ensures that even with disturbances, the previously calculated plan remains mathematically feasible.
- The Affine Combination: Instead of relying solely on a sub-optimal guess $s(x_k)$, the controller considers a blend of this guess and the "shifted" plan from the previous time step. A simple 1D optimization determines a blending factor $\alpha \in [-1, 1]$ to minimize the cost: $$u_k = \alpha \cdot (\text{shifted previous plan}) + (1 - \alpha) \cdot (\text{new sub-optimal guess})$$
This 1D search adds negligible overhead while ensuring the mathematical "descent" required for stability proofs.
Key Insights: The Role of the Blending Factor
The blending factor $\alpha$ serves as a diagnostic tool for the controller's behavior. In nominal conditions without noise, the system relies heavily on the shifted previous plan, with $\alpha$ values clustering near 1.0. When disturbances occur (the robust case), the controller adaptively shifts toward the new sub-optimal guess to compensate, resulting in a broader distribution of $\alpha$ values.
Performance and Robustness
Evaluated on a single mass oscillator simulation, the method demonstrated that any continuous sub-optimal map can be made Input-to-State Stable (ISS). The sub-optimal controller remained highly efficient, with costs typically within 10% of the optimal solution in quiet conditions and rarely exceeding a 25% deviation even under disturbance.
Conclusion
By combining a "fresh" sub-optimal guess with a "proven" previous plan, the authors bridge the gap between computational speed and mathematical reliability. While the current work focuses on linear systems and requires the sub-optimal map to be continuous, it provides a powerful template for making approximate control schemes robust enough for real-world deployment.
Summary updated