Model Predictive Control (MPC) is exceptionally powerful because of a simple premise: if you have a sufficiently accurate mathematical model of a system, you can simulate the future, optimize your control inputs over a forward horizon, execute the first step, and repeat.
But this premise conceals a fundamental vulnerability. In the real world, the internal mathematical model used by the controller and the actual physical system (the plant) are rarely identical. Physical components age and drift, high-frequency dynamics are omitted for computational efficiency, and environmental disturbances introduce unmodeled forces. This discrepancy is known as model-plant mismatch (MPM).
When a predictive controller operates with an inaccurate model, its forward simulations diverge from reality. The optimized control inputs it computes are tailored to a fictional system, which can cause the real-world plant to oscillate, operate suboptimally, or catastrophically violate safety constraints. To deploy predictive controllers reliably, we must understand how to quantify our sensitivity to these errors and design mechanisms to mitigate them.
Quantifying the Limits of Tolerable Error
Before fixing a mismatch, we must understand how much error a predictive controller can tolerate before it degrades or destabilizes. If we can map the sensitivity of a controller to specific parametric errors, we can define a safe operating window.
Lino O. Santos established an offline framework to compute these exact boundaries in A tool to analyze robust stability for model predictive controllers. By applying Nonlinear Programming (NLP) sensitivity analysis, the author demonstrated that we can calculate robust stability bounds for existing controllers without needing to redesign them from scratch. A key insight from this analysis is that directional errors matter immensely: for instance, overestimating the response speed of a physical process drastically narrows the stable operating window compared to underestimating it.
[Internal Model: Fast Response] ---> Predicts rapid stabilization
[Actual Plant: Slow Response] ---> Results in over-correction & oscillations
In real-time optimization contexts, this issue manifests as a gradient mismatch. As explored by Marta Zagorowska et al. in Sensitivity of Online Feedback Optimization to time-varying parameters, when a controller calculates optimization gradients using an inaccurate model, the step it takes may move the physical system away from the true optimum. Formalizing these sensitivity expressions allows engineers to predict exactly how much an objective will degrade when environmental parameters drift.
Active Compensation Layers
If modifying the predictive model is too computationally expensive or complex, an alternative strategy is to alter the physical plant's behavior so that it forces the real system to match the idealized model. This is achieved by inserting an ultra-fast, low-level adaptive control layer between the high-level MPC and the plant.
+------------------------+ Idealized State +------------------------+
| Model Predictive |-------------------------->| Active Compensation |
| Control (High-Level) |<--------------------------| Layer (e.g., L1 / UC) |
+------------------------+ Actual State +------------------------+
|
v Modified Input
+------------------------+
| Physical Plant |
+------------------------+
Angela P. Schoellig et al. leverage this architecture in Adaptive Model Predictive Control for High-Accuracy Trajectory Tracking in Changing Conditions by pairing a standard linear MPC with an $L_1$ adaptive control layer. The $L_1$ layer acts as a rapid disturbance observer, actively canceling out unmodeled dynamics (such as aerodynamic drag or wind gusts) before they can disrupt the system. Because the low-level loop absorbs the discrepancies, the high-level MPC can plan trajectories assuming its simple, idealized model is perfectly accurate.
Similarly, Ilya V. Kolmanovsky introduces Uncertainty Compensation MPC (UC-MPC) in Robust Adaptive MPC Using Uncertainty Compensation. By utilizing fast adaptation to cancel out unmodeled friction or external disturbances, UC-MPC provides tight, guaranteed performance bounds. This allows the controller to satisfy strict safety constraints without the extreme conservatism—such as severely slowing down or restricting the system's state space—that typically plagues traditional robust control methods.
Distorting the Optimization Landscape
Instead of forcing the plant to match the model, can we alter the optimization problem itself so that an incorrect model still yields the correct real-world actions?
Mario Zanon proved a powerful theoretical result answering this question in Data-Driven Economic NMPC Using Reinforcement Learning. The author demonstrated that an Economic Nonlinear MPC scheme can deliver a mathematically optimal control policy even when its underlying physics model is structurally wrong.
This works by using data-driven reinforcement learning to tune the MPC's cost functions and constraints online. If the internal model underestimates a physical penalty (e.g., actuator wear or energy loss), the reinforcement learning layer artificially inflates that penalty in the cost function. By distorting the optimization landscape, the controller is guided to select the true physical optimum despite relying on flawed internal physics.
True Plant Cost Landscape: [_____*_____] (Optimum at center)
Incorrect Model Landscape: [___*_______] (Predicts optimum shifted left)
RL-Adjusted Cost Function: [_______*___] (Artificially shifts landscape)
-----------------------------------------------------------------------------
Resulting MPC Action: [_____*_____] (Hits true plant optimum!)
Alternatively, when the model structure is correct but parameters are unknown, we can update the model iteratively. Mouhacine Benosman et al. combine MPC with model-free Extremum Seeking (ES) in Extremum seeking-based iterative learning linear MPC. By observing the real-world performance of the system across repeated tasks, the ES layer tunes the uncertain parameters online, systematically eliminating the tracking errors and constraint violations caused by static model errors.
Safe Exploration and Cautious Buffers
When a predictive controller must operate in an entirely unfamiliar environment, it faces a paradox: it must explore the system to learn an accurate model, but exploring with an inaccurate model risks violating safety constraints.
To resolve this, controllers must dynamically adjust their safety margins based on their level of uncertainty. Kim P. Wabersich addresses this in Cautious Bayesian MPC: Regret Analysis and Bounds on the Number of Unsafe Learning Episodes by introducing a framework that systematically "tightens" operational constraints when model confidence is low.
High Mismatch (Low Confidence) ---> Large Safety Buffers ---> Conservative Actions
Low Mismatch (High Confidence) ---> Small Safety Buffers ---> Aggressive/Optimal Actions
As the system collects data, the model-plant mismatch shrinks, allowing the controller to safely relax these buffers. By leveraging posterior sampling (Thompson sampling), Kim P. Wabersich et al. show in Bayesian model predictive control: Efficient model exploration and regret bounds using posterior sampling that this exploration-exploitation trade-off can be managed with provable, finite-time regret bounds, ensuring the system safely converges to optimal performance.
Key Foundational Work
A tool to analyze robust stability for model predictive controllers (1999) — contributed 2 concept elements to this direction.
A Tool to Analyze Robust Stability for Constrained MPC (2004) — contributed 1 concept elements to this direction.