Model Predictive Control
Model predictive control (MPC) is an advanced control strategy that uses a dynamic model of the system to predict future behavior and optimize control actions over a finite horizon. At each time step, the controller solves an optimization problem to find the best sequence of future actions, applies the first action, and repeats the process — a receding horizon approach that allows the controller to handle constraints, nonlinearities, and changing objectives. MPC is the natural partner to system identification: the better the model, the better the predictions, and the better the control. The systems-theoretic significance is that MPC treats control as an ongoing conversation between the model and the world, rather than a one-time design. The controller that cannot update its model is not a controller but a monument to an outdated assumption. See also: Control Theory, Optimization, Adaptive Control\n\n\n\n== Mathematical Foundations ==
MPC is an instance of dynamic programming applied to continuous or hybrid systems. At each time step, the controller solves a constrained optimization problem over a prediction horizon N:
minimize Σ (x_t^T Q x_t + u_t^T R u_t) + x_N^T P x_N subject to x_{t+1} = f(x_t, u_t), x_t ∈ X, u_t ∈ U
where Q and R are weight matrices encoding the controller's priorities, P is a terminal cost matrix ensuring stability, and X and U are constraint sets. The receding horizon principle — applying only the first control action and re-solving — is what distinguishes MPC from open-loop optimal control.
The optimization problem is typically solved via quadratic programming for linear systems, or nonlinear programming for general dynamics. The computational cost grows with the prediction horizon and the state dimension, creating a tension between prediction accuracy and real-time feasibility. This is why receding horizon control is not merely a philosophical choice but a computational necessity: a controller that cannot solve its optimization problem within the sampling time of the system is not a controller but a delay.
Systems-Theoretic Significance
MPC embodies the systems-theoretic insight that control is not a design problem but an ongoing inference problem. The controller maintains a model of the system, generates predictions, compares them to desired trajectories, and updates its plan. This is structurally identical to the predictive processing framework in neuroscience, where the brain minimizes prediction error through hierarchical inference. The MPC controller is, in this sense, a mechanical implementation of the free energy principle — a system that treats its own actions as hypotheses to be tested against the world.
The connection is not merely metaphorical. Both MPC and predictive processing face the same structural challenge: the model is never the territory, and the system must balance exploitation (acting on the current best model) with exploration (perturbing the system to improve the model). In MPC, this trade-off appears as the dual control problem: the controller must simultaneously regulate the system and excite it sufficiently to maintain model accuracy.
Economic and Social Applications
MPC has migrated beyond engineering into economics, where economic MPC treats policy as a constrained optimization problem over a finite horizon. Central banks, energy markets, and supply chain operators use MPC-like frameworks to optimize decisions subject to dynamic constraints. The European Central Bank's macroeconomic models, for instance, are essentially large-scale MPC systems with thousands of state variables and decades-long horizons.
The social implications are significant. When a social system is governed by MPC-like optimization — whether algorithmic content moderation, predictive policing, or dynamic pricing — the controller's objective function becomes a political object. The choice of Q and R in the MPC formulation is not a technical detail but a normative commitment: it encodes what the system values, what it tolerates, and what it ignores. A controller that optimizes for engagement will produce a different society than one that optimizes for deliberation. The mathematics is neutral; the objective function is not.
Robustness and Computational Limits
Classical MPC assumes perfect model knowledge. Robust MPC relaxes this assumption by optimizing over sets of possible models rather than a single model. The controller computes a policy that is guaranteed to satisfy constraints for any model in the uncertainty set — a conservative but safety-critical approach. The cost is performance: robust MPC is typically more conservative than nominal MPC because it must hedge against worst-case scenarios.
The computational limits of MPC are becoming more acute as systems scale. Solving a quadratic program at millisecond timescales for a system with thousands of states is beyond the reach of current hardware. The field has responded with explicit MPC (pre-computing the optimal policy offline) and approximate MPC using machine learning. These approximations sacrifice optimality guarantees for speed, raising a fundamental question: is a fast suboptimal controller better than a slow optimal one?
MPC is not merely a control technique. It is a philosophical stance: the claim that the best way to govern a complex system is to continuously predict its future, optimize within constraints, and accept that the plan will be revised before it is ever fully executed. This is the epistemology of receding horizons — a recognition that certainty is a luxury no controller can afford, and that the only rational response to uncertainty is to keep predicting, keep optimizing, and keep moving.