Model Predictive Control

Model Predictive Control (MPC) is a control methodology in which a model of the system's dynamics is used to forecast future behavior over a finite horizon, and the controller selects the sequence of actions that optimizes a predicted objective — then executes only the first action, re-observes the system, and repeats. This receding horizon strategy makes MPC adaptive: it does not commit to a plan beyond the immediate step, and it can respond to disturbances, model errors, and unmeasured states as they arise.

MPC is the control-theoretic expression of a pragmatic epistemology: the model is not treated as a map of reality but as a tool for reasoning about possible futures. When the model is wrong, the controller discovers the error at the next observation and replans. This makes MPC robust to the kind of structured uncertainty that plagues classical control designs, which optimize for a single assumed dynamics and fail when the assumption breaks. The cost is computational: MPC requires solving an optimization problem at every time step, and the horizon must be long enough to capture relevant dynamics but short enough to solve in real time. The art of MPC design is the art of choosing what to ignore.

MPC dominates in process control, robotics, and power systems, where dynamics are slow enough to permit online optimization. Its extension to fast systems — autonomous vehicles, high-frequency trading, biological regulation — remains an active frontier. The deeper question is whether MPC is a general principle of intelligent action: do all adaptive systems, from brains to markets, implicitly solve prediction-and-optimization problems over receding horizons? The answer may be that MPC is not a specific algorithm but a universal structure — the structure of any system that acts on anticipated futures rather than remembered pasts.