Receding Horizon Control

Receding horizon control (RHC) is the implementation principle underlying model predictive control: at each decision point, a controller plans over a finite future horizon, executes only the first action of that plan, and then re-plans from the new state. The horizon recedes because the planning window moves forward in time, never reaching the originally predicted endpoint.

The principle is not merely computational pragmatism. It is a formal recognition that long-term prediction is epistemically unreliable — the further a model projects, the more its errors compound. By re-planning at each step, the controller incorporates the latest observations and discards predictions that have already failed. This is the control-theoretic analogue of the Bayesian principle that priors should be revised when new evidence arrives.

RHC transforms an open-loop optimization problem into a closed-loop feedback policy. The cost is computational: the controller must solve an optimization problem at every time step. The benefit is adaptivity: the controller never commits to a plan it cannot revise. In systems theory, RHC represents the operationalization of a fundamental principle: that intelligent control requires not just prediction but the willingness to abandon prediction on contact with reality.