Feedforward control

Feedforward control is a control strategy in which corrective action is taken before the effect of a disturbance is felt by the system, by using a predictive model of the disturbance or the system's response. Unlike feedback control, which measures the error between desired and actual output and corrects reactively, feedforward control anticipates the disturbance and counteracts it preemptively. The archetypal example is the thermostat that begins heating before sunset because it knows — from a schedule, a sensor, or a learned model — that the temperature is about to drop. The correction arrives before the deviation, not after it.

The concept is simple to state and difficult to implement. Feedforward control requires a model: either a model of the disturbance (so that it can be predicted and canceled) or a model of the system's response to the disturbance (so that the counteraction can be shaped precisely). If the model is accurate, feedforward can eliminate error before it occurs, achieving performance that no feedback loop can match — because feedback, by its nature, must wait for error to exist before it can correct it. If the model is inaccurate, feedforward can amplify the very disturbance it was meant to cancel, injecting energy into the system at precisely the wrong phase.

Feedforward and Feedback as Complementary Strategies

In practice, feedforward and feedback are not alternatives but complements. A well-designed control system typically uses feedforward for the disturbances it can predict and model, and feedback for everything else — the unmodeled dynamics, the sensor noise, the disturbances that were not anticipated. The LISA gravitational wave detector exemplifies this architecture: its time-delay interferometry (TDI) is a form of feedforward control that anticipates and cancels predictable laser phase noise, while lower-level feedback loops stabilize the test masses against unpredictable forces.

The complementary roles have a deep structural logic. Feedforward excels when the world is regular enough to be modeled; feedback excels when the world is irregular enough to surprise the model. Feedforward is the strategy of a system that trusts its understanding of the environment; feedback is the strategy of a system that doubts it. A system that relies solely on feedforward is brittle — a single unmodeled disturbance can destabilize it. A system that relies solely on feedback is sluggish — it must experience error before it can correct it, and the correction always lags behind the disturbance. The optimal controller is not the one that chooses between these strategies but the one that architectures them together.

This complementarity appears across domains. In physiology, allostasis — the predictive regulation of the body — is feedforward: the HPA axis adjusts cortisol in anticipation of predicted stress. Homeostasis — the reactive maintenance of setpoints — is feedback. The healthy organism uses both. In economics, forward guidance by central banks is feedforward: the bank communicates its future policy intentions so that markets can anticipate and price them. The actual interest rate adjustments are feedback. In both cases, the feedforward component shapes expectations; the feedback component corrects the errors that expectations inevitably produce.

The Architecture of Feedforward Control

The canonical feedforward controller consists of three elements: a disturbance model, a system model, and a compensator. The disturbance model predicts the future value of the disturbance based on available information — sensor readings, schedules, or learned patterns. The system model predicts how the system will respond to both the disturbance and the control input. The compensator computes the control input that will cancel the predicted effect of the disturbance on the output.

The design challenge is that the compensator must be the inverse of the system model — it must undo what the system model says the disturbance will do. If the system model is not invertible (because it contains delays, non-minimum-phase zeros, or saturation), exact feedforward compensation is impossible. The controller must then approximate the inverse, accepting some residual error that feedback will handle. This is the universal pattern: feedforward does what it can, and feedback cleans up what feedforward misses.

A special case is model predictive control (MPC), which generalizes feedforward to multiple time steps. MPC uses a system model to predict the future trajectory of the system over a receding horizon, computes the optimal sequence of control inputs that minimizes a cost function, and applies only the first input before recomputing. MPC is feedforward in the sense that it anticipates the future; it is feedback in the sense that it recomputes at every step based on the latest measurement. It is the synthesis of the two strategies, and it has become the dominant paradigm in process control, robotics, and autonomous systems.

Limits and Pathologies of Feedforward

The fundamental limit of feedforward control is model error. A feedforward controller cannot correct for disturbances it does not model, and it cannot correct for modeling error itself — because it has no measurement of the error it is producing. A thermostat that predicts a temperature drop at sunset will overheat the room if a cold front arrives unexpectedly. A self-driving car that predicts the trajectory of a pedestrian based on a model of human walking will fail when the pedestrian behaves in a way the model did not include.

This produces a characteristic pathology: model overconfidence. Systems with effective feedforward control can perform remarkably well in nominal conditions, creating the illusion that the model is complete. When an unmodeled disturbance finally appears, the system has no response repertoire — it has never needed one. The failure is sudden and catastrophic, not gradual. This is the pattern of the 2008 financial crisis: risk models that had performed well for years failed when correlations shifted in ways the models did not anticipate. The feedforward component — the pricing of risk based on predicted defaults — was accurate until it was not.

Another pathology is compensatory escalation. When a feedforward controller's model is slightly wrong, the natural response is to refine the model — to add more variables, more sensors, more data. This can produce systems of enormous complexity whose behavior is opaque even to their designers. The air traffic control system, modern supply chain management, and algorithmic trading platforms all exhibit this pattern: feedforward controllers that grew by accretion, each layer compensating for the limits of the layer below, until the whole system becomes a brittle lattice of interdependent predictions.

Feedforward control is the seductive promise of prediction made mechanical. It works beautifully when the world cooperates with the model, and it fails catastrophically when the world does not. The deepest error in control design is not choosing feedback over feedforward or vice versa — it is believing that a sufficiently sophisticated model can eliminate the need for feedback entirely. No model is complete. No prediction is certain. The system that forgets this, that trusts its feedforward loop so completely that it neglects its feedback loop, is not a system that has solved control. It is a system that has forgotten why control was necessary in the first place.