Feedforward Control

Feedforward control is a regulatory strategy in which a system adjusts its behavior based on predicted future disturbances rather than reacting to present errors. Unlike feedback control, which corrects deviations after they occur, feedforward control anticipates perturbations and acts to prevent them or compensate for them before they affect the system's output. The strategy requires a model of the disturbance — its timing, magnitude, and effect on the system — and the capacity to generate a compensating action that arrives synchronously with the disturbance itself.

The concept is central to cybernetics, control theory, and systems theory, and it appears across scales from molecular biology to global economics. In physiology, the hypothalamus initiates thermoregulatory responses before body temperature actually changes — detecting cold ambient conditions and triggering vasoconstriction and shivering in anticipation. In engineering, a cruise control system that slows the vehicle when it detects an approaching hill is using feedforward; one that slows only after the vehicle has already decelerated is using feedback. In economics, anticipatory monetary policy — raising interest rates before inflation materializes — is feedforward regulation of the economy.

The Architecture of Feedforward

A feedforward controller requires three components that a feedback controller does not:

1. A disturbance sensor. The system must detect the perturbation before it reaches the regulated variable. The sensor need not be a separate physical component; it can be an internal model that predicts disturbance from context. In the allostatic regulation of the HPA axis, the "sensor" is the hippocampal memory of past stressors that modulates the hypothalamic set point.

2. A predictive model. The system must know — or have learned — how the disturbance will affect the regulated variable. This model need not be explicit or conscious. In the simplest case, it is a hardwired reflex: the pupillary response to light onset is feedforward, with the retina's photoreceptors serving as the disturbance sensor and the iris's motor circuits as the model. In complex cases, the model is a learned representation: a chess player who sees a threat and prepares a defense before the threat materializes is using an internal model of the opponent's intentions.

3. A compensating actuator. The system must be able to generate an action that counteracts the predicted effect. The timing is critical: the compensation must arrive when the disturbance arrives, not before (which wastes resources) and not after (which is feedback). The temporal matching of prediction and compensation is what distinguishes feedforward from simple preemptive action.

Feedforward and Feedback: The Dual Architecture

No real system relies exclusively on feedforward or feedback. The two strategies are complementary, and their combination is the architecture of sophisticated regulation.

Feedback is robust but reactive. It corrects errors, but it cannot prevent them. It requires no model of the disturbance, which makes it applicable in unknown environments. But it is always late: the error must occur before correction begins. Feedback controllers also suffer from the waterbed effect: correcting one frequency of disturbance often amplifies another.

Feedforward is proactive but fragile. It prevents errors, but only when the disturbance is predictable and the model is accurate. If the model is wrong — if the disturbance is larger than expected, or different in kind, or arrives at the wrong time — feedforward compensation can make the error worse. A system that overcorrects for a predicted disturbance that does not materialize introduces an error that did not previously exist.

The optimal architecture combines both: feedforward for the disturbances you can predict, feedback for the disturbances you cannot. The feedforward component handles the expected, leaving the feedback component to handle the residual. In adaptive control, the feedback component also serves to update the feedforward model: the prediction errors that escape feedforward correction become training signals for the model. This is precisely what happens in actor-critic learning: the critic's prediction errors update the actor's policy, which is a feedforward controller whose internal model is continuously refined by feedback.

Biological Feedforward: The Fast and the Slow

Biological systems implement feedforward at multiple timescales, creating a hierarchy of anticipatory regulation:

Fast feedforward — reflex arcs, startle responses, predictive saccades. These operate on millisecond timescales and use hardwired models. The vestibulo-ocular reflex, which stabilizes images on the retina during head movement, is a feedforward controller: it predicts the consequences of head motion and generates compensatory eye movements before retinal slip occurs.

Intermediate feedforward — circadian rhythms, hormonal anticipation, immune priming. These operate on timescales of hours to days and use learned or evolved models. The circadian clock anticipates light-dark cycles and pre-adjusts metabolic, cognitive, and immune parameters before the transitions occur. The immune system's memory of prior pathogens is a feedforward model: it anticipates re-exposure and pre-arms the response.

Slow feedforward — developmental programming, evolutionary adaptation, cultural anticipation. These operate on timescales of years to millennia and use models encoded in genes, epigenetic marks, and institutional knowledge. The seasonal timing of flowering in plants is feedforward based on photoperiodic models encoded in the genome. The construction of flood defenses is feedforward based on historical models of river behavior.

The hierarchy is not merely a list. It is a nested architecture in which slower feedforward mechanisms set the context for faster ones. The circadian clock does not directly control muscle reflexes, but it modulates the metabolic state in which reflexes operate. Evolution does not directly control immune memory, but it shapes the repertoire of possible memories. The timescales are coupled, and the coupling is what makes biological regulation adaptive across the enormous range of temporal variation in the environment.

The Limits of Feedforward

Feedforward control faces a characteristic set of failure modes that derive from its dependence on prediction:

Model error — If the internal model of the disturbance is inaccurate, the compensation will be wrong. This is the problem of model lock in anticipatory systems: a system that continues to anticipate a world that no longer exists.

Timing error — If the compensation arrives too early or too late, it fails to synchronize with the disturbance. Early arrival wastes resources; late arrival is feedback with a delay.

Compensation saturation — If the disturbance exceeds the range the feedforward controller was designed to handle, the compensation is insufficient and the feedback controller must manage the residual. This is why biological systems maintain both anticipatory and reactive stress responses: the feedforward HPA axis prepares for predicted stress, but the sympathetic nervous system reacts to actual stress that exceeds the anticipation.

Interference — In systems with multiple feedforward controllers, the compensations can interfere. One controller's compensation may be another controller's disturbance. This is the problem of coordination in complex systems: the feedforward strategies of different subsystems must be aligned, or they will fight each other.

Feedforward and the Free Energy Principle

In Karl Friston's Free Energy Principle, feedforward and feedback are reinterpreted as two modes of inference. Feedforward connections in the brain carry prediction errors — the mismatch between expected and observed sensory states. Feedback connections carry predictions — the internal model's expectations about what should be observed. The sensory cortex is not primarily a feature detector; it is an error detector, comparing incoming signals to predictions and passing the residual upward.

This inversion of the traditional view has profound implications. What appears to be a feedforward sensory pathway is actually a feedback prediction pathway, and what appears to be a feedback motor command is actually a feedforward prediction of proprioceptive consequences. The brain, on this view, is not a passive receiver of sensory data and an active sender of motor commands. It is a predictive machine that generates its own expectations and uses sensory input only to correct them.

The feedforward controller in this framework is not a separate component but the prediction itself. To act is to predict the sensory consequences of action and to minimize the prediction error that would result if the action were not performed. The goal-directed behavior that cybernetics studied as feedback regulation becomes, in the Free Energy Principle, a form of self-fulfilling prophecy: the system predicts a future state and acts to make that prediction true.

The feedforward controller is the optimist of the regulatory world: it believes it can see the future, and it acts on that belief. Sometimes it is right, and the system is flawless. Sometimes it is wrong, and the error is worse than if it had never predicted at all. The wisdom of regulation is knowing when to anticipate and when to wait.