AI Glossary / Deep learning / Intermediate
Feed forward (control)
Definition
Feedforward control determines a control action from a reference signal, a measured disturbance, or other predictive information, without relying on the resulting controlled-output error for that part of the action. It uses a model or known relationship to anticipate what input is needed. Feedforward is often combined with feedback, which corrects observed deviations. This control-engineering concept is distinct from a feedforward neural-network architecture.
What it means
A feedback controller responds after the measured output differs from its target. A feedforward component can act when a known command changes or a disturbance is detected, potentially before the output is substantially affected. For example, a measured increase in incoming material flow may justify a corresponding adjustment to processing input. Its success depends on knowing the relevant relationship well enough and obtaining the information in time.
How it works
The designer models how manipulated inputs and disturbances affect the controlled quantity. A feedforward rule then computes an action intended to achieve the reference or offset a measured disturbance. In dynamic systems, timing and response shapes matter as well as steady-state magnitudes. A feedback loop may be added to compensate for model errors, unmeasured disturbances, or changing conditions. Actuator limits and measurement quality constrain what compensation is physically possible. An ideal mathematical cancellation may be unrealizable when delays or dynamics prevent timely action.
Three practical examples
- Temperature processes: a controller adjusts heating in response to measured incoming flow or temperature changes, while feedback checks the actual controlled temperature.
- Motion control: a planned trajectory supplies expected actuator effort, with feedback correcting deviations from the desired motion.
- Process coordination: a known change in upstream throughput prompts a calculated downstream adjustment, provided the relationship and timing are sufficiently understood.
Think of it this way
Think of preparing an umbrella after seeing a reliable forecast, then adjusting it when actual rain and wind arrive. The anticipatory action resembles feedforward, while responding to observed conditions resembles feedback. The analogy is limited because engineered feedforward uses a defined signal relationship and must account for physical dynamics, not just general anticipation.
Worked example
Consider a simplified static process y = 2u + d, where u is the manipulated input, d is a measured additive disturbance, and y is the output. The desired output is r = 10. A feedforward rule is u = (r minus d)/2. If d = 2, this gives u = 4, and the modeled output is 2 times 4 plus 2 = 10. If the disturbance rises to 4, the rule changes u to 3, again giving the modeled output 10.
Now suppose the real process gain is 1.8 rather than the assumed 2. With d = 4 and u = 3, the actual output is 1.8 times 3 plus 4 = 9.4. Feedforward alone has left an error of 0.6 because its model is imperfect. A properly designed feedback component can respond to that discrepancy. This example ignores dynamics, noise, and actuator limits to isolate the principle; it is not a controller design for a real physical installation.
Illustrative case study
In a fictional heating-process project, engineers add a feedforward term based on an upstream flow sensor. A delayed sensor signal causes the compensating action to arrive after the disturbance has already affected the output. The team revisits timing and model alignment and retains feedback correction. The case illustrates that knowing a disturbance's magnitude is not enough when its arrival time and the actuator's response also matter.
Common misconceptions and mistakes
Feedforward does not guarantee perfect control or eliminate the need for feedback. It cannot directly compensate for every unmeasured disturbance. A reference-only action and a measured-disturbance compensator are different uses of the same general idea. Simply predicting an outcome is not a complete control strategy unless that prediction informs a feasible action. The word feedforward in a neural network describes information-flow structure, while feedforward in control describes how a control action is determined.
Business use
Use feedforward when relevant commands or disturbances are available early enough and their effects can be modeled with useful accuracy. Evaluate whether measurements are reliable and whether the actuator has enough authority to compensate. Test the combined feedforward-feedback behavior over the intended operating range. Monitor changes that invalidate the model. For business readers evaluating automation proposals, ask which signal triggers the anticipatory action and how the system detects and corrects a mismatch between predicted and actual behavior.
People also ask
Does feedforward require a neural network?
No. A simple physical model or calibrated relationship may provide the control rule.
Can feedforward and feedback operate together?
Yes. Feedforward can provide anticipatory action while feedback corrects observed errors.
Is feedforward always based on a forecast of the future?
No. It can use a currently measured disturbance or reference command whose effect is known before the controlled output responds.
Check your understanding
Why does the toy feedforward rule produce 9.4 instead of 10 when the real gain is 1.8?
Answer: The rule divides by the assumed gain of 2. The actual process responds less strongly, leaving a mismatch that feedforward alone does not observe and correct.
Related terms
- Feedback control
- reference signal
- disturbance
- process model
- actuator
- open-loop control
- tracking error
- model mismatch
- control system
Learn next
Solve the static toy equation for the required input and then change the assumed gain. Distinguish anticipatory action from error correction. Next, study how delays, actuator limits, and combined feedback affect the design of a practical control system.
More in Deep learning
- Exponential Function - An exponential function places the variable in the exponent of a fixed positive base.
- Exponential Smoothing State Space Model - An exponential smoothing state space model is a statistical time-series model that expresses exponential smoothing through equations for observations and...
- Extension neural network - An extension neural network, or ENN, is a specialized pattern-recognition approach that combines neural-network ideas with extension theory, also called...
- Feedback Network - In neural-network terminology, a feedback network contains connections through which information can return to an earlier part of the computation, creating...
- Feedforward neural network - A feedforward neural network is a neural network whose prediction computation follows directed connections without cycles.
- Fixed Effects Model - A fixed effects model is a statistical model that represents specified groups or factor levels with their own parameters.
Sources
Astrom and Murray, author-hosted Feedback Systems textbook: https://www.cds.caltech.edu/~murray/books/AM05/pdf/am08-complete_28Sep12.pdf Murray, original feedforward and feedback lecture material: https://www.kiss.caltech.edu/workshops/systems/presentations/murray.pdf Author textbook companion, controller structure: https://www.cds.caltech.edu/~murray/books/AM05/wiki/Frequency_Domain_Design.html
Previous: Feed forward chunking Next: Feed-forward network