Feedback is a process in which the output or consequences of a system return to influence that system’s subsequent behavior. It creates a closed causal loop: an action changes a condition, and the changed condition affects further action. Feedback occurs in engineered devices, living organisms, and environmental systems. It is a central concept in control theory and cybernetics, where interconnected processes are studied in terms of regulation, communication, and dynamic behavior. (fbswiki.org)
Structure of a feedback loop
An engineered feedback loop commonly contains a process being controlled, a sensor that measures its behavior, a controller that determines a response, and an actuator that applies that response. The measured output may be compared with a reference value, or set point. The difference between them is an error signal that guides corrective action. A vehicle’s cruise-control system, for example, measures speed and adjusts propulsion to reduce deviations from the requested speed. (fbswiki.org)
The defining feature is the return path, not the presence of an electronic controller or an explicit numerical target. Mechanical linkages and biochemical interactions can also close a loop. An open-loop system lacks this return of information about the controlled result. Feedforward control, by contrast, acts on a command or measured disturbance without waiting for its effect to appear in the output. Feedback and feedforward can operate together. (fbswiki.org)
Negative and positive feedback
Negative feedback counteracts an initiating change. If a regulated quantity rises, the loop produces an effect tending to lower it; if it falls, the response tends to raise it. This principle underlies much of biological homeostasis. “Negative” describes the direction of the response, not an unfavorable outcome, and does not mean that the regulated quantity must always decrease. (openstax.org)
Positive feedback reinforces an initiating change. An increase produces effects that promote further increase, while a decrease can promote further decrease. It can accelerate a transition or amplify a response. Positive feedback need not continue indefinitely: limiting resources, saturation, opposing processes, or a definite endpoint may terminate it. During childbirth, for example, cervical stretching promotes hormone release and stronger contractions, which produce further stretching until delivery ends the loop. (openstax.org)
The distinction between these signs does not by itself determine stability. Negative feedback can produce oscillations or instability when the response is too strong or poorly timed. Conversely, positive feedback can support bounded switching behavior rather than unlimited growth. An assessment therefore requires the dynamics of the complete loop, not simply the direction of one interaction. (cds.caltech.edu)
Engineering and historical development
A classical mechanical example is the centrifugal governor used with a steam engine. As the shaft rotates faster, rotating weights move outward and operate a linkage that reduces the steam supply. The resulting reduction in driving power opposes the speed increase. In his 1868 paper “On Governors,” James Clerk Maxwell analyzed such regulators mathematically, including conditions under which disturbances decay rather than produce growing oscillations. (en.wikisource.org)
In electrical engineering, feedback can make amplifier behavior less sensitive to component variations. More generally, it can reshape a system’s response, improve disturbance rejection, and stabilize otherwise unstable dynamics. These benefits depend on suitable sensing and controller design; returning a signal to an input does not automatically improve performance. (fbswiki.org)
Norbert Wiener brought feedback into a broad framework linking machines, organisms, and communication in Cybernetics, first published in 1948. The framework emphasized similarities in the organization of feedback systems across different physical implementations, rather than treating regulation as exclusively mechanical or biological. (mitpress.mit.edu)
Biological and climate systems
In physiology, feedback maintains variables within operating ranges rather than holding every quantity perfectly constant. Thermoregulation illustrates this: sensors detect temperature changes, and coordinating mechanisms activate responses that alter heat production or heat loss. The regulated condition remains dynamic, with fluctuations around an operating range. (openstax.org)
Blood-glucose regulation provides another example. Increased glucose concentration stimulates insulin release from the pancreas, promoting glucose uptake and storage. As glucose concentration falls, the stimulus for insulin secretion diminishes. This is a negative-feedback relationship between a measured physiological condition and a response that counteracts its change. (openstax.org)
In the climate system, feedback modifies the response to an initial forcing. The ice–albedo feedback occurs when warming reduces reflective snow or ice cover, exposing darker surfaces that absorb more sunlight and promote additional warming. Water-vapor feedback operates because warming generally increases atmospheric water vapor, strengthening the greenhouse effect. These mechanisms interact with other processes in Earth’s energy budget; an individual reinforcing loop does not imply unlimited warming. (science.nasa.gov)
Feedback in learning systems
In machine learning, feedback can take the form of evaluative information used to modify future behavior. In reinforcement learning, rewards guide changes to an agent’s policy. The feedback need not specify the correct action directly; it may instead indicate how desirable an observed outcome or behavior was. (arxiv.org)
Reinforcement learning from human feedback can use human comparisons between examples of behavior to train a reward model. An agent then learns using the model’s predicted rewards. Here, feedback is an evaluative signal within an iterative learning process, rather than necessarily a continuous physical measurement. Its interpretation depends on the preferences represented, the information supplied to the learner, and the way that information affects subsequent actions. (arxiv.org)