AI Glossary / Graphs and symbolic AI / Intermediate

Closed-loop

Definition

Closed-loop describes a system in which information about the system's behavior feeds back into decisions that influence its subsequent behavior. In control, measurements help a controller adjust actions toward a desired outcome. The defining feature is the feedback connection, not the presence of AI, a particular algorithm, or a guarantee of good performance.

What it means

An open-loop action follows a predetermined instruction without using the resulting output to correct that instruction. A closed-loop process observes what is happening and uses that observation in its next decision. For example, adjusting motion from measured position closes a loop that a fixed sequence of motor commands leaves open.
An AI model can participate as a predictor, controller, or estimator, but a conventional thermostat is also a closed-loop system. The term should not be confused with closed-world reasoning in logic. It also does not mean that a system is isolated from outside disturbances or that a human must review every action.

How it works

A typical feedback-control loop has a target, a measured output, a controller, and an actuator that changes the process. The controller uses available feedback, sometimes including an estimate of hidden state, to choose its action. The process responds, a new measurement arrives, and the cycle continues.
The behavior of the complete loop depends on the process, the decision rule, and the timing and quality of measurements. Delay can cause an action to respond to an outdated condition. Measurement noise can provoke needless corrections. Actuator limits can prevent a requested adjustment. Consequently, closing a loop creates a coupled dynamical system whose stability and performance must be evaluated, rather than automatically fixing an open-loop design.

Three practical examples

• A robot updates its steering from measured position instead of relying only on a precomputed sequence of turns.
• A computing service adjusts allocated resources using observed workload and response behavior, subject to limits and delays.
• A temperature controller changes heating from measured temperature, while a predictive model may help anticipate how quickly the room will respond.

Think of it this way

Think of filling a glass while watching the water level and slowing the pour as it approaches the mark. Watching and adjusting closes the loop. The analogy is limited because industrial or software feedback can be delayed, noisy, and distributed. Even a person can overshoot when the observation arrives too late or the adjustment is too large.

Worked example

Consider a simple simulated position x with target 10. At each step, choose a movement u equal to half the current error: u = 0.5 times (10 minus x). Assume the simulated process updates exactly as x_next = x + u, without noise, delay, or limits.
Starting at x = 2, the error is 8 and the first movement is 4, giving x = 6. The next error is 4, so movement 2 gives x = 8. Then movement 1 gives x = 9, followed by 0.5 giving x = 9.5. The error halves each step under these assumptions.
Now consider a gain of 3 instead of 0.5. The first movement is 24, taking x from 2 to 26. The next error is minus 16, so movement minus 48 takes x to minus 22. The errors grow in magnitude and alternate sign. Both designs are closed-loop; only the first has the illustrated convergent behavior. These are teaching calculations for a specified simulation, not tuning advice for a physical machine.

Illustrative case study

At fictional Cedar Compute, a team builds a demonstration that changes worker capacity in response to a queue. The initial rule reacts strongly to every new measurement. Because adding capacity takes time, several adjustments can be requested before earlier ones take effect.
The team adds a model of that delay to its simulation and reviews the complete feedback behavior rather than judging each adjustment alone. This fictional case illustrates why timing matters. It does not report a production deployment, a measured improvement, or a universal resource-control policy.

Common misconceptions and mistakes

Feedback does not automatically make a system stable, accurate, or robust. An overly aggressive response can amplify fluctuations. A system can also regulate the wrong quantity very effectively if the chosen measurement is a poor proxy for the actual objective.
Closed-loop operation is different from online learning. A controller can react to feedback while its parameters remain fixed; a learning component changes its model or policy. Human review is another separate design choice. Finally, merely collecting logs does not close the operational loop unless that information actually influences subsequent behavior within the defined process.

Business use

Businesses should specify what is measured, what can be changed, how quickly effects appear, and which outcome the loop should maintain. Evaluate disturbances and delayed responses before relying on a feedback policy. For AI-assisted workflows, record which model output affects an action and which observation later corrects it. This helps reviewers distinguish a useful feedback mechanism from a system that simply repeats its own recommendations without learning whether the intended result occurred.

People also ask

Does closed-loop require AI?

No. Many feedback controllers use fixed mathematical rules.

Is a dashboard a closed-loop controller?

Not by itself. Its measurements must feed into actions; a person can form part of that larger loop.

Does feedback eliminate error?

Not necessarily. The achievable error depends on the controller, process, disturbances, measurement quality, and constraints.

Check your understanding

Why does the gain-3 example count as closed-loop even though it diverges?

Answer: Its next action still depends on measured error. Closed-loop describes the feedback structure, while convergence describes a property of that particular loop.

Related terms

• Controller
• Stability
• Online learning
• Feedback control
• open-loop control
• actuator
• sensor
• disturbance
• control policy
• model predictive control

Learn next

Learn stability next. Derive the error update for the simulated process and examine which gains make the error shrink. Then introduce one step of measurement delay to see why the original calculation no longer fully describes the loop.

• Stability

More in Graphs and symbolic AI

• Circumscription - Circumscription is a formal approach to nonmonotonic reasoning that prefers models in which selected properties hold for as few entities as possible...
• Classical planning - Classical planning is the task of finding actions that transform a known initial state into a state satisfying a goal, under a simplified model with...
• Clause - A clause, in the logic used by many symbolic AI systems, is a disjunction of literals: statements joined by logical OR, where each statement is an atom or...
• Closed-world assumption - The closed-world assumption is a reasoning convention that treats relevant facts not established by the knowledge base as false.
• Closeness - Closeness, or closeness centrality, measures how near a node is to other nodes in a graph according to shortest-path distances.
• Clutter - Clutter, in tracking and data association, means detections that do not come from the target or targets of interest and instead create nuisance observations.

Sources

Primary sources: Astrom and Murray, Feedback Systems, author-hosted introduction, https://www.cds.caltech.edu/~murray/FBS/Introduction.html ; author-hosted preface and scope, https://www.cds.caltech.edu/~murray/FBS/Preface.html . The numerical recurrence and fictional example are original illustrations.

Previous: Closed-form matting Next: Closed-world assumption