Is a thermostat a controller?

Yes. It uses temperature information to influence heating or cooling.

Does control require full knowledge of the internal state?

No. Many designs use outputs or estimated states, subject to observability and design assumptions.

Is stability the same as optimality?

No. Stability concerns behavior near or around a specified regime; optimality concerns an objective under stated constraints.

AI Glossary / Graphs and symbolic AI / Intermediate

Control theory

Definition

Control theory studies how to influence a changing system so that its behavior meets specified objectives. It develops models, analysis methods and controllers for tasks such as maintaining a target value, following a trajectory and responding to disturbances while respecting practical limits.

What it means

A controller chooses inputs to a system, often called the plant. The plant could be a motor, room-temperature process or computing service. Sensors provide measurements, and the controller uses those measurements or a model to determine what to do next.

Feedback closes the loop: measured results influence subsequent actions. Open-loop control follows commands without correcting them from the observed output. Feedback can compensate for disturbances and modeling errors, but poor design can also amplify noise or create oscillations.

The objective is more precise than simply make the error small. A design may need acceptable settling time, limited overshoot, bounded actuator effort and robustness to uncertain parameters. A stable response can still be too slow or inaccurate, while a fast response can be unstable outside a narrow test condition.

Control theory overlaps with AI in robotics, decision making and reinforcement learning. Many effective controllers, however, use mathematical models and designed laws without machine learning.

How it works

Start by defining the controlled quantity, available inputs, measurements and operating range. Model how the system changes over time, including relevant delays and disturbances. The model may use differential equations, difference equations or an identified input-output relationship.

Choose a controller structure and analyze the resulting closed-loop behavior. A proportional controller acts on current error. Integral action can address persistent offset under suitable conditions, while derivative action responds to change and requires care with noisy measurements. More advanced methods can use state estimates or optimize future actions over a horizon.

Testing must include limits. An actuator cannot supply unlimited force or computing capacity, and a sensor may be delayed or biased. A controller designed around impossible actions can fail even if its unconstrained equations look favorable.

Robustness asks how well the design tolerates departures from its nominal model. Stability analysis, simulation and physical or operational tests provide different kinds of evidence. Changes to sampling time, plant dynamics or the measurement pipeline can alter the behavior of the complete loop.

Three practical examples

A motor controller adjusts electrical input to track a requested speed despite changes in load.
A building controller varies heating or cooling to keep temperature near a target while accounting for equipment limits.
A computing-service controller adjusts worker capacity in response to queue measurements, considering startup delay and resource limits.

Think of it this way

Think of steering a trolley towards a line while repeatedly looking at its position. Corrections based on observation are feedback; issuing a fixed sequence of pushes without looking is open-loop action.

The analogy does not guarantee that more frequent or stronger correction is better. Delayed observations, momentum and excessive force can cause overshoot or alternating corrections, which is precisely why loop dynamics matter.

Worked example

A proportional controller in a simple discrete model

Use a deliberately simplified system x(t+1) = x(t) + u(t), where x is position and u is the position change applied during one step. The target is r = 10. Choose u(t) = K × (10 - x(t)), with no disturbance or actuator limit in this teaching model.

With K = 0.5 and x(0) = 0, the first input is 5 and x(1) = 5. The next input is 2.5, giving x(2) = 7.5. The next position is 8.75. The error halves each step.

Writing e(t) = 10 - x(t) gives e(t+1) = (1 - K)e(t). Error converges to zero when |1 - K| < 1, meaning 0 < K < 2 for this model.

At K = 2, positions alternate between 0 and 20 rather than settling. At K = 2.5, the error alternates and grows because its multiplier is -1.5. A larger gain therefore does not universally improve control. The interval derived here belongs to this simple equation; delays, different plant dynamics or saturation can change it.

Illustrative case study

Fictional case: Rowan Services adjusts background-worker capacity from queue length. A first prototype adds many workers whenever the queue grows, but new workers take time to start. The controller repeatedly reacts before earlier changes take effect, creating large swings in capacity.

The team models the startup delay, limits adjustment size and evaluates response to changing workloads. It tracks both queue delay and resource consumption rather than judging the controller by one short recovery example.

The lesson is that feedback operates through a dynamic system. A reasonable immediate action can become unreasonable when repeated without accounting for delayed effects.

Common misconceptions and mistakes

Feedback automatically guarantees stability. The combined controller and plant determine stability.
Higher gain always means better performance. It can increase oscillation, noise sensitivity or instability.
A good model eliminates the need for testing. Models omit effects and require validation.
Control theory and reinforcement learning are identical. They overlap but use different methods and assumptions.
Reaching a target once establishes good control. Sustained response to disturbances and operating changes also matters.

Business use

Control theory offers a disciplined way to design systems that continually act on measurements. Its practical value is making objectives, delays, constraints and feedback effects explicit.

Define success using response measures that reflect the service: tracking error, recovery time, resource use or disturbance rejection. Record the conditions under which a controller was analyzed and tested. When AI supplies predictions or policies inside a loop, evaluate the whole loop rather than only prediction accuracy. A small systematic prediction error can have effects that accumulate through repeated action.

Check your understanding

In the worked model, K = 2 gives error multiplier -1. Does the error decay?

Answer: No. Its sign alternates while its magnitude stays constant, so the position oscillates rather than converging.

Related terms

Feedback
Stability
Robustness
PID controller
closed-loop control
open-loop control
state estimation
model predictive control

Learn next

Calculate four steps of the worked model for K = 0.5, 1.5 and 2.5. Compare monotone convergence, alternating convergence and divergence. Then introduce a one-step measurement delay and explain why the original stability calculation no longer describes the modified system.

Stability

Sources

Richard Murray, Caltech introduction to control theory, describing feedback design, stability and performance: https://www.cds.caltech.edu/~murray/courses/cds110/

Åström and Murray, Feedback Systems, author-hosted text: https://www.cds.caltech.edu/~murray/books/AM05/pdf/fbs-public_29Jun19.pdf

The discrete controller calculation and fictional service example are original illustrations, not deployment specifications.

Previous: Control limits Next: Control variates