Every shower tap teaches the same lesson: a system that reacts late cannot be corrected hard. That is the whole of control engineering

Stand under a shower with a single mixer tap in an old building and you will re-derive most of control theory in about ninety seconds. The water is cold. You turn the tap. Nothing happens, because the hot water is still travelling down several metres of pipe. You turn it further. A moment later it is scalding, so you turn it sharply back, and a moment after that it is cold again. Two or three cycles later you settle, having learned to make smaller adjustments and wait.

That experience is not a plumbing defect. It is the central problem of feedback control, and the same pattern produces a geyser that overshoots its setting, an air conditioner that cycles noisily between too cold and too warm, and a motor that hunts around its target speed instead of holding it.

Feedback is easy to describe and hard to do

An open-loop system acts without checking the result: a fan set to speed three, a timer that runs a pump for ten minutes. Simple, predictable, and blind — if something changes, the system neither knows nor cares.

A closed-loop system measures what actually happened and uses the difference to decide what to do next. A sensor reports the current value, the controller compares it with the target to get an error, and it drives an actuator to reduce that error. Thermostats, cruise control, voltage regulators, drone stabilisation and industrial process control are all this same loop.

Described that way it sounds trivially easy: measure the gap, close the gap. The difficulty is entirely in the word next, because the world does not respond instantly.

Delay is what makes correction dangerous

Here is the key idea, and it is worth stating plainly because it is not obvious: instability comes from lag, not from lack of power.

Suppose a system responds immediately. Then you can correct as hard as you like — the moment the error disappears, your correction stops mattering, and nothing overshoots.

Now suppose the system responds slowly. You apply a correction, see no change, and apply more. The original correction is still working its way through, and by the time both arrive the error has already been eliminated — so the extra effort pushes the system past the target in the other direction. Now the error has reversed sign and the controller corrects the other way, late again. That is an oscillation, and it is generated by the controller trying to help.

Every part of a real system adds lag. Thermal mass takes time to heat. Water takes time to travel down a pipe. A motor takes time to accelerate. A sensor takes time to register a change. In digital systems, the processing interval itself adds delay. The more lag a loop contains, the gentler its corrections must be, which is the engineering version of the lesson the shower taught: when the response is late, make smaller adjustments and wait.

A controller that reacts strongly to a system that reacts slowly does not control it. It drives it.

What the three classic terms actually do

The standard controller combines three ways of looking at the error, and each buys something at a cost.

Proportional action makes the correction proportional to the current error: the further off target, the harder the push. On its own it has two characteristic failures. Set the gain too low and the system never quite arrives — there is a permanent offset, because a small remaining error produces only a small correction, which may be just enough to balance whatever is pulling the system away. Set the gain too high and you get the shower: oscillation, and eventually instability.

Integral action accumulates the error over time, so even a small persistent offset eventually builds into a correction large enough to eliminate it. This is what removes the steady-state error that proportional action leaves behind. Its cost is that accumulation is itself a form of lag, which makes the loop slower to settle. It also has a famous failure mode — integral windup — where the actuator is already at its limit but the integral term keeps accumulating, so when the system finally responds there is an enormous stored correction to unwind, and it overshoots badly. A geyser that has been off at the wall, then switched on, can show exactly this.

Derivative action responds to how fast the error is changing rather than how large it is, which lets the controller ease off before it arrives at the target — the damping that stops overshoot. Its cost is noise: differentiation amplifies it, so a jittery sensor signal becomes a violently twitchy output, which is why derivative terms are often filtered or left out entirely in noisy applications.

Tuning these three against each other is a genuine skill and the reason control engineering is a discipline rather than a formula.

Why a thermostat is deliberately imprecise

A household thermostat does something that looks like a defect and is a design decision: it does not switch at a single temperature. It turns the cooling on at one value and off at a different one, a gap called hysteresis or a deadband.

Without it, a system sitting exactly at the set point would switch on and off continuously as the reading drifted by hundredths of a degree — chattering the relay, and in a refrigerator or air conditioner short-cycling the compressor, which is the fastest way to destroy it. The deadband trades a little temperature accuracy for a great deal of mechanical life.

This is bang-bang control: full on or full off, nothing in between. It is cheap, robust and entirely adequate for a fridge, and it is why the temperature in any on-off system is not held at a value but swings slowly around it.

The alternative is to vary the output continuously, which is precisely what an inverter air conditioner or refrigerator does. Instead of running the compressor flat out and then stopping it, it varies compressor speed to match the actual heat load. That is proportional control replacing on-off control, and it produces steadier temperature, less noise, less mechanical stress and lower energy use — the last because a compressor running continuously at partial load is more efficient than one repeatedly starting at full load. The efficiency gain marketed on the label is, underneath, a change in control strategy.

The bug is often the sensor, not the controller

One practical lesson from industrial practice deserves wider circulation: a great many control problems are measurement problems wearing a disguise.

A sensor in the wrong place adds lag to the loop, and lag is the enemy. A thermostat mounted on a wall in direct sun, near a doorway, or above a heat-producing appliance is not measuring the room. A temperature probe placed far downstream of a heater reports what happened rather than what is happening. In each case the controller is behaving correctly on bad information, and no amount of retuning fixes it — the loop has to be shortened.

The complementary trick is feedforward. Feedback is inherently reactive: it cannot act until an error exists. But if a disturbance is predictable, it can be countered before it shows up. Cruise control that uses map data to increase power as a hill begins, rather than waiting to slow down, is acting on feedforward. So is a furnace that increases output when it sees cold material entering. In practice, good systems combine the two: feedforward handles what can be anticipated, feedback cleans up everything else.

Why it matters for students and researchers

Control theory is one of the most portable bodies of knowledge in engineering. The same mathematics governs a chemical reactor, a robot arm, a power grid holding its frequency, a drone staying level, a biological regulatory pathway and an economic policy response. Students frequently meet it as a course full of transfer functions and never see that it is a general theory of acting under delay and uncertainty.

The applied questions are active. Embedded and real-time implementation determines whether a controller that is stable on paper is stable in a device, because sampling rate, jitter and computational delay all enter the loop. Tuning in the field, where models are approximate and plants change with age and load, remains substantially an empirical craft and a good subject for careful study. And the newer question — how far learning-based controllers can be trusted where stability must be guaranteed rather than usually achieved — is genuinely open, and matters most in exactly the applications where failure is expensive.

That coverage of automatic control systems, including embedded control, real-time systems and both digital and analog control, is the stated scope of the International Journal of Advanced Control and System Engineering (ISSN 3049-1452), a peer-reviewed journal launched in 2015. For electrical, electronics and instrumentation students, feedback is worth understanding beyond the exam, because once seen it is visible everywhere — and because the discipline's central lesson generalises well past machinery: when the effect of an action arrives late, react in smaller steps than instinct demands.

Frequently asked questions

Why does a geyser or air conditioner overshoot its set temperature?

Because the system responds with a delay. The controller keeps applying correction while earlier correction is still taking effect, so by the time everything arrives the target has already been passed, and the error reverses.

What causes oscillation in a control system?

The combination of delay and strong correction. If a system reacts slowly, a controller that responds aggressively will repeatedly overshoot in alternating directions. Reducing gain or adding damping stabilises it.

What do the P, I and D terms do?

Proportional corrects in proportion to the current error but can leave a permanent offset or oscillate if too strong. Integral accumulates past error to remove that offset but adds lag and can wind up. Derivative responds to the rate of change to damp overshoot but amplifies sensor noise.

Why does a thermostat switch on and off at different temperatures?

That gap is deliberate hysteresis. Without it the system would switch continuously around the set point, chattering the relay and short-cycling the compressor, which rapidly damages it. A small loss of precision buys a large gain in equipment life.

Why are inverter air conditioners more efficient?

Because they vary compressor speed to match the heat load rather than switching fully on and off. That is proportional control instead of on-off control, which gives steadier temperature, less mechanical stress and better efficiency at partial load.

What is feedforward control?

Acting on a predictable disturbance before it produces an error, rather than waiting to detect the error. Increasing engine power as a hill begins, rather than after the vehicle slows, is feedforward; it is normally combined with feedback, which corrects whatever was not anticipated.