ENGINEERING NOTE

Feedback is a trade-off, not a magic trick.

A small model makes the tension between speed, disturbance rejection and actuator limits visible. The tuning that looks best on paper is the first to struggle when the process is not quite what you assumed.

Low Kah Yi · 27 September 2026 · Updated 25 October 2026

A feedback controller compares a measurement with a target and turns the difference into an action. That loop has a difficult job: it has to correct deviations quickly without amplifying every imperfection in the measurement, the actuator or its own picture of the process. Every tuning choice moves that balance somewhere. None of them removes it.

Three terms, three different jobs.#

Proportional action responds to the error now. On its own it leaves an offset, because it needs an error to produce any output at all. In Control Room, removing integral action from the balanced controller leaves the process settled about a third short of its target, and it stays there.

Integral action accumulates error over time, so it keeps pushing until the offset is gone. That is what makes it useful, and also what makes it dangerous near an actuator limit (more on that below).

Derivative action responds to how fast the measurement is changing, which can damp an approach before it overshoots. It also responds to noise, because noise is fast change. Control Room filters it, and takes it from the process measurement rather than the error, so a change of target does not produce a sudden kick. The presets leave it at zero.

Fast is not the same as good.#

The lab has three presets. On the default process, with a 12-second time constant and a 2.5-second transport delay, they compare like this:

• Conservative: no overshoot, 46 s to rise, not settled within 80 s, total absolute error 24.6.
• Balanced: no overshoot, 9 s to rise, settled in 44 s, total absolute error 10.4.
• Aggressive: 4% overshoot, 5 s to rise, settled in 31 s, total absolute error 7.3.

On this process, the aggressive controller wins every column. If the model were exact, that would be the end of the discussion.

Now change the process.#

Double the transport delay to 5 seconds, which is the kind of error a real process model can easily carry, and run the same controllers again.

• Balanced: 16% overshoot, settled in 54 s, total absolute error 13.2. Worse, but still a controlled response.
• Aggressive: 44% overshoot, and it never settles. The process keeps swinging between 0.66 and 1.34 of its target for the rest of the run, the actuator spends about 77 seconds pinned at a limit, and the total absolute error rises to 42.1.

The aggressive tuning was fast because it relied on the process being exactly as modelled. That reliance is the cost of its speed. A less aggressive controller gives up some nominal performance in exchange for margin: room for the delay, gain or time constant to be different from what was assumed. Choosing a tuning means choosing how much of that margin to keep.

The actuator has limits.#

A control law is linear until a valve, heater or motor reaches the end of its travel. After that, the controller is asking for something the process cannot deliver. If integral action keeps accumulating error during that time, the controller winds up a large internal demand, and has to unwind it before it can respond in the other direction. The result is a long, late overshoot.

Control Room uses conditional integration: it stops accumulating whenever accumulating would push an already saturated output farther into its limit. That prevents windup, but it does not create capacity. In the delayed aggressive case above, the actuator still spends most of the run on a limit, because the tuning keeps demanding more than the process can follow.

A useful experiment.#

1. Open Control Room with the balanced preset and pin it as the comparison.
2. Select the aggressive preset. Compare the responses, including the load disturbance at 90 seconds.
3. Raise the transport delay from 2.5 to 5 seconds and watch both controllers again.
4. Turn on the robustness study. It runs both controllers across 32 synthetic process variations with identical noise, which shows which tuning holds up when the process is uncertain rather than which one wins on a single run.

A shorter rise time is not, by itself, evidence of a better controller. The question is how the controller behaves on the process you actually have, which is never quite the one you modelled.

Know the boundary.#

This is a synthetic first-order process with dead time, a unit setpoint and a bounded actuator. It can add measurement noise, but it excludes nonlinear valves, sampling jitter, interacting loops and plant safety constraints. Its purpose is to build intuition about the trade-offs, not to prescribe tuning for operating equipment.