CASE STUDY / CONTROL ROOM
Make a tuning decision inspectable.
A controller that looks good on one nominal process can behave differently when delay, dynamics or measurement quality change.
Faster.
But more resilient?
One process. Two controllers.
Make the trade-off visible.
The question
How can an interactive experiment show the difference between a fast initial response and a controller that performs consistently across plausible process conditions?
My design decisions
- Compare like with like
- Use identical process variations and random seeds for both controllers. Otherwise an apparent improvement could come from an easier test.
- Keep the interface responsive
- Run repeated simulations in a Web Worker and discard outdated responses when inputs change. Keep a numerical table alongside the graphic.
- Make the result portable
- Encode the parameters, baseline and study settings in a versioned URL. A reviewer can reproduce a result without creating an account.
The model
The simulation uses a unity-gain first-order process with dead time, a bounded actuator and a PID controller with derivative-on-measurement. The baseline equations are tested against an analytic proportional-control solution. Numerical convergence is checked separately.
A paired comparison
Pin a controller, then change its tuning. Both designs run on the same process. In the robustness study, they also receive identical random process variations and measurement noise. A fixed seed makes the comparison repeatable.
Across 32 synthetic trials, the median absolute error is 10.38 for the balanced preset and 7.23 for the aggressive one.
Settings: ±20% variation in time constant and delay; measurement-noise standard deviation 0.01; 180 s runs. These are synthetic model results, not operating-plant measurements.
Results you can reproduce
| Measure | Balanced | Aggressive |
|---|---|---|
| Startup overshoot | 0.0% | 4.2% |
| Rise time | 9.0 s | 4.9 s |
| Absolute error (this run) | 10.39 | 7.25 |
These values are computed from the model when the site builds. The faster response has to be weighed against overshoot, actuator limits and variation across the paired trials. There is no single score that substitutes for those decisions.
Why a Web Worker?
Repeated simulation should not block pointer, keyboard or input events. The main thread owns the controls and SVG. A separate worker returns versioned results; response IDs prevent an older request from overwriting a newer result.
What the evidence does not establish
The model omits actuator nonlinearities, coupled loops and plant safety constraints. Trial percentiles describe the chosen synthetic conditions. They do not establish confidence bounds or safe tuning for real equipment.
Reproduce the experiment →Terms used on this page
- PID controller
- The most common feedback controller: it adjusts an output using the present error, its accumulated history and how fast it is changing.
- Dead time
- The delay between a control action and the first sign of its effect on the measured process.