Blog · 2026-08-23
After Ziegler-Nichols: IMC, Cohen-Coon and the reaction curve
The reaction-curve methods that followed — Cohen-Coon for setpoint response, IMC and lambda for a loop you choose the speed of — and when each is worth the extra effort.
One premise, three answers
Every reaction-curve method starts from the same act: step the output in manual and fit gain, dead time and time constant to what comes back. Ziegler-Nichols, Cohen-Coon and IMC all read the same three numbers from the same bump test, and then disagree about what to do with them.
The disagreement is the interesting part. The three families of rules encode three different ideas about what a tuned loop is for, and choosing between them is mostly a question of which idea matches the job.
Cohen-Coon: fitted for setpoint, applied to everything
Cohen-Coon appeared in 1953, a decade after Ziegler-Nichols, from G.H. Cohen and G.A. Coon. Its formulas were fitted to give a fast, well-damped response to setpoint changes on an FOPDT process, with the three controller terms expressed as functions of the ratio of dead time to time constant.
The method is easy to apply by hand, which kept it in the textbooks. What travels with it less often is the premise: the fit assumes a specific disturbance shape and position, and tuning fitted to that shape is aggressive when the loop's real disturbances arrive from somewhere else.
On an operating plant, load regulation rather than setpoint response is usually the job. Cohen-Coon tuned for the wrong job is a loop that looks good on paper and fights its own process.
IMC and lambda: choose the speed, not the gain
Internal model control comes from a different tradition entirely. Instead of fitting a correlation, it derives the controller from the process model and a chosen closed-loop time constant — one knob, called lambda in the tuning literature, which sets how fast the loop is asked to be.
The appeal is that you stop guessing three numbers and start choosing one. Pick the closed-loop speed and the method hands you the rest, with a stability margin that degrades predictably as lambda is pushed down toward the dead time.
Lambda tuning is simply IMC applied to an FOPDT model, which is why the names travel together. The price of the single knob is a model: IMC running on a poor FOPDT is a confident method wrapped around a fiction, so the bump test that feeds it is not a formality.
When each one earns its place
Cohen-Coon is a commissioning tool for a self-regulating process when nothing is known and setpoint response is genuinely what matters. IMC and lambda are study and commissioning methods for loops whose model you can trust, and they shine where interacting loops need a predictable response rather than a fast one.
Ziegler-Nichols still earns its keep as the fastest way to a first estimate. None of the three should be the last word on a loop you intend to keep running, because none of them measured your loop — they assumed a model of one.
What this means on a plant
The distinction that matters day to day is between tuning by formula and tuning by observation. A formula tells you what the constants should be for the process it assumed; observation tells you what the loop you actually have is doing, which is a different thing.
PIDSnap works from the observation. The bump-test procedure gives you a printable routine for making the step cleanly, the pre-flight check catches structural problems before tuning is even on the table, and the guided sessions adjust the constants from the response that actually comes back.
Questions that come up
Which of these methods is the safest?
The safest is the one that assumes the least and measures the most. Lambda tuning with lambda set comfortably above the dead time is predictable and robust; the aggressive reaction-curve rules trade that margin for speed. Whatever the rule, verify against the actual response before it goes near a live setpoint.
Do I need a process model for lambda tuning?
Yes. Lambda tuning is IMC applied to an FOPDT model, so it needs the three numbers from a bump test — gain, dead time and time constant. If the process does not fit that shape, the model is a fiction and the tuning inherits it.