PID Control

concept
control-theorypidfeedback-loopscontroller

PID (Proportional–Integral–Derivative) is the most widely used closed-loop controller in industrial process control. Developed by Nicolas Minorsky in 1922 for automatic ship steering, it remains the dominant control strategy alongside Model Predictive Control (Wikipedia).

Three components

ComponentResponds toEffect
P (Proportional)Current errorLarger error → stronger correction. Fast but leaves steady-state error.
I (Integral)Accumulated past errorEliminates persistent error by increasing action over time. Can cause overshoot.
D (Derivative)Rate of change of errorDampens overshoot by opposing rapid changes. Sensitive to noise.

The controller output is the sum: u(t) = Kp·e(t) + Ki·∫e(t)dt + Kd·de(t)/dt

Tuning trade-offs

Each gain parameter creates a trade-off:

  • High Kp — fast response but oscillation risk
  • High Ki — eliminates steady-state error but slow to stabilize (integral windup)
  • High Kd — smooth approach but amplifies measurement noise

Classical controllers often require on-site tuning because mathematical models never perfectly match real systems — a core robustness concern.

Connection to the controller pattern

The controller pattern (watch-diff-reconcile) is a simplified PID where only the P component operates: the controller observes the diff and acts proportionally. Kubernetes controllers, for example, don’t accumulate error history (no I) or predict rate of change (no D). This simplicity is deliberate — for software reconciliation, the “plant” is deterministic enough that P-only control converges.

Beyond engineering

Management feedback can be read through the PID lens:

  • P: react to the current gap between target and actual (e.g., sprint velocity below forecast)
  • I: react to a persistent gap that hasn’t closed over multiple cycles (e.g., cumulative OKR shortfall)
  • D: react to the trend — if the gap is shrinking fast, reduce intervention to avoid overshoot

Most management frameworks operate P-only or at best PI. The derivative component requires frequent, reliable measurement — which is why shorter feedback cycles (xettel) enable better control.

Sources