Control Theory

Adaptive Control

Adaptive controllers adjust their own parameters online as the plant changes, maintaining performance without a fixed model.

Controllers That Learn Their Plant

Adaptive control addresses plants whose parameters are unknown or change over time. Instead of a fixed controller tuned for one model, an adaptive controller adjusts its own gains online, using measured data, to track the plant as it evolves. It is control and identification running together in real time.

Two main families

Kronos motion — control room

Direct adaptive control: the controller parameters are adjusted directly to drive a performance error to zero, without an explicit plant estimate.
Indirect adaptive control: the plant parameters are estimated online, and the controller is recomputed from the current estimate at each step.
Model-reference adaptive control (MRAC) is a leading direct method; self-tuning regulators are a leading indirect method.

Model-reference adaptive control

In MRAC, a reference model specifies the desired closed-loop behavior. An adaptation law adjusts the controller parameters to drive the difference between the actual output and the reference model's output toward zero. The MIT rule and Lyapunov-based adaptation laws are common; the Lyapunov approach is preferred because it comes with a stability proof.

The stability challenge

Adaptive control is powerful but delicate. Because the controller parameters change while the loop runs, standard fixed-gain stability analysis does not apply, and Lyapunov methods are used to prove stability of the combined system. Poor design can cause parameter drift, bursting, or instability, especially under disturbances or unmodeled dynamics. Robustness modifications such as the sigma-modification or dead zones guard against these failures.

When to use it

Adaptive control suits systems with large, slow parameter variation that a fixed controller cannot cover, such as aircraft across flight envelopes or processes with changing loads. When the variation is known and can be scheduled against a measured signal, the simpler gain-scheduling approach is often preferred; adaptation is reserved for genuinely unknown or unpredictable change.

The line between adaptive control and machine learning has blurred, with modern methods blending online identification, reinforcement learning, and stability guarantees to keep learning-based controllers safe.