Adaptive Control, Self-Tuning & Gain Scheduling

A fixed controller is designed for a declared family of plant behavior. An adaptive controller goes further: it changes parameters online using measured closed-loop data. That can recover performance when mass, gain, friction, aerodynamics, or operating conditions change, but it also makes the controller a coupled dynamical system whose estimator, update law, and protection mechanisms must be analyzed together.

Three mechanisms that are often conflated #

Robust control and adaptive control are also distinct. Robust control keeps one controller and proves properties over a modeled uncertainty set. Adaptive control attempts to learn or compensate changes during operation. A practical design may combine both, but their evidence is different.

The adaptive loop has two feedback paths #

flowchart TD
  R[Reference r] --> C[Controller with parameters theta]
  C --> P[Plant]
  P --> Y[Measured output y]
  Y --> C
  R --> E[Performance error]
  Y --> E
  E --> U[Adaptation law]
  Y --> U
  U --> C

The ordinary feedback path maps the current error to control. The adaptation path changes the map itself. Even if each frozen value of $\theta$ would produce a stable controller, rapid or poorly conditioned parameter motion can destabilize the combined system.

Model-reference adaptive control #

Model-reference adaptive control (MRAC) declares desired dynamics through a reference model

$$\dot x_m=A_m x_m+B_m r.$$

The plant receives a parameterized control law, for example

$$u=\theta_x^\top x+\theta_r r,$$

and the tracking error is $e=x-x_m$. A Lyapunov-based design chooses an update law such as

$$\dot\theta=-\Gamma\,\phi e^\top P B,$$

where $\phi$ collects regressors, $\Gamma$ sets adaptation rates, and $P$ comes from a Lyapunov equation for the desired error dynamics. The exact form depends on plant assumptions, known control direction, and parameterization.

The important idea is not “gradient descent on error.” The update law is constructed so a Lyapunov function for state and parameter error has a nonpositive derivative under stated assumptions. If actuator saturation, unmodeled dynamics, delay, or sensor noise violate those assumptions, the proof no longer transfers automatically.

Self-tuning regulation separates estimation and design #

An indirect self-tuning regulator repeatedly:

1. estimates plant parameters from recent input/output data;
2. constructs a controller from the estimated model;
3. applies the control action;
4. updates the estimate as new data arrive.

For a discrete model

$$y_k=\varphi_k^\top\theta+v_k,$$

recursive least squares can update $\hat\theta_k$ and its covariance. But closed-loop inputs are correlated with disturbances, and regulation deliberately removes excitation. The estimator can therefore appear numerically confident while important parameter directions remain unobserved.

Gain scheduling is interpolation, not adaptation #

Gain scheduling designs controllers at representative operating points—such as aircraft speed and altitude or engine speed and load—then selects or interpolates them using a scheduling variable $\rho$:

$$u=K(\rho)x.$$

The schedule is predetermined; it does not learn from tracking error. It is attractive when plant variation is repeatable and the scheduling variable is measurable.

The subtlety is that closed-loop stability at every fixed $\rho$ does not prove stability when $\rho(t)$ changes. Interpolation can introduce hidden coupling terms, discontinuities, or excessive gain rate. Analysis must include the allowed rate and path of scheduling-variable variation, not only the grid points.

Adaptation needs bounded semantics #

Practical mechanisms include:

• projection: keep estimates inside a justified parameter set;
• normalization: prevent large regressors from causing oversized updates;
• dead zones: suppress adaptation when error is within the noise/modeling floor;
• rate limits: bound parameter motion between controller cycles;
• anti-windup: account for saturation so the update law does not learn from an action the actuator never produced;
• parameter monitoring: expose drift, covariance, excitation, and constraint activity.

These mechanisms change the adaptive dynamics and belong in the analysis. They are not implementation details that can be added after a proof for the ideal update law.

Worked failure mode: adaptation against saturation #

Suppose a controller requests $u=15$ while the actuator is limited to $u=10$. The plant responds to 10, but an estimator that records the requested command as the applied input attributes the weak response to low plant gain. It increases the controller gain, requests an even larger command, and drives the estimate further from reality.

The data path must distinguish requested, limited, and measured actuator values. The update law must know when the excitation and error are dominated by saturation.

Verification must cover the learning trajectory #

Test not only final tracking but also:

• parameter and covariance trajectories;
• transient response from incorrect initial estimates;
• loss and recovery of excitation;
• sensor bias, noise, delay, and dropouts;
• actuator saturation and rate limits;
• abrupt and gradual plant changes;
• scheduling transitions at maximum allowed rate;
• worst-case execution time of estimation and redesign;
• reset, mode-transfer, and parameter-persistence semantics.

NASA's adaptive-flight-control work emphasizes why safety-critical adoption is difficult: nonlinear adaptation and learning complicate conventional verification, certification, and assurance arguments. Simulation breadth is necessary but not a proof of stability or safe authority.

Choosing the mechanism #

• Use gain scheduling when operating regimes are known, repeatable, and measured.
• Use a self-tuning regulator when a reliable online model can be identified and controller redesign is well posed.
• Use direct adaptation when the uncertainty matches a parameterization for which the combined dynamics can be analyzed.
• Use robust fixed control when the uncertainty envelope is known and online learning adds more risk than information.

Adaptation is valuable when it is narrower than the uncertainty it addresses. “Tune until the error improves” is an experiment, not an adaptive-control design.