Questions: Adaptive Control Theory

Question 1

A direct adaptive control system adjusts a proportional gain Kp in real-time to minimize error. The closed-loop response is stable, but the gain keeps drifting slowly even though the setpoint and disturbances are constant. What is the most likely cause?

• A. The system is unstable and diverging — the gain drift is evidence of loss of control
• B. The input signal lacks sufficient excitation; without variance at the frequency scales where the plant responds, parameter estimates cannot converge and drift randomly
• C. The adaptation rate is too slow; increasing the learning rate will speed convergence
• D. The reference model used for comparison is incorrect

Answer: (B)

Without persistent excitation — variance in the plant's input signal at the relevant frequencies — there is no feedback about the quality of the current parameter estimate. The error is zero at the current instant but may not be zero for all possible disturbances or setpoints. The parameter estimate loses its grounding and drifts according to measurement noise and unmodeled dynamics. This is one of the most important and counterintuitive aspects of adaptive control: stability does not imply convergence of estimates. To guarantee convergence, you need either external probing signals (increasing computational cost and control input), or the natural disturbances must provide sufficient excitation (which is not guaranteed). Modern adaptive control designs mitigate this by requiring only bounded excitation or by accepting slow drift as a design trade-off.

Question 2

In a Model Reference Adaptive System (MRAS), the adaptive controller adjusts parameters to make the plant output track a reference model output. If the reference model dynamics are faster than the plant can physically follow, what happens?

• A. The adaptive control loop automatically slows the reference model to match the plant
• B. The plant output lags the reference, the error is persistent and nonzero, and the parameter estimates can diverge or oscillate because no control input can reduce the error to zero
• C. The system remains stable but the reference model is ignored
• D. The error decreases exponentially until the system reaches the reference model

Answer: (B)

If the reference model bandwidth exceeds the plant's physical capability, the plant cannot catch the reference — the error will be persistent and substantial. The adaptive law (e.g., dKp/dt = -γ·error·input) will continuously adjust parameters, but since the fundamental limitation is the plant's speed, not the control gain, the parameters will oscillate or diverge. This is why MRAS designs require that the reference model be *achievable* by the plant under perfect control. Choosing the reference model is an underappreciated design step: it sets the closed-loop performance target and must be realistic.

Question 3

Persistent excitation is necessary for parameter convergence in adaptive control. A simple way to ensure it is to add a dither signal (high-frequency sinusoid) to the control input. What is the drawback of this approach?

• T. True
• F. False

Answer: True

Dither ensures input variance, guaranteeing parameter convergence, but at the cost of injecting energy into the system: the dither signal appears in the plant output as well, creating a limit cycle around the true parameter values. The plant output oscillates at the dither frequency and amplitude. This may be acceptable in some applications (e.g., slow thermal processes), but unacceptable in others (e.g., precision positioning). Modern adaptive control avoids dither by accepting slower convergence, using intermittent or scheduled excitation, or exploiting natural disturbances (wind, vibration, etc.).

Question 4

A self-tuning regulator estimates plant parameters at each time step using recursive least squares (RLS) and immediately computes new controller gains. Between the time the parameter estimate changes and the time the controller gain is recomputed, the system experiences a transient. What prevents this transient from destabilizing the loop?

• T. True
• F. False

Answer: True

The separation principle (related to certainty equivalence): if the parameter estimation error decreases over time and the true plant is stabilizable, then treating the estimated parameters as if they were true (certainty equivalence) and using the optimal controller for those estimated parameters guarantees closed-loop stability. However, the separation principle holds rigorously only in the ideal case (zero estimation error in the limit, exact optimization, no disturbances). Real self-tuning regulators have finite estimation error at each step, and the separation principle is approximate — careful analysis and gain scheduling (slow parameter updates) are needed to avoid instability.

Question 5

Explain the difference between direct and indirect adaptive control, and why indirect adaptive control (estimate plant parameters, then compute controller) is often more robust despite requiring two sequential steps.

Model answer: Direct adaptive control: estimate the controller parameters directly (e.g., Kp, Ki, Kd) using a feedback law such that error decreases. The controller parameters change at each step. Indirect adaptive control: estimate the plant parameters (e.g., mass, damping, natural frequency) at each step, then compute the optimal controller parameters from the current parameter estimates (certainty equivalence). Direct is simpler (one estimation step), but the relationship between parameter estimates and error is indirect — improving one parameter may worsen another, and the optimization landscape can be multimodal. Indirect is more complex (estimation + control design at each step), but separates concerns: parameter estimation focuses on minimizing model error, and control design is a well-studied problem. If the parameter estimator is robust and accurate, the subsequent controller design is reliable. Modern practice uses indirect when plant parameters have clear physical meaning (easier to validate estimates) and direct when the control law structure is simple (e.g., scalar gain).

The robustness advantage of indirect control comes from decoupling: a poor Kp estimate in direct control can either be too large (destabilizing) or too small (poor performance), and the adaptation law may struggle to correct it. With indirect control, a poor mass estimate leads to wrong control gain, but the subsequent control law computation is transparent and can be validated independently. Some plants have nonminimum-phase or unstable dynamics, where small errors in the estimated zeros or poles can cause control instability; indirect control makes these errors visible, whereas direct control can hide them.