Robust and Adaptive Control

Introduction and Context

Classical control design commonly assumes that a mathematical model accurately represents the plant. In practice, however, parameters vary with operating conditions, neglected dynamics influence the response, sensors and actuators introduce uncertainty, and external disturbances are unavoidable. A controller designed for only one nominal model may therefore become unstable or perform poorly when the real plant differs from the model.

Robust and adaptive control address this problem in complementary ways. Robust control seeks a single controller that maintains stability and acceptable performance for all plants within a specified uncertainty set. Adaptive control changes controller parameters online in response to observed plant behavior. Robust design is especially valuable when uncertainty can be bounded but not identified accurately; adaptive design is attractive when plant parameters vary substantially and can be estimated from measurements.

This chapter develops sensitivity-based analysis, small-gain reasoning, structured uncertainty, H-infinity loop shaping, gain scheduling, model-reference adaptive control (MRAC), and self-tuning regulators. The emphasis is on assumptions, design interpretation, implementation, and limitations rather than on treating any method as universally superior.

Learning Objectives

After studying this chapter, the reader should be able to:

1. Distinguish parametric, unstructured, structured, and uncertainty caused by neglected dynamics.

2. Use sensitivity and complementary sensitivity functions to assess disturbance rejection, tracking, and robustness.

3. Apply the small-gain theorem to formulate robust-stability conditions.

4. Explain structured singular value concepts and their relevance to mixed uncertainty.

5. Formulate the H-infinity loop-shaping problem and interpret weighting functions.

6. Compare gain scheduling, MRAC, and self-tuning regulators.

7. Identify implementation risks, excitation requirements, and certification issues in adaptive control.

8. Select an appropriate robust or adaptive strategy for an engineering application.

1. Uncertainty in Feedback Systems

Consider a nominal plant P(s) and controller K(s) in a negative-feedback loop. The loop transfer function is

L(s) = P(s)K(s).

The sensitivity and complementary sensitivity functions are, respectively,

S(s) = [1 + L(s)]^{-1}, T(s) = L(s)[1 + L(s)]^{-1}.

Where the loop is internally stable, S + T = 1. Sensitivity describes how strongly certain closed-loop quantities respond to disturbances and plant perturbations. Complementary sensitivity describes the transmission of reference signals and high-frequency measurement noise, and it is also central to robustness against high-frequency unmodeled dynamics.

For a standard unity-feedback system, the tracking error due to a reference r and output disturbance d can often be expressed through S, while the output response to the reference is T. Thus, large loop gain generally reduces low-frequency tracking error because |S| becomes small. However, because S + T = 1, reducing sensitivity over one frequency range usually requires increased complementary sensitivity elsewhere. This trade-off is a fundamental limitation, not merely a consequence of a particular design method.

1.1 Types of uncertainty

Parametric uncertainty occurs when a known model structure contains unknown or varying parameters. For example,

P(s,\theta) = \frac{1}{ms^2 + bs + k},

where mass m, damping b, and stiffness k may vary within specified intervals. The uncertainty is structured because its location and physical meaning are known.

Unstructured uncertainty represents modeling error without relying on a detailed description of its internal structure. A common multiplicative representation is

P(s) = P_0(s)[1 + W_m(s)\Delta(s)],

where P_0 is the nominal plant, W_m is a known frequency-dependent bound, and \Delta is unknown but stable with ||\Delta||_\infty \leq 1. The bound W_m should be obtained from experiments, validated models, or conservative engineering estimates.

An additive representation is

P(s) = P_0(s) + W_a(s)\Delta(s).

Multiplicative uncertainty is often more appropriate when model error scales with plant magnitude; additive uncertainty is useful when an absolute error bound is more meaningful. Neither representation is automatically correct at all frequencies.

Dynamic or neglected uncertainty results from omitted resonances, flexible modes, delays, actuator dynamics, and sensor filtering. Such uncertainty is particularly dangerous near crossover frequencies and at frequencies where the controller has high gain.

Disturbance uncertainty concerns unknown external inputs rather than plant dynamics. Robust control must distinguish uncertainty in the model from uncertainty in forcing signals, because the appropriate weighting and performance measures differ.

2. Sensitivity-Based Robustness Analysis

2.1 Disturbance rejection and tracking

For a plant P and controller K, the closed-loop transfer functions include

Y/R = T, \qquad E/R = S,

for reference r and error e under unity feedback. A low-frequency design objective commonly requires |S(j\omega)| to be small in the frequency band containing tracking commands and slowly varying disturbances. Conversely, measurement noise n transmitted to the output is commonly associated with T, so |T(j\omega)| should be small in the high-frequency noise band.

The quantity

M_s = ||S||_\infty = \sup_\omega |S(j\omega)|

is called the peak sensitivity. It is a useful robustness indicator: large M_s usually indicates a lightly damped or fragile closed loop. However, M_s alone does not describe all uncertainty types or actuator limitations.

For multiplicative output uncertainty, the perturbed loop is approximately affected through T. A sufficient robust-stability condition is

||W_m T||_\infty < 1,

provided the uncertainty is stable, unstructured, normalized, and represented as a feedback perturbation in the relevant configuration. This condition follows from the small-gain theorem. It is sufficient and, for full complex unstructured uncertainty, generally tight in the corresponding normalized setting.

For additive uncertainty represented at the plant input or output, the relevant transfer function changes. For example, if P = P_0 + W_a\Delta and the perturbation enters additively at the plant output, a typical sufficient condition involves

||W_a K S||_\infty < 1,

subject to the exact interconnection and signal normalization. This illustrates why uncertainty location must be specified before applying a robustness formula.

2.2 Fundamental trade-offs

The identity S + T = 1 implies that sensitivity and complementary sensitivity cannot both be small at the same frequency. In addition, stable plants with right-half-plane zeros, unstable poles, or delays impose stronger limitations. For example, attempting to obtain very high bandwidth in a plant with a nonminimum-phase zero can generate a large sensitivity peak and poor robustness.

The Bode sensitivity integral provides a qualitative expression of the “waterbed effect.” Under restrictive assumptions, including a stable open-loop plant and suitable high-frequency behavior,

\int_0^\infty \ln |S(j\omega)|\,d\omega = 0.

For unstable poles or other nonminimum-phase features, the integral is modified by additional positive terms. The practical conclusion is that reducing sensitivity over one frequency interval generally causes sensitivity amplification elsewhere. Robust design therefore requires bandwidth selection, not simply maximization of loop gain.

3. Small-Gain Robustness

The small-gain theorem states that the feedback interconnection of two stable systems M and \Delta is internally stable if

||M||_\infty ||\Delta||_\infty < 1.

If ||\Delta||_\infty \leq 1, then the condition becomes ||M||_\infty < 1. The theorem is conservative when uncertainty is structured or real-valued, but it is powerful because it requires only norm bounds and does not require identifying the exact uncertainty.

For a nominal closed-loop system subject to a normalized unstructured perturbation, M is the transfer matrix from the perturbation output to its input. Robust stability is certified if its H-infinity norm is less than one. A frequency-domain interpretation is

\sup_\omega \bar{\sigma}[M(j\omega)] < 1,

where \bar{\sigma} denotes the maximum singular value. For scalar systems, this reduces to the maximum magnitude.

The theorem has important limitations. It assumes the uncertainty bound is valid, the interconnection is correctly represented, and the uncertain block is stable in the chosen formulation. A bound that is too optimistic invalidates the conclusion; a very conservative bound may reject a perfectly usable controller.

4. Structured Robustness and the Structured Singular Value

Unstructured analysis treats all perturbations as if they could coordinate in the worst possible way. Real systems often contain repeated scalar parameters, real-valued gains, repeated dynamic blocks, or separate uncertainty channels. Structured singular value analysis addresses this situation.

Let M(j\omega) be a transfer matrix and let \Delta belong to a prescribed block-diagonal uncertainty structure. The structured singular value is conceptually defined by

\mu_\Delta(M) = \frac{1}{\min\{\bar{\sigma}(\Delta): \det(I-M\Delta)=0\}},

with the convention that \mu_\Delta(M)=0 if no destabilizing perturbation exists. Robust stability for normalized uncertainty is guaranteed when

\sup_\omega \mu_\Delta[M(j\omega)] < 1.

Unlike the ordinary singular value, \mu depends on the uncertainty structure. It can therefore provide less conservative conclusions than the small-gain test for full-block uncertainty. Exact computation is difficult in general; practical software commonly uses upper and lower bounds, D-scalings, and frequency gridding.

Structured analysis is particularly useful for mixed-uncertainty models such as

\Delta = \operatorname{diag}(\delta_1 I,\delta_2 I,\Delta_3),

where \delta_1 and \delta_2 may represent repeated real parameter variations and \Delta_3 a dynamic unstructured block. The analysis should preserve physical correlations where credible. Artificially independent uncertainty blocks can create unrealistic worst cases, while omitting correlations can produce unsafe optimism.

5. H-Infinity Loop Shaping

H-infinity control formulates design as minimization of the worst-case induced gain. For a stable transfer matrix G(s),

||G||_\infty = \sup_{\omega} \bar{\sigma}[G(j\omega)].

In a generalized plant, exogenous inputs w include references, disturbances, and noise; regulated outputs z include weighted tracking errors, control effort, and measured noise effects. The design objective is

\min_K ||T_{zw}(s)||_\infty,

where T_{zw} is the closed-loop transfer from w to z, subject to internal stability.

5.1 Weight selection

Weighting functions convert engineering requirements into frequency-domain penalties. A performance weight W_p may be selected so that

||W_p S||_\infty < 1.

If low-frequency tracking error must be small, W_p is made large at low frequency. A control-effort weight W_u produces a requirement such as

||W_u K S||_\infty < 1.

A multiplicative uncertainty weight W_m gives the robustness condition

||W_m T||_\infty < 1.

Weight selection is not a cosmetic step. If W_p, W_u, and W_m are mutually incompatible, no controller can satisfy the desired bound. A useful workflow is to plot the inverse weights, identify conflicting frequency requirements, and adjust specifications before running synthesis.

The mixed-sensitivity formulation is frequently written as

\left\|\begin{bmatrix} W_p S \\ W_u K S \\ W_m T \end{bmatrix}\right\|_\infty < \gamma,

where \gamma is minimized. The equations assume a compatible generalized-plant interconnection and stable proper weighting functions. In practice, scaling, nonminimum-phase plant features, unstable poles, and controller order must be handled explicitly.

5.2 Loop-shaping interpretation

An alternative approach shapes the singular-value profile of the nominal loop using pre- and post-compensators, followed by robust stabilization. Loop shaping seeks high loop gain in the performance band and low gain where uncertainty or noise dominates. It is especially useful when the designer has a clear frequency-domain understanding of plant and actuator limitations.

H-infinity designs can be high order and sensitive to numerical conditioning. Model reduction must preserve robust performance, and the reduced controller must be revalidated against the original plant and uncertainty set.

6. Gain Scheduling

Gain scheduling uses a family of controllers whose parameters depend on a measured scheduling variable, such as speed, altitude, load, or temperature. If a plant model is available at operating points \rho, one may design K_\rho and interpolate controller parameters or controller outputs.

A simple scheduled controller is

u(t) = K(\rho(t))x(t),

where x is a measured or estimated state. Scheduling is not automatically adaptive: the scheduling variable is generally assumed measurable, and the controller does not necessarily estimate unknown parameters from input-output data.

A valid gain-scheduling design requires:

• Operating-point models that represent the plant over the intended envelope.
• A scheduling variable correlated with the dominant plant variation.
• Smooth interpolation and bumpless transfer.
• Adequate rate limits on scheduling-variable changes.
• Stability analysis for the time-varying or parameter-dependent system.

Pointwise stability of every frozen controller does not by itself guarantee stability during rapid scheduling. Parameter-dependent Lyapunov functions, linear parameter-varying methods, or conservative common Lyapunov functions may be required.

7. Model-Reference Adaptive Control

MRAC specifies a desired reference model and adjusts controller parameters so that the plant follows its response. Consider a plant with known relative degree and uncertain parameters. A simplified first-order example is

\dot{x} = a x + b u,

where b has known sign but a and possibly b are unknown. Let the reference model be

\dot{x}_m = a_m x_m + b_m r,

with a_m < 0. A controller may be parameterized as

u = \hat{\theta}_1 x + \hat{\theta}_2 r,

and the tracking error is e = x - x_m. The adaptive law is derived using a Lyapunov function rather than selected solely by intuition.

For an ideal parameter vector \theta^*, the error dynamics can often be expressed as

\dot e = a_m e + \tilde{\theta}^{T}\phi,

where \tilde{\theta}=\hat{\theta}-\theta^* and \phi is a regressor. Choose

V = \frac{1}{2}e^2 + \frac{1}{2}\tilde{\theta}^{T}\Gamma^{-1}\tilde{\theta},

where \Gamma=\Gamma^T>0 is the adaptation-gain matrix. An update law of the form

\dot{\hat{\theta}} = -\Gamma \phi e

can cancel cross terms in \dot V under compatible sign conventions, producing

\dot V = a_m e^2 \leq 0.

The precise sign and scaling depend on the plant formulation and control-law parameterization. The result establishes boundedness and, under additional conditions, convergence of tracking error; it does not automatically guarantee parameter convergence.

Parameter convergence generally requires persistent excitation of the regressor. Without sufficient excitation, many parameter values may produce nearly identical output behavior. Practical MRAC therefore often incorporates projection, normalization, dead zones, leakage, or sigma modification. These additions improve robustness against noise and unmodeled dynamics but can weaken ideal convergence guarantees.

A major limitation is the sensitivity of unrestricted adaptation to high-frequency unmodeled dynamics. The adaptive law may attempt to compensate for dynamics that the model cannot represent, producing parameter drift or oscillation. Robust adaptive control combines adaptation with bounded uncertainty models, filters, dead zones, or robustifying terms.

8. Self-Tuning Regulators

A self-tuning regulator separates identification and control. It estimates plant parameters online and computes a controller using the current model. The usual architecture contains:

1. A recursive estimator, such as recursive least squares.

2. A controller-design procedure, such as pole placement, minimum-variance control, or generalized predictive control.

3. An optional excitation and monitoring mechanism.

For a discrete-time model, a regression equation can be written as

y(k)=\phi^T(k)\theta+v(k),

where y(k) is the measured output, \phi(k) is a regressor vector, \theta is the parameter vector, and v(k) is modeling noise. Recursive least squares with forgetting factor \lambda, 0<\lambda\leq1, commonly uses

K(k)=\frac{P(k-1)\phi(k)}{\lambda+\phi^T(k)P(k-1)\phi(k)},

\hat\theta(k)=\hat\theta(k-1)+K(k)[y(k)-\phi^T(k)\hat\theta(k-1)],

P(k)=\lambda^{-1}[P(k-1)-K(k)\phi^T(k)P(k-1)].

Here P is the covariance or information-shape matrix. The equations assume a linear regression form and require appropriate initialization and numerical safeguards. A forgetting factor allows tracking of slowly changing parameters but increases sensitivity to noise.

In a certainty-equivalence regulator, the estimated parameters are treated as if they were true, and a controller is recomputed. This can work well when estimates are accurate and the plant remains within the assumed model class. It can fail when the data are insufficiently informative, when the identified model is nonminimum phase, or when controller recomputation produces abrupt changes.

Self-tuning regulators raise the dual-control problem: inputs must both control the plant and provide information for identification. Deliberate probing improves identifiability but may degrade performance or violate safety constraints. In safety-critical systems, exploration must be bounded and justified.

9. Worked Example: Robustness of a First-Order Loop

Consider the nominal plant

P_0(s)=\frac{1}{s+1}

and proportional controller K(s)=k, with k>0. The loop transfer function is L=k/(s+1). The sensitivity and complementary sensitivity functions are

S(s)=\frac{s+1}{s+1+k}, \qquad T(s)=\frac{k}{s+1+k}.

The nominal closed-loop pole is at -(1+k), so every positive k gives nominal stability. Suppose the plant has multiplicative uncertainty bounded by |W_m(j\omega)|=0.1 over the frequency range important for robustness. A sufficient condition is

||W_m T||_\infty < 1.

For this first-order system, |T(j\omega)| is largest at \omega=0 and equals k/(1+k). Therefore,

||W_m T||_\infty =0.1\frac{k}{1+k}<1.

For every finite positive k, the left side is less than 0.1 and the robust-stability condition is satisfied. This result does not mean arbitrarily large gain is desirable. Increasing k reduces low-frequency sensitivity, since |S(0)|=1/(1+k), but increases bandwidth, control effort, and potential sensitivity to uncertainty not captured by the assumed 0.1 bound. If the uncertainty instead grows with frequency, or if an unmodeled delay is present, the conclusion can change substantially. The example demonstrates that robustness conclusions are conditional on the uncertainty model and frequency range.

10. Illustrative Case Study: Adaptive Speed Control of an Electric Drive

Background. Consider a variable-speed electric drive used in a manufacturing line. Its effective inertia changes when tooling is exchanged, while friction varies with temperature. A fixed controller performs well at nominal load but produces excessive settling time and overshoot after a load change.

Problem. The drive must track speed commands while respecting current limits. The controller must remain stable during rapid but bounded load changes, and the system must not inject unsafe exploratory signals.

Analysis. A frequency-domain identification campaign reveals a reliable nominal model up to the intended bandwidth, but uncertainty increases near a mechanical resonance. The load estimate is available from production data, but it is delayed and noisy. This makes an aggressively adaptive high-bandwidth design undesirable.

Approach. A robust inner current loop is first designed using loop shaping. The uncertainty weight is increased near the resonance, forcing the controller to roll off before that frequency. For the outer speed loop, gain scheduling uses measured operating region and estimated load class. A slow MRAC correction is added only to compensate for low-frequency friction variation. Projection bounds the adaptive parameters, a dead zone prevents adaptation on noise-dominated errors, and current saturation is explicitly included in simulation. During commissioning, the scheduled controllers are tested at grid points and during worst-case transitions. The adaptation rate is limited so that it cannot destabilize the validated robust inner loop.

Outcome. The combined architecture provides faster response over the load range than a single fixed controller while retaining a certified robust inner loop. Performance degrades gracefully when the adaptive correction is disabled, which is important for fault handling.

Learning points. Robust and adaptive methods need not be competing alternatives. A robust baseline can protect against unmodeled dynamics, while bounded adaptation or scheduling handles predictable variation. The case also illustrates that actuator saturation, scheduling transients, estimator quality, and fallback behavior are part of the control design—not afterthoughts.

11. Implementation, Verification, and Governance

Robust and adaptive controllers should be evaluated against more than nominal step responses. Recommended tests include parameter sweeps, randomized uncertainty sampling, worst-case frequency analysis, sensor-noise tests, actuator saturation, computation delays, quantization, initialization errors, and abrupt operating-point transitions.

For adaptive systems, engineers should define parameter bounds, reset behavior, data-quality checks, adaptation freeze conditions, and a safe fallback controller. Monitoring should detect loss of excitation, estimator divergence, sensor faults, and excessive control effort. Independent verification is especially important because online adaptation changes the effective controller after deployment.

In safety-critical applications, adaptive behavior may complicate certification and traceability. Requirements should specify allowable controller variation, validated operating envelopes, and evidence that adaptation cannot bypass safety interlocks. Cybersecurity is also relevant: corrupted measurements or commands can drive an adaptive law to learn unsafe parameters. Access control, authenticated data, anomaly detection, and secure update procedures should therefore be integrated into the control architecture.

Artificial intelligence can assist with model identification, uncertainty-bound estimation, or anomaly detection, but data-driven components should not replace stability arguments without qualification. Any learned model must be validated under distribution shifts, sensor faults, and rare operating conditions. Explainability, monitoring, and conservative fallback logic are essential when learning affects control decisions.

12. Best-Practice Design Workflow

A technically sound workflow is:

1. Define performance, disturbance, noise, actuator, and safety requirements.

2. Establish a nominal model and identify the frequency range in which it is credible.

3. Separate parametric variation from unstructured modeling error.

4. Choose uncertainty locations and bounds based on physical evidence.

5. Analyze S, T, singular values, and actuator demand before synthesis.

6. Select robust synthesis, scheduling, adaptation, or a hybrid architecture according to the uncertainty characteristics.

7. Check feasibility of weighting functions before optimization.

8. Include saturation, rate limits, delays, quantization, and estimator dynamics.

9. Validate with deterministic worst cases and statistically meaningful uncertainty samples.

10. Define monitoring, fallback, update, and maintenance procedures.

The central principle is that robustness claims are only as credible as the uncertainty description supporting them. Conversely, adaptive performance claims are only as credible as the excitation, estimator assumptions, and safety mechanisms supporting them.

Challenges and Limitations

Robust control may be conservative, particularly when unstructured bounds ignore known physical structure. H-infinity optimization can produce high-order controllers and difficult weight-selection problems. Structured singular value calculations may be computationally demanding and dependent on uncertain block descriptions.

Gain scheduling can fail when the scheduling variable does not capture the true source of variation or changes too quickly. MRAC can suffer from parameter drift, noise sensitivity, and interaction with unmodeled dynamics. Self-tuning regulators depend on informative data and can generate unsafe controllers when estimation is poor. All methods are affected by saturation, delays, sensor faults, and implementation limits.

No mathematical certificate replaces model validation and engineering judgment. A controller may be robust to the modeled uncertainty while remaining vulnerable to an omitted failure mode. Conversely, a conservative analysis may hide useful performance. The design task is therefore an iterative combination of modeling, analysis, synthesis, experiment, and verification.

Chapter Summary

Robust control designs for bounded uncertainty, whereas adaptive control changes controller parameters in response to observed behavior. Sensitivity S governs error and disturbance attenuation, while complementary sensitivity T governs reference transmission, measurement-noise transmission, and important aspects of multiplicative robustness. The small-gain theorem provides a sufficient norm-based robust-stability test. Structured singular value analysis improves on unstructured tests when uncertainty has known block structure.

H-infinity loop shaping converts tracking, control-effort, noise, and uncertainty requirements into weighted worst-case performance objectives. Gain scheduling handles predictable operating-point variation but requires careful treatment of scheduling transients. MRAC uses a reference model and Lyapunov-based adaptation; parameter convergence requires excitation and practical implementations need robustness modifications. Self-tuning regulators combine online estimation with controller redesign but face identification, excitation, and safety challenges.

Effective engineering practice combines validated uncertainty models, frequency-domain analysis, appropriate synthesis, nonlinear and implementation testing, monitoring, and safe fallback behavior. Robust and adaptive techniques are often strongest when used together rather than in isolation.

Key Terms

Parametric uncertainty: Variation in known model parameters.

Unstructured uncertainty: Modeling error represented without detailed internal structure.

Multiplicative uncertainty: Error represented as a frequency-dependent factor multiplying a nominal plant.

Additive uncertainty: Error represented as a term added to the nominal plant.

Sensitivity function: S=(1+L)^{-1}, relating several disturbances and errors to closed-loop responses.

Complementary sensitivity: T=L(1+L)^{-1}, governing reference and noise transmission in common configurations.

Peak sensitivity: The H-infinity norm of S.

Small-gain theorem: A norm-based sufficient condition for stability of a feedback interconnection.

Structured singular value: A robustness metric that accounts for specified uncertainty structure.

H-infinity norm: The supremum over frequency of the maximum singular value.

Loop shaping: Designing the frequency profile of loop gain to balance performance and robustness.

Gain scheduling: Selecting controller parameters according to a measured operating variable.

Model-reference adaptive control: Adaptation that seeks to make a plant follow a prescribed reference model.

Persistent excitation: A richness condition on signals needed for parameter identification or convergence.

Parameter drift: Unbounded or undesirable parameter movement caused by noise, disturbances, or model mismatch.

Self-tuning regulator: A controller that estimates plant parameters online and redesigns control accordingly.

Certainty equivalence: Treating estimated parameters as though they were exact during controller design.

Generalized plant: An interconnected model used to formulate robust-performance objectives.

Robust stability: Stability maintained for every plant in a specified uncertainty set.

Mixed sensitivity: Simultaneous shaping of weighted S, control sensitivity, and weighted T.

Review Questions

Short Answer Questions

1. What is the difference between parametric and unstructured uncertainty?

2. Define the sensitivity and complementary sensitivity functions.

3. State the small-gain robust-stability condition.

4. Why is persistent excitation important in adaptive control?

5. How does gain scheduling differ from adaptive control?

Long Answer Questions

1. Explain how sensitivity functions support analysis of tracking, disturbance rejection, measurement noise, and robustness.

2. Describe structured singular value analysis and compare it with unstructured small-gain analysis.

3. Formulate and interpret a mixed-sensitivity H-infinity design problem.

4. Derive the role of a Lyapunov function in a basic MRAC design and state its limitations.

5. Explain the architecture, assumptions, and practical risks of a self-tuning regulator.

Analytical / Application Questions

1. A plant has multiplicative uncertainty that increases sharply above its nominal bandwidth. Explain how the uncertainty weight should influence controller loop shaping.

2. Design a control architecture for a system whose inertia varies predictably with operating point but also contains an unmodeled flexible resonance. Justify whether scheduling, robust control, adaptation, or a hybrid is appropriate.

3. An adaptive controller exhibits parameter drift whenever the reference command is constant. Diagnose possible causes and propose modifications that preserve safety and useful tracking performance.

Multiple Choice Questions with Answers

Question 1

For a unity-feedback loop, which expression defines the sensitivity function?

A. S(s)=1+L(s)

B. S(s)=L(s)/(1+L(s))

C. S(s)=1/(1+L(s))

D. S(s)=1/L(s)

Correct Answer: C. S(s)=1/(1+L(s))

Question 2

If normalized unstructured uncertainty satisfies ||Δ||∞≤1, which condition is sufficient for small-gain robust stability?

A. ||M||∞>1

B. ||M||∞<1

C. ||M||∞=0 only

D. ||M||∞≤-1

Correct Answer: B. ||M||∞<1

Question 3

Which function is most directly associated with robustness to multiplicative plant uncertainty?

A. T

B. 1/S

C. K alone

D. The reference signal alone

Correct Answer: A. T

Question 4

What does the structured singular value account for that an ordinary maximum singular value does not?

A. Sampling frequency only

B. The specified structure of uncertainty

C. Controller implementation language

D. The nominal reference command

Correct Answer: B. The specified structure of uncertainty

Question 5

In mixed-sensitivity design, a control-effort weight primarily penalizes which quantity?

A. W_uKS

B. W_uT only

C. W_uP only

D. W_uS^{-1} only

Correct Answer: A. W_uKS

Question 6

Which statement about gain scheduling is correct?

A. It always estimates unknown parameters from input-output data

B. It selects controller behavior using an operating or scheduling variable

C. It guarantees stability during arbitrarily fast scheduling

D. It does not require operating-point models

Correct Answer: B. It selects controller behavior using an operating or scheduling variable

Question 7

What condition is generally associated with parameter convergence in MRAC?

A. Persistent excitation

B. Zero controller gain

C. Infinite sensor noise

D. Removing the reference model

Correct Answer: A. Persistent excitation

Question 8

What is the principal role of a forgetting factor in recursive least squares?

A. To eliminate all measurement noise

B. To allow greater emphasis on recent data

C. To guarantee plant stability independently of the controller

D. To make the plant minimum phase

Correct Answer: B. To allow greater emphasis on recent data

Question 9

For P0(s)=1/(s+1) and K(s)=k>0, what is the nominal closed-loop pole?

A. k-1

B. -k

C. -(1+k)

D. 1+k

Correct Answer: C. -(1+k)

Question 10

Why can unrestricted adaptation be dangerous in the presence of unmodeled high-frequency dynamics?

A. It may attempt to compensate for dynamics outside the assumed model

B. It always reduces controller bandwidth to zero

C. It removes all actuator saturation

D. It makes persistent excitation unnecessary

Correct Answer: A. It may attempt to compensate for dynamics outside the assumed model