Cognitive Radar, AI and Emerging Technologies

Introduction and Context

Radar engineering is moving from systems designed around fixed waveforms, fixed thresholds and predetermined operating modes toward systems that sense their environment, infer uncertainty, and adapt their actions. This transformation is driven by dense electromagnetic environments, spectrum sharing, low-cost digital hardware, autonomous platforms, large data volumes and the increasing availability of machine-learning methods.

A cognitive radar is not simply a radar containing an AI classifier. It is an adaptive sensing system in which observations are used to update an internal representation of the environment, select sensing actions, and evaluate the results. The central design problem is a closed loop:

\[

\text{Observe} \rightarrow \text{Infer} \rightarrow \text{Decide} \rightarrow \text{Act} \rightarrow \text{Learn}

\]

The loop must operate under physical, computational, regulatory and safety constraints. AI can improve detection, classification, waveform selection, interference mitigation and maintenance, but it does not remove the need for electromagnetic theory, statistical detection, calibration, verification and human oversight.

This chapter examines the technical foundations and limitations of cognitive radar, with particular attention to machine learning, distributed sensing, passive radar, joint communication-radar sensing, edge AI and claims concerning quantum radar.

Learning Objectives

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

1. Explain the architecture and operating principles of cognitive radar.

2. Select appropriate machine-learning approaches for detection, classification, adaptation and maintenance.

3. Distinguish supervised, unsupervised, self-supervised and reinforcement learning in radar applications.

4. Analyse distributed, passive and joint communication-radar sensing systems.

5. Evaluate edge-AI architectures, performance metrics and deployment risks.

6. Critically assess quantum-radar claims using established radar and quantum principles.

7. Design a safe, testable and governance-aware adaptive radar workflow.

1. From Conventional to Cognitive Radar

A conventional radar commonly follows a designed sequence: transmit a waveform, receive echoes, perform matched filtering, detect peaks, estimate parameters and track targets. Its operating modes may be selected by a rule table or by an operator. Such systems can be highly reliable because their behaviour is bounded and extensively tested, but they may be inefficient when the environment changes rapidly.

A cognitive radar adds three functional elements:

• A belief or knowledge state: an estimate of target, clutter, interference, channel and sensor conditions, together with uncertainty.
• A policy: a rule, optimization procedure or learned mapping that selects waveform, beam, dwell time, bandwidth, repetition interval or sensor resource.
• A feedback mechanism: the effect of the selected action is measured and used to revise the belief state and future actions.

The knowledge state should not be understood as a perfect model. It may contain target tracks, clutter statistics, interference occupancy, historical false-alarm rates, receiver health indicators and confidence values. A useful abstract representation is

\[

b_t = p(x_t\mid y_{1:t},a_{1:t-1}),

\]

where \(x_t\) denotes the unknown environment state at time \(t\), \(y_{1:t}\) denotes observations up to time \(t\), and \(a_{1:t-1}\) denotes previous sensing actions. This is a probability distribution, or an approximation to one, rather than merely a list of detections. The expression assumes that the system has a probabilistic state-space model or an estimator capable of representing uncertainty.

The action may be selected by optimizing an objective such as information gain, detection probability, track quality, mission utility or energy efficiency. In a simplified form,

\[

a_t^* = \arg\max_{a\in\mathcal A} U(b_t,a),

\]

where \(\mathcal A\) is the set of feasible actions and \(U\) is a utility function. In practice, the utility must include constraints such as peak power, duty cycle, spectrum authorization, computational latency and safety.

Cognitive levels

Cognition can be introduced progressively:

1. Adaptive signal processing: thresholds, filters or waveform parameters respond to measured conditions.

2. Model-based adaptation: an environmental model predicts the effect of candidate actions.

3. Learning-assisted adaptation: a trained model estimates target or interference states.

4. Policy learning: an agent learns action selection from interaction with a simulator or controlled environment.

5. Mission-level cognition: the radar coordinates sensing resources with other sensors and operational objectives.

The last level is the most powerful but also the most difficult to validate. A system that changes its waveform autonomously must remain observable, bounded and recoverable when its model is wrong.

2. Machine Learning Tasks in Radar

2.1 Detection

Radar detection decides whether a measurement contains a target under noise and clutter. Classical detectors use a statistical model and a threshold selected to achieve a desired probability of false alarm, \(P_{FA}\). Constant false alarm rate (CFAR) detectors estimate local background power and adapt the threshold.

Machine learning can assist detection by learning spatial, temporal or spectral structures that are difficult to model analytically. Examples include convolutional networks on range-Doppler maps, recurrent models for sequences, and transformer-based models for long-context data. However, a learned detector should be evaluated at fixed false-alarm operating points, not only by accuracy. A detector that labels almost all cells as target can have high nominal accuracy in an imbalanced data set while being operationally useless.

Important metrics include:

• Probability of detection, \(P_D\).
• Probability of false alarm, \(P_{FA}\).
• Receiver operating characteristic (ROC) and precision-recall curves.
• Detection latency and computational load.
• Performance under distribution shift.

For a binary detector, the confusion matrix contains true positives, false positives, true negatives and false negatives. Precision is \(TP/(TP+FP)\), while recall is \(TP/(TP+FN)\). These definitions assume nonzero denominators and should not replace radar-specific measures such as \(P_D\) at a specified \(P_{FA}\).

2.2 Classification and recognition

Classification assigns a detected object or emission to a category, such as aircraft type, vehicle class, maritime object or emitter family. Features may include micro-Doppler, high-resolution range profiles, polarimetric measurements, angle history and waveform characteristics.

Supervised learning is appropriate when representative labelled data exist. In operational radar, labels are often incomplete, biased toward known targets and collected under limited weather, aspect angle and sensor configurations. Consequently, confidence calibration and open-set recognition are essential. A classifier should be able to state that an observation is unknown rather than forcing it into the nearest known class.

2.3 Waveform and resource adaptation

Waveform adaptation may change frequency, bandwidth, pulse duration, coding, pulse repetition interval, polarization or beam direction. The objective may be to improve range resolution, Doppler ambiguity performance, low-probability-of-intercept behaviour, interference avoidance or track continuity.

A reinforcement-learning formulation defines a state \(s_t\), action \(a_t\), reward \(r_t\), and transition model. A policy \(\pi(a\mid s)\) selects actions. The discounted return is commonly written as

\[

G_t=\sum_{k=0}^{\infty}\gamma^k r_{t+k+1}, \qquad 0\leq\gamma<1,

\]

where \(\gamma\) is a discount factor. This formulation is useful when actions have delayed effects, but the infinite sum assumes bounded rewards and a continuing decision process. Safety constraints cannot be left to the reward function alone; prohibited actions should be excluded or independently monitored.

A practical hybrid architecture uses a model-based scheduler to enforce legal and physical constraints, while machine learning ranks the feasible actions. This is usually easier to verify than unconstrained end-to-end policy learning.

2.4 Interference mitigation

AI may identify interference type, estimate occupancy, predict interference arrival and select notch filters, null directions, frequency changes or cooperative sensing actions. Training data should include intentional jamming, unintentional emissions, receiver nonlinearities and benign changes in propagation. Otherwise, a classifier may learn the identity of a test site or a sensor artefact rather than the interference mechanism.

Adaptive spatial filtering can be based on a sample covariance matrix. If \(\mathbf R\) is the interference-plus-noise covariance matrix and \(\mathbf s\) is the steering vector, a minimum-variance distortionless response beamformer has the form

\[

\mathbf w=\frac{\mathbf R^{-1}\mathbf s}{\mathbf s^H\mathbf R^{-1}\mathbf s},

\]

provided \(\mathbf R\) is nonsingular and accurately estimated. Here \((\cdot)^H\) denotes conjugate transpose. In practice, diagonal loading or regularization is often required when the number of training samples is small or the covariance estimate is ill-conditioned. Machine learning can estimate or classify covariance regimes, but it does not eliminate the need for stable array calibration.

2.5 Predictive maintenance

Radar maintenance traditionally uses scheduled inspection and threshold alarms. Predictive maintenance estimates the probability of failure or remaining useful life from telemetry such as transmitter current, temperature, phase noise, calibration residuals, receiver gain and cooling-system behaviour.

Useful methods include anomaly detection, survival analysis, state-space estimation and supervised failure prediction. The data are usually highly imbalanced: failures are rare, and maintenance interventions censor the natural failure process. Therefore, random train-test splitting may produce misleading results if measurements from the same hardware unit appear in both sets. Splitting by unit and by time gives a more realistic estimate of generalization.

The value of predictive maintenance is not merely classification accuracy. It includes avoided downtime, reduced false work orders, spare-parts optimization and safe maintenance scheduling.

3. Data, Models and Validation

Radar AI development begins with a measurement and labelling strategy. Data should preserve raw or sufficiently rich intermediate representations when possible, including calibration information, signal quality, environmental context and sensor configuration. Labels should record uncertainty and provenance rather than implying unjustified precision.

Model choices

• Linear and generalized linear models: interpretable, efficient and useful for calibrated baselines.
• Decision trees and random forests: effective for mixed features and relatively interpretable.
• Support vector machines: useful for moderate-sized, well-engineered feature sets.
• Convolutional networks: suitable for structured images such as range-Doppler maps.
• Recurrent and temporal-convolution models: useful for evolving tracks and health signals.
• Transformers: useful for long-range dependencies, but potentially expensive and data-hungry.
• Autoencoders and density models: useful for anomaly detection when failure labels are scarce.

A baseline detector, such as a CFAR detector or a matched-filter pipeline, should be retained for comparison. AI should be shown to improve a mission metric, not merely a software benchmark.

Generalization and robustness

Radar data distributions change with aspect angle, weather, sensor hardware, software versions, geographic location and adversarial behaviour. Validation must therefore include:

• held-out geographic regions;
• held-out platforms or hardware units;
• different signal-to-noise ratios and clutter conditions;
• previously unseen target or interference classes;
• temporal drift;
• adversarial or deceptive inputs;
• missing and corrupted data.

Calibration is important. If a model reports 0.8 confidence, approximately 80% of comparable predictions should be correct under the stated conditions. Reliability diagrams, expected calibration error and conformal prediction can help quantify uncertainty, although conformal guarantees depend on exchangeability assumptions that may fail under severe distribution shift.

Explainability should be treated as a diagnostic and governance aid, not as proof of correctness. Saliency maps can be unstable, and a visually plausible explanation may not identify the true causal feature.

4. Distributed and Networked Radar Sensing

A distributed radar network uses multiple spatially separated transmitters, receivers or passive sensors. It can obtain diversity in aspect angle, reduce shadowing and improve localization. Architectures range from loosely coordinated sensors that share tracks to coherent multistatic systems that share phase-referenced samples.

Fusion levels

1. Data-level fusion: raw or near-raw samples are combined. It can provide high performance but demands bandwidth, synchronization and common calibration.

2. Feature-level fusion: extracted features are combined. This reduces communication load but may discard information.

3. Track-level fusion: local tracks and covariances are combined. It is scalable, but correlation between sensor estimates must be handled correctly.

If two estimates of the same target are correlated and treated as independent, fusion becomes overconfident. Covariance intersection is one conservative method when cross-correlation is unknown, although it can be less statistically efficient than fusion with known correlations.

Distributed systems require time synchronization, coordinate registration, sensor health monitoring, cyber protection and resource scheduling. Network latency can produce stale measurements, while a compromised node can inject false tracks. Trust management and robust consensus are therefore part of radar engineering, not merely network administration.

5. Passive Radar

Passive radar does not normally transmit its own dedicated illumination. It exploits existing signals such as digital broadcast, cellular, satellite or communication transmissions. A reference channel records the illuminator signal, while a surveillance channel receives direct and reflected components. Cross-correlation or matched processing estimates bistatic delay and Doppler.

The absence of a transmitter reduces spectrum occupation, electromagnetic visibility and hardware power requirements. It also creates limitations: the illuminator may be unavailable, its waveform may change, transmitter and receiver geometry may be unfavourable, and direct-path leakage and multipath can be strong. Passive radar performance is therefore environment-dependent.

Passive radar is not automatically covert or universally low cost. The receiver, processing chain, synchronization system and site infrastructure still require engineering. AI can classify illuminators, suppress structured interference and detect weak reflections, but it cannot create illumination energy that is not present.

6. Joint Communication and Radar Sensing

Joint communication-radar sensing, often called integrated sensing and communications (ISAC), seeks to use spectrum, hardware and signal processing for both data transmission and environmental sensing. A communication waveform may provide useful bandwidth and network connectivity, while radar sensing can support positioning, mapping, gesture recognition, vehicular safety and industrial automation.

The design involves trade-offs among data rate, sensing resolution, latency, energy, privacy and interference. A waveform optimized for communication may have poor ambiguity properties; a sensing waveform may reduce communication efficiency. Joint design can optimize a weighted objective or impose sensing and communication constraints simultaneously.

In a simplified range-resolution relation for a waveform with effective bandwidth \(B\),

\[

\Delta R \approx \frac{c}{2B},

\]

where \(c\) is propagation speed and the approximation assumes a sufficiently known waveform and a conventional monostatic range measurement. For bistatic geometries or complex propagation, the relationship must be interpreted in terms of bistatic range and the actual ambiguity function.

ISAC also changes privacy and governance. A communications infrastructure capable of sensing people, vehicles or room occupancy may collect information beyond what users expect. Consent, data minimization, access control and purpose limitation must be designed into the system.

7. Edge AI for Radar

Edge AI places inference near the sensor rather than sending all data to a remote cloud. Its advantages include low latency, reduced backhaul bandwidth, resilience to intermittent connectivity and improved control over sensitive measurements.

The constraints are severe: limited memory, thermal budget, power, processor capability and software update mechanisms. Common deployment methods include quantization, pruning, knowledge distillation, low-rank approximation and hardware-specific compilation. Quantization can reduce numerical precision, but its effect on weak-target detection and calibration must be measured rather than assumed.

A suitable edge architecture separates safety-critical signal processing from learning components. For example, a trusted pipeline may perform synchronization, calibration, physical plausibility checks and basic detection, while an AI module ranks tracks or recommends resource allocation. A watchdog can revert to a certified fallback mode when confidence, timing or health limits are violated.

Federated learning may allow multiple radar sites to train a shared model without centralizing raw data. However, model updates can leak information, and non-identical site distributions make convergence difficult. Secure aggregation, update validation and participation controls are necessary.

8. Quantum-Radar Claims: A Critical Assessment

Quantum radar is a broad and often ambiguous term. It may refer to quantum illumination, entangled-photon concepts, quantum-enhanced measurement, or simply highly sensitive classical radar marketed with quantum terminology. These are not equivalent.

Quantum illumination theoretically studies detection of a weakly reflecting object in a noisy environment using correlations between transmitted and retained quantum modes. Under certain idealized assumptions, a quantum strategy can offer an error-exponent advantage over a comparable classical strategy, even when entanglement is lost during propagation. The advantage is generally subtle and depends on receiver design, loss, noise, brightness, bandwidth and the comparison baseline.

Several claims require caution:

• Quantum correlation does not imply unlimited range or perfect detection.
• Entanglement cannot transmit information faster than light.
• A theoretical advantage may disappear with loss, thermal noise, imperfect sources and detector inefficiency.
• A laboratory photon-counting demonstration is not automatically a practical long-range radar.
• Classical coherent radar remains exceptionally capable and often more economical.

A credible evaluation should specify the transmitter, receiver, channel model, noise regime, resource accounting, range, integration time and classical benchmark. The appropriate question is not whether a system is labelled quantum, but whether it provides a reproducible mission-level benefit at acceptable cost and complexity.

9. An Integrated Cognitive-Radar Architecture

A robust architecture can be organized into six layers:

1. Sensing layer: antennas, RF front ends, timing, calibration and health monitoring.

2. Signal-processing layer: filtering, beamforming, matched processing, range-Doppler processing and CFAR.

3. Perception layer: detection, parameter estimation, classification and tracking.

4. Knowledge layer: environmental models, uncertainty, historical data and sensor trust.

5. Decision layer: waveform, beam, dwell, network and maintenance actions subject to constraints.

6. Governance layer: logging, authorization, cybersecurity, human supervision, testing and rollback.

The governance layer must be active during operation, not added after deployment. Every adaptive action should be logged with the input state, selected action, model version, confidence and constraint checks. This enables forensic analysis and certification.

10. Illustrative Case Study: Cognitive Maritime Surveillance Network

Background

Consider an illustrative, fictional coastal surveillance network with three shore-based passive receivers and one active phased-array radar. The mission is to detect small vessels in clutter while sharing spectrum with communication systems.

Problem

Sea clutter varies with wind and weather. Passive receivers experience changing illuminator availability, while the active radar faces intermittent interference. Fixed scheduling produces missed detections during clutter transitions and wastes energy during quiet periods.

Analysis

The network records calibrated range-Doppler features, local clutter statistics, illuminator quality, track continuity and interference indicators. A conventional CFAR detector remains the primary safety baseline. A supervised classifier identifies broad clutter regimes, while a constrained policy selects among a small set of approved active-radar waveforms and dwell times.

The policy objective rewards track continuity and detection probability, penalizes false alarms and energy use, and imposes hard limits on transmit duty cycle, authorized frequencies and maximum latency. Track-level fusion uses reported covariance and accounts for common measurements; when cross-correlation is unknown, a conservative fusion method is used.

Approach and solution

Inference is performed at the edge nodes. Only detections, features, uncertainty and health statistics are transmitted routinely; raw data are retained selectively for audit and retraining. A drift detector compares current feature distributions with validated operating ranges. If drift or model uncertainty exceeds a threshold, the network switches to a certified baseline schedule and alerts an operator.

Training data include different sea states, sensor sites, illuminators and interference types. Evaluation is performed by holding out one site and one weather regime. Metrics include \(P_D\) at fixed \(P_{FA}\), track continuity, energy per monitored sector, latency and false-alert workload.

Outcome

In this illustrative scenario, the adaptive system improves resource use and maintains more stable tracking during changing clutter, but it does not eliminate missed detections. The principal benefit arises from combining physical signal processing, uncertainty-aware fusion and constrained adaptation, rather than from an unconstrained neural network.

Learning points

• AI performance depends on representative environmental coverage.
• A fallback mode is essential for safety and certification.
• Distributed fusion must account for correlated information.
• Mission metrics are more meaningful than generic classification accuracy.
• Logging and operator review support accountability and improvement.

11. Challenges, Ethics and Regulation

Technical challenges

Key challenges include nonstationary clutter, scarce labels, rare-event learning, adversarial manipulation, calibration drift, latency, synchronization, model compression and the sim-to-real gap. A policy trained in simulation may exploit simulator artefacts and fail on real hardware. Hardware-in-the-loop and progressive field testing are therefore required.

Safety and assurance

Adaptive systems should have bounded action spaces, independent monitors, graceful degradation, deterministic fallback behaviour and version-controlled models. Verification should include requirements-based tests, stress tests, red-team evaluation and operationally realistic acceptance criteria.

Ethics and privacy

Radar sensing can reveal movement, occupancy, identity-related patterns and behaviour. Even when radar does not produce conventional images, inferred information may be sensitive. Governance should address lawful purpose, proportionality, retention, access, auditability and human accountability. Automated classification should not be treated as unquestionable evidence in consequential decisions.

Cybersecurity

Threats include data poisoning, adversarial examples, model extraction, command manipulation, compromised sensor nodes and spoofed time sources. Secure boot, signed models, authenticated updates, network segmentation, anomaly monitoring and incident response are essential.

Best-practice development lifecycle

A professional lifecycle consists of:

1. Define the mission metric and operating envelope.

2. Establish physics-based and classical baselines.

3. Collect traceable, diverse and representative data.

4. Train with leakage-resistant data partitioning.

5. Evaluate uncertainty, calibration and distribution shift.

6. Constrain adaptive actions and implement fallback modes.

7. Test in simulation, hardware-in-the-loop and field trials.

8. Monitor drift, health and model performance after deployment.

9. Revalidate every significant model, hardware or environment change.

Contemporary Research Directions

Current research includes self-supervised learning from unlabelled radar streams, foundation models for multimodal sensing, neural receivers, differentiable signal processing, distributed reinforcement learning, semantic communications, reconfigurable intelligent surfaces and cooperative autonomous sensing. These directions are promising but should be judged by reproducibility, computational cost, robustness and operational benefit.

A particularly important research problem is combining learned representations with physical constraints. Physics-informed or model-based learning can improve data efficiency and interpretability, but it must be demonstrated under realistic model mismatch. Another important direction is uncertainty-aware autonomy: the system should know when it lacks sufficient evidence and should request additional sensing or human intervention.

Chapter Summary

Cognitive radar extends conventional radar with an observe-infer-decide-act-learn loop. Its essential elements are an environmental belief state, an action policy, feedback and constraints. Machine learning can assist detection, classification, waveform adaptation, interference mitigation and predictive maintenance, but evaluation must use radar-specific measures such as probability of detection at a controlled false-alarm rate, track continuity, latency and resource use.

Distributed radar improves spatial diversity but introduces synchronization, fusion, correlation and cybersecurity problems. Passive radar exploits existing illuminators and is highly environment-dependent. Joint communication-radar sensing shares spectrum and hardware while creating trade-offs among sensing quality, communication performance and privacy. Edge AI reduces latency and bandwidth requirements but requires model compression, monitoring and safe fallback operation.

Quantum-radar claims must be assessed against explicit channel models, receiver assumptions and classical baselines. No emerging technology removes the need for electromagnetic theory, calibration, statistical validation, governance and human accountability. The most credible cognitive-radar systems combine learned components with established signal processing, physical constraints, uncertainty estimation and disciplined engineering assurance.

Key Terms

Cognitive radar: Radar that uses observations and feedback to adapt sensing actions toward mission objectives.

Belief state: A probabilistic representation of the estimated environment and its uncertainty.

Policy: A rule or mapping that selects an action from an observed state.

Constant false alarm rate: Detection processing designed to maintain a specified false-alarm behaviour under changing background conditions.

Probability of detection: Probability that a target is correctly detected under specified conditions.

Distribution shift: A change between the data distribution used for development and that encountered in operation.

Open-set recognition: Recognition that permits an unknown or novel class instead of forcing a known label.

Reinforcement learning: Learning of action policies through states, actions, rewards and transitions.

Predictive maintenance: Estimation of degradation or failure risk to schedule maintenance proactively.

Distributed radar: A radar network using spatially separated sensing elements.

Multistatic radar: Radar in which transmitter and receiver locations are distinct.

Track-level fusion: Combination of local target tracks and their uncertainty information.

Passive radar: Radar that uses illuminators not dedicated to the radar system.

ISAC: Integrated sensing and communications using shared signals, spectrum or hardware.

Edge AI: Machine-learning inference performed near the sensor or data source.

Model calibration: Agreement between stated prediction confidence and observed correctness.

Fallback mode: A trusted operating mode used when adaptive components are uncertain or faulty.

Quantum illumination: A quantum sensing framework for detecting weak reflectors in noisy environments.

Sim-to-real gap: Performance difference between simulation and physical deployment.

Concept drift: Change over time in the statistical relationship between observations and operational outcomes.

Review Questions

Short Answer Questions

1. What distinguishes a cognitive radar from a conventional adaptive radar?

2. Why is probability of false alarm important when evaluating an AI detector?

3. State two advantages and two limitations of passive radar.

4. What is the purpose of a fallback mode in an edge-AI radar?

5. Why must correlated information be considered in distributed track fusion?

Long Answer Questions

1. Explain the architecture of a cognitive radar and discuss the role of the belief state, policy and feedback loop.

2. Compare supervised learning, anomaly detection and reinforcement learning for radar applications.

3. Discuss the technical and governance challenges of deploying AI for waveform adaptation.

4. Explain the principal trade-offs in integrated sensing and communications.

5. Critically evaluate the practical significance of quantum illumination for radar engineering.

Analytical / Application Questions

1. Design an evaluation plan for a neural detector intended for operation across multiple climates and radar platforms. Specify data partitioning, metrics, uncertainty tests and fallback criteria.

2. A distributed radar network receives two local tracks that may have common measurements. Explain why naïve covariance fusion can be overconfident and propose a conservative strategy.

3. Develop a constrained cognitive-radar policy for an interference-rich environment. Identify the state variables, allowable actions, reward or utility terms, hard constraints and monitoring mechanisms.

Multiple Choice Questions with Answers

Question 1

Which feature most fundamentally distinguishes a cognitive radar?

A. Use of a high-power transmitter

B. Closed-loop adaptation based on observations and an internal knowledge state

C. Operation only with passive illuminators

D. Use of a fixed waveform and fixed threshold

Correct Answer: B. Closed-loop adaptation based on observations and an internal knowledge state

Question 2

Which metric is especially important when comparing radar detectors under operational conditions?

A. Training-set accuracy alone

B. Number of neural-network layers

C. Probability of detection at a specified probability of false alarm

D. File size of the trained model

Correct Answer: C. Probability of detection at a specified probability of false alarm

Question 3

In reinforcement learning for waveform selection, what does the policy determine?

A. The antenna material

B. The action selected for a given state

C. The speed of light

D. The radar’s manufacturing cost

Correct Answer: B. The action selected for a given state

Question 4

What is a major limitation of passive radar?

A. It cannot use any digital processing

B. It requires an available and suitable external illuminator

C. It always transmits at maximum power

D. It is independent of geometry

Correct Answer: B. It requires an available and suitable external illuminator

Question 5

Why can naïve fusion of two correlated tracks be dangerous?

A. It always increases communication bandwidth to zero

B. It can produce unjustifiably small uncertainty estimates

C. It prevents all target detections

D. It eliminates the need for synchronization

Correct Answer: B. It can produce unjustifiably small uncertainty estimates

Question 6

For effective bandwidth B in a conventional monostatic range measurement, approximate range resolution is

A. cB/2

B. 2B/c

C. c/(2B)

D. B/(2c)

Correct Answer: C. c/(2B)

Question 7

Which deployment strategy is most appropriate for safety-critical edge AI?

A. Allow unrestricted actions with no monitoring

B. Use only cloud processing with no local fallback

C. Combine constrained AI decisions with a trusted fallback mode

D. Disable calibration to reduce latency

Correct Answer: C. Combine constrained AI decisions with a trusted fallback mode

Question 8

What is open-set recognition intended to do?

A. Force every observation into a known class

B. Recognize that an observation may belong to an unknown class

C. Remove all uncertainty estimates

D. Replace the radar antenna

Correct Answer: B. Recognize that an observation may belong to an unknown class

Question 9

Which statement about quantum radar is most scientifically defensible?

A. Quantum correlations guarantee unlimited detection range

B. Entanglement permits faster-than-light information transfer

C. Any laboratory quantum experiment automatically outperforms classical radar

D. Claims require comparison with an explicit classical baseline and realistic loss and noise models

Correct Answer: D. Claims require comparison with an explicit classical baseline and realistic loss and noise models

Question 10

Why should radar AI data commonly be partitioned by hardware unit or geographic site?

A. To make data leakage more likely

B. To test generalization to unseen units or environments

C. To eliminate the need for labels

D. To guarantee perfect detection

Correct Answer: B. To test generalization to unseen units or environments