Mobile Sensing, Analytics and Artificial Intelligence

Introduction and Context

Mobile computing systems have evolved from communication platforms into distributed sensing and decision-making infrastructures. Smartphones, wearable devices, connected vehicles, wireless access points, cellular networks, Internet of Things (IoT) devices and edge servers continuously generate observations about people, machines, locations and environments. These observations support positioning, mobility prediction, network optimisation, public-health monitoring, industrial automation, intelligent transportation and context-aware services.

The value of mobile data does not arise merely from its volume. It depends on the quality of sensing, the temporal and spatial resolution of measurements, the ability to combine heterogeneous sources, and the correctness, fairness and accountability of the resulting decisions. Mobile sensing therefore combines signal processing, statistics, machine learning, wireless networking, distributed systems and human-centred governance.

This chapter examines how mobility traces, sensor observations and network telemetry are converted into useful intelligence. It focuses on four major capabilities: positioning, prediction, anomaly detection and optimisation. It then evaluates federated learning, edge artificial intelligence (AI), reinforcement learning and responsible AI in mobile systems.

Learning Objectives

After studying this chapter, a student should be able to:

1. Explain the architecture and data characteristics of mobile sensing systems.

2. Apply basic positioning and sensor-fusion models and identify their assumptions.

3. Select suitable analytical methods for prediction and anomaly detection.

4. Explain how network telemetry supports performance diagnosis and optimisation.

5. Compare cloud, edge, federated and split-learning approaches.

6. Formulate mobile resource-management problems using reinforcement learning.

7. Evaluate privacy, security, fairness, safety and regulatory issues in mobile AI.

8. Design an evidence-based mobile analytics pipeline and assess its limitations.

1. Mobile Sensing as a Distributed Cyber-Physical System

A mobile sensing system consists of sensing devices, communication links, computational resources, analytical models and decision mechanisms. A useful abstraction is:

\[

\mathcal{M}=(D,S,N,C,A,G),

\]

where \(D\) denotes devices, \(S\) sensor sources, \(N\) the communication network, \(C\) computational resources, \(A\) analytical and decision algorithms, and \(G\) governance controls. The abstraction is not a performance equation; it identifies the interacting components that must be designed together.

1.1 Sources of mobile data

Important sources include:

• Inertial sensors: accelerometers, gyroscopes and magnetometers, used for activity recognition, pedestrian dead reckoning and orientation estimation.
• Radio measurements: received signal strength (RSS), channel state information (CSI), time of arrival, angle of arrival, serving-cell identity and handover records.
• Satellite positioning: global navigation satellite system (GNSS) pseudoranges and positions.
• Environmental sensors: temperature, humidity, pressure, light, sound and air-quality sensors.
• Device and application telemetry: battery level, CPU utilisation, application events, packet loss and response time.
• Mobility traces: sequences of locations, cell associations, road segments, visits or transitions over time.
• Network telemetry: throughput, latency, jitter, queue occupancy, retransmissions, signal quality and resource-block utilisation.

These sources differ in sampling rate, accuracy, energy cost, coverage, privacy sensitivity and failure modes. An accelerometer may sample at tens or hundreds of hertz but accumulate integration error. A cellular record may cover a wide area but provide coarse spatial resolution. GNSS can be accurate outdoors but unreliable indoors or in urban canyons.

1.2 Data lifecycle

A robust sensing pipeline includes sensing, time synchronisation, preprocessing, feature construction, transmission, storage, modelling, decision-making and feedback. Errors at early stages propagate downstream. For example, clock offsets can create false velocity estimates; missing labels can make a classifier appear accurate while failing on rare events; and changes in device hardware can produce a distribution shift.

Common preprocessing operations include outlier screening, calibration, interpolation, resampling, coordinate transformation and privacy filtering. However, preprocessing must not remove meaningful rare events. A sudden acceleration may be noise, a collision, or a genuine emergency; its interpretation requires context.

1.3 Spatial and temporal reasoning

Mobile analytics is inherently spatiotemporal. A trace can be represented as

\[

\tau_i=\{(t_k,\mathbf{x}_k,\mathbf{z}_k)\}_{k=1}^{n_i},

\]

where \(t_k\) is time, \(\mathbf{x}_k\) is a position or location representation, and \(\mathbf{z}_k\) is a vector of sensor or network observations. The trace may be irregularly sampled, incomplete and uncertain. Treating it as a simple table can conceal temporal dependence and spatial bias.

Useful representations include trajectories, origin-destination matrices, road-network paths, cell-transition graphs and spatial grids. The appropriate representation depends on the application. A grid may support heat-map analysis, whereas a graph is more suitable for routing or handover prediction.

2. Positioning and Sensor Fusion

Positioning estimates the state of a device, person or vehicle from direct and indirect observations. A state may include position, velocity, heading and sensor biases. Positioning systems must distinguish measurement accuracy from confidence: a numerical estimate without uncertainty is insufficient for safety-critical decisions.

2.1 Geometric positioning

For a two-dimensional position \(\mathbf{p}=(x,y)\) and known anchor \(\mathbf{a}_i=(x_i,y_i)\), an ideal range measurement satisfies

\[

r_i=\|\mathbf{p}-\mathbf{a}_i\|+\epsilon_i,

\]

where \(r_i\) is the measured distance and \(\epsilon_i\) is measurement error. With at least three non-collinear anchors, position can in principle be estimated in two dimensions. In practice, multipath, non-line-of-sight propagation, anchor uncertainty and synchronization errors make the problem nonlinear and noisy.

A weighted least-squares estimate can be written as

\[

\hat{\mathbf{p}}=\arg\min_{\mathbf{p}}\sum_{i=1}^{m}w_i\left(r_i-\|\mathbf{p}-\mathbf{a}_i\|\right)^2,

\]

where \(w_i\geq0\) represents confidence in measurement \(i\). This formulation assumes that the anchor coordinates are known, the measurements refer to a common time, and the selected error model is appropriate. It is not guaranteed to find a global optimum without suitable numerical methods.

2.2 GNSS and radio positioning

GNSS estimates position from signal timing and satellite ephemeris information. Errors arise from satellite-clock errors, atmospheric delay, multipath and receiver noise. Cellular positioning uses cell identity, timing advance, observed signal levels and more advanced time or angle measurements. Wi-Fi and Bluetooth positioning may use fingerprinting, in which radio observations are compared with a site survey database.

Fingerprinting often performs well in complex indoor spaces but requires database maintenance. Changes in access-point placement, people density, furniture and radio conditions can reduce accuracy. A model trained in one building should not automatically be assumed valid in another.

2.3 Inertial navigation and drift

Inertial navigation integrates acceleration and angular-rate measurements. In simplified one-dimensional form,

\[

v(t)=v(t_0)+\int_{t_0}^{t}a(\tau)d\tau,

\]

\[

s(t)=s(t_0)+\int_{t_0}^{t}v(\tau)d\tau.

\]

Here \(a\) is acceleration, \(v\) velocity and \(s\) displacement. These relationships assume that acceleration has been expressed in a consistent reference frame and that gravity and sensor bias have been compensated. Since integration accumulates bias and noise, inertial-only positioning drifts over time.

2.4 Kalman filtering and probabilistic fusion

A linear state-space model is

\[

\mathbf{x}_k=\mathbf{F}_k\mathbf{x}_{k-1}+\mathbf{B}_k\mathbf{u}_k+\mathbf{w}_k,

\]

\[

\mathbf{y}_k=\mathbf{H}_k\mathbf{x}_k+\mathbf{v}_k,

\]

where \(\mathbf{x}_k\) is the hidden state, \(\mathbf{u}_k\) a known input, \(\mathbf{y}_k\) the measurement, and \(\mathbf{w}_k,\mathbf{v}_k\) process and measurement noise. The classical Kalman filter is optimal for linear models with Gaussian, zero-mean, independent noises with known covariance matrices. Extended and unscented variants address certain nonlinear models, but their assumptions and numerical behaviour must still be checked.

Sensor fusion may combine GNSS, inertial measurements, map constraints, Wi-Fi and cellular observations. Fusion improves robustness when sensors fail independently, but correlated errors can make confidence estimates over-optimistic. A system should therefore monitor residuals, sensor health and uncertainty.

Worked Example: Range-Based Positioning

Suppose a device is at \((3,4)\) m and an anchor is at \((0,0)\) m. The ideal range is

\[

r=\sqrt{(3-0)^2+(4-0)^2}=5\text{ m}.

\]

If the measured range is 5.4 m, the scalar range error is 0.4 m. This example illustrates measurement error; it does not determine the two-dimensional position by itself. At least additional independent geometric information is required.

3. Analytics for Prediction

Prediction estimates an unknown future or unobserved value. Mobile prediction tasks include travel time, next cell association, demand, traffic density, battery depletion, radio quality and user activity.

3.1 Feature and target design

A prediction problem should define the prediction horizon, unit of analysis, target variable, available information and evaluation protocol. Features may include recent locations, speed, heading, time of day, day of week, weather, network load and historical transitions. Features must be available at prediction time; using future information creates leakage and invalidates evaluation.

For a sequence of locations, a Markov model approximates

\[

P(X_{t+1}=j\mid X_t=i, X_{t-1},\ldots)\approx P(X_{t+1}=j\mid X_t=i).

\]

This first-order assumption is useful for simple mobility modelling but may be inadequate when destinations, routines or longer history matter. Higher-order Markov models, hidden Markov models, recurrent neural networks, temporal convolutional networks and transformers provide alternatives, with increasing data and computational requirements.

3.2 Evaluation

Classification should use metrics appropriate to class balance and cost. Accuracy is

\[

\text{Accuracy}=\frac{TP+TN}{TP+TN+FP+FN},

\]

where TP, TN, FP and FN denote true positives, true negatives, false positives and false negatives. Accuracy can be misleading for rare events. Precision, recall, F1 score, calibration error and area under a precision-recall curve may be more informative.

For a continuous prediction \(\hat{y}_i\) of target \(y_i\), mean absolute error is

\[

MAE=\frac{1}{n}\sum_{i=1}^{n}|y_i-\hat{y}_i|.

\]

MAE has the same units as the target and is less sensitive to large errors than mean squared error. Evaluation should use temporal or geographical holdout sets where deployment conditions demand generalisation across time or place.

3.3 Uncertainty and distribution shift

A point prediction conceals uncertainty. Prediction intervals, quantile regression, Bayesian methods, ensembles or conformal techniques can communicate uncertainty, subject to their assumptions. Mobile environments change because of software updates, new roads, seasonal behaviour, radio deployment and demographic changes. Monitoring input distributions and error rates is therefore part of model operation, not an optional research activity.

4. Anomaly Detection in Mobile Systems

An anomaly is an observation or pattern that differs materially from an expected distribution or operational baseline. Examples include a sudden traffic surge, abnormal handover frequency, sensor tampering, unusual battery drain, spoofed location or a device behaving differently from its peer group.

4.1 Types of anomaly

• Point anomaly: one observation is abnormal.
• Contextual anomaly: an observation is abnormal in a particular context, such as high traffic at a normally quiet time.
• Collective anomaly: a sequence is abnormal even when individual observations appear ordinary.
• Spatial anomaly: a location or region differs from comparable regions.
• Concept drift: the normal generating process changes over time.

4.2 Methods

Statistical methods estimate normal ranges or probability distributions. Clustering methods identify observations distant from dense groups. Isolation forests exploit the fact that unusual points are often isolated by fewer random partitions. One-class support vector machines learn a boundary around normal data. Autoencoders learn to reconstruct normal patterns; high reconstruction error may indicate an anomaly.

No detector is universally reliable. An unsupervised model may detect novelty rather than danger. A rare but legitimate event can be flagged, while a systematic attack resembling normal traffic can be missed. Thresholds should be chosen using operational costs, available labels and the acceptable false-alarm rate.

A useful operational measure is alert precision: the proportion of alerts confirmed as actionable. High alert volume with low precision causes alarm fatigue. Human investigation, feedback and post-event labelling are essential for improving the detector.

5. Network Telemetry and Mobile Optimisation

Network telemetry turns infrastructure behaviour into observable variables. It supports fault management, capacity planning, quality-of-service assurance, energy management and security operations.

Important indicators include throughput, one-way or round-trip latency, jitter, packet loss, retransmission rate, radio signal quality, handover failure rate, queue length, processor load and energy consumption. These indicators should be interpreted jointly. Low throughput may arise from weak signal, congestion, transport-layer control, server limitation or device power management.

A common service-level relationship is that utilisation influences delay, but simple queueing formulas require explicit assumptions. For an M/M/1 queue with Poisson arrivals, exponential service times and utilisation \(\rho=\lambda/\mu<1\), the mean time in the system is

\[

W=\frac{1}{\mu-\lambda}.

\]

Here \(\lambda\) is the arrival rate and \(\mu\) the service rate, both in jobs per second; \(W\) is in seconds. Real mobile networks rarely satisfy all M/M/1 assumptions, so this formula is a conceptual baseline rather than a direct network-performance law.

Optimisation objectives may combine latency, energy, cost, reliability and fairness. A generic constrained problem is

\[

\min_{\mathbf{a}}\; J(\mathbf{a})

\quad\text{subject to}\quad g_j(\mathbf{a})\leq0,

\]

where \(\mathbf{a}\) represents decisions such as computing placement, transmit power or scheduling, \(J\) is a measurable objective, and \(g_j\) are resource or service constraints. A weighted objective must be justified because weights encode policy priorities; they are not merely technical constants.

6. Federated Learning and Edge AI

6.1 Cloud, edge and federated architectures

In cloud learning, data or large feature sets are sent to central servers. This offers substantial computation and global coordination but introduces backhaul cost, latency and privacy exposure. Edge AI executes inference near the data source, reducing latency and bandwidth use. Federated learning keeps training data on participating devices or sites while exchanging model updates.

In synchronous federated averaging, client \(k\) has \(n_k\) training examples and model parameters \(\mathbf{w}_k

times. If \(n=\sum_k n_k\), the server update is

\[

\mathbf{w}_{t+1}=\sum_{k=1}^{K}\frac{n_k}{n}\mathbf{w}_{t+1}^{(k)}.

\]

This weighted average is appropriate when client updates are compatible and example counts are a meaningful weighting basis. It does not guarantee convergence under highly non-independent and non-identically distributed data, client dropouts, device constraints or adversarial updates.

6.2 Federated-learning challenges

Mobile clients differ in data volume, connectivity, hardware, battery and local behaviour. Their data are non-IID: one phone may observe walking, another driving, and another no examples of the target event. Communication can dominate training cost. Model updates can leak information through gradient or update analysis, and malicious clients may poison the global model.

Mitigations include secure aggregation, differential privacy, robust aggregation, client sampling, update clipping, personalisation and anomaly screening. Each introduces trade-offs. Differential privacy can reduce disclosure risk but may reduce accuracy; secure aggregation hides individual updates from the server but complicates debugging; robust aggregation can resist some outliers but may perform poorly with heterogeneous honest clients.

6.3 Edge AI design

Edge deployment requires model compression, quantisation, pruning, knowledge distillation, hardware-aware compilation and energy profiling. Accuracy alone is insufficient. Designers should report latency distribution, peak memory, energy per inference, thermal effects, update mechanism and failure behaviour. An edge model should degrade safely when input quality is poor or connectivity is lost.

Split learning divides a model between device and server. It can reduce local computation or avoid transmitting raw data, but intermediate activations may still reveal information and communication remains necessary. Architecture selection must therefore be driven by threat models and service requirements rather than by the label “privacy-preserving.”

7. Reinforcement Learning for Mobile Decision-Making

Reinforcement learning (RL) models sequential decisions under uncertainty. A Markov decision process is represented by \((\mathcal{S},\mathcal{A},P,R,\gamma)\), where \(\mathcal{S}\) is the state space, \(\mathcal{A}\) the action space, \(P\) transition dynamics, \(R\) reward, and \(\gamma\in[0,1)\) the discount factor.

A policy \(\pi(a\mid s)\) specifies action probabilities. The action-value function is

\[

Q^\pi(s,a)=\mathbb{E}_\pi\left[\sum_{t=0}^{\infty}\gamma^tR_{t+1}\mid S_0=s,A_0=a\right].

\]

The expectation is over transitions and any policy randomness. RL can support handover selection, spectrum allocation, computation offloading, caching, drone control and battery-aware task scheduling.

The reward must represent the real objective. For example, a reward that strongly favours throughput may cause excessive energy consumption or unfair resource allocation. Multi-objective or constrained RL is often more suitable than a single unconstrained score. Exploration is also dangerous in live networks: random actions can degrade service. Safe exploration, simulation, offline datasets, conservative policy improvement and human override are important.

RL suffers from partial observability, non-stationary environments, sparse rewards and sim-to-real mismatch. A simulator may omit interference, user reactions or failures. Policies should therefore be stress-tested and monitored after deployment.

8. Responsible AI for Mobile Systems

Mobile AI operates close to identity, movement, communication and daily behaviour. Responsible design must address:

• Purpose limitation: collect and use data for specified legitimate purposes.
• Data minimisation: retain only the information needed for the task.
• Consent and lawful basis: provide meaningful information and valid controls where required.
• Privacy: use access controls, encryption, retention limits, anonymisation or pseudonymisation, and privacy-preserving computation where suitable.
• Fairness: test performance across relevant groups, locations, devices and connectivity conditions.
• Transparency: communicate what is measured, inferred and decided.
• Accountability: assign responsibility for model, data and operational decisions.
• Safety and security: protect against spoofing, poisoning, evasion, model theft and unsafe automation.
• Human oversight: provide escalation and appeal mechanisms for consequential decisions.

Location data are particularly sensitive because apparently anonymous trajectories can be re-identified through distinctive movement patterns. Removing names does not automatically make a dataset anonymous. Governance should include a data inventory, threat assessment, access policy, model documentation, impact assessment and incident-response process. Applicable privacy, telecommunications, cybersecurity, consumer-protection and AI regulations vary by jurisdiction and use case; technical teams must involve legal and compliance specialists rather than assume one universal rule.

Fairness is not a single metric. A positioning model may have unequal error by environment; a mobility predictor may underperform for people with irregular travel; and an anomaly detector may generate more alerts in communities with denser monitoring. Evaluation must examine who is represented, who bears errors and whether the application itself is justified.

9. Integrated Illustrative Case: Edge Intelligence for Urban Bus Operations

The following is an illustrative case, not a report of a particular organisation.

Background

A metropolitan bus operator experiences irregular headways, overcrowded vehicles and poor passenger information during peak periods. Buses provide GNSS positions, door events, passenger-count estimates, radio quality and vehicle telemetry. The operator also receives weather and road-incident data.

Problem

The operator must predict arrival times, identify abnormal delays, select communication and edge-computing resources, and preserve passenger privacy. A cloud-only design would transmit large volumes of raw sensor data and might respond too slowly during connectivity disruptions.

Analysis

The data are heterogeneous and asynchronous. GNSS has occasional multipath errors, passenger counts are noisy, and radio quality changes with route and weather. Historical data overrepresent busy routes, so a global model may perform poorly on low-frequency services. Delay is not independent between buses: congestion and bunching create spatial and temporal dependencies.

Approach and solution

The operator first establishes time synchronisation, sensor-quality flags and a route-map representation. An edge model on each bus estimates short-horizon arrival time from recent position, speed, dwell time and route context. A central service aggregates encrypted, quality-controlled updates. Federated training is used when raw passenger and telemetry records should remain on the vehicle or depot. Secure aggregation and update clipping reduce exposure to individual updates.

A temporal prediction model is evaluated using route- and date-based holdout sets. MAE is reported separately for peak and off-peak periods, weather conditions and route groups. An anomaly detector flags unusual dwell time, repeated handover failure and implausible position jumps. Alerts are ranked by operational impact rather than sent indiscriminately to staff.

A constrained RL policy is tested in a digital simulation to recommend bus holding or dispatch adjustments. The policy cannot directly issue unsafe commands; it proposes actions subject to headway, driver, accessibility and service constraints. Human dispatchers approve high-impact interventions.

Outcome

The design reduces raw-data transmission, supports operation during intermittent backhaul and exposes performance differences across routes. It also reveals that a small improvement in average arrival-time error does not necessarily improve passenger experience if rare severe delays remain undetected. The operator therefore adopts tail-error reporting, fairness checks across routes and periodic model recalibration.

Learning points

1. A mobile AI system is a socio-technical system, not only a predictive model.

2. Edge processing can improve responsiveness, but it does not remove governance obligations.

3. Federated learning addresses data locality, not every privacy or security threat.

4. Evaluation must reflect operational subgroups and rare but consequential failures.

5. Automation should be constrained when actions affect safety or public access.

10. Best Practices and Research Directions

A professional mobile analytics project should:

1. Define the operational decision before selecting an algorithm.

2. Record sensor provenance, calibration, timestamps and missingness.

3. Establish baselines before deploying complex deep models.

4. Separate training, validation and test data by time, device or location where appropriate.

5. Report uncertainty, subgroup performance and failure cases.

6. Measure communication, memory, latency and energy, not only accuracy.

7. Use threat modelling for data, models, devices and communication channels.

8. Introduce rollback, monitoring, drift detection and model-update governance.

9. Validate in representative field conditions rather than only laboratory data.

10. Keep humans accountable for consequential decisions.

Contemporary research includes multimodal foundation models for sensor streams, neuromorphic and tinyML hardware, privacy-preserving sensing, semantic communications, graph learning for mobility networks, digital twins for safe RL training, 6G-native AI and continual learning. These directions raise unresolved questions about energy use, reproducibility, intellectual property, model ownership, long-term drift and the environmental cost of large-scale computation.

Challenges and Limitations

Mobile sensing is affected by missing data, calibration differences, sampling bias, device heterogeneity and unreliable connectivity. Mobility traces can encode sensitive relationships even after direct identifiers are removed. Deep models may learn spurious correlations, while RL policies may exploit imperfections in reward design. Federated learning can produce an apparently global model that performs poorly for minority clients. Edge AI may reduce latency but increase the number of devices that must be secured and updated.

Consequently, no single architecture or algorithm is universally superior. The correct design depends on latency, accuracy, energy, privacy, regulatory requirements, connectivity, safety and the consequences of error. Technical evaluation should be combined with stakeholder analysis and operational governance.

Chapter Summary

Mobile sensing systems integrate heterogeneous devices, networks, computation and governance to transform observations into decisions. Positioning combines geometric, radio, satellite and inertial information, but every estimate must be interpreted with uncertainty and sensor assumptions. Prediction uses temporal and spatial structure to forecast mobility, demand, radio conditions and service behaviour; valid evaluation requires leakage prevention, appropriate metrics and tests under distribution shift. Anomaly detection identifies unusual point, contextual, collective and spatial behaviour, but thresholds must reflect operational costs and false-alarm consequences.

Network telemetry supports diagnosis and optimisation of latency, throughput, reliability, energy and resource use. Edge AI reduces latency and bandwidth demand, while federated learning keeps training data local but introduces non-IID data, communication, privacy and robustness challenges. Reinforcement learning provides a framework for sequential mobile decisions, yet reward design, safe exploration, partial observability and sim-to-real transfer remain major concerns. Responsible AI requires privacy, fairness, transparency, accountability, security, safety and human oversight. The central design principle is that mobile AI must be evaluated as a complete socio-technical system rather than as an isolated model.

Key Terms

Mobile sensing: Collection of observations using mobile, wearable, vehicular or distributed devices.

Mobility trace: A time-ordered record of movement, locations or network associations.

Network telemetry: Operational measurements describing communication and computing behaviour.

Sensor fusion: Combining observations from multiple sensors to estimate a state or event.

Fingerprinting: Position estimation by matching observed radio features to a reference database.

State-space model: A model describing hidden system states and their noisy observations.

Kalman filter: A recursive estimator for linear Gaussian state-space models.

Distribution shift: Change between the data-generating conditions used for training and deployment.

Anomaly detection: Identification of observations or patterns that depart from expected behaviour.

Edge AI: AI inference or training performed near the data source.

Federated learning: Collaborative model training in which raw client data remain local.

Secure aggregation: A protocol that allows aggregation of updates without revealing individual updates.

Differential privacy: A formal privacy framework limiting the effect of one record on an output.

Non-IID data: Data distributions that differ across clients, devices or locations.

Reinforcement learning: Learning sequential decisions from interaction and reward.

Markov decision process: A formal model of states, actions, transitions, rewards and discounting.

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

Responsible AI: Design and governance ensuring lawful, fair, transparent, secure and accountable AI.

Model calibration: Agreement between predicted confidence and observed event frequencies.

TinyML: Machine learning designed for highly resource-constrained embedded devices.

Review Questions

Short Answer Questions

1. Why can inertial-only positioning drift over time?

2. Distinguish a point anomaly from a collective anomaly.

3. What is the purpose of secure aggregation in federated learning?

4. Why can accuracy be misleading for rare-event detection?

5. State two operational constraints that should be considered when deploying edge AI.

Long Answer Questions

1. Explain a complete mobile sensing data lifecycle, including synchronisation, preprocessing, modelling and governance.

2. Compare GNSS, radio fingerprinting and inertial positioning in terms of accuracy, coverage, cost and failure modes.

3. Explain federated learning and critically evaluate its privacy, communication and statistical limitations.

4. Formulate the use of reinforcement learning for mobile computation offloading and discuss safe exploration.

5. Discuss how responsible AI principles should influence the design and evaluation of a location-aware mobile service.

Analytical / Application Questions

1. A positioning system combines GNSS, inertial and Wi-Fi observations. Design a sensor-fusion strategy and identify at least four failure conditions that should trigger uncertainty monitoring or fallback behaviour.

2. A network anomaly detector produces many alerts during a city festival. Propose an analysis that distinguishes genuine congestion from contextual anomalies and explain which metrics and additional data you would use.

3. Design an edge-and-federated architecture for activity recognition across heterogeneous smartphones. Specify the data retained locally, the exchanged information, evaluation subgroups, privacy protections and model-update safeguards.

Multiple Choice Questions with Answers

Question 1

Which condition is required for the classical Kalman filter to be optimal in its standard formulation?

A. The system must have no measurement noise

B. The model must be linear with Gaussian noise assumptions

C. All sensors must sample at different rates

D. The state must be directly observable without estimation

Correct Answer: B. The model must be linear with Gaussian noise assumptions

Question 2

What is a principal limitation of inertial-only positioning?

A. It cannot measure acceleration

B. It requires fixed radio anchors

C. Integration causes accumulated drift from bias and noise

D. It operates only when GNSS is available

Correct Answer: C. Integration causes accumulated drift from bias and noise

Question 3

In federated averaging, why are client updates commonly weighted by the number of local examples?

A. To give larger datasets proportional influence

B. To eliminate all non-IID data

C. To guarantee differential privacy

D. To avoid sending model updates

Correct Answer: A. To give larger datasets proportional influence

Question 4

Which metric is generally more informative than accuracy for a highly imbalanced rare-event detector?

A. Screen resolution

B. Precision and recall

C. Clock frequency only

D. Number of model parameters only

Correct Answer: B. Precision and recall

Question 5

What does edge AI primarily seek to reduce?

A. The need for any sensor calibration

B. Local computation and all device energy use

C. Latency and backhaul data transmission

D. The need for model monitoring

Correct Answer: C. Latency and backhaul data transmission

Question 6

In a Markov decision process, what does the reward represent?

A. The complete device hardware specification

B. The immediate utility or cost associated with an action and outcome

C. The number of clients in federated learning

D. The sensor sampling frequency

Correct Answer: B. The immediate utility or cost associated with an action and outcome

Question 7

Which example is a contextual anomaly?

A. A corrupted single data packet in every situation

B. A high traffic level during a normally quiet overnight period

C. A normal temperature reading

D. A model parameter stored on a server

Correct Answer: B. A high traffic level during a normally quiet overnight period

Question 8

What is a major privacy risk of supposedly anonymous mobility traces?

A. They cannot be stored digitally

B. Distinctive movement patterns may enable re-identification

C. They always contain encryption keys

D. They prevent any statistical analysis

Correct Answer: B. Distinctive movement patterns may enable re-identification

Question 9

For an M/M/1 queue with arrival rate λ and service rate μ, which condition is required for the stated steady-state mean delay formula to apply?

A. λ is greater than μ

B. λ equals zero only

C. The utilisation λ/μ is less than one

D. The service rate is unknown

Correct Answer: C. The utilisation λ/μ is less than one

Question 10

Why should reinforcement-learning exploration be constrained in a live mobile network?

A. Exploration can degrade service or violate safety constraints

B. Exploration always improves fairness automatically

C. Rewards are unnecessary in RL

D. Mobile networks have no changing states

Correct Answer: A. Exploration can degrade service or violate safety constraints