Machine Learning Foundations for IoT

Introduction and Context

The Internet of Things (IoT) connects sensors, actuators, devices, gateways, communication networks, cloud platforms, and human users. These systems generate continuous streams of observations about physical processes: temperature, vibration, pressure, energy consumption, location, sound, images, and machine states. Machine learning (ML) provides methods for converting these observations into predictions, classifications, anomaly alerts, recommendations, and control actions.

Applying ML to IoT is not simply a matter of selecting an algorithm and fitting it to a dataset. IoT data are commonly noisy, incomplete, time-dependent, imbalanced, distributed across devices, and affected by changes in the environment. A model that performs well in a laboratory may fail after deployment because sensors drift, operating conditions change, labels are delayed, or communication constraints prevent cloud-based processing. Consequently, IoT machine learning requires an end-to-end perspective that includes sensing, data quality, feature engineering, modelling, validation, deployment, monitoring, and responsible governance.

This chapter develops the foundations needed to apply supervised, unsupervised, and reinforcement learning to IoT data. Particular attention is given to temporal dependence, noisy labels, class imbalance, concept drift, uncertainty, and the difference between predictive performance and operational usefulness.

Learning Objectives

After studying this chapter, you should be able to:

1. Explain how supervised, unsupervised, and reinforcement learning support IoT applications.

2. Design features from time-series and multimodal sensor data.

3. Select suitable models according to the prediction task, data structure, latency, and resource constraints.

4. Construct validation strategies that avoid temporal leakage.

5. Evaluate models using metrics appropriate to imbalanced and cost-sensitive IoT problems.

6. Diagnose the effects of noisy labels, missing data, concept drift, and uncertainty.

7. Propose an end-to-end ML solution for an IoT application while considering safety, privacy, and governance.

1. IoT Data as a Machine-Learning Problem

An IoT observation usually contains a value, a timestamp, a device identifier, and sometimes location, operating mode, or contextual metadata. A useful abstraction is

\[

z_t=(x_t,c_t,d,t),

\]

where \(x_t\) is the sensor measurement vector at time \(t\), \(c_t\) represents contextual variables, \(d\) identifies the device, and \(t\) is the timestamp. The target may be a class, a numerical value, an anomaly indicator, or an action.

IoT datasets have several distinctive properties:

• Temporal dependence: adjacent observations are often correlated. A temperature reading at 10:01 is not independent of the reading at 10:00.
• Heterogeneity: devices may use different sampling rates, units, firmware versions, and calibration procedures.
• Missingness: wireless loss, battery depletion, interference, and maintenance produce gaps.
• Noise and outliers: measurements can be affected by electromagnetic interference, sensor faults, environmental disturbances, or incorrect installation.
• Class imbalance: failures and safety incidents are generally much rarer than normal operation.
• Resource constraints: edge devices may have limited memory, computation, energy, and bandwidth.
• Distribution change: seasonal conditions, ageing equipment, user behaviour, and software updates may alter the data distribution.

These characteristics influence every stage of the ML lifecycle. For example, randomly splitting rows into training and test sets can place nearly identical observations from one event in both sets, producing an overly optimistic result.

2. Learning Paradigms for IoT

2.1 Supervised Learning

In supervised learning, a model learns a relationship between inputs and labelled targets. For a dataset of pairs \((x_i,y_i)\), the model \(f_\theta\) is trained by minimizing a loss function:

\[

\hat{\theta}=\arg\min_\theta \frac{1}{n}\sum_{i=1}^{n}L(f_\theta(x_i),y_i)+\lambda\Omega(\theta).

\]

Here, \(n\) is the number of training examples, \(L\) measures prediction error, \(\Omega(\theta)\) is a regularization term, and \(\lambda\geq0\) controls its strength. The equation describes empirical risk minimization with regularization; it assumes that examples and labels are sufficiently defined for the chosen loss, although they need not be independent in all practical settings.

Typical IoT tasks include:

• Classification: identifying normal versus faulty operation, recognizing activity, or classifying machine states.
• Regression: predicting energy demand, remaining useful life, temperature, or air quality.
• Forecasting: predicting future sensor values or system loads.
• Sequence labelling: identifying the state associated with each time interval.

Suitable models include logistic regression, decision trees, random forests, gradient-boosted trees, support vector machines, convolutional neural networks, recurrent neural networks, and transformer-based models. The best choice depends on sample size, interpretability requirements, temporal structure, latency, and available hardware.

2.2 Unsupervised Learning

Unsupervised learning operates without target labels or uses only limited labels. It is particularly valuable because labelled failure data are expensive and rare. Common applications include:

• clustering devices or operating regimes;
• detecting unusual behaviour;
• compressing high-dimensional sensor streams;
• discovering latent structure;
• identifying sensor groups with similar behaviour.

Clustering methods such as k-means require a selected number of clusters and are sensitive to scaling and outliers. Density-based methods can identify irregularly shaped groups and isolated observations but may be sensitive to neighbourhood parameters. Principal component analysis (PCA) projects centred data into directions of maximum variance. If \(X\) is a centred data matrix, PCA obtains eigenvectors of its covariance matrix; the selected components summarize variation, but high variance is not necessarily equivalent to high diagnostic importance.

Anomaly detection may use isolation forests, one-class support vector machines, statistical control limits, autoencoders, or forecasting residuals. An anomaly score is not automatically a fault probability. It indicates unusualness relative to a reference distribution and must be interpreted using operational context.

2.3 Reinforcement Learning

Reinforcement learning (RL) addresses sequential decision-making. An agent observes a state \(s_t\), selects an action \(a_t\), receives a reward \(r_{t+1}\), and transitions to a new state \(s_{t+1}\). The objective is to maximize expected discounted return:

\[

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

\]

The discount factor \(\gamma\) expresses the relative importance of future rewards. RL is applicable to HVAC control, energy storage, traffic signal coordination, adaptive sampling, and industrial scheduling. It is not appropriate to deploy an exploratory policy directly in a safety-critical physical system without constraints and testing. Exploration can cause equipment damage, unsafe conditions, or unacceptable service quality.

IoT RL systems often use a simulator, a digital twin, offline historical data, or a constrained policy. Important design questions include whether the state is sufficiently observable, whether the reward represents all relevant costs, and whether the environment changes over time. Reward functions should include safety, energy, comfort, and maintenance considerations rather than rewarding only a convenient short-term objective.

3. From Sensor Streams to Learning Features

Raw sensor values are rarely the most effective representation. Feature engineering converts observations into variables that expose physically and statistically meaningful patterns.

3.1 Windowing and Temporal Features

A continuous stream is commonly divided into windows of length \(W\), possibly with overlap. For a signal \(x_1,\ldots,x_W\), common features include:

• mean: \(\bar{x}=\frac{1}{W}\sum_{t=1}^{W}x_t\);
• variance: \(s^2=\frac{1}{W-1}\sum_{t=1}^{W}(x_t-\bar{x})^2\), when \(W>1\);
• minimum, maximum, range, median, and quantiles;
• root mean square: \(\sqrt{\frac{1}{W}\sum_{t=1}^{W}x_t^2}\);
• slope from a fitted time trend;
• lag differences, rolling changes, and autocorrelation;
• frequency-domain energy or dominant frequencies;
• cross-sensor ratios, correlations, and physical residuals.

Window length determines the time scale captured by the model. A short window may miss slow degradation, whereas a long window may dilute a brief fault and increase detection latency. Window boundaries must respect the prediction point: features must be calculated only from information available at the time the prediction would have been made.

3.2 Domain and Physics-Based Features

Engineering knowledge can produce features that are more robust and interpretable than arbitrary transformations. Examples include pressure differences across a filter, power factor, vibration-band energy, temperature gradients, and deviation from a physical balance equation. Physics-based features can also support hybrid models that combine mechanistic relationships with ML residual correction.

3.3 Scaling, Encoding, and Selection

Distance-based and gradient-based models may be sensitive to feature scale. Standardization transforms a feature using training-set mean and standard deviation:

\[

z=\frac{x-\mu_{train}}{\sigma_{train}},

\]

where \(\sigma_{train}>0\). The parameters must be estimated from training data only and then applied unchanged to validation, test, and production data. Tree-based models generally require less scaling, but scaling may still simplify pipelines.

Categorical variables can be one-hot encoded or represented with carefully regularized embeddings. Feature selection may remove redundant, unstable, or unavailable variables. Selection must occur inside the training process or cross-validation procedure; selecting features using the complete dataset can leak information from the test set.

4. Model Selection for IoT Applications

Model selection should consider more than accuracy. Relevant criteria include:

• predictive performance and calibration;
• inference latency and memory footprint;
• energy consumption and communication cost;
• interpretability and auditability;
• robustness to missing and noisy values;
• retraining and update requirements;
• privacy and security exposure;
• ease of deployment at the edge or in the cloud.

A simple baseline is essential. For forecasting, a persistence model may predict the next value as the latest observed value. For classification, a majority-class model or rule-based alarm provides a reference. A sophisticated model should be adopted only if it offers meaningful improvement over an appropriate baseline under realistic evaluation.

A useful deployment architecture is hybrid. Initial filtering and feature extraction may occur on the device or gateway, while heavier inference and model management occur in the cloud. Edge inference reduces latency and bandwidth use, but introduces challenges involving model compression, hardware variation, and secure updates.

5. Validation Under Temporal Dependence

5.1 Why Random Splits Can Fail

Randomly assigning rows to training and test sets assumes that the split preserves the intended deployment situation. In IoT, it may place future information in training and past information in testing. Overlapping windows can be especially problematic: two windows from the same physical event may share most observations.

A time-ordered split trains on earlier periods and evaluates on later periods. For example, training may use months 1–6, validation months 7–8, and testing months 9–10. A rolling-origin evaluation repeatedly trains on an expanding or moving historical window and tests on the next period. If the operational system is retrained periodically, this evaluation better represents deployment.

Group-based splitting is needed when observations from the same device, person, machine, or site are strongly related. A model intended for new devices should be tested on devices excluded from training. In some applications both time and group separation are required.

5.2 Leakage Controls

Common leakage sources include:

• using future sensor readings in a feature;
• calculating normalization parameters using all data;
• allowing post-failure maintenance records into pre-failure prediction;
• randomly splitting overlapping windows;
• using a label that would only become available after the prediction deadline;
• selecting a threshold on the test set.

A reproducible pipeline should record the prediction time, information availability time, feature-generation code, split logic, preprocessing parameters, and model version.

6. Evaluation and Decision Thresholds

For binary classification, define true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN). Accuracy is

\[

\frac{TP+TN}{TP+TN+FP+FN},

\]

but it can be misleading under severe imbalance. Precision and recall are

\[

Precision=\frac{TP}{TP+FP},\qquad Recall=\frac{TP}{TP+FN},

\]

provided the relevant denominators are nonzero. The F1 score is the harmonic mean of precision and recall:

\[

F1=\frac{2(Precision)(Recall)}{Precision+Recall},

\]

when precision plus recall is nonzero. Precision answers “How many alerts were correct?” while recall answers “How many actual events were detected?”

For rare events, precision-recall curves can be more informative than receiver operating characteristic curves. However, threshold selection must reflect operational costs. If a missed fire is much more costly than an unnecessary inspection, a lower threshold may be justified. If interventions are expensive, precision may receive greater emphasis. Evaluation should include alert rate, detection delay, false alarms per device-day, and performance by device, site, season, and operating mode.

For regression, mean absolute error is

\[

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

\]

and root mean squared error is

\[

RMSE=\sqrt{\frac{1}{n}\sum_{i=1}^{n}(y_i-\hat{y}_i)^2}.

\]

MAE is expressed in the target’s units and treats errors linearly. RMSE gives greater influence to large errors. Neither is universally superior. For energy forecasting, percentage metrics may be unstable when actual values approach zero; a scaled or absolute metric may be preferable.

A well-calibrated probabilistic classifier should produce probabilities that correspond approximately to observed frequencies. Calibration can be examined with reliability diagrams and quantitative calibration measures. Calibration is important when probabilities drive risk-based decisions rather than simple rankings.

7. Noisy Labels, Missing Data, and Imbalance

7.1 Noisy Labels

IoT labels may come from maintenance logs, operator reports, threshold rules, or delayed diagnoses. A recorded fault time may be approximate, and “normal” data may contain unobserved faults. Label noise can cause a model to learn reporting practices rather than physical conditions.

Good practice includes defining the labelling protocol, recording label confidence and time uncertainty, reviewing borderline cases, and separating confirmed events from weak labels. Robust losses, sample weighting, label smoothing, or semi-supervised learning may help, but they do not replace improved data collection. Evaluation labels should ideally be independently verified.

7.2 Missing Data

Missingness can be completely random, conditionally related to observed variables, or associated with the unobserved value itself. A disconnected device, for example, may be more likely to disappear during a fault. Imputation should therefore be treated as a modelling decision, not merely a preprocessing convenience.

Options include forward filling for suitable slowly changing variables, interpolation for short gaps, model-based imputation, explicit missingness indicators, and models that natively handle missing values. Imputation must not use future data when making real-time predictions. Long gaps should often trigger a data-quality state rather than be hidden by aggressive imputation.

7.3 Class Imbalance

Approaches to rare events include class-weighted losses, targeted sampling, anomaly detection, threshold adjustment, and collecting more positive examples. Oversampling must be performed within the training portion only. Synthetic examples should respect physical constraints; blindly interpolating sensor vectors may create impossible states.

The correct response depends on the goal. If every rare failure must be detected, recall and detection delay may dominate. If maintenance capacity is limited, the number and precision of alerts may be more important. Metrics and thresholds should be selected with domain stakeholders.

8. Concept Drift and Uncertainty

8.1 Concept Drift

Concept drift occurs when the relationship between inputs and targets changes over time. Sensor ageing may shift measurements; a new production process may change normal behaviour; climate or occupancy may alter energy patterns; and a firmware update may change sampling characteristics.

Drift may affect the input distribution \(P(X)\), the target distribution \(P(Y)\), or the conditional relationship \(P(Y|X)\). Monitoring should track feature distributions, missingness, calibration, residuals, alert rates, and confirmed outcomes. Drift detection is not itself proof that the model is invalid; it is a signal for investigation.

Responses include recalibration, threshold adjustment, retraining with recent data, adaptive or online learning, device-specific models, and maintaining a stable reference model for comparison. Automatic retraining requires safeguards against incorporating corrupted or adversarial data.

8.2 Uncertainty

Prediction uncertainty is essential in safety- and maintenance-related IoT systems. Aleatoric uncertainty reflects inherent randomness or measurement noise. Epistemic uncertainty reflects limited knowledge, such as unfamiliar operating conditions or sparse training data.

Methods for representing uncertainty include prediction intervals, quantile regression, ensembles, Bayesian models, conformal prediction, and Monte Carlo dropout. A prediction interval should be interpreted according to its construction and coverage assumptions; it is not a guarantee under arbitrary distribution shift. Models should be able to abstain or defer to a human when uncertainty is high. An uncertainty-aware system may issue “inspect,” “continue monitoring,” or “safe to automate” decisions rather than forcing a brittle binary answer.

9. Worked Example: Predictive Maintenance for an Industrial Pump

Consider a pump fitted with vibration, motor-current, pressure, temperature, and flow sensors. The maintenance team wants at least 90% recall for bearing failures while keeping false alarms below a manageable level.

Problem definition. The prediction deadline is set to 24 hours before confirmed failure. A positive example is a time window whose future contains a verified bearing failure within that horizon. Windows after maintenance are excluded until the pump returns to stable operation.

Feature construction. For each ten-minute window, the system calculates vibration RMS, vibration-band energy, temperature slope, pressure difference, current variability, flow deviation, and missingness indicators. Features use only data available at the window end. Training-set means and standard deviations are stored for later transformation.

Data splitting. Data from the first six months are used for training, the next month for threshold selection, and the final month for testing. Windows from the same failure event are kept in one partition. A second experiment holds out one pump to evaluate generalization to a new asset.

Models. A persistence and rule-based alarm provide baselines. Logistic regression provides an interpretable benchmark; a random forest captures nonlinear interactions; and a gradient-boosted model is tested for predictive improvement. A complex sequence model is considered only if it improves event-level performance enough to justify its operational cost.

Evaluation. The team reports recall, precision, false alarms per pump-month, detection delay, and calibration. If the model produces several alerts during the same failure episode, they are consolidated into one event for operational reporting. The threshold is selected on validation data to meet the recall target as closely as possible without excessive false alarms. The final test set is used once for unbiased performance estimation.

Deployment. A gateway computes features and sends a compact feature vector to the cloud. The cloud returns a risk score and uncertainty estimate. High-risk, high-confidence cases create work orders; uncertain cases request additional inspection data. The system logs model version, features, score, threshold, operator response, and eventual maintenance outcome.

Outcome and learning points. In practice, the most valuable improvement may come from correcting failure timestamps and separating maintenance-related sensor patterns from genuine degradation. The example illustrates that an ML model is only one component of an operational decision system. Success requires a valid target, leakage-free evaluation, event-level metrics, integration with maintenance workflows, and continuous monitoring.

10. Responsible and Secure IoT Machine Learning

IoT ML can affect safety, employment, insurance, access to services, and privacy. Sensor data may reveal household routines, worker activity, location, health conditions, or industrial secrets. Data collection should follow purpose limitation, data minimization, access control, retention rules, and applicable privacy law.

Security threats include adversarial sensor manipulation, model extraction, poisoned training data, compromised gateways, and unauthorized model updates. Secure boot, signed models, encrypted communication, device identity management, least-privilege access, and audit logs are important controls. Federated learning can reduce centralization of raw data, but it does not automatically guarantee privacy or eliminate poisoning and communication risks.

Models should be evaluated for unequal error rates across relevant devices, sites, user groups, or operating conditions. Explainability is especially important when operators must justify an intervention. Explanations should be faithful to the model and presented in terms meaningful to the responsible professional, such as “increased vibration-band energy and temperature slope,” rather than unsupported causal claims.

11. Contemporary Directions and Best Practices

Current IoT ML development increasingly combines edge AI, TinyML, self-supervised learning, digital twins, multimodal foundation models, federated learning, and continual learning. Self-supervised methods can learn representations from large volumes of unlabelled sensor data, after which smaller labelled datasets are used for task-specific adaptation. Digital twins can provide simulated data and safe testing environments, but simulation-to-reality differences must be measured.

Recommended practice is to:

1. Define the operational decision, prediction horizon, and cost of errors before choosing a model.

2. Establish strong, simple baselines.

3. Audit timestamps, labels, units, calibration, missingness, and device identity.

4. Build leakage-resistant temporal and group-aware splits.

5. Engineer features using domain knowledge and information available at prediction time.

6. Evaluate event-level performance, uncertainty, calibration, latency, and resource use.

7. Test across devices, sites, seasons, and operating modes.

8. Monitor drift, data quality, alerts, and real-world outcomes after deployment.

9. Provide rollback, human escalation, and safe-failure mechanisms.

10. Document data provenance, model limitations, versioning, and governance responsibilities.

Chapter Summary

Machine learning extends IoT from passive monitoring to prediction, detection, decision support, and control. Supervised learning uses labelled examples for classification, regression, and forecasting; unsupervised learning discovers structure and unusual behaviour when labels are limited; and reinforcement learning addresses sequential control but requires careful safety and reward design. IoT feature engineering commonly uses windows, statistical summaries, trends, frequency information, cross-sensor relationships, and physics-based variables.

IoT models must be evaluated under realistic temporal and device conditions. Random row-level splits can cause leakage, while rolling-origin and group-aware evaluation better represent deployment. Accuracy alone is inadequate for rare events; precision, recall, F1, calibration, detection delay, false-alarm rates, and operational costs should be considered. Noisy labels, missing data, imbalance, concept drift, and uncertainty are central engineering concerns rather than exceptional cases.

A successful IoT ML system includes data governance, secure deployment, monitoring, human oversight, and procedures for retraining or rollback. The central professional principle is to optimize the complete decision process—not merely the score of an isolated model.

Key Terms

Supervised learning: Learning from input examples paired with target labels.

Unsupervised learning: Learning structure from data without task labels.

Reinforcement learning: Learning sequential actions from states, rewards, and transitions.

Feature engineering: Transforming raw data into informative model inputs.

Windowing: Dividing a continuous stream into finite intervals for analysis.

Temporal leakage: Using information unavailable at the time of prediction.

Concept drift: Change over time in data distributions or input-target relationships.

Class imbalance: Unequal prevalence of target classes.

Precision: Proportion of predicted positives that are true positives.

Recall: Proportion of actual positives that are detected.

Calibration: Agreement between predicted probabilities and observed frequencies.

Anomaly detection: Identifying observations that differ substantially from a reference pattern.

Aleatoric uncertainty: Uncertainty arising from inherent randomness or measurement noise.

Epistemic uncertainty: Uncertainty caused by limited knowledge or data.

Edge inference: Running model computation near the data source.

Data drift: Change in the distribution of model inputs.

Label noise: Errors or uncertainty in target annotations.

Rolling-origin evaluation: Repeated time-ordered training and future testing.

Digital twin: A computational representation of a physical system used for monitoring or experimentation.

Model governance: Policies and controls for responsible model development, deployment, and monitoring.

Review Questions

Short Answer Questions

1. Why can random train-test splitting be misleading for IoT time-series data?

2. Distinguish precision from recall in an IoT fault-detection system.

3. What is concept drift, and give one IoT example?

4. Why should preprocessing parameters be estimated using training data only?

5. State one advantage and one risk of edge inference.

Long Answer Questions

1. Compare supervised, unsupervised, and reinforcement learning using appropriate IoT applications.

2. Explain how you would engineer features from a vibration sensor for predictive maintenance.

3. Describe a leakage-resistant validation strategy for a multi-device IoT dataset.

4. Discuss methods for addressing noisy labels, missing data, and severe class imbalance.

5. Explain why uncertainty and calibration matter when ML outputs drive maintenance or safety decisions.

Analytical / Application Questions

1. An alarm model detects 18 of 20 genuine failures but generates 42 false alarms. Calculate its recall and precision, and discuss whether it should be deployed without changing its threshold.

2. Design an evaluation plan for an energy-forecasting model deployed across 50 buildings with seasonal demand and occasional sensor outages.

3. A deployed anomaly detector shows a steadily increasing alert rate after a firmware update. Propose a diagnostic and remediation process that distinguishes genuine equipment deterioration from data or concept drift.

Multiple Choice Questions with Answers

Question 1

Which learning paradigm is most directly suited to predicting whether an IoT machine will fail within 24 hours when verified failure labels are available?

A. Supervised learning

B. Unsupervised learning

C. Reinforcement learning

D. Association-rule mining

Correct Answer: A. Supervised learning

Question 2

What is the main reason for using time-ordered splits in IoT model evaluation?

A. They guarantee a larger training set

B. They prevent future information from entering training

C. They eliminate all sensor noise

D. They make every class equally frequent

Correct Answer: B. They prevent future information from entering training

Question 3

A classifier detects 15 of 20 actual failures. What is its recall?

A. 25%

B. 50%

C. 75%

D. 95%

Correct Answer: C. 75%

Question 4

Which feature is most directly a frequency-domain feature for vibration analysis?

A. Median temperature

B. Number of missing readings

C. Dominant vibration frequency

D. Device identifier

Correct Answer: C. Dominant vibration frequency

Question 5

Why can accuracy be misleading for rare fault detection?

A. Accuracy cannot be calculated for classifiers

B. A majority-class model may appear strong while missing most faults

C. Accuracy always penalizes false positives more than false negatives

D. Accuracy is measured only in physical units

Correct Answer: B. A majority-class model may appear strong while missing most faults

Question 6

Which situation is an example of temporal leakage?

A. Using a training-set mean to standardize test data

B. Using a future maintenance record to construct a past prediction feature

C. Using a vibration RMS feature calculated from past readings

D. Evaluating a model on a later month

Correct Answer: B. Using a future maintenance record to construct a past prediction feature

Question 7

What does concept drift describe?

A. A permanent reduction in model memory usage

B. A change in the data or input-target relationship over time

C. A method for converting categories into numbers

D. Randomly deleting sensor observations

Correct Answer: B. A change in the data or input-target relationship over time

Question 8

Which metric is defined as TP divided by TP plus FP?

A. Recall

B. Specificity

C. Precision

D. Accuracy

Correct Answer: C. Precision

Question 9

Which action is most appropriate when a model is highly uncertain about an unfamiliar operating condition?

A. Force a high-risk decision regardless of uncertainty

B. Ignore the prediction and delete the observation

C. Use a defined abstention or human-escalation procedure

D. Treat the uncertainty as proof of failure

Correct Answer: C. Use a defined abstention or human-escalation procedure

Question 10

In reinforcement learning for HVAC control, what does the reward function represent?

A. The list of sensor identifiers

B. The objective signal balancing outcomes such as comfort and energy use

C. The normalization parameter for temperature

D. The number of training labels

Correct Answer: B. The objective signal balancing outcomes such as comfort and energy use