PaperlayerView exact revision on arXiv

Source-linked AI summary

Wi-Fi Meets ML: A Survey on Improving IEEE 802.11 Performance with Machine Learning

Szymon Szott, Katarzyna Kosek-Szott, Piotr Gawłowicz, Jorge Torres Gómez, Boris Bellalta, Anatolij Zubow, Falko Dressler

arXiv:2109.04786v5cs.NI

TL;DR

Next-generation Wi-Fi is difficult to configure because many interacting parameters must support dense deployments and demanding applications. This survey systematically synthesizes ML research across Wi-Fi performance areas, reviewing over 250 papers. It finds frequent use of supervised and reinforcement learning, with deep learning increasingly common, and identifies open challenges for future work.

  • Problem

    Growing Wi-Fi complexity, densification, distributed management, and coexistence create performance-optimization challenges, while real-world validation and distributed operation remain limited.

  • Method

    The paper uses a structured survey of ML applications across core features, newer mechanisms, management, coexistence, and multi-hop Wi-Fi.

  • Results

    Supervised and reinforcement learning are used most often; Q-learning, multi-armed bandits, and neural networks dominate, while about half of recent papers implement deep learning.

  • Takeaways & Limitations

    The survey identifies open challenges across Wi-Fi performance areas and provides future research directions for applying ML to Wi-Fi.

Abstract

from arXiv · show

Wireless local area networks (WLANs) empowered by IEEE 802.11 (Wi-Fi) hold a dominant position in providing Internet access thanks to their freedom of deployment and configuration as well as the existence of affordable and highly interoperable devices. The Wi-Fi community is currently deploying Wi-Fi 6 and developing Wi-Fi 7, which will bring higher data rates, better multi-user and multi-AP support, and, most importantly, improved configuration flexibility. These technical innovations, including the plethora of configuration parameters, are making next-generation WLANs exceedingly complex as the dependencies between parameters and their joint optimization usually have a non-linear impact on network performance. The complexity is further increased in the case of dense deployments and coexistence in shared bands. While classical optimization approaches fail in such conditions, machine learning (ML) is able to handle complexity. Much research has been published on using ML to improve Wi-Fi performance and solutions are slowly being adopted in existing deployments. In this survey, we adopt a structured approach to describe the various Wi-Fi areas where ML is applied. To this end, we analyze over 250 papers in the field, providing readers with an overview of the main trends. Based on this review, we identify specific open challenges and provide general future research directions.

I. INTRODUCTION

Wi-Fi is widely deployed because it is flexible, inexpensive, and interoperable, but successive IEEE 802.11 innovations are making configuration increasingly complex. Dense deployments and demanding applications further increase the challenge of timely, effective network tuning.

  • Wi-Fi holds a dominant position in wireless Internet access because of flexible deployment, unlicensed-band operation, and affordable, interoperable devices.
  • Wi-Fi 6 is state of the art in new consumer products, while Wi-Fi 7 is under development.
  • Successive 802.11 mechanisms introduce many parameters whose configuration dependencies make next-generation Wi-Fi increasingly complex.
  • Future Wi-Fi must support ultra-low-latency and ultra-high-reliability traffic, making timely transmission-setting updates important.
  • Traditional algorithms struggle to find adequate configurations in the enormous search space within acceptable time and computational resources.

A. Need for ML in Wi-Fi

Wi-Fi complexity, densification, and distributed deployment create performance-management problems that motivate machine-learning approaches. Existing surveys do not provide a dedicated, detailed account of ML for improving IEEE 802.11 performance.

  • Increasing Wi-Fi complexity, uncoordinated deployment, distributed management, and densification may negatively affect future 802.11 networks.
  • Machine learning is presented as a candidate approach because it can learn from training data without explicit programming.
  • Existing general wireless-networking and 5G surveys do not detail Wi-Fi performance improvement with ML from multiple perspectives.
  • The paper argues that Wi-Fi’s complexity warrants a survey dedicated specifically to ML-based performance improvement.

B. Methodology

The survey began with a systematic literature review across five bibliographic databases, then expanded the candidate set through manual additions and citation analysis.

  • The initial search combined Wi-Fi, 802.11, WLAN, and machine-learning terms in abstracts across IEEE Xplore, ACM, Elsevier, Wiley, and MDPI.
  • The database search yielded 1189 papers before out-of-scope papers were removed.
  • Additional papers were added manually, usually through cross-citation analysis.
  • SpringerLink was not searchable at this stage because its interface did not support searching within published-paper abstracts, so papers from it were added manually.

C. Survey Scope and Contributions

The survey organizes ML applications across core Wi-Fi features, newer mechanisms, management, shared-band coexistence, and multi-hop settings. It reviews more than 250 papers, identifies trends and open challenges, and outlines future research directions.

  • The survey structure covers core Wi-Fi features, newer features, management issues, shared-band coexistence, and multi-hop Wi-Fi.
  • The authors use a structured approach to describe the Wi-Fi performance areas where ML is applied.
  • The review analyzes over 250 papers to summarize existing work and major trends in applying ML to Wi-Fi performance problems.
  • Open challenges are identified for every Wi-Fi performance area covered in Sections III–VII.
  • The survey provides general future research directions intended to guide new work in this developing area.

II. RELATED SURVEYS

Related surveys cover ML across diverse wireless-network types and applications, but Wi-Fi performance improvement receives only partial or superficial treatment. This survey organizes that scattered literature around Wi-Fi-specific performance areas.

  • Scope of related surveys: Existing surveys address ML for Wi-Fi, wireless sensor, cognitive-radio, mesh, heterogeneous, 5G, and 6G networks.They span network operation, applications, resource management, spectrum sharing, and related wireless functions.
  • Survey organization: The survey’s Figure 2 classifies research areas where ML improves Wi-Fi network performance and indicates the ML methods reported for each area.The survey structure covers core Wi-Fi features and multiple performance-related functions.
  • Scope of related surveys: Wi-Fi-related survey coverage often emphasizes analytics, quality indicators, applications, security, or coexistence rather than detailed performance improvement.The cited literature includes monitoring, user satisfaction, localization, activity detection, intrusion detection, and shared-spectrum concerns.

A. Wi-Fi-related Surveys

Wi-Fi-related surveys cover analytics, quality assessment, Wi-Fi applications, and security. They examine how ML extracts knowledge from network data and improves QoS, QoE, localization, activity detection, and intrusion detection.

  • Wi-Fi analytics surveys examine ML models that extract knowledge from big data streams generated by large-scale wireless networks.
  • Other surveys use ML to estimate QoS, QoE, and their cross-correlation.
  • Indoor localization and human activity detection are the two main Wi-Fi-based application areas covered by these surveys.
  • Wi-Fi security surveys address attacks exploiting vulnerabilities and ML-based intrusion detection systems that improve IDS autonomy and accuracy.

B. Wireless Communications-related Surveys

Wireless-communications surveys cover ML fundamentals, protocol-layer applications, network-specific problems, and transfer or federated learning. However, Wi-Fi-specific mechanisms and their integration with ML remain only briefly or superficially covered in much of this literature.

  • A broad survey reviews supervised, unsupervised, and reinforcement learning across wireless applications, but mentions IEEE 802.11 only briefly.
  • Other surveys report ML applications across PHY, MAC, and network layers, alongside MEC, SDN, NFV, 5G, LPWAN, and related network types.
  • Wireless-network surveys also examine MAC-layer resource management, network-layer mobility management, application-layer localization, AP power saving, and indoor localization.
  • Surveys of WSNs, CRNs, MANETs, and IoT focus on network-specific problems; only CRN surveys discuss Wi-Fi applications such as coexistence and channel selection.
  • Specific details on integrating ML techniques with Wi-Fi mechanisms are covered only superficially.
  • Transfer-learning and federated-learning surveys address wireless constraints including limited labels, changing conditions, device capacity, and ML requirements.

C. Wi-Fi and 5G-related Surveys

Surveys in 5G and related wireless areas apply ML to infer patterns, predict trends, and optimize parameters across protocol layers. Their applications include network management, performance improvement, signal identification, spectrum sharing, and unlicensed-band use.

  • 5G surveys focus on PHY, MAC, and network-layer tasks including interference identification, link-quality prediction, and traffic-demand estimation.
  • ML solutions analyze and manage mobile networks through state prediction, traffic classification, call-detail mining, and radio-signal analysis.
  • Other applications aim to improve mobile-system and IoT performance.
  • ML is used to identify wireless modulations and technologies.
  • ML supports fair and efficient spectrum sharing in 5G and future 6G networks.
  • ML is also used to maximize the potential of unlicensed bands for Industry 4.0 applications.

D. Summary

The survey maps existing work on machine learning for Wi-Fi-related topics while focusing on performance improvement in the 802.11 protocol family. It also defines exclusions and summarizes core-feature studies by evaluation method and comparison baseline.

  • Existing surveys: Existing surveys cover Wi-Fi-related topics and ML models, but provide few details on using ML to improve 802.11 performance.Application-oriented topics include human activity detection, indoor localization, and network security.
  • Scope: The survey excludes dedicated Wi-Fi applications, energy efficiency, and network security from its non-performance scope.Examples include device positioning, power-saving protocols, and detecting selfishly configured devices.
  • Scope: The survey also does not explain how individual ML methods operate, directing readers to dedicated literature for those methodological details.Its focus is on ML applications for Wi-Fi performance rather than general ML-method descriptions.
  • Core Wi-Fi features: Table II summarizes studies that improve core Wi-Fi features with ML using theoretical, simulation, or experimental evaluation.The table defines the evaluation abbreviations as theoretical (E), simulation (S), and experimental (E).
  • Core Wi-Fi features: Reported ML improvements are compared with state-of-the-art methods or, when unspecified, with IEEE 802.11.This comparison rule provides the reference point for interpreting the summarized studies.

III. CORE WI-FI FEATURES

ML is applied across core Wi-Fi mechanisms to adapt channel access, link configuration, and PHY-layer parameters under changing conditions and competing objectives. Reported studies improve throughput, fairness, latency, scalability, or rate selection, although broader network-level evaluation remains needed.

  • Channel Access: Channel-access studies use supervised and reinforcement learning to tune contention-window values for collision reduction, utilization, throughput, fairness, and traffic prioritization.Objectives include reduced collisions, improved utilization, throughput, network utility, lower energy, and fairer access.
  • Channel Access: 153.9% throughput, 64% latency, and 19.34% fairness improvements were reported for random-forest CW balancing versus 802.11ac.The method monitors busy time, user occupancy, and sent frames in indoor 802.11ac scenarios with up to eight stations.
  • Channel Access: 200% higher throughput and 33% lower end-to-end delay were reported versus DCF in a heavily loaded scenario with 100 users.The weighting mechanism balances aggressive small-CW users against non-aggressive large-CW users.
  • Channel Access: DQN and DDPG maintained stable throughput as station counts increased, whereas 802.11ax throughput decreased by up to 28%.The centralized DRL solution was evaluated with one AP and up to 50 stations using collision-probability history and throughput rewards.
  • Channel Access: PDS learning converged faster than Q-learning by exploiting known system transitions and improved throughput under moderate network load.The method uses prior knowledge of contention windows and transmission-buffer occupancy to eliminate some exploration actions.
  • Link Configuration and PHY Trade-offs: ML-based PHY and link configuration methods address rate-selection trade-offs involving channel conditions, bandwidth, spatial streams, frame aggregation, and guard intervals.Reported examples include at least 27% goodput improvement, slight throughput gains from online SGI learning, and up to 358% higher throughput from MAB-based configuration.

C. Frame Aggregation

Machine learning is applied to frame aggregation and related Wi-Fi configuration problems to adapt transmission settings under changing channel and collision conditions. Reported studies use supervised and reinforcement learning across throughput, goodput, energy, and aggregation-level objectives.

  • C. Frame Aggregation: 55% highest goodput improvement is reported for a random forest regressor trained with real measurements from a Wi-Fi scenario containing up to 10 stations.
  • C. Frame Aggregation: At least 13% throughput improvement over baseline algorithms is reported when an ANN predicts MCS levels and estimates the best aggregation level.The ANN is trained from packets received across all available rates within a 1 s window.
  • C. Frame Aggregation: Close to 100% accuracy is reported for logistic regression estimating the aggregation level from hardware-level timestamps on non-rooted client hardware.
  • C. Frame Aggregation: Jointly optimizing frame size and contention window improves throughput over optimizing frame size alone in simulations with 10 mobile users.An ANN provides a gradient indicating the direction of optimal frame and contention-window sizes.

D. Channel Bonding

Channel bonding is difficult because wider channels can increase contention in dense deployments and the best width, channel set, and allocation depend on scenario-specific conditions. The surveyed ML approaches learn channel-bonding actions, predict performance, or coordinate related multi-band and multi-link decisions.

  • Channel-bonding motivation: Wider channels increase transmission rates but can notably increase contention between neighboring networks in dense scenarios.IEEE 802.11 channel widths progress from 40 MHz in 802.11n to 80, 160, and 320 MHz in later amendments.
  • Channel-bonding motivation: The appropriate channel width and channel selection depend on contending devices, BSS loads, and available channels rather than one universal configuration.
  • Learning-based bonding: MAB-based channel-bonding methods learn channel widths and bonding strategies from experience, including station needs and access categories.UCB-based dynamic bonding produced gains higher than 100% in some testbed cases, while ITE improved static and dynamic bandwidth channel access.
  • Learning-based bonding: Lightweight stateless MABs are reported as efficient for rapid adaptation, avoiding broad or meaningless state definitions in realistic channel-bonding scenarios.
  • Learning-based bonding: DRL-based on-demand channel bonding uses multi-agent training to allocate channels according to expected load and performance, minimizing latency.The strategy outperformed rules based on aggregating all channels or randomly selecting one or two channels.
  • Related advanced mechanisms: ML also supports multi-band traffic balancing, multi-link operation, distributed MU-MIMO, and full-duplex user pairing by predicting channel conditions or optimizing joint decisions.Examples include occupancy-based interface selection, DRL for channel allocation and AP clustering, and neural pairing that handles varying input lengths without retraining within an expected range.

F. Open Challenges

The survey identifies open challenges across advanced Wi-Fi mechanisms and connectivity management, especially the gap between promising ML studies and realistic, evolving deployments. Future work must address practical deployment, mixed traffic and QoS, coordination assumptions, and emerging multi-link operation.

  • General open challenges: Many advanced Wi-Fi features remain only partially studied because they are recent, complex, and still under development.The survey identifies room for new ML techniques and for addressing uncovered aspects.
  • Beamforming: Robust COTS solutions are still required for low-latency beam-sector alignment, while outdoor ML resource coordination needs updating for IEEE 802.11ay.
  • Multi-user communication: Multi-user ML research should address spatial-stream and RU allocation under realistic traffic patterns and QoS requirements.Future traffic, contention, and environmental predictions may support resource pre-reservation for sensitive traffic and worst-case latency.
  • Spatial reuse: Spatial-reuse methods developed for decentralized BSSs must be extended for IEEE 802.11be TXOP sharing and cooperative schemes.
  • Channel bonding: Channel-bonding results require further testing and comparison, including combinations with OFDMA RU allocation.The survey suggests complex DRL techniques may be suitable for this coupled problem.
  • Multi-link operation: Multi-link operation creates open problems in channel switching, link selection, occupancy-aware traffic handling, and flow distribution.
  • Connectivity management: Connectivity management spans channel, band, and AP selection, while traffic and link-health prediction can prepare configuration changes and help minimize outage probability.
  • Channel allocation: Dense Wi-Fi channel allocation must account for interference relationships and changing traffic loads, with ML methods using centralized, graph-based, and monitoring-based approaches.Reported examples include DRL with contention graphs and passive monitoring using LASSO and OLS.

B. Management Architectures

ML-based management architectures address distributed Wi-Fi configuration by learning network conditions and adapting control decisions across access points, traffic, coexistence, and connection health. The surveyed approaches report improvements in performance, but coexistence research remains limited in real-world validation and joint adaptation of both technologies.

  • Management architectures: AI-based control planes simplify distributed Wi-Fi management by coordinating configuration decisions across multiple network devices.They are designed to handle the large parameter space involved in managing Wi-Fi networks.
  • Applications: ML-based Wi-Fi management has been applied to traffic prediction, congestion classification, anomaly detection, QoS estimation, and AP or channel configuration.Reported methods include SVMs, clustering, kNN, decision trees, and neural models.
  • Network coexistence: In coexistence settings, ML modifies LTE-U duty-cycle or subframe behavior and can schedule Wi-Fi transmissions during detected LTE-U OFF periods.WiPlus uses k-means clustering and spectral scans to estimate LTE-U timing and available Wi-Fi airtime.
  • Open challenges: The survey identifies simulation-heavy evaluation and insufficient joint adaptation of Wi-Fi and cellular nodes as open coexistence challenges.Real-testbed validation is needed because simulations may omit important radio effects.

A. Tool Chains

Research on ML-enhanced Wi-Fi relies mainly on simulation and software frameworks that connect network environments to learning agents. Existing tools support ns-3, GNU Radio, SDR testbeds, and next-generation Wi-Fi simulation, but seamless general integration remains incomplete.

  • Evaluation practice: Approximately 80% of reviewed studies validate ML-based Wi-Fi solutions through network simulation, while only about one quarter use analytical investigations or real testbeds.ns-3 accounts for 10% of simulation use; SDR platforms dominate experimental studies, while COTS Wi-Fi hardware is rarely used.
  • Simulation frameworks: ns3-gym connects ns-3 network models to reinforcement-learning agents through an environment gateway, proxy, and standard agent interface.The framework separates scenario configuration in ns-3 from agent development using standard libraries.
  • Open tooling needs: The survey calls for standardized interfaces and simulator support that incorporate ML execution time into virtual network time.Existing efforts such as ns3-gym and Komondor do not yet provide general-purpose default ML integration.
  • Experimental frameworks: GrGym integrates GNU Radio programs with OpenAI Gym, enabling learning agents to interact with environments that can include real SDR transmissions.Unlike purely simulated environments, it can expose real wireless testbeds to the agent.
  • Next-generation Wi-Fi tools: Komondor provides direct ML support for simulating complex next-generation Wi-Fi environments, including a subset of 802.11ax.It is intended for ML-based studies of advanced Wi-Fi scenarios.

B. Datasets

Wi-Fi ML research depends heavily on locally collected datasets, while open and standardized resources remain scarce. The survey therefore emphasizes shared data-collection procedures, reproducible datasets, and testbeds spanning diverse wireless conditions.

  • Future dataset development: The survey recommends large, high-quality shared datasets and standardized collection procedures across wireless platforms and testbeds.Using multiple platforms may reduce model overfitting and improve learning performance.
  • Dataset availability: Researchers created labelled datasets in 49 papers, whereas only 6 papers used publicly available datasets and only 6 datasets were released.This limits the availability of common resources for training and comparison.
  • Existing resources: Open datasets include Wi-Fi measurements, 802.11ac performance data, packet-level Wi-Fi and cellular traces, and experimental mmWave waveforms.Examples are available through CRAWDAD, IEEE DataPort, individual research pages, and other repositories.
  • Dataset content: Available datasets mostly contain raw measurements or sniffed traffic traces for tasks such as anomaly detection, activity recognition, traffic classification, and human detection.They also support application-oriented Wi-Fi ML beyond protocol optimization.
  • Testbeds: Open-access testbeds such as Orbit, COSMOS, and POWDER provide hardware and software support for experimental evaluation of ML-based Wi-Fi solutions.These platforms can help bridge the gap between simulation and real-network studies.

E. Reference Evaluation Scenarios and Performance Metrics

The survey finds that ML-Wi-Fi results are difficult to compare because studies lack common evaluation scenarios, reproducible tooling, and consistent real-network validation. It therefore calls for standardized, dynamic benchmarks, robust models, transparent operation, and newer distributed learning approaches.

  • Reference scenarios: Most published ML-Wi-Fi studies report performance gains, but the absence of common scenarios prevents direct comparison and firm conclusions about progress.Reference scenarios should range from simple training environments to large, heterogeneous, and dynamic conditions.
  • Reproducibility: Reproducible evaluation requires detailed datasets and shared scenarios that preserve environmental conditions and ML-process responses across researchers.The survey links common scenarios directly to the reproducibility challenge.
  • Simulation tools: Current simulators lack seamless ML integration, including standard interfaces between Wi-Fi components and ML functions and explicit modeling of ML execution time.ns-3 modules and Komondor are identified as partial efforts rather than complete solutions.
  • Real-world validation: Real-network platforms and testbeds are needed to validate ML-enhanced Wi-Fi functionality under conditions simulators cannot accurately reproduce.The survey treats such platforms as crucial before deployment.
  • Robustness and safety: Robust ML solutions should prevent unpredictable behavior even when sacrificing general performance in particular scenarios.The concern is especially important when ML controls an individual feature or the whole Wi-Fi network.
  • Interpretability: Black-box ML can impede adoption because opaque decisions risk improper or unsafe behavior in wireless network control.The survey identifies transparency and human control as important limitations of current models.
  • Emerging ML methods: Federated learning and transfer learning are identified as emerging directions for distributed Wi-Fi learning and reuse of models across related tasks.Transfer learning can reduce development time when learned features generalize to the second task.

X. CONCLUSION

The survey synthesizes over 250 ML-based Wi-Fi solutions across basic and complex networking functions, coexistence, multi-hop settings, and research tools. It finds supervised and reinforcement learning most prevalent, while DL and newer paradigms are gaining attention; future work should address joint optimization and coexistence fairness.

  • Over 250 ML-based solutions are reviewed across Wi-Fi features, coexistence, multi-hop settings, and available research tools and datasets.
  • Observed ML trends: Supervised learning and reinforcement learning are frequently used, whereas unsupervised learning is less popular.Q-learning, multi-armed bandits, and neural networks are among the most common mechanisms.
  • Observed ML trends: About half of the most recent papers implement deep learning, with supervised and reinforcement learning the most common DL techniques.
  • Emerging paradigms: Federated learning and transfer learning are recent additions to Wi-Fi research that may distribute learning tasks and improve training speed.
  • Open directions: Future research should jointly optimize wider ranges of Wi-Fi features and examine coexistence between ML-controlled and legacy networks because channel access may become unfair.
Loading 2109.04786v5…