PaperlayerView exact revision on arXiv

Source-linked AI summary

Artificial Intelligence Enabled Wireless Networking for 5G and Beyond: Recent Advances and Future Challenges

Cheng-Xiang Wang, Marco Di Renzo, Slawomir Stańczak, Sen Wang, Erik G. Larsson

arXiv:2001.08159v1cs.NIeess.SP

TL;DR

B5G networks create unstructured, data-intensive design and operation problems that motivate AI/ML methods. The paper surveys how AI/ML can be applied across channel processing, physical-layer research, network management, algorithms, applications, and standards, and identifies future research challenges. Its principal outcome is a comprehensive synthesis of recent advances and open directions for combining AI/ML with B5G wireless networks.

  • Problem

    B5G network design and optimization must address unstructured, seemingly intractable problems involving large amounts of data.

  • Method

    The paper provides a comprehensive survey of AI/ML applications and developments across B5G channel processing, physical-layer research, network management, algorithms, applications, and standards.

  • Results

    The survey synthesizes recent advances and discusses challenges and potential future research directions for combining AI/ML with B5G wireless networks.

  • Takeaways & Limitations

    AI/ML is presented as a framework for addressing B5G wireless-network problems across design, optimization, operation, and physical-layer processing.

Abstract

from arXiv · show

The fifth generation (5G) wireless communication networks are currently being deployed, and beyond 5G (B5G) networks are expected to be developed over the next decade. Artificial intelligence (AI) technologies and, in particular, machine learning (ML) have the potential to efficiently solve the unstructured and seemingly intractable problems by involving large amounts of data that need to be dealt with in B5G. This article studies how AI and ML can be leveraged for the design and operation of B5G networks. We first provide a comprehensive survey of recent advances and future challenges that result from bringing AI/ML technologies into B5G wireless networks. Our survey touches different aspects of wireless network design and optimization, including channel measurements, modeling, and estimation, physical-layer research, and network management and optimization. Then, ML algorithms and applications to B5G networks are reviewed, followed by an overview of standard developments of applying AI/ML algorithms to B5G networks. We conclude this study by the future challenges on applying AI/ML to B5G networks.

I. INTRODUCTION

5G and B5G networks face growing data, scale, and complexity demands that conventional design and operation may not handle effectively. This survey examines how AI/ML can support B5G network design, optimization, and management across several technical areas.

  • Motivation: AI technologies can address unstructured, seemingly intractable B5G problems involving large amounts of data.The paper frames AI as a way to support the design and optimization of 5G and B5G wireless networks.
  • Motivation: Dynamic channels and interference make channel information difficult to model, while ML can learn unknown channel information from communication data and prior knowledge.This targets channel and interference modeling in B5G scenarios.
  • Motivation: Increasing access-point density creates a need for global resource optimization and fine system tuning, but coupled resources and parameters make existing approaches difficult to apply.The paper identifies deep learning and probabilistic learning as possible ways to model nonlinear correlations and estimate system parameters.
  • Motivation: Learning-based adaptive configuration can identify behavioral patterns, anticipate traffic, and plan ahead instead of only reacting to unexpected scenarios.This is presented as a potential benefit of ML for network operation.
  • Scope and contribution: Most existing AI algorithms are not specifically designed for wireless networks, making direct application to B5G difficult.This limitation motivates the paper’s focus on combining AI technologies with B5G wireless networks.
  • Scope and contribution: The article surveys AI/ML applications across channel measurements, modeling and estimation, physical-layer research, network management and optimization, algorithms, and standards.It also discusses future challenges for applying AI/ML to B5G networks.

II. CHANNEL MEASUREMENTS, MODELING, AND ESTIMATION FOR B5G NETWORKS USING AI TECHNOLOGIES

B5G channel modeling must handle diverse frequency bands, complex propagation, and new scenarios where conventional assumptions can be inaccurate or computationally expensive. AI/ML methods use measurement and environmental data to support data-driven modeling, feature extraction, clustering, estimation, and classification.

  • Channel modeling: B5G’s diverse frequency bands and complex propagation make channel modeling more difficult than in simpler scenarios.The cited passages identify sub-6 GHz, millimeter wave, terahertz, and optical bands as part of this diversity.
  • Channel modeling: Conventional channel models rely on assumptions and approximations, while new scenarios require time-consuming measurements to characterize unfamiliar channel behavior.The paper notes that modeling new scenarios requires conducting channel measurements.
  • AI/ML methods: Channel features can be extracted from measurement data and environmental information, enabling data-driven prediction, CIR modeling, MPC clustering, parameter estimation, and scenario classification.The paper describes combining data-driven and model-based methods to balance accuracy and computational complexity.
  • AI/ML methods: FNN and RBF-NN models were evaluated with real and synthetic channel data and showed good potential for channel modeling.The example includes prediction of path loss and RMS delay spread.
  • AI/ML methods: CNN-based processing can use extracted MPC parameters such as amplitude, delay, and Doppler frequency to classify wireless channels.The CNN output is the wireless-channel class.

B. Channel Estimation Associated with ML

ML-based channel estimation addresses difficult CSI acquisition conditions involving fast fading, nonlinear channels, high mobility, and pilot overhead. Reviewed approaches include SVR, deep learning, sparse Bayesian learning, and a proposed direction toward generalized schemes usable across scenarios.

  • Challenges: Channel estimation is needed for quality of service and efficient information transmission, but fast fading, nonlinear channels, mobility, and pilot overhead complicate CSI acquisition.The pilot-length versus estimation-accuracy trade-off is especially relevant for massive MIMO and ultra-dense networks.
  • ML-based estimation: A 2D nonlinear complex SVR with an RBF kernel was proposed for accurate estimation in fast-fading time-varying multipath channels.The passage identifies the method and its target setting but provides no numerical result.
  • ML-based estimation: Deep learning can learn channel structure and estimate channels from large numbers of training samples in beamspace mmWave massive MIMO systems.Sparse Bayesian learning can exploit spatial sparsity to identify scatterer-path angles and gains.
  • Future direction: A generalized ML-based channel-estimation scheme is identified as a future direction because it could be used across different scenarios without further training.Constructing it requires pre-collected data from different environments to learn channel features.

III. PHYSICAL-LAYER RESEARCH FOR B5G NETWORKS USING AI TECHNOLOGIES

AI/ML can exploit data from large antenna arrays for physical-layer sensing, inference, and signal processing when conventional models are unavailable, inaccurate, or insufficient. The section also reviews applications to massive MIMO, including sparse Bayesian learning for pilot-contamination-related CSI estimation.

  • Sensing and inference: Large antenna arrays provide substantial baseband data that can support inference about environmental conditions beyond conventional communication functions.Examples include detecting moving objects, estimating road traffic, counting people, and intrusion monitoring.
  • Sensing and inference: ML is suited to analyzing the large data volumes generated by large antenna arrays for sensing and inference tasks.Deep-learning methods from image processing and video analytics are identified as potential tools.
  • Future research: Future work should combine physical modeling with ML foundations, evaluate on simulated channel models and real experimental data, and explore deep neural networks and dictionary learning.The paper presents this combination as a direction for advancing physical-layer research.
  • Sensing and inference: AI-based sensing may support open-space, indoor, and through-the-wall inference tasks that conventional model-based signal processing cannot accomplish.Gesture recognition is also identified as a possible application.
  • Massive MIMO signal processing: Pilot contamination from interference between adjacent cells can limit accurate CSI acquisition in massive MIMO systems.Sparse Bayesian learning exploits beamspace channel sparsity, and the reviewed method performs better than conventional CSI estimators in pilot contamination.
  • Massive MIMO signal processing: Sparse Bayesian learning achieves better performance than conventional CSI estimators in terms of pilot contamination.The surrounding discussion frames sparse recovery as an important issue for Bayesian compressive sensing.

C. Data-driven Localization in Wireless Networks

Data-driven localization uses machine learning on raw sensing and communication data to reduce dependence on labor-intensive channel updates and adapt to changing wireless conditions. Accurate positioning can support context awareness, location-based network management, and improved location-based services.

  • Limitations of Existing Methods: Current positioning techniques use channel information and fingerprinting, but channel characteristics require frequent updates because transmission impediments vary over time.Examples include path loss, interference, and blockage.
  • Limitations of Existing Methods: Periodic, long-term channel maintenance is time-consuming and labor-intensive, particularly for large-scale B5G systems.The burden grows as the scale and diversity of sensing and communication data increase.
  • Motivation: Data-driven localization positions devices and users by learning from raw sensing and communication data with ML algorithms.The approach is motivated by increasing sensing and communication data in B5G systems.
  • Data-driven Localization: Wireless channel locations can be continuously updated and improved by learning automatically from crowd-sourced big data collected by many mobile devices.This supports adaptive maintenance of localization information.
  • Applications: Accurate localization results can provide better location-based services and support context awareness and location-based network management.The paper identifies positioning as valuable for both service delivery and network operations.

B. Proactive Wireless Networking for Online Software Networks Orchestration

Future wireless networks must move beyond reactive service handling because heterogeneous, software-defined architectures and diverse requirements demand prediction and proactive resource allocation. AI-enabled prediction can support online network slicing and more effective use of spectrum and network resources.

  • Network Slicing: Network slicing creates customized network pipes for services with diverse functionality, performance, and isolation requirements.It is presented as a fundamental necessity for future cellular architectures.
  • From Reactive to Proactive Networking: Current 5G networks use a reaction principle that passively responds to incoming demands, but this is inadequate for future service capabilities.The limitation is especially relevant when services have stringent requirements such as ultra-low latency.
  • From Reactive to Proactive Networking: Future networks mix traffic from logically independent services, creating highly dynamic conditions that the reaction principle may make unmanageable.The networks are described as heterogeneous software-defined networks.
  • Proactive Networking: Prediction capabilities enable networks to anticipate future conditions and proactively allocate resources.This contrasts with passively serving demands only after they arrive.
  • Proactive Networking: Predicting traffic patterns and off-peak spectrum availability can improve allocation of incoming demands over a given time window.The approach uses network slicing to allocate resources online according to predicted user behavior.
  • From Reactive to Proactive Networking: A shift from reactive to proactive network design can be enabled with AI.The paper connects AI-based prediction with more anticipatory network operation.

V. AI ALGORITHMS AND APPLICATIONS FOR B5G NETWORKS

AI algorithms for B5G must support distributed operation across cloud, fog, and edge environments while adapting to changing information sources and communication constraints. Key challenges include balancing complexity, latency, and reliability and learning evolving relationships among network entities.

  • Distributed Architectures: Conventional centralized processing can become unsuitable when many devices and limited fronthaul or backhaul links restrict information exchange with the cloud.These conditions motivate local execution or minimal exchange with the cloud.
  • Distributed Architectures: Decentralized functional architectures should adapt dynamically to network requirements, with lightweight deep learning applicable across cloud, fog, and edge networks.The cloud is the data and computing center, while fog and edge networks distribute processing closer to devices.
  • Distributed Learning: Distributed learning, classification, and signal processing must adapt to the number and type of information sources and the available communication bandwidth.This extends beyond conventional end-user and device execution settings.
  • Distributed Learning: Dynamic edge computing requires combining decentralized and centralized algorithms while trading off complexity, latency, and reliability.The paper identifies data fusion, compression, and distributed decision-making as needed supporting methods.
  • Dynamic Network Inference: Scalable solutions are needed to learn relationships among network entities and their time evolution in distributed settings.The paper highlights online learning methods as a direction for this complex task.

B. ML Algorithms for Ultra-fast Training and Inference

B5G requires ML models with ultra-fast training and inference to meet high processing rates and ultra-low latency, while many communication devices also impose tight resource constraints. Promising directions include hardware implementation, complexity reduction, and distributed computing for lightweight models.

  • Ultra-fast Training and Inference: B5G networks require ultra-fast ML training and inference because they must process data at high rates for ultra-low latency.The speed requirement applies particularly to model inference.
  • Acceleration Directions: Hardware implementation may accelerate ML training while reducing power consumption and increasing efficiency.This is one of two directions identified for improving training speed.
  • Acceleration Directions: Reducing ML algorithm complexity while retaining reasonable accuracy is another direction for faster training.The approach targets a practical balance between computational cost and model performance.
  • Embedded Systems: Communication systems include resource-constrained embedded and IoT devices with limited storage, computational power, and energy resources.ML algorithms must therefore operate effectively under these device limitations.
  • Embedded Systems: Developing lightweight ML, especially deep learning models, for embedded systems is challenging but potentially rewarding.Combining ML with distributed computing is identified as a possible direction.
  • Development Tools: High-level ML development libraries and toolboxes are identified as an additional research direction.The paper also relates ML development to fog and edge computing frameworks.

VI. AI/ML FOR B5G NETWORKS IN STANDARDS AND STUDY GROUPS

AI/ML standardization for B5G is progressing through multiple organizations and study groups, but no standard or baseline ML algorithm has yet been established. These efforts address architectures, data, algorithms, network management, and automation.

  • No standard or baseline ML algorithm has been established, and suitable ML algorithms for B5G remain unclear.
  • ITU, 3GPP, FuTURE, TIP, and 5G PPP have initiated preliminary AI/ML activities for future and 5G networks.
  • The ITU ML5G focus group is developing reports and specifications covering interfaces, architectures, protocols, algorithms, and data formats.
  • 3GPP's NWDAF could become a central analytics point in the 5G core network, although its standardization remains at an early stage.
  • TIP and 5G PPP efforts apply AI/ML to network planning, operations, customer behavior, and network management to improve automation and service experience.
  • These standards and study-group developments target physical-layer and network-management applications that could boost wireless-network performance.

VII. CONCLUSIONS AND FUTURE CHALLENGES

The paper surveys how AI and ML can address unstructured and seemingly intractable problems in future B5G wireless networks. It synthesizes advances, applications, standards activity, and future research challenges across major network-design areas.

  • The study investigates how AI and ML can solve unstructured and seemingly intractable problems in future B5G wireless communication networks.
  • The survey covers channel measurements, modeling, estimation, physical-layer research, and network management and optimization.
  • It discusses challenges and potential future research directions for combining AI/ML with wireless networks.
  • The paper introduces ML algorithms and their applications to B5G networks.
  • It provides an overview of AI/ML developments undertaken by standards organizations and study groups for B5G systems.

Distinguished Lecturer of the IEEE Communications Society (COMSOC) and IEEE Vehicular Technology

The supplied material is primarily biographical and figure-caption content rather than a coherent research section. It identifies authors' affiliations and research interests and labels four figures concerning AI/ML applications in B5G.

  • The material includes biographical information about researchers in wireless communications, signal processing, information theory, robotics, and machine learning.
  • Figure 1 is labeled as covering research aspects that bring AI technologies into B5G wireless networks.
  • Figure 2 compares measured and predicted path loss and RMS DS, with FNN and RBF-NN listed as prediction methods.
  • Figure 3 concerns optimal demodulation thresholds for small and large ISI values.
  • Figure 4 depicts deep-learning applications in cloud, fog, and edge-computing networks.
Loading 2001.08159v1…