PaperlayerView exact revision on arXiv

Source-linked AI summary

Application of Machine Learning in Wireless Networks: Key Techniques and Open Issues

Yaohua Sun, Mugen Peng, Yangcheng Zhou, Yuzhe Huang, Shiwen Mao

arXiv:1809.08707v2cs.NIcs.LG

TL;DR

Wireless networks generate large, dynamic, and difficult-to-model optimization problems, while prior surveys leave gaps in technique and application coverage. This paper comprehensively surveys ML methods across wireless resource management, networking, mobility management, and localization, organizes studies by technique, compares them with traditional approaches, and identifies open issues. Reported studies include a 125% improvement over uniform power allocation and a 50.79% femtocell-capacity increase over smart power control, alongside broader evidence that ML supports autonomous and lower-complexity wireless control.

  • Problem

    Wireless networks are increasingly data-rich and dynamic, while existing algorithms may waste data and require complex repeated optimization or unavailable network information.

  • Method

    The paper comprehensively surveys ML techniques and their applications across wireless resource management, networking, mobility management, and localization, organizing studies by method and comparing traditional approaches.

  • Results

    The survey reports that learning-based approaches can improve wireless control, including 125% improvement over uniform power allocation and 50.79% higher average femtocell capacity than smart power control.

  • Takeaways & Limitations

    The surveyed evidence highlights ML's use for lower-complexity decisions, autonomous resource allocation, and control under incomplete network information.

Abstract

from arXiv · show

As a key technique for enabling artificial intelligence, machine learning (ML) is capable of solving complex problems without explicit programming. Motivated by its successful applications to many practical tasks like image recognition, both industry and the research community have advocated the applications of ML in wireless communication. This paper comprehensively surveys the recent advances of the applications of ML in wireless communication, which are classified as: resource management in the MAC layer, networking and mobility management in the network layer, and localization in the application layer. The applications in resource management further include power control, spectrum management, backhaul management, cache management, beamformer design and computation resource management, while ML based networking focuses on the applications in clustering, base station switching control, user association and routing. Moreover, literatures in each aspect is organized according to the adopted ML techniques. In addition, several conditions for applying ML to wireless communication are identified to help readers decide whether to use ML and which kind of ML techniques to use, and traditional approaches are also summarized together with their performance comparison with ML based approaches, based on which the motivations of surveyed literatures to adopt ML are clarified. Given the extensiveness of the research area, challenges and unresolved issues are presented to facilitate future studies, where ML based network slicing, infrastructure update to support ML based paradigms, open data sets and platforms for researchers, theoretical guidance for ML implementation and so on are discussed.

I. INTRODUCTION

The introduction motivates ML for wireless networks by contrasting data-rich, dynamic environments with conventional algorithms and surveys prior work's gaps. It positions this paper as a broad, layer-spanning synthesis of ML techniques, applications, and open issues.

  • 5G wireless networks must support higher data rates, broader coverage, more connected devices, lower latency, lower energy consumption, and improved localization.
  • Growing application, user, and network data creates opportunities for performance gains, but many existing algorithms cannot process or exploit it effectively.
  • ML learns from data or environment feedback without explicit programming through supervised, unsupervised, and reinforcement learning paradigms.
  • ML can extract spatial, sequential, and robust fingerprint features from noisy or time-varying RSSI for mobility management and indoor localization.
  • Earlier surveys often omit deep learning, deep reinforcement learning, transfer learning, or broad cross-layer coverage, motivating this survey's comprehensive scope.

2) K Nearest Neighbors:

This section introduces unsupervised learning and core reinforcement-learning concepts used in wireless-network studies, including clustering, Q-learning, multi-armed bandits, and actor-critic learning.

  • 2) K Nearest Neighbors:: Supervised learning learns input-to-output mappings from labeled data, whereas unsupervised learning discovers hidden structure from unlabeled data.
  • 2) K Nearest Neighbors:: K-means partitions n data points into K clusters by assigning each point to its nearest mean and iteratively updating cluster means.
  • 2) K Nearest Neighbors:: Q-learning estimates discounted cumulative rewards for state-action pairs so an agent can select the action with the largest learned Q value.
  • 2) K Nearest Neighbors:: Multi-armed bandits balance exploitation of the current best action against exploration for potentially larger future rewards.
  • 2) K Nearest Neighbors:: Actor-critic learning combines an actor that selects actions with a critic that updates state values using a time-difference error.

3) Actor-Critic Learning:

The section describes neural-network architectures and learning procedures that extend conventional ML for wireless applications. It covers actor-critic, joint utility-strategy estimation, DRL, DNNs, RNNs, CNNs, and autoencoders.

  • 3) Actor-Critic Learning:: Joint utility and strategy estimation updates each agent's expected utility and action-selection probabilities in the same iteration.
  • 3) Actor-Critic Learning:: The joint utility-strategy method is fully distributed when rewards can be calculated locally, supporting autonomous multi-agent decisions.
  • 3) Actor-Critic Learning:: DRL uses a deep Q network to approximate optimal Q values from high-dimensional sensory data, with replay memory and a target network stabilizing training.
  • 3) Actor-Critic Learning:: DNNs transform inputs through successive layers of weighted neurons and activation functions, with backpropagation and gradient-descent methods optimizing parameters.
  • 3) Actor-Critic Learning:: RNN outputs depend on current and former inputs, while CNNs learn local patterns and spatial hierarchies through convolution and pooling.
  • 3) Actor-Critic Learning:: Autoencoders learn compressed, robust input representations and commonly retain the encoder as a feature extractor after training.

4) Auto-encoder:

The section connects transfer learning, learning assumptions, and ML-based resource management with concrete power-control applications. It also contrasts learned approaches with heuristic alternatives and identifies implementation constraints.

  • 4) Auto-encoder:: Transfer learning reuses knowledge from one domain to accelerate learning in another, but differences between tasks or environments can make former knowledge harmful.
  • 4) Auto-encoder:: Supervised models implicitly assume that mappings learned from training data remain effective in future environments.
  • 4) Auto-encoder:: Dynamic wireless states make repeated execution of complex, assumption-dependent resource-management algorithms costly, motivating ML-based policies and approximators.
  • 4) Auto-encoder:: Q-learning-based power control lets Pico and macro base stations autonomously optimize inter-cell interference coordination with limited coordination.
  • 4) Auto-encoder:: Distributed Q-learning manages aggregated interference from cognitive radios, including partial observations handled through belief states.

2) Supervised Learning Based Approaches:

Machine-learning approaches address wireless resource-management problems using neural networks, reinforcement learning, and distributed decision-making. Reported applications target faster power allocation, spectrum coordination, and adaptive resource use under incomplete or changing network information.

  • Power Control: Deep neural networks approximate high-complexity WMMSE power allocation to provide similar performance with much lower computation time.Training uses channel realizations and WMMSE-generated power allocations as labeled data.
  • Power Control: CNN-based power control maps normalized channel gains to power allocations and can subsequently optimize spectral or energy efficiency.The CNN is first trained to approximate WMMSE, then its loss function is adjusted for SE or EE.
  • Power Control: Distributed Q learning and joint utility-and-strategy estimation support self-organizing power control in cognitive-radio and heterogeneous networks.Q values may be represented tabularly or with neural networks, while transfer learning can improve adaptation to dynamic environments.
  • Spectrum Management: Distributed learning methods manage spectrum access through potential-game Q learning, multi-armed bandits, and joint utility-and-strategy estimation under limited information.These formulations address channel allocation, interference constraints, unknown node counts, and decentralized operation.
  • Spectrum Management: Echo-state-network multi-agent reinforcement learning improves the sum rate of the 50th-percentile users by up to 167% compared to Q learning.Each base station uses ESNs to estimate immediate and expected utilities while broadcasting only limited action information.

2) Lessons Learned:

The surveyed backhaul and caching studies use reinforcement learning and neural prediction to adapt network decisions to congestion, user demand, and heterogeneous resource conditions. Their lessons emphasize reward design, decentralized adaptation, popularity estimation, and direct policy optimization.

  • Backhaul Management: Reinforcement learning can adapt cell-range-extension parameters to relieve macrocell backhaul congestion while improving user QoE.Q learning minimizes a cost designed to maximize total backhaul utilization while satisfying each cell’s capacity constraint.
  • Backhaul Management: Joint utility-and-strategy estimation balances prefetching potentially requested files against serving urgent traffic in backhaul management.In one formulation, the learning scheme converges to a unique equilibrium point.
  • Backhaul Management: Reward functions can combine system-centric and user-centric objectives through a weighted difference between backhaul resource utilization measures.Distributed reinforcement learning also enables heterogeneous backhaul-link selection, improving rate and delay in relay scenarios.
  • Cache Management: Caching studies use decentralized reinforcement learning, deep reinforcement learning, and popularity prediction to adapt placement decisions across cells and time.DRL can jointly orchestrate computing, networking, and cache resources, while other approaches directly optimize caching policies.

1) Reinforcement Learning Based Approaches:

The survey covers machine-learning methods for caching, content-demand prediction, computation offloading, and fast resource allocation. Across these applications, learning either predicts demand or maps network states directly to adaptive decisions, with reported gains over selected conventional baselines.

  • Cache Management: Deep reinforcement learning jointly orchestrates computing, networking, and cache resources in connected vehicular networks.The agent assigns vehicles to base stations and decides whether requested content should be cached.
  • Cache Management: Echo state networks learn content-request distributions and mobility patterns from limited training data for caching decisions.One reported scheme improves sum effective capacity by 27.8% and 30.7% over random caching with and without clustering, respectively.
  • Cache Management: Content popularity estimation combines 3D CNN feature extraction, SVM representation, and regression before determining cached video portions.This pipeline differs from approaches that predict popularity directly with neural networks.
  • Computation Resource Management: Double-DQN-based deep reinforcement learning optimizes computation offloading without requiring network-dynamics information and addresses state-space explosion.The surveyed formulation considers time-varying channel quality, task arrivals, and device energy status.
  • Resource Allocation: Supervised learning transforms resource management into classification by labeling scenario feature vectors according to their optimal or near-optimal allocation solutions.The resulting classifier quickly determines resource allocation for a new scenario; KNN beam allocation improves sum rate over a state-of-the-art method.

1) Lessons Learned:

Machine learning approaches address dynamic wireless networking problems including user association and BS switching by adapting to traffic, channel, and service conditions. The surveyed lessons emphasize exploiting temporal regularity, incorporating relevant costs and constraints, and balancing learned knowledge with exploration.

  • User Association: Vehicular user association can exploit spatial-temporal traffic regularity through online reinforcement learning to balance loads across heterogeneous BSs.The approach uses initial multi-armed-bandit learning followed by history-based reinforcement learning.
  • User Association: Fuzzy Q learning supports distributed user association under backhaul-capacity and user-QoE constraints, while MAB methods must balance exploration and exploitation.The surveyed lessons report that fuzzy Q learning can outperform Q learning in continuous state spaces.
  • BS Switching: Actor-critic and Q-learning methods formulate BS switching as sequential decisions over traffic states, optimizing energy consumption without requiring prior traffic-load knowledge.Some formulations also include transmission gains or switching-transition power in the feedback cost.
  • BS Switching: Fuzzy Q learning can improve energy efficiency under fluctuating user density while preserving network capacity and coverage probability.The method controls SBS sensing and sleep-wake behavior through an optimized sensing probability.

2) Transfer Learning Based Approaches:

The surveyed approaches use reinforcement learning, clustering, and neural networks to manage dynamic routing and switching decisions. Transfer learning reuses prior policies, while routing methods learn from bottleneck-channel quality, network traffic, or experienced congestion.

  • Transfer Learning Based Approaches: Transfer learning combined with actor-critic reinforcement learning reuses historical or neighboring-region policy values to accelerate BS-switching decisions.The surveyed result reports improvements over reinforcement learning alone in energy saving and convergence speed.
  • Transfer Learning Based Approaches: Transfer learning can negatively influence network decisions when source and target tasks differ, motivating mechanisms that reduce prior-knowledge influence over time.The surveyed discussion also notes that WiFi network scenarios were not covered by some prior BS-sleeping studies.
  • Routing: In cognitive-radio routing, Q-learning can use bottleneck-link channel availability as reward feedback to guide route selection.An average-Q-value variant selects more stable routes than traditional reinforcement learning, with performance verified on a test bed.
  • Routing: Q-learning geographic routing in high-mobility robotic networks can improve packet delivery ratio while reducing network overhead.The reward incorporates packet travel speed, and the result was confirmed through NS-3 simulation.
  • Routing: DNN- and CNN-based routing learn from OSPF-generated traffic data or experienced congestion to predict next routers and avoid congested or faulty routing decisions.The surveyed synthesis associates DNN routing with lower signaling overhead and higher throughput, while CNN routing is trained online.

2) Supervised Learning Based Approaches:

Supervised, unsupervised, and reinforcement learning methods are applied to clustering and routing-related decisions in dynamic wireless networks. Neural networks reduce cluster-search complexity, K-means groups nodes by similarity, and deep reinforcement learning selects clusters under changing conditions.

  • Routing: CNN-based routing evaluates traffic patterns to predict whether routing strategies will cause congestion, continuing until a non-congested strategy is identified.The surveyed discussion contrasts online CNN routing with OSPF and reports that it can avoid past faulty routing decisions.
  • Routing: For cognitive-radio ad hoc routing, network characteristics such as network size have slight impacts on secondary-user end-to-end delay and packet loss rate.This observation is reported for Q-learning-based routing.
  • Clustering: DNNs can identify users or BSs that need not join clusters, reducing the search space for cluster formation.This supports optimization when the number of potential user clusters grows exponentially with the number of served users.
  • Clustering: K-means clustering groups BSs or users using similarity features such as location, traffic load, or RSRQ before optimizing switching or coordination.For SBSs, intra-cluster coordination can increase the number of switched-off BSs by offloading users to active BSs.
  • Clustering: Deep reinforcement learning can select cluster members directly in dynamic networks with time-varying channel state information and cache states.The method is presented for cache-enabled opportunistic interference alignment and link scheduling.

V. MACHINE LEARNING BASED MOBILITY MANAGEMENT

The survey organizes ML-based mobility management around handover optimization, user association, mobility prediction, and RF-aware decision-making. It also covers ML applications to indoor localization, including KNN, SVM, RVM, neural networks, and autoencoders.

  • Mobility management applications include user mobility prediction and handover parameter optimization.
  • Fuzzy Q learning adjusts handover parameters using call drop rate and handover signaling load as state information.
  • Deep reinforcement learning can optimize multi-user handovers after clustering users by mobility patterns with unsupervised learning.
  • Unsupervised shapelets identify useful RSRP patterns for cell-edge RF awareness, supporting handover parameter setting.
  • 70% fewer unnecessary handovers were reported for a self-organizing-map method that uses RSRP and angle-of-arrival information.
  • The surveyed localization methods include KNN, online SVM, RVM, ELM, and neural networks using RSS or CSI information.

2) Unsupervised Learning Based Approaches:

Unsupervised learning is applied to extract patterns from wireless measurements and to formulate wireless problems as clustering or feature-discovery tasks. The survey emphasizes that ML is appropriate when the problem can be abstracted into a suitable task and essential data are available.

  • Autoencoders extract robust information from RSS or CSI data, contributing to higher indoor localization accuracy.
  • KNN-based indoor localization compares measured RSS vectors with fingerprints collected at reference points.
  • For static KNN localization, k = 1 and k = 2 are reported as better choices than other neighborhood sizes.
  • The survey identifies regression, classification, clustering, and Markov decision making as suitable abstractions for ML-based wireless problems.

B. Training Data Availability

Applying ML in wireless networks requires suitable training data, acceptable response and training times, and manageable implementation complexity. Technique selection also depends on the problem type and the relative trade-offs among available ML methods.

  • Training data may come from repeated traditional-algorithm runs, cloud-collected network parameters, recorded traffic patterns, or reinforcement-learning rewards.
  • Essential training data must be acquirable before ML is applied to a wireless-network problem.
  • Response time and training time are distinct constraints, and training must keep pace with changes in the communication environment.
  • A trained DNN made a power-control decision for 30 users in 0.0149 ms on average without GPUs.
  • Fuzzy Q learning has low implementation complexity because it uses stored Q values, simple operations, and readily available network KPIs.
  • For regression and classification, KNN is simple, SVM supports robust low-complexity binary classification, and deep neural networks strengthen feature extraction with more data.
  • Actor-critic learning can learn an explicit stochastic policy that may be useful in non-Markov environments.

VIII. ALTERNATIVES TO MACHINE LEARNING AND MOTIVATIONS

The survey compares ML-based wireless methods with heuristic, search-based, optimization-based, and conventional association alternatives. These comparisons motivate ML through improved performance, competitive results, or lower signaling, computing, and information-acquisition costs.

  • 125% performance improvement over uniform power allocation was reported for Q learning, while Q learning improved average femtocell capacity by 50.79% over smart power control.
  • Reinforcement learning reached nearly the same average cell throughput as centralized dynamic spectrum sharing without information sharing between base stations.
  • Distributed reinforcement learning achieved similar spectral efficiency to exhaustive search while avoiding its large cloud-computing burden.
  • Reinforcement-learning cache management achieved 13% and 56% higher per-base-station utility than random caching and time-averaged popularity caching, respectively.
  • KNN beam allocation approached optimal average sum rate and outperformed the low-complexity beam allocation algorithm as training data increased.
  • Multi-armed-bandit association achieved performance similar to exhaustive search with lower complexity and lower information-acquisition overhead.

H. Alternatives for BS Switching Control

The survey compares ML-based wireless-network methods with traditional schemes across routing, clustering, mobility management, localization, and resource management. It highlights performance gains, lower complexity, reduced information requirements, and better adaptation as recurring motivations for ML adoption.

  • Routing: Learning-based routing can achieve lower end-to-end delay than shortest-path and PU-aware shortest-path baselines.
  • Routing: Deep learning based routing reduces average per-hop delay by around 93% compared with OSPF routing.
  • Clustering: Deep reinforcement learning improves user sum rate in time-varying CSI environments compared with static-environment user-selection schemes.
  • Mobility management: Deep reinforcement learning mitigates the ping-pong effect while maintaining a high data rate compared with RSSI-based handover.
  • Motivations for ML: ML is adopted to reduce algorithmic complexity, operate with incomplete network knowledge, and enable self-organization with lower signaling overhead.
  • Learning robust patterns: Neural networks learn spatial, sequential, and fingerprint patterns that support power control, mobility management, and localization.

IX. CHALLENGES AND OPEN ISSUES

The survey identifies unresolved issues spanning network slicing, infrastructure, standardized research resources, transfer learning, and theoretical guarantees. These issues concern practical deployment, fair evaluation, scalability across environments, and reliability of ML-based wireless systems.

  • Infrastructure update: ML deployment requires updated infrastructure, including edge GPUs, storage, and network-function virtualization.
  • ML-based network slicing: ML-based network slicing should learn service-demand mappings to resource-allocation plans while coordinating computing, caching, backhaul, fronthaul, and radio resources.
  • Standard datasets and environments: Common wireless-network problems, labeled and unlabeled datasets, and well-defined reinforcement-learning environments are needed for fair comparisons.
  • Theoretical guidance: Open questions include hyperparameter selection, training-dataset size, deep-learning performance bounds, and model generalization; rigorous theory is needed for stability guarantees.
  • Transfer learning: Transfer learning can accelerate adaptation to new environments and maintain good performance with limited training data, but training cost remains a practical concern.
Loading 1809.08707v2…