PaperlayerView exact revision on arXiv

Source-linked AI summary

AI based Service Management for 6G Green Communications

Bomin Mao, Fengxiao Tang, Kawamoto Yuichi, Nei Kato

arXiv:2101.01588v2cs.NI

TL;DR

The paper addresses the challenge of improving energy efficiency in 6G networks whose expanding infrastructure, diversified service requirements, and dynamic energy harvesting complicate management. It surveys heuristic, ML, and DL approaches for AI-based service management and energy harvesting, while examining their cooperation with mathematical models. The survey identifies AI-based management as a route to simplifying optimization and anticipating network changes, alongside open challenges including computation overhead, privacy, hardware, and deployment.

  • Problem

    Growing 6G infrastructure, diversified QoS requirements, and dynamic energy harvesting make energy-efficient power control and network management increasingly complex.

  • Method

    The paper surveys heuristic, ML, and DL techniques for green communications across network management and energy-harvesting scenarios.

  • Results

    The surveyed AI approaches simplify conventional iterative optimization, learn network-control relationships, and support prediction of future network changes.

  • Takeaways & Limitations

    AI-based green communications require systematic consideration of energy use, data privacy, computation complexity, hardware design, and network deployment.

  • Takeaways & Limitations

    Conventional optimization is constrained by large solution spaces, multi-factor non-convex or NP-hard formulations, and frequent failures.

Abstract

from arXiv · show

Green communications have always been a target for the information industry to alleviate energy overhead and reduce fossil fuel usage. In current 5G and future 6G era, there is no doubt that the volume of network infrastructure and the number of connected terminals will keep exponentially increasing, which results in the surging energy cost. It becomes growing important and urgent to drive the development of green communications. However, 6G will inevitably have increasingly stringent and diversified requirements for Quality of Service (QoS), security, flexibility, and even intelligence, all of which challenge the improvement of energy efficiency. Moreover, the dynamic energy harvesting process, which will be adopted widely in 6G, further complicates the power control and network management. To address these challenges and reduce human intervene, Artificial Intelligence (AI) has been widely recognized and acknowledged as the only solution. Academia and industry have conducted extensive research to alleviate energy demand, improve energy efficiency, and manage energy harvesting in various communication scenarios. In this paper, we present the main considerations for green communications and survey the related research on AI-based green communications. We focus on how AI techniques are adopted to manage the network and improve energy harvesting toward the green era. We analyze how state-of-the-art Machine Learning (ML) and Deep Learning (DL) techniques can cooperate with conventional AI methods and mathematical models to reduce the algorithm complexity and optimize the accuracy rate to accelerate the applications in 6G. Finally, we discuss the existing problems and envision the challenges for these emerging techniques in 6G.

I. INTRODUCTION

6G’s expanding infrastructure, services, and device population intensify ICT energy demands, while shorter THz coverage increases the required number of base stations. The paper surveys AI-based management across three communication scenarios to improve energy efficiency and handle increasingly complex optimization problems.

  • I. INTRODUCTION: 6G may sharply increase energy consumption through expanded infrastructure, indoor THz base-station deployment, cloud and edge services, and complex AI-enabled network management.THz transmission ranges are envisioned to shrink from 100 m for mmWave to 10 m, requiring more base stations for coverage.
  • I. INTRODUCTION: Energy-efficient network design and energy harvesting are the two main solution directions, including solar, wind, vibration, RF harvesting, and intelligent reflecting surfaces.RF harvesting can support simultaneous information and energy transmission, while IRSs can reflect wasted signals to improve SINR.
  • 3) Advantages of AI Methods:: Heuristic algorithms, ML, and DL are adopted to simplify iterative mathematical optimization, learn complex parameter relationships, and predict network changes for automatic management.ML/DL models can learn power-control and resource-allocation policies and enable network adjustments before parameter changes cause performance deterioration.
  • I. INTRODUCTION: The survey organizes AI-based green-communication research around cellular network communications, machine type communications, and computation oriented communications.These scenarios involve base-station scheduling, access and routing for constrained devices, and computation offloading, resource allocation, and server deployment.
  • 2) Limitations of Conventional Methods:: Energy-efficiency optimization jointly involves power control, scheduling, resource allocation, network design, and user association, making many formulations non-convex or NP-hard.Large solution spaces, diverse 6G service requirements, nonlinear parameter relationships, and node mobility constrain conventional mathematical optimization.

B. Scope

This survey examines AI-based green communications across three 6G service scenarios, covering energy management from both communication and computation perspectives. It reviews AI models, their performance-enhancement techniques, and challenges including computation overhead, security, and deployment.

  • Scope: The survey focuses on AI-based approaches for reducing energy cost and improving energy efficiency across cellular, machine-type, and computation-oriented communications.It extends beyond surveys centered on definite networks and organizes research around three 6G communication services.
  • Scope: It summarizes communication components and techniques for alleviating energy demand and improving energy efficiency.
  • AI methods: It introduces conventional AI models alongside state-of-the-art machine learning and deep learning methods for energy management and network performance.
  • Scope: The analysis covers green ICT from communication and computation viewpoints across THz cellular networks, SA-GINs, DCNs, VANETs, and IoTs.
  • AI methods: The paper examines AI-model design, including common techniques and mathematical methods intended to improve AI accuracy.
  • Challenges: It identifies overwhelming computation overhead, security issues, and practical deployment as challenges for AI-based 6G green communications.

A. Traditional AI Algorithms

Traditional AI techniques for green communications include non-data-based heuristics and data-based ML, while ANN-derived models support increasingly complex optimization tasks.

  • Heuristic Algorithms: Heuristic algorithms seek good-enough solutions to NP-hard problems within limited time using shortcuts rather than exhaustive greedy search.PSO, ACO, and GA are representative heuristic methods.
  • Heuristic Algorithms: Particle Swarm Optimization updates particle positions using individual and globally best-known locations to search for satisfactory solutions.Its limitations include local optima in high-dimensional spaces and low convergence rate.
  • Heuristic Algorithms: Ant Colony Optimization simulates artificial ants whose recorded positions and solution quality guide later iterations toward better routes.The method has been studied for improving energy efficiency in network applications.
  • Heuristic Algorithms: Genetic Algorithms evolve candidate solutions through chromosome crossover, mutation, selection, and fitness evaluation.These operations generate new candidate populations while directing the search toward expected solutions.
  • Machine Learning Algorithms: Common ML techniques for green communications include regression analysis, SVM, and K-means clustering, while ANN-based models underpin modern ML/DL development.Regression maps labeled inputs to outputs, SVM supports supervised classification or regression, and K-means partitions observations by nearest cluster center.

C. Future Perspective AI Learning Methods

Future 6G AI learning methods address dynamic, complex network management through reinforcement learning, transfer learning, and federated learning, while conventional methods remain useful in resource-limited settings.

  • C. Future Perspective AI Learning Methods: Advanced learning methods are motivated by future networks’ complex scenarios and dynamics beyond traditional supervised and unsupervised learning.The paper highlights reinforcement, transfer, and federated learning as methods likely to attract increasing attention.
  • C. Future Perspective AI Learning Methods: Reinforcement Learning learns through trial and error, with an agent selecting actions from states and receiving rewards or penalties.Deep Reinforcement Learning uses DL models to map states to actions when Q-value tables become infeasible.
  • C. Future Perspective AI Learning Methods: Transfer learning reuses knowledge from a related problem by fine-tuning a new model or training only part of it.This can reduce computation consumption and required training data.
  • C. Future Perspective AI Learning Methods: Federated learning trains models across distributed servers or devices while keeping local training data and sharing model parameters with a central controller.The controller integrates parameters, and edge devices download updated models for prediction or periodic updating.
  • C. Future Perspective AI Learning Methods: DL is suited to more complex problems, but heuristic algorithms and shallow ML remain appropriate for some resource-limited scenarios because their computation complexity is lower.The paper also identifies imitation learning and quantum machine learning as promising techniques not introduced in this section.

1) Power Consumption Modeling of BSs:

The paper models BS power consumption through sleep and active-state components, then defines energy efficiency using transmission rate per power and identifies controllable network factors.

  • 1) Power Consumption Modeling of BSs:: A BS’s power consumption includes constant sleep power and active-mode power comprising additional fixed consumption and usage-dependent transmission power.The model represents BS activity with a binary active-or-sleep parameter and usage rate η.
  • 1) Power Consumption Modeling of BSs:: Reducing BS energy consumption requires switching idle BSs to sleep mode and minimizing usage while active.The model separates basic sleep functions from computation, backhaul, power-supply, and transmission-related consumption.
  • 2) Energy Efficiency Measurement:: Energy efficiency is measured as transmission rate divided by power consumption, using bit-per-Joule as the unit.The discussion derives this measure for cellular-network UEs under a multi-cell interference model.
  • 2) Energy Efficiency Measurement:: UE energy efficiency depends on assigned bandwidth, channel gain, transmission power, and interference, while noise and static power consumption are usually constant.These dependencies motivate optimizing channel and bandwidth allocation, power control, and transmission scheduling.
  • 1) Power Consumption Modeling of BSs:: In 6G, THz penetration loss and uneven traffic distribution motivate minimizing the number and transmit power of working BSs, while BS switching must avoid QoS deterioration.The paper identifies deployment and management, power control, resource allocation, and renewable energy as major green-network directions.

1) Base Station Deployment:

AI-assisted base-station deployment and work-state management address changing traffic, propagation, and service conditions while balancing energy consumption, coverage, load, and QoS.

  • Base-station deployment: Deployment conditions vary with user density, propagation, physical surroundings, and climate, motivating more automatic placement strategies.
  • Base-station deployment: Hybrid approaches use ML to predict network parameters or evaluate fitness functions, while GA searches deployment policies iteratively.
  • Work-state management: Traffic-aware BS switching reduces energy consumption but requires user reassociation and careful scheduling to preserve qualified connections.
  • Work-state management: RL and transfer learning improve flexibility by selecting BS work modes and transferring learned spectrum-assignment knowledge to user association.
  • Work-state management: Switching off low-usage BSs can create coverage holes or overload nearby BSs, so AI methods balance energy efficiency against QoS through reward or cost functions.
  • Work-state management: Q-learning can jointly support load balancing and energy saving by coordinating BS activation with user association or offloading.

C. Power Control and Resource Allocation

AI-based power control and resource allocation target energy efficiency while accounting for interference, service requirements, channel conditions, and large-scale optimization complexity.

  • General power control: Transmit power control and resource allocation are critical because they affect interference and system energy efficiency in emerging 6G technologies.
  • General power control: Neural networks can map service arrival rates and packet information to transmit power and channel allocation for energy minimization.
  • General power control: Transfer learning fine-tunes selected network layers for non-stationary channels or changed service types, reducing the data needed for adaptation.For multiple services, newly added layers can be fine-tuned using a few training samples.
  • General power control: Offline ANN training from Branch-and-Bound solutions enables online transmit-power calculation and is reported robust to mismatches between training and real data.
  • Beamforming and cell sleeping: DNN-based joint cell sleeping and coordinated beamforming achieves advantages in power saving and QoS satisfaction over no-sleep and equivalent-association strategies.
  • Beamforming and cell sleeping: RL and Meaning Field Game methods guide beam-steering and hybrid beamforming when conventional numerical methods face large action and state spaces.

3) MIMO:

AI methods support energy-efficient MIMO, NOMA, and renewable-energy network management through learning-based allocation, clustering, prediction, and control.

  • MIMO: Unsupervised learning predicts pilot-power allocation from large-scale channel fading coefficients to mitigate pilot contamination in distributed massive MIMO.
  • MIMO: Game theory combined with RL controls MIMO transmit power to suppress jamming, eavesdropping, and spoofing while considering required energy efficiency.
  • MIMO: Cross-Entropy algorithms iteratively update analog-beamformer distributions and select elite beamformers that minimize total transmit power.
  • NOMA: Semi-supervised neural networks learn allocation strategies from channel gains using labeled and unlabeled data for NOMA radio-resource management.
  • NOMA: K-means clustering separates users among base stations in THz MIMO-NOMA networks, addressing severe path spreading and molecular absorption losses.
  • Renewable-energy management: AI tracks dynamic renewable harvesting and optimizes switching, scheduling, association, power control, and resource allocation in energy-harvesting networks.
  • Renewable-energy management: Centralized DRL manages harvesting-enabled SBS states using harvested energy, battery levels, traffic, throughput, and delay to balance energy efficiency and QoS.
  • Renewable-energy management: DQL is reported to achieve the best performance in energy saving and system outage among the compared battery-management approaches for highly dense scenarios.

E. summary

MTC energy consumption depends on transmission and reception circuitry plus distance-dependent power amplification, motivating sleep control and transmit-power minimization.

  • AI models regress complex relationships among network parameters and can replace massive iterative computation after training with conventional-method data.
  • MTC energy optimization targets access, routing, relay, and energy-dynamics management for numerous battery-constrained devices.
  • Power consumption modeling: A single-hop MTC power model accounts for transmission and reception circuitry and power-amplifier consumption, with multi-hop consumption derived by summing hop costs.
  • Power consumption modeling: Power-amplifier consumption varies with hardware, bias, load, frequency, output power, and transmission distance, while circuit terms are usually treated as constants.
  • Green MTC strategies put idle nodes into sleep mode and minimize required transmit power and transmission time for active nodes.

B. Energy-Efficient Network Access

AI-based methods optimize access configurations across heterogeneous MTC scenarios, targeting energy efficiency alongside connectivity, QoS, latency, and reliability. Research covers terrestrial, satellite, and UAV-assisted access using clustering, reinforcement learning, and deep reinforcement learning.

  • Terrestrial, satellite, and UAV access: MTC access research addresses heterogeneous QoS requirements and energy constraints across cellular, IEEE 802.15.4, LoRa, NB-IoT, satellite, and UAV platforms.These technologies support scenarios ranging from stable cellular connections to remote-area satellite coverage and mobile UAV access.
  • Terrestrial access configurations: Q-learning and clustering optimize access timing, node grouping, preamble selection, and transmission scheduling to reduce energy use, collisions, or access inefficiency.Examples include duty-cycle control in IEEE 802.15.4, K-means scheduling in LoRa, and adaptive Q-learning for NB-IoT access parameters.
  • Summary: The surveyed works are organized in a table of AI techniques for optimizing network access in green MTCs.The table complements the representative terrestrial, satellite, and UAV studies discussed in the section.
  • Access through satellites: Satellite IoT studies apply deep reinforcement learning to channel allocation and combine deep learning with optimization to improve energy efficiency while maintaining QoS or throughput.Satellite links face large path loss, while NOMA systems jointly address power allocation, rate control, and SIC decoding order.
  • UAV-assisted access: UAV-assisted IoT access uses reinforcement learning and actor-critic models to optimize trajectories, channel allocation, and transmit power according to network conditions.The actor-critic architecture selects channels and transmit powers for IoT uplinks when the UAV trajectory is fixed.

C. Energy-Efficient Transmissions

AI-based transmission methods optimize routing, relay selection, and device-to-device decisions under energy, latency, reliability, mobility, and spectrum constraints. The surveyed approaches use reinforcement learning, supervised learning, learning automata, and multi-armed bandits.

  • Transmission design: Transmission-path design in MTC must account for energy dynamics, node mobility, spectrum efficiency, and network performance in addition to selecting routes or relays.Cooperative transmission makes routing and relay choices directly relevant to both network performance and power consumption.
  • Routing: Temporal-difference learning updates routing tables using Boltzmann exploration, while Q-learning selects underwater forwarding nodes using residual energy and depth.The underwater formulation balances end-to-end delay and energy consumption through long-term rewards.
  • Routing: Supervised MLP learning optimizes transmission range rather than directly predicting the next hop, improving routing performance while minimizing energy consumption.This approach uses labeled network data to learn a transmission-range setting for low-power IoT networks.
  • Relay and D2D: Relay and D2D studies use supervised learning, Q-learning, and multi-armed bandits to select cooperation policies or neighbors under energy and performance objectives.Objectives include long-term utility, energy efficiency, throughput, packet delivery, latency, and residual-energy constraints.
  • Summary: The surveyed transmission studies are compiled in a table covering AI techniques for transmissions in green MTCs.The table summarizes the representative routing, relay, and D2D research discussed in this section.

D. Energy Harvesting and Sharing

Energy harvesting can reduce grid and fossil-fuel demand for MTC devices, but unpredictable harvested power and battery dynamics complicate network control. AI methods therefore predict battery or harvesting states and optimize access, power, security, or wake-up decisions.

  • Energy harvesting foundations: Renewable and RF energy harvesting are identified as two major approaches for powering MTC terminals toward green 6G communications.Renewable sources include solar, wind, and tide, while RF harvesting recovers otherwise unusable energy from RF signals.
  • Energy-state prediction: Dynamic harvesting power makes energy-efficiency optimization difficult, motivating AI methods that estimate future available power and battery status.These predictions support network configuration decisions for access and transmission control.
  • Renewable energy harvesting: A two-layer LSTM-DQN predicts battery levels before access control, and simulations report improved system sum rate and resulting energy efficiency.The first LSTM layer supplies predicted battery information as part of the access-control state.
  • Joint access and power control: A later DRL design combines LSTM battery prediction with actor-critic and DQN control for joint access and power decisions.The control state includes channel power gain, predicted battery level, and historical information.
  • RF harvesting and sharing: AI-based energy-harvesting research also addresses robust power estimation, wake-up scheduling, and active-time control when RF energy or channel information is uncertain.Linear regression, ANN, and robust Bayesian learning are among the methods used in these settings.

B. Energy-Efficient Cloud and Edge Computing

Green cloud and edge computing research uses AI to balance energy consumption, latency, resource capacity, workload placement, and privacy across heterogeneous computation platforms. The surveyed methods optimize offloading, resource allocation, server provisioning, association, and task scheduling.

  • Scope of computation management: Computation energy management focuses on offloading decisions, CPU/GPU resource allocation, and server placement across local, edge, fog, and cloud platforms.These platforms differ in latency, energy consumption, capacity, and communication overhead.
  • Computation offloading: Heuristic algorithms, deep learning, and deep reinforcement learning optimize fog/cloud offloading, active or passive offloading, and dependent-task CPU allocation under latency and energy objectives.The formulations include MINLP, multi-action policies, and weighted rewards combining energy consumption with latency.
  • Privacy-preserving learning: Federated learning combined with deep reinforcement learning optimizes partial offloading for energy-harvesting IoT nodes while keeping sensing data local and reducing transmission overhead.The edge node receives trained-agent parameters rather than raw sensing data.
  • Dynamic scheduling and learning efficiency: Imitation learning accelerates branch-and-bound scheduling in dynamic vehicular edge computing, while transfer learning reduces DNN training time by two orders of magnitude and approximates optimal allocation.The imitation-learning agent follows an expert policy learned offline; the transfer-learning DNN trains with limited samples.
  • Resource and platform selection: Other studies apply reinforcement learning, federated SVMs, and particle swarm optimization to hybrid-powered server provisioning, user association, and joint access-network and edge-cloud selection.These approaches target energy, latency, capacity, and task-size or workload constraints.

3) Edge Server and Virtual Machine Placement:

AI-based placement methods optimize edge, cloudlet, VM, and container deployments to reduce energy consumption while respecting service or resource constraints.

  • Edge Server and Virtual Machine Placement: Idle-state power is a major source of network energy waste, motivating AI-based optimization of active servers and virtual-machine deployment.Placement methods aim to improve resource utilization while meeting service requirements.
  • Edge Server and Virtual Machine Placement: PSO-based edge-server placement reduced energy consumption by more than 10% on a Shanghai Telecom dataset.The approach also considered an access-delay threshold for base stations.
  • Edge Server and Virtual Machine Placement: ACO-based VM placement minimizes active servers and balances resource utilization through a bipartite-graph formulation and global-search information.The method distributes pheromone information between VMs, servers, and VMs sharing a server.
  • Edge Server and Virtual Machine Placement: K-means clustering searches cloudlet location centers by grouping mobile devices, with energy consumption modeled through the number of deployed cloudlets.The placement objective is to minimize deployed cloudlets and improve energy efficiency.
  • Edge Server and Virtual Machine Placement: Discrete PSO for heterogeneous virtualized data centers lessened energy consumption by 13%-23% while satisfying service requirements.Two-dimensional encoding and local fitness were used to accelerate convergence and reduce search time.

2) Delivery:

AI methods optimize content delivery across access, UAV-assisted, core, and joint caching-delivery settings, while practical deployment remains challenging in 6G.

  • Delivery: DNN-based user clustering maps channel coefficients and requested data amounts to low-computation scheduling decisions that minimize base-station transmit power.The scheduling must satisfy stringent delay requirements and support real-time operations.
  • Delivery: DRL optimizes cache-enabled UAV trajectories to improve content delivery for intelligent transportation systems.The UAVs hover over vehicles to satisfy requests when local vehicle caches lack required content.
  • Delivery: ACO optimizes core-network data forwarding by reducing content-retrieval hops, lowering energy consumed by routers and links.The method divides CDNs into domains and models data and hello-message packets as ant types.
  • Delivery: DRL jointly minimizes latency and energy cost for content caching and delivery in radio access networks using a reward over users, SBSs, MBSs, and cloud servers.The approach adopts actor-critic and DDPG models.
  • Open Research Issues: Current AI-based green-communication research still faces computation-complexity, hardware-compatibility, and data-security challenges in moving toward practical 6G applications.The paper identifies practical deployment as an ongoing research priority.
  • Energy-Efficient Space-Air-Ground Integrated Networks: Future green management must address heterogeneous, mobile, renewable-powered SAGIN components, while current research mainly focuses on single layers.Whole-system AI orchestration is challenged by factor characterization and extreme execution overhead.

C. AI-based Energy-Efficient Transmissions

Energy-efficient transmission research must coordinate multiple communication modes, dynamic harvesting, security, and AI resource use. The paper surveys AI techniques for these coupled management problems while highlighting their practical trade-offs.

  • AI-based Energy-Efficient Transmissions: Hybrid transmission scheduling must coordinate cellular, WiFi, D2D, routing, relaying, backscatter, and IRS-assisted communication in multi-agent multi-task environments.Most existing AI studies address single communication scenarios, with limited work on hybrid scenarios.
  • Energy Harvesting: AI can optimize BS power control and transmission scheduling for RF harvesting and UAV-BS trajectories for harvesting efficiency.These methods target uncontrollable or partially controllable energy-harvesting settings.
  • Energy Harvesting: RF harvesting enables energy sharing to reduce outages and waste when devices cannot store additional incoming energy.SWIPT has been widely studied, especially for machine-type communications, although harvesting can cause performance loss.
  • Security: AI-based threat detection, transmit-power control, and resource allocation can address attacks and network jammers that degrade transmission reliability and energy efficiency.The passage links adversarial activity to privacy threats, transmission failures, and deteriorated energy efficiency.
  • Open Research Issues: AI-based green communications must account for training and inference energy, data security, algorithm complexity, hardware design, and the energy-efficiency–accuracy trade-off.The paper notes that complex AI models may consume more energy than traditional methods.
Loading 2101.01588v2…