A Vision of Self-Evolving Network Management for Future Intelligent Vertical HetNet
Tasneem Darwish, Gunes Karabulut Kurt, Halim Yanikomeroglu, Gamini Senarath, Peiying Zhu
11 A Vision of Self-Evolving Network Managementfor Future Intelligent Vertical HetNet
Tasneem Darwish, Gunes Karabulut Kurt, Halim Yanikomeroglu, Gamini Senarath, and Peiying Zhu
Abstract —Future integrated terrestrial-aerial-satellite net-works will have to exhibit some unprecedented characteristicsfor the provision of both communications and computationservices, and security for a tremendous number of deviceswith very broad and demanding requirements in an almost-ubiquitous manner. Although 3GPP introduced the concept ofself-organization networks (SONs) in 4G and 5G documents toautomate network management, even this progressive conceptwill face several challenges as it may not be sufficiently agile incoping with the immense levels of complexity, heterogeneity, andmobility in the envisioned beyond-5G integrated networks. Inthe presented vision, we discuss how future integrated networkscan be intelligently and autonomously managed to efficientlyutilize resources, reduce operational costs, and achieve the targetedQuality of Experience (QoE) . We introduce the novel conceptof ”self-evolving networks (SENs)” framework, which utilizesartificial intelligence, enabled by machine learning (ML) algo-rithms, to make future integrated networks fully intelligent andautomated with respect to the provision, adaptation, optimization,and management aspects of networking, communications, andcomputation. To envisage the concept of SEN in future integratednetworks, we use the Intelligent Vertical Heterogeneous Network(I-VHetNet) architecture as our reference. The paper discussesfive prominent communications and computation scenarios whereSEN plays the main role in providing automated networkmanagement. Numerical results provide an insight on how theSEN framework improves the performance of future integratednetworks. The paper presents the leading enablers and examinesthe challenges associated with the application of SEN concept infuture integrated networks.
I. I
NTRODUCTION
To address the ever-increasing user demands of extremelyhigh data rates with extremely low latency in an almost-ubiquitous manner, it has become essential to envision newways of integrating terrestrial, aerial, and satellite networks.In the coming years, networks will undergo an unprecedentedtransformation that will make them substantially different fromprevious generations. This will require a radical paradigmshift in the way networks and services are managed andcontrolled. There is an emerging need to handle the increasein the overall complexity resulting from the transformation ofnetworks into programmable, software-driven, service-basedand holistically managed architectures, and the unprecedentedagility and mobility, where users and base stations (BSs) willbe mobile in three dimensions [1], [2].Conventional model-based networks’ control and manage-ment solutions will not be suitable for these networks. Math-ematical models are not dynamic and can not adapt to thevariable environments of future networks. There is no singlemodel that can cover all possible scenarios. In this paper, we provide a vision on how future networkscan be managed in order to efficiently utilise resources,reduce operational cost, and achieve the coveted QoE . Wepropose the self-evolving networks (SENs) framework thatutilizes data-driven intelligent real-time control to automatenetwork management. For 4G and 5G, 3GPP proposed theconcept of self-organizing network (SON) (Rel. 8-14) andmachine learning (ML) empowered SON (Rel. 15-16), whichfocused on the autonomous configuration, optimization, andhealing of an existing network that has a predefined setof radio resources. However, future networks will requirea paradigm shift from classical SON, whereby the networkadapts its functions to specific environment states, into a self-evolving network that can maintain its performance, underhighly dynamic and complex environments. SEN will drivethe network management from self-organization to continuousand automatic evolvement and be able to automatically reactto unknown environments and triggers, requiring self-adaptiveand resilient learning mechanisms. Unlike SON, SEN will beable to self-manage a network of networks that spans acrossmultiple operators and ecosystems (e.g., satellite, aerial, andterrestrial networks). In addition, SEN will resolve conflictsand manage the coordination among the several entities inthe future integrated networks. Moreover, SEN will considerthe provision, optimization, and management of both com-munication and computational resources. Table I provides acomparison between SON and SEN concepts.Artificial intelligence, enabled by ML algorithms, willwork as the core of SEN engine and will be powered byboth the communications environment’s collected data (e.g.,spatial and temporal traffic distributions, user preferences,and mobility patterns) and external sources improvementssuch as novel technologies, emerging network components,and advanced communication services. We expect that MLwill be effective in learning from experience and detectingchanges. Thus, with such knowledge and self-awareness, con-tinuous, intelligent, and automated decision-making can bemade to evolve the network. For example, intelligent decisionscan be made to inject more communications/computationresources/components, add services, or exploit technologyadvances in the network when there are expected demandsfor extra high data rates, ultra low latency edge computing, ornew emerging applications. Fundamentally, SEN will supportcommunication networks through intelligent and automatedmanagement, and communication networks will support theself-awareness, and the distributed and collaborative comput-ing in SEN.We utilize the intelligent vertical heterogeneous network (I- a r X i v : . [ c s . N I] S e p TABLE I: A comparison of SON and SEN characteristics
Characteristic SON SENScope
Operates in designatednetwork or operator. Across networks, opera-tors, and ecosystems.
Dimensions
Communication and con-trol. Communication, comput-ing, control, caching, se-curity & privacy.
AI/ML im-plementation
Centralized or partiallydistributed (within a net-work or ecosystem). Fully distributed acrossnetworks, operators, andecosystems.
Intelligencedeployment
Adds intelligence at thenetwork edge and core Integrates intelligence intothe fabric of future VHet-Net.
Coordinationand conflictmanagement
Lacks coordination andconflict avoidance amongautonomic managementfunctions of a SON Ensures conflict-freeand coordinated inter-working of multipleautonomic functionsand multiple SONs thatoperate simultaneously inthe same or interactingnetworks, operators, andecosystems.
Level of se-curity & pri-vacy
Security & privacy can bemanaged within a networkor operator domain. Security & privacywill be managed acrossnetworks, operators, andecosystems.
DistributedAI/ML
Intelligence is appliedin centralised or semi-centralised manner. AI/ML application canbe distributed acrossnetworks, operators, andecosystems, where edgecomputing and UEsresources are used aswell.
VHetNet) architecture, which is an extension of the work in[1], as a reference architecture to reflect the concept of SENon future integrated networks. The I-VHetNet architecturejointly performs communications and computation tasks byfully utilizing satellite, aerial, and terrestrial networks. Figure3 shows the SEN framework with I-VHetNet.The contributions of this paper are as follows: • We provide a vision of the SEN management frame-work, which is necessary to automate the managementof services and network operation of future integratednetworks. • Based on the concept of SEN framework and the I-VHetNet architecture, we introduce five scenarios whereSEN plays a vital role in providing automated networkmanagement. • We present simulation results for a simple yet non trivialsystem to compare the performance of SON and SEN fordata offloading and computing in future networks. • The main enablers and challenges associated with theapplication of SEN in future networks are highlighted.In the next section, we introduce the concept of SENs anddiscuss how it differs from SONs. Section III, presents anoverview of I-VHetNet. In Section IV, SENs enablers arediscussed, and Section V presents the envisioned five sce-narios, where SEN framework plays a vital role in providingautomated network management along with numerical results.Several critical issues and challenges are discussed in SectionVI, and our conclusions are presented in Section VII. II. S
ELF -E VOLVING N ETWORKS
Future integrated networks will provide immense heteroge-neous communication and computation resources. However,it will be almost impossible to achieve the coveted keyperformance indicators (KPIs) without intelligent and fullyautomated management of network services and operation.The adoption of manual or semi-manual network operationand management approaches in the evolved future communi-cation networks will result in failing to achieve the requiredKPIs even though extensive communication and computationresources are available.Recently, some ideas have emerged that call for the utiliza-tion of AI/ML in managing services and network operation.The knowledge defined network concept introduced in [3]utilizes Software-Defined Networking (SDN) and NetworkAnalytics to facilitate the adoption of AI techniques in thecontext of network operation and control. The ETSI Zero-touch network and Service Management (ZSM) group isformed with the goal to accelerate the definition of the requiredarchitecture and solutions in order to achieve full end-to-endautomation of network and service management in the contextof 5G. 3GPP has discussed the SON concept in Release8 (LTE) and in all subsequent releases. SON refers to theautomation of communications networks and the minimizationof human intervention in the network management process[4]. SON provides the capabilities of self-configuration atthe network deployment phase, self-optimization of networkparameters, and self-healing to prevent/detect/correct networkfailures. SON presents significant limitations relative to thechallenges facing future networks. The challenges are sum-marized in Figure 1. Table I provides a comparison betweenthe characteristics of SON and SEN.To fulfill the requirements of future integrated networks,network management should go beyond the concept of execut-ing the pre-defined network management functions and shouldbe able to automatically react to unknown environments andtriggers. In SEN, network operation and service managementautomation will evolve through time. The network will notonly learn the new environment but it will also be able tolearn how to learn and what to learn.The concept of SENs implements multi-level intelligent net-work management policies, which can perform across differentdomains, networks, operators, and even ecosystems (e.g.,cellular or satellites ecosystems). SEN concept is supportedby advances in ML (e.g., federated learning, online learning,continual learning), the availability of edge and distributedcollaborative computing, the agility and mobility of networkresources, and the softwarization of network resource man-agement. The individual network entities (microscopic level)of a SEN interact locally with each other in a distributedpeer-to-peer fashion resulting in the evolving structure andfunctionality of the overall SEN system (macroscopic level).Designing individual entities’ simple behaviours that willresult in the sophisticated organization and high performanceof the overall SEN is critical. SEN embodies a special entitythat resolves conflicts and manages the coordination amongthe individual network entities. The SEN’s conflict avoidance
Challenges facing future networksVery high complexity associated with the full integrationof satellite, aerial, and terrestrial networks, nodeheterogeneity, agility, and 3D mobility of users and BSs.The need to support tremendous traffic to/from -trillions of UEs (user equipments) including IoT devices.The need to improve users' quality-of-experience (QoE)by enabling Tbps speeds, and reduced latency.The need to intelligently virtualize and manageresources dynamically. The integration of user device capabilities in networkcommunication or computation aspects.The requirement to enable computational and cachingservices at different levels of the network (i.e. cloudcomputing, fog/edge computing, user devices clusters).
Fig. 1: Challenges facing future networks.and coordination management entity works on multi-levelsspanning from single network domain to multiple operatorsand ecosystems.Figure 2 shows the evolution engine of SEN. The enginecycle starts by collecting data, such as network status, datatraffic, and mobility patterns, through users and networkdevices, sensors, and external sources (e.g., news and weatherforecast). The massive collected data may go through somepre-processing procedures (e.g., cleaning, reductions, transfor-mations). Different types of ML algorithms can be applied.SEN engine’s core utilizes both special pre-designed MLmodels that can be obtained from the SEN repository andadapted models that can be modified on the spot to meetnew requirements. In the dynamic environment of wirelessnetworks, fast online learning algorithms executed at thenetwork edge and distributed among UEs will be necessaryto provide fast intelligent and adaptive responses to delay-sensitive applications. Offline ML is important for predictionand planning purposes. Afterwards, the selected/designed MLmodel can be used to make automated and intelligent deci-sions, such as automatically allocating or retrieving resources,ordering a UAV-BS, forming a new temporary network, ad-justing beamforming parameters, preparing for handoff, andoffloading computations to fog/cloud computing. Moreover,a decision can be made to use a development tool fromthe repository (e.g., a special ML model, a coordinationscheme, or a network component). Periodically, the networkperformance and user satisfaction need to be measured thenthe network can intelligently decide to perform a new cycleof self-evolution.To cope with changes in the network environment, SENhas a development repository of new adopted technologies,agile and mobile resources, coordination schemes, networkcomponents, services, intelligent decision-making models, andspecially designed ML algorithms. SEN performs a continuousexamination, scanning, assessing and predicting changes in application/service requirements, users’ needs, and networkstatus. When a need or a change is detected, SEN selectivelyobtains the suitable development tool from the repositoryand adaptively exploits it to meet the variable requirements.Basically, SEN has the characteristic of ”autonomous drivingnetworks”.The distributed interaction of SEN entities eliminates theeffect of single point of failure, and the system can repair orcorrect damages without external help. The combination of theadaptability of SENs with their distributed nature presents twomajor advantages: robustness against failure and scalability.The continuous evolution of SENs increases the reliability ofthe network. SENs can provide end-to-end network automa-tion that is not limited to optimizing network configurationparameters, but can reach the level of automatically forminga temporary communications network (i.e., through mobileand agile BSs) to fulfil the demands of a specific area for acertain time. Through ML, SENs can enable automated SDNreprogramming, network function virtualization (NFV), anddynamic network slicing (NS) to match highly variable userdemands and efficiently utilize resources.III. O
VERVIEW OF
I-VH ET N ET ARCHITECTURE
The I-VHetNet architecture, shown in Figure 3, not onlyintegrates the terrestrial-aerial-satellite networks, but it alsoincorporates intelligence and provides a computation andcaching platform to enable multi-level edge computing. Infuture networks, BSs might not only be for communications,they might also be utilized for computation or storage. Thedistributed computing resources in I-VHetNet facilitate theapplication of ML algorithms. The SEN framework supportsthe intelligence and computational dimensions of I-VHetNet.
A. Game-Changing Components of I-VHetNeta) UAV as a BS:
UAV-BS concept was mentioned in3GPP Release 17 documents TS22.125 and TS22.261 [5].Under time and energy constraints of fast UAV-BS deploy-ment, several tangled complex decisions must be made veryfast, including load balancing, radio resource management,route management, and beamforming. An intelligent and au-tomated management approach is necessary to enable theself-deployment of UAV-BSs, handle their fast mobility andhandoffs, and manage the connections of the hundreds orpossibly thousands of users that are served by the UAV-BS. b) UAV as a User Equipment (UE):
The use of UAV-UEs (e.g., cargo drones) is already supported through existingterrestrial networks [6]. The main future challenge is thescalability of existing solutions when the number of UAV-UEs reaches into the millions. Mobility management in a 3Dscenario for such a huge number of UAVs will certainly be abig challenge. It is expected that SEN will play an importantrole in these setups through automating and optimizing theprocesses of radio resource management, and mobility man-agement across networks and operators.
Learning training &ML modeldesign Intelligent & automateddecision-makingEvaluateperformance
Collect data (environment sensing & external sources)
DatacleaningDatareductionUnsupervised learning Reinforcementlearning
ModelEvaluationMetrics Designspecial MLmodels
Cross-validation evaluation
Increase capacity using UAV BSAdjust beamforming offloadcomputationto fog Prepare for handoffData rateQoE DelayDecisionaccuracyenergy-efficiencySensorscontrolsystemsAntennas userdevices
Mobile client vehiclesDeep learningSupervised learning Testingdata
Update ML modelsin repository Utilize UEs incommunication& computation
IntelligentdecisionsObtain networks orservices from otheroperators Selectfromrepository special MLmodels IntelligentoptimizationNewtechnologiesAllocateresources
Serviceusage toresourceusage map Diagnosefaults UAV UEs/ BSsExternal sources
Fig. 2: The evolutionary cycle of SEN’s evolution engine. c) High Altitude Platform Station (HAPS) Systems:
AHAPS will be a principal quasi-stationary network element inthe aerial network. Current HAPS deployments target an alti-tude of 18-21 km with a coverage of a radius 50-100 km [7].The HAPS has emerged as a viable aerial network componentdue to the evolution in communications technologies and theadvances in solar panel efficiency, lightweight composite mate-rials, autonomous avionics, and antennas [8]. With free-spaceoptical (FSO) secure communications, several HAPS systemscan form a powerful backbone network and enable an ultralow latency backhaul connectivity for UAVs and various aerialnetwork elements. Manual or semi-automated management insuch a complicated system will limit its capabilities, waste itsresources, and increase its operational costs. Therefore, SENautomated management is vital for HAPS systems. d) LEO Satellites:
In the near future an immense numberof low earth orbit (LEO) satellites are going to be orbitingthe earth to provide global connectivity and Internet accessto users everywhere. As a key player in our SEN I-VHetNet,satellite networks are self-controlled and self-managed withautomated decision-making capabilities. This can be achievedby incorporating the SEN automated management in LEOsatellite networks.
B. Distinctive Characteristics of I-VHetNeta) Intelligence Dimension:
Intelligence is necessary tosupport the cocept of SEN in I-VHetNet. The highest andmost powerful level of intelligence is executed in the cloudcomputing centers located in the core network. The middlelevel is at the edge/fog computing facilities near the users,
Backhaul B a c k h a u l A d d i t i o n a l C a p a c i t y / C o m p u t a t i o n B e a m D M a ss i v e M I M O bea m Cargodrones
Sparsely populated spaces
3D Massive MIMO beams
HAPSnetwork
Aerial Network
LEO
Satellite Network
GEO MEO
Ground Station
Core Network S e l f- ev o l v i ng n e t w o r k ( SE N ) f r a m e w o r k Free Space Optical LinkRF LinkFiber Optic Communication
Satellite SON
Drone Network
AI/MLBig dataanalyticsAI/MLBig dataanalyticsAI/MLBig dataanalytics Ege/ FogComputing
Aerial SON
Ege/ FogComputing
TerrestrialSON
Ege/ FogComputing
Distributed&CollaborativeComputing
Conflictavoidance&Coordinationmanagemententity
Communication for IntelligenceIntelligence for Communications
SEN Engine polar region
Fig. 3: Intelligent vertical heterogeneous network (I-VHetNet) architecture with distributed intelligence and computationalcapabilities. The terrestrial layer consists of the conventional BSs. UAVs, flying aircrafts, and high altitude platform station(HAPS) systems are the main components of aerial networks. UAVs can be used either as an aerial BS or as user equipment.where fast and intelligent computation can be done. The thirdlevel is the level of UEs (e.g., autonomous vehicles and cargodrones), where such smart devices can collaborate among eachother and with the edge computing nodes to achieve distributedintelligent learning and decision-making. b) A Group of Self-Evolving Networks with Distributedand Intelligent Decision-Making:
I-VHetNet consists of agroup of SENs that collectively form a large integrated SEN,which can create, organize, control, manage, and sustainitself autonomously by using the evolution engine. This willcreate high adaptability to changes in the network environmentand increase the scalability, robustness, and fault-tolerance.Such networks can be formed horizontally (i.e., within oneof the three integrated layers) or vertically (across multipleintegrated layers). SEN can automate the processes of SDNprogramming, NFV, and NS to make network softwarizationmore dynamic and intelligent to satisfy the needs of a highlyvariable communications environment. c) Multi-Level Computing and Caching:
I-VHetNetprovides computational and caching capabilities at multiplelevels to serve future applications (e.g., augmented reality)that require high computational capabilities. The cloud levelprovides the highest computational power and storage capac-ity. The network-edge level supports delay-sensitive applica- tions through mobile-edge/fog computing. The lowest levelis provided by managing the collaborative computing andresource sharing among UEs. The I-VHetNet computationaland caching dimension not only supports user applications, italso supports the intelligent automation functionality in SENs.On the other hand, the SEN framework can automaticallymanage and self-allocate the required communications andcomputational resources to fulfil the constantly changing userdemands. With the distributed computational resources in I-VHetNet, implementing distributed ML algorithms will bepractical and feasible. d) Dynamic, 3D, and Agile Topology:
I-VHetNet inte-grates terrestrial, aerial, and spatial networks in a 3D topologywhere everything can move including BSs (e.g., UAV &satellite). As I-VHetNet consists of a group of SENs, thiswill allow forming, splitting, and slicing of networks basedon changes in user demands. With the agility and flexibility ofSENs, we do not have to over-engineer or excessively densifythe terrestrial network to provide high throughput rates ortemporary coverage, which are necessary only for a short time(often unpredictable). SEN evolving characteristic can play asignificant role in managing the I-VHetNet’s resources andtopology. e) Seamless Connectivity Anywhere, Anytime, and forEverything:
I-VHetNet architecture extends the coverage ofcommunications networks not only to the entire globe butalso to the surrounding air and space. Through the SENframework, seamless connectivity can be achieved by realizingfull coverage and required communication capacity with opti-mized mobility management. In particular, satellites and HAPSsystems will play a significant role in solving the problemof coverage in areas where there is no network infrastructure(e.g., rural areas, offshore platforms, ships, submarines). InSEN, the mobility of resources and users can be intelligentlymanaged to achieve seamless connectivity anywhere, anytime,and for everything.IV. S
ELF -E VOLVEMENT E NABLERS IN FUTURE NETWORKS
A. Massive Volume of Data
Massive volumes of data will be generated from sensors,surveillance cameras, smart gadgets, vehicles, UAVs, HAPSsystems, and satellites. The collected data is the preciousfuel of data analytics and ML algorithms. Such data can beused to reveal trends, hidden patterns, unseen correlations,and achieve automated decision making. It can also be usedto continuously learn about user behaviour and enable thenetwork to proactively adapt to changes in the communicationsenvironment. However, data anonymization techniques areessential to maintain user privacy.
B. Softwarization Paving the Way for Intelligence
Softwarization is expected to bring the benefits of pro-grammability into network management and control [9]. How-ever, by using intelligent decisions obtained through ML algo-rithms to apply network softwarization, this moves the networkcontrol and management to the intelligentization dimension.For example, intelligent SDN can be redefined automaticallyand dynamically on the basis of intelligent decisions to adaptto changes in the communications environment. Similarly,NS can be done on the basis of future demands predictions.Moreover, the process of defining or programming the networkusing SDN, NS, or NFV can be automated through MLalgorithms.
C. ML Science Advances
Currently, there are a number of powerful ML algorithms,such as deep neural network and reinforcement learning, whichresemble the human brain learning process of trial and error.In addition, new ML algorithms are emerging such as metalearning and continual learning, where a dynamic ML modelcan be modified and adapted through re-configuring someparameters. Moreover, research is progressing on the conceptsof learning how to learn and what to learn.For resource-limited UEs, some simplified novel ML algo-rithms (e.g., compressed deep neural networks learning) havebeen proposed. FastGRNN and FastRNN are algorithms toimplement recurrent neural networks (RNNs), and gated RNNsinto tiny devices [10].
D. Edge and Fog Computing Capabilities
The computational power of edge/fog computing can beused to execute ML algorithms on behalf of resource-limiteddevices such as smart phones or sensors [11]. The SEN frame-work’s distributed and collaborative computing component canutilize the edge computing resources to realize the function-alities of SEN. In SEN, communications enable distributedintelligence and intelligence improves communications perfor-mance. Offloading computations to the network edge has manyadvantages. First, data and computations can be processedlocally which reduces the response delay and enables real-time data-driven applications [12]. Second, offloading to theedge reduces traffic and congestion towards cloud data centers.Third, edge/fog computing supports mobility-aware applica-tions as it considers user mobility. Fourth, as data do not haveto travel through many nodes in the network, user privacy anddata security are more preserved. Edge/fog computing can beperformed at different levels of computational capabilities andlocality by dedicated nodes or in a collaborative way throughexploiting the aggregated UEs’ computational capabilities anddistributed intelligence.
E. Collaborative Computing and Distributed ML
In big data centers, complex ML jobs are divided intosmall tasks that are executed in parallel on multiple virtualor physical machines. This makes the idea of collaborativecomputing [13] feasible by distributing the tasks of ML amonga group of collaborating fog nodes or UEs. As a leadingalternative to centralized ML algorithms, federated learningtechniques can provide a platform to achieve distributed MLwith high prediction accuracy in a privacy-preserving manner[14]. To realize the concept of distributed intelligence inSENs, the edge computing and the aggregated computationalresources of UEs can be utilized to form a ”collaborativeedge/fog cluster”, where data can be shared under privacy andsecurity preserving techniques (e.g., federated learning). Sincethe resources in such clusters are distributed, reliable and fastcommunications among the participants is crucial.V. E
NVISIONED SCENARIOS OF
SEN
BASED ON
I-VH ET N ET ARCHITECTURE
In this section, we present some scenarios to explain howthe SEN framework can support I-VHetNet.
A. Intelligent Network Selection
Through SEN framework high load balancing across net-works and BSs can be achieved by learning user demandsand mobility patterns. SEN can make automated decisionsof choosing the optimal serving cells, network, or even thenetwork components and architectures that match the requiredservices loading situation (e.g., choosing a flat architecture,an edge/cloud server, and a suitable core network parts,etc). Using the network self-awareness information obtainedthrough continuous learning of network’s status, I-VHetNetcan utilize the SEN functionality to cluster users based ontheir required QoE, mobility patterns, and device capabilities,
HAPSnetwork
Aerial Network
LEO
Satellite Network
Ground Station
RF Link
UAV-BS Networks
User AUser B C a r go - d r on e UA V - U E Network options (a) A dd i t i on a l C a p ac i t y B ea m HAPSnetwork
Aerial Network
UAV-BS Networks
Extended network coverage(temporary network)
LEO
Satellite Network
RF Link
Rural area
Capacity beam (b)
HAPSnetwork
Aerial Network
LEO
Satellite Network
Ground Station
RF Link
UAV-BS Networks
BeamPrevious Beam AB Vehicle movement direction (c) Aerial Network
Communication and offloading through satellite networks.Communication and offloading through HAPS systemsData center (2)
LEO
Satellite Network
Data center (1)
HAPSnetwork
Data center (3)Communication and offloading in terrestrial networks (d)
Aerial Network
RF Link
UAV-BS Networks
Cargo-droneUAV-UE Communication within data centerAerial data center Big data centerCommunication between aerial and ground data centerComputation offloading and data collection at aerial data center
LEO
Satellite Network (e)
Fig. 4: (a) Intelligent network selection for users A and B. As UEs might not be able to carry out such complicated intelligentdecision-making procedure, SEN framework can autonomously make such a decision while considering all available networks,operators, and even ecosystems. (b) Based on intelligent prediction/detection, SEN framework can make automated decisionsto extend network capacity by a HAPS (left), and extend network coverage by a temporary network of UAV-BSs (right). Suchextensions can be achieved in a timely manner and for a certain duration in order to intelligently manage network resources. (c)Intelligent and dynamic beamforming for mobile network entities, where SEN framework can automatically adjust the formedbeam through different technologies, domains, networks, and operators to adapt to the mobility of the user. (d) Distributeddata offloading and computing towards three data centres with different computing capabilities and are accessible through oneof the integrated networks (terrestrial, aerial, and satellites). In this scenario, SEN framework plays the role of managing theautomated and optimized distribution based on network conditions and data sizes in order to minimize the overall offloadingand computing delays while utilizing network resources efficiently. (e) A HAPS network as an aerial mini data center, whereSEN will provide the automated management of the mini data centre and coordinate data offloading and processing with theground data centre. (a) (b)
Fig. 5: (a) Data offloading and computing delays for 300 users. The highest delays are when offloading the data for computing inthe farthest data centre (3) which is reachable through satellites only. In the integrated networks of I-VHetNet, SEN frameworkoptimally distributes the data and computation offloading among the three data centres and achieves the lowest delays (I-VHetNet SEN). (b) Data offloading and computing delays for 3000 users. Compared to Fig. 5 (a), it is obvious that theterrestrial data centre (1) is overloaded, which results in extra delays (even higher than data centre (2) that is reachable throughHAPS). This is because SON cannot optimally utilize the resources across different networks and ecosystems. Although thenumber of users has increased by a factor of 10, SEN manages to achieve the lowest delays.and then intelligently make decisions to provide the bestcommunications network selection. SEN can play significantrole in optimizing the handoff target and timing to guaranteeseamless connection. Figure 4a illustrates this scenario.
B. Extending Network Capacity/Coverage
I-VHetNet consists of a group of SENs. The massivecollected network data can be utilized to predict future networkevents, whereby proactive actions can be performed to avoiddelays or network failures. In addition, with advanced ML,SENs have the ability to learn new environments and deal withunprecedented network situations. Thus, the network has theability to extend its coverage or capacity by using mobile BSs,splitting itself into two or more networks, merging with othernetworks, and/or forming a new network. For example, withdeep learning, the spatial and temporal wireless traffic patternscan be used to match the network’s capacity to user demandsby establishing a new temporary network through collaborativemobile BSs without over-engineering the network’s physicalresources, as shown in Figure 4b.
C. Intelligent Beamforming
In I-VHetNet, communications will depend heavily onbeamforming in order to mitigate interference [15]. The SEN framework can be used to optimize beamforming parametersand achieve dynamic beamsteering across different networks,and operators. As a SEN, multiple coordinated HAPS systemsor UAV-BSs in I-VHetNet, which are equipped with multi-antenna arrays, can form distributed MIMO-network. Throughintelligent and distributed control, and with conflict avoidanceentity of SEN extremely precise beams can be created thatcan track the user mobility while limiting the interference, asdepicted in Figure 4c. To accurately change the beam con-figuration based on changes in the communications networkand user mobility, reinforcement learning or continual learningapproaches can help incorporate intelligence in sequentialdecision-making processes such as these.
D. Distributed Data Offloading and Computation throughIntegrated Networks
Through the intelligent classification of the various require-ments of different services, the SEN framework can optimallymake decisions to distributed data offloading and computingacross several networks, operators, and even ecosystems. Forexample, Figure 4d describes the scenario of having three datacentres, where the first one (1) has the lowest computationalcapability and it is accessible through terrestrial networks, thesecond one (2) has better computational capabilities but it is accessible through HAPS, and the third one (3) has the highestcomputational capabilities but it is accessible by satellites only.By employing the SEN framework, an optimal decision canbe made to distribute the data offloading and computationacross the three data centres, which are accessible throughthree different networks or ecosystems (e.g., terrestrial, aerial,satellites), taking into account the delays of both communi-cation and computation. Figure 5 compares the performanceof SON and SEN in the data offloading and computingscenario, where SON concept is applied independently ineach of terrestrial, aerial, and satellite networks while theSEN functionality can be across the integrated I-VHetNet.The simulation results show the significant SEN frameworkperformance improvement in comparison to SON.
E. Aerial Mini Data Centers
HAPS systems equipped with powerful processors andconnected with high-speed FSO links may collectively form amini aerial data center and be an aerial network intelligenceenabler, as illustrated in Figure 4e. Through the distributed andcollaborative computing entity of SEN framework, the self-managed aerial data centers can provide near-user computationservices for aerial networks users (e.g., drones) by allowingaerial network elements with limited resources to offload theirintelligent algorithms computations. Analyzing data in the skywill reduce response delays and decrease the burden on theaerial-to-ground communications links, which can be easilydisrupted by the fast speed of aerial network elements. Dueto its relatively large footprint, a HAPS can collect data fromlarge portions of the aerial network to use it in supportingthe self-evolution of aerial networks. These data centers canalso provide a backup computational facility in emergencyscenarios.VI. C
HALLENGES AND RESEARCH DIRECTIONS
Incorporating intelligence in future networks and realizingthe concept of SEN faces many challenges that require inten-sive research work. In this section we discuss some of themost important challenges in the following points: • Real-time and online learning algorithms:
Most exist-ing ML algorithms require relatively long convergencetimes. However, in future communication networks theenvironment may change rapidly and many applicationsmay require fast decision-making that adapt to changesin the network. Online and continual learning algorithmsshould be further enhanced and adapted to suit the futureintegrated network environment. • Learning what to learn and how to learn:
In manysituations, a given node in a network does not needto learn the full network environment. Learning withincertain scope and time frame is sufficient in many cases.Choosing an appropriate data scope and duration isimportant to avoid learning and processing unnecessarydata. This issue is quite important for network nodes withhigh mobility or limited processing resources. In currentML applications this is done in a manual way by thedevelopers or engineers. However, to realize the concept of SEN, emerging meta learning, continual learning andthe concept learning how to learn need to be adopted andenhanced to adapt to the dynamic environment. Learninghow to learn and what to learn should be done in anautomated way, which needs intensive research work. • Information sharing and learning across networks,operators, and ecosystems:
To implement the SENframework, new policies are required to control datasharing and collection across different networks, oper-ators domains, and ecosystems. Thus, the networks andcommunications community should invest in developingnovel ML algorithms that are designed for distributedlearning scenarios. It would be a major enhancement toextend federated learning concept to work on differentlevels and scales. Privacy and security is a main concernin this area. In addition, new schemes are needed tohandle the emerging issues such as data ownership rights,data credibility, data trading, and data pricing whilemaintaining users’ privacy and rights. • Intelligent and standardized conflict resolution algo-rithms:
In the environment of SENs, conflicts mightarise among entities with contradicting objectives (e.g.,minimizing delays and sharing resources). In addition,SENs will involve several operators and each one ofthem has the goal of maximizing his own profits, whichmight result in greedy behavior among operators wherethe user might pay the price. Nevertheless, resolving con-flicts among different operators, networks, or ecosystemswith heterogeneous technologies and different operationalpolicies is a very complex task. In this domain, intelli-gent conflict resolution algorithms need to be developedwhich can mimic the human way of thinking in similarsituations. In addition, standardised conflict resolutionalgorithms are also necessary to ensure compatibilitybetween different systems.VII. S
UMMARY
Providing extra resources in future networks without intel-ligent and automated management may not fulfil the require-ments of the emerging applications and expected QoS levels.Efficient integration of terrestrial, aerial, and satellite networksof the future will necessitate intelligent, automated, adaptive,and real-time control, optimization, and management. To thisend, we introduced the self-evolving networks (SENs) frame-work. Empowered with AI/ML, SEN framework can makeintelligent, adaptive, and automated decisions • to manage heterogeneous network dynamically and intel-ligently in a distributed manner across different domains,networks, and ecosystems; • to resolve conflicts and manage coordination amongseveral automated network entities; • to satisfy the QoE requirements of an enormous numberof a broad range of UEs (including IoT devices); • to handle the high levels of heterogeneity, agility, and 3Dmobility of both UEs and BSs; and • to utilize the UE assets (integrated with those of thenetwork) in the provision of communications and com-putation services. We constructed five prominent I-VHetNet communications andcomputation scenarios, which model several aspects of SENs.In addition, we listed the leading enablers and the associatedchallenges. R
EFERENCES[1] M. Alzenad and H. Yanikomeroglu, “Coverage and rate analysis forvertical heterogeneous networks (VHetNets),”
IEEE Transactions onWireless Communications , vol. 18, no. 12, pp. 5643–5657, Dec. 2019.[2] L. Zhang and N. Ansari, “A framework for 5g networks with in-band full-duplex enabled drone-mounted base-stations,”
IEEE WirelessCommunications , vol. 26, no. 5, pp. 121–127, Feb. 2019.[3] A. Mestres, A. Rodriguez-Natal, J. Carner, P. Barlet-Ros, E. Alarc´on,M. Sol´e, V. Munt´es-Mulero, D. Meyer, S. Barkai, M. J. Hibbett et al. , “Knowledge-defined networking,”
ACM SIGCOMM ComputerCommunication Review , vol. 47, no. 3, pp. 2–10, 2017.[4] J. Moysen and L. Giupponi, “From 4G to 5G: Self-organized networkmanagement meets machine learning,”
Computer Communications
IEEE Wireless Communi-cations , vol. 26, no. 1, pp. 18–27, Feb. 2019.[7] T. Tozer and D. Grace, “High-altitude platforms for wireless commu-nications,”
Electronics & Communication Engineering Journal , vol. 13,no. 3, pp. 127–137, Jun. 2001.[8] G. Kurt, M. G. Khoshkholgh, S. Alfattani, A. Ibrahim, T. S. Darwish,M. S. Alam, H. Yanikomeroglu, and A. Yongacoglu, “A vision andframework for the high altitude platform station (HAPS) networks ofthe future,” arXiv preprint arXiv:2007.15088 , 2020.[9] K. Wang, Y. Wang, D. Zeng, and S. Guo, “An SDN-based architecturefor next-generation wireless networks,”
IEEE Wireless Communications ,vol. 24, no. 1, pp. 25–31, Feb. 2017.[10] A. Kusupati, M. Singh, K. Bhatia, A. Kumar, P. Jain, and M. Varma,“FastGRNN: A fast, accurate, stable and tiny kilobyte sized gatedrecurrent neural network,” in
Advances in Neural Information ProcessingSystems (NIPS) , 2018, pp. 9017–9028.[11] T. S. J. Darwish and K. A. Bakar, “Fog based intelligent transportationbig data analytics in the Internet of vehicles environment: motivations,architecture, challenges, and critical issues,”
IEEE Access , vol. 6, pp.15 679–15 701, 2018.[12] S. Chen, P. Gong, B. Wang, A. Anpalagan, M. Guizani, and C. Yang,“Edge AI for heterogeneous and massive IoT networks,” in
IEEE 19thInternational Conference on Communication Technology (ICCT) , 2019,pp. 350–355.[13] H. Lee, J. Lee, Y. C. Lee, H. Han, and S. Kang, “Mobile collaborativecomputing on the fly,”
Pervasive and Mobile Computing , vol. 58, pp.1–19, Aug. 2019.[14] Z. Zhao, C. Feng, H. H. Yang, and X. Luo, “Federated-learning-enabled intelligent fog radio access networks: Fundamental theory, keytechniques, and future trends,”
IEEE Wireless Communications , vol. 27,no. 2, pp. 22–28, May 2020.[15] H. Vaezy, M. S. H. Abad, O. Ercetin, H. Yanikomeroglu, M. J. Omidi,and M. M. Naghsh, “Beamforming for maximal coverage in mmwavedrones: A reinforcement learning approach,”