1

Wireless Edge Computing with Latencyand Reliability Guarantees

Mohammed S. Elbamby, Cristina Perfecto, Chen-Feng Liu, Student Member, IEEE,Jihong Park, Sumudu Samarakoon, Xianfu Chen, Member, IEEE and Mehdi Bennis, Senior Member, IEEE

F

Abstract—Edge computing is an emerging concept based on distribut-ing computing, storage, and control services closer to end networknodes. Edge computing lies at the heart of the fifth generation (5G)wireless systems and beyond. While current state-of-the-art networkscommunicate, compute, and process data in a centralized manner (atthe cloud), for latency and compute-centric applications, both radio ac-cess and computational resources must be brought closer to the edge,harnessing the availability of computing and storage-enabled small cellbase stations in proximity to the end devices. Furthermore, the networkinfrastructure must enable a distributed edge decision-making servicethat learns to adapt to the network dynamics with minimal latency andoptimize network deployment and operation accordingly. This article willprovide a fresh look to the concept of edge computing by first discussingthe applications that the network edge must provide, with a specialemphasis on the ensuing challenges in enabling ultra-reliable and low-latency edge computing services for mission-critical applications suchas virtual reality (VR), vehicle-to-everything (V2X), edge artificial intelli-gence (AI), and so forth. Furthermore, several case studies where theedge is key are explored followed by insights and prospect for futurework.

1 INTRODUCTION

THE ever increasing requirements of wireless services inMedia & Entertainment (M&E), as well as in healthcare

and wellbeing demands are transforming the way data iscommunicated and processed. Future networks are antic-ipated to support massive number of connected devicesrequesting a variety of different services such as mobilevideo streaming, virtual and augmented reality (AR/VR),as well as mission-critical applications. Such services requiredata, computation, and storage to be performed more oftenwith ultra-high success rate and minimal latency. Multi-access edge computing (MEC) has emerged as an infras-tructure that enables data processing and storage at thenetwork edge as a means to cut down the latency betweenthe network nodes and the remote servers that typicallyexisted in cloud computing architectures [1]. Instead, edge

Mohammed S. Elbamby, Chen-Feng Liu, Jihong Park, and Sumudu Sama-rakoon are with the Centre for Wireless Communications, University of Oulu,90014 Oulu, Finland (emails: firstname.surname@oulu.fi).Mehdi Bennis is with the Centre for Wireless Communications, Universityof Oulu, 90014 Oulu, Finland, and also with the Department of ComputerScience and Engineering, Kyung Hee University, Seoul 17104, South Korea(e-mail: mehdi.bennis@oulu.).Cristina Perfecto, is with the University of the Basque Country (UPV/EHU),Spain. (email: cristina.perfecto@ehu.eus)Xianfu Chen is with VTT Technical Research Centre of Finland, P.O. Box1100, FI-90571 Oulu, Finland (e-mail: xianfu.chen@vtt.fi).

computing can be provided as a service at the network edgeto minimize the service latency, network complexity, andsave the device nodes’ energy and battery consumption.

Edge networking in cellular systems aims to efficientlyprovide the required connectivity, data access, bandwidth,and computation resources to end devices [2], [3]. Edge basestations in proximity of network users will not only relaycontent from and to the network core, but will help executethe users processing tasks, provide customized content andcomputing services, and control the connectivity and inter-action between coupled network nodes.

In essence, the performance of edge computing is pre-dominantly assessed through two main components, com-munication between the edge server and the end device,and the processing at the edge server. Further, severaloptimization aspects are considered to optimize these twocomponents. Optimizing the communication part can beexplored through wireless bandwidth and power allocation,edge server selection, computation task distribution, tasksplitting, and partial task offloading. For the processing part,computation cycle allocation, task queuing and prioritiza-tion, joint computing, and predictive computing are criticalfactors to optimize the computing efficiency.

The focus of the fifth generation (5G) cellular networkshas shifted from merely increasing the data communicationrate to providing service-specific performance guarantees interms of ultra-reliability and low latency. This shift is fueledby the emergence of new use cases that require genuinesupport to critical and latency-sensitive communication ser-vices. Nonetheless, ultra-reliability and low latency are oftenseen as contradictory requirements [4], compelling the useof distinctive set of tools to be efficiently realized. Yet, theseindividually challenging per se requirements are anticipatedto be met together for networks of diverse topologies andheterogeneous services.

This article discusses the feasibility and potential ofproviding edge computing services with latency and reli-ability guarantees. In particular, it first sheds light on theservices that can be offered from edge computing networks.It follows by looking into how ultra reliable low latencycommunication (URLLC) contributes to and benefits fromedge computing. The article proceeds by presenting selecteduse cases that reflect the interplay between edge computingand URLLC. Finally, the article ends with our concludingremarks and future works.

arX

iv:1

905.

0531

6v1

[cs

.NI]

13

May

201

9

2

2 EDGE COMPUTING SERVICES

Legacy network architectures relied on centrally located andcentrally-controlled servers with high computational andstorage powers to provide on-demand computing to net-work devices [5]. These servers could support high numberof network nodes over a large geographical area. However,the large distance between the cloud computing server andend-user device results in higher service latency. Moreover,the centralized architecture limited the ability to providecontext-aware service, and to preserve the user data privacy.Future wireless networks are evolving towards supportingnew set of applications that require minimal latency andhigh level of service personalization. This motivated theshift towards distributed networking architectures wherethe network resources are available close to users at thenetwork edge. Edge computing aims to provide comput-ing, content, and connectivity services closer to the datasource and consumption points. It is applicable to scenarioswith different network environments and use cases. Thisdiversity led to several implementations that did not followspecific standard or interoperability. The European Telecom-munications Standards Institute (ETSI) has been working onsolving this issue through providing an efficient standard-ized MEC that can be integrated across several applicationsand service providers [6]. MEC also enables providers todeploy edge computing services on top of wireless mobilenetworks. This will allow cellular operators to integratecomputing into the services provided to their users. In thisregard, the term edge networking refers to the action andprocess of serving a user or device at the network edge.

2.1 Content at the Edge

The idea of leveraging the network edge as a contentstorage has gained popularity in the last few years [7].Existing popularity patterns on the contents requested bynetwork users motivated developing proactive networks.A proactive server can predict popular contents, prefetchthem from the core network, and have them stored andreadily available at the network edge, hence cutting downdelivery times once users request them. Proactive networksrequire efficient methods to predict the popularity of thecontent to be cached, as well as high storage capacity tocache this content. Edge caching not only minimizes theservice latency but also the load on the backhaul networkby prefetching the popular content in the off-peak times [8]–[10]. Further, we envision that the notion of edge contentwill be extended to include new types of data that canbe served from the network edge to support the new usecases. One application to which the future network edgewill provide information is the distributed machine learningapplication. The tight latency requirements and the need forminimizing the information exchange mandate the develop-ment of distributed machine intelligence schemes in whichedge servers play a major rule. Edge machine learning[11], [12] will allow end users to locally develop their ownmachine learning models instead of relying on centralizedapproaches. However, ”machine learning applications” relyon information from other network nodes that affect theirstate and utility. The network edge role here will be to bring

the information necessary for enhancing or complementingthe local model close to the user.

2.2 Computing at the EdgeProcessing is becoming as an important commodity tocellular applications as content. The use of applicationsranging from smart factory, self-driving vehicles, to virtualand augmented reality are growing by the day and arebecoming more resource greedy and less latency tolerant.While part of the computing load of these applications isserved using their local processing units, constraints on size,portability, battery life-time or lack of full access to taskdata limit the ability to locally execute computing tasks.Edge computing promises to pool powerful yet proximatecomputing resources at the network edge, as well as toprovide connectivity and seamless information exchangebetween neighboring nodes. It is also set to allow for therealization of various 5G verticals that require low-latencyand high-reliable computing, such as virtual reality (VR)and mission-critical Internet of Things (IoT) applications.Yet, there are several components that need to be addressedto realize low-latency and high-reliable edge computing.Executing computing tasks at the edge often requires thetask data to be offloaded to the edge server before execu-tion. This introduces communication delay that adds to theservice latency. In addition, how to queue and schedule thecomputing tasks at the edge server plays a major role inthe queuing and processing latency. Our vision is that theavailability of more data and computing power will shapehow the edge network performs computing. Similar in veinto proactive content caching, where knowledge of userspreferences and future interests allow for prefeching of theircontent, data availability and machine learning will help tospeed up computing the tasks of network nodes. Predictingvehicles future locations and path allows the edge networkto proactively render and deliver its high-definition livemap. In VR applications, predicting users future field ofview (FoV) allows rendering the corresponding part of its360◦ frame with minimal latency. Several other enablers arevital to achieve ultra-reliable and low-latency computing,such as task replications, parallel, and coded computing,which will be addressed in detail in the following section.

2.3 Control at the EdgeMost of the existing cloud and edge computing architecturesrely on centralized decision-making schemes which requiresall the network nodes to send their local states data to acentral controller. Instead, distributed decision making, inwhich the decision-making process is distributed among theedge servers will allow for low latency, and privacy pre-serving operation [13], which is essential for mission-criticalapplications. Indeed, the control of the network devicesperformance requires policies that adapt to their local states.This can be challenging for scenarios where the local statedynamically varies due to highly dynamic environment ordue to the nature of the application, such as in mission-critical applications. Reinforcement learning (RL) solutionscan provide efficient control policies that maximize thesystem rewards by finding policies that map those dynam-ically changing states into actions. These decision-making

3

URLLC Enablers for

Edge Computing

Control Panel

proximity-based

computing

highcapacity

links(mmWave)

proactivecomputing

edge ML&

federatedlearning

multi-connectivity

taskreplication

paralleland codedcomputing

extremeevent

control

Figure 1. Breakdown of key URLLC enablers for edge computing, exemplified over an Industry 4.0/Smart Factory ecosystem that includes cyber-physical systems, IoT and MEC.

policies need to take into account the effect of actions onthe environment and update the reward accordingly. Incentralized architectures, classical reinforcement learning isoften performed offline, not taking into account reliabilityin decision making for example under noisy feedback. Edgecontrol can provide robust decision-making, where multi-agent RL architectures can be used to provide communi-cations efficient methods that take latency and reliabilityinto account in dynamic and mission-critical environments.Latency stems from the local state exchanges between edgedevices, in which the overhead due to the state exchangeincreases exponentially with the number of devices. This canbe addressed using the mean-field game (MFG) theory [14],which can tackle this by approximating the average state asa collection of agents’ instantaneous states.

3 URLLC ENABLERS AND CHALLENGES

3.1 URLLC overview

The prime focus of the recent groundswell of mission crit-ical applications such as autonomous vehicles, immersiveVR/AR experiences, industrial automation, and robotics,is to provide services with guaranteed high reliability andlow latency. Therein, latency deductions in channel estima-tions, information exchange among the network elements,decision making, computation tasks completion, and mem-ory access within devices have utmost importance. Alongwith them, guaranteed low-latency in operations, ensuringconnectivity, and speed-precision-and-accuracy of computa-tions are essential to assure the reliability of mission criticalapplications. Due to the on-device constraints on storage,processing capability, and availability and accessibility ofnetwork resources, it is mandatory to utilize the edgeservers to maintain the quality-of-service in mission criticalapplications. To support the communication among user

devices within mission critical applications and the edgeservers, URLLC, that has been introduced as one of the mainservice in 5G systems, plays a pivotal role. In this section,we identify the key enablers of reliability and low-latencyin wireless edge computing networks, and the challengestowards realizing each of them. Moreover, in Table 1, wesummarize the issues and enablers of providing latency andreliability guarantees in wireless edge computing networks,as well as the applications and use cases these enablers aretargeting.

3.2 URLLC Enablers for Edge Computing

3.2.1 Low latency Enablers

There are several components that contribute to latency inedge networking. In this regard, enabling low latency re-quires several techniques to be implemented and integratedtogether at different levels of edge networking systems. Atthe communication level, proximity-based computing andmillimeter wave (mmWave) links play major roles in reduc-ing task offloading latency from edge devices to servers byreducing distance attenuation and providing broad band-width with high directionality, respectively. In addition,mmWave also enables wireless backhauling [15], [16] thatfacilitates edge servers’ prefetching popular content withlow latency. At the processing level, proactive computingprovides significant latency reduction while maximizingresource efficiency by avoiding repetitive and redundanton-demand computing [17]–[19]. Next, coded computingis effective in reducing parallel computing latency, whicheliminates the dependency of processing tasks, thereby min-imizing the worst-case latency due to a straggling task. Lastbut not least, machine learning (ML) is crucial in supportinglow-latency mission critical applications, by empowering

4

Demands/Challenges Enablers MEC applications and use cases

Low la-tency

bandwidth, backhauls mmWave extended reality, vehicular edge com-puting (Sec. 4.1.1 and Sec. 4.1.2)

low propagation delay proximity based computing deep reinforcement learning based taskoffloading (Sec. 4.2, use case 4)

computing power, taskdependency

parallel and coded computing [50], [51]

low propagation delay,energy efficiency

proactive computing use case 6 in Sec. 4.2

low prediction delay edge machine learning edge computing for federated learning(use case 1 in Sec. 4.2)

Highrelia-bility

channel intermittency multi-connectivity, task replication use case 6 in Sec. 4.2 and [19], [44]

low communication cost,data privacy

federated learning edge computing for federated learning(use case 4 in Sec. 4.2)

rare event detection extreme event control extreme value theoretic edge comput-ing and vehicular federated learning(use cases 2 and 3 in Sec. 4.2)

Table 1Challenges and enablers of realizing low latency and high reliability in wireless edge computing.

edge servers and devices to locally carry out their decision-making.

Low Latency Enabler 1. High capacity mmWave links:Driven by the spectrum shortage below 6 GHz, communi-cations in the radio frequencies encompassing the electro-magnetic spectrum from 30 to 300 GHz, i.e. the mmWave orInternational Telecommunications Union (ITU)’s extremelyhigh frequency (EHF) band, have been attracting a growingattention [20]–[22], to the point of being currently consid-ered the most important technology to achieve the 10 Gbpspeak data rates foreseen for the upcoming 5G systems [23].Having abundant available spectrum, the main appeal ofmmWave communications comes from the use of generousbandwidths that –ranging from the 0.85GHz in the 28GHzband to 5 GHz in the 73GHz band–are more than ten timesgreater than Long Term Evolution (LTE)’s 20 MHz cellularchannel [24], and grant an important channel capacity in-crease [25].

However, signal propagation at these frequencies isharsh and inherently different from that at the microwaveband [26] experiencing 1) higher pathloss for equal antennagains due to a stronger atmospheric attenuation wherebysignals are more prone to being absorbed by foliage andrain, 2) higher penetration losses as mmWaves are blockedwhen trying to pass through walls, buildings, or obsta-cles, and 3) higher transmit power consumptions than inlower bands to preserve an equal signal-to-noise ratio (SNR)unless directional antennas together with advanced sig-nal processing that includes massive input massive output(MIMO) [27] and beamforming (BF) techniques are used.Notably, due to the shorter wavelengths in mmWave bandsit is possible to pack more antennas at the transmitter andreceiver devices and, thanks to the spatial degrees of free-dom afforded, use analog or hybrid BF –fully digital BF im-plies having one dedicated radio-frequency (RF) chain perantenna which currently discourages its use in mmWaves

due to the unaffordable power consumption and costs– tobuild a radiation pattern with narrow beams which will besubsequently steered towards the receivers while the energyradiated through the sidelobes is minimized or negligible.

To administer high capacity links with mmWaves, trans-mitters’ and receivers’ mainlobes need to be preciselyaligned towards each other if favored with a clear, unob-structed, line-of-sight (LOS) path. In practice, when a mobileuser equipment (MUE) is in the connected state, uplink (UL)control channels are used to periodically feed back to thebase station (BS) its best transmit beam index; similarlydownlink (DL) control channels are used to report MUEs’best transmit beams. Data transmission is then performedthrough the best beam pair. However, during initial accessand handover, i.e. in random access, such information onthe best beams is not available which hinders taking fullbenefit from BF. Henceforth, in analog BF, to discover andthen maintain the best transmit-receive beam pairs, a seriesof techniques referred to as beamtraining or beamsearching,are applied. Then, beam tracking is performed to adaptthe beamforming, e.g., due to MUEs’ movement leadingto transmitter-receiver beam misalignments. Nevertheless,a full new directional channel discovery process will needto be triggered if the signal-to-interference-plus-noise ratio(SINR) drops below a certain threshold due to e.g., block-ages and/or interference [28]. As analog BF employs a singleRF chain, it is challenging to adjust the beam to channelconditions, leading to some performance loss. Moreover,analog BF does not provide multiplexing gains as it canonly operate a single data stream. Therefore, to bring allthe benefits of mmWave while benefiting from multiplexinggains for MEC, MIMO hybrid BF architectures, which strikea balance between performance, complexity, and power con-sumption, should be considered. Finally, as adaptive beam-forming requires precise channel state information (CSI),one of the key challenges for mmWave to work as a low-latency enabler for MEC lies on the availability of expedited

5

CSI acquisition schemes together with directionality-awaremobility and beam management procedures [29].

In the next subsection a series of reliability enablerswill be discussed to reduce the delay incurred to coun-teract the intermittent blockages and temporal disruptionsof the mmWave channel. Largely, these techniques are inline with the idea of overbooking radio resources as aprotection against channel vulnerability [30] or to considerrisk-sensitive approaches [31].

Low Latency Enabler 2. Proximity-based Computing: Re-ducing the distance between the application and the MECserver is a key latency enabler. This idea is motivated by theconcept of bringing the transmitter and the receiver closerto one another yielding capacity improvements [21]. Withthe low proximity between the application and MEC server,over-the-air latency that has a significant contribution to theend-to-end (E2E), sometimes dominating over the comput-ing latency, can be greatly minimized.

Network densification, the concept of dense deploymentof small cells, remote radio units, and relay heads that hasbeen an attractive research interest during recent years [32]–[37], plays a major role in proximity-based computing.While boosting the capacity and coverage, the dense de-ployment of access points offers the opportunity of intro-ducing additional computing resources at the network edge.Henceforth, the user devices in the network are capable ofuploading their computational tasks to access points anddownload the corresponding outputs after the processingwith high data rates yielding lower latencies.

Another proximity-based computing technique is mobil-ity assisted MEC. Therein, networks of connected vehicles,unmanned autonomous vehicle (UAV), and robots withhigh processing power can assist the computational tasksof the users [38], [39]. The high processing power of abovedevices that are dedicated to users provides low computa-tional latencies. Moreover, their flexible connectivity withthe users due to the mobility and high data rates thereindue to the proximity offer lower communication latencies,yielding reduced E2E latencies.

Computing location swapping is another proximity-based computing method. Therein, groups of users coexistin either physical (located close by) or virtual spaces (in-teract and/or share computing tasks). In his regard, prox-imity alone provides low communication latency, yet couldyield poorly utilized computational resources. Combiningthe user groups in virtual space and their physical loca-tions, some users can swap their associated MEC serversto improve both computing and communication latencies,resulting better E2E performance [40].

Although the proximity-based computing enables lowlatency in MEC, the concept itself brings up new challengesto the network design and resource optimization therein.The increased interference is one of the challenges in bothnetwork densification and computing location swapping.Due to the limited availability of both communication andcomputation resource, increased interference may degradeboth uplink and downlink communication yielding in-creased E2E latency [41]. In this regard, interference avoid-ance, management, and mitigation techniques as well as useof higher frequency channels are viable remedies. Another

challenge is the frequent handover due to the dynamicsof environment and user mobility [41], [42]. While han-dover may incur undesirable latencies, the concept of multi-connectivity (MC) can be utilized, in which users receivecomputing assistance from several MEC servers.

Low Latency Enabler 3. Edge Machine Learning: Inference(or prediction) capabilities with low latency is one of themain reason for ML to be popular in MEC as well asseveral other communication applications such as coding,beamforming, resource optimization, caching, scheduling,routing, and security [43]–[46]. While the majority of theML-based communication system design literature is rootedon the centralized and offline ML techniques, the upturnof mission critical applications for massive number of con-nected devices demands for the intelligence at the networkedge [11], [47]. In contrast to conventional centralized MLdesigns, the edge ML is capable of generating inferencewithin an instance at the edge devices, presenting the op-portunity to greatly reduce the E2E latency in MEC appli-cations. Such intelligence at the edge devices can 1) predictthe uncertainties in channel dynamics, communication andcomputation resource availability, interference, and networkcongestion at the local devices; 2) explore and learn aboutthe network environment with minimal additional signal-ing overheads; and 3) characterize and model the networkbehavior in which the system performance is analyzed.At the MEC servers, such prior knowledge provides theopportunities to smartly schedule their computing resourcesand share the results with the corresponding user devices.Furthermore, at the events of connectivity losses, edge MLat the user devices allows the decision making within thedevices using the forecast on system behaviors, allowinguninterrupted end-user service experiences. This ability tooperate offline/off-grid can reduce the number of latency-critical parallel tasks at the MEC server, in which network-wide end user experience is improved.

The challenge of enabling low latency in MEC via edgeML relies on the training latency and inference accuracytherein. In the distributed setting, each edge device lacksthe access to the large global training data set, in whichtraining over local data can degrade the inference accu-racy. To improve the inference accuracy, edge ML devicesmay need often cooperation among one another or witha centralized helper, which incurs additional overheadsand thus, increased training latency. In this regard, furtherinvestigations need to be carried out to optimize the tradeoffbetween training latency and inference accuracy dependingon the design architectures, communication models, andapplication requirement.

Low Latency Enabler 4. Proactive Computing: Althoughedge computing is capable of minimizing the latency in-duced due to the high propagation delay of cloud com-puting, it still experiences delay due to offloading the taskdata to the edge server, processing delay, as well a queuingdelay for both operations. While these delays are inevitablein some cases, there exists situations in which the task hasalready been executed before for another user at a differenttime. Take for example an AR case in which visitors of aspecific spot in an exhibition or museum request a specifictask of augmenting an object to the view of this spot, or

6

the task of object identification by multiple vehicles in anintelligent transportation systems (ITSs) system. Executingthese tasks redundantly each time it is requested is certainlynot resource efficient, and is causing higher delays to thesetasks as well as other tasks sharing these resources. Here,executing and caching the results of these tasks in advance,such that they are served when requested with minimallatency, can be a major latency minimizer.

The ideas of prefetching tasks [48] and proactive com-puting [17], [18] aim to develop techniques that learnsand predicts which tasks are to be requested in the futureand pre-compute them. Indeed, the success of proactivecomputing lies on a well-aimed choice of which tasks toproactively compute and which are to leave for real-timeprocessing. Essentially, this involves developing efficientprediction methods that studies the popularity patterns ofthe computing tasks to decide on which tasks to prefetch.The idea also relies on the availability of storage capabilitiesat the edge servers [49].

Low Latency Enabler 5. Parallel and Coded Computing:The computing task data can be distributed over multipleservers in different edge computing scenarios. For example,in a smart vehicle scenario where the navigation map datacan be partly stored in several edge servers. Parallel execu-tion of computing tasks over multiple servers significantlyimpacts the efficiency and speed of task execution. More-over, it eliminates the need to collect the full task dataset in asingle entity. For example, partial offloading can be performedwhere only a partition of the task is offloaded to where itsrequired input data is available [5]. The implementation ofparallel computing depends on the correlation between thetask partitions, i.e., only partitions that are not dependent oneach other can be executed in parallel, whereas dependenttasks have to be executed sequentially. task dependencygraph models and task partitioning [5], [50] are used totackle the inter-dependency between the different task par-titions.

A challenge in realizing parallel computing, however, isthe resulting high inter-server communication load. More-over, it suffers from the straggling effect, where a missingresult from a single node delays the entire computationprocess. The concept of coded computing has shown toaddress both of these challenges [51]. Through exploitingthe redundancy in the task partitions execution at differ-ent servers, coded multicast messages, e.g. via maximumdistance separable (MDS) codes, can be used to deliver theresults of the missing partitions simultaneously to multipleservers. This approach significantly reduced the amountof data that has to be communicated between the servers,at the expense of more redundant task executions at eachserver. Coded computing also helps in minimizing the over-all computing latency through minimum latency codes. Inconventional parallel computing task, each server executesa partition of the task and returns its result to the client. Inthis model, one delayed or failed partition will cause a delayor failure to the entire task. Alternatively, by generatingredundant task data that are coded combinations of theoriginal task data and executing these coded tasks, the resultcan be recovered by decoding the data from only a subsetof the servers, eliminating the effect of a delayed or failed

result. Optimizing the creation of the redundant coded tasksenables an inverse linear trade-off between the computinglatency and computing load [52].

3.2.2 High Reliability Enablers

For MEC to fulfill its role and run applications on devicesbehalf, i.e. offloading the computing, it needs to be ableto operate below stringent latency values, which are un-achievable in traditional mobile cloud computing (MCC)systems or too demanding to be run locally due to excessivecomputational and communication power

In this regard, to exploit both the high capacity of 5Gmobile connections and the extensive computing capabili-ties located at the edge cloud, the concept of reliability isintroduced with a two-fold interpretation: In the first place,we find the classical notion of reliability related to error-robustness guarantees. As such, it allows to be tackled atdifferent layers, including the reliability of the wireless linkat the physical layer (PHY). Another fundamental notionof reliability, that has been widely adopted for wirelesscommunications and standardization bodies as the ThirdPartnership Project (3GPP), is that of reliability understoodas a probabilistic bound over the latency.

Understood in its most classical form, it is commonthat a toll in return for ensuring high reliability will haveto be paid in the form additional/increased delays. Forinstance, at the PHY layer the use of parity, redundancy,and re-transmission will increase the latency. Also, in multi-user environments allocating multiple sources to a singleuser while clearly beneficial at an individual level, couldpotentially impact the experienced latency of the remainingusers.

Next, we will set forth some of the enablers for bothnotions of reliability.

High Reliability Enabler 1. Multi-Connectivity: Comparedto wired transmissions, in wireless environments temporaryoutages are common due to impairments in the SINR. Theseoriginate from, among others, stochasticity of the wirelesschannels, fluctuating levels of interference, or mobility of theMUEs. The term multi-connectivity (MC) [53] encompassesseveral techniques developed with the overarching aim ofenhancing effective data rates and the mobility robustness,i.e. the reliability, of wireless links. For that purpose, MCexploits different forms of diversity to cut down on the num-ber of failed handovers, dropped connections and, generallyspeaking, radio-link failures (RLFs) that might cause serviceinterruptions [54], [55].

MC solutions are classified as intra or inter frequency,i.e., depending on whether they operate using the same fre-quency or, otherwise, combine multiple carrier frequencies.Examples of the former include coordinated multi-point(CoMP) [56] transmissions and single frequency networks(SFNs) [57]. CoMP involves a set of techniques that exploitrather than mitigate inter-cell interference (ICI) to improvethe performance at the cell edge. On performing joint pro-cessing, dynamic point selection (JP/DPS) or coordinatedscheduling and beamforming (CS/CB) in the UL/DL, BSseffectively operate as if assembled in a distributed multipleantenna system. SFNs embody a form of synchronous mul-ticell transmission whereby various sources use the same

7

time and frequency resource to non-coherently transmitsignals to a receiver. The multiple received copies will bethen constructively combined if their propagation delays aretightly bounded or, else, will induce inter-symbol interfer-ence (ISI) [58].

As for inter-frequency MC, carrier aggregation (CA) [59]and dual connectivity (DC) are its most noteworthy ex-amples. In CA contiguous or non-contiguous componentcarriers, possibly allocated to several different BSs, arecombined and the scheduling and interference manage-ment orchestrated over these frequency bands aiming toenhance the resulting system’s capacity. As for DC, thisframework provides solutions for inter-frequency, for het-erogeneous networks (HetNets) scenarios, and for differentwireless standards MC so that a user equipment (UE) willbe simultaneously connected, respectively, in two differentfrequencies, to two different types of BSs or two differ-ent wireless standards [60]. Recently, the idea of DC formmWave and microwave bands has been proposed [33],[61] as an effective approach to facilitate cellular mmWaveinitial access (IA) [62] as well as mmWave handover [63].In like manner, mmWave and sub 6 GHz DC can teamtogether to augment the reliability of the mmWave work-ing as fallback to compensate eventual mmWave channelvulnerability, e.g. to blocking events. Finally, the benefits ofintegrating communication interface diversity for reliabilitypurposes are also studied in [64] in the context of machinetype communications (MTC).

SFN operation is proposed in use case 6 detailed inSection 4.2. The goal is to protect against mmWave channelintermittence by increasing the rate of those links betweenthe millimeter wave access points (mmAPs) and the virtualreality players (VRPs) that, otherwise, would jeopardize theimmersive experience.

High Reliability Enabler 2. Task Replication: While MCcan boost the reliability in the presence of channel fluctua-tions, it requires coordination between the different serversthat are connected to the end user. However, when coor-dination is not possible, reliability can still be enhancedthrough the task replication. Similar to packet replication indata communication, a user can offload a computing taskto multiple servers that are not connected to each otherand receive the result from whichever has the result readyfirst. This mechanism provides more guarantees of taskexecution, at the expense of reduced system capacity, due tothe under-utilization of computing servers. One realizationof this concept is proposed in [65], namely, hedged requests,is when the user sends one replica of the task to the serverthat is believed to be most suitable, then follows by sendinganother replica to an additional server after some delay.Completion pending remaining requests are canceled oncea result is received from any server.

While task replication is can be efficient in ensuring thereliability in in the case of channel dynamics, it incurs sig-nificant additional load. To combat this, one can offload thetask to an additional server only when the delay from thefirst server exceeds a certain threshold [65] This approachis investigated in [19]. Therein, it shown that imposingsuch condition can significantly curb the latency variability

without inducing much additional load.

High Reliability Enabler 3. Federated Machine Learning:While performing ML inference at the network edge yieldslow latency, distributed training of their ML models acrossdifferent edge nodes improves the inference reliability. To bespecific, each learning agent optimizes its ML model duringthe training phase so as to maximize the inference accuracyover locally available training data. The measured inferenceaccuracy at the training phase is however not always identi-cal to the inference accuracy at the test phase, primarily be-cause of unseen training data samples. This accuracy gap isknown as the generalization error that measures the inferencereliability under unseen data samples [66]. A straightfor-ward way to reduce the generalization error is exchangingtraining data samples among edge nodes. Data exchange,however, incurs extra communication and computation cost,and may not be available for user-generated private data. Toaddress this problem, federated learning (FL) has recentlybeen proposed [67], [68], in which edge nodes exchange andaggregate their local ML models, thereby preserving dataprivacy, avoiding extra computation, and reducing com-munication overhead when ML model sizes are sufficientlysmaller than data sizes.

FL is still a nascent field of research, calling for co-designing communication, computation, and ML architec-tures [11], [47]. For instance, the original FL algorithm hasthe communication payload size being proportional to theML model sizes, and thus cannot deal with deep neuralnetwork models. Proper model compression and parameterquantization techniques are thus needed, while trading theincreased communication efficiency off against the reducedaccuracy. Furthermore, the server in current FL algorithmssimply aggregates uploaded local models, although it hashigher computation resources compared to the edge devices.Along with these FL architectures, computing task offload-ing, task scheduling, and resource allocations should bejointly optimized towards achieving reliability under uncer-tainties on MEC operations, including unseen data samples,channel fluctuations, and time-varying communication andcomputation resources.

High Reliability Enabler 4. Extreme Event Control: Asmentioned previously, one reliability notion is the proba-bility of violation or failure over a latency bound, whichcan be mathematically expressed as Pr(Latency > Lbound).This probability ranges from 10−3 to 10−9, depending onthe mission-critical application in 5G networks [69]. Tomeet the ultra-reliability requirements, we should focus onthe extreme events with very low occurrence probabilities.However, in classical communication systems, the designedapproaches are based on the expected metrics, e.g., averagerate and average latency, in which the random event real-izations with higher probability distribution function (PDF)values dominate the system performance. In other words,the conventional average-based approaches are inadequatefor enhancing reliability performance, and instead we needto take into account the metrics or statistics, which arerelated to or affect the extreme events, such as

• worst-case measurement, e.g., largest latency in thenetwork,

8

• tail/decay behavior of the complementary cumulativedistribution function (CCDF),

• very low bound violation probability,• threshold deviation and its higher-order statistics, e.g.,

variance,

while designing the URLLC-enabled MEC systems. To ana-lytically analyze these metrics and statistics, extreme valuetheory (EVT) [70], [71] is a useful methodology for math-ematical characterization and, thus, provides a powerfulframework for extreme event control. Let us introduce thefundamental theorems in EVT as follows, which character-ize the aforementioned metrics and their statistics.

Theorem 1 (Fisher–Tippett–Gnedenko theorem [70]). Weconsider n independent and identically distributed (i.i.d.) samplesfrom a random variable X , i.e., X1, · · · , Xn

i.i.d.∼ X and defineZn := max{X1, · · · , Xn}. If Zn converges to a non-degeneratedistribution as n → ∞, we can approximate the limit as a gen-eralized extreme value (GEV) distribution which is characterizedby a location parameter µ ∈ R, a scale parameter σ > 0, and ashape parameter ξ ∈ R.

Among them, the shape parameter governs the GEVdistributions’ tail behaviors [71], which are sorted into threetypes depending on the value of ξ.

1) When ξ > 0, the GEV distribution has a heavy-tailedCCDF which is more weighted than an exponentialfunction.

2) When ξ = 0, the GEV distribution has a light tail, inwhich the CCDF has a thinner tail than an exponentialfunction.

3) When ξ < 0, the GEV distribution is short-tailed. That is,the CCDF has a finite upper endpoint at z = µ− σ/ξ.

When ξ ≥ 0, the upper endpoint of the CCDF approachesinfinity.

Theorem 2 (von Mises conditions [71]). In Theorem1, the characteristic parameters (µ, σ, ξ) of the approximatedGEV distribution can be asymptotically found as per µ =limn→∞

F−1X (1 − 1/n), σ = limn→∞

1nfX(F−1

X (1−1/n)) , and ξ =

−1− limx→∞

[1−FX(x)]f′X(x)

[fX(x)]2 .

Theorem 3 (Pickands–Balkema–de Haan theorem [70]).Consider the random variable X in Theorem 1 and a thresholdd. As d → F−1X (1), the CCDF of the excess value Y |X>d =X − d > 0 can be approximated as a generalized Pareto dis-tribution (GPD) whose mean and variance are σ/(1 − ξ) and

σ2

(1−ξ)2(1−2ξ) , respectively.

Analogously to the GEV distribution, the GPD is char-acterized by a scale parameter σ > 0 and a shape pa-rameter ξ ∈ R. In Theorems 1 and 3, ξ is identical whileσ = σ+ξ(µ−d). Note that Theorems 1 and 2 provide a wayto characterize the worst-case metric and its tail behavior,whereas Theorem 3 is directly related to the bound violationand its statistics. Since the characteristic parameters of theGEV distribution and GPD are identical or related, theresults of these three theorems are complementary to oneanother.

Nevertheless, some tradeoffs and dilemmas exist whenwe apply the results of EVT and estimate the characteristic

parameters. For example, we need to trade off data availabil-ity, which affects the performance, convergence speed, andestimation accuracy. Specifically, given N i.i.d. realizationsof X (i.e., N/n realizations of Zn), larger n theoreticallygives the better approximation of the GEV distributionbut slows down the convergence of parameter estimationsdue to the less availability of data samples of Zn. Thesimilar tradeoff between high threshold d and availabilityof threshold-exceeding data can be found from Theorem 3.Additionally, if the distribution of X , e.g., delay of a singleuser, is unknown beforehand, this agnostic makes Theorem2 difficult to characterize the network-wide largest delay.Fortunately, thanks to the mature development in the MLfield, the aforementioned issues can be tackled by using theML approaches, in which unsupervised learning providesa way to infer a mathematical expression of the unknowndistribution, while the lack of available data is addressed inan FL manner by aggregating and averaging the estimatedcharacteristic parameters of all distributed devices.

4 APPLICATIONS AND USE CASES

In this section, we elaborate on some of the prospectiveservices and applications for whom offloading their com-puting tasks to the edge significantly improves their per-formance in terms of latency and reliability. In particular,we focus on two scenarios where offloading task computingto the network edge will be beneficial: 1) when end usershave limited computing capabilities, e.g., VR head mounteddevices (HMDs)); and 2) when end users have sufficientcomputing and energy resources, but are accessible onlyto a fraction of the entire information for the computationinput, e.g., vehicular edge computing scenarios. We followby presenting different edge computing use cases in whichthe URLLC enablers are utilized.

4.1 Edge Computing Applications

4.1.1 Extended RealityExtended reality (XR) is an umbrella term that covers allvirtual or combined real-virtual environments, includingVR, AR and mixed reality (MR). These environments differin the nature of the content a user sees or interacts with.While VR describes environments where users are fullyimmerse in a virtual world, AR refers to the view of a virtualenvironment that is merged or supplemented by elementsor inputs from the real-world. AR can be categorized as aspecial case of the more general MR, which refers to theenvironments that mixes together real and virtual elementsthat can interact with each other.

XR is anticipated to be one of the leading applications toleverage edge computing. Providing high quality XR expe-rience comes with high computation resource demand. Atthe same time, XR applications are highly sensitive to delay.Typically, a maximum E2E delay, also known as motion-to-photon (MTP) delay, of 15-20 milliseconds can be toleratedin VR. Higher delay values trigger what is known as motionsickness, resulting from a visual-motor sensory conflict. Thismakes it unrealistic to rely on remote cloud servers forprocessing. On the other hand, Processing XR locally onthe user device has several complications. First, XR devices,

9

such as HMDs and smartphones are often equipped withlimited compute capabilities. This limitation is due to thedevice size, manufacturing cost, as well as to limit theheat generated from powering the device. Second, runningapplications on different types of devices, with differenthardware, operating systems, and platforms is a challengingtask. For these reasons, existing standalone XR devices oftenprovide limited content quality. Standalone VR headsetsoperate with reduced frame resolution and frame rate [72],whereas AR headsets such as Microsoft HoloLens restrictthe amount of renderable polygons [73].

For these reasons, the success of XR requires providinghigh computation and storage resources close to the endusers. In this regard, edge computing is an intuitive solutionto provide such services [74]. Today’s most powerful VRheadsets rely on edge computers to perform sophisticatedrendering. However, wired connections are still used be-tween the headsets and the edge servers, due to the high raterequirement of VR applications. This limits the mobility andconvenience of VR users and hence decrease the quality-of-experience (QoE).

The need for a better XR QoE and the advancement inwireless communication capabilities motivate the develop-ment of wireless XR systems that incorporate powerful edgecomputers and high capacity wireless links [44], [74]–[77].MmWave communication can provide large spectrum andhigh data rates, making it a solid candidate for wirelessXR. Moreover, the directionality of mmWave links allowfor leveraging multi-user transmission techniques such asmulticasting and broadcasting to deliver common and cor-related content to multiple users in a way that minimizesthe communication delay. However, directional mmWavelinks suffer outages due to signal blockage. This affectsthe link signal quality and increases the channel variability,and hence decreases the link reliability. MC can be a viablesolution to provide robust mmWave communication. UsingMC, an XR user maintains multiple simultaneous commu-nication links with multiple servers.

4.1.2 Vehicular Edge Computing and V2X/V2V for ADAS:

Future autonomous driving vehicles comprised as nodesof the Internet of Vehicles (IoV), a larger mobility networkwhich can be considered as an extended application of theIoT to ITSs [78], will operate as hubs integrating multipletechnologies and consuming and producing massive vol-umes of data [79]. The advanced driver-assistance systems(ADASs) to be equipped in these vehicles, especially thosepertaining to the area of traffic safety, heavily depend onreliable and instantaneous decision-making processes thathinge on inputs from multiple sensory data sources, includ-ing laser imaging detection and ranging (LIDAR), automo-tive radar, image processing, computer vision, etc. [80]. Asan example, we can think of successful object identificationfrom LIDAR point clouds or speed and trajectory predic-tion for dynamic objects moving within a vehicle’s vicin-ity. Hereof, it is essential that these vehicles are equippedwith powerful computing and processing capabilities toswiftly handle high data volumes rather than solely re-lying on cloud services that, in the above example, may

classify the objects or predict trajectories from raw datawith higher accuracy but, possibly, incurring to do so inunacceptable delays. Moreover, for next-generation ADASit is envisaged that vehicles will communicate with eachother as well as with an increasingly intelligent roadway in-frastructure through the use of vehicle-to-everything (V2X)and vehicle-to-vehicle (V2V) communications, ultimatelyexploiting high capacity mmWave links [81], [82]. Conse-quently, the cumbersome volume of locally generated datacould be exacerbated by the acquisition of data from boththe environment and from surrounding vehicles.

Indeed, vehicular edge computing will play a pivotalrole to support delay-sensitive as well as future emergingmultimedia-rich applications in vehicular networks, whichis buttressed by the growing body of literature devoted tothe area of content-centric applications of vehicular MEC[83]–[86] that are frequently combined with ML to providereliability as edge analytics [87], to leverage huge volumesof information [86] or to provide an integrated frameworkfor dynamic orchestration of networking, caching, and com-puting resources in next generation vehicular networks [88].

Being not nearly as tightly constrained by size or by theaccess to a power supply as their counterpart IoT devicesor smartphones, the computational and storage capabilitiesin vehicular terminals could allow them to run locally orcollaboratively, using vehicles as the infrastructures for com-munication and computation as proposed in [89], resource-hungry applications1. In this regard, provided that com-puting and processing capabilities may not be the limitingfactor, a second advantage of running these applications inthe network edge is substantiated by the availability of datacollected from multiple vehicles in edge servers. Access tothis information raw or preprocessed can augment individ-ual vehicles’ situational awareness by extending their ownsensing range. Resorting to edge contents can thus providea bigger picture at acceptable delays.

The later idea is exemplified in the third usecase inupcoming Section 4.2 where the information from differentvehicles is combined in the network edge following FLprinciples and used to refine a global model for transmissionqueue length distribution for the purpose of providing ultra-reliable low-latency V2V communications.

4.2 Use CasesNext, we present different case studies in which the URLLCenablers are utilized in edge computing settings.

Use case 1. Edge Computing for Federated Machine Learn-ing: As addressed in Sect. 3.2.1 and 3.2.2, edge ML isenvisaged to be a key enabler for URLLC, in which bothinference and training processes of ML models, e.g., neuralnetworks (NNs), are pushed down to the network edge [11].This direction of edgeML has been fueled by FL [67], [68],[91]–[94] under a data split architecture (see Fig. 2(a)), whereedge devices collectively train local models with their ownuser-generated data via a coordinating edge server that

1. However, the longer product’s life-span in the automotive indus-try, according to the US Department of Transportation as of 2018 theaverage age of on-the-road vehicles is over 11 years [90], could quicklyturn onboard central processing unit (CPU)/graphical processing unit(GPU) processing capabilities obsolete.

10

device

cloud server

MSI

ensemble

➁sample➀

➂

training dataedge server

ensemble

local MSI➀

➁

➁global MSI

edge server

aggregate

intermediate MSI➀

➁

(a) Data split.

device

cloud server

MSI

ensemble

➁sample➀

➂

training dataedge server

ensemble

local MSI➀

➁

➁global MSI

edge server

aggregate

➀

➁

intermediate MSI

➂

(b) Model split.

Figure 2. Edge ML architectural splits: (a) data split and (b) model split.

aggregates locally computed model updates, referred to asmodel state information (MSI). The MEC framework canfurther improve FL by its co-design with training architec-tures and algorithms. In view of this, on the one hand, eachedge device is able to optimize the MSI type dependingon the NN model size and channel quality. As done inFL, one can exchange the model parameter MSI whosepayload size is proportional to the model size, which isnot feasible for deep NNs under poor channel conditions.Alternatively, one can exchange model output MSI whosepayload size is independent of the model size, referred toas federated distillation (FD) [95]. As shown in Fig. 3(a),this fundamentally results in FD’s incomparably smallercommunication payload per MSI exchange than FL, and canthereby better cope with poor channel conditions.

On the other hand, the edge server can assist in the train-ing process by exploiting its extra computation and commu-nication resources. A compelling example is to rectify thenon-IID training dataset incurred by the user-generated na-ture of data, wherein entirely un-correlated (non-identical)and/or too similar (non-independent) data samples acrossdevices negate the benefit of distributed training [96]. Tothis end, in federated augmentation (FAug) [95], the edgeserver first collects few seed samples from edge devices,and oversamples them (e.g., via Google’s image search forvisual data) through its fast connection to the Internet. Then,the edge server can utilize its high computing power fortraining a generative model (e.g., conditional generativeadversarial network (GAN) [97]). Downloading the trainedgenerator empowers each device to locally augment defi-cient data samples until reaching an IID training dataset.With FAug, both FL and FD yield higher test accuracy asshown in Fig. 3(b), at the cost of slight increase in commu-nication cost as illustrated in Fig. 3(a).

Lastly, a very deep NN (e.g., Inception V4 NN modelconsuming 44.3 GB [98]) cannot fit into a single device’smemory, and has to be partitioned into multiple segmentsstored across edge devices and server, i.e., model split (seeFig.2(b)). Here, the model’s local and offloaded computa-tions should be orchestrated over wireless links, by optimiz-ing the partitioning strategy based on the NN’s topologyand constituent layers. This calls for a novel MEC frame-work that takes into account not only communication andcomputation resources but also NN forward and backwardpropagation dynamics intertwined with channel dynamics.

4 5 6 7 8 9 10Number of Devices

105

106

107

108

109

1010

1011

Sum

Com

mun

icat

ion

Cos

t (bi

ts/e

poch

)

FL+FAug (IID)FL (non-IID)FD+FAug (IID)FD (non-IID)

FD

FL

non-IID

(a) Communication cost.

4 5 6 7 8 9 10Number of Devices

0.65

0.7

0.75

0.8

0.85

0.9

0.95

1

Tes

t Acc

urac

y

FL+FAug (IID)FL (non-IID)FL+FAug (IID)FL (non-IID)

FL

FD

non-IID

(b) Test accuracy.

Figure 3. Communication cost and inference accuracy of federatedlearning (FL) and federated distillation (FD) with or without federatedaugmentation (FAug) in the MNIST classification problem, where eachdevice stores a 5-layer convolutional neural network (CNN). For FAug,the conditional generative adversarial network (GAN) consists of a 4-layer generator NN and another 4-layer discriminator NN.

Use case 2. Extreme Event-Controlled MEC: For the ex-treme event-controlling computation and communicationco-design in [99], [100], we studied a multi-user MEC sce-nario as shown in Fig. 4, in which multiple MEC serverswith different computation capabilities are deployed. In thissetting, the UE manages its local resource (i.e., total powerbudget) for computation and communication, i.e., task of-floading, while the MEC server schedules its computationalresources for the UEs’ offloaded tasks. Herein, we considerthe length of the task queue as a latency measurement sincequeuing latency can be reflected by the queue length. Forthe reliability concerns, we are concerned about the boundviolation probability and higher-order statistics of thresholddeviation as highlighted in high reliability enabler 4. In thisregard, we first impose a constraint on the queue length2

bound violation probability as

limT→∞

1

T

T∑t=1

Pr(Q(t) > d

)≤ ε� 1. (1)

2. The notation Q generalizes the lengths of all task queues at theUEs and MEC servers.

11

Local-computation queue

𝑄𝑄(𝑡𝑡)Outgoing tasks for computation

Queue length

𝐴𝐴(𝑡𝑡) 𝑅𝑅(𝑡𝑡)Task arrivals

MEC server

UE

MEC serverMEC server

UE UE

Offloaded-task queue

Figure 4. Extreme Event-Controlled MEC architecture.

0 10 20 30 40 50 60 70 80 90 100 110 120

Excess queue length X|Q>3.96×104 (kbit)

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

CCDF

Numerial resultApproximated GPD with (σ, ξ) = (9× 103, 0.23)

(a)

0 0.4 0.8 1.2 1.6 2 2.4 2.8 3.2 3.6 4

Processing density (cycle/bit)×10

4

0

20

40

60

80

100

120

140

160

180

200

99th

percentile

(kbit)

Non-MECFully-offloading, 30 UEsLocal-computation queue, 30 UEsTask-offloading queue, 30 UEs

(b)

0.4 0.8 1.2 1.6 2 2.4 2.8 3.2 3.6 4

Processing density (cycle/bit)×10

4

100

101

102

Meanan

dstan

darddeviation

(SD)of

exceedan

ce(kbit)

Mean, fully-offloadingSD, fully-offloadingMean, Task-offloading queueSD, Task-offloading queue

(c)

Figure 5. (a) Tail distributions of the excess queue length and the approximated GPD of exceedances, (b) 99th percentile of the queue length, and(c) mean and standard deviation of exceedances over the 99th percentile queue length, versus processing density.

Here, d and ε are the given bound and tolerable violationprobability. Let us further focus on the excess value over thebound d, which is denoted by X(t)|Q(t)>d = Q(t) − d > 0.By applying Theorem 3, we approximate the exceedancesas a GPD with the characteristic parameters (σ, ξ). Themean and variance are E

[X(t)|Q(t) > d

]≈ σ

1−ξ and

Var(X(t)|Q(t) > d

)≈ σ2

(1−ξ)2(1−2ξ) , respectively. We canfind that the smaller σ and ξ are, the smaller the meanvalue and variance. Since the approximated GPD is justcharacterized by the scale and shape parameters, we im-pose thresholds on these two parameters, i.e., σ ≤ σth

and ξ ≤ ξth. Subsequently, applying the two parameterthresholds and Var(X) = E[(X)2]− E[X]2, we consider theconditional constraints on the mean and second moment ofthe excess queue length

limT→∞

1

T

T∑t=1

E[X(t)|Q(t)>d

]≤ σth

1− ξth, (2)

limT→∞

1

T

T∑t=1

E[[X(t)]2|Q(t)>d

]≤

2(σth)2(

1− ξth)(1− 2ξth

) . (3)

Taking into account the above three requirements for theextreme events, we trade off the UE’s computation power

and communication power in the extreme event-controllingcomputation and communication co-design.

The effectiveness of characterizing threshold deviationby the Pickands–Balkema–de Haan theorem, i.e., Theorem 3,is verified in Fig. 5(a). Therein, Pr(Q > d) = 3.4×10−3 withd = 3.96 × 104. Additionally, in contrast with the schemeswithout edge computing and without local computationcapability, the extreme event-controlling approach achievesthe better performance, in terms of the extreme event-relatedmetrics shown in Fig. 5(b) and Fig. 5(c), in the consideredMEC system.

Use case 3. EVT/FL Ultra-Reliable Low-Latency V2V Com-munication: The idea of how to combine extreme valuetheory (EVT) and FL to enable URLLC in vehicular com-munication networks, referred as extFL, is discussed in ourpreliminary study [101], and illustrated in Fig. 6. Here, vehi-cles observe their queue length samples and utilize the taildistribution of queue lengths at the vehicular transmittersover the whole edge network to optimize their transmissiondecisions such that the worst-case queue lengths are mini-mized while ensuring reliability in terms of queuing latency.The analytical parametric model of the aforementioned tail

12

1

Collect from vehiclesexcess samples

2

Upload localmodel

(∇dfduu ,du, |Qu|)

3

Download globalmodel

(∇dfd,d,

∑u |Qu|)

4Calculate Tx power and

RB allocationfor V2V commun.

Based on the local or the global model

VUEu

···

V2V link

Perform ModelAveraging7→

Global Model

···

Federated

Learning

a

Updatelocal learning model

(∇dfduu ,du, |Qu|)

b

Enforce new policy

VUEu

transmissionqueue length

q0queue

threshold

time

···

···

···

···

···

···

···

···

···

···

RSU

Vehicular Network

···

···

···

···

···

GoalLearn σ, ξ fromdistributedly

contributed queueexcess samplesto improve the

accuracy

q0

distribution

GPD (σ,ξ)

Figure 6. Operational structure of EVT parametric FL (extFL).

distribution is obtained via EVT. Naturally, the evaluationof above parameters is carried out by gathering all queuelength samples at a central controller, the MEC server, withthe additional costs of communication and computationoverheads. In contrast to the centralized approach, here, FLis used to reduce the communication payload by allowingindividual vehicles to learn the tail distribution by exchang-ing a simplified model (two gradient values) instead of theirraw local queue length samples, i.e. enabling URLLC withthe aid of ML at the edge devices.

The goal is thus to minimize the network-wide powerconsumption of a set of vehicular user equipments (vUEs)while ensuring low queuing latencies with high reliability.However, there still exists worst-case vUEs experiencinghigh latencies with a low probability whose performancelosses are captured by extreme events pertaining to vehiclesqueue lengths exceeding a predefined threshold with non-negligible probability. The principles of EVT characterize thetail distribution of the queue lengths exceeding a predefinedthreshold by a generalized Pareto distribution with twoparameters scale and shape, respectively. The concepts inmaximum likelihood estimate (MLE) are used along FL toestimate the scale and shape parameters of the queue taildistribution locally at each vUEs over the queue lengthsamples. Therein, occasionally, local estimations and thegradients of MLE, known as local model at each vUEs areshared with the MEC server. The MEC server does modelaveraging and shares the global model with the vUEs toupdate their local estimations. Using the knowledge of thetail distribution over the network, the transmit power ofeach vUE is optimized to reduce the worst-case queuingdelays.

Fig. 7(a) compares the amount of data exchanged andthe achieved V2V communication reliability of extFL witha centralized tail distribution estimation model, denoted as

CEN. Note that the CEN method requires all vUEs to uploadall their queue length samples to the RSU and to receivethe estimated GPD parameters. In contrast, in extFL, vUEsupload their locally estimated learning models and receivethe global estimation of the model. As a result, extFLyields equivalent or better end user reliability comparedto CEN for denser networks while reducing the amountof data exchange among vUEs and the RSU. The worst-case vUEs queue lengths, i.e., queue lengths exceeding q0,are compared in Fig. 7(b). Here, the mean indicates theaverage queuing latency of the worst-case vUEs while thevariance highlights the uncertainty of the latency. As thenumber of vUEs increases, it can be noted that both themean and variance in extFL are lower than the ones in CEN.The reason for above improvement is the reduced traininglatency in extFL over CEN.

Use case 4. Deep Reinforcement Learning for OptimizedEdge Computing Task Offloading: The task offloadingdecision-making in edge computing networks is a chal-lenging task in the presence of environmental dynamics.This situation is aggravated in ultra-dense networks, wheresolutions to break the curse of dimensionality is desperatelyneeded. In the works [102], [103], a discrete-time Markovdecision process was adopted to model the problem ofexpected long-term MEC performance optimization in anultra-dense radio access network, where a number of BSs areavailable for computation task offloading. For a represen-tative wireless charging enabled MUE, whether to executean arriving computation task at the local mobile deviceor to offload the task for edge server execution via oneof the BSs should adapt to the environment dynamics inan intelligent manner. These environment dynamics mayconsist of random computation task arrivals, time-varyingcommunication qualities between the MU and the BSs andthe sporadic energy availability at the mobile device. The

13

(a)

(b)

Figure 7. Comparison between CEN and extFL. (a) The amount ofdata exchanged between RSU and VUEs (left axis) and the achievedreliability (right axis). (b) Mean and variance of the worst-case VUEqueue lengths.

challenges for the problem-solving lie in the lack of anya priori knowledge of any environment dynamic statisticsalong with the high dimensional state space. A deep rein-forcement learning technique shows the power of achievingan optimal solution.

More specifically, the objective of the MUE is to minimizean expected infinite-horizon discounted cost given by

Q(s, a) = E

[ ∞∑t=1

(γ)t−1 · c(st, at

)|s1 = s, a1 = a

], (4)

where γ ∈ [0, 1) is the discount factor, while the immediatecost c(st, at) after performing an action at under a statest at each time slot t takes into account the incurred taskexecution delay and the penalty of failing to process anarriving computation task. Once we obtain the optimal Q-function, the optimal action a∗ can be made by the MUEfollowing a∗ = argminaQ(s, a) under a state s. Insteadof using a conventional Q-learning to find the optimalQ-function, we resort to a deep-Q network (DQN) [104]Q(s, a;θ) to approximate Q(s, a) with θ being the set of pa-rameters of the neural network. The procedure of the deep

reinforcement learning for MEC performance optimizationis briefly depicted as in Fig. 8.

In Fig. 9, we compare the average cost performance fromthe Proposed deep reinforcement learning algorithm withthree baselines: 1) Local – Whenever a computation taskarrives, the MUE executes it at the local mobile deviceusing the queued energy units; 2) Server – All arrivingcomputation tasks are offloaded to the edge server for com-puting via the BSs with the best communication qualities;and 3) Greedy – When the computation task queue aswell as the energy queue are not empty at a time slot, theMUE decides to execute the task locally or at the cloud toachieve the minimum immediate cost. We configure a DQNof one hidden layer with 512 neurons. The replay memoryis assumed to have a capacity of 5000 and we select thesize of the mini-batch as 100. From Fig. 9, we can clearlysee that compared to the baselines, the deep reinforcementlearning algorithm realizes best performance in averagecost. A higher task arriving probability ρ indicates a longeraverage task execution delay, hence a larger average cost.As the average energy arrival rate increases, the averagecost improves due to the decreased failure of processing anarriving computation task.

Use case 5. Edge ML Enabled 360◦ VR Multicast Transmis-sion

Our previous work in [44] considered merging ML andmmWave multicasting to optimize the proactive wirelessstreaming of FoV-based high definition (HD) 360◦ videos ina multi-user VR environment with low latency guarantees.Hereof, the use of edge ML to predict users’ FoV in advanceis pivotal to leverage inter-user correlations and curb thelatency. These predicted correlations will ultimately driveboth how contents are transmitted and the beamformingdecisions at the mmWave base stations.

A VR theater scenario consisting of a network of VRusers watching different HD 360◦ VR videos streamed inthe mmWave band over a set of distributed small cellbase stations (SBSs) is studied. The SBSs will report users’6 degrees-of-freedom (6DoF) pose as well as CSI and pro-duce multiple spatially orthogonal beams to serve sharedFoV video content to groups of users (multicast) or a singlebeam (unicast) following the scheduling decisions adoptedat the edge controller. By optimizing video frame admissionand user scheduling, the goal is to provide a highly reliablebroadband service for VR users that deliver HD videos witha latency that is below the MTP latency limits with very highprobability.

To achieve this proactive content transmission and per-form a head movement pattern recognition predicting users’upcoming tiled-FoV, a sequential learning model basedon gated recurrent units (GRUs) [105], [106] is selected.Specifically, GRUs are a form of recurrent neural networks(RNNs) that include a double gating mechanism to governthe impact of past hidden states over the new output statesand effectively tackle long-term dependencies. To that pur-pose, an architecture based on 2 layers of GRU cells witha hidden state size equal to 512 separated by a rectifiedlinear unit (ReLU) activation are stacked. The output is thenfed to a serial to parallel (S/P) layer and to a dense neurallayer. Given the multi-label nature of the learning model, a

14

Internet Edge Server

Fiber Links

Base Stations

ComputationOffloading

Mobile User

Wireless Charging Energy Queue

Task QueueCurrent State

Loss and Gradient

Parameter Updating

ɛ-Greedy Policy

Mini-

Batch

⁞ ⁞

⁞ ⁞

DQN

⁞ ⁞

⁞ ⁞

Target DQN

Rep

lay

Mem

ory

Current State

Action

Immediate Cost

Next State

Current State

Action

Immediate Cost

Next State

Copy

Action

Deep Reinforcement Learning Loop

Figure 8. Illustration of deep reinforcement learning for mobile-edge computing performance optimization.

Figure 9. Average cost per time slot versus average energy arrival rateunder MILD (ρ = 0.3) and HEAVY (ρ = 0.5) task arrival probabilities,respectively represented with solid and dashed lines.

sigmoid activation layer maps the N sized dense output tothe N logits, one for each tile in the equirectangular (EQR)projection of the 360◦ VR video frame, which are binarizedwith a cutoff layer such that

yfpu,n =

{1, σ(Wdh

(2)f + bd)n ≥ γth,

0, otherwise,(5)

where Wd, bd are the weights and biases of the dense fully-connected layer and γth is the threshold value for the cutofflayer. The predicted FoV for a user u and frame indexfp = f + TH is retrieved as N fp

u = {n ∈ [1, ..., N ] : yfpu,n =

1}.Fig. 10 provides an overview of the building. The output

of the DRNN is fed to a user clustering module and Theformer constitutes one of the inputs for a scheduler the Lya-punov Drift plus penalty approach. In addition to our pro-posed scheme MPROAC+, the performance of three referencebaselines with reactive unicast and multicast, and proactivemulticast transmission capabilities, correspondingly, UREAC,MREAC, and MPROAC is evaluated. Our proposed approach

incorporates a penalty whereby quality is trade in exchangefor not violating a maximum latency bound. For simulationpurposes, a small size theatre with capacity for 50 userswith SBSs are located at ceiling level in its upper 4 cornersis selected. Fig. 11 evaluates the impact of the requestedHD video quality by representing the average and 99th

percentile delay, the HD delivery rate and Jaccard indexmeasured while 30 users watch one out of the 3 availableVR videos for an increasing requested video chunk size.

Fig. 11 clearly shows the tradeoff between frame delayand HD streaming rate. As the chunk size increases, theaverage and 99th percentile delays increase for the differ-ent schemes. Moreover, comparing UREAC with the otherschemes, it is shown that multicasting brings 40 − 50%increase in the HD rate and 33 − 70% latency reductionthrough the utilization of shared FoVs of different users.By delivering the predicted frames in advance, both theMPROAC and MPROAC+ minimize the average delay withoutsacrificing the HD quality rate. Moreover, our proposedMPROAC+ scheme is shown to also keep the worst delayvalues bounded due to imposing the constraint over thelatency.

The tradeoff between frame delay and quality is furtherillustrated the results for different values of the Lyapunovparameter Vδ are compared; as Vδ increases, the schedulingalgorithm prioritizes maximizing users’ HD delivery rate,whereas at lower values of the scheduler prioritizes keepingthe delay bounded with high probability. This comes at theexpense of having lower HD delivery rate.

Lastly, the Jaccard similarity in Fig. 11(d) illustrates thetradeoffs between effective vs. transmitted contents. At lowtraffic loads, the Jaccard index is low, which is due to thelarge amount of excess data delivered due to transmittingan estimated user/cluster level FoV. As the traffic loadincreases, the proactive schemes transmit more real-timeframes, which increases the Jaccard index. The Jaccard indexdecreases again at higher traffic loads as the effect of missedframes increases (once the average delay is close to reachingthe deadline, as can be seen in Fig. 11(a)).

15

Scheduler

Edge Controller

Deep RecurrentNeural Network with

GRU units

UserClustering

OUT: Schedulingcontents&users

7→SBSs

ReportCSI

CoordinatedHD video chunktransmission

Latency-aware

OptimizationVR

VR

VR

SBS b1

SBS b2

SBS b3

VR

VR

VR

VR

VR

VR

VR

Inputlayer

Outputlayer

Hiddenlayers

Real-timerequests Recursive

feedback

Wired/wirelesslink to edge controller

OUT: Tiled(multilabel)

FoV PredictionOUT:User clusters

&Cluster level FoV

3DoF pose

Report6DoFpose

proactivemulticast

real-timeunicast

VR

Tiled-FoV

proactivemulticast

··· }

Head orientationangles

(yaw, pitch, roll)

Figure 10. Operational structure and building blocks of the edge controller that coordinates the DRNN FoV prediction-aided proactive content qualityadaptation for the mmWave 360◦ VR video streaming.

Figure 11. (a) Average delay, (b) 99th percentile delay, (c) HD delivery rate and (d) Jaccard index performance in sT-3v, respectively, as a functionof the HD chunk size, for V =3 videos, K=2×V clusters, TH=5 frames, and Lyapunov trade-off Vδ=1·108 and Vδ=1·109.

Use case 6. MEC Enabled Multi-User VR Gaming ArcadeWe consider a practical use case of wireless VR to delivera low latency service to multi-user scenario of users play-ing VR video games in a gaming arcade. This scenario,that is fully detailed in our previous work [74], is highlydemanding due to the tight latency tolerance in VR aswell as the state dynamics of the user due to the game-specific actions taken by themselves or by other players thataffect what content should be shown to them. The users areserved wirelessly through multiple mmAPs wired to edgecomputing and storage servers. These servers receive theusers 3-dimensional (3D) location coordinates, their 3D posethat consists of roll, pitch, and yaw angles, and their game-related actions. The servers will render the correspondingframes in HD resolution and deliver it wirelessly to users.Hence, the latency consists of the processing latency at theserver and the communication latency to deliver the HDframes expressed as

Duf (t) = ξfu(Dcpuf (t) +Dcm

uf (t) + τEP), (6)

where ξfu represents a binary indicator that equals 1 whenthe HD video frame is delivered to VRP u and equals 0 if thelow quality (LQ) frame is delivered, Dcp

uf and Dcmuf are the

computing and communication delays of HD frame f ini-tiated from user u, and τEP is the processing latency which

accounts for the edge server processing, storage processing,and the UL transmission of user pose and action data. Letthe computing delay Dcp

uf be expressed as follows:

Dcpuf (t) =

(κLHD

fu

ce+Wuf (t)

)zfu(t)(1− yfu(t)), (7)

where ce is the computation capability of edge server e,zfu(t) and yfu(t) indicate that the video frame f of user u isscheduled for computing, and is cached in the fog networkat time instant t, respectively, and Wuf is the computationwaiting time of HD frame f of user u in the service queue,defined as Q(t). Furthermore, let the communications delayDcmuf be given as

Dcmuf (t)=argmin

du

Dcpuf (t)+du∑

t′=Dcpuf (t)+1

(Ttru(t

′) ≥ LHDfu

), (8)

where the argmin function is to find the minimum numberof time slots needed for the video frame f to be delivered.

Here, we study two enablers to minimize the latencyand boost the reliability of the VR gaming experience. Forthe computing latency, we investigate how prior knowledgeof users’ future pose using prediction methods affects thecomputing the latency. We leverage results from previousworks as in [107] that state that the user’s future pose inthe next hundreds of milliseconds can be predicted withhigh accuracy to proactively predict, render, and cache the

16

Edge Computing andStorage Platform

mmAPa′mmAPa

Detail view of theGaming Arcade

LOS linkto mmAPa

LOS linkto mmAPa′′

NLOS(blocked) linkto mmAPa′mmAPa′′′

Temporal beammisalignmentto mmAPa′

Gaming Arcade Gaming Arcade

Gaming ArcadeGaming Arcade

Figure 12. Representation of a group of VR gaming arcades where HD frame computation is offloaded to a MEC platform such that input actionsof the VRPs might impact the virtual environment shown to a subset of the remaining VRPs. The detailed view of the bottom arcade also illustratesseveral LOS and NLOS mmWave link states e.g., link blockage and VRP and mmAP beam misalignment.

users upcoming frames, subject to computation and storageresource availability. For the communication parts, the useof MC is considered to associate a user with more than onemmAPs if the SINR with its serving mmAP falls below agiven threshold. Specifically, SFN operation is consideredwhere multiple mmAP use the same frequency and timeresource to transmit to the intended user.

Fig. 13 compares the communications and computing la-tency of our PROPOSED scheme that considers both enablersof proactive computing and MC, with BASELINE-1 thatdoes not have either of the two enablers, and BASELINE-2that considers only proactive computing. By looking intothe computing latency in Fig. 13, we can see that theschemes with proactive computing significantly minimizesthe computing latency, whereas a look at the communicationlatency shows the gain achieved using MC. Comparing thecommunication latency of BASELINE-1 and BASELINE-2also shows that the proactive computing, that improved thecomputing performance, also slightly increases the commu-nication latency. This is due to having to send additionaldata due to the errors in prediction, in which the correctdata has to be retransmitted in real time.

5 CONCLUSIONS AND FUTURE OUTLOOK

Edge computing is an essential component of future wire-less networks, in which several challenges need to beovercome to realize the vision of ultra-reliable and low-latency edge computing. Chief to this vision is leveragingmultiple high reliability and low-latency enablers appliedfor different types of services and use cases. In this article,we have discussed edge networking services and examinedkey enablers to achieve low-latency and high reliabilitynetworking. Moreover, we showcased how the networkresources can be optimized for a selection of use casescharacterized by their shared need for edge networking.

Figure 13. The communication delay (solid lines) and computing delay(dashed lines) for different schemes as the number of players varies foran arcade of 16 mmAPs, each equipped with an edge computing unit.

As the vision of 5G starts to materialize beyond its initialinception towards imminent first commercial deployments,we envision a realization of edge computing hand in handwith the development of URLLC and distributed artificialintelligence (AI) able to deal with dynamic and heteroge-neous environments, provide seamless computing, content,and control services, while preserving data privacy andsecurity.

REFERENCES

[1] W. Shi, J. Cao, Q. Zhang, Y. Li, and L. Xu, “Edge computing:Vision and challenges,” IEEE Internet of Things Journal, vol. 3,no. 5, pp. 637–646, Oct 2016.

17

[2] G. Klas, “Edge computing and the role of cellular networks,”Computer, vol. 50, no. 10, pp. 40–49, 2017.

[3] A. Ndikumana, N. H. Tran, T. M. Ho, Z. Han, W. Saad, D. Niyato,and C. S. Hong, “Joint communication, computation, caching,and control in big data multi-access edge computing,” ArXivpreprint, vol. abs/1803.11512, 2018.

[4] B. Soret, P. Mogensen, K. I. Pedersen, and M. C. Aguayo-Torres,“Fundamental tradeoffs among reliability, latency and through-put in cellular networks,” in 2014 IEEE Globecom Workshops (GCWkshps). IEEE, 2014, pp. 1391–1396.

[5] Y. Mao, C. You, J. Zhang, K. Huang, and K. B. Letaief, “Asurvey on mobile edge computing: The communication perspec-tive,” IEEE Commun. Surveys Tuts., vol. 19, no. 4, pp. 2322–2358,Fourthquarter 2017.

[6] ETSI, “Multi-access Edge Computing (MEC),” https://www.etsi.org/technologies/multi-access-edge-computing, 2018, [Online;accessed 04-02-2019].

[7] E. Bastug, M. Bennis, and M. Debbah, “Living on the edge: Therole of proactive caching in 5g wireless networks,” IEEE Commun.Mag., vol. 52, no. 8, pp. 82–89, Aug. 2014.

[8] E. Bastug, M. Bennis, and M. Debbah, “Social and spatial proac-tive caching for mobile data offloading,” in 2014 IEEE Interna-tional Conference on Communications Workshops (ICC), June 2014,pp. 581–586.

[9] S. Tamoor-ul-Hassan, M. Bennis, P. H. J. Nardelli, and M. Latva-aho, “Caching in wireless small cell networks: A storage-bandwidth tradeoff,” IEEE Communications Letters, vol. 20, no. 6,pp. 1175–1178, June 2016.

[10] M. S. ElBamby, M. Bennis, W. Saad, and M. Latva-aho, “Content-aware user clustering and caching in wireless small cell net-works,” in 2014 11th International Symposium on Wireless Commu-nications Systems (ISWCS), Aug 2014, pp. 945–949.

[11] J. Park, S. Samarakoon, M. Bennis, and M. Debbah, “Wireless net-work intelligence at the edge,” Proceedings of the IEEE (submitted)[Online]. Arxiv preprint, vol. abs/1812.02858, Dec. 2018.

[12] E. Li, Z. Zhou, and X. Chen, “Edge intelligence: On-demanddeep learning model co-inference with device-edge synergy,” inProceedings of the Workshop on Mobile Edge Communications, ser.MECOMM ’18. ACM, 2018.

[13] Y. Sahni, J. Cao, S. Zhang, and L. Yang, “Edge mesh: A newparadigm to enable distributed intelligence in internet of things,”IEEE Access, vol. 5, pp. 16 441–16 458, 2017.

[14] J.-M. Lasry and P.-L. Lions, “Mean field games,” Japanese Journalof Mathematics, vol. 2, no. 1, pp. 229–260, Mar. 2007. [Online].Available: https://doi.org/10.1007/s11537-007-0657-8

[15] C. Dehos, J. L. Gonzlez, A. D. Domenico, D. Ktnas, and L. Dus-sopt, “Millimeter-wave access and backhauling: the solution tothe exponential data traffic increase in 5G mobile communica-tions systems?” IEEE Commun. Mag., vol. 52, no. 9, pp. 88–95,Sep. 2014.

[16] T. K. Vu, M. Bennis, S. Samarakoon, M. Debbah, and M. Latva-aho, “Joint in-band backhauling and interference mitigation in5G heterogeneous networks,” in European Wireless 2016; 22thEuropean Wireless Conference. VDE, 2016, pp. 1–6.

[17] M. S. Elbamby, M. Bennis, and W. Saad, “Proactive edge com-puting in latency-constrained fog networks,” in 2017 EuropeanConference on Networks and Communications (EuCNC), June 2017,pp. 1–6.

[18] J. Oueis, “Joint communication and computation resourcesallocation for cloud-empowered future wireless networks,”Theses, Universite Grenoble Alpes, Feb. 2016. [Online].Available: https://tel.archives-ouvertes.fr/tel-01366449

[19] M. S. Elbamby, M. Bennis, W. Saad, M. Latva-aho, and C. S. Hong,“Proactive edge computing in fog networks with latency and reli-ability guarantees,” EURASIP Journal on Wireless Communicationsand Networking, vol. 2018, no. 1, p. 209, Aug 2018.

[20] T. S. Rappaport, S. Sun, R. Mayzus, H. Zhao, Y. Azar, K. Wang,G. N. Wong, J. K. Schulz, M. Samimi, and F. Gutierrez, “Millime-ter wave mobile communications for 5G cellular: It will work!”IEEE Access, vol. 1, pp. 335–349, 2013.

[21] J. G. Andrews, S. Buzzi, W. Choi, S. V. Hanly, A. Lozano, A. C. K.Soong, and J. C. Zhang, “What Will 5G Be?” IEEE J. Sel. AreasCommun., vol. 32, no. 6, pp. 1065–1082, 6 2014.

[22] F. Boccardi, R. W. Heath, A. Lozano, T. Marzetta, and P. Popovski,“Five disruptive technology directions for 5G,” IEEE Commun.Mag., vol. 52, no. 2, pp. 74–80, Feb. 2014.

[23] M. Xiao, S. Mumtaz, Y. Huang, L. Dai, Y. Li, M. Matthaiou,G. K. Karagiannidis, E. Bjrnson, K. Yang, C. I, and A. Ghosh,“Millimeter wave communications for future mobile networks,”IEEE J. Sel. Areas Commun., vol. 35, no. 9, pp. 1909–1935, 9 2017.

[24] T. S. Rappaport, Y. Xing, G. R. MacCartney, A. F. Molisch,E. Mellios, and J. Zhang, “Overview of millimeter wave com-munications for fifth-generation (5g) wireless networkswith afocus on propagation models,” IEEE Transactions on Antennas andPropagation, vol. 65, no. 12, pp. 6213–6230, Dec 2017.

[25] M. R. Akdeniz, Y. Liu, M. K. Samimi, S. Sun, S. Rangan, T. S.Rappaport, and E. Erkip, “Millimeter wave channel modelingand cellular capacity evaluation,” IEEE J. Sel. Areas Commun.,vol. 32, no. 6, pp. 1164–1179, Jun. 2014.

[26] I. A. Hemadeh, K. Satyanarayana, M. El-Hajjar, and L. Hanzo,“Millimeter-wave communications: Physical channel models, de-sign considerations, antenna constructions, and link-budget,”IEEE Commun. Surveys Tuts., vol. 20, no. 2, pp. 870–913, Sec-ondquarter 2018.

[27] A. L. Swindlehurst, E. Ayanoglu, P. Heydari, and F. Capolino,“Millimeter-wave massive MIMO: the next wireless revolution?”IEEE Commun. Mag., vol. 52, no. 9, pp. 56–62, Sep. 2014.

[28] M. Rebato, J. Park, P. Popovski, E. D. Carvalho, , and M. Zorzi,“Stochastic geometric coverage analysis in mmWave cellularnetworks with realistic channel and antenna radiation models,”ArXiv preprint, vol. abs/1806.04193, Jun. 2018.

[29] M. Giordani, M. Polese, A. Roy, D. Castor, and M. Zorzi,“A tutorial on beam management for 3GPP NR at mmWavefrequencies,” IEEE Commun. Surveys Tuts., 2019. [Online].Available: https://ieeexplore.ieee.org/document/8458146

[30] S. Barbarossa, E. Ceci, and M. Merluzzi, “Overbooking Radioand Computation Resources in mmW-Mobile Edge Computingto Reduce Vulnerability to Channel Intermittency,” in Proc. Eur.Conf. on Networks and Commun. (EuCNC), 2017, pp. 1–5. [Online].Available: https://doi.org/10.1109/EuCNC.2017.7980746

[31] T. K. Vu, M. Bennis, M. Debbah, M. Latva-aho, and C. S. Hong,“Ultra-reliable communication in 5G mmWave networks: A risk-sensitive approach,” IEEE Communications Letters, vol. 22, no. 4,pp. 708–711, Apr. 2018.

[32] M. Kamel, W. Hamouda, and A. Youssef, “Ultra-dense networks:A survey,” IEEE Trans. Control Syst. Technol., vol. 18, no. 4, pp.2522–2545, Fourthquarter 2016.

[33] J. Park, S.-L. Kim, and J. Zander, “Tractable resource managementwith uplink decoupled millimeter-wave overlay in ultra-densecellular networks,” IEEE Trans. Wireless Commun., vol. 15, no. 6,pp. 4362–4379, Jun. 2016.

[34] J. Park, S. Jung, S.-L. Kim, M. Bennis, and M. Debbah, “User-centric mobility management in ultra-dense cellular networksunder spatio-temporal dynamics,” in Proc. of IEEE GLOBECOM,Washington, DC, USA, Dec. 2016.

[35] J. Park, , S.-L. Kim, M. Bennis, and M. Debbah, “Spatio-temporalnetwork dynamics framework for energy-efficient ultra-densecellular networks,” in Proc. of IEEE GLOBECOM, Washington,DC, USA, Dec. 2016.

[36] H. Kim, J. Park, M. Bennis, S. Kim, and M. Debbah, “Ultra-dense edge caching under spatio-temporal demand and networkdynamics,” in Proc. of IEEE ICC, Paris, France, May 2017, pp. 1–7.

[37] S. Samarakoon, M. Bennis, W. Saad, M. Debbah, and M. Latva-aho, “Ultra dense small cell networks: Turning density intoenergy efficiency,” IEEE J. Sel. Areas Commun., vol. 34, no. 5, pp.1267–1280, May 2016.

[38] B. Li, Z. Fei, and Y. Zhang, “UAV communications for 5G andbeyond: Recent advances and future trends,” IEEE Internet ofThings Journal, pp. 1–1, 2019.

[39] F. Hagenauer, C. Sommer, T. Higuchi, O. Altintas, and F. Dressler,“Vehicular micro clouds as virtual edge servers for efficient datacollection,” in Proceedings of the 2Nd ACM International Workshopon Smart, Autonomous, and Connected Vehicular Systems and Ser-vices, ser. CarSys ’17. New York, NY, USA: ACM, 2017, pp.31–35.

[40] J. Park, P. Popovski, and O. Simeone, “Minimizing latency tosupport VR social interactions over wireless cellular systems viabandwidth allocation,” vol. 7, no. 5, pp. 776–779, Oct 2018.

[41] B. Romanous, N. Bitar, A. Imran, and H. Refai, “Network den-sification: Challenges and opportunities in enabling 5G,” in 2015IEEE 20th International Workshop on Computer Aided Modelling andDesign of Communication Links and Networks (CAMAD), Sep. 2015,pp. 129–134.

18

[42] R. Arshad, H. ElSawy, S. Sorour, T. Y. Al-Naffouri, andM. Alouini, “Handover management in dense cellular networks:A stochastic geometry approach,” in 2016 IEEE InternationalConference on Communications (ICC), May 2016, pp. 1–7.

[43] Y. Wang, M. Narasimha, and R. W. H. Jr, “MmWave beam predic-tion with situational awareness: A machine learning approach,”ArXiv preprint, vol. abs/1805.08912, May 2018.

[44] C. Perfecto, M. S. ElBamby, J. Del Ser, and M. Bennis, “Tamingthe latency in multi-user VR 360◦: A QoE-aware deep learning-aided multicast framework,” ArXiv preprint, vol. abs/1811.07388,2018.

[45] N. Kato, Z. M. Fadlullah, B. Mao, F. Tang, O. Akashi, T. Inoue,and K. Mizutani, “The deep learning vision for heterogeneousnetwork traffic control: Proposal, challenges, and future perspec-tive,” IEEE Wireless Commun. Mag., vol. 24, no. 3, pp. 146–153,June 2017.

[46] Q. Mao, F. Hu, and Q. Hao, “Deep learning for intelligent wire-less networks: A comprehensive survey,” IEEE CommunicationsSurveys Tutorials, vol. 20, no. 4, pp. 2595–2621, jun 2018.

[47] G. Zhu, D. Liu, Y. Du, C. You, J. Zhang, and K. Huang, “Towardsan intelligent edge: Wireless communication meets machinelearning,” ArXiv preprint, vol. abs/1809.00343, 2018.

[48] S. Ko, K. Huang, S. Kim, and H. Chae, “Live prefetching formobile computation offloading,” IEEE Trans. Wireless Commun.,vol. 16, no. 5, pp. 3057–3071, May 2017.

[49] E. Bastug, M. Bennis, M. Medard, and M. Debbah, “Towardinterconnected virtual reality: Opportunities, challenges, and en-ablers,” IEEE Commun. Mag., vol. 55, no. 6, pp. 110–117, June2017.

[50] X. Chen, Q. Shi, L. Yang, and J. Xu, “Thriftyedge: Resource-efficient edge computing for intelligent iot applications,” IEEENetw., vol. 32, no. 1, pp. 61–65, Jan 2018.

[51] S. Li, M. A. Maddah-Ali, and A. S. Avestimehr, “Coded mapre-duce,” in 2015 53rd Annual Allerton Conference on Communication,Control, and Computing (Allerton), Sep. 2015, pp. 964–971.

[52] ——, “Coding for distributed fog computing,” IEEE Commun.Mag., vol. 55, no. 4, pp. 34–40, April 2017.

[53] A. Ravanshid, P. Rost, D. S. Michalopoulos, V. V. Phan, H. Bakker,D. Aziz, S. Tayade, H. D. Schotten, S. Wong, and O. Holland,“Multi-connectivity functional architectures in 5g,” in IEEE Int.Conf. Commun. Workshops (ICC), May 2016, pp. 187–192.

[54] B. Soret, P. Mogensen, K. I. Pedersen, and M. C. Aguayo-Torres,“Fundamental tradeoffs among reliability, latency and through-put in cellular networks,” in Proc. IEEE Global Telecommun. Conf.(GLOBECOM) Workshops, 2014, pp. 1391–1396.

[55] A. Wolf, P. Schulz, M. Drpinghaus, J. C. S. S. Filho, and G. Fet-tweis, “How reliable and capable is multi-connectivity?” IEEETrans. Commun., pp. 1–1, 2018.

[56] 3GPP TR36.819, “Coordinated multi-point operation for LTEphysical layer aspects,” Tech. Rep. 36.819 Rel-11, 9 2013.

[57] M. Eriksson, “Dynamic single frequency networks,” IEEE J. Sel.Areas Commun., vol. 19, no. 10, pp. 1905–1914, Oct 2001.

[58] F. B. Tesema, A. Awada, I. Viering, M. Simsek, and G. P. Fettweis,“Multiconnectivity for Mobility Robustness in Standalone 5GUltra Dense Networks with Intrafrequency Cloud Radio Access,”Wirel. Commun. Mob. Comput., vol. 2017, pp. 1–17, Jan 2017.

[59] 3GPP TR 36.823, “Evolved Universal Terrestrial Radio Access (E-UTRA); Carrier Aggregation Enhancements; UE and BS radiotransmission and reception,” Tech. Rep. 36.823 Rel-11, 11 2013.

[60] F. B. Tesema, A. Awada, I. Viering, M. Simsek, and G. P. Fet-tweis, “Mobility modeling and performance evaluation of multi-connectivity in 5g intra-frequency networks,” in 2015 IEEE Globe-com Workshops (GC Wkshps), Dec 2015, pp. 1–6.

[61] M. Giordani, M. Mezzavilla, S. Rangan, and M. Zorzi, “Multi-connectivity in 5G mmWave cellular networks,” in 2016 Mediter-ranean Ad Hoc Networking Workshop (Med-Hoc-Net), June 2016, pp.1–7.

[62] M. Giordani, M. Mezzavilla, and M. Zorzi, “Initial access in 5GmmWave cellular networks,” IEEE Commun. Mag., vol. 54, no. 11,pp. 40–47, Nov 2016.

[63] M. Polese, M. Giordani, M. Mezzavilla, S. Rangan, and M. Zorzi,“Improved handover through dual connectivity in 5G mmWavemobile networks,” IEEE J. Sel. Areas Commun., vol. 35, no. 9, pp.2069–2084, Sep. 2017.

[64] J. J. Nielsen and P. Popovski, “Latency analysis of systems withmultiple interfaces for ultra-reliable M2M communication,” in

2016 IEEE 17th International Workshop on Signal Processing Ad-vances in Wireless Communications (SPAWC), July 2016, pp. 1–6.

[65] J. Dean and L. A. Barroso, “The tail at scale,” Commun. ACM,vol. 56, no. 2, pp. 74–80, Feb. 2013.

[66] O. Bousquet, U. von Luxburg, and G. Ratsch, Introduction toStatistical Learning Theory. Berlin, Heidelberg: Springer BerlinHeidelberg, 2004, pp. 169–207.

[67] J. Konecny, H. B. McMahan, F. X. Yu, P. Richtarik, A. T. Suresh,and D. Bacon, “Federated learning: strategies for improving com-munication efficiency,” in Proc. of NIPS Wksp. PMPML, Barcelona,Spain, Dec. 2016.

[68] H. B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A.y Arcas, “Communication-efficient learning of deep networksfrom decentralized data,” in Proc. of AISTATS, Fort Lauderdale,FL, USA, Apr. 2017.

[69] M. Bennis, M. Debbah, and V. Poor, “Ultra-reliable and low-latency wireless communication: Tail, risk and scale,” Proc. IEEE,vol. 106, no. 10, pp. 1834–1853, Oct. 2018.

[70] S. Coles, An Introduction to Statistical Modeling of Extreme Values.London, U.K.: Springer, 2001.

[71] L. de Haan and A. Ferreira, Extreme Value Theory: An Introduction.New York, NY, USA: Springer, 2006.

[72] M. S. Elbamby, C. Perfecto, M. Bennis, and K. Doppler, “Towardlow-latency and ultra-reliable virtual reality,” IEEE Netw., vol. 32,no. 2, pp. 78–84, Mar. 2018.

[73] M. Schneider, J. Rambach, and D. Stricker, “Augmented realitybased on edge computing using the example of remote live sup-port,” in 2017 IEEE International Conference on Industrial Technology(ICIT), Mar. 2017, pp. 1277–1282.

[74] M. S. Elbamby, C. Perfecto, M. Bennis, and K. Doppler, “Edgecomputing meets millimeter-wave enabled VR: Paving the wayto cutting the cord,” in 2018 IEEE Wireless Commun. and Netw.Conf. (WCNC), 4 2018, pp. 1–6.

[75] ABI Research and Qualcomm, “Augmented and virtual reality:The first wave of 5g killer apps,” White paper, Feb. 2017.

[76] J. Park and M. Bennis, “URLLC-eMBB slicing to support VRmultimodal perceptions over wireless cellular systems,” ArXivpreprint, vol. abs/1805.00142, May 2018.

[77] A. AL-Shuwaili and O. Simeone, “Energy-efficient resource al-location for mobile edge computing-based augmented realityapplications,” vol. 6, no. 3, pp. 398–401, apr 2017.

[78] L. Nanjie, “Internet of vehicles: Your next connection,” in WinWinMagazine, Huawei Communicate, Ed., Dec. 2011.

[79] P. Gao, R. Hensley, and A. Zielke, “A road mapto the future for the auto industry,” McKinseyQuarterly, Oct. 2014. [Online]. Available: https://www.mckinsey.com/industries/automotive-and-assembly/our-insights/a-road-map-to-the-future-for-the-auto-industry

[80] C. Perfecto, J. Del Ser, M. Bennis, and M. N. Bilbao, “BeyondWYSIWYG: Sharing contextual sensing data through mmWaveV2V communications,” in 2017 Eur. Conf. on Netw. and Commun.(EuCNC), 6 2017, pp. 1–6.

[81] C. Perfecto, J. Del Ser, and M. Bennis, “Millimeter-wave V2Vcommunications: Distributed association and beam alignment,”IEEE J. Sel. Areas Commun., vol. 35, no. 9, pp. 2148–2162, 9 2017.

[82] C. Perfecto, J. Del Ser, and M. Bennis, “On the interplay betweenscheduling interval and beamwidth selection for low-latency andreliable v2v mmwave communications,” in 2017 20th Conf. onInnovations in Clouds, Internet and Networks (ICIN). IEEE, 2017,pp. 1–8.

[83] M. Amadeo, C. Campolo, and A. Molinaro, “Information-centricnetworking for connected vehicles: a survey and future perspec-tives,” IEEE Commun. Mag., vol. 54, no. 2, pp. 98–104, 2 2016.

[84] K. Zhang, Y. Mao, S. Leng, Y. He, and Y. Zhang, “Mobile-edge computing for vehicular networks: A promising networkparadigm with predictive off-loading,” IEEE Veh. Technol. Mag.,vol. 12, no. 2, pp. 36–44, 6 2017.

[85] Q. Yuan, H. Zhou, J. Li, Z. Liu, F. Yang, and X. S. Shen, “Towardefficient content delivery for automated driving services: An edgecomputing solution,” IEEE Netw., vol. 32, no. 1, pp. 80–86, 1 2018.

[86] Z. Zhou, H. Yu, C. Xu, Z. Chang, S. Mumtaz, and J. Rodriguez,“Begin: Big data enabled energy-efficient vehicular edge comput-ing,” IEEE Commun. Mag., vol. 56, no. 12, pp. 82–89, 12 2018.

[87] A. F. Khosrowshah, U. Challita, and W. Saad, “Deep learningfor reliable mobile edge analytics in intelligent transportationsystems: An overview,” IEEE Veh. Technol. Mag., pp. 1–1, 2019.

19

[88] Y. He, N. Zhao, and H. Yin, “Integrated networking, caching, andcomputing for connected vehicles: A deep reinforcement learningapproach,” IEEE Trans. Veh. Technol., vol. 67, no. 1, pp. 44–55, 12018.

[89] X. Hou, Y. Li, M. Chen, D. Wu, D. Jin, and S. Chen, “Vehicularfog computing: A viewpoint of vehicles as the infrastructures,”IEEE Trans. Veh. Technol., vol. 65, no. 6, pp. 3860–3873, Jun. 2016.

[90] “Consumer Reports - Make Your Car Last 200,000 Miles,” [on-line, Accessed: 2019-01-15], https://www.consumerreports.org/car-repair-maintenance/make-your-car-last-200-000-miles/.

[91] H. Kim, J. Park, M. Bennis, and S.-L. Kim, “On-device federatedlearning via blockchain and its latency analysis,” ArXiv preprint,vol. abs/1808.03949, Aug. 2018.

[92] M. M. Amiri and D. Gunduz, “Machine learning at the wirelessedge: Distributed stochastic gradient descent over-the-air,” ArXivpreprint, vol. abs/1901:00844, Jan. 2019.

[93] S. Ha, J. Zhang, O. Simeone, and J. Kang, “Coded federatedcomputing in wireless networks with straggling devices andimperfect csi,” ArXiv preprint, vol. abs/1901.05239, 2019.

[94] S. Wang, T. Tuor, T. Salonidis, and K. K. Leung, “Adaptive feder-ated learning in resource constrained edge computing systems,”ArXiv preprint, vol. abs/1804.05271, Aug. 2018.

[95] E. Jeong, S. Oh, H. Kim, J. Park, M. Bennis, and S.-L. Kim,“Federated distillation and augmentation under non-IID privatedata,” presented at NIPS Wksp. MLPCD, Dec. 2018.

[96] I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning. MITPress, 2016.

[97] M. Mirza and S. Osindero, “Conditional generative adversarialnets,” ArXiv preprint, vol. abs/1411.1784, Nov. 2014.

[98] L. Wang, J. Ye, Y. Zhao, W. Wu, A. Li, S. L. Song,Z. Xu, and T. Kraska, “Superneurons: Dynamic gpu memorymanagement for training deep neural networks,” The ACMSpecial Interest Group on Programming Languages, vol. 53,no. 1, pp. 41–53, Feb. 2018. [Online]. Available: http://doi.acm.org/10.1145/3200691.3178491

[99] C.-F. Liu, M. Bennis, and H. V. Poor, “Latency and reliability-aware task offloading and resource allocation for mobile edgecomputing,” in Proc. IEEE Global Commun. Conf. Workshops, Dec.2017, pp. 1–7.

[100] C.-F. Liu, M. Bennis, M. Debbah, and H. V. Poor, “Dynamic taskoffloading and resource allocation for ultra-reliable low-latencyedge computing,” IEEE Trans. Commun., vol. 67, 2019. [Online].Available: http://arxiv.org/abs/1812.08076

[101] S. Samarakoon, M. Bennis, W. Saad, and M. Debbah, “Distributedfederated learning for ultra-reliable low-latency vehicular com-munications,” ArXiv preprint, vol. abs/1807.08127, Jul. 2018.

[102] X. Chen, H. Zhang, C. Wu, S. Mao, Y. Ji, and M. Bennis,“Performance optimization in mobile-edge computing via deepreinforcement learning,” in 2018 IEEE 88th Vehicular TechnologyConference (VTC-Fall), Aug. 2018.

[103] ——, “Optimized computation offloading performance in virtualedge computing systems via deep reinforcement learning,” IEEEInternet Things J., Early Access Article 2018.

[104] V. Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G.Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland, G. Ostro-vski, S. Petersen, C. Beattie, A. Sadik, I. Antonoglou, H. King,D. Kumaran, D. Wierstra, S. Legg, , and D. Hassabis, “Human-level control through deep reinforcement learning,” Nature, vol.518, no. 7540, pp. 529–533, February 2015.

[105] K. Cho, B. van Merrienboer, C. Gulcehre, D. Bahdanau,F. Bougares, H. Schwenk, and Y. Bengio, “Learning phrase rep-resentations using RNN encoder–decoder for statistical machinetranslation,” in Proc. Conf. Emp. Methods Natural Lang. Process.(EMNLP), Oct. 2014, pp. 1724–1734.

[106] J. Chung, C. Gulcehre, K. Cho, and Y. Bengio, “Empirical evalua-tion of gated recurrent neural networks on sequence modeling,”in NIPS 2014 Wksh. on Deep Learning, 2014.

[107] F. Qian, L. Ji, B. Han, and V. Gopalakrishnan, “Optimizing 360◦video delivery over cellular networks,” in Proc. 5th Workshopon All Things Cellular: Operations, Applications and Challenges, ser.ATC ’16, New York, NY, USA, 2016, pp. 1–6.

Mohammed S. Elbamby received the B.Sc. de-gree (Hons.) in Electronics and CommunicationsEngineering from the Institute of Aviation En-gineering and Technology, Egypt, in 2010, andthe M.Sc. degree in Communications Engineer-ing from Cairo University, Egypt, in 2013. Heis currently pursuing the Dr.Tech. degree withthe University of Oulu. After receiving the M.Sc.degree, he joined the Centre for Wireless Com-munications, University of Oulu. His research in-terests include resource optimization, uplink and

downlink configuration, fog networking, and caching in wireless cellularnetworks. He received the Best Student Paper Award from the EuropeanConference on Networks and Communications in 2017.

Cristina Perfecto (S’15) received her B.Sc.and M.Sc. in Telecommunication Engineeringfrom the University of the Basque Country(UPV/EHU) in 2000. She is currently a col-lege associate professor with the Department ofCommunications Engineering at the UPV/EHU.Her research interests lie on millimeter wavecommunications and in the application of ma-chine learning in 5G networks. She is currentlyworking towards her Ph.D. focused on the ap-plication of multidisciplinary computational intel-

ligence techniques in radio resource management for 5G.

Chen-Feng Liu (S’17) received the B.S. degreefrom National Tsing Hua University, Hsinchu,Taiwan, in 2009, and the M.S. degree in com-munications engineering from National ChiaoTung University, Hsinchu, in 2011. He is cur-rently pursuing the Ph.D. degree with the Uni-versity of Oulu, Oulu, Finland. In 2012, he joinedAcademia Sinica as a Research Assistant. In2014, he was a Visiting Researcher with theSingapore University of Technology and Design,Singapore. He was also a Visiting Ph.D. Student

with the University of Houston, Houston, TX, USA, and New York Uni-versity, New York, NY, USA, in 2016 and 2018, respectively. His currentresearch interests include 5G communications, mobile edge computing,ultra-reliable low latency communications, and wireless artificial intelli-gence.

Jihong Park received the B.S. and Ph.D. de-grees from Yonsei University, Seoul, South Ko-rea, in 2009 and 2016, respectively. From 2016to 2017, he was a Post-Doctoral Researcherwith Aalborg University, Denmark. He was aVisiting Researcher with Hong Kong Polytech-nic University; KTH, Sweden; Aalborg University,Denmark; and New Jersey Institute of Technol-ogy, USA, in 2013, 2015, 2016, and 2017, re-spectively. He is currently a Post-Doctoral Re-searcher with the University of Oulu, Finland.

His research interests include ultra-dense/ultra-reliable/massive-MIMOsystem designs using stochastic geometry and network economics. Hispapers on tractable ultra-dense network analysis received the IEEEGLOBECOM Student Travel Grant in 2014, the IEEE Seoul SectionStudent Paper Contest Bronze Prize in 2014, and the 6th IDIS-ETNEWS(The Electronic Times) Paper Contest Award sponsored by the Ministryof Science, ICT, and Future Planning of Korea.

20

Sumudu Samarakoon (S’08-AM’18) receivedhis B. Sc. Degree (Hons.) in Electronic andTelecommunication Engineering from the Uni-versity of Moratuwa, Sri Lanka in 2009, the M.Eng. degree from the Asian Institute of Tech-nology, Thailand in 2011, and Ph. D. degree inCommunication Engineering from University ofOulu, Finland in 2017. He is currently working inCentre for Wireless Communications, Universityof Oulu, Finland as a post doctoral researcher.His main research interests are in heteroge-

neous networks, small cells, radio resource management, reinforcementlearning, and game theory. In 2016, he received the Best Paper Award atthe European Wireless Conference and Excellence Awards for innova-tors and the outstanding doctoral student in the Radio Technology Unit,CWC, University of Oulu.

Xianfu Chen received his Ph.D. degree in Sig-nal and Information Processing, from the De-partment of Information Science and ElectronicEngineering at Zhejiang University, Hangzhou,China, in March 2012. He is currently a SeniorScientist with the VTT Technical Research Cen-tre of Finland Ltd, Oulu, Finland. His researchinterests cover various aspects of wireless com-munications and networking, with emphasis onnetwork virtualization, software-defined radio ac-cess networks, green communications, central-

ized and decentralized resource allocation, dynamic spectrum access,and the application of machine learning to wireless communications.

Mehdi Bennis (S’07-AM’08-SM’15) received hisM.Sc. degree in Electrical Engineering jointlyfrom the EPFL, Switzerland and the EurecomInstitute, France in 2002. From 2002 to 2004,he worked as a research engineer at IMRA-EUROPE investigating adaptive equalization al-gorithms for mobile digital TV. In 2004, he joinedthe Centre for Wireless Communications (CWC)at the University of Oulu, Finland as a researchscientist. In 2008, he was a visiting researcherat the Alcatel-Lucent chair on flexible radio, SU-

PELEC. He obtained his Ph.D. in December 2009 on spectrum sharingfor future mobile cellular systems. Currently Dr. Bennis is an AssociateProfessor at the University of Oulu and Academy of Finland researchfellow. His main research interests are in radio resource management,heterogeneous networks, game theory and machine learning in 5Gnetworks and beyond. He has co-authored one book and publishedmore than 100 research papers in international conferences, journalsand book chapters. He was the recipient of the prestigious 2015 FredW. Ellersick Prize from the IEEE Communications Society, the 2016Best Tutorial Prize from the IEEE Communications Society and the2017 EURASIP Best paper Award for the Journal of Wireless Commu-nications and Networks. Dr. Bennis serves as an editor for the IEEETransactions on Wireless Communication.