Empowering Mobile Edge Computing by Exploiting Reconfigurable Intelligent Surface
IIEEE DRAFT 1
Empowering Mobile Edge Computing by ExploitingReconfigurable Intelligent Surface
Tong Bai,
Member, IEEE , Cunhua Pan,
Member, IEEE ,Chao Han,
Student Member, IEEE , and Lajos Hanzo,
Fellow, IEEE
Abstract —Along with the proliferation of sensors and of smartdevices, an explosive volume of data will be generated. However,restricted by their limited physical sizes and low manufacturingcosts, these wireless devices are typically equipped with limitedcomputational capabilities and battery lives and thus incapableof processing data time-efficiently. To overcome this issue, theparadigm of mobile edge computing (MEC) is proposed, wherewireless devices may offload all or a fraction of their computationtasks to their nearby computing nodes deployed at the networkedge. At the time of writing, the benefits of MEC systems havenot been fully exploited, predominately because the computationoffloading link is still far from the perfect. In this article, wepropose to empower the MEC systems by exploiting the emergingtechnique of reconfigurable intelligent surfaces, which is capableof reconfiguring the wireless propagation environments and henceof enhancing the offloading links. The beneficial role of RISscan be exploited by jointly optimizing both the RISs as wellas communications and computing resource allocations of MECsystems, which imposes new research challenges on the systemicdesign and thus necessitates a specific investigation. Against thisbackground, this article provides an overview of RIS-assistedMEC systems and highlights their four use cases as well astheir design challenges and solutions. Then their advantageousperformance is validated with the aid of a specific case study.Finally, a guide on future research opportunities is elucidated.
I. I
NTRODUCTION
In the forthcoming Internet-of-Things (IoT) era, myriads ofmachines, sensors, and electronics are envisioned to be con-nected through the Internet [1]. By processing and analyzingthe data collected, the IoT technology enables monitoring andcontrol of the physical world in diverse scenarios, such as,automated home appliances, smart communities, autonomousdriving, intelligent transportation, industrial automation, andemergency management. This innovative paradigm facilitatesa novel interaction pattern among “things” and humans, andwill substantially reshape people’s daily lives. Softbank Grouphas predicted that a trillion devices will be connected to theInternet and will create trillion US dollar in value by .In contrast to the conventional cloud computing systemwhere devices usually act as data consumers, say, downloading The financial support of the ERC-Advanced Fellow Grant of QuantComProject is gratefully acknowledged.T. Bai is with the School of Cyber Science and Technology, BeihangUniversity, Beijing 100191, China. (e-mail: [email protected]).C. Pan is with the School of Electronic Engineering and ComputerScience, Queen Mary University of London, London, E1 4NS, U.K. (e-mail:[email protected])C. Han is with the School of Electronic and Information Engineering,Beihang University, Beijing 100191, China (e-mail: [email protected]).L. Hanzo is with the School of Eletronics and Computer science, Universityof Southampton, SO17 1BJ, UK. (e-mail: [email protected]) a Netflix video, the devices in the IoT era become hetero-geneous and immense data are envisioned to be produced.For example, a raw data rate of . / s is required for avirtual reality handset having a × resolution [2],while Instagram users post nearly , new photos everysingle minute. If all raw data were offloaded to the cloud, itwould cause excessive latency due to the long end-to-end delayand severe traffic congestion in the core network. Against thisbackground, the mobile edge computing (MEC) paradigm isproposed for extending the functionality of the conventionalcloud computing towards the edge of the network. Here “edge”is defined as the computing, storage, and networking resourcesalong the path between data sources and cloud data centers.For example, the gateway in a smart home is the edge betweenhome appliances and cloud, while a cloudlet is the edgebetween mobile devices and cloud. With the assistance of edgecomputing nodes, the raw data can be offloaded to edge nodesand then processed or stored without going through the lengthycore networks to the cloud, which may substantially relieve thetraffic congestion of the core network and eliminate the end-to-end delay along the path between the edge and the cloud,especially beneficial to the delay-sensitive applications.At the time of writing, however, the potential of MECsystems has not been fully exploited, predominantly becausethe computation offloading link is far from the perfect. Moreexplicitly, the devices at the cell edge or behind blockagesusually suffer from low offloading rates, which induces longlatency or high energy consumption on computation offload-ing. This low offloading rate also implies that a large fractionof computing resources at MEC servers have to be idle dueto the limited volume of the tasks received, which limits theoverall computation rate of MEC systems. Furthermore, sincethe data offloaded by devices usually contains customers’privacy, the offloading security should also be guaranteed.Therefore, it is imperative to empower MEC systems byenhancing their computation offloading links.The recent advances of programmable meta-materiel [3]enable the manufacturing of reconfigurable intelligent surfaces(RIS), used for enhancing wireless communications systems.An RIS comprises of a large number of low-cost passivereflecting elements, each of which is capable of imposing anamplitude and/or a phase shift to the signal reflected, thuscollaboratively modifying the signal propagation environment.Specifically, if the direct link between wireless devices andedge nodes is in the face of obstacles, the data can be offloadedvia the RIS-aided virtual LoS link. Reflection-based beam-forming can be realized by jointly optimizing the reflection a r X i v : . [ ee ss . SP ] F e b EEE DRAFT 2 coefficients of RISs, both for enhancing the offloading rateof the devices at the cell edge and for ensuring the compu-tation offloading security from the perspective of physical-layer security. Thus, the system performance of MEC canbe substantially improved with the aid of RISs. Furthermore,in contrast to the conventional transceivers, RISs eliminatethe radio-frequency (RF) chain and operate in a near-passivemanner. Hence, they can be densely deployed with scalablecost and low energy consumption, fulfilling the ubiquitousnessof MEC services. Finally, the amalgamation between RISs andexisting MEC systems does not necessitate the redesign of theprotocols and of hardware implementation.The rest of this paper is organized as follows. In Section II,we detail the RIS-assisted MEC system and provide its fouruse cases as well as two pertinent research challenges. Thebeneficial role of RISs in MEC systems is presented with theaid of case study in Section III. In Section IV, we highlight anumber of future research opportunities. Finally, the paper isconcluded in Section V.II. R
ECONFIGURABLE I NTELLIGENT S URFACE E MPOWERING M OBILE E DGE C OMPUTING
In order to shed light on the operation of an RIS-assistedsystem, we commence with a brief introduction to the funda-mentals of RISs. Following this, we elucidate the RIS-assistedMEC system and outline its potential use cases as well as thepenitent challenges and solutions.
A. Fundamentals of RIS
An RIS, also known as an intelligent reflecting surface, is asoftware-controlled artificial surface that can be programmedto alter its electromagnetic response [3]. As shown in Fig. 1,the RIS comprises of an RIS controller and a large number ofpassive reflection elements. Both the amplitude and the phaseshift of the incident signals can be adjusted in each reflectionelement, which is realized by leveraging electronically tunablemeta-atoms, e.g. the electronic devices of positive-intrinsic-negative (PIN) diodes, field-effect transistors (FETs) or micro-electromechanical system (MEMS) switches. Upon adopting amicro-controller, e.g. a field-programmable gate array (FPGA)board, the reflection elements’ electromagnetic responses canbe concurrently adjusted by controlling their switch state anddirect current bias voltages. As such, we may establish a con-trollable communications link, where the wireless propagationenvironment can be adapted in a real-time manner.The comparative advantages of RISs over other alternatives,namely full-duplex relays and active metasurfaces, are elabo-rated on as follows. Full-duplex relays resemble RISs in full-duplex transmission and in exploiting multipath diversity gain.However, since data detection and transmission are operatedsimultaneously, full-duplex relays inevitably suffer from self-interference (SI). By contrast, owing to their intrinsic nature ofpassive reflection, RISs are free from SI, eliminating the spe-cific design of SI cancellation and the corresponding latency.Active metasurfaces [4] are capable of providing exceptionalcontrollability of signals, which however is energy-consuming
Fig. 1: Illustration of a reconfigurable intelligent surface (RIS), which typicallycomprises a field-programmable gate array (FPGA) controller and a largenumber of reflection elements. Here PIN refers to positive-intrinsic- negative.Fig. 2: Illustration of a reconfigurable intelligent surface (RIS)-assisted mobileedge computing (MEC) system, where a fraction of computational tasks areoffloaded from wireless devices to the mobile edge computing node via anaccess point (AP) with the aid of an RIS, for enabling edge computing orboth local computing and edge computing. and incurs high computational complexity for real-time con-figuration. As seen in this discussion, RISs has the advantageon their scalable cost, energy efficiency, and computationalcomplexity, which enables them advisable for MEC systemsfor fulfilling their ubiquitous services.
B. RIS-Assisted MEC Systems
Fig. 2 illustrates an RIS-assisted MEC system, whichcomprises of an access point (AP) connected with the edgecomputing node using high-speed optical fiber, an RIS, and anumber of wireless devices, each of which has specific com-putation tasks to be processed. If a wireless device is incapableof processing all of these tasks due to its limited power supplyand/or restricted computational capability, a fraction of or allof the tasks can be offloaded to the edge computing nodethrough both the direct channel and the RIS-induced reflectionchannel between the device and the AP. The edge computingnode is equipped with an edge manager, used for makingdecisions for scheduling the volume of computation tasksto be offloaded from the wireless devices and for allocatingboth the computation resource at the edge node and thecommunications resource for each wireless device, accordingto both the channel and computation information, such asthe instantaneous channel state information, the computationalcapabilities of wireless devices and of the edge node, as wellas the quality of experience (QoE) required by the specificapplications.
EEE DRAFT 3
Fig. 3: Use Cases of reconfigurable intelligent surface (RIS)-assisted mobile edge computing (MEC) systems, including shortening processing latency, improvingenergy efficiency, enhancing total completed task-input bits, and secure computation offloading.
In contrast to the conventional MEC system where thewireless propagation environment is uncontrollable, the de-ployment of RISs in MEC systems provides a new degree offreedom. Specifically, the reflection coefficients of RISs canbe jointly designed both with the communications and withcomputing resource allocation, for reaping the gap betweenthe performance of the existing MEC systems and the stringentrequirements of the emerging applications.
C. Use Cases
Fig. 3 highlights four potential use cases of RISs in MECsystems, which are detailed as follows.
1) Shortening Processing Latency:
Low latency is requiredby a variety of applications, such as augmented/virtual reality,online gaming, remote desktop, health care, and connectedvehicles. The processing latency in MEC systems is the su-perposition of the delay induced by computation offloading, bycomputing at edge nodes, and by feeding computational resultsback. Utilizing RISs is capable of shortening the processinglatency in all these three stages [5]. To elaborate using anexample of collective perception of environment in intelligenttransportation systems [6], both video and image gleaned byvehicles are required to be processed collaboratively at edgenodes in an almost real-time manner, for facilitating the driverassistance and traffic management. However, vehicles’ compu-tation offloading links are often in the face of obstacles, e.g.trucks and mansions. As shown in Fig. 3, RISs are deployedfor establishing a virtual LoS link, which may substantiallyenhance the offloading rate and thus reduce the offloading delay. RISs also contribute to the reduction of the computingdelay. Specifically, the volume of data offloaded to edge nodesis largely dependent on the channel gain of offloading links.The assistance of RISs enables more data to be offloadedto edge nodes, where the data can be processed more time-efficiently than local computing at the vehicles. Furthermore,the feedback delay can be reduced with the aid of the virtualLoS link again.
2) Improving Energy Efficiency:
In MEC systems, devices’energy efficiency is defined as the ratio of the number of bitsprocessed to the sum of devices’ energy consumption bothon local computing and on computation offloading, which isof paramount importance to the IoT devices equipped withshort battery lives. As shown in Fig. 3, the deployment ofRISs is capable of substantially improving devices’ energy effi-ciency. Specifically, RISs contribute both virtual array gain andreflection-based beamforming gain to computation offloadinglinks. To elaborate, the virtual array gain can be achieved bycombining all the signal reflected from RIS elements, while thereflection-based beamforming gain is realized by proactivelyadjusting both the amplitude of and phase shift of the signalreflected. Hence, compared with the system without deployingRISs, the same offloading rate can be attained by using a loweroffloading power in RIS-assisted MEC systems. Furthermore,the emerging technique of energy harvesting enables wirelessdevices to be charged wirelessly. Hence, an innovative wirelesspowered MEC architecture has been proposed for prolongingthe network life, where the devices’ batteries are replenishedrelying on wireless energy transfer from hybrid APs in the first
EEE DRAFT 4 stage while computation offloading is conducted in the secondstage. Given that the aforementioned gains can be exploited inboth stages, the energy efficiency of the hybrid AP can alsobe improved [7].
3) Enhancing Total Completed Task-Input Bits:
In MECsystems, total completed task-input bits is defined as the totalnumber of bits that can be processed in a time slot given a lim-ited energy budget. It plays a crucial role in indicating whetherthe system considered is capable of supporting computation-intensive applications, such as face/speech recognition and bigdata analysis. In the conventional MEC system, although edgecomputing nodes are deployed in the vicinity of devices forproviding computation services, a large fraction of compu-tation resources in the edge nodes have to be idle, becauselocal computing instead of computation offloading is optedwhen offloading links are imperfect. As shown in Fig. 3, uponrefining the wireless propagation environment relying on RISs,more data can be offloaded to edge nodes, whose powerfulcomputing capabilities can be exploited for supporting theaforementioned computation-intensive applications [8].
4) Secure Computation Offloading:
In IoT networks, thedata offloaded to edge nodes usually contains environmentperception information or people’s personal behaviors, whichis also possibly valuable to a third party. Given that wirelesssignal propagation can be accessed both by authorized usersand by adversaries, offloading links in MEC systems are underthe risk of eavesdropping, which jeopardizes data securityand privacy [9]. Furthermore, the systemic information, suchas equipment ID, communication protocol, and the addressof wireless nodes can also be obtained by the adversariesusing capturing packet tools. To overcome these securityissues in MEC systems, as shown in Fig. 3, RISs can bedeployed in the vicinity of wireless devices or eavesdroppers.On one hand, the reflection coefficients of the RIS can bejointly designed with the multi-user detection (MUD) of theAP for enhancing the legitimate wireless devices’ signal-to-interference-plus-noise ratio (SINR). On the other hand, theSINR at the eavesdropper may be reduced by jointly designingthe reflection responses of the RIS together with a jammer forforming a beam to concentrate the artificial noise receivedat the eavesdropper. Both of or the combination of thesetwo schemes are capable of enhancing the secrecy capacity,thereby ensuring the security of the offloading link [10], froma physical-layer perspective.
D. Challenges and Solutions
Deploying RISs in MEC systems imposes new challengeson its system design, which are discussed as follows, togetherwith their potential solutions.
1) RIS Reflection Coefficients Design:
The optimizationproblem of RIS reflection coefficients is typically a non-convex quadratically constrained quadratic programs subjectto the unit-modulus constraint imposed on the reflection phaseshifts, which can be readily solved by using the conven-tional semi-definite relaxation technique [11]. However, itincurs high computation complexity due to a large numberof randomizations for guaranteeing the approximation of the
Fig. 4: Top view of the locations of the RIS-assisted MEC systems forsimulation settings. optimal objective value. As a compromise, a locally optimalsolution can be obtained with the aid of low-complexityalgorithms in an iterative manner, such as the majorization-minimization algorithm, the complex circle manifold method,and successive convex optimization. Furthermore, given thatthe channel state information (CSI) in RIS-assisted systemsis not always available owing to its high dimension, boththe robust optimization technique and the machine learningalgorithm [12] can be applied in the imperfect CSI scenario.
2) Joint Communications and Computation Optimization:
Since diverse applications are supported by MEC systems,it is imperative to adapt both communications and computa-tion resource allocation to their heterogeneous requirementsin a sophisticated manner. Specifically, the communicationresource allocation includes multiple access schemes, spectraland temporal resource allocation to wireless devices, RISconfiguration, and the design of the MUD matrix at theAP, while the computation resource allocation involves thecomputation offloading decision, the design of the centralprocessing unit (CPU) cycling frequencies of wireless devices,and the computation resource allocation at the edge nodes[13]. Since various variables are optimized in this joint systemdesign, the optimization problem cannot be solved using asingle optimization technique. Hence, a combination of convexand of non-convex optimization techniques have to be invoked.However, if the optimization problem formulated is NP-hard,the emerging techniques of game theory and of machinelearning can be invoked.III. C
ASE S TUDY : E
XPLOITING
RIS
S FOR S HORTENING P ROCESSING L ATENCY IN
MEC S
YSTEMS
As shown in Fig. 4, we consider an RIS-assisted MECsystem where K single-antenna wireless devices may opt foroffloading a fraction of or all their computational tasks to theedge computing nodes via an M -antenna AP with the aid of anRIS having N reflection elements. The distance between theRIS and the AP is denoted by R , while the wireless devices arein the circle associated with a radius of r . It is assumed that thedirect device-AP channel follows Rayleigh fading associatedwith a path loss exponent of . , while the device-RIS andRIS-AP channels are dominated by the LoS link associatedwith a path loss exponent of . . The computation offloadingtakes place over a given frequency band of in the sametime resource. The total number of bits to be processed variesin the range of
250 Kb to
350 Kb . The maximum cycling ratesof the edge node is
50G cycles / s . Both the communications EEE DRAFT 5 L a t e n c y ( m s )
10 15 20 25 30 35 40 45 50N
Optimized Reflection CoefficientsRandom Reflection CoefficientsWithout RIS (a) L a t e n c y ( m s )
200 220 240 260 280 300d (m)
Optimized Reflection CoefficientsRandom Reflection CoefficientsWithout RIS (b)Fig. 5: Simulation results of the latency versus the number of reflection elements and the distance between devices and the AP. (a): d = 280 m ; (b): N = 30 . resource including the MUD matrix and the RIS reflectioncoefficients as well as the computing resources such as thevolume to be offloaded and CPU cycling frequencies at theedge nodes are jointly optimized for shortening the processingdelay. In order to demonstrate the benefit of deploying an RISin MEC systems, we compare the following three schemesunder the setup of M = 5 , K = 4 , R = 300 m and r = 10 m . • Without RIS:
The computation tasks are offloaded via thedirect device-AP link. • Random Reflection Coefficients:
The computation of-floading takes place with the aid both of the direct device-AP and of the reflection-based device-RIS-AP link, wherethe RIS reflection coefficients are set as random values. • Optimized Reflection Coefficients:
The computation of-floading takes place with the aid both of the direct device-AP and of the reflection-based device-RIS-AP link, wherethe RIS reflection coefficients are optimized relying onthe majorization-minimization algorithm.As mentioned in Section II-C1, the processing delay inMEC systems comprises of the delay induced by computationoffloading, by computing at edge nodes, and by computationfeedback. Since computation feedback is usually of a smallvolume, we neglect the time spent on result feedback. Fig. 5ashows the simulation results of the processing latency versusthe number of reflection elements. It can be seen that theperformance gap between the schemes of “Without RIS” andof “RandPhase” increases upon increasing the number ofreflection elements, which implies that RISs are capable ofenhancing the performance of MEC systems even withoutcarefully designing RIS reflection responses. This is mainlydue to the virtual array gain, i.e. collecting the signal reflectedfrom RIS reflection elements. Furthermore, the performancegap between the schemes of “With RIS” and of “RandPhase” is
12 ms when N = 10 , while it becomes
43 ms when N = 50 ,which implies that a well design of RIS reflection responsesmay provide a beamforming gain, and that a higher number of reflection elements leads to an increase of the beamforminggain. Exploiting these two types of gain together, RISs arecapable of eminently reducing the processing latency in MECsystems. Fig. 5b plots the simulation results of the processinglatency versus the distance between wireless devices and theAP. It can be seen that the advantage of deploying RISsgradually increases when the distance between the wirelessdevices and the RIS becomes small, where the device-RIS-APlink dominates the computation offloading, which consolidatesthat a higher gain can be achieved in the near-RIS area.IV. F UTURE R ESEARCH O PPORTUNITIES
The state-of-the-art research in RIS-assisted MEC systemsmainly focuses on the MEC system where a single RISis deployed or a specific performance metric is considered.However, aimed at providing the ubiquitous service, futureMEC systems are envisioned to collaborate with a substantialnumber of RISs, accommodating diverse demands of wire-less devices involved, which necessitates specific systematicdesigns. In this section, the future research opportunities areelucidated as follows.
A. RIS Deployment in MEC Systems
Ideally, ubiquitous MEC services can be provided bydensely deploying both MEC servers and RISs, which how-ever entails excessive hardware expenditure and maintenancecost. Therefore, it is imperative to conceive a cost-efficientdeployment strategy. Given that the coordination between RISsand other network participators is perpetually required, thedistribution of MEC servers, of APs, and of wireless devicesshould be taken into consideration when deploying RISs. Onone hand, given an area served by MEC, only a fraction ofall APs are usually co-located with MEC servers and differentMEC servers typically have diverse computational capabilities.Owing to RISs’ hotspot performance, their locations shouldbe carefully decided for striking the trade-off between the
EEE DRAFT 6 delay caused by the wireless traffic jamming at the APs co-located with MEC servers and that induced by the data trans-mission and by the task scheduling in backhaul connectingthe MEC servers and those APs without equipping with anycomputational resources. On the other hand, the computationdemands are normally non-uniformly distributed, prohibitingthe usage of the conventional homogenenous Poisson pointprocess (HPPP) model for planing both MEC servers andRISs. One possible solution is to leverage the Ginibre α -determinantal point process, for characterizing the clusteringbehaviors [14]. B. Joint Communications and Computation Optimization forHeterogeneous Applications and Wireless Devices
Given a diverse variety of applications and of wirelessdevices in RIS-assisted MEC systems, communications andcomputation designs focusing on a specific requirement areincapable of meeting all their demands. Therefore, a unifiedcommunications and computation design should be proposed.Two facts have to be carefully considered when we conceivethe joint optimization. One is that the applications operateddetermine the options of wireless devices’ offloading deci-sions, which affects the communications resource manage-ment policy. The other is that different applications imposediverse QoE requirements on MEC systems. For example,smart manufacturing exhibits stringent delay specifications,but loose energy consumption requirements, whereas the MECservices for health monitoring are expected to accommodateboth energy consumption and offloading security requirements.In this case, how to jointly optimize both RIS reflectioncoefficients as well as the communications and computationresource allocation remains an open issue, calling for thesystematic exploration of the entire Pareto font.
C. On-Demand Device-RIS-Edge Association
Device association in conventional cellular networks, whichis generally implemented by relying on opting the AP as-sociated with the maximum signal-to-interference-plus-noiseratio (SINR), cannot be directly applied in RIS-assisted MECsystems, mainly for the following reasons. Firstly, owing totheir intrinsic beamforming capabilities, RISs are capable ofboosting up the SINR of the link between a wireless device anda particular AP, especially in the emerging cell-free systems,which introduces a new degree-of-freedom in network asso-ciation. Secondly, the SINR values of offloading links onlydecide upon the computation offloading rate, while wirelessdevices’ computation offloading decision is also jointly deter-mined by the applications being operated, by wireless devices’computational capabilities and battery lives, as well as by theidle states of MEC servers. Thirdly, in practice, the volume ofcomputational tasks may vary extensively over time, leadingto the dynamically fluctuating processing delay or energyconsumption at wireless devices. Therefore, it is imperativeto address the computation-task dynamics by conceiving anon-demand device-RIS-edge association scheme. One possiblesolution is to resemble the user-centric network associationstrategy [15] while considering the specific role of RISs.
D. Specific System Design for Coexistence of Offloading Dataand Conventional Cellular Data
Owing to their ubiquitous services, cellular networks aretypically leveraged for computation offloading in MEC sys-tems. However, this may eminently reshape the conven-tional perception of uplink/downlink utilization. Specifically,downlink wireless traffic dominates the conventional cellularnetworks, while MEC systems mainly rely on the uplinktransmission. Therefore, in order to support the coexistenceof offloading data and of conventional cellular data, it isimperative to conceive a sophisticated full-duplex systems.Fortunately, as a benefit of their intrinsically passive reflectionnature, RISs are ideally capable of enhancing both the uplinkand downlink transmission in a simultaneous manner bycollaborating with full-duplex APs. Bearing in mind that incontrast to cellular users whose QoE is only dependent on thecommunications quality, QoE of MEC users is also dependentboth on computation resource allocation at edge nodes and onwireless devices’ features, e.g. local computational capabilitiesand battery lives, the joint communications and computationdesign necessitates a profound rethinking for fully exploitingthe beneficial role of RISs in the face of coexistence ofoffloading data and conventional cellular data.V. C
ONCLUSIONS
In this article, we discussed the potential of RISs in improv-ing the MEC systems in terms of processing latency, energyefficiency, total completed task-input bits, and secure compu-tation offloading. Particularly, the beneficial role of RISs wasvalidated numerically, by exemplifying the processing delayshortening in MEC systems. Since RIS-assisted MEC systemsremain largely unexplored, we also provide a guide for theirfuture research opportunities.R
EFERENCES[1] T. Hong, W. Zhao, R. Liu, and M. Kadoch, “Space-air-ground iotnetwork and related key technologies,”
IEEE Wireless Communications ,vol. 27, no. 2, pp. 96–104, 2020.[2] B. Han, “Mobile immersive computing: Research challenges and theroad ahead,”
IEEE Communications Magazine , vol. 57, pp. 112–118,Oct. 2019.[3] T. J. Cui, M. Q. Qi, X. Wan, J. Zhao, and Q. Cheng, “Codingmetamaterials, digital metamaterials and programmable metamaterials,”
Light: Science & Applications , vol. 3, no. 10, pp. e218–e218, 2014.[4] S. Hu, F. Rusek, and O. Edfors, “Beyond massive MIMO: The potentialof data transmission with large intelligent surfaces,”
IEEE Transactionson Signal Processing , vol. 66, pp. 2746–2758, May 2018.[5] T. Bai, C. Pan, Y. Deng, M. Elkashlan, A. Nallanathan, and L. Hanzo,“Latency minimization for intelligent reflecting surface aided mobileedge computing,”
IEEE Journal on Selected Areas in Communications ,vol. 38, pp. 2666–2682, Nov. 2020.[6] S. Zhou, Y. Sun, Z. Jiang, and Z. Niu, “Exploiting moving intelligence:Delay-optimized computation offloading in vehicular fog networks,”
IEEE Communications Magazine , vol. 57, pp. 49–55, May 2019.[7] T. Bai, C. Pan, H. Ren, Y. Deng, M. Elkashlan, and A. Nallanathan,“Resource allocation for intelligent reflecting surface aided wirelesspowered mobile edge computing in OFDM systems.” [Online]. Avail-able: https://arxiv.org/abs/2003.05511.[8] Z. Chu, P. Xiao, M. Shojafar, D. Mi, J. Mao, and W. Hao, “Intelligent re-flecting surface assisted mobile edge computing for Internet of Things,”
IEEE Wireless Communications Letters , Early Access.[9] T. Bai, J. Wang, Y. Ren, and L. Hanzo, “Energy-efficient computationoffloading for secure UAV-edge-computing systems,”
IEEE Transactionson Vehicular Technology , vol. 68, pp. 6074–6087, June 2019.
EEE DRAFT 7 [10] X. Lu, E. Hossain, T. Shafique, S. Feng, H. Jiang, and D. Niyato,“Intelligent reflecting surface enabled covert communications in wirelessnetworks,”
IEEE Network , vol. 34, pp. 148–155, Oct. 2020.[11] Z.-Q. Luo, W.-K. Ma, A. M.-C. So, Y. Ye, and S. Zhang, “Semidefiniterelaxation of quadratic optimization problems,”
IEEE Signal ProcessingMagazine , vol. 27, pp. 20–34, May 2010.[12] S. Khan, K. S. Khan, N. Haider, and S. Y. Shin, “Deep-learning-aideddetection for reconfigurable intelligent surfaces.” [Online]. Available:https://arxiv.org/abs/1910.09136.[13] F. Zhou, C. You, and R. Zhang, “Delay-optimal scheduling for IRS-aided mobile edge computing,”
IEEE Wireless Communications Letters ,Early Access.[14] Y. Mao, C. You, J. Zhang, K. Huang, and K. B. Letaief, “A surveyon mobile edge computing: The communication perspective,”
IEEECommunications Surveys & Tutorials , vol. 19, pp. 2322–2358, FourthQuarter 2017.[15] S. Chen, F. Qin, B. Hu, X. Li, and Z. Chen, “User-centric ultra-densenetworks for 5G: Challenges, methodologies, and directions,”