Ready Player One: UAV Clustering based Multi-Task Offloading for Vehicular VR/AR Gaming
Long Hu, Yuanwen Tian, Jun Yang, Tarik Taleb, Lin Xiang, Yixue Hao
UUNDER REVIEW: IEEE NETWORK, VOL. XX, NO. YY, MONTH 20XX 1
Ready Player One: UAV Clustering basedMulti-Task Offloading for Vehicular VR/ARGaming
Long Hu, Yuanwen Tian, Jun Yang, Tarik Taleb, Lin Xiang, Yixue Hao
Abstract —With rapid development of unmanned aerial vehicle(UAV) technology, application of the UAVs for task offloading hasreceived increasing interest in the academia. However, real-timeinteraction between one UAV and the mobile edge computing(MEC) node is required for processing the tasks of mobile endusers, which significantly increases the system overhead and isunable to meet the demands of large-scale artificial intelligence(AI) based applications. To tackle this problem, in this article, wepropose a new architecture for UAV clustering to enable efficientmulti-modal multi-task task offloading. By the proposed architec-ture, the computing, caching and communication resources arecollaboratively optimized using AI based decision-making. Thisnot only increases the efficiency of UAV clusters, but also providesinsight into the fusion of computation and communication.
Index Terms —Task Offloading, Unmanned Aerial Vehicle,Artificial Intelligence, Cognitive Computing
I. I
NTRODUCTION
With rapid advances in mobile computing and wirelesscommunication technologies, the demands of mobile end usershave been largely met by deploying edge computing [1] andother solutions on the ground, while employing Internet of Ve-hicles as a supplementary infrastructure to accommodate un-manned applications [2], [3]. Recently, researchers have shownincreasing interest towards processing in the air complex tasksas automatic cruise, aerial photography, and precision targetidentification. However, traditional architecture for enablingcollaboration between one unmanned aerial vehicle (UAV)and mobile edge computing (MEC) is not applicable. Firstly,the miniaturization of UAV severely limits its computation,caching and communication (3C) capabilities. That is, for thesame cost, the computing capabilities of UAVs are inferior tothose of autonomous vehicles on the ground. Secondly, due tofrequent interactions between the UAV and the MEC duringprocessing tasks of mobile end users, the battery power ofthe UAV will be drained rapidly, resulting in low processingefficiency. Thus, the architectures of the ground-based systemscannot be directly applied in aerial systems, highlighting thedemand for new computing architectures in UAV scenes.
L. Hu, Y. Tian, J. Yang and Y. Hao are with School of Computer Science andTechnology, Huazhong University of Science and Technology, China. (Email:[email protected], [email protected], junyang [email protected])T. Taleb is with Aalto University, Finland, Centre for Wireless Com-munications (CWC), University of Oulu, Finland, and The Computer andInformation Security Department, Sejong University, South Korea. (Email:tarik.taleb@aalto.fi)L. Xiang is with University of Luxembourg, Luxembourg. (Email:[email protected])Yixue Hao is the corresponding author. (Email: [email protected])
Due to their flexible deployment, UAVs have attractedextensive research activities. For example, researchers haveinvestigated cooperating UAVs for providing communicationcoverage. In particular, Motlagh et al. [4] investigate a UAV-aided MEC system for crowd surveillance scene. Mozaffari etal. [5] propose an efficient deployment scheme for providingcoverage to ground users by exploiting multiple UAVs aswireless base stations. Lyn et al. [6] propose a UAV-aided hy-brid network architecture to assist ground base station (GBS),which can exploit UAV-aided offloading for both throughputgains and cost savings. However, these works [4]– [6] havenot explored the role of UAV clusters nor the applicationof artificial intelligence (AI) technology. Other researchershave considered the problem of one UAV processing AI taskssuch as disaster relief task, precision target identification,etc. For example, Zhao et al. [7] propose a deep learningalgorithm for applying a UAV to identify wildfire. However,for wide-range mountain fires in reality, e.g., the large-scaleCalifornia mountain fires in November 2018, UAV clustersneed to complete the disaster relief task in a timely manner.Schwarzrock et al. [8] propose an efficient task allocationscheme for UAV clusters based on the swarm intelligence.The task investigated in [8] can be decomposed into computa-tion, caching and communication to achieve the collaborativeoptimization of resources. The aforementioned works [4]– [8]have promoted the development of UAV technology. However,in the scenes of large-scale mobile users, heavy task loadwill lead to high delay. To tackle this challenge, we proposea UAV collaboration framework to offload multiple complextasks and consider the coordination of computation, cachingand communication resources [9], where the efficiency of UAVteams is maximized using AI based decisions.We consider the virtual reality/augmented reality (VR/AR)gaming scene as shown in Fig. 1. With the development ofAI technology, the number of mobile users and the demandfor high-quality user experiences are increasing rapidly. Inthe VR/AR hybrid gaming scene described in “Ready PlayerOne” [10], the driver and the passengers may enjoy a real-time experience from augmented visual effects while the caris moving at high speed in the physical environment. Wearabledevices [11] can be utilized to augment user experience.However, as the user distribution is changing dynamicallyin real-time, deploying static/fixed edge computing nodes inthe state-of-the-art networks fails to meet their computingdemands. As a result, a large number of computing tasks maypile up in hot spots. UAVs flexibly deployed for tracing the a r X i v : . [ c s . N I] A p r NDER REVIEW: IEEE NETWORK, VOL. XX, NO. YY, MONTH 20XX 2
Fig. 1. Illustration for vehicular VR/AR gaming scene enabled by UAV clustering based multi-task offloading mobile users provide a promising solution to tackle this issue.Based on the real-time high resolution videos in peripheralphysical scenes, UAVs can facilitate virtual scene processingand provide the users with personalized experience. However,UAVs are expensive and a large number of UAVs may causestrong mutual interference in the air. Therefore, enhancing theefficiency of UAVs is crucial for meeting the requirements oflarge-scale mobile users and, at the same time, guaranteeinghigh-quality user experience. Several important characteristicsof the proposed architecture are listed as follows. • Multi-task offloading:
The traditional scheme [7] onlyconsiders a single UAV for processing a single task.For VR/AR applications, although the UAV-aided MECsystem [4] can mitigate this problem, it depends heavilyon the infrastructure and, hence, is not applicable herein.In contrast, by our proposed scheme, one UAV canserve multiple tasks. In particular, the results of eachtask can be partially reused to serve other tasks in anopportunistic manner. As a result, the proposed schemecan significantly enhance the efficiency of UAVs on alarge scale. • Collaboration of UAV clusters:
By our proposedscheme, one task can be jointly processed using multipleUAVs. The UAV network consists of multiple dynamicresources. Considering that each UAV may have differentloads while processing different tasks, the computationresources of the idle UAVs can be shared with to over- loaded UAVs to improve resource utilization. • Joint optimization of computing, caching and com-munication resources:
The completion of one task issuccessful only if sufficient computing, communicationand storage resources are available. The VR/AR taskmay easily fail in the traditional single-UAV scene asthe resources are fixed and limited. By considering UAVcollaborations in multi-task offloading scenes, the UAVclusters form a dynamic resource pool. Meanwhile, theircomputing, caching and communication resources can beshared with each other, in a dynamic and flexible manner,to balance the utilization of resource. • AI based decision-making:
By the traditional scheme,the interaction between one UAV and MEC causes heavyoverhead. When considering multiple-UAVs collaborativeoperation, each UAV must perceive and forecast the mo-bility of neighboring UAVs and the dynamic resources ofthe UAV network, before the task offloading decisions aremade on this basis. This causes many open problems suchas high delay, task failure. In fact, considering UAV coop-eration in multi-task offloading scenes can achieve jointoptimization of computing, caching and communicationresources of UAV clusters. Moreover, AI based decision-making is crucial to enhance the utilization of availableresources for maximization of the system performance.The contributions of this work are as follows. • We investigate the fusion of computation and communi-
NDER REVIEW: IEEE NETWORK, VOL. XX, NO. YY, MONTH 20XX 3 cation in UAV networks. The current research of mobileUAV networks has only focused on the communicationaspects of UAVs or the task-processing capability of oneUAV. Different from the UAV literature, we considerthe entire large-scale applications of vehicular VR/ARgaming, and propose a new research direction by thefusion of two fields. • Moreover, we construct a novel architecture, called UAV-M3T, for UAV clusters to collaboratively perform dif-ferent tasks. Under this architecture, the trajectory, taskoffloading and network resource allocation for the coop-erating UAVs within the clusters can be jointly optimized. • Finally, we propose an AI based decision-making frame-work to facilitate UAV cooperation and joint optimizationof computing, caching and communication resources. Inthis framework, deployments of UAV clusters both inadvance based on historical data mining and in real-time based on real-time perception are considered. Ex-perimental evaluation reveals that our proposed strategycan effectively improve the collaboration of UAV clusters.In the remainder of this article, we present the proposedUAV collaboration architecture for multiple task scenarios inSection II. Moreover, in Section III we introduce the resourcecoordination method for cooperating UAVs and discuss itsadvantages. Furthermore, the dynamic deployment schemeand its experimental evaluation are elaborated in Section IV.Finally, Section V concludes the paper and discusses someinteresting future work.II. UAV
CLUSTERING BASED M ULTI -M ODAL M ULTI -T ASK (UAV-M3T) O
FFLOADING A RCHITECTURE
A. Architecture of UAV-aided MEC task offloading
Recently, the task-processing mode of UAV-aided MECnetworks has been proposed in [4], whereby the computingtask beyond a user’s processing capability is offloaded to theUAV. The architecture is illustrated in Fig. 2. If the UAVhas limited computing capability available, the task is thenoffloaded to the ground MEC server for processing. In thelatter case, the UAV is used as a repeater to efficiently offloadof the user’s computing task to the MEC server when e.g. theuser has a poor communication connection to the MEC server.However, the UAV-aided MEC networks may fail to meetthe users’ required quality of experience in several key scenes.For example, for the VR/AR application scene in wild, desert,and complex topographies, the ground MEC network maynot be conveniently and reliably built. If the ground MECsystem is absent, the UAV-assisted MEC network architecturefails to promptly address the situation. On the other hand,even if the ground MEC exists, the users may distribute ina large area such that it is difficult to fully offload the tasksto the static/fixed MEC, which degrades the users’ quality ofexperience. To tackle these issues, it is necessary to adopta flexible task processing architecture based on e.g. UAVclustering.
B. Architecture of UAV-M3T task offloading
We construct a novel architecture, called UAV-M3T, forUAV clusters to collaboratively perform different tasks. Thearchitecture is illustrated in Fig. 3.1)
UAV-O2O Mode (One UAV to One Task) : The simplestmode in UAV-M3T offloads one user task to one UAV thathas sufficient computing, caching and communicationcapabilities for processing. This mode has the lowestcost but can still fully exploit the advantages of UAVclusters. We note that the one UAV-aided MEC servicemode discussed in Section II-A is essentially a result ofintroducing the MEC server as backup resources into theUAV-O2O mode.2)
UAV-O2M Mode (One UAV to Multi-Task) : The UAV-O2M mode differs from the UAV-O2O mode in that theformer does not process the tasks separately, but can reusethe tasks fo improve the users’ quality of experience. Anexample of the UAV-O2M mode is illustrated in Fig. 3. Ifthe UAV-O2O mode is adopted, the tasks of users Adam,Bob and Cindy will be processed by UAVs A, B andC, respectively. This significantly reduces the processingefficiency of the UAVs. For example, the data collectiontasks from a group of neighboring users within the sametime window are usually the same. Therefore, the multi-user data collection task can be delegated to one UAV forsaving computing resources. As shown in Fig. 3, sinceAdam and Bob are in the same area, the data collectedat UAV A can be transmitted to UAV B for computing,while the computation results of UAV B can be directlyfed back to and used at both Adam and Bob. In thisway, the UAV-O2M mode utilizes the technique of taskresuing.3)
UAV-M2O Mode (Multi-UAV to One Task) : In theUAV-M2O mode, multiple UAVs collaboratively processone task. As shown in Fig. 3, the VR/AR gaming taskof user Adam is allocated to UAVs A, B and C for jointprocessing. In particular, the landscape data collected byUAV A is first transmitted to UAV B for processing. IfUAV B has only limited computing resources and failsto serve all the task requests of Adam, a portion of thetasks will be then offloaded to UAV C. Finally, UAV Cwill utilize its idle computing resources to process thetask of Adam jointly with UAV B. As a result, the M2Omode can efficiently utilize the network resources of theUAV clusters by enabling cooperation among neighboringUAVs. This significantly improves the users’ quality ofexperience and leads to efficient resource allocation.4)
UAV-M2M Mode (Multi-UAV to Multi-Task) : TheUAV-M2M hybrid service mode combines the UAV-O2Mmode and the UAV-M2O mode. The hybrid service modeis the most common mode of UAV clusters cooperationin processing multi-task scenes. Multi-agent system hasbeen investigated in communication systems [12], alongwith agent-based implementation on smart objects in IoTsystems [13]. By adopting the UAV-M2M mode, thetrajectory, task offloading and network resource allocationfor the cooperating UAVs within the clusters can be
NDER REVIEW: IEEE NETWORK, VOL. XX, NO. YY, MONTH 20XX 4
Fig. 2. Architecture of UAV-aided MEC task offloadingFig. 3. Architecture of UAV clustering based multi-modal multi-task offloading (UAV-M3T)TABLE IP
ERFORMANCE COMPARISON OF THREE ARCHITECTURES : MEC
WITHOUT
UAV, UAV-
AIDED
MEC
AND
UAV=M3TArchitecture Deploymentdynamicity Resourceflexibility Real-timeresponse Intelligentdecision-making Cost User’s quality ofexperienceMEC without UAV N/A Medium Medium Limited Low LowUAV-aided MEC Limited Medium Medium Limited Medium MediumUAV-M3T High High High High High High
NDER REVIEW: IEEE NETWORK, VOL. XX, NO. YY, MONTH 20XX 5 jointly optimized.Research into UAV-M3T architecture is promising for futureAI based applications. Although the UAV-M3T architecture inthe hybrid service mode has a relatively high deployment cost,it can significantly improve users’ quality of experience andprovide brand new market returns for the service provider.Table I presents a comparison of the three architectures. Infact, several projects on facilitating the UAV-aided MEC ap-plications have been recently launched by Google, Facebook,Amazon and Huawei. It is expected that the deploymentcost of UAV clusters will be continuously reduced in thefuture. Moreover, the advent of advanced Beyond 5G (B5G)technology will facilitate a widespread deployment of UAVsto meet users’ rising requirements on quality of experience.III. C
OORDINATION OF C OMPUTING , C
ACHING AND C OMMUNICATION R ESOURCES
The key performance indices of UAVs include capacity, de-lay, energy, reliability, and cost, etc. The quality of experiencemeasures customer’s satisfaction level, which depends on thepersonal preference of the user, environment and service. Dur-ing the task processing, the actual tasks themselves are multi-modal. Due to their heterogeneity, different tasks demand fordifferent computation, caching, and communication resources.In the proposed system, the deployment of computation,caching, and communication resources using UAV clusters hasadvantages in the following aspects. • Amount of information collected:
Even if UAVs servedifferent independent objects, the collected informationcan be highly redundant due to the requirements ofthe same business such as the VR/AR gaming scene.Therefore, the data validity can be enhanced by meansof data reusing, content caching and task migration etc. • Real-time performance:
For UAV-aided MEC architec-ture, a large quantity of information collected by theUAVs needs to be transmitted back to the ground withoutcompression. This causes communication disruptions andfails the tasks when the bandwidth is insufficient. For themulti-UAV clustering based collaboration architectures,many tasks are compressed and processed in real-timeduring the UAVs’ flight before been offloaded, thereforecan reduce the communication delay of data transmission. • Decision capability:
Due to its limited computing,caching and communication capabilities, a single smallUAV can only support limited network decision-making.UAV-M3T architecture can realize decision based onnetwork resources and mobility, as stated in Section IV.Next, for the performance of decisions, the tasks withvery high requirements on performance can be completedby ensemble learning. However, one UAV deploying theensemble learning fails to meet the user requirements ofreal-time performance due to high computing cost. • Efficiency:
For complex application scenes such asVR/AR gaming scenes, enhancing the efficiency of UAVwill reduce costs. On the one hand, efficient data col-lection and task processing can be achieved by taskreusing, content caching and other strategies. On the other hand, UAV clusters collaboration including multi-UAV data collection, resource allocation coordination,and intelligent decision-making can enhance the overallresource efficiency of UAV clusters.By considering the cooperation between UAVs in multi-user scenes, we can achieve efficient sharing of computation,caching, and communication resources among the UAVs toincrease the system throughput. Meanwhile, the data and sig-naling exchanges between cooperating UAVs can be reformedusing e.g. device-to-device (D2D) connections. The price isan increased transmission delay as the resources need to beoffloaded to other terminals using the D2D communicationbetween UAVs. Thus, in case of multiple users, the trade-off between the cooperation gains and the resulting systemoverhead needs to be investigated. For this purpose, we assumethat UAV clusters within the same organization are connectedby D2D and that one user’s task is completed by a designatedUAV. For notational convenience, we assume that only oneuser requests a VR/AR gaming task. Let r ti,j be the commu-nication data rate between UAVs i and j . Moreover, β ti,j and κ ti,j are the amounts of computation offloading and cachingcontent conveyed from the UAV i to UAV j , respectively. Forthe considered VR/AR gaming scene, we optimize the averagedelay of the UAVs subject to the energy capacity of eachUAV. We denote the UAV serving the requesting terminal as“master” UAV of the task and the UAV connecting the masterUAV to provide 3C resources as the “slave” UAV. The delay D UE of the requesting user terminal accounts for both theaverage computating delay, which includes the computationdelays in the master and slave the UAVs and the latencyof D2D connection setup, and the average communicationdelay, which is the time needed to offload the tasks betweendifferent nodes. Moreover, E UAVi denotes energy consumptionof task computing and D2D transmission of
U AV i . E MAXi denotes the maximum energy for
U AV i . The resulting resourceallocation for UAV collaboration optimization problem isformulated as follows, min r,β,κ D UE s.t. E UAVi ≤ E MAXi , i = 1 , , . . . , n . The solution of such optimization problem has been inves-tigated in [14]. We can adopt online algorithms to determinethe optimal resource allocation when collaboration betweenUAVs is enabled. Furthermore, this problem formulating canbe extended to include collaboration between UAV clusters.When multiple tasks arrive simultaneously, the UAV clusterscan collaboratively optimize their 3C resources in the samemanner as above.IV. D
YNAMIC D EPLOYMENT S TRATEGY OF
UAVC
LUSTERING B ASED ON I NTELLIGENT D ECISIONS
A. Deployment in advance based on historical data mining
The task assignment and resource allocation for UAVs canbe deployed a priori before the actual task requests are known,as shown in the right part of Fig. 4. This pre-deployment
NDER REVIEW: IEEE NETWORK, VOL. XX, NO. YY, MONTH 20XX 6
Fig. 4. Dynamic deployment strategy for UAV clustering enhances the user experience because it can improve thenetwork capacity and reduce the likelihood of network con-gestion. For the VR/AR gaming scene, the tasks requested bydifferent users in the same location usually contain redundantinformation about the physical environment. Hence, historicaland social data can be utilized via data mining to forecastthe demands. If the data mining result indicates that a largenumber of users make similar task requests within a timewindow, the UAV can cache the results and reuse them toserve the user demands at subsequent times. In this way, thedelay and energy consumption are reduced simultaneously.Next, the mobility of the UAVs can be forecasted peri-odically by collecting the trace data of the UAV clusters inthe historical time period to optimize resource allocation. For3C resource coordination in multi-task scenes, the resourcesavailable at a given UAV and at its neighboring UAVs shouldbe jointly considered. When the mobility of UAVs is high, theconnection between UAVs may be interrupted due to increasedlikelihood of link outage. In this case, dynamic adjustment of3C resources is crucial to improve the efficiency of resourceallocation.
B. Deployment in real-time based on real-time perception
However, the historical and social data cannot accuratelyforecast the user demands due to their dynamic nature. Thus,adaptive adjustment of the UAV clustering via real-timescheduling must also be made based on real-time perceptionto improve the network capacity and the user experience,as shown in the left part of Fig. 4. Sine each UAV in theUAV clusters may have a different path, the data perceivedand the knowledge learned at the UAVs are different. In thisway, the analysis of mutual information between UAVs canbe enhanced using machine learning. When communication,computation and storage capacities change dynamically, therelevant real-time network status should be further analyzed inreal time such that more UAVs will be sent to the hot spots.The real-time collaboration and tracking optimization areconducted by several UAVs to balance the network resources.By considering multi-modal data, our optimization problem
UAV 1 D a t a P a c k a g e N u m b e r ( P e r S ) UAV 2
UAV 3
Fig. 5. Real-time task load forecasting for three UAVs collaboration is generally non-convex due to non-convex constraints. Thetraditional optimization scheme has long task duration. Inview of high complexity of deep reinforcement learning, weshould design the lightweight deep reinforcement learningalgorithm for decision-making of UAV clusters. In our futurework, we will try combing the Lyapunov optimization anddeep reinforcement learning to further improve 3C resourceallocation.
C. Experiment: LSTM-based multi-UAVs load forecasting
We adapt the real-time load forecasting and investigate thecooperation and resource coordination among multiple UAVs.We then investigate on how multiple UAVs coordinate theresources with limited communication resource between eachother. In real scenes, players need to be served by multi-UAVs in UAV clusters simultaneously to carry out globalresource scheduling on UAV clusters and estimate the loadof the UAV nodes in the next time period. The UAV nodedefines a time series data, which can be forecasted usingrecurrent neural network (RNN) model. It has been shownthat RNNs are able to analyze deep semantic expressionand time series information in data mining. However, RNNssuffer from a poor forecast capability when the resource loadchanges at large rate. To maintain the long-term memory ofthe RNN, we use long short term memory (LSTM) network toeliminate the dependence of the forecast model on abnormaldata. In the experiment, we forecast the changes of the UAVcommunication load state as an example.Fig. 5 shows the communication load state serving 3 UAVsto one player. The communication load changes of each UAVin the next time period are forecasted using the data of an hourprior to current time point as the reference, and forecast thecurrent load change trend. After obtaining the communicationload trend of each UAV, player divides the flexible task into aseries of subtasks. Meanwhile, coordination of 3C resources isconsidered during task offloading to reduce the task delay andenergy consumption. From our selected time window, around250 second in particular, it can be observed that UAV 2 was
NDER REVIEW: IEEE NETWORK, VOL. XX, NO. YY, MONTH 20XX 7 trying to share some task load from UAV 1 while UAV 3stayed relatively stable.V. C
ONCLUSION AND FUTURE WORK
In this paper, we propose a new architecture, referred to asUAV-M3T, for vehicular VR/AR gaming. The UAV-M3T ar-chitecture utilizes AI based decision making for collaborativeoptimization of the UAV team and the network resources and,hence, improves the task performance and resource efficiencyof the UAVs. Our proposed scheme has extensive applicationsin the military industry as well as city and business appli-cations. However, many research challenges also need to betackled. For example, we should consider improving resourcecoordination of UAVs in more complex scenes such as taskmigration [15] and investigate efficient algorithms for dynamicdeployment of UAV clusters, which are left as future work.A
CKNOWLEDGEMENT
This work was supported by the National Natural ScienceFoundation of China (Grant61802138, Grant 61802139), theChina Postdoctoral Science Foundation (No. 2018M632859).This work was partially supported by the Academy of Fin-land 6Genesis Flagship (Grant No. 318927) and the Primo-5G project, that has received funding from the EuropeanUnions Horizon 2020 Research and Innovation Programmeunder Grant Agreement No.815191. The work of L. Xiang issupported by the European Research Council (ERC) projectAGNOSTIC. R
EFERENCES[1] M. Chen, Y. Hao, “Task Offloading for Mobile Edge Computing inSoftware Defined Ultra-dense Network”,
IEEE Journal on Selected Areasin Communications , vol. 36, No. 3, pp. 587–597, Mar. 2018.[2] M. Chen, Y. Tian, G. Fortino, J. Zhang and I. Humar, “Cognitive Internetof Vehicles”,
Computer Communications , vol. 120, pp. 58–70, 2018.[3] H. Hellaoui, O. Bekkouche, M. Bagaa, and T. Taleb, “Aerial ControlSystem for Spectrum Efficiency in UAV to Cellular Communications”,
IEEE Communications Magazine , vol. 56, no. 10, pp. 108–113, Oct. 2018.[4] N.H. Motlagh, M. Bagaa, and T. Taleb, “UAV-based IoT Platform: ACrowd Surveillance Use Case”,
IEEE Communications Magazine , vol.55, no. 2, pp. 128–134, Feb. 2017.[5] M. Mozaffari, W. Saad, M. Bennis and M. Debbah, “Efficient Deploymentof Multiple Unmanned Aerial Vehicles for Optimal Wireless Coverage”,
IEEE Communications Letters , vol. 20, no. 8, pp. 1647–1650, Aug. 2016.[6] J. Lyu, Y. Zeng and R. Zhang, “UAV-aided Offloading for CellularHotspot”,
IEEE Transactions on Wireless Communications , 2018.[7] Y. Zhao, J. Ma, X. Li and J. Zhang, “Saliency Detection and DeepLearning-Based Wildfire Identification in UAV Imagery”,
Sensors , vol.18, no. 3, 2018.[8] J. Schwarzrock, I. Zacarias, A. L.C. Bazzan, et al., “Solving taskallocation problem in multi Unmanned Aerial Vehicles systems usingSwarm intelligence”,
Engineering Applications of Artificial Intelligence ,vol. 72, pp. 10–20, 2018.[9] M. Chen, Y. Hao, L. Hu, M. Hossain, A. Ghoneim, “Edge-CoCaCo: To-wards Joint Optimization of Computation, Caching and Communicationon Edge Cloud”,
IEEE Wireless Communications, vol. 25, no. 3, pp. 21–27, June 2018.[10] E. Cline, “Ready player one”, New York: Crown, 2011, ISBN: 978-0307887436.[11] M. Chen, J. Zhou, G. Tao, J. Yang, L. Hu, “Wearable Affective Robot”,
IEEE Access , vol. 6, pp. 64766–64776, 2018.[12] G. Fortino, A. Garro, W. Russo, “An integrated approach for the devel-opment and validation of multi-agent systems”,
International Journal ofComputer Systems Science & Engineering , 20 (4), 259–271. 2005. [13] N.H. Motlagh, T. Taleb, and O. Arrouk, “Low-Altitude Unmanned AerialVehicles-Based Internet of Things Services: Comprehensive Survey andFuture Perspectives”,
IEEE J. on IoT , vol. 3, no. 6, pp. 899-922, Dec.2016.[14] M. Chen, Y. Hao, K. Lin, L. Hu, Z. Yuan, “Label-less Learning forTraffic Control in an Edge Network”,
IEEE Network , vol. 32, no. 6, pp.8–14, 2018.[15] M. Chen, W. Li, G. Fortino, Y. Hao, L. Hu, I. Humar, “A DynamicService-Migration Mechanism in Edge Cognitive Computing”,