Temporal Information Services in Large-Scale Vehicular Networks … · 2019-03-18 · Physical...

14
218 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 20, NO. 1, JANUARY 2019 Temporal Information Services in Large-Scale Vehicular Networks Through Evolutionary Multi-Objective Optimization Penglin Dai, Member, IEEE, Kai Liu , Member, IEEE, Liang Feng , Member, IEEE, Haijun Zhang, Member, IEEE, Victor Chung Sing Lee, Member, IEEE, Sang Hyuk Son, Fellow, IEEE , and Xiao Wu, Member, IEEE Abstract—Temporal information services are critical in imple- menting emerging intelligent transportation systems. Neverthe- less, it is challenging to realize timely temporal data update and dissemination due to an intermittent wireless connection and a limited communication bandwidth in dynamic vehicular networks. Some previous studies have considered the temporal data dissemination in vehicular networks, but they are limited to the service region, which is inside the coverage of roadside units. To enhance system scalability, it is imperative to exploit the synergic effect of vehicle-to-infrastructure (V2I) and vehicle-to- vehicle (V2V) communications for providing efficient temporal information services in such an environment. With the above motivations, we propose a novel system architecture to enable efficient data scheduling in hybrid V2I/V2V communications by having the global knowledge of network resources of the system. On this basis, we formulate a temporal data upload and dissemination (TDUD) problem, aiming at optimizing two conflict objectives simultaneously, which are enhancing the data quality and improving the delivery ratio. Furthermore, we propose an evolutionary multi-objective algorithm called MO-TDUD, which consists of a decomposition scheme for han- Manuscript received April 23, 2017; revised October 17, 2017 and December 30, 2017; accepted February 4, 2018. Date of publication March 13, 2018; date of current version December 21, 2018. This work was supported in part by NSFC under Grant 61572088, Grant 61603064, Grant 61572156, and Grant 61772436, in part by the Fundamental Research Funds for the Central Universities under Project 2018CDQYJSJ0034, in part by the Chongqing Application Foundation and Research in Cutting- Edge Technologies under Grant cstc2017jcyjAX0026, in part by the ICT Research and Development Program through MSIP/IITP (Resilient Cyber- Physical Systems Research) under Grant 14-824-09-013, in part by the GRL Program through NRF under Grant 2013K1A1A2A02078326, in part by the DGIST Research and Development Program (CPS Global Center) through MSIP, and in part by the Shenzhen Science and Technology Program under Grant JCYJ20170413105929681. The Associate Editor for this paper was F. Chu. (Corresponding authors: Kai Liu; Xiao Wu.) P. Dai and X. Wu are with the School of Information Science and Technology, Southwest Jiaotong University, Chengdu 611756, China (e-mail: [email protected]; [email protected]). K. Liu and L. Feng are with the Key Laboratory of Dependable Service Computing in Cyber Physical Society, Ministry of Education, Chongqing University, Chongqing 400040, China, and also with the College of Com- puter Science, Chongqing University, Chongqing 400040, China (e-mail: [email protected]; [email protected]). H. Zhang is with the Shenzhen Graduate School, Harbin Institute of Technology, Shenzhen 518055, China (e-mail: [email protected]). V. C. S. Lee is with the Department of Computer Science, City University of Hong Kong, Hong Kong (e-mail: [email protected]). S. H. Son is with the Department of Information and Communica- tion Engineering, Daegu Gyeongbuk Institute of Science and Technology, Daegu 42988, South Korea (e-mail: [email protected]). Digital Object Identifier 10.1109/TITS.2018.2803842 dling multiple objectives, a scalable chromosome representa- tion for TDUD solution encoding, and an evolutionary oper- ator designed for TDUD solution reproduction. The proposed MO-TDUD can be adaptive to different requirements on data quality and delivery ratio by selecting the best solution from the derived Pareto solutions. Last but not least, we build the simulation model and implement MO-TDUD for performance evaluation. The comprehensive simulation results demonstrate the superiority of the proposed solution. Index Terms— Vehicular networks, temporal information services, evolutionary multi-objective optimization. I. I NTRODUCTION V EHICULAR networks promise to enhance future intel- ligent transportation systems (ITSs) in terms of safety, efficiency and sustainability via V2V (Vehicle-to-Vehicle) and V2I (Vehicle-to-Infrastructure) communications [1]. Particu- larly, temporal data services are the underling technologies for many types of emerging ITS applications, such as road reservation [2], routing planning [3], and infotainment ser- vices [4], where data services have to be completed within a certain time window [5]. On the other hand, due to the highly dynamic nature of traffic systems, many applications require the access of temporal information where data quality may decrease over time, such as road traffic information, location- based service information, etc. Clearly, to guarantee the service quality, timely update of temporal information is expected [6]. Further, due to distributed RSU resources and dynamic vehicle mobility, data dissemination suffers from unpredictable delay in hybrid V2I/V2V communications. Overall, it is imperative yet non-trivial to design an efficient data service mechanism for temporal information services by exploring the joint effect of V2I and V2V communications in such a dynamic and intermittent connected vehicular communication environment. In previous research, great efforts have been paid on net- work communication and data dissemination issues in vehic- ular networks. For example, some studies [7], [8] focused on designing interference-free approaches for reducing packet collision in V2I and V2V communications. Some other stud- ies [9], [10] mainly focused on designing routing strategies to reduce the delivery delay and improve transmission reliability. A few studies [11]–[13] considered the temporal information 1524-9050 © 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

Transcript of Temporal Information Services in Large-Scale Vehicular Networks … · 2019-03-18 · Physical...

Page 1: Temporal Information Services in Large-Scale Vehicular Networks … · 2019-03-18 · Physical Systems Research) under Grant 14-824-09-013, in part by the GRL Program through NRF

218 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 20, NO. 1, JANUARY 2019

Temporal Information Services in Large-ScaleVehicular Networks Through Evolutionary

Multi-Objective OptimizationPenglin Dai, Member, IEEE, Kai Liu , Member, IEEE, Liang Feng , Member, IEEE,

Haijun Zhang, Member, IEEE, Victor Chung Sing Lee, Member, IEEE,Sang Hyuk Son, Fellow, IEEE, and Xiao Wu, Member, IEEE

Abstract— Temporal information services are critical in imple-menting emerging intelligent transportation systems. Neverthe-less, it is challenging to realize timely temporal data updateand dissemination due to an intermittent wireless connectionand a limited communication bandwidth in dynamic vehicularnetworks. Some previous studies have considered the temporaldata dissemination in vehicular networks, but they are limitedto the service region, which is inside the coverage of roadsideunits. To enhance system scalability, it is imperative to exploit thesynergic effect of vehicle-to-infrastructure (V2I) and vehicle-to-vehicle (V2V) communications for providing efficient temporalinformation services in such an environment. With the abovemotivations, we propose a novel system architecture to enableefficient data scheduling in hybrid V2I/V2V communicationsby having the global knowledge of network resources of thesystem. On this basis, we formulate a temporal data uploadand dissemination (TDUD) problem, aiming at optimizing twoconflict objectives simultaneously, which are enhancing thedata quality and improving the delivery ratio. Furthermore,we propose an evolutionary multi-objective algorithm calledMO-TDUD, which consists of a decomposition scheme for han-

Manuscript received April 23, 2017; revised October 17, 2017 andDecember 30, 2017; accepted February 4, 2018. Date of publicationMarch 13, 2018; date of current version December 21, 2018. This workwas supported in part by NSFC under Grant 61572088, Grant 61603064,Grant 61572156, and Grant 61772436, in part by the Fundamental ResearchFunds for the Central Universities under Project 2018CDQYJSJ0034,in part by the Chongqing Application Foundation and Research in Cutting-Edge Technologies under Grant cstc2017jcyjAX0026, in part by the ICTResearch and Development Program through MSIP/IITP (Resilient Cyber-Physical Systems Research) under Grant 14-824-09-013, in part by the GRLProgram through NRF under Grant 2013K1A1A2A02078326, in part by theDGIST Research and Development Program (CPS Global Center) throughMSIP, and in part by the Shenzhen Science and Technology Program underGrant JCYJ20170413105929681. The Associate Editor for this paper was F.Chu. (Corresponding authors: Kai Liu; Xiao Wu.)

P. Dai and X. Wu are with the School of Information Scienceand Technology, Southwest Jiaotong University, Chengdu 611756, China(e-mail: [email protected]; [email protected]).

K. Liu and L. Feng are with the Key Laboratory of Dependable ServiceComputing in Cyber Physical Society, Ministry of Education, ChongqingUniversity, Chongqing 400040, China, and also with the College of Com-puter Science, Chongqing University, Chongqing 400040, China (e-mail:[email protected]; [email protected]).

H. Zhang is with the Shenzhen Graduate School, Harbin Institute ofTechnology, Shenzhen 518055, China (e-mail: [email protected]).

V. C. S. Lee is with the Department of Computer Science, City Universityof Hong Kong, Hong Kong (e-mail: [email protected]).

S. H. Son is with the Department of Information and Communica-tion Engineering, Daegu Gyeongbuk Institute of Science and Technology,Daegu 42988, South Korea (e-mail: [email protected]).

Digital Object Identifier 10.1109/TITS.2018.2803842

dling multiple objectives, a scalable chromosome representa-tion for TDUD solution encoding, and an evolutionary oper-ator designed for TDUD solution reproduction. The proposedMO-TDUD can be adaptive to different requirements on dataquality and delivery ratio by selecting the best solution fromthe derived Pareto solutions. Last but not least, we build thesimulation model and implement MO-TDUD for performanceevaluation. The comprehensive simulation results demonstratethe superiority of the proposed solution.

Index Terms— Vehicular networks, temporal informationservices, evolutionary multi-objective optimization.

I. INTRODUCTION

VEHICULAR networks promise to enhance future intel-ligent transportation systems (ITSs) in terms of safety,

efficiency and sustainability via V2V (Vehicle-to-Vehicle) andV2I (Vehicle-to-Infrastructure) communications [1]. Particu-larly, temporal data services are the underling technologiesfor many types of emerging ITS applications, such as roadreservation [2], routing planning [3], and infotainment ser-vices [4], where data services have to be completed within acertain time window [5]. On the other hand, due to the highlydynamic nature of traffic systems, many applications requirethe access of temporal information where data quality maydecrease over time, such as road traffic information, location-based service information, etc. Clearly, to guarantee the servicequality, timely update of temporal information is expected [6].Further, due to distributed RSU resources and dynamic vehiclemobility, data dissemination suffers from unpredictable delayin hybrid V2I/V2V communications. Overall, it is imperativeyet non-trivial to design an efficient data service mechanismfor temporal information services by exploring the joint effectof V2I and V2V communications in such a dynamic andintermittent connected vehicular communication environment.

In previous research, great efforts have been paid on net-work communication and data dissemination issues in vehic-ular networks. For example, some studies [7], [8] focusedon designing interference-free approaches for reducing packetcollision in V2I and V2V communications. Some other stud-ies [9], [10] mainly focused on designing routing strategies toreduce the delivery delay and improve transmission reliability.A few studies [11]–[13] considered the temporal information

1524-9050 © 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

Page 2: Temporal Information Services in Large-Scale Vehicular Networks … · 2019-03-18 · Physical Systems Research) under Grant 14-824-09-013, in part by the GRL Program through NRF

DAI et al.: TEMPORAL INFORMATION SERVICES IN LARGE-SCALE VEHICULAR NETWORKS 219

services, where the quality of temporal data will degradeover time. They analyzed the data quality in terms of temporalfeatures and proposed efficient scheduling methods for tempo-ral data services with real-time requirements. However, thesestudies only considered the services within RSUs’ coverage.In addition, existing solutions can not be adaptive to dynamicrequirements of vehicular application scenarios by exploitingboth V2I and V2V communication resources.

Distinguished from previous works, this paper focuseson investigating temporal information services in large-scalevehicular networks. Specifically, the concept of “large-scale”refers to both the physical service range and the logical servicetype. First, in terms of the service range, this paper considersthe scenario where multiple RSUs distributed in large-scaleof geographic areas are cooperated to provide informationservices. Second, with respect to the service types, we considerthe hybrid of V2I and V2V communications for providing dataservice cooperatively. Meanwhile, we consider the temporalityfeature of data items, as well as the heavy service workload onboth data uploading and downloading. In particular, passingvehicles can sense temporal information along their drivingtrajectories and upload up-to-date data items to RSUs viaV2I communication. Further, multiple RSUs are connectedvia a backbone network and any data update in the localdatabase of one RSU will be synchronized in other peers.With the above description, there are two main problems tobe investigated. First, due to the temporality of information,data items maintained in RSUs have to be kept up-to-dateby timely uploading from passing vehicles. Second, a datadissemination strategy has to be designed to enhance thedata delivery ratio. To sum up, due to the high mobility ofvehicles and the limited wireless bandwidth, it is non-trivial tosolve the above two problems simultaneously and efficiently.

In this paper, we present a novel service architecture tomaximize the wireless bandwidth efficiency for both V2Iand V2V communications with a central scheduler in thebackbone. The scheduler maintains the states of all the devices,including vehicle velocities, locations, etc., by collecting thebeacon messages from vehicles. Meanwhile, the scheduler isable to collect information from each RSU, including databasestatus, bandwidth status, etc. Based on the global knowledgeof both vehicles and RSUs, the scheduler can estimate thenetwork connectivity among different devices in the nearfuture [14]. Further, the central scheduler can adaptively makethe data dissemination decisions, including the forwardingrules of RSUs and vehicles based on the dynamic vehicularenvironment. With above motivations, this work is dedicatedto optimizing both quality of service request and data deliveryratio in the hybrid of V2I/V2V vehicular communicationenvironments.

The main contributions of this paper are summarized asfollows.

• We consider a novel temporal information service sce-nario in vehicular networks, where vehicles with up-to-date information will be scheduled to upload data items tothe RSUs when they are passing by. Meanwhile, vehiclesmay also submit service requests to RSUs for particularapplications, and based on certain scheduling policy, these

vehicles can be served via the hybrid of V2I and V2Vcommunications.

• We present a system architecture to support efficient dataservices for the above described scenario. In such anarchitecture, real-time vehicle statuses, including GPSpositions, velocities, driving directions and cache/requestinformation, etc., are collected by RSUs via the beaconmessages of vehicles. Then, each RSU will report its localinformation to the central scheduler via the high speedbackbone network. Based on the collected information,the central scheduler is capable of estimating the networkconnectivity and instructing the behaviors of static RSUsas well as mobile vehicles so as to best exploit the V2Iand V2V communication resources cooperatively.

• We quantitatively define service quality of the requestsand delivery ratio of the system, respectively. Then,we analyze that the objectives of maximizing the servicequality and the delivery ratio are in conflict with eachother. On this basis, we formulate the problem called tem-poral data uploading and dissemination (TDUD), whichaims to optimize the two objectives simultaneously.

• We propose a problem-specific multi-objective evolution-ary algorithm called MO-TDUD, which includes chro-mosome encoding, population initialization, evolutionaryoperators and local refinement subroutines. The proposedMO-TDUD can generate a set of pareto-solutions withoutextra overhead, which makes it scalable and adaptable todifferent application requirements.

• We build the simulation model to simulate and give acomprehensive performance evaluation. The simulationresults demonstrate the effectiveness and the scalabilityof the proposed MO-TDUD.

The rest of this paper is organized as follows. Section IIreviews the related work. In Section III, we present thesystem model. Section IV formulates the TDUD problem andSection V proposes the algorithm. In section VI, we buildthe simulation model and give the performance evaluation.Section VII concludes the work and discusses future researchdirections.

II. RELATED WORK

Data dissemination in vehicular networks have receivedconsiderable attention in the past decade. For better high-lighting the distinguishing of this work, we classify relatedworks into three categories based on communication types,namely, V2I, V2V and hybrid V2I/V2V communications.First, for data service in V2I scenario, all requested data itemshave to be retrieved within the RSUs’ coverage. Specifically,Liu et al. [15] investigated real-time service within a singleRSU and designed two online algorithms to satisfy the staticand dynamic snapshot consistency requirements. Further, tem-porality of data is one of the key features in dynamic vehicularenvironment. Zhang et al. [13] proposed an architecture toenable temporal data services within a single RSU’s coverage,where the RSU is responsible for assigning data uploadtasks and broadcasting temporal data to passing vehicles.Dai et al. [16] quantified the freshness of temporal data and

Page 3: Temporal Information Services in Large-Scale Vehicular Networks … · 2019-03-18 · Physical Systems Research) under Grant 14-824-09-013, in part by the GRL Program through NRF

220 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 20, NO. 1, JANUARY 2019

proposed a priority-based scheduling (PBS) algorithm to strikea balance between improving service ratio and enhancing dataquality in a single RSU. Further, [17] extended the data serviceto multi-RSU scenario, where a sharing scheme of multi-ple RSUs’ residual bandwidth is designed to address querystarvation inside the multi-item query problem. However, allthe above studies only considered the data service within theRSU’s service range.

In the scenario of V2V communication, vehicles act asrelay nodes and data transmission from source vehicle todestination vehicle is a multi-hop process via multiple relaynodes. Particularly, plenty of studies [18]–[20] focused oninterference problem in V2V communication. Tchouankemand Lorenzen [18] investigated the impact of co-channelinterference in V2V communication under line-of-sight andnon-line-of-sight conditions. Wisitpongphan et al. [19] pro-posed a routing protocol by utilizing a timer-based broad-cast suppression technique at network layer, which reduces70 percent broadcast redundancy on a well connected net-work. In addition, great efforts [21]–[24] have been paid onrouting issues in dynamic network topology. Tonguz et al. [21]designed a local-information-based multi-hop protocol, whichenhances reliability and efficiency of data dissemination inhighway scenarios. Abdou et al. [22] took tools of simulatorand evolutionary algorithm to optimize the broadcast para-meters, including relay probability and retransmission time.To reduce transmission delay, Zhao and Cao [23] proposeda vehicle-assisted data delivery (VADD) protocol, which for-wards the packet to the destination with the lowest data-delivery delay by predicting the vehicle mobility. To improvemulticast efficiency, Jiang et al. [24] proposed an architec-ture to support efficient multicast in V2V communication,which makes forwarding decisions by estimating inter-vehicleencounter probability. The above studies focused on the effi-ciency and throughput of data transmission in V2V com-munication without considering temporality feature of dataitself.

Further, a number of studies focused on exploiting bothV2I and V2V resources improve the performance of dataservice. Zhang et al. [7] proposed a graph-based resource-sharing scheme to improve data throughput by allowinginterference-free transmission via both V2V and V2I com-munication link simultaneously. Wu et al. [25] designed alocation-based crowdsensing framework in hybrid V2I/V2Vnetworks. By retrieving the number and location of avail-able RSU resources, a routing switch mechanism combingRSU resources and ad hoc solutions, is designed to guar-antee quality of service under various network connectivity.Zhao et al. [26] proposed an analytical model integrating prop-agation distance, one hop transmission range and distributionof vehicles, to improve the network connectivity probabilityin multi-hop broadcasting scenario. Liu et al. [27] proposedan architecture where a centralized scheduler implemented atthe RSU is capable of scheduling data dissemination in V2Iand V2V channel simultaneously. The above research workfocused on improving the performance of data service byintegrating V2I and V2V resources, without considering thecharacteristics of temporal data service with exploiting the

cooperation of multiple RSUs and vehicles in a centralizedscheduling way.

For temporal information services in vehicular net-works, we have done preliminary studies in our previouswork [11], [12]. Specifically, [11] considered the serviceof real-time requests with multiple dependent data itemsin the service range of a single RSU. An adaptive-updaterequest scheduling (AURS) is designed to allocate bandwidthdynamically for data update and data broadcast, which aimsat serving requests with maximizing data quality. In [12],we further extended the temporal data service to multiple-RSU scenario. A multi-objective algorithm is implemented toenable the cooperation of multiple RSUs to enhance overallsystem performance by balancing the data update task amongmultiple RSUs. However, previous studies only consideredV2I based data services. Distinguished from previous works,we investigate in a new service scenario where vehicles cannot only be served via V2I communication inside RSU’scoverage, but also can share data items with each other viaV2V communication when they are out of RSU’s coverage.Accordingly, the target of the system is to maximize band-width efficiency of the hybrid of V2I/V2V communications.Specifically, we designed a centralized architecture, wherea central scheduler is capable of managing the behaviorsof RSUs, as well as distributed vehicles in a centralizedway based on the estimated network connectivity of differentdevices. In addition, a multi-objective algorithm integratingboth cooperative data update operations and hybrid V2I/V2Vdata forwarding rules, is designed to optimize service qualityand delivery ratio simultaneously. As far as we know, thisis the first work which aims to satisfy requirement on servicequality and delivery ratio adaptively by designing a centralizedarchitecture and developing a problem-specific multi-objectivealgorithm within exploiting V2I and V2V resources in theconcerned application scenario.

III. SYSTEM MODEL

In this section, we present a novel architecture to efficientlysupport temporal information services in large-scale vehicularnetworks. As shown in Fig. 1, the RSUs are connected viathe high speed backbone network. Accordingly, we assumethat any data update in the local database of one RSU willbe synchronized in other peers. The central scheduler inthe backbone network is responsible for making schedulingdecision, including data broadcast strategy of all the RSUsand the vehicles. The procedure of the scheduling policyoptimization is described as follows. First, the beacon messageperiodically broadcast by vehicles is overheard by RSUs,which is synchronized to the central scheduler via the back-bone network. The beacon message includes GPS position,velocity and cache/request information of vehicles in thesystem. Second, based on the mobility information of vehicles,the central scheduler estimates the network connectivity ofvehicle-to-RSU and inter-vehicle communication in the nearfuture. Further, based on the collected cache and requestinformation of vehicles, the scheduler will make a sequence ofdata broadcast decisions via both V2I and V2V communica-tions. Third, the message containing the scheduling policy is

Page 4: Temporal Information Services in Large-Scale Vehicular Networks … · 2019-03-18 · Physical Systems Research) under Grant 14-824-09-013, in part by the GRL Program through NRF

DAI et al.: TEMPORAL INFORMATION SERVICES IN LARGE-SCALE VEHICULAR NETWORKS 221

Fig. 1. A service architecture for temporal information service in large-scale vehicular networks.

delivered to RSUs and vehicles via wired connection and V2Icommunication, respectively. Once receiving the schedulingdecision, the vehicles and RSUs broadcast the data itemsaccordingly via either V2I or V2V communications.

The temporal information service is one of the criticalapplications in vehicular networks. In the concerned scenario,the vehicles may request different types of temporal informa-tion, such as road traffic information, shopping mall promotionand location-based information, etc. Meanwhile, the requesteddata items have to be delivered within certain tolerated delay.For example, the vehicles driving through certain urban areamay be interested in local information, such as availableparking lots or current promotions in shopping malls nearby.In order to complete such a service, a maximum tolerateddelay needs to be associated based on certain factors suchas driver expectations or driving speed. On the other hand,since the quality of temporal information degrades over time,timely data update is required to keep the information useful.A time stamp is recorded to reflect the freshness of temporaldata. Vehicles are capable of sensing and caching the up-to-date information along their trajectories, and they will sendbeacon messages (including cache and request information)periodically to the central scheduler via backbone. Based onthe global knowledge, the scheduler instructs the RSUs toassign upload tasks to the vehicles in their respective coverage,and then the vehicles will perform data update based on thescheduling decisions. Besides, the scheduler will periodicallydetermine the broadcast strategy at RSUs and vehicles, so thatthe service can be adaptable to dynamic traffic conditions.

In the following, we give an example to better illustrate thepresented system architecture as well as to reveal the chan-llenges of designing an efficient scheduling approach. First,we introduce the service scanrio. As shown in Fig.2, the setof RSUs and the set of vehicles are denoted by R = {r1, r2, r3}and V = {v1, v2, v3, v4}, respectively. Further, the set ofdatabase is denoted by D = {d1, d2, . . . , d5}. For simplicity,

the data quality is characterized into two levels, high (denotedby H) and low (denoted by L). Assume that both the uploadingand downloading for one data time takes one time slot. Theglobal knowledge of central scheduler includes the followinginformation: a) local database information of each RSU r j .For example, the local information of RSU r1 is denotedby {d1:L, d2:H , d3:H , d4:H , d5:L}, which indicates thatd1 and d5 have low quality and d2, d3 and d4 have highquality; b) the cached/requested information and enter/leavetime of each vehicle vk . For example, the “Request” columnand the “Cache” column of v3 is “d4” and “d1:H , d2:H ”,respectively, which indicates that v3 requests d4 and cachesd1 and d2. Further, the “Enter Time” column and the “LeaveTime” column of v3 are t0 and t3, which indicates that v3enters r2 at t0 and leaves r2 at t3. The “Deadline” column ofv3 is t3, which represents that v3 has to retrieve all requesteddata before t3. Further, assuming that v1 is predicted to meetv4 at certain time based on the knowledge of vehicle drivingdirections, velocities and trajectories, etc.

The scheduler is supposed to make the following schedulingdecisions, as shown in Fig. 2. Solution 1 is an optimalsolution. At time t0, r1 broadcasts d2 to v1 and r2 assignsupdate task for v3 to upload d1 for enhancing quality of d1.Simultaneously, d1 with high quality is synchronized in otherlocal databases via the backbone network. Then, v1 in thecoverage of r1 receives d1 with high quality at t2. It is notedthat v3 receives d3 from r2 at t2 for the purpose of dataforwarding. v4 finally receives d3 at time t3 with high qualitybefore the deadline t4. The column of Result 1 shows theperformance of Solution 1. As observed, all vehicles retrievethe requested data items with high quality. On the contrary,Solution 2 is a broadcast-first strategy. For each RSU, it giveshigher priority to the service requests compared with updaterequests. As shown in the column of Result 2, v1 receives d1with low data quality and v2 misses d1, which leads to poorservice performance. We can observe from the above example

Page 5: Temporal Information Services in Large-Scale Vehicular Networks … · 2019-03-18 · Physical Systems Research) under Grant 14-824-09-013, in part by the GRL Program through NRF

222 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 20, NO. 1, JANUARY 2019

Fig. 2. An example of scheduling.

TABLE I

SUMMARY OF NOTATIONS

that it is not trivial to design an optimal scheduling policyto achieve high delivery ratio while maintaining satisfactoryservice quality, especially when the application requirementchanges dynamically.

IV. PROBLEM FORMULATION

A. Preliminary

The set of the data items is denoted by D and the setof vehicles is denoted by V . For a vehicle vk ∈ V , the setof cached data items at time t is denoted by Ck (t). The set ofrequested data items by vehicle vk is denoted by Ak and therequest submission time is denoted by stk . At time t , the setof outstanding requests of vehicle vk is denoted by Ak (t).Further, the maximum tolerated delay for serving vehicle vk

is denoted by sdk . Accordingly, we have Ak (t) ∩ Ck (t) = ∅and Ak (t) , Ck (t) ⊆ D. The set of RSUs is denoted by R.The set of vehicles in the coverage of RSU ri at time t isdenoted by Vi (t). The dwelling time of a vehicle vk ∈ Vi (t)

in the coverage of RSU ri is denoted by I rik and the time

interval when vk and vl are within their V2V communicationrange is denoted by I v

kl . Note that the values of I rik and I v

klcan be estimated based on available trajectory predictiontechniques [28], [29]. The primary notations in this paper aresummarized in Table I.

B. Quality of Service

In this section, we establish the data quality model toquantitatively evaluate the freshness of temporal information.Specifically, for data d j , we define the data quality functionas Qd j (t) = f (ts j , ld j , t), where ts j is the time stamp whend j is generated, ld j is the data valid period and t is the currenttime, which indicates that the value of d j will be outdatedat ts j + ld j . During

[ts j , ts j + ld j

], the quality of d j will

degrade over time if it is not updated in due course. Clearly,the degradation of data quality depends on both ts j and ld j .If the value of ts j is closer to the current time t , it indicates

Page 6: Temporal Information Services in Large-Scale Vehicular Networks … · 2019-03-18 · Physical Systems Research) under Grant 14-824-09-013, in part by the GRL Program through NRF

DAI et al.: TEMPORAL INFORMATION SERVICES IN LARGE-SCALE VEHICULAR NETWORKS 223

that d j is generated more recently. On the other hand, a largervalue of ld j represents that the valid period of d j is longer,which indicates that Qd j (t) degrades more slowly. For thesake of better exhibition, Qd j (t) is normalized into the rangeof [0,1], where 1 means the data has no loss of quality (when itis generated) and 0 means the data value is invalid. In addition,Qri

d j(t) and Qvk

d j(t) represent the data quality of d j maintained

in the database of RSU ri and cached by the vehicle vk ,respectively. If vk uploads its cached d j to RSU ri , the valueof Qri

d j(t) would be updated by Qvk

d j(t).

To emphasize the generality of our analysis, we adopt thegeneral form of the data quality model when analyzing theproblem, while giving a specific form in Section V whenimplementing the simulation model.

C. Temporal Data Uploading and Dissemination(TDUD) Problem

Given the service interval [0, T ], let x = {Or1 , . . . ,

Ori , . . . , Ov1 , . . . , Ovk , . . .}

be the set of operations, whereOri represents the operation sequence of the RSU ri , i =1, 2, . . . , �R� and Ovk represents the operation sequence ofthe vehicle vk , k = 1, 2, . . . , �V �. The two-tuple (ol, sl ) ∈ Ori

represents the lth operation of ri , where ol denotes the operandand sl denotes the operation instruction. Specifically, sl = 0indicates the data upload and sl = 1 indicates the databroadcast. Accordingly, the start time tl of the lth operation

(ol, sl ) is computed by tl =l−1∑

k=1(τ1 + sk (τ2 − τ1)), where τ1

and τ2 represent the time taken for broadcasting one data itemby a vehicle and a RSU, respectively. Similarly, the two-tuple(om, tm) ∈ Ovk represents the mth operation of vk and tmrepresents the start time of the mth operation. Due to thelimited bandwidth and vehicular mobility, several constraintsare imposed on data operations, which are described asfollows.

When the RSU ri broadcasts ol at time tl , the vehi-cle vk acquires oi if it satisfies the following conditions:1) vk requests the data, which has not yet been received, ol ∈Ak (tl); 2) the delivery delay does not exceed the maximumtolerated delay of vk , i.e., tl + τ2 − stk ≤ sdk ; 3) vk cancomplete the retrieval of ol before it leaves the coverage ofRSU ri , i.e., [tl , tl + τ2] ∈ I r

ik . In order to evaluate the benefitof broadcasting ol , we define the Beneficial Vehicle Set asfollows:

Definition 1: Beneficial Vehicle Set: When ri broadcastsdata d j at time t, the beneficial vehicle set is defined as theset of vehicles in the coverage of ri , which can acquire d j ,expressed as follows:

Brid j

(t) = {vk∣∣[t, t + τ2] ∈ I r

ik∩ d j ∈ Ak (t) ∩ t + τ2 − stk

≤ sdk, ∀vk ∈ Vi (t)} (1)

As the data quality degrades over time, RSUs will performdata update operation to keep the freshness of the data.When RSU ri performs a data update operation (ol, 0) attime tl , it will assign the data upload task to a vehicle forenhancing Qri

ol (t). A vehicle vk can be selected as a candidateif it satisfies the following conditions: 1) ol belongs to the

set of cached data, i.e., ol ∈ Ck (tl); 2) vk can completethe data upload of ol before it leaves the coverage of ri ,i.e., [tl , tl + τ1] ⊆ I r

ik . To achieve best performance, a RSUalways chooses the candidate with the best data quality.Therefore, in order to evaluate the benefit of data update,we define the Best Upload Quality as follows:

Definition 2: Best Upload Quality: When RSU ri updatesdata quality of d j at time t, the best upload quality of d j atRSU ri at time t is defined as the highest data quality cachedby the candidate vehicles, expressed as follows:

Quid j

(t) = max∀vk∈Vi (t)

{Qvk

d j(t)

∣∣[t, t + τ1] ⊆ I r

ik ∩ d j ∈ Ck (t)}

(2)

Then, the vehicle assigned for uploading data d j at RSU ri isdetermined as follows:

v∗j = arg max

∀vk∈Vi (t)

{Qvk

d j(t)

∣∣[t, t + τ1] ⊆ I r

ik ∩ d j ∈ Ck (t)}

(3)

If d j is not cached by any vehicles in the coverage of ri ,then Qui

d j(t) is set to 0. Therefore, after data update opera-

tion (ol, 0), the data quality of ol in the database Qriol (tl + τ1)

is updated to Quiol (tl + τ1). Simultaneously, the data quality

of ol in other local databases will be synchronized via thehigh-speed backbone network accordingly.

When two vehicles vk and vl are in the V2V communicationrange and outside the coverage of RSUs, they have chanceto exchange cached data items with each other via V2Vcommunication. When vk broadcasts om at time tm , vl acquiresom if it satisfies the following conditions: 1) om belongs to theset of outstanding request of vl at time tm , i.e., om ∈ Al (tm);2) the delivery delay does not exceed the maximum tolerateddelay of vl , i.e., tm +τ1 −stl ≤ sdl ; 3) data transmission of om

is completed during I vkl , i.e., [tm , tm + τ1] ⊆ I v

kl . In addition,to avoid transmission collision, vk and vl should not broadcastsimultaneously during I v

kl . Therefore, for any two operations(om, tm) ∈ Ovk , (on, tn) ∈ Ovl during I v

kl , the collisionavoidance constraint is expressed as follows: [tm , tm + τ1] ∩[tn, tn + τ1] = ∅. Similarly, to evaluate the broadcast benefit,we compute the beneficial vehicle set Bvk

om (t) of data om

broadcast by vehicle vk at time t .In order to quantitatively measure the system performance,

we define two metrics, namely, delivery ratio and averageservice quality.

Definition 3: Delivery Ratio (DR): Given an operation x,SR is defined as the ratio of the satisfied service requests tothe total number of submitted service requests during serviceinterval [0, T ], which is computed by:

f1 (x)=

�R�∑

i=1

∥∥Ori

∥∥

l=1sl ∗ ∥

∥Briol (tl)

∥∥+

�V �∑

k=1

∥∥Ovk

∥∥

m=1

∥∥Bvk

om (tm)∥∥

∀vk∈V�Ak (0)� (4)

Definition 4: Average Service Quality (ASQ): Given anoperation x, ASQ is defined as the mean value of service

Page 7: Temporal Information Services in Large-Scale Vehicular Networks … · 2019-03-18 · Physical Systems Research) under Grant 14-824-09-013, in part by the GRL Program through NRF

224 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 20, NO. 1, JANUARY 2019

qualities of all the satisfied service requests during serviceinterval [0, T ], which is computed by:

f2 (x) =

�R�∑

i=1

∥∥Ori

∥∥∑

l=1Qri

ol (tl) ∗ sl ∗ ∥∥Bri

ol (tl)∥∥

+�V �∑

k=1

∥∥Ovk

∥∥∑

m=1Qvk

om (tm) ∗ ∥∥Bvk

om (tm)∥∥

�R�∑

i=1

∥∥Ori

∥∥

l=1sl ∗ ∥∥Bri

ol (tl)∥∥ +

�V �∑

k=1

∥∥Ovk

∥∥

m=1

∥∥Bvkom (tm)

∥∥

(5)

From the system point of view, it desires both high deliveryratio and service quality. With the above definitions, we for-mulate the TDUD as a two-objective optimization problem asfollows:

maximize F (x)=( f1 (x) , f2 (x))

subject to

∥∥∥Or j

∥∥∥

l=1

(τ1+sl (τ2−τ1)) ≤ T, j =1, 2, . . . , �R�

τ1 · ∥∥Ov

k

∥∥ ≤ T, k = 1, 2, . . . , �V �

∩i=m,n

[ti , ti + τ1] = ∅, i f ∪i=m,n

[ti , ti +τ1]∈ I vkl ,

∀ (om, tm) ∈ Ok , (on, tn) ∈ Ol (6)

V. ALGORITHM DESIGN

In this section, we propose an evolutionary algorithm,namely, Multi-Objective Temporal Data Uploading andDissemination (MO-TDUD), aiming at optimizing both theDR and the ASQ simultaneously. First, we present the basicidea of MO-TDUD. Then, we explain the mechanism of eachMO-TDUD component in detail.

A. Multi-Objective DecompositionFirst, we use Weighted Sum [30] to decompose the multi-

objective optimization problem in Eq. 6 into N scalar opti-mization subproblems. Let {w1, . . . , wN } be a set of evenlyspread weight vectors, where wk = (

wk1, wk

2

)and wk

1 +wk

2 = 1. Each weight vector wk corresponds to a scalaroptimization problem. Therefore, for each weight vector wk ,the objective of the kth corresponding subproblem is thenformulated as follows:

g(

x∣∣∣wk, z

)=

2∑

i=1

wki fi (x), subject to x ∈ � (7)

where � is the decision space. For each weight vector wk ,the best solution for the kth subproblem during the searchprocess is maintained for population evolution in the nextgeneration.

B. Chromosome EncodingA problem-specific chromosome encoding form is designed

to represent the solution of T DU D problem, which includesthe operations of both RSUs and vehicles during service inter-val [0, T ]. Fig. 3 shows an example of chromosome encod-ing. The rectangles of Ori and Ovk represent the operation

Fig. 3. Encoding form of the chromosome.

sequences of RSU ri and vehicle vk , respectively and the arrowindicates the order of the operation sequences. For example,Or2 represents the scheduling for RSU r2. Specifically, accord-ing to the ordered operations of r2, it represents updatingd6 and d3 in sequence, followed by broadcasting d3, and finallyupdating d7. The total time consumed by the operations of oneRSU should not exceed T . Similarly, the operations of v1 areto broadcast d6 at time t1, and then broadcast d5 at t3. Theoperations of v2 are to broadcast d1 at t2, and then broadcastd7 at time t4. When v1 and v2 are in the V2V communicationduring [0, T ], the time interval of broadcasting each data

should not be overlapped, i.e.,4∩

i=1[ti , ti + τ1] = ∅.

C. MO-TDUD Procedure

First, the procedure of MO-TDUD is outlined as follows.Step 1 Initialization: Set up initial parameters and generate

initial population.Step 2 Update: Update the population based on particularly

designed mechanisms, including selection, mutation operatorsand a local refinement method.

Step 3 Output: Check the stopping criterion. If the stoppingcriterion is not satisfied, go back to Step 2. Otherwise, outputthe set of non-dominated solutions and select the best solutionin this set according to the given weight vector.

In the following, we give the detailed steps of theMO-TDUD, where the key components of problem-specificoperators are elaborated.

1) Initialization: The initialization consists of three parts.First, we initialize the non-dominated set E P as an emptyset, which is used for archiving the non-dominated solutionsin due course. Let xn and xm denote two solutions, since ourformulated problem in Section IV is a maximization problem,F (xn) dominates F (xm) if and only if fi (xn) ≥ fi (xm) , i =1, 2 and fi (xn) > fi (xm) for at least one index i ∈ {1, 2}.We call a solution x a non-dominated solution if no solutionsx � ∈ � whose F

(x �) can dominate F (x). The non-dominated

set E P consists of the solutions, which cannot be dominatedby any other solutions in the population.

Second, for each weight vector wk , we generate theneighborhood set which contains M neighborhood vec-tors. For any two weight vectors wn, wm ,∀n, m ≤ N ,we compute the Euclidean distance between wn and wm ,

i.e., �wm − wn� =√

2∑

i=1

(wm

i − wni

)2. Then, for each weight

vector wk , we choose the M closest weight vectors to wk asits neighborhood set N B

(wk

) = (wk1 , wk2 , . . . , wkM

). The

Page 8: Temporal Information Services in Large-Scale Vehicular Networks … · 2019-03-18 · Physical Systems Research) under Grant 14-824-09-013, in part by the GRL Program through NRF

DAI et al.: TEMPORAL INFORMATION SERVICES IN LARGE-SCALE VEHICULAR NETWORKS 225

neighborhood set of the kth subproblem consists of all the sub-problems associated with the weight vectors from N B

(wk

).

For the kth subproblem, only the current solutions to itsneighborhood subporblems are exploited for optimizing asubproblem.

Third, we design a particular method to initialize thepopulation. A solution x consists of Ori for each RSU ri anddata operation Ovk for each vehicle Ovk . We first generatethe initial data operations Ori for each RSU ri under a givenweight vector wk . For each candidate operation

(d j , s j

) ∈ Ori ,we evaluate the operation benefit in two aspects: broadcastbenefit and update benefit. For a broadcast operation

(d j , s j

),

the broadcast benefit is denoted by �1((

d j , s j), t

), which is

equal to the product of Qrid j

(t) and the number of vehicles

in Brid j

(t). For a data update operation, it is equal to zero sincedata update operation brings no benefit on serving requests.To sum up, the broadcast benefit of an operation

(d j , s j

)at

time t is computed as follows:

�1((

d j , s j), t

) = Qrid j

(t) ·∥∥∥Bri

d j(t)

∥∥∥ · s j (8)

Similarly, the update benefit of an operation(d j , s j

)is com-

puted as follows:

�2((

d j , s j), t

)

=βd j ·(

Quid j

(t)−Qrid j

(t))·⎛

⎝�R�∑

i=1

∥∥∥Bri

d j(t)

∥∥∥+ε

⎠·(1−s j)

(9)

where βd j is the popularity and ε is a constant (0 < ε < 1).βd j is defined as the ratio of the number of service requestsasking for d j to the total number of service requests, whichgives higher priority to more popular data items. For anupdate operation

(d j , s j

), the update benefit synthesizes the

popularity, the augmented data quality and the total requestnumber in the system. On the other hand, for a broadcastoperation, the update benefit is equal to zero. Combing Eq. 8and Eq. 9, the benefit of each candidate data operation undera given wk is formulated as follows:

�((

d j , s j), t

∣∣∣wk

)= max

1≤i≤2

(wk

i · �i((

d j , s j), t

))(10)

For each wk , we initialize the data operation set Ori asfollows: a) Sort the candidate operations of each RSU ri indescending order. b) For each RSU ri , the operations areselected iteratively based on the sorted order until the totalcumulative time reaches the bound T . Second, we initializethe data operations Ovk during V2V communication. For eachtime interval I v

kl = [etkl , ltkl ], the data to be broadcast duringI vkl is determined as follows: a) The candidate data set (CDS)

of each vk is determined by Ck (etkl) ∩ Al (etkl ). b) The vk

with longer remaining time stk + sdk − t broadcasts the datain CDS first since the other party is more urgent to receivethe data. c) The d j with the highest data quality in CDS isselected iteratively until the total cumulative time achieves thebound ltkl − etkl . If the cumulative time has not yet reach thebound, the other vehicle continues to select the data in its CDSby repeating steps a) to c). After step c), the broadcast data

set Ovk→vl of vk and Ovl→vk of vl during I vkl is determined.

Accordingly, during service interval [0,T], the broadcast data

set of vk is set as Ovk = �V �∪l=1

Ovk→vl . The initial data operation

set of Ovk maintains the same for each weight vector wk .2) Update: For each index k, k = 1, 2, . . . , N , the update

includes the following procedures:1. Reproduction: the two weight vectors wn and wm are

randomly selected from N B(wk

)and the two corresponding

solutions xn and xm are chosen as parent solutions. Then,a new solution y is generated from xn and xm by the designedgenetic operators. First, we apply uniform crossover operatoron xn and xm to generate one child solution y. By usinguniform crossover operator, each Ori and Ovi of solutiony is randomly selected from xn and xm iteratively. Then,the mutation operator is applied to y. It mutates each Ori andOvi of solution y independently with mutation probability ρ.If one Ori (Ovi ) is selected to be mutated, then one operationof Ori (Ovi ) is randomly selected to be mutated. The selectedoperation is replaced by a data operation randomly selectedfrom the candidate data operation set of ri (vi ).

2. Local refinement: a local refinement method is proposedto generate a new solution y � based on solution y. Themotivation is to reduce the unnecessary data upload operationand further enhance bandwidth utilization. Rule 1, if multipledata update operations

(d j , 0

)of one data d j appear in the

solution y, only the data update operation with the maximumQri

d j(t) is remained and other data update operations are

replaced by other unselected data operations according to thebenefit order in Eq. 10. Rule 2, if there exists an updateoperation

(d j , 0

)in the Ori ,

(d j , 1

)of any other RSUs are

arranged after(d j , 0

)of the Ori .

3. Update neighborhood solution: for each weight vectorwl ∈ N B (k), we compare the objective stated in Eq. 7 basedon solution xl and the new solution y �. If g

(xl

∣∣wl , z

) ≤g

(y � ∣∣wl , z

), then the corresponding solution xl is replaced

by y �.4. Update E P: for any solution x ∈ E P , if F (x) is

dominated by F(y �), then x is removed from E P . If there

is no x ∈ E P which satisfies that F (x) dominates F(y �),

then add y � to E P .3) Output: The iteration is terminated and output the non-

dominated set E P if it satisfies that the iteration numberreaches the maximum value, which is pre-defined basedon domain-knowledge. Once the non-dominate set E P isobtained, the scheduler chooses a solution x from E P . In orderto adaptively fulfill the given requirement on delivery ratioand service quality, a weight vector is given to determinethe final solution. In particular, given w� = (

w1�, w2

�) andw1

� + w2� = 1, wi

� ≥ 0, i = 1, 2, we evaluate the priority ofa solution x ∈ E P as follows:

P (x) =2∑

i=1

fi (x)wi� (11)

Then, the final solution x∗ is chosen with the maximum valueP (x∗) = max∀x∈E P

{P (x)}. In practice, the weight vector can

be evaluated by certain evaluation method based on statistical

Page 9: Temporal Information Services in Large-Scale Vehicular Networks … · 2019-03-18 · Physical Systems Research) under Grant 14-824-09-013, in part by the GRL Program through NRF

226 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 20, NO. 1, JANUARY 2019

Fig. 4. Simulation modules.

information of particular application scenarios. In addition,in order to enhance the V2V communication chance, the vehi-cle in the coverage of RSUs may cache non-requested dataitems. When the vehicle vk receives a data broadcast by theRSU at time t , it will first check whether the data item belongsto Ak (t). If it is true, the vehicle vk will accept the data itemand service is completed. Otherwise, vk will further checkwhether the data item belongs to ∩

Ikl �=0Al (t). If it is true,

the vehicle vk will accept the data item and forward it tothe destination vehicle when they are in V2V communicationrange. Otherwise, the vehicle vk discards it.

VI. PERFORMANCE EVALUATION

A. Simulation Model

In this section, we build the simulation model based on thesystem architecture described in Section III for performanceevaluation. Specifically, as shown in Fig. 4, the simulationmodel integrates the real-world map, the traffic simulatorSUMO and the scheduling module. Specifically, a real-worldmap, extracted from a 7km × 7km area of Hi-tech Zone,in Chengdu, China, is downloaded from OpenStreetMap andimported to establish the traffic scenario in SUMO [31].Then, SUMO is adopted to simulate vehicle mobility andgenerate vehicle traces. The scheduling module as well as theproposed algorithm is implemented based on C programming.The arrival of vehicles follows the Poisson process and thearrival rate is denoted by λ. For each vehicle, the numberof cached data items is randomly generated in the rangeof cn and the number of requested data items is randomlygenerated in the range of rn. The data valid period is randomlygenerated in the range of vp. The data access pattern followsthe commonly used Zipf distribution [32] with a skewnessparameter θ . Further, the maximum tolerated delay of eachrequested data is set to T Lmax . The total number of dataitems maintained in the database is 200. The time units takenfor broadcasting one data item via V2I and V2V communi-cation are set to 1s and 2s, respectively. This is reasonablebecause according to DSRC [33], it supports the data ratefrom 3∼27 Mpbs, depending on the modulation technique and

TABLE II

DEFAULT VALUE

vehicle speed, and hence it is sufficient to transmit a data itemwith normal sizes (e.g. in the order of KB) in second order.The default parameters and the corresponding descriptionsare summarized in Table II. For algorithm implementation,the population size and maximum iteration number is set to500 and 100, respectively. Further, mutation probability is setto 0.01 and the weight vector of MO-TDUD is set to (0.3, 0.7).The above parameter configurations are commonly adoptedin relevant literatures, which can refer to [34]–[36]. Unlessstated otherwise, the simulation is conducted under the defaultsetting.

To quantitatively evaluate the data quality, a commonly usedlinear function [37] is adopted, which is a typical settingfor evaluating quality of temporal information in vehicularnetworks. The formulation is expressed as follows:

Qd j (t) ={

1 − t−t s jld j

ts j ≤ t ≤ ts j + ld j

0 t < ts j or t > ts j + ld j(12)

Further, for performance comparison, we have implementedtwo alternative solutions: a) PBS algorithm [16], which deter-mines the operations of downloading and uploading in eachtime slot based on evaluating the priority of each operation bysynthesizing the number of pending requests, service dead-line and data quality. b) Alternative-EDF, which alternativelychooses update operation or broadcast operation in each timeslot. The order of update (or broadcast) operation is sortedby the deadline, namely, earliest deadline first. We collect thefollowing statistics from the simulation: the set of requesteddata items by each vehicle Avk , the set of received data itemsby each vehicle Cvk and data quality Qvk

diof each received

data item. On this basis, we consider the two objectives forevaluating system performance, namely, delivery ratio (DR),average service quality (ASQ), which are computed asfollows:

DR =∑

∀vk

∥∥Cvk

∥∥/

∀vk

∥∥Avk

∥∥ (13)

ASQ =∑

∀vk

∀di∈Cvk

∥∥∥Qvk

di

∥∥∥/

∀vk

∥∥Cvk

∥∥ (14)

Page 10: Temporal Information Services in Large-Scale Vehicular Networks … · 2019-03-18 · Physical Systems Research) under Grant 14-824-09-013, in part by the GRL Program through NRF

DAI et al.: TEMPORAL INFORMATION SERVICES IN LARGE-SCALE VEHICULAR NETWORKS 227

Fig. 5. Performance evaluation under different data valid periods. (a) Deliveryratio under different data valid periods. (b) Average service quality underdifferent data valid periods.

B. Experimental Results

1) Effect of Data Valid Periods: First, we compare the algo-rithms under different data valid periods. A longer data validperiod indicates slower degradation of data quality. Fig. 5(a)compares the DR of the three algorithms under different datavalid periods. When the data valid period increases, the DR ofboth MO-TDUD and PBS also increases gradually since boththe algorithms give more bandwidth for data broadcast. How-ever, the DR of AEDF remains the same because AEDF onlyconsiders the service deadline in scheduling while ignoring thevariation of data valid period. Due to the slower degradation ofdata quality, the ASQ of three algorithms increases graduallywith an increasing of data valid period, as shown in Fig. 5(b).This set of results demonstrate that MO-TDUD achievesbetter overall system performance under various data validperiods.

2) Effect of Service Workloads: Fig. 6 compares the algo-rithm performance under the different service workloads. Themore data items are requested by each vehicle, the heavierservice workload is. Fig. 6(a) compares the DR of the threealgorithms under different service workloads. As expected,the DR of three algorithms decreases gradually when theservice workload is getting higher. Fig. 6(b) compares the ASQof the three algorithms under different service workloads. TheASQ of MO-TDUD and PBS increases gradually because theincreasing of service workloads brings more benefit of broad-cast effect. That is, one broadcast data item has the potentialto serve more vehicles, which improves the bandwidth effi-ciency. However, the ASQ of PBS degrades gradually. This isbecause PBS allocates more bandwidth for data broadcasting,which is observed from Fig. 6(a) where the gap between

Fig. 6. Performance evaluation under different system workloads.(a) Delivery ratio under different number of requests per vehicle. (b) Averageservice quality under different number of requests per vehicle.

MO-TDUD and PBS is getting narrower in a heavier workloadenvironment. Therefore, the ASQ of PBS decreases. This set ofexperimental results demonstrate the scalability of MO-TDUDunder different service workloads.

3) Effect of Maximum Tolerated Delay: Fig. 7 comparesthe algorithm performance under different maximum tolerateddelays. As noted in Figs. 7(a) and 7(b), the DR and ASQof PBS and AEDF almost remain the same since these twostrategies do not consider the tolerated delay in scheduling.As shown in Fig. 7(a), at first, the DR of MO-TDUD increasesdramatically, which is mainly because the service time windowis very small at the beginning, and an increasing of theservice time window will give better performance of algo-rithms. When the maximum tolerated delay keeps increasing,the performance gain of DR slows down since the maximumtolerated delay is long enough and it is no longer a majorconstraint in scheduling. As demonstrated, the MO-TDUDshows better performance against other two algorithms underdifferent requirements on the maximum tolerated delay.

4) Effect of Traffic Workloads: This part evaluates theperformance of algorithms under different traffic workloads.Meanwhile, to give a comprehensive performance analysis,we also evaluate the adaptiveness of MO-TDUD under dif-ferent weight factors, wk = (

wk1, wk

2

), k = 1, 2, . . . .. Specif-

ically, wk1 ranges from 0.1 to 0.7 with interval 0.05, which

indicates that the assigned weight for DR from low to high.Fig. 8 compares the DR and ASQ of three algorithms underdifferent traffic workloads. Each rectangular dot represents theresult of MO-TDUD under a given weight vector. Accordingto Fig. 8, we have the following observations. First, whenwk

1 increases, the DR of MO-TDUD increases and the ASQ

Page 11: Temporal Information Services in Large-Scale Vehicular Networks … · 2019-03-18 · Physical Systems Research) under Grant 14-824-09-013, in part by the GRL Program through NRF

228 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 20, NO. 1, JANUARY 2019

Fig. 7. Performance evaluation under different maximum tolerated delays.(a) Delivery ratio under different maximum tolerated delays. (b) Averageservice quality under different maximum tolerated delays.

of MO-TDUD decreases. Clearly, assigning high weight onone metric will adversely result in the performance of otherone. Second, the variation of two metrics is not obviously attwo borders of the curve because at the beginning, the effectof changing weight on one part is too small to affect thescheduling preference between data broadcast and data updateoperations. Third, the points of both PBS and AEDF are belowthe curve of MO-TDUD in all cases in Fig. 8, which indicatesthat we can always choose a point in the curve whose ASQand DR are both higher than PBS. It is because MO-TDUDis capable of achieving pareto-solution frontier during themuti-objective optimization search. Fourth, when the trafficworkload increases as shown from Fig. 8(a)∼8(d), we canobserve that the DR of all the algorithms is decreasing. That isbecause the increasing of vehicle density gives higher serviceworkload. The above analysis conclusively demonstrates thatMO-TDUD is able to provide adaptive solutions to strike a bestbalance between delivery ratio and service quality dependingon any specific application requirement.

5) Effect of Traffic Scenarios: In this part, we evaluate theeffectiveness of MO-TDUD under different traffic scenarios.Specifically, Figs.9(a)∼9(c) show three different traffic sce-narios, which are extracted from the core areas of Xichengand Docheng districts in Beijing City, Binjiang district inHangzhou City and Futian district in Shenzhen City, respec-tively. Accordingly, the scales of the three traffic scenarios are108km2, 50 km2 and 30 km2, respectively. Further, the defaultsettings (see Table II) are adopted to simulate the vehiclemobility in all traffic scenarios.

Further, in order to give an insight into the perfor-mance of the particularly designed MO-TDUD algorithm,

Fig. 8. Performance evaluation under different traffic workloads.(a) λ = 1200veh/h. (b) λ = 1500veh/h. (c) λ = 1800veh/h. (d) λ =2100veh/h.

we implement three multi-objective evolutionary algorithms,i.e., SPEA-II [38], NSGA-II [39] and MOEA/D [30] forperformance comparison. Specifically, we implement thethree algorithms by adopting random population initialization

Page 12: Temporal Information Services in Large-Scale Vehicular Networks … · 2019-03-18 · Physical Systems Research) under Grant 14-824-09-013, in part by the GRL Program through NRF

DAI et al.: TEMPORAL INFORMATION SERVICES IN LARGE-SCALE VEHICULAR NETWORKS 229

Fig. 9. Different traffic scenarios. (a) Docheng and Xicheng districts in Beijing. (b) Binjiang district in Hangzhou. (c) Futian district in Shenzhen.

Fig. 10. Performance evaluation under different traffic scenarios. (a) Simulation results in scenario Fig.9(a). (b) Simulation results in scenario Fig.9(b).(c) Simulation results in scenario Fig.9(c).

and standard genetic operators, including binary tournamentselection, uniform crossover and bit mutation operators. Then,the same solution selection method used in MO-TDUD isadopted to select the best solutions from non-dominate set.In the simulation, all the evolutionary algorithms adopt thesame default parameter settings. To give a comprehensive per-formance analysis, we evaluate the adaptiveness of algorithmsunder different weight factors, wk = {wk

1, wk2}, where wk

1ranges from 0.1 to 0.9 with interval 0.1. Figs.10(a)∼10(c)show the performance of the four algorithms correspond-ing to the simulation scenarios of Figs.9(a)∼9(c), respec-tively. Particularly, in the figures, each rectangle, triangle,star and circle represent the results of MO-TDUD, SPEA-II,NSGA-II and MOEA/D under one weight factor, respectively.According to Fig.10, we have the following observations.First, the three naive solutions which are based on SPEA-II,NSGA-II and MOEA/D perform competitively. For example,as shown in Fig.10(a), compared with MOEA/D and SPEA-II,NSGA-II achieves better DR but less ASQ. On the contrary,MOEA/D can achieve better ASQ but lower DR comparedwith NSGA-II and SPEA-II. Second, it can be observedthat the Pareto optimal of MO-TDUD significantly dominatesthe solutions of the other algorithms in all the scenarios.Third, although MO-TDUD adopts the same framework withMOEA/D, we note that MO-TDUD achieves much betterperformance than MOEA/D in terms of both DR and ASQ. Thisconfirmed the importance of designing particular evolutionarymechanisms for solving the encountered problem of interests.Based on the above analysis, we may safely conclude that

MO-TDUD achieves best performance under various simula-tion scenarios.

VII. CONCLUSION AND FUTURE WORK

This paper presented a system architecture for temporaldata services in large-scale vehicular networks via the hybridV2I/V2V communications. In such an architecture, a cen-tral scheduler makes the scheduling policy for both RSUsand vehicles based on the collected global system status.On this basis, we formulated the TDUD problem, whichaimed to enhance both the service quality and the deliveryratio. Further, to solve the derived problem, we proposeda multi-objective evolutionary algorithm called MO-TDUD.In particular, in MO-TDUD, we designed a decompositionscheme to tackle the complexity incurred by multiple objec-tives, a scalable representation for MO-TDUD encoding,and the evolutionary operators for MO-TDUD reproduction.By adjusting the weight vectors, MO-TDUD can adaptivelyselect the solution from the non-dominate set to strike thebest balance between the ASQ and the DR based on differentapplication requirements. Lastly, we built a simulation modeland provided a comprehensive performance evaluation. Thesimulation results demonstrated the superiority and scalabilityof MO-TDUD.

In the future work, we would like to consider the wirelesscommunication issues resided in lower layers of the vehicularnetwork, such as packet loss and interference, to establish amore realistic system model. Also, the multi-hop V2V com-munication is expected to be incorporated to further enhance

Page 13: Temporal Information Services in Large-Scale Vehicular Networks … · 2019-03-18 · Physical Systems Research) under Grant 14-824-09-013, in part by the GRL Program through NRF

230 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 20, NO. 1, JANUARY 2019

the system performance. Finally, we would like to further lookinto the effect of the parameters for evolutionary algorithms soas to deeply understand the nature of designing multi-objectivealgorithm for problems in vehicular environments.

REFERENCES

[1] J. Santa, A. F. Gómez-Skarmeta, and M. Sánchez-Artigas, “Architectureand evaluation of a unified V2V and V2I communication systembased on cellular networks,” Comput. Commun., vol. 31, no. 12,pp. 2850–2861, 2008.

[2] P. Su and B. B. Park, “Auction-based highway reservation system anagent-based simulation study,” Transp. Res. C, Emerg. Technol., vol. 60,pp. 211–226, Nov. 2015.

[3] M. Wang, H. Shan, R. Lu, R. Zhang, X. Shen, and F. Bai, “Real-timepath planning based on hybrid-VANET-enhanced transportation system,”IEEE Trans. Veh. Technol., vol. 64, no. 5, pp. 1664–1678, May 2015.

[4] A. Baiocchi and F. Cuomo, “Infotainment services based on push-modedissemination in an integrated VANET and 3G architecture,” J. Commun.Netw., vol. 15, no. 2, pp. 179–190, Apr. 2013.

[5] A. Bazzi and A. Zanella, “Position based routing in crowd sensingvehicular networks,” Ad Hoc Netw., vol. 36, pp. 409–424, Jan. 2016.

[6] S. H. Bouk, G. Kim, S. H. Ahmed, and D. Kim, “Hybrid adaptivebeaconing in vehicular ad hoc networks: A survey,” Int. J. Distrib. SensorNetw., vol. 11, no. 5, p. 390360, May 2015.

[7] R. Zhang, X. Cheng, Q. Yao, C.-X. Wang, Y. Yang, and B. Jiao,“Interference graph-based resource-sharing schemes for vehicular net-works,” IEEE Trans. Veh. Technol., vol. 62, no. 8, pp. 4028–4039,Oct. 2013.

[8] P. Fazio, F. De Rango, C. Sottile, and C. Calafate, “A new chan-nel assignment scheme for interference-aware routing in vehicularnetworks,” in Proc. IEEE 73rd Veh. Technol. Conf. (VTC Spring),May 2011, pp. 1–5.

[9] Y. Zeng, K. Xiang, D. Li, and A. V. Vasilakos, “Directional routing andscheduling for green vehicular delay tolerant networks,” Wireless Netw.,vol. 19, no. 2, pp. 161–173, 2013.

[10] H. Tong, X. Wu, and J. Zheng, “A destination information basedprobabilistic routing protocol for vehicular sensor networks,” in Proc.IEEE Int. Conf. Commun. (ICC), Jun. 2013, pp. 1419–1423.

[11] P. Dai, K. Liu, L. Feng, Q. Zhuge, V. C. S. Lee, and S. H. Son,“Adaptive scheduling for real-time and temporal information servicesin vehicular networks,” Transp. Res. C, Emerg. Technol., vol. 71,pp. 313–332, Oct. 2016.

[12] P. Dai, K. Liu, L. Feng, Q. Zhuge, V. C. S. Lee, and S. H. Son, “Towardsreal-time and temporal information services in vehicular networks viamulti-objective optimization,” in Proc. IEEE 41st Conf. Local Comput.Netw. (LCN), Nov. 2016, pp. 671–679.

[13] Y. Zhang, J. Zhao, and G. Cao, “Service scheduling of vehicle-roadsidedata access,” Mobile Netw. Appl., vol. 15, no. 1, pp. 83–96, 2010.

[14] Z. He, D. Zhang, and J. Liang, “Cost-efficient sensory data transmissionin heterogeneous software-defined vehicular networks,” IEEE Sensors J.,vol. 16, no. 20, pp. 7342–7354, Oct. 2016.

[15] K. Liu, V. C. S. Lee, J. K. Y. Ng, S. H. Son, and E. H.-M. Sha, “Schedul-ing temporal data with dynamic snapshot consistency requirement invehicular cyber-physical systems,” ACM Trans. Embedded Comput.Syst., vol. 13, no. 5s, Oct. 2014, Art. no. 163.

[16] P. Dai, K. Liu, E. Sha, Q. Zhuge, V. Lee, and S. H. Son, “Vehicleassisted data update for temporal information service in vehicularnetworks,” in Proc. IEEE 18th Int. Conf. Intell. Transp. Syst., Sep. 2015,pp. 2545–2550.

[17] G. G. M. N. Ali, E. Chan, and W. Li, “Supporting real-time multipledata items query in multi-RSU vehicular ad hoc networks (VANETs),”J. Syst. Softw., vol. 86, no. 8, pp. 2127–2142, 2013.

[18] H. Tchouankem and T. Lorenzen, “Measurement-based evaluationof interference in vehicular ad-hoc networks at urban intersections,”in Proc. IEEE Int. Conf. Commun. Workshop (ICCW), Jun. 2015,pp. 2381–2386.

[19] N. Wisitpongphan, O. K. Tonguz, J. S. Parikh, P. Mudalige, F. Bai,and V. Sadekar, “Broadcast storm mitigation techniques in vehicularad hoc networks,” IEEE Wireless Commun., vol. 14, no. 6, pp. 84–94,Dec. 2007.

[20] P. S. Bithas, G. P. Efthymoglou, and A. G. Kanatas, “A cooperativerelay selection scheme in V2V communications under interference andoutdated CSI,” in Proc. IEEE 27th Annu. Int. Symp. Pers., Indoor, MobileRadio Commun. (PIMRC), Sep. 2016, pp. 1–6.

[21] O. K. Tonguz, N. Wisitpongphan, and F. Bai, “DV-CAST: A distributedvehicular broadcast protocol for vehicular ad hoc networks,” IEEEWireless Commun., vol. 17, no. 2, pp. 47–57, Apr. 2010.

[22] W. Abdou, A. Henriet, C. Bloch, D. Dhoutaut, D. Charlet, and F. Spies,“Using an evolutionary algorithm to optimize the broadcasting methodsin mobile ad hoc networks,” J. Netw. Comput. Appl., vol. 34, no. 6,pp. 1794–1804, Nov. 2011.

[23] J. Zhao and G. Cao, “VADD: Vehicle-assisted data delivery in vehic-ular ad hoc networks,” IEEE Trans. Veh. Technol., vol. 57, no. 3,pp. 1910–1922, May 2008.

[24] R. Jiang, Y. Zhu, X. Wang, and L. M. Ni, “TMC: Exploiting trajectoriesfor multicast in sparse vehicular networks,” IEEE Trans. Parallel Distrib.Syst., vol. 26, no. 1, pp. 262–271, Jan. 2015.

[25] D. Wu, Y. Zhang, J. Luo, and R. Li, “Efficient data dissemination bycrowdsensing in vehicular networks,” in Proc. IEEE 22nd Int. Symp.Quality Service (IWQoS), May 2014, pp. 314–319.

[26] J. Zhao, Y. Chen, and Y. Gong, “Study of connectivity probabilityof vehicle-to-vehicle and vehicle-to-infrastructure communication sys-tems,” in Proc. IEEE 83rd Veh. Technol. Conf. (VTC Spring), May 2016,pp. 1–4.

[27] K. Liu, J. K. Y. Ng, V. C. S. Lee, S. H. Son, and I. Stojmenovic,“Cooperative data scheduling in hybrid vehicular ad hoc networks:VANET as a software defined network,” IEEE/ACM Trans. Netw.,vol. 24, no. 3, pp. 1759–1773, Jun. 2016.

[28] P. Pathirana, A. Savkin, and S. Jha, “Location estimation and trajectoryprediction for cellular networks with mobile base stations,” IEEE Trans.Veh. Technol., vol. 53, no. 6, pp. 1903–1913, Nov. 2004.

[29] C. Barrios, Y. Motai, and D. Huston, “Trajectory estimationsusing smartphones,” IEEE Trans. Ind. Electron., vol. 62, no. 12,pp. 7901–7910, Dec. 2015.

[30] Q. Zhang and H. Li, “MOEA/D: A multiobjective evolutionary algorithmbased on decomposition,” IEEE Trans. Evol. Comput., vol. 11, no. 6,pp. 712–731, Dec. 2007.

[31] M. Behrisch, L. Bieker, J. Erdmann, and D. Krajzewicz, “SUMO—Simulation of urban mobility: An overview,” in Proc. 3rd Int. Conf.Adv. Syst. Simulation (SIMUL), 2011, pp. 1–6.

[32] G. K. Zipf, Human Behavior and the Principle of Least Effort: An Intro-duction to Human Ecology. Cambridge, MA, USA: Addison Wesley,2016.

[33] J. B. Kenney, “Dedicated short-range communications (DSRC) standardsin the United States,” Proc. IEEE, vol. 99, no. 7, pp. 1162–1182,Jul. 2011.

[34] M. Fenton, D. Lynch, S. Kucera, H. Claussen, and M. O’Neill, “Loadbalancing in heterogeneous networks using an evolutionary algorithm,”in Proc. IEEE Congr. Evol. Comput. (CEC), May 2015, pp. 70–76.

[35] J.-S. Liu, S.-Y. Wu, and K.-M. Chiu, “Path planning of a data mulein wireless sensor network using an improved implementation ofclustering-based genetic algorithm,” in Proc. IEEE Symp. Comput. Intell.Control Autom. (CICA), Apr. 2013, pp. 30–37.

[36] N. Okati, M. R. Mosavi, and H. Behroozi, “A genetic approach in relay-jammer selection and power allocation for physical layer security,” inProc. 8th Int. Symp. Telecommun. (IST), Sep. 2016, pp. 374–379.

[37] C. Anagnostopoulos and S. Hadjiefthymiades, “Delay-tolerant deliveryof quality information in ad hoc networks,” J. Parallel Distrib. Comput.,vol. 71, no. 7, pp. 974–987, 2011.

[38] E. Zitzler, M. Laumanns, and L. Thiele, “SPEA2: Improving the strengthPareto evolutionary algorithm,” Dept. Elect. Eng., ETH Zurich, Zurich,Switzerland, Tech. Rep. TIK-103, 2001.

[39] K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan, “A fast and elitistmultiobjective genetic algorithm: NSGA-II,” IEEE Trans. Evol. Comput.,vol. 6, no. 2, pp. 182–197, Apr. 2002.

Penglin Dai received the B.S. degree in mathemat-ics and applied mathematics and the Ph.D. degreein computer science from Chongqing University,Chongqing, China, in 2012 and 2017, respectively.He is currently an Assistant Professor with theSchool of Information Science and Technology,Southwest Jiaotong University, Chengdu, China. Hisresearch interests include intelligent transportationsystems and vehicular cyber-physical systems.

Page 14: Temporal Information Services in Large-Scale Vehicular Networks … · 2019-03-18 · Physical Systems Research) under Grant 14-824-09-013, in part by the GRL Program through NRF

DAI et al.: TEMPORAL INFORMATION SERVICES IN LARGE-SCALE VEHICULAR NETWORKS 231

Kai Liu (S’07–M’12) received the Ph.D. degreein computer science from City University ofHong Kong in 2011. From 2010 to 2011, he was aVisiting Scholar with the Department of ComputerScience, University of Virginia, Charlottesville, VA,USA. From 2011 to 2014, he was a Post-DoctoralFellow with Nanyang Technological University,Singapore, the City University of Hong Kong, andHong Kong Baptist University. He is currently anAssistant Professor with the College of ComputerScience, Chongqing University, Chongqing, China.

His research interests include mobile computing, vehicular networks, andintelligent transportation systems.

Liang Feng received the Ph.D. degree in computa-tional intelligence and machine learning from theSchool of Computer Engineering, Nanyang Tech-nological University, Singapore, in 2014. He wasa Post-Doctoral Research Fellow with the Compu-tational Intelligence Graduate Laboratory, NanyangTechnological University. He is currently anAssistant Professor with the College of ComputerScience, Chongqing University, Chongqing, China.His research interests include computational andartificial intelligence, memetic computing, big dataoptimization, learning, and transfer learning.

Haijun Zhang (M’13) received the B.Eng. andmaster’s degrees from Northeastern University,Shenyang, China, in 2004 and 2007, respectively,and the Ph.D. degree from the Department of Elec-tronic Engineering, City University of Hong Kong,Hong Kong, in 2010. He was a Post-DoctoralResearch Fellow with the Department of Electricaland Computer Engineering, University of Windsor,Windsor, ON, Canada, from 2010 to 2011. Since2012, he has been with the Shenzhen Gradu-ate School, Harbin Institute of Technology, China,

where he is currently an Associate Professor of computer science. His currentresearch interests include neural networks, multimedia data mining, machinelearning, computational advertising, and evolutionary computing.

Victor Chung Sing Lee (M’92) received the Ph.D.degree in computer science from City Universityof Hong Kong, Hong Kong, in 1997. He is cur-rently an Assistant Professor with the Department ofComputer Science, City University of Hong Kong.His research interests include data disseminationin vehicular networks, real-time databases, andperformance evaluation. He is a member of theACM and the IEEE Computer Society. He was theChair of the IEEE Hong Kong Section ComputerChapter (2006–2007).

Sang Hyuk Son (M’85–SM’98–F’13) received theB.S. degree in electronics engineering from SeoulNational University, the M.S. degree from KAIST,and the Ph.D. degree in computer science fromUniversity of Maryland at College Park. He hasbeen a Professor with the Computer Science Depart-ment, University of Virginia, and a WCU ChairProfessor with Sogang University. He has been aVisiting Professor with KAIST, the City Universityof Hong Kong, the École Centrale de Lille, France,Linköping University, Sweden, and the University

of Skövde, Sweden. He is currently the President of the Daegu GyeongbukInstitute of Science and Technology.

His research interests include cyber physical systems, real-time and embed-ded systems, database and data services, and wireless sensor networks. Hehas authored or co-authored over 340 papers and edited/authored four booksin these areas. His research has been funded by the Korean Government,National Research Foundation, National Science Foundation, DARPA, Officeof Naval Research, Department of Energy, National Security Agency, andIBM. He is a member of the Korean Academy of Science and Technologyand the National Academy of Engineering of Korea. He is a FoundingMember of the ACM/IEEE CPS Week, and serving as a member of theSteering Committee for the IEEE RTCSA and Cyber Physical SystemsWeek. He received the Outstanding Contribution Award from the CyberPhysical Systems Week in 2012. He has served on the Editorial Boardof ACM Transactions on Cyber Physical Systems, IEEE TRANSACTIONS

ON COMPUTERS, IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED

SYSTEMS, and Real-Time Systems Journal.

Xiao Wu (S’05–M’08) received the B.Eng. and M.S.degrees in computer science from Yunnan Univer-sity, Yunnan, China, in 1999 and 2002, respectively,and the Ph.D. degree in computer science from CityUniversity of Hong Kong, Hong Kong, in 2008.

He was with the Institute of Software, ChineseAcademy of Sciences, Beijing, China, from 2001 to2002. He was a Research Assistant and a SeniorResearch Associate with the City University ofHong Kong from 2003 to 2004 and from 2007 to2009, respectively. From 2006 to 2007, he was with

the School of Computer Science, Carnegie Mellon University, Pittsburgh,PA, USA, as a Visiting Scholar, and with the School of Information andComputer Science, University of California at Irvine, Irvine, CA, USA, as aVisiting Associate Professor from 2015 to 2016. He is currently a Professorwith Southwest Jiaotong University, Chengdu, China. He is the AssistantDean of the School of Information Science and Technology, and the Head ofthe Department of Computer Science and Technology. His research interestsinclude multimedia information retrieval, image/video computing, and datamining. He received the second prize of Natural Science Award from theMinistry of Education, China, in 2016.